diff --git a/.github/workflows/create-tag.yml b/.github/workflows/create-tag.yml
index b80a539ae..064dc9688 100644
--- a/.github/workflows/create-tag.yml
+++ b/.github/workflows/create-tag.yml
@@ -22,6 +22,7 @@ on:
           - cron
           - database
           - email
+          - eval
           - harness
           - hermes
           - http
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 268f53ac1..274ab64d2 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -17,6 +17,7 @@ on:
       - 'cron/v*'
       - 'database/v*'
       - 'email/v*'
+      - 'eval/v*'
       - 'harness/v*'
       - 'hermes/v*'
       - 'http/v*'
diff --git a/README.md b/README.md
index 469e6a0c7..f76e1c965 100644
--- a/README.md
+++ b/README.md
@@ -41,6 +41,7 @@ npx skills add iii-hq/iii --all
 | [`acp`](acp/) | Rust | Agent Client Protocol surface — stdio JSON-RPC, exposes iii agents as ACP sessions. |
 | [`approval-gate`](approval-gate/) | Rust | Human-in-the-loop approval gate — evaluates each function call (continue / deny / hold), holds pending calls for a human, and emits `approval::pending-*` events. Binds the harness `pre_trigger` hook. See [`approval-gate/architecture/`](approval-gate/architecture/). |
 | [`harness`](harness/) | Node | TS port of the iii harness stack — bundles `harness` (provider registry + credentials/settings/permissions via the `configuration` worker), `turn-orchestrator`, `hook-fanout`, `models-catalog`, the `provider-*` workers, `llm-budget`, and `context-compaction` as one pnpm monorepo. Approval is delegated to the standalone `approval-gate` worker via the `pre_trigger` hook. Conversations persist in `session-manager`. See [`harness/README.md`](harness/README.md). |
+| [`eval`](eval/) | Rust | Durable same-model A/B evaluation for prompts and system prompts — runs paired harness sessions, delegates correctness to iii evaluator functions, and reports pass rates with token, cost, latency, function-call, trace, and span metrics. |
 | [`codex`](codex/) | Rust | OpenAI Codex as an iii worker — `codex::*` spawn the codex CLI for headless turns, mirror raw thread events onto `codex::events`, and stream AgentEvent frames onto `agent::events`. |
 | [`grok`](grok/) | Rust | xAI Grok CLI as an iii worker — `grok::*` spawn the grok CLI for headless turns (`grok --print --output-format streaming-json`), mirror raw events onto `grok::events`, and stream AgentEvent frames onto `agent::events`. |
 | [`devin`](devin/) | Rust | Devin as an iii worker: `devin::run` drives the local devin CLI and streams AgentEvent frames onto `agent::events`, `devin::session::*` wrap the Devin cloud session lifecycle, and `devin::api` reaches any v3 endpoint. |
diff --git a/approval-gate/Cargo.lock b/approval-gate/Cargo.lock
index 46a733694..6101c32c9 100644
--- a/approval-gate/Cargo.lock
+++ b/approval-gate/Cargo.lock
@@ -545,7 +545,7 @@ dependencies = [
 
 [[package]]
 name = "harness"
-version = "1.5.2"
+version = "1.6.1"
 dependencies = [
  "anyhow",
  "async-trait",
diff --git a/approval-gate/iii.worker.yaml b/approval-gate/iii.worker.yaml
index be1da10a2..0ccb4618b 100644
--- a/approval-gate/iii.worker.yaml
+++ b/approval-gate/iii.worker.yaml
@@ -8,7 +8,7 @@ tags: [approval, human-in-the-loop, permissions, policy, security]
 description: Policy and decision surface for human-held function calls — pre_trigger gate, pending inbox, per-session permission settings, and two notification trigger types.
 
 dependencies:
-  state: "^0.21.2"
+  state: "^0.21.3"
   configuration: "^0.21.6"
   iii-directory: "^1.0.0"
   session-manager: "^1.0.0"
diff --git a/console/web/e2e/ui-send.spec.ts b/console/web/e2e/ui-send.spec.ts
index c628f8634..70c35176b 100644
--- a/console/web/e2e/ui-send.spec.ts
+++ b/console/web/e2e/ui-send.spec.ts
@@ -9,7 +9,8 @@ test('sends and renders a streamed turn through the Console', async ({
   const completed = stack.waitForTurnCompleted()
   await openSession(page, stack)
   const composer = page.getByLabel('message composer')
-  await composer.fill(stack.ready.message)
+  await composer.pressSequentially(stack.ready.message)
+  await expect(composer).toHaveText(stack.ready.message)
   await page.getByRole('button', { name: 'send message' }).click()
 
   await expect(
diff --git a/eval/Cargo.lock b/eval/Cargo.lock
new file mode 100644
index 000000000..1e0c7ff72
--- /dev/null
+++ b/eval/Cargo.lock
@@ -0,0 +1,2603 @@
+# This file is automatically @generated by Cargo.
+# It is not intended for manual editing.
+version = 4
+
+[[package]]
+name = "ahash"
+version = "0.8.12"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5a15f179cd60c4584b8a8c596927aadc462e27f2ca70c04e0071964a73ba7a75"
+dependencies = [
+ "cfg-if",
+ "getrandom 0.3.4",
+ "once_cell",
+ "serde",
+ "version_check",
+ "zerocopy",
+]
+
+[[package]]
+name = "aho-corasick"
+version = "1.1.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ddd31a130427c27518df266943a5308ed92d4b226cc639f5a8f1002816174301"
+dependencies = [
+ "memchr",
+]
+
+[[package]]
+name = "anstream"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "824a212faf96e9acacdbd09febd34438f8f711fb84e09a8916013cd7815ca28d"
+dependencies = [
+ "anstyle",
+ "anstyle-parse",
+ "anstyle-query",
+ "anstyle-wincon",
+ "colorchoice",
+ "is_terminal_polyfill",
+ "utf8parse",
+]
+
+[[package]]
+name = "anstyle"
+version = "1.0.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "940b3a0ca603d1eade50a4846a2afffd5ef57a9feac2c0e2ec2e14f9ead76000"
+
+[[package]]
+name = "anstyle-parse"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "52ce7f38b242319f7cabaa6813055467063ecdc9d355bbb4ce0c68908cd8130e"
+dependencies = [
+ "utf8parse",
+]
+
+[[package]]
+name = "anstyle-query"
+version = "1.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "40c48f72fd53cd289104fc64099abca73db4166ad86ea0b4341abe65af83dadc"
+dependencies = [
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "anstyle-wincon"
+version = "3.0.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "291e6a250ff86cd4a820112fb8898808a366d8f9f58ce16d1f538353ad55747d"
+dependencies = [
+ "anstyle",
+ "once_cell_polyfill",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "anyhow"
+version = "1.0.104"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "330a5ed07fa54e4702c9d6c4174f74427fc0ef6e214bbd677ae50a5099946470"
+
+[[package]]
+name = "async-trait"
+version = "0.1.91"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ae36dc4177970ef04fde5178d3e2429882def40e57a451f919c098f72baa6cec"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 3.0.3",
+]
+
+[[package]]
+name = "atomic-waker"
+version = "1.1.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1505bd5d3d116872e7271a6d4e16d81d0c8570876c8de68093a09ac269d8aac0"
+
+[[package]]
+name = "autocfg"
+version = "1.5.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f2032f911046de80f0a198e0901378627c33f59ea0ac00e363d481118bd70a53"
+
+[[package]]
+name = "base64"
+version = "0.22.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "72b3254f16251a8381aa12e40e3c4d2f0199f8c6508fbecb9d91f575e0fbb8c6"
+
+[[package]]
+name = "bit-set"
+version = "0.5.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0700ddab506f33b20a03b13996eccd309a48e5ff77d0d95926aa0210fb4e95f1"
+dependencies = [
+ "bit-vec",
+]
+
+[[package]]
+name = "bit-vec"
+version = "0.6.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "349f9b6a179ed607305526ca489b34ad0a41aed5f7980fa90eb03160b69598fb"
+
+[[package]]
+name = "bitflags"
+version = "2.13.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b588b76d00fde79687d7646a9b5bdf3cc0f655e0bbd080335a95d7e96f3587da"
+
+[[package]]
+name = "block-buffer"
+version = "0.10.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3078c7629b62d3f0439517fa394996acacc5cbc91c5a20d8c658e77abd503a71"
+dependencies = [
+ "generic-array",
+]
+
+[[package]]
+name = "bstr"
+version = "1.13.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1f7dc094d718f2e1c1559ad110e27eeaae14a5465d3d56dd6dbd793079fbd530"
+dependencies = [
+ "memchr",
+ "serde_core",
+]
+
+[[package]]
+name = "bumpalo"
+version = "3.20.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "72f5acc6cb2ba439de613abc23857ec3d78374d8ed5ac84e9d11336e87da8649"
+
+[[package]]
+name = "bytecount"
+version = "0.6.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "175812e0be2bccb6abe50bb8d566126198344f707e304f45c648fd8f2cc0365e"
+
+[[package]]
+name = "bytes"
+version = "1.12.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fc652a48c352aef3ea3aed32080501cf3ef6ed5da78602a020c991775b0aff04"
+
+[[package]]
+name = "cc"
+version = "1.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c89588d05638b5b4594a3348a2d6c20277e43a7f5c5202b05cc56888475a47b8"
+dependencies = [
+ "find-msvc-tools",
+ "shlex",
+]
+
+[[package]]
+name = "cfg-if"
+version = "1.0.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9330f8b2ff13f34540b44e946ef35111825727b38d33286ef986142615121801"
+
+[[package]]
+name = "cfg_aliases"
+version = "0.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f079e83a288787bcd14a6aea84cee5c87a67c5a3e660c30f557a3d24761b3527"
+
+[[package]]
+name = "chacha20"
+version = "0.10.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d524456ba66e72eb8b115ff89e01e497f8e6d11d78b70b1aa13c0fbd97540a81"
+dependencies = [
+ "cfg-if",
+ "cpufeatures 0.3.0",
+ "rand_core 0.10.1",
+]
+
+[[package]]
+name = "clap"
+version = "4.6.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d91e0c145792ef73a6ad36d27c75ac09f1832222a3c209689d90f534685ee5b7"
+dependencies = [
+ "clap_builder",
+ "clap_derive",
+]
+
+[[package]]
+name = "clap_builder"
+version = "4.6.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f09628afdcc538b57f3c6341e9c8e9970f18e4a481690a64974d7023bd33548b"
+dependencies = [
+ "anstream",
+ "anstyle",
+ "clap_lex",
+ "strsim",
+]
+
+[[package]]
+name = "clap_derive"
+version = "4.6.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d012d2b9d65aca7f18f4d9878a045bc17899bba951561ba5ec3c2ba1eed9a061"
+dependencies = [
+ "heck",
+ "proc-macro2",
+ "quote",
+ "syn 3.0.3",
+]
+
+[[package]]
+name = "clap_lex"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c8d4a3bb8b1e0c1050499d1815f5ab16d04f0959b233085fb31653fbfc9d98f9"
+
+[[package]]
+name = "colorchoice"
+version = "1.0.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1d07550c9036bf2ae0c684c4297d503f838287c83c53686d05370d0e139ae570"
+
+[[package]]
+name = "core-foundation"
+version = "0.10.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b2a6cd9ae233e7f62ba4e9353e81a88df7fc8a5987b8d445b4d90c879bd156f6"
+dependencies = [
+ "core-foundation-sys",
+ "libc",
+]
+
+[[package]]
+name = "core-foundation-sys"
+version = "0.8.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "773648b94d0e5d620f64f280777445740e61fe701025087ec8b57f45c791888b"
+
+[[package]]
+name = "cpufeatures"
+version = "0.2.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "59ed5838eebb26a2bb2e58f6d5b5316989ae9d08bab10e0e6d103e656d1b0280"
+dependencies = [
+ "libc",
+]
+
+[[package]]
+name = "cpufeatures"
+version = "0.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8b2a41393f66f16b0823bb79094d54ac5fbd34ab292ddafb9a0456ac9f87d201"
+dependencies = [
+ "libc",
+]
+
+[[package]]
+name = "crypto-common"
+version = "0.1.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "78c8292055d1c1df0cce5d180393dc8cce0abec0a7102adb6c7b1eef6016d60a"
+dependencies = [
+ "generic-array",
+ "typenum",
+]
+
+[[package]]
+name = "data-encoding"
+version = "2.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a4ae5f15dda3c708c0ade84bfee31ccab44a3da4f88015ed22f63732abe300c8"
+
+[[package]]
+name = "deranged"
+version = "0.5.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7cd812cc2bc1d69d4764bd80df88b4317eaef9e773c75226407d9bc0876b211c"
+
+[[package]]
+name = "digest"
+version = "0.10.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9ed9a281f7bc9b7576e61468ba615a66a5c8cfdff42420a70aa82701a3b1e292"
+dependencies = [
+ "block-buffer",
+ "crypto-common",
+]
+
+[[package]]
+name = "displaydoc"
+version = "0.2.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1ac70aa55017e108007fbaf5aa0f54b021c98f92ff8af59d42eda9da96e3dd4f"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "dyn-clone"
+version = "1.0.20"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d0881ea181b1df73ff77ffaaf9c7544ecc11e82fba9b5f27b262a3c73a332555"
+
+[[package]]
+name = "equivalent"
+version = "1.0.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "877a4ace8713b0bcf2a4e7eec82529c029f1d0619886d18145fea96c3ffe5c0f"
+
+[[package]]
+name = "errno"
+version = "0.3.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "39cab71617ae0d63f51a36d69f866391735b51691dbda63cf6f96d042b63efeb"
+dependencies = [
+ "libc",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "eval"
+version = "0.1.0"
+dependencies = [
+ "anyhow",
+ "async-trait",
+ "clap",
+ "harness",
+ "iii-console-ui",
+ "iii-helpers",
+ "iii-sdk",
+ "schemars",
+ "serde",
+ "serde_json",
+ "serde_yaml",
+ "sha2",
+ "thiserror",
+ "tokio",
+ "tracing",
+ "tracing-subscriber",
+ "uuid",
+]
+
+[[package]]
+name = "fancy-regex"
+version = "0.13.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "531e46835a22af56d1e3b66f04844bed63158bc094a628bec1d321d9b4c44bf2"
+dependencies = [
+ "bit-set",
+ "regex-automata",
+ "regex-syntax",
+]
+
+[[package]]
+name = "find-msvc-tools"
+version = "0.1.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5baebc0774151f905a1a2cc41989300b1e6fbb29aff0ceffa1064fdd3088d582"
+
+[[package]]
+name = "form_urlencoded"
+version = "1.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cb4cb245038516f5f85277875cdaa4f7d2c9a0fa0468de06ed190163b1581fcf"
+dependencies = [
+ "percent-encoding",
+]
+
+[[package]]
+name = "fraction"
+version = "0.15.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e076045bb43dac435333ed5f04caf35c7463631d0dae2deb2638d94dd0a5b872"
+dependencies = [
+ "lazy_static",
+ "num",
+]
+
+[[package]]
+name = "futures-channel"
+version = "0.3.33"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "262590f4fe6afeb0bc83be1daa64e52657fe185690a958af7f3ad0e92085c5ae"
+dependencies = [
+ "futures-core",
+]
+
+[[package]]
+name = "futures-core"
+version = "0.3.33"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2cd50c473c80f6d7c3670a752354b8e569b1a7cbfdc0419ec88e5edad85e0dc7"
+
+[[package]]
+name = "futures-executor"
+version = "0.3.33"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6754879cc9f2c66f88c6e5c35344bb0bdb0708b0352b1201815667c7eabc7458"
+dependencies = [
+ "futures-core",
+ "futures-task",
+ "futures-util",
+]
+
+[[package]]
+name = "futures-macro"
+version = "0.3.33"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2d6d3cde68c518367be28956066ddfef33813991b77a55005a69dae04bf3b10b"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "futures-sink"
+version = "0.3.33"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e34418ac499d6305c2fb5ad0ed2f6ac998c5f8ca209b4510f7f94242c647e307"
+
+[[package]]
+name = "futures-task"
+version = "0.3.33"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b231ed28831efb4a61a08580c4bc233ec56bc009f4cd8f52da2c3cb97df0c109"
+
+[[package]]
+name = "futures-util"
+version = "0.3.33"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a77a90a256fce34da66415271e30f94ee91c57b04b8a2c042d9cf3220179deaa"
+dependencies = [
+ "futures-core",
+ "futures-macro",
+ "futures-sink",
+ "futures-task",
+ "pin-project-lite",
+ "slab",
+]
+
+[[package]]
+name = "generic-array"
+version = "0.14.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "85649ca51fd72272d7821adaf274ad91c288277713d9c18820d8499a7ff69e9a"
+dependencies = [
+ "typenum",
+ "version_check",
+]
+
+[[package]]
+name = "getrandom"
+version = "0.2.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ff2abc00be7fca6ebc474524697ae276ad847ad0a6b3faa4bcb027e9a4614ad0"
+dependencies = [
+ "cfg-if",
+ "js-sys",
+ "libc",
+ "wasi",
+ "wasm-bindgen",
+]
+
+[[package]]
+name = "getrandom"
+version = "0.3.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "899def5c37c4fd7b2664648c28120ecec138e4d395b459e5ca34f9cce2dd77fd"
+dependencies = [
+ "cfg-if",
+ "libc",
+ "r-efi 5.3.0",
+ "wasip2",
+]
+
+[[package]]
+name = "getrandom"
+version = "0.4.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "300e883d756b2e4ec94e02791f39b04b522276138852cfc41d9fb7e904106099"
+dependencies = [
+ "cfg-if",
+ "js-sys",
+ "libc",
+ "r-efi 6.0.0",
+ "rand_core 0.10.1",
+ "wasm-bindgen",
+]
+
+[[package]]
+name = "globset"
+version = "0.4.19"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e47d37d2ae4464254884b60ab7071be2b876a9c35b696bd018ddcc76847309cd"
+dependencies = [
+ "aho-corasick",
+ "bstr",
+ "log",
+ "regex-automata",
+ "regex-syntax",
+]
+
+[[package]]
+name = "harness"
+version = "1.6.1"
+dependencies = [
+ "anyhow",
+ "async-trait",
+ "clap",
+ "globset",
+ "iii-helpers",
+ "iii-sdk",
+ "jsonschema",
+ "schemars",
+ "serde",
+ "serde_json",
+ "serde_yaml",
+ "thiserror",
+ "tokio",
+ "tracing",
+ "tracing-subscriber",
+ "uuid",
+]
+
+[[package]]
+name = "hashbrown"
+version = "0.17.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ed5909b6e89a2db4456e54cd5f673791d7eca6732202bbf2a9cc504fe2f9b84a"
+
+[[package]]
+name = "heck"
+version = "0.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2304e00983f87ffb38b55b444b5e3b60a884b5d30c0fca7d82fe33449bbe55ea"
+
+[[package]]
+name = "hostname"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "617aaa3557aef3810a6369d0a99fac8a080891b68bd9f9812a1eeda0c0730cbd"
+dependencies = [
+ "cfg-if",
+ "libc",
+ "windows-link",
+]
+
+[[package]]
+name = "http"
+version = "1.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6970f50e31d6fc17d3fa27329444bfa74e196cf62e95052a3f6fee181dba6425"
+dependencies = [
+ "bytes",
+ "itoa",
+]
+
+[[package]]
+name = "http-body"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ca2a8f2913ee65f60facd6a5905613afaa448497a0230cc41ce022d93290bc2c"
+dependencies = [
+ "bytes",
+ "http",
+]
+
+[[package]]
+name = "http-body-util"
+version = "0.1.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e9f41fd6a08e4d4ec69df65976da761afd5ad5e58a9d4acb46bd1c953a9e3ff2"
+dependencies = [
+ "bytes",
+ "futures-core",
+ "http",
+ "http-body",
+ "pin-project-lite",
+]
+
+[[package]]
+name = "httparse"
+version = "1.10.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6dbf3de79e51f3d586ab4cb9d5c3e2c14aa28ed23d180cf89b4df0454a69cc87"
+
+[[package]]
+name = "hyper"
+version = "1.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d22053281f852e11534f5198498373cbb59295120a20771d90f7ed1897490a72"
+dependencies = [
+ "atomic-waker",
+ "bytes",
+ "futures-channel",
+ "futures-core",
+ "http",
+ "http-body",
+ "httparse",
+ "itoa",
+ "pin-project-lite",
+ "smallvec",
+ "tokio",
+ "want",
+]
+
+[[package]]
+name = "hyper-rustls"
+version = "0.27.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "33ca68d021ef39cf6463ab54c1d0f5daf03377b70561305bb89a8f83aab66e0f"
+dependencies = [
+ "http",
+ "hyper",
+ "hyper-util",
+ "rustls",
+ "tokio",
+ "tokio-rustls",
+ "tower-service",
+ "webpki-roots",
+]
+
+[[package]]
+name = "hyper-util"
+version = "0.1.20"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "96547c2556ec9d12fb1578c4eaf448b04993e7fb79cbaad930a656880a6bdfa0"
+dependencies = [
+ "base64",
+ "bytes",
+ "futures-channel",
+ "futures-util",
+ "http",
+ "http-body",
+ "hyper",
+ "ipnet",
+ "libc",
+ "percent-encoding",
+ "pin-project-lite",
+ "socket2",
+ "tokio",
+ "tower-service",
+ "tracing",
+]
+
+[[package]]
+name = "icu_collections"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2984d1cd16c883d7935b9e07e44071dca8d917fd52ecc02c04d5fa0b5a3f191c"
+dependencies = [
+ "displaydoc",
+ "potential_utf",
+ "utf8_iter",
+ "yoke",
+ "zerofrom",
+ "zerovec",
+]
+
+[[package]]
+name = "icu_locale_core"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "92219b62b3e2b4d88ac5119f8904c10f8f61bf7e95b640d25ba3075e6cac2c29"
+dependencies = [
+ "displaydoc",
+ "litemap",
+ "tinystr",
+ "writeable",
+ "zerovec",
+]
+
+[[package]]
+name = "icu_normalizer"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c56e5ee99d6e3d33bd91c5d85458b6005a22140021cc324cea84dd0e72cff3b4"
+dependencies = [
+ "icu_collections",
+ "icu_normalizer_data",
+ "icu_properties",
+ "icu_provider",
+ "smallvec",
+ "zerovec",
+]
+
+[[package]]
+name = "icu_normalizer_data"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "da3be0ae77ea334f4da67c12f149704f19f81d1adf7c51cf482943e84a2bad38"
+
+[[package]]
+name = "icu_properties"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bee3b67d0ea5c2cca5003417989af8996f8604e34fb9ddf96208a033901e70de"
+dependencies = [
+ "icu_collections",
+ "icu_locale_core",
+ "icu_properties_data",
+ "icu_provider",
+ "zerotrie",
+ "zerovec",
+]
+
+[[package]]
+name = "icu_properties_data"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8e2bbb201e0c04f7b4b3e14382af113e17ba4f63e2c9d2ee626b720cbce54a14"
+
+[[package]]
+name = "icu_provider"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "139c4cf31c8b5f33d7e199446eff9c1e02decfc2f0eec2c8d71f65befa45b421"
+dependencies = [
+ "displaydoc",
+ "icu_locale_core",
+ "writeable",
+ "yoke",
+ "zerofrom",
+ "zerotrie",
+ "zerovec",
+]
+
+[[package]]
+name = "idna"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3b0875f23caa03898994f6ddc501886a45c7d3d62d04d2d90788d47be1b1e4de"
+dependencies = [
+ "idna_adapter",
+ "smallvec",
+ "utf8_iter",
+]
+
+[[package]]
+name = "idna_adapter"
+version = "1.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cb68373c0d6620ef8105e855e7745e18b0d00d3bdb07fb532e434244cdb9a714"
+dependencies = [
+ "icu_normalizer",
+ "icu_properties",
+]
+
+[[package]]
+name = "iii-console-ui"
+version = "0.1.0"
+dependencies = [
+ "iii-sdk",
+ "schemars",
+ "serde",
+ "serde_json",
+ "tokio",
+ "tracing",
+]
+
+[[package]]
+name = "iii-helpers"
+version = "0.21.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c0d84d5c149ae4404365a79feca28aa66f6a7dbed56423b4b8c4e2421e0b5add"
+dependencies = [
+ "futures-util",
+ "opentelemetry",
+ "opentelemetry-http",
+ "opentelemetry_sdk",
+ "reqwest",
+ "schemars",
+ "serde",
+ "serde_json",
+ "sysinfo",
+ "tokio",
+ "tokio-tungstenite",
+ "tracing",
+ "uuid",
+]
+
+[[package]]
+name = "iii-sdk"
+version = "0.21.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "07dd060fddcc9153b0dd07c038a14cf172ce15ce1d4edb98155563ed55b2caba"
+dependencies = [
+ "async-trait",
+ "futures-util",
+ "hostname",
+ "iii-helpers",
+ "reqwest",
+ "schemars",
+ "serde",
+ "serde_json",
+ "thiserror",
+ "tokio",
+ "tokio-tungstenite",
+ "tracing",
+ "uuid",
+]
+
+[[package]]
+name = "indexmap"
+version = "2.14.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d466e9454f08e4a911e14806c24e16fba1b4c121d1ea474396f396069cf949d9"
+dependencies = [
+ "equivalent",
+ "hashbrown",
+]
+
+[[package]]
+name = "ipnet"
+version = "2.12.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d98f6fed1fde3f8c21bc40a1abb88dd75e67924f9cffc3ef95607bad8017f8e2"
+
+[[package]]
+name = "is_terminal_polyfill"
+version = "1.70.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a6cb138bb79a146c1bd460005623e142ef0181e3d0219cb493e02f7d08a35695"
+
+[[package]]
+name = "iso8601"
+version = "0.6.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e1082f0c48f143442a1ac6122f67e360ceee130b967af4d50996e5154a45df46"
+dependencies = [
+ "nom",
+]
+
+[[package]]
+name = "itoa"
+version = "1.0.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8f42a60cbdf9a97f5d2305f08a87dc4e09308d1276d28c869c684d7777685682"
+
+[[package]]
+name = "js-sys"
+version = "0.3.103"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "53b44bfcdb3f8d5837a46dae1ca9660a837176eee74a28b229bc626816589102"
+dependencies = [
+ "cfg-if",
+ "futures-util",
+ "wasm-bindgen",
+]
+
+[[package]]
+name = "jsonschema"
+version = "0.18.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fa0f4bea31643be4c6a678e9aa4ae44f0db9e5609d5ca9dc9083d06eb3e9a27a"
+dependencies = [
+ "ahash",
+ "anyhow",
+ "base64",
+ "bytecount",
+ "fancy-regex",
+ "fraction",
+ "getrandom 0.2.17",
+ "iso8601",
+ "itoa",
+ "memchr",
+ "num-cmp",
+ "once_cell",
+ "parking_lot",
+ "percent-encoding",
+ "regex",
+ "serde",
+ "serde_json",
+ "time",
+ "url",
+ "uuid",
+]
+
+[[package]]
+name = "lazy_static"
+version = "1.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bbd2bcb4c963f2ddae06a2efc7e9f3591312473c50c6685e1f298068316e66fe"
+
+[[package]]
+name = "libc"
+version = "0.2.189"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3eaf3ede3fee6db1a4c2ee091bf8a8b4dccdc6d17f656fb07896ee72867612f2"
+
+[[package]]
+name = "litemap"
+version = "0.8.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "92daf443525c4cce67b150400bc2316076100ce0b3686209eb8cf3c31612e6f0"
+
+[[package]]
+name = "lock_api"
+version = "0.4.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "224399e74b87b5f3557511d98dff8b14089b3dadafcab6bb93eab67d3aace965"
+dependencies = [
+ "scopeguard",
+]
+
+[[package]]
+name = "log"
+version = "0.4.33"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0ceec5bc11778974d1bcb055b18002eba7f4b3518b6a0081b3af5f21666da9ad"
+
+[[package]]
+name = "lru-slab"
+version = "0.1.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "112b39cec0b298b6c1999fee3e31427f74f676e4cb9879ed1a121b43661a4154"
+
+[[package]]
+name = "matchers"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d1525a2a28c7f4fa0fc98bb91ae755d1e2d1505079e05539e35bc876b5d65ae9"
+dependencies = [
+ "regex-automata",
+]
+
+[[package]]
+name = "memchr"
+version = "2.8.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cf8baf1c55e62ffcace7a9f06f4bd9cd3f0c4beb022d3b367256b91b87513d98"
+
+[[package]]
+name = "mio"
+version = "1.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "30d65c71f1ce40ab09135ce117d742b9f8a19ff91a41a8b57ed50bc2de59c427"
+dependencies = [
+ "libc",
+ "wasi",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "nom"
+version = "8.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "df9761775871bdef83bee530e60050f7e54b1105350d6884eb0fb4f46c2f9405"
+dependencies = [
+ "memchr",
+]
+
+[[package]]
+name = "ntapi"
+version = "0.4.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c3b335231dfd352ffb0f8017f3b6027a4917f7df785ea2143d8af2adc66980ae"
+dependencies = [
+ "winapi",
+]
+
+[[package]]
+name = "nu-ansi-term"
+version = "0.50.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7957b9740744892f114936ab4a57b3f487491bbeafaf8083688b16841a4240e5"
+dependencies = [
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "num"
+version = "0.4.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "35bd024e8b2ff75562e5f34e7f4905839deb4b22955ef5e73d2fea1b9813cb23"
+dependencies = [
+ "num-bigint",
+ "num-complex",
+ "num-integer",
+ "num-iter",
+ "num-rational",
+ "num-traits",
+]
+
+[[package]]
+name = "num-bigint"
+version = "0.4.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c89e69e7e0f03bea5ef08013795c25018e101932225a656383bd384495ecc367"
+dependencies = [
+ "num-integer",
+ "num-traits",
+]
+
+[[package]]
+name = "num-cmp"
+version = "0.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "63335b2e2c34fae2fb0aa2cecfd9f0832a1e24b3b32ecec612c3426d46dc8aaa"
+
+[[package]]
+name = "num-complex"
+version = "0.4.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "73f88a1307638156682bada9d7604135552957b7818057dcef22705b4d509495"
+dependencies = [
+ "num-traits",
+]
+
+[[package]]
+name = "num-conv"
+version = "0.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "521739c6d2bac4aa25192232afe6841231376b2b26d4d9fae5ecf8ca5772e441"
+
+[[package]]
+name = "num-integer"
+version = "0.1.46"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7969661fd2958a5cb096e56c8e1ad0444ac2bbcd0061bd28660485a44879858f"
+dependencies = [
+ "num-traits",
+]
+
+[[package]]
+name = "num-iter"
+version = "0.1.46"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c92800bd69a1eac91786bcfe9da64a897eb72911b8dc3095decbd07429e8048b"
+dependencies = [
+ "num-integer",
+ "num-traits",
+]
+
+[[package]]
+name = "num-rational"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f83d14da390562dca69fc84082e73e548e1ad308d24accdedd2720017cb37824"
+dependencies = [
+ "num-bigint",
+ "num-integer",
+ "num-traits",
+]
+
+[[package]]
+name = "num-traits"
+version = "0.2.19"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "071dfc062690e90b734c0b2273ce72ad0ffa95f0c74596bc250dcfd960262841"
+dependencies = [
+ "autocfg",
+]
+
+[[package]]
+name = "objc2-core-foundation"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2a180dd8642fa45cdb7dd721cd4c11b1cadd4929ce112ebd8b9f5803cc79d536"
+dependencies = [
+ "bitflags",
+]
+
+[[package]]
+name = "objc2-io-kit"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "33fafba39597d6dc1fb709123dfa8289d39406734be322956a69f0931c73bb15"
+dependencies = [
+ "libc",
+ "objc2-core-foundation",
+]
+
+[[package]]
+name = "once_cell"
+version = "1.21.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9f7c3e4beb33f85d45ae3e3a1792185706c8e16d043238c593331cc7cd313b50"
+
+[[package]]
+name = "once_cell_polyfill"
+version = "1.70.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "384b8ab6d37215f3c5301a95a4accb5d64aa607f1fcb26a11b5303878451b4fe"
+
+[[package]]
+name = "openssl-probe"
+version = "0.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7c87def4c32ab89d880effc9e097653c8da5d6ef28e6b539d313baaacfbafcbe"
+
+[[package]]
+name = "opentelemetry"
+version = "0.31.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b84bcd6ae87133e903af7ef497404dda70c60d0ea14895fc8a5e6722754fc2a0"
+dependencies = [
+ "futures-core",
+ "futures-sink",
+ "js-sys",
+ "pin-project-lite",
+ "thiserror",
+ "tracing",
+]
+
+[[package]]
+name = "opentelemetry-http"
+version = "0.31.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d7a6d09a73194e6b66df7c8f1b680f156d916a1a942abf2de06823dd02b7855d"
+dependencies = [
+ "async-trait",
+ "bytes",
+ "http",
+ "opentelemetry",
+ "reqwest",
+]
+
+[[package]]
+name = "opentelemetry_sdk"
+version = "0.31.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e14ae4f5991976fd48df6d843de219ca6d31b01daaab2dad5af2badeded372bd"
+dependencies = [
+ "futures-channel",
+ "futures-executor",
+ "futures-util",
+ "opentelemetry",
+ "percent-encoding",
+ "rand 0.9.5",
+ "thiserror",
+ "tokio",
+ "tokio-stream",
+]
+
+[[package]]
+name = "parking_lot"
+version = "0.12.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "93857453250e3077bd71ff98b6a65ea6621a19bb0f559a85248955ac12c45a1a"
+dependencies = [
+ "lock_api",
+ "parking_lot_core",
+]
+
+[[package]]
+name = "parking_lot_core"
+version = "0.9.12"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2621685985a2ebf1c516881c026032ac7deafcda1a2c9b7850dc81e3dfcb64c1"
+dependencies = [
+ "cfg-if",
+ "libc",
+ "redox_syscall",
+ "smallvec",
+ "windows-link",
+]
+
+[[package]]
+name = "percent-encoding"
+version = "2.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9b4f627cb1b25917193a259e49bdad08f671f8d9708acfd5fe0a8c1455d87220"
+
+[[package]]
+name = "pin-project-lite"
+version = "0.2.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a89322df9ebe1c1578d689c92318e070967d1042b512afbe49518723f4e6d5cd"
+
+[[package]]
+name = "potential_utf"
+version = "0.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0103b1cef7ec0cf76490e969665504990193874ea05c85ff9bab8b911d0a0564"
+dependencies = [
+ "zerovec",
+]
+
+[[package]]
+name = "powerfmt"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "439ee305def115ba05938db6eb1644ff94165c5ab5e9420d1c1bcedbba909391"
+
+[[package]]
+name = "ppv-lite86"
+version = "0.2.21"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "85eae3c4ed2f50dcfe72643da4befc30deadb458a9b590d720cde2f2b1e97da9"
+dependencies = [
+ "zerocopy",
+]
+
+[[package]]
+name = "proc-macro2"
+version = "1.0.107"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "985e7ec9bb745e6ce6535b544d84d6cd6f7ad8bd711c398938ae983b91a766d9"
+dependencies = [
+ "unicode-ident",
+]
+
+[[package]]
+name = "quinn"
+version = "0.11.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0c1a41e437b6bbd489372cd4971de128e85c855f56c57f283d20ff016cf7c0a8"
+dependencies = [
+ "bytes",
+ "cfg_aliases",
+ "pin-project-lite",
+ "quinn-proto",
+ "quinn-udp",
+ "rustc-hash",
+ "rustls",
+ "socket2",
+ "thiserror",
+ "tokio",
+ "tracing",
+ "web-time",
+]
+
+[[package]]
+name = "quinn-proto"
+version = "0.11.16"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2f4bfc015262b9df63c8845072ce59068853ff5872180c2ce2f13038b970e560"
+dependencies = [
+ "bytes",
+ "getrandom 0.4.3",
+ "lru-slab",
+ "rand 0.10.2",
+ "rand_pcg",
+ "ring",
+ "rustc-hash",
+ "rustls",
+ "rustls-pki-types",
+ "slab",
+ "thiserror",
+ "tinyvec",
+ "tracing",
+ "web-time",
+]
+
+[[package]]
+name = "quinn-udp"
+version = "0.5.15"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "35a133f956daabe89a61a685c2649f13d82d5aa4bd5d12d1277e1072a21c0694"
+dependencies = [
+ "cfg_aliases",
+ "libc",
+ "once_cell",
+ "socket2",
+ "tracing",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "quote"
+version = "1.0.47"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1fbf4db142a473a8d80c26bbf18454ed458bf8d26c8219c331daecfdbd079001"
+dependencies = [
+ "proc-macro2",
+]
+
+[[package]]
+name = "r-efi"
+version = "5.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "69cdb34c158ceb288df11e18b4bd39de994f6657d83847bdffdbd7f346754b0f"
+
+[[package]]
+name = "r-efi"
+version = "6.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f8dcc9c7d52a811697d2151c701e0d08956f92b0e24136cf4cf27b57a6a0d9bf"
+
+[[package]]
+name = "rand"
+version = "0.9.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b9ef1d0d795eb7d84685bca4f72f3649f064e6641543d3a8c415898726a57b41"
+dependencies = [
+ "rand_chacha",
+ "rand_core 0.9.5",
+]
+
+[[package]]
+name = "rand"
+version = "0.10.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c7f5fa3a058cd35567ef9bfa5e75732bee0f9e4c55fa90477bef2dfcdbc4be80"
+dependencies = [
+ "chacha20",
+ "getrandom 0.4.3",
+ "rand_core 0.10.1",
+]
+
+[[package]]
+name = "rand_chacha"
+version = "0.9.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d3022b5f1df60f26e1ffddd6c66e8aa15de382ae63b3a0c1bfc0e4d3e3f325cb"
+dependencies = [
+ "ppv-lite86",
+ "rand_core 0.9.5",
+]
+
+[[package]]
+name = "rand_core"
+version = "0.9.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "76afc826de14238e6e8c374ddcc1fa19e374fd8dd986b0d2af0d02377261d83c"
+dependencies = [
+ "getrandom 0.3.4",
+]
+
+[[package]]
+name = "rand_core"
+version = "0.10.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "63b8176103e19a2643978565ca18b50549f6101881c443590420e4dc998a3c69"
+
+[[package]]
+name = "rand_pcg"
+version = "0.10.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "caa0f4137e1c0a72f4c651489402276c8e8e1cf081f3b0ba156d2cbeef09e86a"
+dependencies = [
+ "rand_core 0.10.1",
+]
+
+[[package]]
+name = "redox_syscall"
+version = "0.5.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ed2bf2547551a7053d6fdfafda3f938979645c44812fbfcda098faae3f1a362d"
+dependencies = [
+ "bitflags",
+]
+
+[[package]]
+name = "regex"
+version = "1.13.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f020237b6c8eed93db2e2cb53c00c60a8e1bc73da7d073199a1180401450218d"
+dependencies = [
+ "aho-corasick",
+ "memchr",
+ "regex-automata",
+ "regex-syntax",
+]
+
+[[package]]
+name = "regex-automata"
+version = "0.4.16"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8fcfdb36bda0c880c5931cdc7a2bcdc8ba4556847b9d912bca70bc94708711ad"
+dependencies = [
+ "aho-corasick",
+ "memchr",
+ "regex-syntax",
+]
+
+[[package]]
+name = "regex-syntax"
+version = "0.8.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d6f6ff9a378485b298a5286656da665ba74413d36db0979633275d2e708145d4"
+
+[[package]]
+name = "reqwest"
+version = "0.12.28"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "eddd3ca559203180a307f12d114c268abf583f59b03cb906fd0b3ff8646c1147"
+dependencies = [
+ "base64",
+ "bytes",
+ "futures-core",
+ "http",
+ "http-body",
+ "http-body-util",
+ "hyper",
+ "hyper-rustls",
+ "hyper-util",
+ "js-sys",
+ "log",
+ "percent-encoding",
+ "pin-project-lite",
+ "quinn",
+ "rustls",
+ "rustls-pki-types",
+ "serde",
+ "serde_json",
+ "serde_urlencoded",
+ "sync_wrapper",
+ "tokio",
+ "tokio-rustls",
+ "tower",
+ "tower-http",
+ "tower-service",
+ "url",
+ "wasm-bindgen",
+ "wasm-bindgen-futures",
+ "web-sys",
+ "webpki-roots",
+]
+
+[[package]]
+name = "ring"
+version = "0.17.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a4689e6c2294d81e88dc6261c768b63bc4fcdb852be6d1352498b114f61383b7"
+dependencies = [
+ "cc",
+ "cfg-if",
+ "getrandom 0.2.17",
+ "libc",
+ "untrusted",
+ "windows-sys 0.52.0",
+]
+
+[[package]]
+name = "rustc-hash"
+version = "2.1.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6b1e7f9a428571be2dc5bc0505c13fb6bf936822b894ec87abf8a08a4e51742d"
+
+[[package]]
+name = "rustls"
+version = "0.23.42"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3c54fcab019b409d04215d3a17cb438fd7fbf192ee61461f20f4fe18704bc138"
+dependencies = [
+ "once_cell",
+ "ring",
+ "rustls-pki-types",
+ "rustls-webpki",
+ "subtle",
+ "zeroize",
+]
+
+[[package]]
+name = "rustls-native-certs"
+version = "0.8.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dab5152771c58876a2146916e53e35057e1a4dfa2b9df0f0305b07f611fdea4d"
+dependencies = [
+ "openssl-probe",
+ "rustls-pki-types",
+ "schannel",
+ "security-framework",
+]
+
+[[package]]
+name = "rustls-pki-types"
+version = "1.15.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2f4925028c7eb5d1fcdaf196971378ed9d2c1c4efc7dc5d011256f76c99c0a96"
+dependencies = [
+ "web-time",
+ "zeroize",
+]
+
+[[package]]
+name = "rustls-webpki"
+version = "0.103.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "61c429a8649f110dddef65e2a5ad240f747e85f7758a6bccc7e5777bd33f756e"
+dependencies = [
+ "ring",
+ "rustls-pki-types",
+ "untrusted",
+]
+
+[[package]]
+name = "rustversion"
+version = "1.0.23"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cf54715a573b99ac80df0bc206da022bcd442c974952c7b9720069370852e21f"
+
+[[package]]
+name = "ryu"
+version = "1.0.23"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9774ba4a74de5f7b1c1451ed6cd5285a32eddb5cccb8cc655a4e50009e06477f"
+
+[[package]]
+name = "schannel"
+version = "0.1.29"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "91c1b7e4904c873ef0710c1f407dde2e6287de2bebc1bbbf7d430bb7cbffd939"
+dependencies = [
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "schemars"
+version = "0.8.22"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3fbf2ae1b8bc8e02df939598064d22402220cd5bbcca1c76f7d6a310974d5615"
+dependencies = [
+ "dyn-clone",
+ "schemars_derive",
+ "serde",
+ "serde_json",
+]
+
+[[package]]
+name = "schemars_derive"
+version = "0.8.22"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "32e265784ad618884abaea0600a9adf15393368d840e0222d101a072f3f7534d"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "serde_derive_internals",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "scopeguard"
+version = "1.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "94143f37725109f92c262ed2cf5e59bce7498c01bcc1502d7b9afe439a4e9f49"
+
+[[package]]
+name = "security-framework"
+version = "3.7.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b7f4bc775c73d9a02cde8bf7b2ec4c9d12743edf609006c7facc23998404cd1d"
+dependencies = [
+ "bitflags",
+ "core-foundation",
+ "core-foundation-sys",
+ "libc",
+ "security-framework-sys",
+]
+
+[[package]]
+name = "security-framework-sys"
+version = "2.17.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6ce2691df843ecc5d231c0b14ece2acc3efb62c0a398c7e1d875f3983ce020e3"
+dependencies = [
+ "core-foundation-sys",
+ "libc",
+]
+
+[[package]]
+name = "serde"
+version = "1.0.229"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4148590afebada386688f18773da617792bf2ef03ffc1e4cbd2b1d45b023e0ba"
+dependencies = [
+ "serde_core",
+ "serde_derive",
+]
+
+[[package]]
+name = "serde_core"
+version = "1.0.229"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "67dca2c9c51e58a4791a4b1ed58308b39c64224d349a935ab5039aa360942a48"
+dependencies = [
+ "serde_derive",
+]
+
+[[package]]
+name = "serde_derive"
+version = "1.0.229"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e7a5d71263a5a7d47b41f6b3f06ba276f10cc18b0931f1799f710578e2309348"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 3.0.3",
+]
+
+[[package]]
+name = "serde_derive_internals"
+version = "0.29.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "18d26a20a969b9e3fdf2fc2d9f21eda6c40e2de84c9408bb5d3b05d499aae711"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "serde_json"
+version = "1.0.151"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c841b55ecdae098c80dcae9cf767f6f8a0c2cdb3416bbef72181df4d0fe73f14"
+dependencies = [
+ "itoa",
+ "memchr",
+ "serde",
+ "serde_core",
+ "zmij",
+]
+
+[[package]]
+name = "serde_urlencoded"
+version = "0.7.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d3491c14715ca2294c4d6a88f15e84739788c1d030eed8c110436aafdaa2f3fd"
+dependencies = [
+ "form_urlencoded",
+ "itoa",
+ "ryu",
+ "serde",
+]
+
+[[package]]
+name = "serde_yaml"
+version = "0.9.34+deprecated"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6a8b1a1a2ebf674015cc02edccce75287f1a0130d394307b36743c2f5d504b47"
+dependencies = [
+ "indexmap",
+ "itoa",
+ "ryu",
+ "serde",
+ "unsafe-libyaml",
+]
+
+[[package]]
+name = "sha1"
+version = "0.10.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a978451301f4db1d02937a4ab3ccce137717b81826e79b7d49ffe3244a13c3b8"
+dependencies = [
+ "cfg-if",
+ "cpufeatures 0.2.17",
+ "digest",
+]
+
+[[package]]
+name = "sha2"
+version = "0.10.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a7507d819769d01a365ab707794a4084392c824f54a7a6a7862f8c3d0892b283"
+dependencies = [
+ "cfg-if",
+ "cpufeatures 0.2.17",
+ "digest",
+]
+
+[[package]]
+name = "sharded-slab"
+version = "0.1.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f40ca3c46823713e0d4209592e8d6e826aa57e928f09752619fc696c499637f6"
+dependencies = [
+ "lazy_static",
+]
+
+[[package]]
+name = "shlex"
+version = "2.0.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f8fadd59c855ef2080decdef8ff161eb6661b86933c9d82e5ba29dc602a55aba"
+
+[[package]]
+name = "signal-hook-registry"
+version = "1.4.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c4db69cba1110affc0e9f7bcd48bbf87b3f4fc7c61fc9155afd4c469eb3d6c1b"
+dependencies = [
+ "errno",
+ "libc",
+]
+
+[[package]]
+name = "slab"
+version = "0.4.12"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0c790de23124f9ab44544d7ac05d60440adc586479ce501c1d6d7da3cd8c9cf5"
+
+[[package]]
+name = "smallvec"
+version = "1.15.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8ed6a63f02c8539c91a8685a86f4099661ba3da017932f6ebbea6de3f0fa7c90"
+
+[[package]]
+name = "socket2"
+version = "0.6.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c3d1e2c7f27f8d4cb10542a02c49005dbd6e93095799d6f3be745fae9f8fedd4"
+dependencies = [
+ "libc",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "stable_deref_trait"
+version = "1.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6ce2be8dc25455e1f91df71bfa12ad37d7af1092ae736f3a6cd0e37bc7810596"
+
+[[package]]
+name = "strsim"
+version = "0.11.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7da8b5736845d9f2fcb837ea5d9e2628564b3b043a70948a3f0b778838c5fb4f"
+
+[[package]]
+name = "subtle"
+version = "2.6.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "13c2bddecc57b384dee18652358fb23172facb8a2c51ccc10d74c157bdea3292"
+
+[[package]]
+name = "syn"
+version = "2.0.119"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "872831b642d1a07999a962a351ed35b955ea2cfc8f3862091e2a240a84f17297"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "unicode-ident",
+]
+
+[[package]]
+name = "syn"
+version = "3.0.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "53e9bae58849f64dfa4f5d5ae372c8341f7305f82a3868709269343628b659a3"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "unicode-ident",
+]
+
+[[package]]
+name = "sync_wrapper"
+version = "1.0.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0bf256ce5efdfa370213c1dabab5935a12e49f2c58d15e9eac2870d3b4f27263"
+dependencies = [
+ "futures-core",
+]
+
+[[package]]
+name = "synstructure"
+version = "0.13.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "728a70f3dbaf5bab7f0c4b1ac8d7ae5ea60a4b5549c8a5914361c99147a709d2"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "sysinfo"
+version = "0.38.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "92ab6a2f8bfe508deb3c6406578252e491d299cbbf3bc0529ecc3313aee4a52f"
+dependencies = [
+ "libc",
+ "memchr",
+ "ntapi",
+ "objc2-core-foundation",
+ "objc2-io-kit",
+ "windows",
+]
+
+[[package]]
+name = "thiserror"
+version = "2.0.19"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "09a43598840e33d5b0331f38c5e30d13bb11c11210a4b58f0d9b18a5a5eefcd9"
+dependencies = [
+ "thiserror-impl",
+]
+
+[[package]]
+name = "thiserror-impl"
+version = "2.0.19"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "43cbfe0cf76104d42a574802844187e84a305e531ed54455f11fbde0f10541cd"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 3.0.3",
+]
+
+[[package]]
+name = "thread_local"
+version = "1.1.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1ad99c4c6d32803332c548b1af0540b357b3f5fc0be8f6c6bfe8b2e6ae784070"
+dependencies = [
+ "cfg-if",
+]
+
+[[package]]
+name = "time"
+version = "0.3.54"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3e1d5e639ff6bab73cb6885cc7e7b1de96c3f32c68ec55f3952614bec1092244"
+dependencies = [
+ "deranged",
+ "num-conv",
+ "powerfmt",
+ "serde_core",
+ "time-core",
+ "time-macros",
+]
+
+[[package]]
+name = "time-core"
+version = "0.1.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9e1c906769ad99c88eaa54e728060edef082f8e358ff32030cb7c7d315e81109"
+
+[[package]]
+name = "time-macros"
+version = "0.2.32"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7e689342a48d2ea927c87ea50cabf8594854bf940e9310208848d680d668ed85"
+dependencies = [
+ "num-conv",
+ "time-core",
+]
+
+[[package]]
+name = "tinystr"
+version = "0.8.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c8323304221c2a851516f22236c5722a72eaa19749016521d6dff0824447d96d"
+dependencies = [
+ "displaydoc",
+ "zerovec",
+]
+
+[[package]]
+name = "tinyvec"
+version = "1.12.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bb4ebadaa0af04fab11ae01eb5f9fdb5f9c5b875506e210e71c07873528baa7f"
+dependencies = [
+ "tinyvec_macros",
+]
+
+[[package]]
+name = "tinyvec_macros"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1f3ccbac311fea05f86f61904b462b55fb3df8837a366dfc601a0161d0532f20"
+
+[[package]]
+name = "tokio"
+version = "1.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "202caea871b69668250d242070849eb495be178ed697a3e98aebce5bc81a0bed"
+dependencies = [
+ "bytes",
+ "libc",
+ "mio",
+ "pin-project-lite",
+ "signal-hook-registry",
+ "socket2",
+ "tokio-macros",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "tokio-macros"
+version = "2.7.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6328af13490e73a9b4694030fafd93f8c8c6a9dede33e821c3fc63eddf8042ba"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "tokio-rustls"
+version = "0.26.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1729aa945f29d91ba541258c8df89027d5792d85a8841fb65e8bf0f4ede4ef61"
+dependencies = [
+ "rustls",
+ "tokio",
+]
+
+[[package]]
+name = "tokio-stream"
+version = "0.1.19"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a3d06f0b082ba57c26b79407372e57cf2a1e28124f78e9479fe80322cf53420b"
+dependencies = [
+ "futures-core",
+ "pin-project-lite",
+ "tokio",
+]
+
+[[package]]
+name = "tokio-tungstenite"
+version = "0.28.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d25a406cddcc431a75d3d9afc6a7c0f7428d4891dd973e4d54c56b46127bf857"
+dependencies = [
+ "futures-util",
+ "log",
+ "rustls",
+ "rustls-native-certs",
+ "rustls-pki-types",
+ "tokio",
+ "tokio-rustls",
+ "tungstenite",
+]
+
+[[package]]
+name = "tower"
+version = "0.5.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ebe5ef63511595f1344e2d5cfa636d973292adc0eec1f0ad45fae9f0851ab1d4"
+dependencies = [
+ "futures-core",
+ "futures-util",
+ "pin-project-lite",
+ "sync_wrapper",
+ "tokio",
+ "tower-layer",
+ "tower-service",
+]
+
+[[package]]
+name = "tower-http"
+version = "0.6.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4cfcf7e2740e6fc6d4d688b4ef00650406bb94adf4731e43c096c3a19fe40840"
+dependencies = [
+ "bitflags",
+ "bytes",
+ "futures-util",
+ "http",
+ "http-body",
+ "pin-project-lite",
+ "tower",
+ "tower-layer",
+ "tower-service",
+ "url",
+]
+
+[[package]]
+name = "tower-layer"
+version = "0.3.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "121c2a6cda46980bb0fcd1647ffaf6cd3fc79a013de288782836f6df9c48780e"
+
+[[package]]
+name = "tower-service"
+version = "0.3.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8df9b6e13f2d32c91b9bd719c00d1958837bc7dec474d94952798cc8e69eeec3"
+
+[[package]]
+name = "tracing"
+version = "0.1.44"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "63e71662fa4b2a2c3a26f570f037eb95bb1f85397f3cd8076caed2f026a6d100"
+dependencies = [
+ "pin-project-lite",
+ "tracing-attributes",
+ "tracing-core",
+]
+
+[[package]]
+name = "tracing-attributes"
+version = "0.1.31"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7490cfa5ec963746568740651ac6781f701c9c5ea257c58e057f3ba8cf69e8da"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "tracing-core"
+version = "0.1.36"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "db97caf9d906fbde555dd62fa95ddba9eecfd14cb388e4f491a66d74cd5fb79a"
+dependencies = [
+ "once_cell",
+ "valuable",
+]
+
+[[package]]
+name = "tracing-log"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ee855f1f400bd0e5c02d150ae5de3840039a3f54b025156404e34c23c03f47c3"
+dependencies = [
+ "log",
+ "once_cell",
+ "tracing-core",
+]
+
+[[package]]
+name = "tracing-subscriber"
+version = "0.3.23"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cb7f578e5945fb242538965c2d0b04418d38ec25c79d160cd279bf0731c8d319"
+dependencies = [
+ "matchers",
+ "nu-ansi-term",
+ "once_cell",
+ "regex-automata",
+ "sharded-slab",
+ "smallvec",
+ "thread_local",
+ "tracing",
+ "tracing-core",
+ "tracing-log",
+]
+
+[[package]]
+name = "try-lock"
+version = "0.2.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e421abadd41a4225275504ea4d6566923418b7f05506fbc9c0fe86ba7396114b"
+
+[[package]]
+name = "tungstenite"
+version = "0.28.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8628dcc84e5a09eb3d8423d6cb682965dea9133204e8fb3efee74c2a0c259442"
+dependencies = [
+ "bytes",
+ "data-encoding",
+ "http",
+ "httparse",
+ "log",
+ "rand 0.9.5",
+ "rustls",
+ "rustls-pki-types",
+ "sha1",
+ "thiserror",
+ "utf-8",
+]
+
+[[package]]
+name = "typenum"
+version = "1.20.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b6f5e870be6c3b371b77fe0ee0bafb859fa4964b4404c27de1d380043c4dda20"
+
+[[package]]
+name = "unicode-ident"
+version = "1.0.24"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e6e4313cd5fcd3dad5cafa179702e2b244f760991f45397d14d4ebf38247da75"
+
+[[package]]
+name = "unsafe-libyaml"
+version = "0.2.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "673aac59facbab8a9007c7f6108d11f63b603f7cabff99fabf650fea5c32b861"
+
+[[package]]
+name = "untrusted"
+version = "0.9.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8ecb6da28b8a351d773b68d5825ac39017e680750f980f3a1a85cd8dd28a47c1"
+
+[[package]]
+name = "url"
+version = "2.5.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ff67a8a4397373c3ef660812acab3268222035010ab8680ec4215f38ba3d0eed"
+dependencies = [
+ "form_urlencoded",
+ "idna",
+ "percent-encoding",
+ "serde",
+]
+
+[[package]]
+name = "utf-8"
+version = "0.7.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "09cc8ee72d2a9becf2f2febe0205bbed8fc6615b7cb429ad062dc7b7ddd036a9"
+
+[[package]]
+name = "utf8_iter"
+version = "1.0.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b6c140620e7ffbb22c2dee59cafe6084a59b5ffc27a8859a5f0d494b5d52b6be"
+
+[[package]]
+name = "utf8parse"
+version = "0.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "06abde3611657adf66d383f00b093d7faecc7fa57071cce2578660c9f1010821"
+
+[[package]]
+name = "uuid"
+version = "1.24.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bf3923a6f5c4c6382e0b653c4117f48d631ea17f38ed86e2a828e6f7412f5239"
+dependencies = [
+ "getrandom 0.4.3",
+ "js-sys",
+ "serde_core",
+ "wasm-bindgen",
+]
+
+[[package]]
+name = "valuable"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ba73ea9cf16a25df0c8caa16c51acb937d5712a8429db78a3ee29d5dcacd3a65"
+
+[[package]]
+name = "version_check"
+version = "0.9.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0b928f33d975fc6ad9f86c8f283853ad26bdd5b10b7f1542aa2fa15e2289105a"
+
+[[package]]
+name = "want"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bfa7760aed19e106de2c7c0b581b509f2f25d3dacaf737cb82ac61bc6d760b0e"
+dependencies = [
+ "try-lock",
+]
+
+[[package]]
+name = "wasi"
+version = "0.11.1+wasi-snapshot-preview1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ccf3ec651a847eb01de73ccad15eb7d99f80485de043efb2f370cd654f4ea44b"
+
+[[package]]
+name = "wasip2"
+version = "1.0.4+wasi-0.2.12"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b67efb37e106e55ce722a510d6b5f9c17f083e5fc79afc2badeb12cc313d9487"
+dependencies = [
+ "wit-bindgen",
+]
+
+[[package]]
+name = "wasm-bindgen"
+version = "0.2.126"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4b067c0c11094aef6b7a801c1e34a26affafdf3d051dba08456b868789aaf9a4"
+dependencies = [
+ "cfg-if",
+ "once_cell",
+ "rustversion",
+ "wasm-bindgen-macro",
+ "wasm-bindgen-shared",
+]
+
+[[package]]
+name = "wasm-bindgen-futures"
+version = "0.4.76"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c62df1340f32221cb9c54d6a27b030e3dba64361d4a95bed55f9aacb44da291d"
+dependencies = [
+ "js-sys",
+ "wasm-bindgen",
+]
+
+[[package]]
+name = "wasm-bindgen-macro"
+version = "0.2.126"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "167ce5e579f6bcf889c4f7175a8a5a585de84e8ff93976ce393efa5f2837aab1"
+dependencies = [
+ "quote",
+ "wasm-bindgen-macro-support",
+]
+
+[[package]]
+name = "wasm-bindgen-macro-support"
+version = "0.2.126"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f3997c7839262f4ef12cf90b818d6340c18e80f263f1a94bf157d0ec4420380e"
+dependencies = [
+ "bumpalo",
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+ "wasm-bindgen-shared",
+]
+
+[[package]]
+name = "wasm-bindgen-shared"
+version = "0.2.126"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dc1b4cb0cc549fcf58d7dfc081778139b3d283a081644e833e84682ad71cea24"
+dependencies = [
+ "unicode-ident",
+]
+
+[[package]]
+name = "web-sys"
+version = "0.3.103"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8622dcb61c0bcc9fffa6938bed81210af2da9a7e4a1a834b2e37a59b6dfb6141"
+dependencies = [
+ "js-sys",
+ "wasm-bindgen",
+]
+
+[[package]]
+name = "web-time"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5a6580f308b1fad9207618087a65c04e7a10bc77e02c8e84e9b00dd4b12fa0bb"
+dependencies = [
+ "js-sys",
+ "wasm-bindgen",
+]
+
+[[package]]
+name = "webpki-roots"
+version = "1.0.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7dcd9d09a39985f5344844e66b0c530a33843579125f23e21e9f0f220850f22a"
+dependencies = [
+ "rustls-pki-types",
+]
+
+[[package]]
+name = "winapi"
+version = "0.3.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5c839a674fcd7a98952e593242ea400abe93992746761e38641405d28b00f419"
+dependencies = [
+ "winapi-i686-pc-windows-gnu",
+ "winapi-x86_64-pc-windows-gnu",
+]
+
+[[package]]
+name = "winapi-i686-pc-windows-gnu"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ac3b87c63620426dd9b991e5ce0329eff545bccbbb34f3be09ff6fb6ab51b7b6"
+
+[[package]]
+name = "winapi-x86_64-pc-windows-gnu"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "712e227841d057c1ee1cd2fb22fa7e5a5461ae8e48fa2ca79ec42cfc1931183f"
+
+[[package]]
+name = "windows"
+version = "0.62.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "527fadee13e0c05939a6a05d5bd6eec6cd2e3dbd648b9f8e447c6518133d8580"
+dependencies = [
+ "windows-collections",
+ "windows-core",
+ "windows-future",
+ "windows-numerics",
+]
+
+[[package]]
+name = "windows-collections"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "23b2d95af1a8a14a3c7367e1ed4fc9c20e0a26e79551b1454d72583c97cc6610"
+dependencies = [
+ "windows-core",
+]
+
+[[package]]
+name = "windows-core"
+version = "0.62.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b8e83a14d34d0623b51dce9581199302a221863196a1dde71a7663a4c2be9deb"
+dependencies = [
+ "windows-implement",
+ "windows-interface",
+ "windows-link",
+ "windows-result",
+ "windows-strings",
+]
+
+[[package]]
+name = "windows-future"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e1d6f90251fe18a279739e78025bd6ddc52a7e22f921070ccdc67dde84c605cb"
+dependencies = [
+ "windows-core",
+ "windows-link",
+ "windows-threading",
+]
+
+[[package]]
+name = "windows-implement"
+version = "0.60.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "053e2e040ab57b9dc951b72c264860db7eb3b0200ba345b4e4c3b14f67855ddf"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "windows-interface"
+version = "0.59.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3f316c4a2570ba26bbec722032c4099d8c8bc095efccdc15688708623367e358"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "windows-link"
+version = "0.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f0805222e57f7521d6a62e36fa9163bc891acd422f971defe97d64e70d0a4fe5"
+
+[[package]]
+name = "windows-numerics"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6e2e40844ac143cdb44aead537bbf727de9b044e107a0f1220392177d15b0f26"
+dependencies = [
+ "windows-core",
+ "windows-link",
+]
+
+[[package]]
+name = "windows-result"
+version = "0.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7781fa89eaf60850ac3d2da7af8e5242a5ea78d1a11c49bf2910bb5a73853eb5"
+dependencies = [
+ "windows-link",
+]
+
+[[package]]
+name = "windows-strings"
+version = "0.5.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7837d08f69c77cf6b07689544538e017c1bfcf57e34b4c0ff58e6c2cd3b37091"
+dependencies = [
+ "windows-link",
+]
+
+[[package]]
+name = "windows-sys"
+version = "0.52.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "282be5f36a8ce781fad8c8ae18fa3f9beff57ec1b52cb3de0789201425d9a33d"
+dependencies = [
+ "windows-targets",
+]
+
+[[package]]
+name = "windows-sys"
+version = "0.61.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ae137229bcbd6cdf0f7b80a31df61766145077ddf49416a728b02cb3921ff3fc"
+dependencies = [
+ "windows-link",
+]
+
+[[package]]
+name = "windows-targets"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9b724f72796e036ab90c1021d4780d4d3d648aca59e491e6b98e725b84e99973"
+dependencies = [
+ "windows_aarch64_gnullvm",
+ "windows_aarch64_msvc",
+ "windows_i686_gnu",
+ "windows_i686_gnullvm",
+ "windows_i686_msvc",
+ "windows_x86_64_gnu",
+ "windows_x86_64_gnullvm",
+ "windows_x86_64_msvc",
+]
+
+[[package]]
+name = "windows-threading"
+version = "0.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3949bd5b99cafdf1c7ca86b43ca564028dfe27d66958f2470940f73d86d75b37"
+dependencies = [
+ "windows-link",
+]
+
+[[package]]
+name = "windows_aarch64_gnullvm"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "32a4622180e7a0ec044bb555404c800bc9fd9ec262ec147edd5989ccd0c02cd3"
+
+[[package]]
+name = "windows_aarch64_msvc"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "09ec2a7bb152e2252b53fa7803150007879548bc709c039df7627cabbd05d469"
+
+[[package]]
+name = "windows_i686_gnu"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8e9b5ad5ab802e97eb8e295ac6720e509ee4c243f69d781394014ebfe8bbfa0b"
+
+[[package]]
+name = "windows_i686_gnullvm"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0eee52d38c090b3caa76c563b86c3a4bd71ef1a819287c19d586d7334ae8ed66"
+
+[[package]]
+name = "windows_i686_msvc"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "240948bc05c5e7c6dabba28bf89d89ffce3e303022809e73deaefe4f6ec56c66"
+
+[[package]]
+name = "windows_x86_64_gnu"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "147a5c80aabfbf0c7d901cb5895d1de30ef2907eb21fbbab29ca94c5b08b1a78"
+
+[[package]]
+name = "windows_x86_64_gnullvm"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "24d5b23dc417412679681396f2b49f3de8c1473deb516bd34410872eff51ed0d"
+
+[[package]]
+name = "windows_x86_64_msvc"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "589f6da84c646204747d1270a2a5661ea66ed1cced2631d546fdfb155959f9ec"
+
+[[package]]
+name = "wit-bindgen"
+version = "0.57.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1ebf944e87a7c253233ad6766e082e3cd714b5d03812acc24c318f549614536e"
+
+[[package]]
+name = "writeable"
+version = "0.6.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1ffae5123b2d3fc086436f8834ae3ab053a283cfac8fe0a0b8eaae044768a4c4"
+
+[[package]]
+name = "yoke"
+version = "0.8.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "709fe23a0424b6a435d82152b1bd3fdfb0833487d5fa90d05d42762a9891fef5"
+dependencies = [
+ "stable_deref_trait",
+ "yoke-derive",
+ "zerofrom",
+]
+
+[[package]]
+name = "yoke-derive"
+version = "0.8.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "de844c262c8848816172cef550288e7dc6c7b7814b4ee56b3e1553f275f1858e"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+ "synstructure",
+]
+
+[[package]]
+name = "zerocopy"
+version = "0.8.55"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b5a105cd7b140f6eeec8acff2ea38135d3cab283ada58540f629fe51e46696eb"
+dependencies = [
+ "zerocopy-derive",
+]
+
+[[package]]
+name = "zerocopy-derive"
+version = "0.8.55"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0fe976fb70c78cd64cccfe3a6fc142244e8a77b70959b30faf9d0ac37ee228eb"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "zerofrom"
+version = "0.1.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0ec05a11813ea801ff6d75110ad09cd0824ddba17dfe17128ea0d5f68e6c5272"
+dependencies = [
+ "zerofrom-derive",
+]
+
+[[package]]
+name = "zerofrom-derive"
+version = "0.1.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "11532158c46691caf0f2593ea8358fed6bbf68a0315e80aae9bd41fbade684a1"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+ "synstructure",
+]
+
+[[package]]
+name = "zeroize"
+version = "1.9.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e13c156562582aa81c60cb29407084cdb54c4164760106ab78e6c5b0858cf64e"
+
+[[package]]
+name = "zerotrie"
+version = "0.2.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0f9152d31db0792fa83f70fb2f83148effb5c1f5b8c7686c3459e361d9bc20bf"
+dependencies = [
+ "displaydoc",
+ "yoke",
+ "zerofrom",
+]
+
+[[package]]
+name = "zerovec"
+version = "0.11.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "90f911cbc359ab6af17377d242225f4d75119aec87ea711a880987b18cd7b239"
+dependencies = [
+ "yoke",
+ "zerofrom",
+ "zerovec-derive",
+]
+
+[[package]]
+name = "zerovec-derive"
+version = "0.11.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "625dc425cab0dca6dc3c3319506e6593dcb08a9f387ea3b284dbd52a92c40555"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "zmij"
+version = "1.0.23"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "29666d0abbfad1e3dc4dcf6144730dd3a3ab225bbbdac83319345b1b44ccfc1b"
diff --git a/eval/Cargo.toml b/eval/Cargo.toml
new file mode 100644
index 000000000..055b20f1c
--- /dev/null
+++ b/eval/Cargo.toml
@@ -0,0 +1,36 @@
+[workspace]
+
+[package]
+name = "eval"
+version = "0.1.0"
+edition = "2021"
+publish = false
+
+[[bin]]
+name = "eval"
+path = "src/main.rs"
+
+[lib]
+name = "eval"
+path = "src/lib.rs"
+
+[dependencies]
+harness = { path = "../harness" }
+iii-sdk = "=0.21.6"
+iii-helpers = "=0.21.6"
+iii-console-ui = { path = "../crates/console-ui" }
+tokio = { version = "1", features = ["rt-multi-thread", "macros", "sync", "signal", "time"] }
+serde = { version = "1", features = ["derive"] }
+serde_json = "1"
+anyhow = "1"
+thiserror = "2"
+tracing = "0.1"
+tracing-subscriber = { version = "0.3", features = ["fmt", "env-filter"] }
+clap = { version = "4", features = ["derive", "env"] }
+async-trait = "0.1"
+schemars = "0.8"
+sha2 = "0.10"
+uuid = { version = "1", features = ["v4"] }
+
+[dev-dependencies]
+serde_yaml = "0.9"
diff --git a/eval/README.md b/eval/README.md
new file mode 100644
index 000000000..fcec889a3
--- /dev/null
+++ b/eval/README.md
@@ -0,0 +1,79 @@
+# eval
+
+`eval` compares two prompt or system-prompt variants under the same model and
+execution settings. It runs control and treatment through the iii harness,
+evaluates each result with an iii function, and reports correctness together
+with tokens, cost, latency, function-call, trace, and span metrics.
+
+## Start an evaluation
+
+```json
+{
+  "dimension": "prompt",
+  "model": {
+    "model": "codex/gpt-5.6-luna",
+    "provider": "openai-codex",
+    "system_prompt_strategy": "override"
+  },
+  "control": {
+    "label": "baseline",
+    "prompt": "Reply with exactly OK.",
+    "system_prompt": "Follow the user request exactly."
+  },
+  "treatment": {
+    "label": "candidate",
+    "prompt": "Return exactly the text OK.",
+    "system_prompt": "Follow the user request exactly."
+  },
+  "evaluator": {
+    "function_id": "eval::assert::exact",
+    "arguments": {
+      "expected": "OK"
+    }
+  }
+}
+```
+
+Call `eval::start` with the request. It returns an `evaluation_id`
+immediately. Use `eval::status` for occasional progress checks and
+`eval::result` for the terminal report, or bind to the `eval::completed`
+trigger type.
+
+Three runs per variant are scheduled by default. Their order alternates by
+pair to reduce order bias. A candidate is eligible only when every treatment
+run passes and its pass count does not regress against control. Efficiency
+metrics are descriptive and never select a winner automatically.
+
+## Public functions
+
+- `eval::start` — validate, persist, and enqueue an evaluation.
+- `eval::list` — list recent evaluations as lightweight summaries.
+- `eval::status` — inspect progress without loading the full report.
+- `eval::result` — read the normalized request and terminal comparison report.
+- `eval::cancel` — cancel the active harness session and remaining runs.
+- `eval::delete` — delete a terminal evaluation and its session indexes.
+- `eval::assert::exact` — built-in deep JSON/string equality evaluator.
+
+Evaluator functions receive the output, `harness::metrics`, run identity, and
+caller-supplied arguments. They return `{ passed, score?, reason?, details? }`.
+Evaluators should be deterministic and idempotent because durable delivery is
+at-least-once.
+
+## Console UI
+
+When the console worker is running, `eval` injects an **eval** page at
+`#/ext/eval-benchmarks`. The page creates prompt or system-prompt comparisons,
+tracks durable progress, restores recent reports after reload, and compares
+correctness, token, cost, latency, function-call, trace, and span metrics.
+
+The model picker reads the live `router::models::list` catalog and falls back
+to manual model/provider entry when the catalog is unavailable. Exact-value
+evaluation is built in; any evaluator function can also be selected by id with
+JSON arguments. Harness policies and output/metadata options remain collapsed
+until needed.
+
+## Boundaries
+
+The worker intentionally does not implement an agent loop, model router,
+metrics collector, trace collector, test DSL, or LLM judge. Those concerns
+remain in the harness, engine, and user-provided evaluator functions.
diff --git a/eval/build.rs b/eval/build.rs
new file mode 100644
index 000000000..705915d28
--- /dev/null
+++ b/eval/build.rs
@@ -0,0 +1,137 @@
+use std::path::{Path, PathBuf};
+use std::process::Command;
+use std::time::SystemTime;
+
+fn main() {
+    println!(
+        "cargo:rustc-env=TARGET={}",
+        std::env::var("TARGET").unwrap()
+    );
+    for path in [
+        "ui/page.tsx",
+        "ui/styles.css",
+        "ui/src",
+        "ui/build.mjs",
+        "ui/package.json",
+        "ui/tsconfig.json",
+        "../pnpm-lock.yaml",
+    ] {
+        println!("cargo:rerun-if-changed={path}");
+    }
+
+    let manifest_dir = PathBuf::from(env!("CARGO_MANIFEST_DIR"));
+    let ui_dir = manifest_dir.join("ui");
+    let assets = [ui_dir.join("dist/page.js"), ui_dir.join("dist/styles.css")];
+    if assets
+        .iter()
+        .all(|asset| asset.exists() && dist_is_fresh(asset, &ui_dir))
+    {
+        return;
+    }
+
+    if std::env::var_os("SKIP_UI_BUILD").is_some() {
+        for asset in &assets {
+            assert!(
+                asset.exists(),
+                "SKIP_UI_BUILD set but {} is missing",
+                asset.display()
+            );
+        }
+        return;
+    }
+
+    let pnpm = locate_pnpm();
+    run(&pnpm, &["install"], &ui_dir);
+    run(&pnpm, &["build"], &ui_dir);
+    for asset in &assets {
+        assert!(
+            asset.exists(),
+            "UI build finished but {} is missing",
+            asset.display()
+        );
+    }
+}
+
+fn run(program: &Path, args: &[&str], directory: &Path) {
+    let status = Command::new(program)
+        .args(args)
+        .current_dir(directory)
+        .status()
+        .unwrap_or_else(|error| {
+            panic!(
+                "failed to run {} in {}: {error}",
+                program.display(),
+                directory.display()
+            )
+        });
+    assert!(
+        status.success(),
+        "{} exited with {status}",
+        program.display()
+    );
+}
+
+fn dist_is_fresh(asset: &Path, ui_dir: &Path) -> bool {
+    let Ok(asset_time) = asset.metadata().and_then(|metadata| metadata.modified()) else {
+        return false;
+    };
+    for path in [
+        ui_dir.join("page.tsx"),
+        ui_dir.join("styles.css"),
+        ui_dir.join("build.mjs"),
+        ui_dir.join("package.json"),
+        ui_dir.join("tsconfig.json"),
+        ui_dir.join("../../pnpm-lock.yaml"),
+    ] {
+        if path.exists() {
+            match path.metadata().and_then(|metadata| metadata.modified()) {
+                Ok(modified) if modified <= asset_time => {}
+                _ => return false,
+            }
+        }
+    }
+    subtree_is_older(&ui_dir.join("src"), asset_time)
+}
+
+fn subtree_is_older(root: &Path, ceiling: SystemTime) -> bool {
+    let Ok(entries) = std::fs::read_dir(root) else {
+        return false;
+    };
+    for entry in entries.flatten() {
+        let path = entry.path();
+        let Ok(metadata) = entry.metadata() else {
+            return false;
+        };
+        if metadata.is_dir() {
+            if !subtree_is_older(&path, ceiling) {
+                return false;
+            }
+        } else {
+            match metadata.modified() {
+                Ok(modified) if modified <= ceiling => {}
+                _ => return false,
+            }
+        }
+    }
+    true
+}
+
+fn locate_pnpm() -> PathBuf {
+    if let Ok(explicit) = std::env::var("PNPM") {
+        return PathBuf::from(explicit);
+    }
+    let names: &[&str] = if cfg!(windows) {
+        &["pnpm.cmd", "pnpm.exe", "pnpm"]
+    } else {
+        &["pnpm"]
+    };
+    for directory in std::env::split_paths(&std::env::var_os("PATH").unwrap_or_default()) {
+        for name in names {
+            let candidate = directory.join(name);
+            if candidate.is_file() {
+                return candidate;
+            }
+        }
+    }
+    panic!("pnpm not found on PATH; build eval/ui first or set PNPM");
+}
diff --git a/eval/iii.worker.yaml b/eval/iii.worker.yaml
new file mode 100644
index 000000000..149af3d93
--- /dev/null
+++ b/eval/iii.worker.yaml
@@ -0,0 +1,15 @@
+iii: v1
+name: eval
+language: rust
+deploy: binary
+manifest: Cargo.toml
+bin: eval
+tags: [evaluation, benchmark, prompt, system-prompt, ab-test]
+description: Compare prompt or system-prompt variants on the same model using durable harness runs and correctness-first reports.
+
+dependencies:
+  harness: "^1.5.5"
+  state: "^0.21.3"
+  queue: "^0.2.0"
+  cron: "^0.21.0"
+  iii-observability: "^0.21.6"
diff --git a/eval/src/contract.rs b/eval/src/contract.rs
new file mode 100644
index 000000000..2d7f5ca30
--- /dev/null
+++ b/eval/src/contract.rs
@@ -0,0 +1,619 @@
+use std::collections::BTreeMap;
+
+use harness::functions::metrics::SessionMetricsResponseV1;
+use harness::prompt::{Mode, SystemPromptStrategy};
+use harness::types::model::ThinkingLevel;
+use harness::types::output::OutputContract;
+use harness::types::turn::FunctionPolicy;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+use serde_json::{json, Value};
+
+use crate::error::EvalError;
+use crate::limits::EvalLimitsV1;
+use crate::report::EvalReportV1;
+
+pub const DEFAULT_RUNS: u32 = 1;
+pub const MAX_RUNS: u32 = 20;
+pub const DEFAULT_LIST_LIMIT: u32 = 50;
+pub const MAX_LIST_LIMIT: u32 = 100;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum ComparisonDimensionV1 {
+    Prompt,
+    SystemPrompt,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum VariantRoleV1 {
+    Control,
+    Treatment,
+}
+
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum ExecutionOrderV1 {
+    #[default]
+    BalancedControlFirst,
+    BalancedTreatmentFirst,
+}
+
+impl ExecutionOrderV1 {
+    pub fn roles(self, iteration: u32) -> [VariantRoleV1; 2] {
+        let control_first = match self {
+            Self::BalancedControlFirst => iteration % 2 == 1,
+            Self::BalancedTreatmentFirst => iteration.is_multiple_of(2),
+        };
+        if control_first {
+            [VariantRoleV1::Control, VariantRoleV1::Treatment]
+        } else {
+            [VariantRoleV1::Treatment, VariantRoleV1::Control]
+        }
+    }
+
+    pub fn reversed(self) -> Self {
+        match self {
+            Self::BalancedControlFirst => Self::BalancedTreatmentFirst,
+            Self::BalancedTreatmentFirst => Self::BalancedControlFirst,
+        }
+    }
+}
+
+impl VariantRoleV1 {
+    pub fn as_str(self) -> &'static str {
+        match self {
+            Self::Control => "control",
+            Self::Treatment => "treatment",
+        }
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvalVariantV1 {
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub label: Option<String>,
+    pub prompt: String,
+    /// `null` disables the system prompt for this variant. An empty string
+    /// keeps the harness behavior of resolving the provider or built-in prompt.
+    pub system_prompt: Option<String>,
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvalModelConfigV1 {
+    pub model: String,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub provider: Option<String>,
+    #[serde(default)]
+    pub system_prompt_strategy: SystemPromptStrategy,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub mode: Option<Mode>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub thinking_level: Option<ThinkingLevel>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub provider_options: Option<BTreeMap<String, Value>>,
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvaluatorSpecV1 {
+    pub function_id: String,
+    #[serde(default = "empty_object")]
+    pub arguments: Value,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+pub struct EvalStartRequestV1 {
+    pub dimension: ComparisonDimensionV1,
+    pub model: EvalModelConfigV1,
+    pub control: EvalVariantV1,
+    pub treatment: EvalVariantV1,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub evaluator: Option<EvaluatorSpecV1>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub runs: Option<u32>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub execution_order: Option<ExecutionOrderV1>,
+    #[serde(default)]
+    pub limits: EvalLimitsV1,
+    /// Shared function policy for both variants. Omitted means deny all.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub functions: Option<FunctionPolicy>,
+    /// Shared output contract for both variants. Omitted means text.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub output: Option<OutputContract>,
+    /// Shared harness metadata, including an optional filesystem scope.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub metadata: Option<Value>,
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct NormalizedEvalRequestV1 {
+    pub dimension: ComparisonDimensionV1,
+    pub model: EvalModelConfigV1,
+    pub control: EvalVariantV1,
+    pub treatment: EvalVariantV1,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub evaluator: Option<EvaluatorSpecV1>,
+    pub runs: u32,
+    #[serde(default)]
+    pub execution_order: ExecutionOrderV1,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub source_evaluation_id: Option<String>,
+    pub limits: EvalLimitsV1,
+    pub functions: FunctionPolicy,
+    pub output: OutputContract,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub metadata: Option<Value>,
+}
+
+impl EvalStartRequestV1 {
+    pub fn normalize(self) -> Result<NormalizedEvalRequestV1, EvalError> {
+        validate_non_empty("model.model", &self.model.model)?;
+        if self
+            .model
+            .provider
+            .as_deref()
+            .is_some_and(|provider| provider.trim().is_empty())
+        {
+            return Err(EvalError::InvalidRequest(
+                "model.provider cannot be empty when supplied".into(),
+            ));
+        }
+        validate_non_empty("control.prompt", &self.control.prompt)?;
+        validate_non_empty("treatment.prompt", &self.treatment.prompt)?;
+        if let Some(evaluator) = &self.evaluator {
+            validate_non_empty("evaluator.function_id", &evaluator.function_id)?;
+        }
+        if self
+            .control
+            .label
+            .as_deref()
+            .is_some_and(|label| label.trim().is_empty())
+            || self
+                .treatment
+                .label
+                .as_deref()
+                .is_some_and(|label| label.trim().is_empty())
+        {
+            return Err(EvalError::InvalidRequest(
+                "variant labels cannot be empty when supplied".into(),
+            ));
+        }
+
+        match self.dimension {
+            ComparisonDimensionV1::Prompt => {
+                if self.control.system_prompt != self.treatment.system_prompt {
+                    return Err(EvalError::InvalidRequest(
+                        "prompt comparison requires identical system prompts".into(),
+                    ));
+                }
+                if self.control.prompt == self.treatment.prompt {
+                    return Err(EvalError::InvalidRequest(
+                        "prompt comparison requires different prompts".into(),
+                    ));
+                }
+            }
+            ComparisonDimensionV1::SystemPrompt => {
+                if self.control.prompt != self.treatment.prompt {
+                    return Err(EvalError::InvalidRequest(
+                        "system_prompt comparison requires identical prompts".into(),
+                    ));
+                }
+                if self.control.system_prompt == self.treatment.system_prompt {
+                    return Err(EvalError::InvalidRequest(
+                        "system_prompt comparison requires different system prompts".into(),
+                    ));
+                }
+            }
+        }
+
+        let runs = self.runs.unwrap_or(DEFAULT_RUNS);
+        if !(1..=MAX_RUNS).contains(&runs) {
+            return Err(EvalError::InvalidRequest(format!(
+                "runs must be between 1 and {MAX_RUNS}"
+            )));
+        }
+        self.limits.validate()?;
+
+        Ok(NormalizedEvalRequestV1 {
+            dimension: self.dimension,
+            model: self.model,
+            control: self.control,
+            treatment: self.treatment,
+            evaluator: self.evaluator,
+            runs,
+            execution_order: self.execution_order.unwrap_or_default(),
+            source_evaluation_id: None,
+            limits: self.limits,
+            functions: self.functions.unwrap_or_default(),
+            output: self.output.unwrap_or_default(),
+            metadata: self.metadata,
+        })
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum EvalStatusV1 {
+    Queued,
+    Running,
+    Completed,
+    Failed,
+    Cancelled,
+}
+
+impl EvalStatusV1 {
+    pub fn is_terminal(self) -> bool {
+        matches!(self, Self::Completed | Self::Failed | Self::Cancelled)
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum EvalRunStatusV1 {
+    Pending,
+    Running,
+    Completed,
+    Failed,
+    Cancelled,
+}
+
+impl EvalRunStatusV1 {
+    pub fn is_terminal(self) -> bool {
+        matches!(self, Self::Completed | Self::Failed | Self::Cancelled)
+    }
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvalStartResponseV1 {
+    pub evaluation_id: String,
+    pub status: EvalStatusV1,
+}
+
+#[derive(Debug, Clone, Default, Deserialize, JsonSchema)]
+pub struct EvalListRequestV1 {
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub limit: Option<u32>,
+}
+
+impl EvalListRequestV1 {
+    pub fn normalized_limit(&self) -> Result<usize, EvalError> {
+        let limit = self.limit.unwrap_or(DEFAULT_LIST_LIMIT);
+        if !(1..=MAX_LIST_LIMIT).contains(&limit) {
+            return Err(EvalError::InvalidRequest(format!(
+                "limit must be between 1 and {MAX_LIST_LIMIT}"
+            )));
+        }
+        Ok(limit as usize)
+    }
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvalSummaryV1 {
+    pub evaluation_id: String,
+    pub status: EvalStatusV1,
+    pub dimension: ComparisonDimensionV1,
+    pub model: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub provider: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub control_label: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub treatment_label: Option<String>,
+    pub total_runs: u32,
+    pub terminal_runs: u32,
+    pub passed_runs: u32,
+    pub failed_runs: u32,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub eligible: Option<bool>,
+    pub created_at: i64,
+    pub updated_at: i64,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub completed_at: Option<i64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub error: Option<String>,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvalListResponseV1 {
+    pub evaluations: Vec<EvalSummaryV1>,
+}
+
+#[derive(Debug, Clone, Deserialize, JsonSchema)]
+pub struct EvaluationIdRequestV1 {
+    pub evaluation_id: String,
+}
+
+#[derive(Debug, Clone, Deserialize, JsonSchema)]
+pub struct EvalRerunRequestV1 {
+    pub evaluation_id: String,
+    #[serde(default)]
+    pub reverse_order: bool,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct ActiveRunV1 {
+    pub run_id: String,
+    pub role: VariantRoleV1,
+    pub iteration: u32,
+    pub session_id: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub turn_id: Option<String>,
+    pub started_at: i64,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvalStatusResponseV1 {
+    pub evaluation_id: String,
+    pub status: EvalStatusV1,
+    pub total_runs: u32,
+    pub terminal_runs: u32,
+    pub passed_runs: u32,
+    pub failed_runs: u32,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub active: Option<ActiveRunV1>,
+    pub created_at: i64,
+    pub updated_at: i64,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub completed_at: Option<i64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub error: Option<String>,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvalResultResponseV1 {
+    pub status: EvalStatusV1,
+    pub request: NormalizedEvalRequestV1,
+    pub progress: crate::report::EvalProgressV1,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub report: Option<EvalReportV1>,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvalCancelResponseV1 {
+    pub cancelled: bool,
+    pub status: EvalStatusV1,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvalDeleteResponseV1 {
+    pub deleted: bool,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+pub struct EvaluatorInputV1 {
+    pub evaluation_id: String,
+    pub run_id: String,
+    pub role: VariantRoleV1,
+    pub session_id: String,
+    pub output: Value,
+    pub metrics: SessionMetricsResponseV1,
+    #[serde(default = "empty_object")]
+    pub arguments: Value,
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvaluatorResponseV1 {
+    pub passed: bool,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub score: Option<f64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub reason: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub details: Option<Value>,
+}
+
+impl EvaluatorResponseV1 {
+    pub fn validate(self) -> Result<Self, EvalError> {
+        if self
+            .score
+            .is_some_and(|score| !score.is_finite() || !(0.0..=1.0).contains(&score))
+        {
+            return Err(EvalError::InvalidRequest(
+                "evaluator score must be a finite number between 0 and 1".into(),
+            ));
+        }
+        Ok(self)
+    }
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+pub struct StepRequestV1 {
+    pub evaluation_id: String,
+    pub step: u64,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct StepResponseV1 {
+    pub skipped: bool,
+    pub status: EvalStatusV1,
+}
+
+#[derive(Debug, Clone, Default, Deserialize, JsonSchema)]
+pub struct WakeEventV1 {
+    #[serde(default)]
+    pub session_id: String,
+    #[serde(default)]
+    pub terminal: bool,
+}
+
+#[derive(Debug, Clone, Serialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct WakeResponseV1 {
+    pub woke: bool,
+}
+
+#[derive(Debug, Clone, Default, Deserialize, JsonSchema)]
+pub struct SweepEventV1 {
+    #[serde(default)]
+    pub scheduled_at: Option<i64>,
+    #[serde(default)]
+    pub scheduled_time: Option<String>,
+}
+
+#[derive(Debug, Clone, Serialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct SweepResponseV1 {
+    pub swept: u64,
+}
+
+fn validate_non_empty(name: &str, value: &str) -> Result<(), EvalError> {
+    if value.trim().is_empty() {
+        Err(EvalError::InvalidRequest(format!("{name} cannot be empty")))
+    } else {
+        Ok(())
+    }
+}
+
+fn empty_object() -> Value {
+    json!({})
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn request(dimension: ComparisonDimensionV1) -> EvalStartRequestV1 {
+        EvalStartRequestV1 {
+            dimension,
+            model: EvalModelConfigV1 {
+                model: "model".into(),
+                provider: Some("provider".into()),
+                system_prompt_strategy: SystemPromptStrategy::Override,
+                mode: Some(Mode::Agent),
+                thinking_level: None,
+                provider_options: None,
+            },
+            control: EvalVariantV1 {
+                label: None,
+                prompt: "control".into(),
+                system_prompt: Some("system".into()),
+            },
+            treatment: EvalVariantV1 {
+                label: None,
+                prompt: "treatment".into(),
+                system_prompt: Some("system".into()),
+            },
+            evaluator: Some(EvaluatorSpecV1 {
+                function_id: "eval::assert::exact".into(),
+                arguments: json!({"expected": "OK"}),
+            }),
+            runs: None,
+            execution_order: None,
+            limits: EvalLimitsV1::default(),
+            functions: None,
+            output: None,
+            metadata: None,
+        }
+    }
+
+    #[test]
+    fn list_limit_defaults_and_stays_bounded() {
+        assert_eq!(
+            EvalListRequestV1::default().normalized_limit().unwrap(),
+            DEFAULT_LIST_LIMIT as usize
+        );
+        assert_eq!(
+            EvalListRequestV1 { limit: Some(1) }
+                .normalized_limit()
+                .unwrap(),
+            1
+        );
+        assert!(EvalListRequestV1 { limit: Some(0) }
+            .normalized_limit()
+            .is_err());
+        assert!(EvalListRequestV1 {
+            limit: Some(MAX_LIST_LIMIT + 1)
+        }
+        .normalized_limit()
+        .is_err());
+    }
+
+    #[test]
+    fn list_request_accepts_engine_metadata() {
+        let request: EvalListRequestV1 =
+            serde_json::from_value(json!({"limit": 5, "_caller_worker_id": "console"})).unwrap();
+        assert_eq!(request.limit, Some(5));
+    }
+
+    #[test]
+    fn prompt_comparison_changes_only_prompt_and_defaults_to_one_run() {
+        let normalized = request(ComparisonDimensionV1::Prompt).normalize().unwrap();
+        assert_eq!(normalized.runs, 1);
+        assert_eq!(normalized.functions, FunctionPolicy::default());
+        assert_eq!(normalized.output, OutputContract::Text);
+    }
+
+    #[test]
+    fn system_prompt_comparison_changes_only_system_prompt() {
+        let mut request = request(ComparisonDimensionV1::SystemPrompt);
+        request.treatment.prompt = request.control.prompt.clone();
+        request.treatment.system_prompt = Some("new system".into());
+        assert!(request.normalize().is_ok());
+    }
+
+    #[test]
+    fn system_prompt_comparison_accepts_default_vs_disabled() {
+        let mut request = request(ComparisonDimensionV1::SystemPrompt);
+        request.treatment.prompt = request.control.prompt.clone();
+        request.control.system_prompt = Some(String::new());
+        request.treatment.system_prompt = None;
+        assert!(request.normalize().is_ok());
+    }
+
+    #[test]
+    fn rejects_identical_or_two_dimension_changes() {
+        let mut identical = request(ComparisonDimensionV1::Prompt);
+        identical.treatment.prompt = identical.control.prompt.clone();
+        assert!(identical.normalize().is_err());
+
+        let mut both = request(ComparisonDimensionV1::Prompt);
+        both.treatment.system_prompt = Some("different".into());
+        assert!(both.normalize().is_err());
+    }
+
+    #[test]
+    fn runs_are_bounded() {
+        let mut zero = request(ComparisonDimensionV1::Prompt);
+        zero.runs = Some(0);
+        assert!(zero.normalize().is_err());
+        let mut too_many = request(ComparisonDimensionV1::Prompt);
+        too_many.runs = Some(MAX_RUNS + 1);
+        assert!(too_many.normalize().is_err());
+    }
+
+    #[test]
+    fn evaluator_score_is_bounded() {
+        assert!(EvaluatorResponseV1 {
+            passed: true,
+            score: Some(1.0),
+            reason: None,
+            details: None,
+        }
+        .validate()
+        .is_ok());
+        assert!(EvaluatorResponseV1 {
+            passed: true,
+            score: Some(1.1),
+            reason: None,
+            details: None,
+        }
+        .validate()
+        .is_err());
+    }
+}
diff --git a/eval/src/error.rs b/eval/src/error.rs
new file mode 100644
index 000000000..51fd6259c
--- /dev/null
+++ b/eval/src/error.rs
@@ -0,0 +1,78 @@
+use iii_sdk::errors::Error as SdkError;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum EvalPhaseV1 {
+    Setup,
+    Send,
+    Await,
+    Collect,
+    Evaluate,
+    Limit,
+    Cancel,
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvalFailureV1 {
+    pub phase: EvalPhaseV1,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub code: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub function_id: Option<String>,
+    pub message: String,
+}
+
+impl EvalFailureV1 {
+    pub fn new(phase: EvalPhaseV1, message: impl Into<String>) -> Self {
+        Self {
+            phase,
+            code: None,
+            function_id: None,
+            message: message.into(),
+        }
+    }
+
+    pub fn function(
+        phase: EvalPhaseV1,
+        function_id: impl Into<String>,
+        message: impl Into<String>,
+    ) -> Self {
+        Self {
+            phase,
+            code: None,
+            function_id: Some(function_id.into()),
+            message: message.into(),
+        }
+    }
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum EvalError {
+    #[error("invalid request: {0}")]
+    InvalidRequest(String),
+    #[error("evaluation not found: {0}")]
+    NotFound(String),
+    #[error("evaluation conflict: {0}")]
+    Conflict(String),
+    #[error("dependency error: {0}")]
+    Dependency(String),
+    #[error("state error: {0}")]
+    State(String),
+    #[error("serialization error: {0}")]
+    Serialization(String),
+}
+
+impl From<serde_json::Error> for EvalError {
+    fn from(error: serde_json::Error) -> Self {
+        Self::Serialization(error.to_string())
+    }
+}
+
+impl From<EvalError> for SdkError {
+    fn from(error: EvalError) -> Self {
+        SdkError::Handler(error.to_string())
+    }
+}
diff --git a/eval/src/events.rs b/eval/src/events.rs
new file mode 100644
index 000000000..3c9fa29ed
--- /dev/null
+++ b/eval/src/events.rs
@@ -0,0 +1,154 @@
+use std::collections::HashMap;
+use std::sync::{Arc, Mutex};
+
+use async_trait::async_trait;
+use iii_sdk::errors::Error;
+use iii_sdk::protocol::TriggerRequest;
+use iii_sdk::trigger::{TriggerConfig, TriggerHandler};
+use iii_sdk::{IIIClient, RegisterTriggerType, TriggerAction};
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+use serde_json::{json, Value};
+
+use crate::contract::EvalStatusV1;
+
+pub const COMPLETED: &str = "eval::completed";
+
+#[derive(Debug, Clone, Default, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct CompletedBindingConfigV1 {
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub evaluation_id: Option<String>,
+}
+
+#[derive(Debug, Clone)]
+struct Binding {
+    function_id: String,
+    evaluation_id: Option<String>,
+    metadata: Option<Value>,
+}
+
+#[derive(Clone, Default)]
+struct SubscriberSet {
+    inner: Arc<Mutex<HashMap<String, Binding>>>,
+}
+
+impl SubscriberSet {
+    fn insert(&self, config: TriggerConfig) -> Result<(), String> {
+        let raw = if config.config.is_null() {
+            json!({})
+        } else {
+            config.config
+        };
+        let filter: CompletedBindingConfigV1 = serde_json::from_value(raw)
+            .map_err(|error| format!("invalid eval::completed config: {error}"))?;
+        self.inner
+            .lock()
+            .unwrap_or_else(|poisoned| poisoned.into_inner())
+            .insert(
+                config.id,
+                Binding {
+                    function_id: config.function_id,
+                    evaluation_id: filter.evaluation_id,
+                    metadata: config.metadata,
+                },
+            );
+        Ok(())
+    }
+
+    fn remove(&self, id: &str) {
+        self.inner
+            .lock()
+            .unwrap_or_else(|poisoned| poisoned.into_inner())
+            .remove(id);
+    }
+
+    fn snapshot(&self) -> Vec<Binding> {
+        self.inner
+            .lock()
+            .unwrap_or_else(|poisoned| poisoned.into_inner())
+            .values()
+            .cloned()
+            .collect()
+    }
+}
+
+struct CompletedTriggerHandler {
+    set: SubscriberSet,
+}
+
+#[async_trait]
+impl TriggerHandler for CompletedTriggerHandler {
+    async fn register_trigger(&self, config: TriggerConfig) -> Result<(), Error> {
+        self.set.insert(config).map_err(Error::Handler)
+    }
+
+    async fn unregister_trigger(&self, config: TriggerConfig) -> Result<(), Error> {
+        self.set.remove(&config.id);
+        Ok(())
+    }
+}
+
+#[derive(Clone)]
+pub struct EvalEvents {
+    iii: Arc<IIIClient>,
+    completed: SubscriberSet,
+}
+
+impl EvalEvents {
+    pub fn register(iii: &Arc<IIIClient>) -> Self {
+        let completed = SubscriberSet::default();
+        let _ = iii.register_trigger_type(
+            RegisterTriggerType::new(
+                COMPLETED,
+                "An evaluation reached a terminal status.",
+                CompletedTriggerHandler {
+                    set: completed.clone(),
+                },
+            )
+            .trigger_request_format::<CompletedBindingConfigV1>(),
+        );
+        Self {
+            iii: iii.clone(),
+            completed,
+        }
+    }
+
+    pub async fn emit_completed(
+        &self,
+        evaluation_id: &str,
+        status: EvalStatusV1,
+        eligible: Option<bool>,
+    ) {
+        let mut payload = json!({
+            "evaluation_id": evaluation_id,
+            "status": status,
+            "timestamp": crate::ids::now_ms(),
+        });
+        if let Some(eligible) = eligible {
+            payload["eligible"] = Value::Bool(eligible);
+        }
+        for binding in self.completed.snapshot() {
+            if binding
+                .evaluation_id
+                .as_deref()
+                .is_some_and(|filter| filter != evaluation_id)
+            {
+                continue;
+            }
+            let request = TriggerRequest {
+                function_id: binding.function_id,
+                payload: payload.clone(),
+                action: Some(TriggerAction::Void),
+                timeout_ms: None,
+            };
+            let result = match binding.metadata {
+                Some(metadata) => self.iii.trigger(request.metadata(metadata)).await,
+                None => self.iii.trigger(request).await,
+            };
+            if let Err(error) = result {
+                tracing::warn!(%evaluation_id, %error, "eval::completed delivery failed");
+            }
+        }
+    }
+}
diff --git a/eval/src/functions.rs b/eval/src/functions.rs
new file mode 100644
index 000000000..417686377
--- /dev/null
+++ b/eval/src/functions.rs
@@ -0,0 +1,195 @@
+use std::sync::Arc;
+
+use iii_sdk::errors::Error;
+use iii_sdk::{IIIClient, RegisterFunction};
+use serde_json::json;
+
+use crate::contract::{
+    EvalListRequestV1, EvalRerunRequestV1, EvalStartRequestV1, EvaluationIdRequestV1,
+    EvaluatorInputV1, StepRequestV1, SweepEventV1, WakeEventV1,
+};
+use crate::runtime::Deps;
+
+pub const START_ID: &str = "eval::start";
+pub const LIST_ID: &str = "eval::list";
+pub const RERUN_ID: &str = "eval::rerun";
+pub const STATUS_ID: &str = "eval::status";
+pub const RESULT_ID: &str = "eval::result";
+pub const CANCEL_ID: &str = "eval::cancel";
+pub const DELETE_ID: &str = "eval::delete";
+pub const EXACT_ID: &str = "eval::assert::exact";
+pub const NORMALIZED_TEXT_ID: &str = "eval::assert::normalized_text";
+pub const STEP_ID: &str = "eval::step";
+pub const WAKE_ID: &str = "eval::on-turn-completed";
+pub const SWEEP_ID: &str = "eval::sweep";
+
+pub fn register_all(iii: &Arc<IIIClient>, deps: &Deps) {
+    let current = deps.clone();
+    iii.register_function(
+        START_ID,
+        RegisterFunction::new_async(move |request: EvalStartRequestV1| {
+            let deps = current.clone();
+            async move {
+                crate::runtime::start(&deps, request)
+                    .await
+                    .map_err(Error::from)
+            }
+        })
+        .description(
+            "Start a durable same-model A/B evaluation. Exactly one dimension may change: \
+             `prompt` or `system_prompt`. Returns evaluation_id immediately; use eval::status, \
+             eval::result, or the eval::completed trigger instead of polling in a tight loop.",
+        ),
+    );
+
+    let current = deps.clone();
+    iii.register_function(
+        RERUN_ID,
+        RegisterFunction::new_async(move |request: EvalRerunRequestV1| {
+            let deps = current.clone();
+            async move {
+                crate::runtime::rerun(&deps, request)
+                    .await
+                    .map_err(Error::from)
+            }
+        })
+        .description(
+            "Repeat a terminal evaluation from its persisted request. Set reverse_order to invert \
+             the balanced A/B order. The rerun receives a new evaluation ID and fresh sessions.",
+        ),
+    );
+
+    let current = deps.clone();
+    iii.register_function(
+        LIST_ID,
+        RegisterFunction::new_async(move |request: EvalListRequestV1| {
+            let deps = current.clone();
+            async move {
+                crate::runtime::list(&deps, request)
+                    .await
+                    .map_err(Error::from)
+            }
+        })
+        .description("List recent evaluations as lightweight summaries, newest first."),
+    );
+
+    let current = deps.clone();
+    iii.register_function(
+        STATUS_ID,
+        RegisterFunction::new_async(move |request: EvaluationIdRequestV1| {
+            let deps = current.clone();
+            async move {
+                crate::runtime::status(&deps, request)
+                    .await
+                    .map_err(Error::from)
+            }
+        })
+        .description("Read evaluation progress without loading the complete report."),
+    );
+
+    let current = deps.clone();
+    iii.register_function(
+        RESULT_ID,
+        RegisterFunction::new_async(move |request: EvaluationIdRequestV1| {
+            let deps = current.clone();
+            async move {
+                crate::runtime::result(&deps, request)
+                    .await
+                    .map_err(Error::from)
+            }
+        })
+        .description(
+            "Read an evaluation result. The report is present only after terminal completion.",
+        ),
+    );
+
+    let current = deps.clone();
+    iii.register_function(
+        CANCEL_ID,
+        RegisterFunction::new_async(move |request: EvaluationIdRequestV1| {
+            let deps = current.clone();
+            async move {
+                crate::runtime::cancel(&deps, request)
+                    .await
+                    .map_err(Error::from)
+            }
+        })
+        .description("Cancel an evaluation and its active harness turn."),
+    );
+
+    let current = deps.clone();
+    iii.register_function(
+        DELETE_ID,
+        RegisterFunction::new_async(move |request: EvaluationIdRequestV1| {
+            let deps = current.clone();
+            async move {
+                crate::runtime::delete(&deps, request)
+                    .await
+                    .map_err(Error::from)
+            }
+        })
+        .description("Delete a terminal evaluation report and its session indexes."),
+    );
+
+    iii.register_function(
+        EXACT_ID,
+        RegisterFunction::new_async(move |input: EvaluatorInputV1| async move {
+            crate::runtime::exact(input).map_err(Error::from)
+        })
+        .description(
+            "Built-in deterministic evaluator: deep-compare output with arguments.expected.",
+        ),
+    );
+
+    iii.register_function(
+        NORMALIZED_TEXT_ID,
+        RegisterFunction::new_async(move |input: EvaluatorInputV1| async move {
+            crate::runtime::normalized_text(input).map_err(Error::from)
+        })
+        .description(
+            "Built-in deterministic evaluator: compare text after normalizing case, \
+             whitespace, and surrounding punctuation.",
+        ),
+    );
+
+    let current = deps.clone();
+    iii.register_function(
+        STEP_ID,
+        RegisterFunction::new_async(move |request: StepRequestV1| {
+            let deps = current.clone();
+            async move {
+                crate::runtime::step(&deps, request)
+                    .await
+                    .map_err(Error::from)
+            }
+        })
+        .description("Internal durable evaluation step.")
+        .metadata(json!({ "internal": true, "trace_hidden": true })),
+    );
+
+    let current = deps.clone();
+    iii.register_function(
+        WAKE_ID,
+        RegisterFunction::new_async(move |event: WakeEventV1| {
+            let deps = current.clone();
+            async move {
+                crate::runtime::wake(&deps, event)
+                    .await
+                    .map_err(Error::from)
+            }
+        })
+        .description("Internal harness turn-completed wake-up.")
+        .metadata(json!({ "internal": true, "trace_hidden": true })),
+    );
+
+    let current = deps.clone();
+    iii.register_function(
+        SWEEP_ID,
+        RegisterFunction::new_async(move |_event: SweepEventV1| {
+            let deps = current.clone();
+            async move { crate::runtime::sweep(&deps).await.map_err(Error::from) }
+        })
+        .description("Internal recovery and timeout sweep.")
+        .metadata(json!({ "internal": true, "trace_hidden": true })),
+    );
+}
diff --git a/eval/src/ids.rs b/eval/src/ids.rs
new file mode 100644
index 000000000..4afbd37e7
--- /dev/null
+++ b/eval/src/ids.rs
@@ -0,0 +1,51 @@
+use sha2::{Digest, Sha256};
+use uuid::Uuid;
+
+use crate::contract::VariantRoleV1;
+
+pub fn evaluation_id() -> String {
+    format!("eval_{}", Uuid::new_v4().simple())
+}
+
+pub fn run_id(role: VariantRoleV1, iteration: u32) -> String {
+    format!("{}-{iteration}", role.as_str())
+}
+
+pub fn session_id(evaluation_id: &str, role: VariantRoleV1, iteration: u32) -> String {
+    let suffix = match role {
+        VariantRoleV1::Control => "c",
+        VariantRoleV1::Treatment => "t",
+    };
+    format!("{evaluation_id}_{suffix}_{iteration}")
+}
+
+pub fn send_idempotency_key(evaluation_id: &str, role: VariantRoleV1, iteration: u32) -> String {
+    format!("eval:{evaluation_id}:{}:{iteration}:send", role.as_str())
+}
+
+pub fn finalization_idempotency_key(
+    evaluation_id: &str,
+    role: VariantRoleV1,
+    iteration: u32,
+) -> String {
+    format!(
+        "eval:{evaluation_id}:{}:{iteration}:finalize",
+        role.as_str()
+    )
+}
+
+pub fn sha256_json(value: &serde_json::Value) -> String {
+    let bytes = serde_json::to_vec(value).unwrap_or_default();
+    format!("{:x}", Sha256::digest(bytes))
+}
+
+pub fn sha256_text(value: &str) -> String {
+    format!("{:x}", Sha256::digest(value.as_bytes()))
+}
+
+pub fn now_ms() -> i64 {
+    std::time::SystemTime::now()
+        .duration_since(std::time::UNIX_EPOCH)
+        .map(|duration| duration.as_millis() as i64)
+        .unwrap_or(0)
+}
diff --git a/eval/src/lib.rs b/eval/src/lib.rs
new file mode 100644
index 000000000..217e01c69
--- /dev/null
+++ b/eval/src/lib.rs
@@ -0,0 +1,16 @@
+//! Durable A/B evaluation of prompts and system prompts over harness turns.
+
+pub mod contract;
+pub mod error;
+pub mod events;
+pub mod functions;
+pub mod ids;
+pub mod limits;
+pub mod locks;
+pub mod manifest;
+pub mod queue;
+pub mod report;
+pub mod runtime;
+pub mod state;
+pub mod surface;
+pub mod ui;
diff --git a/eval/src/limits.rs b/eval/src/limits.rs
new file mode 100644
index 000000000..6d45509a0
--- /dev/null
+++ b/eval/src/limits.rs
@@ -0,0 +1,261 @@
+use std::time::Duration;
+
+use harness::functions::metrics::SessionMetricsResponseV1;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::error::{EvalError, EvalFailureV1, EvalPhaseV1};
+
+pub const DEFAULT_INVOCATION_TIMEOUT_SECONDS: u64 = 120;
+pub const DEFAULT_SCENARIO_TIMEOUT_SECONDS: u64 = 600;
+pub const DEFAULT_MAX_TURNS: u32 = 1_000;
+pub const DEFAULT_MAX_OUTPUT_TOKENS_PER_CALL: u64 = 8_192;
+pub const AGENT_QUALITY_DEFAULT_MAX_TOTAL_TOKENS: u64 = 100_000;
+
+#[derive(Debug, Clone, Copy, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvalLimitsV1 {
+    #[serde(default)]
+    pub execution: ExecutionLimitsV1,
+    #[serde(default)]
+    pub evaluation: EvaluationLimitsV1,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct ExecutionLimitsV1 {
+    #[serde(default = "default_invocation_timeout_seconds")]
+    pub invocation_timeout_seconds: u64,
+    #[serde(default = "default_scenario_timeout_seconds")]
+    pub scenario_timeout_seconds: u64,
+    #[serde(default = "default_max_turns")]
+    pub max_turns: u32,
+    #[serde(default = "default_max_output_tokens_per_call")]
+    pub max_output_tokens_per_call: u64,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub max_total_tokens: Option<u64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub max_cost_usd: Option<f64>,
+}
+
+#[derive(Debug, Clone, Copy, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvaluationLimitsV1 {
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub max_function_call_errors: Option<u64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub max_error_spans: Option<u64>,
+}
+
+impl Default for ExecutionLimitsV1 {
+    fn default() -> Self {
+        Self {
+            invocation_timeout_seconds: default_invocation_timeout_seconds(),
+            scenario_timeout_seconds: default_scenario_timeout_seconds(),
+            max_turns: default_max_turns(),
+            max_output_tokens_per_call: default_max_output_tokens_per_call(),
+            max_total_tokens: None,
+            max_cost_usd: None,
+        }
+    }
+}
+
+fn default_invocation_timeout_seconds() -> u64 {
+    DEFAULT_INVOCATION_TIMEOUT_SECONDS
+}
+
+fn default_scenario_timeout_seconds() -> u64 {
+    DEFAULT_SCENARIO_TIMEOUT_SECONDS
+}
+
+fn default_max_turns() -> u32 {
+    DEFAULT_MAX_TURNS
+}
+
+fn default_max_output_tokens_per_call() -> u64 {
+    DEFAULT_MAX_OUTPUT_TOKENS_PER_CALL
+}
+
+impl EvalLimitsV1 {
+    pub fn validate(self) -> Result<(), EvalError> {
+        for (name, value) in [
+            (
+                "limits.execution.invocation_timeout_seconds",
+                self.execution.invocation_timeout_seconds,
+            ),
+            (
+                "limits.execution.scenario_timeout_seconds",
+                self.execution.scenario_timeout_seconds,
+            ),
+            (
+                "limits.execution.max_output_tokens_per_call",
+                self.execution.max_output_tokens_per_call,
+            ),
+        ] {
+            if value == 0 {
+                return Err(EvalError::InvalidRequest(format!(
+                    "{name} must be greater than zero"
+                )));
+            }
+        }
+        if self.execution.max_turns == 0 {
+            return Err(EvalError::InvalidRequest(
+                "limits.execution.max_turns must be greater than zero".into(),
+            ));
+        }
+        if self.execution.max_total_tokens == Some(0) {
+            return Err(EvalError::InvalidRequest(
+                "limits.execution.max_total_tokens must be greater than zero when supplied".into(),
+            ));
+        }
+        if self
+            .execution
+            .max_cost_usd
+            .is_some_and(|value| !value.is_finite() || value < 0.0)
+        {
+            return Err(EvalError::InvalidRequest(
+                "limits.execution.max_cost_usd must be a finite non-negative number".into(),
+            ));
+        }
+        Ok(())
+    }
+
+    pub fn failures(self, metrics: &SessionMetricsResponseV1) -> Vec<EvalFailureV1> {
+        let mut failures = Vec::new();
+        if let Some(limit) = self.execution.max_total_tokens {
+            match (metrics.totals.input_tokens, metrics.totals.output_tokens) {
+                (Some(input), Some(output)) => {
+                    let total = input.saturating_add(output);
+                    if total > limit {
+                        failures.push(EvalFailureV1::new(
+                            EvalPhaseV1::Limit,
+                            format!("total tokens {total} exceeded limit {limit}"),
+                        ));
+                    }
+                }
+                _ => failures.push(EvalFailureV1::function(
+                    EvalPhaseV1::Collect,
+                    "harness::metrics",
+                    "total-token validation requires input_tokens and output_tokens",
+                )),
+            }
+        }
+
+        if let Some(limit) = self.execution.max_cost_usd {
+            match metrics.totals.cost_usd {
+                Some(cost) if cost > limit => failures.push(EvalFailureV1::new(
+                    EvalPhaseV1::Limit,
+                    format!("cost ${cost:.6} exceeded limit ${limit:.6}"),
+                )),
+                Some(_) => {}
+                None => failures.push(EvalFailureV1::function(
+                    EvalPhaseV1::Collect,
+                    "harness::metrics",
+                    "cost validation requires cost_usd from every model generation",
+                )),
+            }
+        }
+
+        if let Some(limit) = self.evaluation.max_function_call_errors {
+            if metrics.totals.function_call_errors > limit {
+                failures.push(EvalFailureV1::new(
+                    EvalPhaseV1::Limit,
+                    format!(
+                        "function call errors {} exceeded limit {limit}",
+                        metrics.totals.function_call_errors
+                    ),
+                ));
+            }
+        }
+
+        if let Some(limit) = self.evaluation.max_error_spans {
+            match metrics.traces.as_ref() {
+                Some(traces) if traces.error_span_count > limit => {
+                    failures.push(EvalFailureV1::new(
+                        EvalPhaseV1::Limit,
+                        format!(
+                            "error spans {} exceeded limit {limit}",
+                            traces.error_span_count
+                        ),
+                    ));
+                }
+                Some(_) => {}
+                None => failures.push(EvalFailureV1::function(
+                    EvalPhaseV1::Collect,
+                    "harness::metrics",
+                    "error-span validation requires in-memory observability metrics",
+                )),
+            }
+        }
+        failures
+    }
+}
+
+impl ExecutionLimitsV1 {
+    pub fn invocation_timeout(self) -> Duration {
+        Duration::from_secs(self.invocation_timeout_seconds)
+    }
+
+    pub fn scenario_timeout(self) -> Duration {
+        Duration::from_secs(self.scenario_timeout_seconds)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use harness::functions::metrics::SessionUsageTotalsV1;
+
+    #[test]
+    fn defaults_leave_optional_tree_budgets_unset() {
+        let limits = EvalLimitsV1::default();
+        assert_eq!(limits.execution.max_turns, 1_000);
+        assert_eq!(limits.execution.max_output_tokens_per_call, 8_192);
+        assert_eq!(limits.execution.max_total_tokens, None);
+        assert_eq!(limits.evaluation.max_function_call_errors, None);
+        limits.validate().unwrap();
+    }
+
+    #[test]
+    fn partial_limits_resolve_operational_defaults() {
+        let limits: EvalLimitsV1 =
+            serde_json::from_value(serde_json::json!({"execution": {}, "evaluation": {}})).unwrap();
+
+        assert_eq!(limits, EvalLimitsV1::default());
+    }
+
+    #[test]
+    fn function_call_errors_are_only_limited_when_explicitly_configured() {
+        let metrics = SessionMetricsResponseV1 {
+            root_session_id: "session".into(),
+            complete: true,
+            totals: SessionUsageTotalsV1 {
+                function_call_errors: 3,
+                ..Default::default()
+            },
+            by_session: Vec::new(),
+            traces: None,
+        };
+
+        assert!(EvalLimitsV1::default().failures(&metrics).is_empty());
+
+        let mut limits = EvalLimitsV1::default();
+        limits.evaluation.max_function_call_errors = Some(0);
+        assert_eq!(limits.failures(&metrics).len(), 1);
+    }
+
+    #[test]
+    fn validation_rejects_zero_and_invalid_cost() {
+        let mut limits = EvalLimitsV1::default();
+        limits.execution.scenario_timeout_seconds = 0;
+        assert!(limits.validate().is_err());
+
+        let mut limits = EvalLimitsV1::default();
+        limits.execution.max_total_tokens = Some(0);
+        assert!(limits.validate().is_err());
+
+        let mut limits = EvalLimitsV1::default();
+        limits.execution.max_cost_usd = Some(f64::NAN);
+        assert!(limits.validate().is_err());
+    }
+}
diff --git a/eval/src/locks.rs b/eval/src/locks.rs
new file mode 100644
index 000000000..b13cb959a
--- /dev/null
+++ b/eval/src/locks.rs
@@ -0,0 +1,29 @@
+use std::collections::HashMap;
+use std::sync::{Arc, Mutex, Weak};
+
+use tokio::sync::OwnedMutexGuard;
+
+#[derive(Clone, Default)]
+pub struct EvalLocks {
+    map: Arc<Mutex<HashMap<String, Weak<tokio::sync::Mutex<()>>>>>,
+}
+
+impl EvalLocks {
+    pub async fn guard(&self, evaluation_id: &str) -> OwnedMutexGuard<()> {
+        let lock = {
+            let mut map = self
+                .map
+                .lock()
+                .unwrap_or_else(|poisoned| poisoned.into_inner());
+            match map.get(evaluation_id).and_then(Weak::upgrade) {
+                Some(lock) => lock,
+                None => {
+                    let lock = Arc::new(tokio::sync::Mutex::new(()));
+                    map.insert(evaluation_id.to_string(), Arc::downgrade(&lock));
+                    lock
+                }
+            }
+        };
+        lock.lock_owned().await
+    }
+}
diff --git a/eval/src/main.rs b/eval/src/main.rs
new file mode 100644
index 000000000..e16597e6c
--- /dev/null
+++ b/eval/src/main.rs
@@ -0,0 +1,125 @@
+use std::sync::Arc;
+
+use anyhow::{Context, Result};
+use clap::Parser;
+use iii_helpers::observability::OtelConfig;
+use iii_sdk::protocol::RegisterTriggerInput;
+use iii_sdk::runtime::WorkerMetadata;
+use iii_sdk::trigger::Trigger;
+use iii_sdk::{register_worker, InitOptions};
+use serde_json::json;
+
+use eval::events::EvalEvents;
+use eval::locks::EvalLocks;
+use eval::runtime::Deps;
+use eval::{functions, manifest, queue, state, ui};
+
+#[derive(Parser, Debug)]
+#[command(
+    name = "eval",
+    about = "Durable same-model A/B prompt evaluation over harness turns."
+)]
+struct Cli {
+    #[arg(long, env = "III_URL", default_value = "ws://127.0.0.1:49134")]
+    url: String,
+    #[arg(long)]
+    manifest: bool,
+}
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    tracing_subscriber::fmt()
+        .with_env_filter(
+            tracing_subscriber::EnvFilter::try_from_default_env()
+                .unwrap_or_else(|_| tracing_subscriber::EnvFilter::new("info")),
+        )
+        .init();
+
+    let cli = Cli::parse();
+    if cli.manifest {
+        println!(
+            "{}",
+            serde_json::to_string_pretty(&manifest::build_manifest()).unwrap()
+        );
+        return Ok(());
+    }
+
+    let iii = Arc::new(register_worker(
+        &cli.url,
+        InitOptions {
+            metadata: Some(WorkerMetadata {
+                runtime: "rust".into(),
+                version: env!("CARGO_PKG_VERSION").into(),
+                name: "eval".into(),
+                os: std::env::consts::OS.into(),
+                pid: Some(std::process::id()),
+                telemetry: None,
+                ..WorkerMetadata::default()
+            }),
+            otel: Some(OtelConfig::default()),
+            ..InitOptions::default()
+        },
+    ));
+
+    let events = EvalEvents::register(&iii);
+    let deps = Deps {
+        iii: iii.clone(),
+        locks: EvalLocks::default(),
+        events,
+    };
+    functions::register_all(&iii, &deps);
+    ui::register(&iii);
+    queue::ensure_run_queue(&iii)
+        .await
+        .context("ensuring eval-run queue")?;
+
+    let _trigger_handles = bind_triggers(&iii);
+
+    match state::list_jobs(&iii).await {
+        Ok(jobs) => {
+            for job in jobs.iter().filter(|job| !job.status.is_terminal()) {
+                if let Err(error) = queue::enqueue_step(&iii, &job.evaluation_id, job.step).await {
+                    tracing::warn!(
+                        evaluation_id = %job.evaluation_id,
+                        %error,
+                        "eval startup recovery could not enqueue step"
+                    );
+                }
+            }
+        }
+        Err(error) => tracing::warn!(%error, "eval startup recovery could not list jobs"),
+    }
+
+    tracing::info!("eval ready");
+    tokio::signal::ctrl_c().await?;
+    iii.shutdown_async().await;
+    Ok(())
+}
+
+fn bind_triggers(iii: &Arc<iii_sdk::IIIClient>) -> Vec<Trigger> {
+    let mut handles = Vec::new();
+    for (trigger_type, function_id, config) in [
+        ("harness::turn-completed", functions::WAKE_ID, json!({})),
+        (
+            "cron",
+            functions::SWEEP_ID,
+            json!({ "expression": "*/15 * * * * *" }),
+        ),
+    ] {
+        match iii.register_trigger(RegisterTriggerInput {
+            trigger_type: trigger_type.into(),
+            function_id: function_id.into(),
+            config,
+            metadata: None,
+        }) {
+            Ok(handle) => handles.push(handle),
+            Err(error) => tracing::warn!(
+                trigger_type,
+                function_id,
+                %error,
+                "eval trigger binding failed; recovery remains available on restart"
+            ),
+        }
+    }
+    handles
+}
diff --git a/eval/src/manifest.rs b/eval/src/manifest.rs
new file mode 100644
index 000000000..c6d16e044
--- /dev/null
+++ b/eval/src/manifest.rs
@@ -0,0 +1,22 @@
+use serde::Serialize;
+
+#[derive(Serialize)]
+pub struct ModuleManifest {
+    pub name: String,
+    pub version: String,
+    pub description: String,
+    pub default_config: serde_json::Value,
+    pub supported_targets: Vec<String>,
+}
+
+pub fn build_manifest() -> ModuleManifest {
+    ModuleManifest {
+        name: env!("CARGO_PKG_NAME").to_string(),
+        version: env!("CARGO_PKG_VERSION").to_string(),
+        description:
+            "Durable same-model A/B evaluation for prompts and system prompts over harness turns."
+                .to_string(),
+        default_config: serde_json::json!({}),
+        supported_targets: vec![env!("TARGET").to_string()],
+    }
+}
diff --git a/eval/src/queue.rs b/eval/src/queue.rs
new file mode 100644
index 000000000..bc723d548
--- /dev/null
+++ b/eval/src/queue.rs
@@ -0,0 +1,89 @@
+use std::time::Duration;
+
+use iii_sdk::protocol::TriggerRequest;
+use iii_sdk::{IIIClient, TriggerAction};
+use serde_json::{json, Value};
+
+use crate::contract::StepRequestV1;
+use crate::error::EvalError;
+
+pub const RUN_QUEUE: &str = "eval-run";
+const DEFINE_TIMEOUT_MS: u64 = 5_000;
+const DEFINE_ATTEMPTS: u32 = 20;
+const DEFINE_RETRY_BACKOFF_MS: u64 = 250;
+
+pub async fn ensure_run_queue(iii: &IIIClient) -> Result<(), EvalError> {
+    let payload = run_queue_definition();
+    let mut last_error = String::new();
+    for attempt in 1..=DEFINE_ATTEMPTS {
+        match iii
+            .trigger(TriggerRequest {
+                function_id: "queue::define".into(),
+                payload: payload.clone(),
+                action: None,
+                timeout_ms: Some(DEFINE_TIMEOUT_MS),
+            })
+            .await
+        {
+            Ok(_) => return Ok(()),
+            Err(error) => {
+                last_error = error.to_string();
+                if attempt < DEFINE_ATTEMPTS {
+                    tokio::time::sleep(Duration::from_millis(DEFINE_RETRY_BACKOFF_MS)).await;
+                }
+            }
+        }
+    }
+    Err(EvalError::Dependency(format!(
+        "queue::define failed after {DEFINE_ATTEMPTS} attempts: {last_error}"
+    )))
+}
+
+pub async fn enqueue_step(
+    iii: &IIIClient,
+    evaluation_id: &str,
+    step: u64,
+) -> Result<(), EvalError> {
+    let payload = serde_json::to_value(StepRequestV1 {
+        evaluation_id: evaluation_id.into(),
+        step,
+    })?;
+    iii.trigger(TriggerRequest {
+        function_id: "eval::step".into(),
+        payload,
+        action: Some(TriggerAction::Enqueue {
+            queue: RUN_QUEUE.into(),
+        }),
+        timeout_ms: None,
+    })
+    .await
+    .map(|_| ())
+    .map_err(|error| EvalError::Dependency(format!("enqueue eval::step: {error}")))
+}
+
+fn run_queue_definition() -> Value {
+    json!({
+        "queue": RUN_QUEUE,
+        "config": {
+            "type": "fifo",
+            "message_group_field": "evaluation_id",
+            "concurrency": 4,
+            "max_retries": 3,
+            "backoff_ms": 1_000,
+            "poll_interval_ms": 100
+        }
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn eval_queue_is_fifo_and_grouped_by_evaluation() {
+        let definition = run_queue_definition();
+        assert_eq!(definition["queue"], "eval-run");
+        assert_eq!(definition["config"]["message_group_field"], "evaluation_id");
+        assert_eq!(definition["config"]["type"], "fifo");
+    }
+}
diff --git a/eval/src/report.rs b/eval/src/report.rs
new file mode 100644
index 000000000..4bd09430e
--- /dev/null
+++ b/eval/src/report.rs
@@ -0,0 +1,699 @@
+use harness::functions::metrics::SessionMetricsResponseV1;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+use serde_json::Value;
+
+use crate::contract::{
+    ComparisonDimensionV1, EvalModelConfigV1, EvalRunStatusV1, ExecutionOrderV1,
+    NormalizedEvalRequestV1, VariantRoleV1,
+};
+use crate::error::EvalFailureV1;
+use crate::ids;
+use crate::limits::EvalLimitsV1;
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvalBenchmarkV1 {
+    pub wall_time_ms: u64,
+    pub sessions: u64,
+    pub turns: u64,
+    pub function_calls: u64,
+    pub function_call_errors: u64,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub input_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub output_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub total_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub cache_read_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub cache_write_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub reasoning_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub cost_usd: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub trace_count: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub span_count: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub error_span_count: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub trace_duration_ms: Option<u64>,
+}
+
+impl EvalBenchmarkV1 {
+    pub fn from_metrics(metrics: &SessionMetricsResponseV1, wall_time_ms: u64) -> Option<Self> {
+        metrics.complete.then(|| Self {
+            wall_time_ms,
+            sessions: metrics.by_session.len() as u64,
+            turns: metrics.totals.turns,
+            function_calls: metrics.totals.function_calls,
+            function_call_errors: metrics.totals.function_call_errors,
+            input_tokens: metrics.totals.input_tokens,
+            output_tokens: metrics.totals.output_tokens,
+            total_tokens: metrics
+                .totals
+                .input_tokens
+                .zip(metrics.totals.output_tokens)
+                .map(|(input, output)| input.saturating_add(output)),
+            cache_read_tokens: metrics.totals.cache_read_tokens,
+            cache_write_tokens: metrics.totals.cache_write_tokens,
+            reasoning_tokens: metrics.totals.reasoning_tokens,
+            cost_usd: metrics.totals.cost_usd,
+            trace_count: metrics.traces.as_ref().map(|traces| traces.trace_count),
+            span_count: metrics.traces.as_ref().map(|traces| traces.span_count),
+            error_span_count: metrics
+                .traces
+                .as_ref()
+                .map(|traces| traces.error_span_count),
+            trace_duration_ms: metrics.traces.as_ref().map(|traces| traces.duration_ms),
+        })
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvalRunReportV1 {
+    pub run_id: String,
+    pub role: VariantRoleV1,
+    pub iteration: u32,
+    #[serde(default)]
+    pub execution_position: u32,
+    #[serde(default)]
+    pub pair_position: u8,
+    pub status: EvalRunStatusV1,
+    pub session_id: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub turn_id: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub passed: Option<bool>,
+    pub started_at: i64,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub completed_at: Option<i64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub output: Option<Value>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub metrics: Option<SessionMetricsResponseV1>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub evaluation: Option<crate::contract::EvaluatorResponseV1>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub benchmark: Option<EvalBenchmarkV1>,
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub failures: Vec<EvalFailureV1>,
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct VariantArtifactV1 {
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub label: Option<String>,
+    pub prompt_sha256: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub system_prompt_sha256: Option<String>,
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvaluatorArtifactV1 {
+    pub function_id: String,
+    pub arguments_sha256: String,
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct SharedArtifactsV1 {
+    pub model_sha256: String,
+    pub function_policy_sha256: String,
+    pub limits_sha256: String,
+    pub output_sha256: String,
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct RoleOrderSensitivityV1 {
+    pub first_runs: u32,
+    pub first_passed: u32,
+    pub second_runs: u32,
+    pub second_passed: u32,
+    pub differs: bool,
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct OrderSensitivityV1 {
+    pub detected: bool,
+    pub control: RoleOrderSensitivityV1,
+    pub treatment: RoleOrderSensitivityV1,
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct VariantAggregateV1 {
+    pub runs: u32,
+    pub evaluated_runs: u32,
+    pub passed: u32,
+    pub pass_rate: f64,
+    pub benchmarked_runs: u32,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_score: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_wall_time_ms: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_input_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_output_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_total_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_reasoning_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_function_calls: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_function_call_errors: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_cost_usd: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_trace_count: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_span_count: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_error_span_count: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_trace_duration_ms: Option<u64>,
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct AggregateDeltaV1 {
+    pub pass_rate: f64,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_score: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_wall_time_ms: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_total_tokens: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_reasoning_tokens: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_function_calls: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_function_call_errors: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_cost_usd: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_trace_count: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_span_count: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_error_span_count: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_trace_duration_ms: Option<f64>,
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvalReportV1 {
+    pub schema_version: String,
+    pub evaluation_id: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub source_evaluation_id: Option<String>,
+    pub execution_order_policy: ExecutionOrderV1,
+    pub effective_execution_order: Vec<String>,
+    pub dimension: ComparisonDimensionV1,
+    pub model: EvalModelConfigV1,
+    pub limits: EvalLimitsV1,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub evaluator: Option<EvaluatorArtifactV1>,
+    pub control: VariantArtifactV1,
+    pub treatment: VariantArtifactV1,
+    pub shared_artifacts: SharedArtifactsV1,
+    pub control_aggregate: VariantAggregateV1,
+    pub treatment_aggregate: VariantAggregateV1,
+    /// Treatment minus control. Negative efficiency deltas are improvements.
+    pub delta: AggregateDeltaV1,
+    pub order_sensitivity: OrderSensitivityV1,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub eligible: Option<bool>,
+    pub runs: Vec<EvalRunReportV1>,
+    pub created_at: i64,
+    pub completed_at: i64,
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct EvalProgressV1 {
+    pub effective_execution_order: Vec<String>,
+    pub control_aggregate: VariantAggregateV1,
+    pub treatment_aggregate: VariantAggregateV1,
+    pub delta: AggregateDeltaV1,
+    pub order_sensitivity: OrderSensitivityV1,
+    pub runs: Vec<EvalRunReportV1>,
+}
+
+pub fn build_progress(runs: &[EvalRunReportV1]) -> EvalProgressV1 {
+    let control_aggregate = aggregate(runs, VariantRoleV1::Control);
+    let treatment_aggregate = aggregate(runs, VariantRoleV1::Treatment);
+    EvalProgressV1 {
+        effective_execution_order: effective_order(runs),
+        delta: delta(&control_aggregate, &treatment_aggregate),
+        order_sensitivity: order_sensitivity(runs),
+        control_aggregate,
+        treatment_aggregate,
+        runs: runs.to_vec(),
+    }
+}
+
+pub fn build_report(
+    evaluation_id: &str,
+    request: &NormalizedEvalRequestV1,
+    runs: Vec<EvalRunReportV1>,
+    created_at: i64,
+    completed_at: i64,
+) -> EvalReportV1 {
+    let control_aggregate = aggregate(&runs, VariantRoleV1::Control);
+    let treatment_aggregate = aggregate(&runs, VariantRoleV1::Treatment);
+    let eligible = request.evaluator.as_ref().map(|_| {
+        treatment_aggregate.runs > 0
+            && treatment_aggregate.evaluated_runs == treatment_aggregate.runs
+            && treatment_aggregate.passed == treatment_aggregate.evaluated_runs
+            && treatment_aggregate.passed >= control_aggregate.passed
+    });
+    let delta = delta(&control_aggregate, &treatment_aggregate);
+
+    EvalReportV1 {
+        schema_version: "1".into(),
+        evaluation_id: evaluation_id.into(),
+        source_evaluation_id: request.source_evaluation_id.clone(),
+        execution_order_policy: request.execution_order,
+        effective_execution_order: effective_order(&runs),
+        dimension: request.dimension,
+        model: request.model.clone(),
+        limits: request.limits,
+        evaluator: request
+            .evaluator
+            .as_ref()
+            .map(|evaluator| EvaluatorArtifactV1 {
+                function_id: evaluator.function_id.clone(),
+                arguments_sha256: ids::sha256_json(&evaluator.arguments),
+            }),
+        control: artifact(&request.control),
+        treatment: artifact(&request.treatment),
+        shared_artifacts: SharedArtifactsV1 {
+            model_sha256: ids::sha256_json(
+                &serde_json::to_value(&request.model).unwrap_or_default(),
+            ),
+            function_policy_sha256: ids::sha256_json(
+                &serde_json::to_value(&request.functions).unwrap_or_default(),
+            ),
+            limits_sha256: ids::sha256_json(
+                &serde_json::to_value(request.limits).unwrap_or_default(),
+            ),
+            output_sha256: ids::sha256_json(
+                &serde_json::to_value(&request.output).unwrap_or_default(),
+            ),
+        },
+        control_aggregate,
+        treatment_aggregate,
+        delta,
+        order_sensitivity: order_sensitivity(&runs),
+        eligible,
+        runs,
+        created_at,
+        completed_at,
+    }
+}
+
+pub fn aggregate(runs: &[EvalRunReportV1], role: VariantRoleV1) -> VariantAggregateV1 {
+    let selected: Vec<_> = runs
+        .iter()
+        .filter(|run| run.role == role && run.status.is_terminal())
+        .collect();
+    let benchmarks: Vec<_> = selected
+        .iter()
+        .filter_map(|run| run.benchmark.as_ref())
+        .collect();
+    let run_count = selected.len() as u32;
+    let evaluated_runs = selected.iter().filter(|run| run.passed.is_some()).count() as u32;
+    let passed = selected
+        .iter()
+        .filter(|run| run.passed == Some(true))
+        .count() as u32;
+    VariantAggregateV1 {
+        runs: run_count,
+        evaluated_runs,
+        passed,
+        pass_rate: if evaluated_runs == 0 {
+            0.0
+        } else {
+            passed as f64 / evaluated_runs as f64
+        },
+        benchmarked_runs: benchmarks.len() as u32,
+        median_score: median_f64(
+            selected
+                .iter()
+                .filter_map(|run| run.evaluation.as_ref()?.score),
+        ),
+        median_wall_time_ms: median_u64(benchmarks.iter().map(|value| value.wall_time_ms)),
+        median_input_tokens: median_u64(benchmarks.iter().filter_map(|value| value.input_tokens)),
+        median_output_tokens: median_u64(benchmarks.iter().filter_map(|value| value.output_tokens)),
+        median_total_tokens: median_u64(benchmarks.iter().filter_map(|value| value.total_tokens)),
+        median_reasoning_tokens: median_u64(
+            benchmarks.iter().filter_map(|value| value.reasoning_tokens),
+        ),
+        median_function_calls: median_u64(benchmarks.iter().map(|value| value.function_calls)),
+        median_function_call_errors: median_u64(
+            benchmarks.iter().map(|value| value.function_call_errors),
+        ),
+        median_cost_usd: median_f64(benchmarks.iter().filter_map(|value| value.cost_usd)),
+        median_trace_count: median_u64(benchmarks.iter().filter_map(|value| value.trace_count)),
+        median_span_count: median_u64(benchmarks.iter().filter_map(|value| value.span_count)),
+        median_error_span_count: median_u64(
+            benchmarks.iter().filter_map(|value| value.error_span_count),
+        ),
+        median_trace_duration_ms: median_u64(
+            benchmarks
+                .iter()
+                .filter_map(|value| value.trace_duration_ms),
+        ),
+    }
+}
+
+fn effective_order(runs: &[EvalRunReportV1]) -> Vec<String> {
+    runs.iter()
+        .map(|run| {
+            format!(
+                "{}:{}:{}",
+                run.execution_position,
+                run.role.as_str(),
+                run.iteration
+            )
+        })
+        .collect()
+}
+
+fn order_sensitivity(runs: &[EvalRunReportV1]) -> OrderSensitivityV1 {
+    fn role(runs: &[EvalRunReportV1], role: VariantRoleV1) -> RoleOrderSensitivityV1 {
+        let evaluated: Vec<_> = runs
+            .iter()
+            .filter(|run| run.role == role && run.passed.is_some())
+            .collect();
+        let first: Vec<_> = evaluated
+            .iter()
+            .filter(|run| run.pair_position == 1)
+            .collect();
+        let second: Vec<_> = evaluated
+            .iter()
+            .filter(|run| run.pair_position == 2)
+            .collect();
+        let first_passed = first.iter().filter(|run| run.passed == Some(true)).count() as u32;
+        let second_passed = second.iter().filter(|run| run.passed == Some(true)).count() as u32;
+        let differs = !first.is_empty()
+            && !second.is_empty()
+            && first_passed as u64 * second.len() as u64
+                != second_passed as u64 * first.len() as u64;
+        RoleOrderSensitivityV1 {
+            first_runs: first.len() as u32,
+            first_passed,
+            second_runs: second.len() as u32,
+            second_passed,
+            differs,
+        }
+    }
+    let control = role(runs, VariantRoleV1::Control);
+    let treatment = role(runs, VariantRoleV1::Treatment);
+    OrderSensitivityV1 {
+        detected: control.differs || treatment.differs,
+        control,
+        treatment,
+    }
+}
+
+fn artifact(variant: &crate::contract::EvalVariantV1) -> VariantArtifactV1 {
+    VariantArtifactV1 {
+        label: variant.label.clone(),
+        prompt_sha256: ids::sha256_text(&variant.prompt),
+        system_prompt_sha256: variant.system_prompt.as_deref().map(ids::sha256_text),
+    }
+}
+
+fn delta(control: &VariantAggregateV1, treatment: &VariantAggregateV1) -> AggregateDeltaV1 {
+    AggregateDeltaV1 {
+        pass_rate: treatment.pass_rate - control.pass_rate,
+        median_score: difference_f64(control.median_score, treatment.median_score),
+        median_wall_time_ms: difference_u64(
+            control.median_wall_time_ms,
+            treatment.median_wall_time_ms,
+        ),
+        median_total_tokens: difference_u64(
+            control.median_total_tokens,
+            treatment.median_total_tokens,
+        ),
+        median_reasoning_tokens: difference_u64(
+            control.median_reasoning_tokens,
+            treatment.median_reasoning_tokens,
+        ),
+        median_function_calls: difference_u64(
+            control.median_function_calls,
+            treatment.median_function_calls,
+        ),
+        median_function_call_errors: difference_u64(
+            control.median_function_call_errors,
+            treatment.median_function_call_errors,
+        ),
+        median_cost_usd: difference_f64(control.median_cost_usd, treatment.median_cost_usd),
+        median_trace_count: difference_u64(
+            control.median_trace_count,
+            treatment.median_trace_count,
+        ),
+        median_span_count: difference_u64(control.median_span_count, treatment.median_span_count),
+        median_error_span_count: difference_u64(
+            control.median_error_span_count,
+            treatment.median_error_span_count,
+        ),
+        median_trace_duration_ms: difference_u64(
+            control.median_trace_duration_ms,
+            treatment.median_trace_duration_ms,
+        ),
+    }
+}
+
+fn difference_u64(control: Option<u64>, treatment: Option<u64>) -> Option<f64> {
+    control
+        .zip(treatment)
+        .map(|(control, treatment)| treatment as f64 - control as f64)
+}
+
+fn difference_f64(control: Option<f64>, treatment: Option<f64>) -> Option<f64> {
+    control.zip(treatment).and_then(|(control, treatment)| {
+        let delta = treatment - control;
+        delta.is_finite().then_some(delta)
+    })
+}
+
+fn median_u64(values: impl IntoIterator<Item = u64>) -> Option<u64> {
+    let mut values: Vec<_> = values.into_iter().collect();
+    if values.is_empty() {
+        return None;
+    }
+    values.sort_unstable();
+    let middle = values.len() / 2;
+    if values.len() % 2 == 1 {
+        Some(values[middle])
+    } else {
+        Some(values[middle - 1].saturating_add(values[middle]) / 2)
+    }
+}
+
+fn median_f64(values: impl IntoIterator<Item = f64>) -> Option<f64> {
+    let mut values: Vec<_> = values
+        .into_iter()
+        .filter(|value| value.is_finite())
+        .collect();
+    if values.is_empty() {
+        return None;
+    }
+    values.sort_by(f64::total_cmp);
+    let middle = values.len() / 2;
+    if values.len() % 2 == 1 {
+        Some(values[middle])
+    } else {
+        Some((values[middle - 1] + values[middle]) / 2.0)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::contract::{
+        ComparisonDimensionV1, EvalModelConfigV1, EvalVariantV1, EvaluatorResponseV1,
+        EvaluatorSpecV1,
+    };
+    use harness::prompt::SystemPromptStrategy;
+    use harness::types::output::OutputContract;
+    use harness::types::turn::FunctionPolicy;
+    use serde_json::json;
+
+    fn request(runs: u32) -> NormalizedEvalRequestV1 {
+        NormalizedEvalRequestV1 {
+            dimension: ComparisonDimensionV1::Prompt,
+            model: EvalModelConfigV1 {
+                model: "model".into(),
+                provider: Some("provider".into()),
+                system_prompt_strategy: SystemPromptStrategy::Override,
+                mode: None,
+                thinking_level: None,
+                provider_options: None,
+            },
+            control: EvalVariantV1 {
+                label: None,
+                prompt: "old".into(),
+                system_prompt: Some("system".into()),
+            },
+            treatment: EvalVariantV1 {
+                label: None,
+                prompt: "new".into(),
+                system_prompt: Some("system".into()),
+            },
+            evaluator: Some(EvaluatorSpecV1 {
+                function_id: "judge".into(),
+                arguments: json!({}),
+            }),
+            runs,
+            execution_order: ExecutionOrderV1::default(),
+            source_evaluation_id: None,
+            limits: EvalLimitsV1::default(),
+            functions: FunctionPolicy::default(),
+            output: OutputContract::Text,
+            metadata: None,
+        }
+    }
+
+    fn run(role: VariantRoleV1, iteration: u32, passed: bool, tokens: u64) -> EvalRunReportV1 {
+        EvalRunReportV1 {
+            run_id: format!("{}-{iteration}", role.as_str()),
+            role,
+            iteration,
+            execution_position: 1,
+            pair_position: 1,
+            status: EvalRunStatusV1::Completed,
+            session_id: "session".into(),
+            turn_id: Some("turn".into()),
+            passed: Some(passed),
+            started_at: 1,
+            completed_at: Some(2),
+            output: Some(json!("OK")),
+            metrics: None,
+            evaluation: Some(EvaluatorResponseV1 {
+                passed,
+                score: Some(if passed { 1.0 } else { 0.0 }),
+                reason: None,
+                details: None,
+            }),
+            benchmark: Some(EvalBenchmarkV1 {
+                wall_time_ms: tokens,
+                sessions: 1,
+                turns: 1,
+                function_calls: 0,
+                function_call_errors: 0,
+                input_tokens: Some(tokens),
+                output_tokens: Some(1),
+                total_tokens: Some(tokens),
+                cache_read_tokens: None,
+                cache_write_tokens: None,
+                reasoning_tokens: None,
+                cost_usd: None,
+                trace_count: None,
+                span_count: None,
+                error_span_count: None,
+                trace_duration_ms: None,
+            }),
+            failures: Vec::new(),
+        }
+    }
+
+    #[test]
+    fn medians_and_treatment_minus_control_deltas_are_reported() {
+        let runs = vec![
+            run(VariantRoleV1::Control, 1, true, 10),
+            run(VariantRoleV1::Treatment, 1, true, 8),
+            run(VariantRoleV1::Control, 2, true, 20),
+            run(VariantRoleV1::Treatment, 2, true, 12),
+            run(VariantRoleV1::Control, 3, true, 30),
+            run(VariantRoleV1::Treatment, 3, true, 16),
+        ];
+        let report = build_report("eval-1", &request(3), runs, 1, 2);
+        assert_eq!(report.eligible, Some(true));
+        assert_eq!(report.control_aggregate.median_total_tokens, Some(20));
+        assert_eq!(report.treatment_aggregate.median_total_tokens, Some(12));
+        assert_eq!(report.delta.median_total_tokens, Some(-8.0));
+    }
+
+    #[test]
+    fn treatment_must_pass_every_run() {
+        let runs = vec![
+            run(VariantRoleV1::Control, 1, true, 10),
+            run(VariantRoleV1::Treatment, 1, true, 10),
+            run(VariantRoleV1::Control, 2, false, 10),
+            run(VariantRoleV1::Treatment, 2, false, 10),
+        ];
+        assert_eq!(
+            build_report("eval-1", &request(2), runs, 1, 2).eligible,
+            Some(false)
+        );
+    }
+
+    #[test]
+    fn null_outputs_are_not_included_in_pass_rate() {
+        let mut control = run(VariantRoleV1::Control, 1, false, 10);
+        control.output = Some(Value::Null);
+        control.passed = None;
+        control.evaluation = None;
+        let mut treatment = run(VariantRoleV1::Treatment, 1, false, 10);
+        treatment.output = Some(Value::Null);
+        treatment.passed = None;
+        treatment.evaluation = None;
+
+        let report = build_report("eval-1", &request(1), vec![control, treatment], 1, 2);
+        assert_eq!(report.control_aggregate.evaluated_runs, 0);
+        assert_eq!(report.treatment_aggregate.evaluated_runs, 0);
+        assert_eq!(report.control_aggregate.passed, 0);
+        assert_eq!(report.eligible, Some(false));
+    }
+
+    #[test]
+    fn report_without_evaluator_has_no_eligibility_decision() {
+        let mut request = request(1);
+        request.evaluator = None;
+        let report = build_report(
+            "eval-1",
+            &request,
+            vec![
+                run(VariantRoleV1::Control, 1, false, 10),
+                run(VariantRoleV1::Treatment, 1, false, 10),
+            ],
+            1,
+            2,
+        );
+        assert_eq!(report.evaluator, None);
+        assert_eq!(report.eligible, None);
+    }
+
+    #[test]
+    fn progress_uses_completed_runs_and_detects_order_sensitivity() {
+        let mut control_first = run(VariantRoleV1::Control, 1, true, 10);
+        control_first.pair_position = 1;
+        let mut control_second = run(VariantRoleV1::Control, 2, false, 20);
+        control_second.pair_position = 2;
+        let mut pending = run(VariantRoleV1::Treatment, 2, true, 30);
+        pending.status = EvalRunStatusV1::Running;
+        pending.passed = None;
+
+        let progress = build_progress(&[control_first, control_second, pending]);
+        assert_eq!(progress.control_aggregate.runs, 2);
+        assert_eq!(progress.treatment_aggregate.runs, 0);
+        assert!(progress.order_sensitivity.detected);
+        assert!(progress.order_sensitivity.control.differs);
+    }
+}
diff --git a/eval/src/runtime.rs b/eval/src/runtime.rs
new file mode 100644
index 000000000..cca0cdd4a
--- /dev/null
+++ b/eval/src/runtime.rs
@@ -0,0 +1,1053 @@
+use std::sync::Arc;
+
+use harness::functions::metrics::{SessionMetricsRequestV1, SessionMetricsResponseV1};
+use harness::functions::send::{MessageInput, SendOptions, SendRequest, SendResponse, SessionInit};
+use harness::functions::session_tree::{SessionTreeRequestV1, SessionTreeResponseV1};
+use harness::functions::status::{StatusReport, StatusRequest};
+use harness::functions::stop::{StopRequest, StopResponse};
+use harness::functions::teardown::{TeardownRequestV1, TeardownResponseV1};
+use harness::types::turn::TurnStatus;
+use iii_sdk::protocol::TriggerRequest;
+use iii_sdk::IIIClient;
+use serde::de::DeserializeOwned;
+use serde::Serialize;
+use serde_json::{json, Value};
+
+use crate::contract::{
+    EvalCancelResponseV1, EvalDeleteResponseV1, EvalListRequestV1, EvalListResponseV1,
+    EvalRerunRequestV1, EvalResultResponseV1, EvalRunStatusV1, EvalStartRequestV1,
+    EvalStartResponseV1, EvalStatusV1, EvaluationIdRequestV1, EvaluatorInputV1,
+    EvaluatorResponseV1, NormalizedEvalRequestV1, StepRequestV1, StepResponseV1, SweepResponseV1,
+    VariantRoleV1, WakeEventV1, WakeResponseV1,
+};
+use crate::error::{EvalError, EvalFailureV1, EvalPhaseV1};
+use crate::events::EvalEvents;
+use crate::locks::EvalLocks;
+use crate::report::{build_progress, build_report, EvalBenchmarkV1};
+use crate::state::{EvalJobRecordV1, SessionIndexV1};
+use crate::{ids, queue, state};
+
+#[derive(Clone)]
+pub struct Deps {
+    pub iii: Arc<IIIClient>,
+    pub locks: EvalLocks,
+    pub events: EvalEvents,
+}
+
+pub async fn start(
+    deps: &Deps,
+    request: EvalStartRequestV1,
+) -> Result<EvalStartResponseV1, EvalError> {
+    let request = request.normalize()?;
+    create_job(deps, request).await
+}
+
+async fn create_job(
+    deps: &Deps,
+    request: NormalizedEvalRequestV1,
+) -> Result<EvalStartResponseV1, EvalError> {
+    let evaluation_id = ids::evaluation_id();
+    let now = ids::now_ms();
+    let job = EvalJobRecordV1 {
+        schema_version: "1".into(),
+        evaluation_id: evaluation_id.clone(),
+        runs: state::build_run_plan(&evaluation_id, &request),
+        request,
+        status: EvalStatusV1::Queued,
+        step: 0,
+        next_index: 0,
+        active_index: None,
+        active_waited_for_descendants: false,
+        active_finalization_sent: false,
+        report: None,
+        error: None,
+        created_at: now,
+        updated_at: now,
+        completed_at: None,
+    };
+    state::put_job(&deps.iii, &job).await?;
+    queue::enqueue_step(&deps.iii, &evaluation_id, 0).await?;
+    Ok(EvalStartResponseV1 {
+        evaluation_id,
+        status: EvalStatusV1::Queued,
+    })
+}
+
+pub async fn rerun(
+    deps: &Deps,
+    request: EvalRerunRequestV1,
+) -> Result<EvalStartResponseV1, EvalError> {
+    let source = state::get_job(&deps.iii, &request.evaluation_id)
+        .await?
+        .ok_or_else(|| EvalError::NotFound(request.evaluation_id.clone()))?;
+    if !source.status.is_terminal() {
+        return Err(EvalError::Conflict(format!(
+            "{} is still {:?}; rerun it after completion or cancellation",
+            source.evaluation_id, source.status
+        )));
+    }
+    let mut persisted = source.request.clone();
+    persisted.source_evaluation_id = Some(source.evaluation_id);
+    if request.reverse_order {
+        persisted.execution_order = persisted.execution_order.reversed();
+    }
+    create_job(deps, persisted).await
+}
+
+pub async fn status(
+    deps: &Deps,
+    request: EvaluationIdRequestV1,
+) -> Result<Option<crate::contract::EvalStatusResponseV1>, EvalError> {
+    Ok(state::get_job(&deps.iii, &request.evaluation_id)
+        .await?
+        .map(|job| job.status_response()))
+}
+
+pub async fn list(
+    deps: &Deps,
+    request: EvalListRequestV1,
+) -> Result<EvalListResponseV1, EvalError> {
+    let limit = request.normalized_limit()?;
+    let mut jobs = state::list_jobs(&deps.iii).await?;
+    jobs.sort_by(|left, right| {
+        right
+            .created_at
+            .cmp(&left.created_at)
+            .then_with(|| right.evaluation_id.cmp(&left.evaluation_id))
+    });
+    Ok(EvalListResponseV1 {
+        evaluations: jobs
+            .into_iter()
+            .take(limit)
+            .map(|job| job.summary_response())
+            .collect(),
+    })
+}
+
+pub async fn result(
+    deps: &Deps,
+    request: EvaluationIdRequestV1,
+) -> Result<Option<EvalResultResponseV1>, EvalError> {
+    Ok(state::get_job(&deps.iii, &request.evaluation_id)
+        .await?
+        .map(|job| EvalResultResponseV1 {
+            status: job.status,
+            progress: build_progress(&job.runs),
+            request: job.request,
+            report: job.report,
+        }))
+}
+
+pub async fn cancel(
+    deps: &Deps,
+    request: EvaluationIdRequestV1,
+) -> Result<EvalCancelResponseV1, EvalError> {
+    let _guard = deps.locks.guard(&request.evaluation_id).await;
+    let mut job = state::get_job(&deps.iii, &request.evaluation_id)
+        .await?
+        .ok_or_else(|| EvalError::NotFound(request.evaluation_id.clone()))?;
+    if job.status.is_terminal() {
+        return Ok(EvalCancelResponseV1 {
+            cancelled: false,
+            status: job.status,
+        });
+    }
+
+    if let Some(index) = job.active_index {
+        if let Some(run) = job.runs.get(index) {
+            stop_harness_tree(deps, &run.session_id, &job).await;
+            let _ = harness_teardown(deps, &run.session_id, &job).await;
+            let _ = state::delete_session_index(&deps.iii, &run.session_id).await;
+        }
+    }
+
+    let now = ids::now_ms();
+    for run in &mut job.runs {
+        if !run.status.is_terminal() {
+            run.status = EvalRunStatusV1::Cancelled;
+            run.completed_at = Some(now);
+            run.failures.push(EvalFailureV1::new(
+                EvalPhaseV1::Cancel,
+                "evaluation cancelled",
+            ));
+        }
+    }
+    job.active_index = None;
+    job.status = EvalStatusV1::Cancelled;
+    job.updated_at = now;
+    job.completed_at = Some(now);
+    job.report = Some(build_report(
+        &job.evaluation_id,
+        &job.request,
+        job.runs.clone(),
+        job.created_at,
+        now,
+    ));
+    state::put_job(&deps.iii, &job).await?;
+    deps.events
+        .emit_completed(&job.evaluation_id, job.status, Some(false))
+        .await;
+    Ok(EvalCancelResponseV1 {
+        cancelled: true,
+        status: job.status,
+    })
+}
+
+pub async fn delete(
+    deps: &Deps,
+    request: EvaluationIdRequestV1,
+) -> Result<EvalDeleteResponseV1, EvalError> {
+    let _guard = deps.locks.guard(&request.evaluation_id).await;
+    let Some(job) = state::get_job(&deps.iii, &request.evaluation_id).await? else {
+        return Ok(EvalDeleteResponseV1 { deleted: false });
+    };
+    if !job.status.is_terminal() {
+        return Err(EvalError::Conflict(format!(
+            "{} is still {:?}; cancel or wait before deleting it",
+            job.evaluation_id, job.status
+        )));
+    }
+    for run in &job.runs {
+        state::delete_session_index(&deps.iii, &run.session_id).await?;
+    }
+    state::delete_job(&deps.iii, &job.evaluation_id).await?;
+    Ok(EvalDeleteResponseV1 { deleted: true })
+}
+
+pub fn exact(input: EvaluatorInputV1) -> Result<EvaluatorResponseV1, EvalError> {
+    let expected = input.arguments.get("expected").ok_or_else(|| {
+        EvalError::InvalidRequest(
+            "eval::assert::exact requires evaluator arguments.expected".into(),
+        )
+    })?;
+    let passed = &input.output == expected;
+    Ok(EvaluatorResponseV1 {
+        passed,
+        score: Some(if passed { 1.0 } else { 0.0 }),
+        reason: Some(if passed {
+            "output exactly matched arguments.expected".into()
+        } else {
+            format!(
+                "output did not exactly match arguments.expected (actual={}, expected={})",
+                compact(&input.output),
+                compact(expected)
+            )
+        }),
+        details: None,
+    })
+}
+
+pub fn normalized_text(input: EvaluatorInputV1) -> Result<EvaluatorResponseV1, EvalError> {
+    let expected = input
+        .arguments
+        .get("expected")
+        .and_then(Value::as_str)
+        .ok_or_else(|| {
+            EvalError::InvalidRequest(
+                "eval::assert::normalized_text requires string evaluator arguments.expected".into(),
+            )
+        })?;
+    let Some(actual) = input.output.as_str() else {
+        return Ok(EvaluatorResponseV1 {
+            passed: false,
+            score: Some(0.0),
+            reason: Some(format!(
+                "output was not text (actual={}, expected={})",
+                compact(&input.output),
+                compact(&Value::String(expected.into()))
+            )),
+            details: None,
+        });
+    };
+    let actual_normalized = normalize_text(actual);
+    let expected_normalized = normalize_text(expected);
+    let passed = actual_normalized == expected_normalized;
+    Ok(EvaluatorResponseV1 {
+        passed,
+        score: Some(if passed { 1.0 } else { 0.0 }),
+        reason: Some(if passed {
+            "normalized output matched arguments.expected".into()
+        } else {
+            format!(
+                "normalized output did not match arguments.expected (actual={}, expected={})",
+                compact(&Value::String(actual_normalized)),
+                compact(&Value::String(expected_normalized))
+            )
+        }),
+        details: None,
+    })
+}
+
+fn normalize_text(value: &str) -> String {
+    value
+        .split_whitespace()
+        .collect::<Vec<_>>()
+        .join(" ")
+        .trim_matches(|character: char| character.is_ascii_punctuation())
+        .to_lowercase()
+}
+
+pub async fn wake(deps: &Deps, event: WakeEventV1) -> Result<WakeResponseV1, EvalError> {
+    if !event.terminal || !event.session_id.starts_with("eval_") {
+        return Ok(WakeResponseV1 { woke: false });
+    }
+    let Some(index) = state::get_session_index(&deps.iii, &event.session_id).await? else {
+        return Ok(WakeResponseV1 { woke: false });
+    };
+    let Some(job) = state::get_job(&deps.iii, &index.evaluation_id).await? else {
+        return Ok(WakeResponseV1 { woke: false });
+    };
+    if job.status.is_terminal() {
+        return Ok(WakeResponseV1 { woke: false });
+    }
+    queue::enqueue_step(&deps.iii, &job.evaluation_id, job.step).await?;
+    Ok(WakeResponseV1 { woke: true })
+}
+
+pub async fn sweep(deps: &Deps) -> Result<SweepResponseV1, EvalError> {
+    let jobs = state::list_jobs(&deps.iii).await?;
+    let mut swept = 0;
+    for job in jobs.iter().filter(|job| !job.status.is_terminal()) {
+        match queue::enqueue_step(&deps.iii, &job.evaluation_id, job.step).await {
+            Ok(()) => swept += 1,
+            Err(error) => {
+                tracing::warn!(
+                    evaluation_id = %job.evaluation_id,
+                    %error,
+                    "eval recovery sweep could not enqueue step"
+                );
+            }
+        }
+    }
+    Ok(SweepResponseV1 { swept })
+}
+
+pub async fn step(deps: &Deps, request: StepRequestV1) -> Result<StepResponseV1, EvalError> {
+    let _guard = deps.locks.guard(&request.evaluation_id).await;
+    let Some(mut job) = state::get_job(&deps.iii, &request.evaluation_id).await? else {
+        return Ok(StepResponseV1 {
+            skipped: true,
+            status: EvalStatusV1::Failed,
+        });
+    };
+    if job.status.is_terminal() || request.step != job.step {
+        return Ok(StepResponseV1 {
+            skipped: true,
+            status: job.status,
+        });
+    }
+
+    if job.status == EvalStatusV1::Queued {
+        job.status = EvalStatusV1::Running;
+        job.updated_at = ids::now_ms();
+        state::put_job(&deps.iii, &job).await?;
+    }
+
+    if let Some(index) = job.active_index {
+        if run_timed_out(&job, index) {
+            time_out_active(deps, &mut job, index).await?;
+            return advance(deps, job).await;
+        }
+        if job.runs[index].turn_id.is_none() {
+            launch_active(deps, &mut job, index).await?;
+            if job.runs[index].status.is_terminal() {
+                return advance(deps, job).await;
+            }
+            return Ok(StepResponseV1 {
+                skipped: false,
+                status: job.status,
+            });
+        }
+        match reconcile_active(deps, &mut job, index).await? {
+            ReconcileOutcome::StillRunning => {
+                return Ok(StepResponseV1 {
+                    skipped: false,
+                    status: job.status,
+                })
+            }
+            ReconcileOutcome::Terminal => return advance(deps, job).await,
+        }
+    }
+
+    if job.next_index < job.runs.len() {
+        let index = job.next_index;
+        job.next_index += 1;
+        job.active_index = Some(index);
+        job.runs[index].status = EvalRunStatusV1::Running;
+        job.runs[index].started_at = ids::now_ms();
+        job.updated_at = ids::now_ms();
+        let session_index = SessionIndexV1 {
+            evaluation_id: job.evaluation_id.clone(),
+            run_id: job.runs[index].run_id.clone(),
+        };
+        state::put_session_index(&deps.iii, &job.runs[index].session_id, &session_index).await?;
+        state::put_job(&deps.iii, &job).await?;
+        launch_active(deps, &mut job, index).await?;
+        if job.runs[index].status.is_terminal() {
+            return advance(deps, job).await;
+        }
+        return Ok(StepResponseV1 {
+            skipped: false,
+            status: job.status,
+        });
+    }
+
+    finalize(deps, &mut job).await?;
+    Ok(StepResponseV1 {
+        skipped: false,
+        status: job.status,
+    })
+}
+
+async fn launch_active(
+    deps: &Deps,
+    job: &mut EvalJobRecordV1,
+    index: usize,
+) -> Result<(), EvalError> {
+    let run = job.runs[index].clone();
+    let variant = match run.role {
+        VariantRoleV1::Control => &job.request.control,
+        VariantRoleV1::Treatment => &job.request.treatment,
+    };
+    let request = SendRequest {
+        session_id: Some(run.session_id.clone()),
+        message: MessageInput::Text(variant.prompt.clone()),
+        model: Some(job.request.model.model.clone()),
+        provider: job.request.model.provider.clone(),
+        idempotency_key: Some(ids::send_idempotency_key(
+            &job.evaluation_id,
+            run.role,
+            run.iteration,
+        )),
+        session: Some(SessionInit {
+            title: Some(format!(
+                "Evaluation {}: {} {}",
+                job.evaluation_id,
+                run.role.as_str(),
+                run.iteration
+            )),
+            metadata: Some(json!({
+                "evaluation_id": job.evaluation_id,
+                "eval_run_id": run.run_id,
+                "eval_role": run.role,
+                "eval_iteration": run.iteration,
+            })),
+        }),
+        options: Some(send_options(job, variant)),
+    };
+
+    let response: Result<SendResponse, EvalError> = trigger(
+        deps,
+        "harness::send",
+        request,
+        job.request.limits.execution.invocation_timeout_seconds,
+    )
+    .await;
+    match response {
+        Ok(response) if response.accepted && response.session_id == job.runs[index].session_id => {
+            job.runs[index].turn_id = Some(response.turn_id);
+            job.updated_at = ids::now_ms();
+            state::put_job(&deps.iii, job).await
+        }
+        Ok(response) => {
+            fail_run(
+                &mut job.runs[index],
+                EvalRunStatusV1::Failed,
+                EvalFailureV1::function(
+                    EvalPhaseV1::Send,
+                    "harness::send",
+                    format!(
+                        "unexpected response: accepted={}, session_id={}",
+                        response.accepted, response.session_id
+                    ),
+                ),
+            );
+            state::put_job(&deps.iii, job).await
+        }
+        Err(error) => {
+            // If the response was lost after acceptance, the deterministic
+            // session may already exist. Preserve it for reconciliation.
+            let status = harness_status(deps, &job.runs[index].session_id, job).await?;
+            if let Some(status) = status {
+                job.runs[index].turn_id = status.turn_id;
+                state::put_job(&deps.iii, job).await
+            } else {
+                fail_run(
+                    &mut job.runs[index],
+                    EvalRunStatusV1::Failed,
+                    EvalFailureV1::function(EvalPhaseV1::Send, "harness::send", error.to_string()),
+                );
+                state::put_job(&deps.iii, job).await
+            }
+        }
+    }
+}
+
+enum ReconcileOutcome {
+    StillRunning,
+    Terminal,
+}
+
+async fn reconcile_active(
+    deps: &Deps,
+    job: &mut EvalJobRecordV1,
+    index: usize,
+) -> Result<ReconcileOutcome, EvalError> {
+    let session_id = job.runs[index].session_id.clone();
+    let Some(status) = harness_status(deps, &session_id, job).await? else {
+        return Ok(ReconcileOutcome::StillRunning);
+    };
+    if !status.status.is_terminal() {
+        return Ok(ReconcileOutcome::StillRunning);
+    }
+
+    let now = ids::now_ms();
+    let wall_time_ms = elapsed_ms(job.runs[index].started_at, now);
+    let metrics = harness_metrics(deps, &session_id, job).await?;
+    if !metrics.complete {
+        if !job.active_waited_for_descendants {
+            job.active_waited_for_descendants = true;
+            job.updated_at = now;
+            state::put_job(&deps.iii, job).await?;
+        }
+        return Ok(ReconcileOutcome::StillRunning);
+    }
+
+    if needs_finalization(
+        status.status,
+        status.result_error.is_none(),
+        status.expects_wake,
+        job.active_waited_for_descendants,
+        job.active_finalization_sent,
+    ) {
+        send_finalization(deps, job, index).await?;
+        return Ok(ReconcileOutcome::StillRunning);
+    }
+
+    let output = status.result.clone().unwrap_or(Value::Null);
+    job.runs[index].output = Some(output.clone());
+    job.runs[index].completed_at = Some(now);
+    match status.status {
+        TurnStatus::Completed if status.result_error.is_none() => {
+            let limit_failures = job.request.limits.failures(&metrics);
+            job.runs[index].benchmark = EvalBenchmarkV1::from_metrics(&metrics, wall_time_ms);
+            job.runs[index].metrics = Some(metrics.clone());
+            if output.is_null() || job.request.evaluator.is_none() {
+                job.runs[index].passed = None;
+                job.runs[index].failures.extend(limit_failures);
+                job.runs[index].status = EvalRunStatusV1::Completed;
+            } else {
+                let evaluator = job
+                    .request
+                    .evaluator
+                    .clone()
+                    .expect("evaluator checked above");
+                let evaluator_input = EvaluatorInputV1 {
+                    evaluation_id: job.evaluation_id.clone(),
+                    run_id: job.runs[index].run_id.clone(),
+                    role: job.runs[index].role,
+                    session_id: session_id.clone(),
+                    output,
+                    metrics: metrics.clone(),
+                    arguments: evaluator.arguments,
+                };
+                let evaluation: Result<EvaluatorResponseV1, EvalError> = trigger(
+                    deps,
+                    &evaluator.function_id,
+                    evaluator_input,
+                    job.request.limits.execution.invocation_timeout_seconds,
+                )
+                .await;
+                match evaluation.and_then(EvaluatorResponseV1::validate) {
+                    Ok(evaluation) => {
+                        job.runs[index].passed =
+                            Some(evaluation.passed && limit_failures.is_empty());
+                        job.runs[index].evaluation = Some(evaluation);
+                        job.runs[index].failures.extend(limit_failures);
+                        job.runs[index].status = EvalRunStatusV1::Completed;
+                    }
+                    Err(error) => {
+                        fail_run(
+                            &mut job.runs[index],
+                            EvalRunStatusV1::Failed,
+                            EvalFailureV1::function(
+                                EvalPhaseV1::Evaluate,
+                                evaluator.function_id,
+                                error.to_string(),
+                            ),
+                        );
+                    }
+                }
+            }
+        }
+        TurnStatus::Cancelled => {
+            store_metrics(job, index, metrics, wall_time_ms);
+            fail_run(
+                &mut job.runs[index],
+                EvalRunStatusV1::Cancelled,
+                EvalFailureV1::function(
+                    EvalPhaseV1::Await,
+                    "harness::status",
+                    status
+                        .result_error
+                        .or_else(|| Some("harness run was cancelled".into()))
+                        .unwrap_or_default(),
+                ),
+            );
+        }
+        _ => {
+            store_metrics(job, index, metrics, wall_time_ms);
+            fail_run(
+                &mut job.runs[index],
+                EvalRunStatusV1::Failed,
+                EvalFailureV1::function(
+                    EvalPhaseV1::Await,
+                    "harness::status",
+                    status
+                        .result_error
+                        .unwrap_or_else(|| format!("harness run ended as {:?}", status.status)),
+                ),
+            );
+        }
+    }
+    job.updated_at = now;
+    state::put_job(&deps.iii, job).await?;
+    Ok(ReconcileOutcome::Terminal)
+}
+
+fn needs_finalization(
+    status: TurnStatus,
+    result_ok: bool,
+    expects_wake: bool,
+    waited_for_descendants: bool,
+    finalization_sent: bool,
+) -> bool {
+    status == TurnStatus::Completed
+        && result_ok
+        && (expects_wake || waited_for_descendants)
+        && !finalization_sent
+}
+
+async fn send_finalization(
+    deps: &Deps,
+    job: &mut EvalJobRecordV1,
+    index: usize,
+) -> Result<(), EvalError> {
+    let run = job.runs[index].clone();
+    let variant = match run.role {
+        VariantRoleV1::Control => &job.request.control,
+        VariantRoleV1::Treatment => &job.request.treatment,
+    };
+    let response: SendResponse = trigger(
+        deps,
+        "harness::send",
+        SendRequest {
+            session_id: Some(run.session_id.clone()),
+            message: MessageInput::Text(
+                "[eval-finalize] All descendant sessions are now terminal. Inspect the durable \
+                 outcomes, re-check every completion condition, finish any required final report, \
+                 remove every trigger or subscription created by this run, and return the \
+                 definitive result. Do not return another progress update or start unrelated work."
+                    .into(),
+            ),
+            model: Some(job.request.model.model.clone()),
+            provider: job.request.model.provider.clone(),
+            idempotency_key: Some(ids::finalization_idempotency_key(
+                &job.evaluation_id,
+                run.role,
+                run.iteration,
+            )),
+            session: None,
+            options: Some(send_options(job, variant)),
+        },
+        job.request.limits.execution.invocation_timeout_seconds,
+    )
+    .await?;
+    if !response.accepted || response.session_id != run.session_id {
+        return Err(EvalError::Dependency(format!(
+            "harness::send finalization returned accepted={}, session_id={}",
+            response.accepted, response.session_id
+        )));
+    }
+    job.runs[index].turn_id = Some(response.turn_id);
+    job.active_finalization_sent = true;
+    job.updated_at = ids::now_ms();
+    state::put_job(&deps.iii, job).await
+}
+
+fn send_options(job: &EvalJobRecordV1, variant: &crate::contract::EvalVariantV1) -> SendOptions {
+    SendOptions {
+        system_prompt: variant.system_prompt.clone(),
+        system_prompt_strategy: if variant.system_prompt.is_none() {
+            harness::prompt::SystemPromptStrategy::Disabled
+        } else {
+            job.request.model.system_prompt_strategy
+        },
+        mode: job.request.model.mode,
+        max_turns: Some(job.request.limits.execution.max_turns),
+        max_output_tokens: Some(job.request.limits.execution.max_output_tokens_per_call),
+        max_total_tokens: job.request.limits.execution.max_total_tokens,
+        max_cost_usd: job.request.limits.execution.max_cost_usd,
+        thinking_level: job.request.model.thinking_level,
+        provider_options: job.request.model.provider_options.clone(),
+        output: Some(job.request.output.clone()),
+        functions: Some(job.request.functions.clone()),
+        metadata: job.request.metadata.clone(),
+    }
+}
+
+fn store_metrics(
+    job: &mut EvalJobRecordV1,
+    index: usize,
+    metrics: SessionMetricsResponseV1,
+    wall_time_ms: u64,
+) {
+    job.runs[index].benchmark = EvalBenchmarkV1::from_metrics(&metrics, wall_time_ms);
+    job.runs[index].metrics = Some(metrics);
+}
+
+async fn time_out_active(
+    deps: &Deps,
+    job: &mut EvalJobRecordV1,
+    index: usize,
+) -> Result<(), EvalError> {
+    let run = job.runs[index].clone();
+    stop_harness_tree(deps, &run.session_id, job).await;
+    fail_run(
+        &mut job.runs[index],
+        EvalRunStatusV1::Failed,
+        EvalFailureV1::new(
+            EvalPhaseV1::Await,
+            format!(
+                "run exceeded {}s wall-clock limit",
+                job.request.limits.execution.scenario_timeout_seconds
+            ),
+        ),
+    );
+    state::put_job(&deps.iii, job).await
+}
+
+async fn advance(deps: &Deps, mut job: EvalJobRecordV1) -> Result<StepResponseV1, EvalError> {
+    if let Some(index) = job.active_index.take() {
+        let session_id = job.runs[index].session_id.clone();
+        if let Err(error) = harness_teardown(deps, &session_id, &job).await {
+            job.runs[index].failures.push(EvalFailureV1::function(
+                EvalPhaseV1::Collect,
+                "harness::teardown",
+                error.to_string(),
+            ));
+        }
+        state::delete_session_index(&deps.iii, &job.runs[index].session_id).await?;
+    }
+    job.active_waited_for_descendants = false;
+    job.active_finalization_sent = false;
+    job.step = job.step.saturating_add(1);
+    job.updated_at = ids::now_ms();
+    if job.next_index >= job.runs.len() {
+        finalize(deps, &mut job).await?;
+    } else {
+        state::put_job(&deps.iii, &job).await?;
+        queue::enqueue_step(&deps.iii, &job.evaluation_id, job.step).await?;
+    }
+    Ok(StepResponseV1 {
+        skipped: false,
+        status: job.status,
+    })
+}
+
+async fn finalize(deps: &Deps, job: &mut EvalJobRecordV1) -> Result<(), EvalError> {
+    let now = ids::now_ms();
+    job.status = EvalStatusV1::Completed;
+    job.updated_at = now;
+    job.completed_at = Some(now);
+    job.report = Some(build_report(
+        &job.evaluation_id,
+        &job.request,
+        job.runs.clone(),
+        job.created_at,
+        now,
+    ));
+    state::put_job(&deps.iii, job).await?;
+    let eligible = job.report.as_ref().and_then(|report| report.eligible);
+    deps.events
+        .emit_completed(&job.evaluation_id, job.status, eligible)
+        .await;
+    Ok(())
+}
+
+fn fail_run(
+    run: &mut crate::report::EvalRunReportV1,
+    status: EvalRunStatusV1,
+    failure: EvalFailureV1,
+) {
+    run.status = status;
+    run.passed = None;
+    run.completed_at = Some(ids::now_ms());
+    run.failures.push(failure);
+}
+
+fn run_timed_out(job: &EvalJobRecordV1, index: usize) -> bool {
+    let started_at = job.runs[index].started_at;
+    started_at > 0
+        && ids::now_ms().saturating_sub(started_at)
+            > job.request.limits.execution.scenario_timeout_seconds as i64 * 1_000
+}
+
+async fn harness_status(
+    deps: &Deps,
+    session_id: &str,
+    job: &EvalJobRecordV1,
+) -> Result<Option<StatusReport>, EvalError> {
+    trigger(
+        deps,
+        "harness::status",
+        StatusRequest {
+            session_id: session_id.into(),
+        },
+        job.request.limits.execution.invocation_timeout_seconds,
+    )
+    .await
+}
+
+async fn harness_metrics(
+    deps: &Deps,
+    session_id: &str,
+    job: &EvalJobRecordV1,
+) -> Result<SessionMetricsResponseV1, EvalError> {
+    trigger(
+        deps,
+        "harness::metrics",
+        SessionMetricsRequestV1 {
+            root_session_id: session_id.into(),
+        },
+        job.request.limits.execution.invocation_timeout_seconds,
+    )
+    .await
+}
+
+async fn harness_teardown(
+    deps: &Deps,
+    session_id: &str,
+    job: &EvalJobRecordV1,
+) -> Result<TeardownResponseV1, EvalError> {
+    trigger(
+        deps,
+        "harness::teardown",
+        TeardownRequestV1 {
+            root_session_id: session_id.into(),
+        },
+        job.request.limits.execution.invocation_timeout_seconds,
+    )
+    .await
+}
+
+async fn stop_harness_tree(deps: &Deps, root_session_id: &str, job: &EvalJobRecordV1) {
+    let timeout_seconds = job.request.limits.execution.invocation_timeout_seconds;
+    let tree: Result<SessionTreeResponseV1, EvalError> = trigger(
+        deps,
+        "harness::session-tree",
+        SessionTreeRequestV1 {
+            root_session_id: root_session_id.into(),
+        },
+        timeout_seconds,
+    )
+    .await;
+    let mut session_ids = tree
+        .map(|tree| {
+            tree.sessions
+                .into_iter()
+                .rev()
+                .map(|node| node.session_id)
+                .collect::<Vec<_>>()
+        })
+        .unwrap_or_else(|_| vec![root_session_id.into()]);
+    if !session_ids
+        .iter()
+        .any(|session_id| session_id == root_session_id)
+    {
+        session_ids.push(root_session_id.into());
+    }
+    for session_id in session_ids {
+        let _: Result<StopResponse, _> = trigger(
+            deps,
+            "harness::stop",
+            StopRequest {
+                session_id,
+                turn_id: None,
+            },
+            timeout_seconds,
+        )
+        .await;
+    }
+}
+
+async fn trigger<I, O>(
+    deps: &Deps,
+    function_id: &str,
+    input: I,
+    timeout_seconds: u64,
+) -> Result<O, EvalError>
+where
+    I: Serialize,
+    O: DeserializeOwned,
+{
+    let timeout = std::time::Duration::from_secs(timeout_seconds);
+    let request = TriggerRequest {
+        function_id: function_id.into(),
+        payload: serde_json::to_value(input)?,
+        action: None,
+        timeout_ms: Some(timeout.as_millis().min(u64::MAX as u128) as u64),
+    };
+    let value = match tokio::time::timeout(timeout, deps.iii.trigger(request)).await {
+        Ok(Ok(value)) => value,
+        Ok(Err(error)) => {
+            return Err(EvalError::Dependency(format!(
+                "{function_id} failed: {error}"
+            )))
+        }
+        Err(_) => {
+            return Err(EvalError::Dependency(format!(
+                "{function_id} exceeded {timeout_seconds}s invocation timeout"
+            )))
+        }
+    };
+    serde_json::from_value(value)
+        .map_err(|error| EvalError::Serialization(format!("{function_id} response: {error}")))
+}
+
+fn elapsed_ms(started_at: i64, completed_at: i64) -> u64 {
+    completed_at.saturating_sub(started_at).max(0) as u64
+}
+
+fn compact(value: &Value) -> String {
+    let rendered = serde_json::to_string(value).unwrap_or_else(|_| "<unserializable>".into());
+    const LIMIT: usize = 512;
+    if rendered.chars().count() <= LIMIT {
+        rendered
+    } else {
+        format!("{}…", rendered.chars().take(LIMIT).collect::<String>())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use harness::functions::metrics::{SessionMetricsResponseV1, SessionUsageTotalsV1};
+
+    use super::*;
+
+    fn evaluator_input(output: Value, expected: Value) -> EvaluatorInputV1 {
+        EvaluatorInputV1 {
+            evaluation_id: "eval".into(),
+            run_id: "run".into(),
+            role: VariantRoleV1::Control,
+            session_id: "session".into(),
+            output,
+            metrics: SessionMetricsResponseV1 {
+                root_session_id: "session".into(),
+                complete: true,
+                totals: SessionUsageTotalsV1::default(),
+                by_session: Vec::new(),
+                traces: None,
+            },
+            arguments: json!({ "expected": expected }),
+        }
+    }
+
+    #[test]
+    fn exact_evaluator_compares_json_values() {
+        assert!(
+            exact(evaluator_input(json!({"ok": true}), json!({"ok": true})))
+                .unwrap()
+                .passed
+        );
+        assert!(
+            !exact(evaluator_input(json!("OK"), json!("NO")))
+                .unwrap()
+                .passed
+        );
+    }
+
+    #[test]
+    fn exact_evaluator_requires_expected_argument() {
+        let mut input = evaluator_input(json!("OK"), json!("OK"));
+        input.arguments = json!({});
+        assert!(exact(input).is_err());
+    }
+
+    #[test]
+    fn normalized_text_ignores_case_whitespace_and_surrounding_punctuation() {
+        let result = normalized_text(evaluator_input(
+            json!("  Olá,   mundo!  "),
+            json!("olá, mundo"),
+        ))
+        .unwrap();
+        assert!(result.passed);
+    }
+
+    #[test]
+    fn normalized_text_rejects_different_text_and_non_text_output() {
+        assert!(
+            !normalized_text(evaluator_input(json!("hello"), json!("goodbye")))
+                .unwrap()
+                .passed
+        );
+        assert!(
+            !normalized_text(evaluator_input(json!({"text": "hello"}), json!("hello")))
+                .unwrap()
+                .passed
+        );
+    }
+
+    #[test]
+    fn normalized_text_requires_string_expected_argument() {
+        assert!(
+            normalized_text(evaluator_input(json!("hello"), json!({"text": "hello"}))).is_err()
+        );
+    }
+
+    #[test]
+    fn finalization_runs_once_after_an_early_root_result() {
+        assert!(needs_finalization(
+            TurnStatus::Completed,
+            true,
+            false,
+            true,
+            false
+        ));
+        assert!(!needs_finalization(
+            TurnStatus::Completed,
+            true,
+            false,
+            false,
+            false
+        ));
+        assert!(needs_finalization(
+            TurnStatus::Completed,
+            true,
+            true,
+            false,
+            false
+        ));
+        assert!(!needs_finalization(
+            TurnStatus::Completed,
+            true,
+            true,
+            true,
+            true
+        ));
+        assert!(!needs_finalization(
+            TurnStatus::Failed,
+            true,
+            true,
+            true,
+            false
+        ));
+        assert!(!needs_finalization(
+            TurnStatus::Completed,
+            false,
+            true,
+            true,
+            false
+        ));
+    }
+}
diff --git a/eval/src/state.rs b/eval/src/state.rs
new file mode 100644
index 000000000..87da8eeea
--- /dev/null
+++ b/eval/src/state.rs
@@ -0,0 +1,443 @@
+use iii_sdk::protocol::TriggerRequest;
+use iii_sdk::IIIClient;
+use serde::{Deserialize, Serialize};
+use serde_json::{json, Value};
+
+use crate::contract::{
+    ActiveRunV1, EvalRunStatusV1, EvalStatusResponseV1, EvalStatusV1, EvalSummaryV1,
+    NormalizedEvalRequestV1,
+};
+use crate::error::EvalError;
+use crate::report::{EvalReportV1, EvalRunReportV1};
+
+pub const JOB_SCOPE: &str = "eval_job";
+pub const SESSION_SCOPE: &str = "eval_session";
+const DISPATCH_TIMEOUT_MS: u64 = 30_000;
+
+fn is_false(value: &bool) -> bool {
+    !*value
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct EvalJobRecordV1 {
+    pub schema_version: String,
+    pub evaluation_id: String,
+    pub request: NormalizedEvalRequestV1,
+    pub status: EvalStatusV1,
+    pub step: u64,
+    pub next_index: usize,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub active_index: Option<usize>,
+    /// The active root returned while descendants were still running. Once
+    /// they settle, the eval asks that same root for one definitive result.
+    #[serde(default, skip_serializing_if = "is_false")]
+    pub active_waited_for_descendants: bool,
+    #[serde(default, skip_serializing_if = "is_false")]
+    pub active_finalization_sent: bool,
+    pub runs: Vec<EvalRunReportV1>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub report: Option<EvalReportV1>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub error: Option<String>,
+    pub created_at: i64,
+    pub updated_at: i64,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub completed_at: Option<i64>,
+}
+
+impl EvalJobRecordV1 {
+    pub fn status_response(&self) -> EvalStatusResponseV1 {
+        let terminal_runs = self
+            .runs
+            .iter()
+            .filter(|run| run.status.is_terminal())
+            .count() as u32;
+        let passed_runs = self
+            .runs
+            .iter()
+            .filter(|run| run.passed == Some(true))
+            .count() as u32;
+        let failed_runs = self
+            .runs
+            .iter()
+            .filter(|run| run.status.is_terminal() && run.passed == Some(false))
+            .count() as u32;
+        let active = self.active_index.and_then(|index| {
+            self.runs.get(index).map(|run| ActiveRunV1 {
+                run_id: run.run_id.clone(),
+                role: run.role,
+                iteration: run.iteration,
+                session_id: run.session_id.clone(),
+                turn_id: run.turn_id.clone(),
+                started_at: run.started_at,
+            })
+        });
+        EvalStatusResponseV1 {
+            evaluation_id: self.evaluation_id.clone(),
+            status: self.status,
+            total_runs: self.runs.len() as u32,
+            terminal_runs,
+            passed_runs,
+            failed_runs,
+            active,
+            created_at: self.created_at,
+            updated_at: self.updated_at,
+            completed_at: self.completed_at,
+            error: self.error.clone(),
+        }
+    }
+
+    pub fn summary_response(&self) -> EvalSummaryV1 {
+        let progress = self.status_response();
+        EvalSummaryV1 {
+            evaluation_id: self.evaluation_id.clone(),
+            status: self.status,
+            dimension: self.request.dimension,
+            model: self.request.model.model.clone(),
+            provider: self.request.model.provider.clone(),
+            control_label: self.request.control.label.clone(),
+            treatment_label: self.request.treatment.label.clone(),
+            total_runs: progress.total_runs,
+            terminal_runs: progress.terminal_runs,
+            passed_runs: progress.passed_runs,
+            failed_runs: progress.failed_runs,
+            eligible: self.report.as_ref().and_then(|report| report.eligible),
+            created_at: self.created_at,
+            updated_at: self.updated_at,
+            completed_at: self.completed_at,
+            error: self.error.clone(),
+        }
+    }
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct SessionIndexV1 {
+    pub evaluation_id: String,
+    pub run_id: String,
+}
+
+pub fn build_run_plan(
+    evaluation_id: &str,
+    request: &NormalizedEvalRequestV1,
+) -> Vec<EvalRunReportV1> {
+    let mut runs = Vec::with_capacity(request.runs as usize * 2);
+    for iteration in 1..=request.runs {
+        for (pair_index, role) in request
+            .execution_order
+            .roles(iteration)
+            .into_iter()
+            .enumerate()
+        {
+            runs.push(EvalRunReportV1 {
+                run_id: crate::ids::run_id(role, iteration),
+                role,
+                iteration,
+                execution_position: runs.len() as u32 + 1,
+                pair_position: pair_index as u8 + 1,
+                status: EvalRunStatusV1::Pending,
+                session_id: crate::ids::session_id(evaluation_id, role, iteration),
+                turn_id: None,
+                passed: None,
+                started_at: 0,
+                completed_at: None,
+                output: None,
+                metrics: None,
+                evaluation: None,
+                benchmark: None,
+                failures: Vec::new(),
+            });
+        }
+    }
+    runs
+}
+
+pub async fn get_job(
+    iii: &IIIClient,
+    evaluation_id: &str,
+) -> Result<Option<EvalJobRecordV1>, EvalError> {
+    let value = state_get(iii, JOB_SCOPE, evaluation_id).await?;
+    if value.is_null() {
+        return Ok(None);
+    }
+    serde_json::from_value(value)
+        .map(Some)
+        .map_err(|error| EvalError::State(format!("parse {JOB_SCOPE}/{evaluation_id}: {error}")))
+}
+
+pub async fn put_job(iii: &IIIClient, job: &EvalJobRecordV1) -> Result<(), EvalError> {
+    state_set(
+        iii,
+        JOB_SCOPE,
+        &job.evaluation_id,
+        serde_json::to_value(job)?,
+    )
+    .await
+}
+
+pub async fn list_jobs(iii: &IIIClient) -> Result<Vec<EvalJobRecordV1>, EvalError> {
+    let value = state_list(iii, JOB_SCOPE).await?;
+    Ok(parse_list(&value))
+}
+
+pub async fn delete_job(iii: &IIIClient, evaluation_id: &str) -> Result<(), EvalError> {
+    state_delete(iii, JOB_SCOPE, evaluation_id).await
+}
+
+pub async fn put_session_index(
+    iii: &IIIClient,
+    session_id: &str,
+    index: &SessionIndexV1,
+) -> Result<(), EvalError> {
+    state_set(iii, SESSION_SCOPE, session_id, serde_json::to_value(index)?).await
+}
+
+pub async fn get_session_index(
+    iii: &IIIClient,
+    session_id: &str,
+) -> Result<Option<SessionIndexV1>, EvalError> {
+    let value = state_get(iii, SESSION_SCOPE, session_id).await?;
+    if value.is_null() {
+        return Ok(None);
+    }
+    serde_json::from_value(value)
+        .map(Some)
+        .map_err(|error| EvalError::State(format!("parse {SESSION_SCOPE}/{session_id}: {error}")))
+}
+
+pub async fn delete_session_index(iii: &IIIClient, session_id: &str) -> Result<(), EvalError> {
+    state_delete(iii, SESSION_SCOPE, session_id).await
+}
+
+async fn state_get(iii: &IIIClient, scope: &str, key: &str) -> Result<Value, EvalError> {
+    iii.trigger(TriggerRequest {
+        function_id: "state::get".into(),
+        payload: json!({ "scope": scope, "key": key }),
+        action: None,
+        timeout_ms: Some(DISPATCH_TIMEOUT_MS),
+    })
+    .await
+    .map_err(|error| EvalError::State(format!("state::get {scope}/{key}: {error}")))
+}
+
+async fn state_set(iii: &IIIClient, scope: &str, key: &str, value: Value) -> Result<(), EvalError> {
+    iii.trigger(TriggerRequest {
+        function_id: "state::set".into(),
+        payload: json!({ "scope": scope, "key": key, "value": value }),
+        action: None,
+        timeout_ms: Some(DISPATCH_TIMEOUT_MS),
+    })
+    .await
+    .map(|_| ())
+    .map_err(|error| EvalError::State(format!("state::set {scope}/{key}: {error}")))
+}
+
+async fn state_list(iii: &IIIClient, scope: &str) -> Result<Value, EvalError> {
+    iii.trigger(TriggerRequest {
+        function_id: "state::list".into(),
+        payload: json!({ "scope": scope }),
+        action: None,
+        timeout_ms: Some(DISPATCH_TIMEOUT_MS),
+    })
+    .await
+    .map_err(|error| EvalError::State(format!("state::list {scope}: {error}")))
+}
+
+async fn state_delete(iii: &IIIClient, scope: &str, key: &str) -> Result<(), EvalError> {
+    iii.trigger(TriggerRequest {
+        function_id: "state::delete".into(),
+        payload: json!({ "scope": scope, "key": key }),
+        action: None,
+        timeout_ms: Some(DISPATCH_TIMEOUT_MS),
+    })
+    .await
+    .map(|_| ())
+    .map_err(|error| EvalError::State(format!("state::delete {scope}/{key}: {error}")))
+}
+
+fn parse_list<T: serde::de::DeserializeOwned>(value: &Value) -> Vec<T> {
+    let candidates: Vec<&Value> = match value {
+        Value::Array(items) => items.iter().collect(),
+        Value::Object(map) => {
+            if let Some(Value::Array(items)) = map.get("values").or_else(|| map.get("items")) {
+                items.iter().collect()
+            } else {
+                map.values().collect()
+            }
+        }
+        _ => return Vec::new(),
+    };
+    candidates
+        .into_iter()
+        .filter_map(|value| serde_json::from_value(value.clone()).ok())
+        .collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use harness::prompt::SystemPromptStrategy;
+    use harness::types::output::OutputContract;
+    use harness::types::turn::FunctionPolicy;
+    use serde_json::json;
+
+    use super::*;
+    use crate::contract::{
+        ComparisonDimensionV1, EvalModelConfigV1, EvalVariantV1, EvaluatorSpecV1, VariantRoleV1,
+    };
+    use crate::limits::EvalLimitsV1;
+
+    fn request(runs: u32) -> NormalizedEvalRequestV1 {
+        NormalizedEvalRequestV1 {
+            dimension: ComparisonDimensionV1::Prompt,
+            model: EvalModelConfigV1 {
+                model: "model".into(),
+                provider: None,
+                system_prompt_strategy: SystemPromptStrategy::Override,
+                mode: None,
+                thinking_level: None,
+                provider_options: None,
+            },
+            control: EvalVariantV1 {
+                label: None,
+                prompt: "a".into(),
+                system_prompt: Some("system".into()),
+            },
+            treatment: EvalVariantV1 {
+                label: None,
+                prompt: "b".into(),
+                system_prompt: Some("system".into()),
+            },
+            evaluator: Some(EvaluatorSpecV1 {
+                function_id: "judge".into(),
+                arguments: json!({}),
+            }),
+            runs,
+            execution_order: crate::contract::ExecutionOrderV1::default(),
+            source_evaluation_id: None,
+            limits: EvalLimitsV1::default(),
+            functions: FunctionPolicy::default(),
+            output: OutputContract::Text,
+            metadata: None,
+        }
+    }
+
+    #[test]
+    fn run_plan_alternates_pair_order() {
+        let runs = build_run_plan("eval_1", &request(3));
+        let roles: Vec<_> = runs.iter().map(|run| run.role).collect();
+        assert_eq!(
+            roles,
+            [
+                VariantRoleV1::Control,
+                VariantRoleV1::Treatment,
+                VariantRoleV1::Treatment,
+                VariantRoleV1::Control,
+                VariantRoleV1::Control,
+                VariantRoleV1::Treatment,
+            ]
+        );
+        assert_eq!(runs[0].session_id, "eval_1_c_1");
+        assert_eq!(runs[2].session_id, "eval_1_t_2");
+        assert_eq!(runs[0].pair_position, 1);
+        assert_eq!(runs[1].pair_position, 2);
+        assert_eq!(runs[5].execution_position, 6);
+    }
+
+    #[test]
+    fn reversed_run_plan_preserves_roles_and_inverts_each_pair() {
+        let mut request = request(2);
+        request.execution_order = crate::contract::ExecutionOrderV1::BalancedTreatmentFirst;
+        let runs = build_run_plan("eval_2", &request);
+        let roles: Vec<_> = runs.iter().map(|run| run.role).collect();
+        assert_eq!(
+            roles,
+            [
+                VariantRoleV1::Treatment,
+                VariantRoleV1::Control,
+                VariantRoleV1::Control,
+                VariantRoleV1::Treatment,
+            ]
+        );
+        assert_eq!(runs[0].run_id, "treatment-1");
+        assert_eq!(runs[1].run_id, "control-1");
+    }
+
+    #[test]
+    fn list_parser_accepts_all_state_shapes() {
+        let index = json!({"evaluation_id": "e", "run_id": "r"});
+        assert_eq!(parse_list::<SessionIndexV1>(&json!([index])).len(), 1);
+        assert_eq!(
+            parse_list::<SessionIndexV1>(&json!({"values": [index]})).len(),
+            1
+        );
+        assert_eq!(
+            parse_list::<SessionIndexV1>(&json!({"key": index})).len(),
+            1
+        );
+    }
+
+    #[test]
+    fn summary_projects_progress_without_the_full_job() {
+        let mut request = request(1);
+        request.control.label = Some("baseline".into());
+        request.treatment.label = Some("candidate".into());
+        let mut runs = build_run_plan("eval_1", &request);
+        runs[0].status = EvalRunStatusV1::Completed;
+        runs[0].passed = Some(true);
+        let job = EvalJobRecordV1 {
+            schema_version: "1".into(),
+            evaluation_id: "eval_1".into(),
+            request,
+            status: EvalStatusV1::Running,
+            step: 1,
+            next_index: 1,
+            active_index: Some(1),
+            active_waited_for_descendants: false,
+            active_finalization_sent: false,
+            runs,
+            report: None,
+            error: None,
+            created_at: 10,
+            updated_at: 20,
+            completed_at: None,
+        };
+
+        let summary = job.summary_response();
+        assert_eq!(summary.control_label.as_deref(), Some("baseline"));
+        assert_eq!(summary.treatment_label.as_deref(), Some("candidate"));
+        assert_eq!(summary.total_runs, 2);
+        assert_eq!(summary.terminal_runs, 1);
+        assert_eq!(summary.passed_runs, 1);
+        assert_eq!(summary.failed_runs, 0);
+        assert_eq!(summary.eligible, None);
+    }
+
+    #[test]
+    fn completed_run_without_evaluation_is_neither_passed_nor_failed() {
+        let request = request(1);
+        let mut runs = build_run_plan("eval_1", &request);
+        runs[0].status = EvalRunStatusV1::Completed;
+        runs[0].output = Some(Value::Null);
+        let job = EvalJobRecordV1 {
+            schema_version: "1".into(),
+            evaluation_id: "eval_1".into(),
+            request,
+            status: EvalStatusV1::Running,
+            step: 1,
+            next_index: 1,
+            active_index: Some(1),
+            active_waited_for_descendants: false,
+            active_finalization_sent: false,
+            runs,
+            report: None,
+            error: None,
+            created_at: 10,
+            updated_at: 20,
+            completed_at: None,
+        };
+
+        let status = job.status_response();
+        assert_eq!(status.terminal_runs, 1);
+        assert_eq!(status.passed_runs, 0);
+        assert_eq!(status.failed_runs, 0);
+    }
+}
diff --git a/eval/src/surface.rs b/eval/src/surface.rs
new file mode 100644
index 000000000..ff266cdb1
--- /dev/null
+++ b/eval/src/surface.rs
@@ -0,0 +1,49 @@
+use schemars::JsonSchema;
+
+use crate::contract::{
+    EvalCancelResponseV1, EvalDeleteResponseV1, EvalListRequestV1, EvalListResponseV1,
+    EvalRerunRequestV1, EvalResultResponseV1, EvalStartRequestV1, EvalStartResponseV1,
+    EvalStatusResponseV1, EvaluationIdRequestV1, EvaluatorInputV1, EvaluatorResponseV1,
+    StepRequestV1, StepResponseV1, SweepEventV1, SweepResponseV1, WakeEventV1, WakeResponseV1,
+};
+use crate::functions::{
+    CANCEL_ID, DELETE_ID, EXACT_ID, LIST_ID, NORMALIZED_TEXT_ID, RERUN_ID, RESULT_ID, START_ID,
+    STATUS_ID, STEP_ID, SWEEP_ID, WAKE_ID,
+};
+
+pub struct FunctionSpec {
+    pub function_id: &'static str,
+    pub request_schema: schemars::schema::RootSchema,
+    pub response_schema: schemars::schema::RootSchema,
+}
+
+fn schema_of<T: JsonSchema>() -> schemars::schema::RootSchema {
+    schemars::r#gen::SchemaSettings::draft07()
+        .into_generator()
+        .into_root_schema_for::<T>()
+}
+
+fn spec<Req: JsonSchema, Resp: JsonSchema>(function_id: &'static str) -> FunctionSpec {
+    FunctionSpec {
+        function_id,
+        request_schema: schema_of::<Req>(),
+        response_schema: schema_of::<Resp>(),
+    }
+}
+
+pub fn catalog() -> Vec<FunctionSpec> {
+    vec![
+        spec::<EvalStartRequestV1, EvalStartResponseV1>(START_ID),
+        spec::<EvalRerunRequestV1, EvalStartResponseV1>(RERUN_ID),
+        spec::<EvalListRequestV1, EvalListResponseV1>(LIST_ID),
+        spec::<EvaluationIdRequestV1, Option<EvalStatusResponseV1>>(STATUS_ID),
+        spec::<EvaluationIdRequestV1, Option<EvalResultResponseV1>>(RESULT_ID),
+        spec::<EvaluationIdRequestV1, EvalCancelResponseV1>(CANCEL_ID),
+        spec::<EvaluationIdRequestV1, EvalDeleteResponseV1>(DELETE_ID),
+        spec::<EvaluatorInputV1, EvaluatorResponseV1>(EXACT_ID),
+        spec::<EvaluatorInputV1, EvaluatorResponseV1>(NORMALIZED_TEXT_ID),
+        spec::<StepRequestV1, StepResponseV1>(STEP_ID),
+        spec::<WakeEventV1, WakeResponseV1>(WAKE_ID),
+        spec::<SweepEventV1, SweepResponseV1>(SWEEP_ID),
+    ]
+}
diff --git a/eval/src/ui.rs b/eval/src/ui.rs
new file mode 100644
index 000000000..a4285cc1e
--- /dev/null
+++ b/eval/src/ui.rs
@@ -0,0 +1,46 @@
+//! Injectable console page for creating and inspecting prompt evaluations.
+
+use std::sync::Arc;
+
+use iii_console_ui::ConsoleUi;
+use iii_sdk::IIIClient;
+
+pub const PAGE_PATH: &str = "eval/page.js";
+pub const STYLES_PATH: &str = "eval/styles.css";
+
+const PAGE_JS: &str = include_str!(concat!(env!("CARGO_MANIFEST_DIR"), "/ui/dist/page.js"));
+const STYLES_CSS: &str = include_str!(concat!(env!("CARGO_MANIFEST_DIR"), "/ui/dist/styles.css"));
+
+fn console_ui() -> ConsoleUi {
+    ConsoleUi::new("eval")
+        .script(PAGE_PATH, PAGE_JS)
+        .style(STYLES_PATH, STYLES_CSS)
+}
+
+pub fn register(iii: &Arc<IIIClient>) {
+    console_ui().register(iii);
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn ui_builder_accepts_the_assets() {
+        let _ = console_ui();
+    }
+
+    #[test]
+    fn embedded_page_is_nonempty_esm() {
+        assert!(PAGE_JS.contains("export"), "built page.js looks wrong");
+    }
+
+    #[test]
+    fn embedded_styles_are_scoped() {
+        assert!(
+            STYLES_CSS.contains(r#"[data-iii-ui="eval"]"#)
+                || STYLES_CSS.contains("[data-iii-ui=eval]"),
+            "built styles.css must be scoped under the eval UI attribute"
+        );
+    }
+}
diff --git a/eval/tests/golden/schemas/eval.assert.exact.json b/eval/tests/golden/schemas/eval.assert.exact.json
new file mode 100644
index 000000000..9e3e0a983
--- /dev/null
+++ b/eval/tests/golden/schemas/eval.assert.exact.json
@@ -0,0 +1,370 @@
+{
+  "function_id": "eval::assert::exact",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "definitions": {
+      "SessionMetricsResponseV1": {
+        "additionalProperties": false,
+        "properties": {
+          "by_session": {
+            "items": {
+              "$ref": "#/definitions/SessionUsageV1"
+            },
+            "type": "array"
+          },
+          "complete": {
+            "type": "boolean"
+          },
+          "root_session_id": {
+            "type": "string"
+          },
+          "totals": {
+            "$ref": "#/definitions/SessionUsageTotalsV1"
+          },
+          "traces": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/SessionTraceMetricsV1"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "description": "Trace/span aggregates when the engine's in-memory observability exporter is available. Usage metrics remain available when it is not."
+          }
+        },
+        "required": [
+          "by_session",
+          "complete",
+          "root_session_id",
+          "totals"
+        ],
+        "type": "object"
+      },
+      "SessionTraceMetricsV1": {
+        "additionalProperties": false,
+        "properties": {
+          "by_session": {
+            "items": {
+              "$ref": "#/definitions/SessionTraceUsageV1"
+            },
+            "type": "array"
+          },
+          "duration_ms": {
+            "description": "Elapsed window from the first observed span to the last observed span.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "error_span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "trace_count": {
+            "description": "Distinct traces across the root session and all descendants.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "by_session",
+          "duration_ms",
+          "error_span_count",
+          "span_count",
+          "trace_count"
+        ],
+        "type": "object"
+      },
+      "SessionTraceUsageV1": {
+        "additionalProperties": false,
+        "properties": {
+          "depth": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "duration_ms": {
+            "description": "Elapsed window from the session's first observed span to its last.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "error_span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "parent_session_id": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "session_id": {
+            "type": "string"
+          },
+          "span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "trace_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "depth",
+          "duration_ms",
+          "error_span_count",
+          "session_id",
+          "span_count",
+          "trace_count"
+        ],
+        "type": "object"
+      },
+      "SessionUsageTotalsV1": {
+        "additionalProperties": false,
+        "properties": {
+          "cache_read_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cache_write_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "function_call_errors": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "function_calls": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "input_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "output_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "reasoning_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "sessions": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "turns": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "function_call_errors",
+          "function_calls",
+          "sessions",
+          "turns"
+        ],
+        "type": "object"
+      },
+      "SessionUsageV1": {
+        "additionalProperties": false,
+        "properties": {
+          "cache_read_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cache_write_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "depth": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "function_call_errors": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "function_calls": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "input_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "output_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "parent_session_id": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "reasoning_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "session_id": {
+            "type": "string"
+          },
+          "turns": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "depth",
+          "function_call_errors",
+          "function_calls",
+          "session_id",
+          "turns"
+        ],
+        "type": "object"
+      },
+      "VariantRoleV1": {
+        "enum": [
+          "control",
+          "treatment"
+        ],
+        "type": "string"
+      }
+    },
+    "properties": {
+      "arguments": {
+        "default": {}
+      },
+      "evaluation_id": {
+        "type": "string"
+      },
+      "metrics": {
+        "$ref": "#/definitions/SessionMetricsResponseV1"
+      },
+      "output": true,
+      "role": {
+        "$ref": "#/definitions/VariantRoleV1"
+      },
+      "run_id": {
+        "type": "string"
+      },
+      "session_id": {
+        "type": "string"
+      }
+    },
+    "required": [
+      "evaluation_id",
+      "metrics",
+      "output",
+      "role",
+      "run_id",
+      "session_id"
+    ],
+    "title": "EvaluatorInputV1",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "additionalProperties": false,
+    "properties": {
+      "details": true,
+      "passed": {
+        "type": "boolean"
+      },
+      "reason": {
+        "type": [
+          "string",
+          "null"
+        ]
+      },
+      "score": {
+        "format": "double",
+        "type": [
+          "number",
+          "null"
+        ]
+      }
+    },
+    "required": [
+      "passed"
+    ],
+    "title": "EvaluatorResponseV1",
+    "type": "object"
+  }
+}
diff --git a/eval/tests/golden/schemas/eval.assert.normalized_text.json b/eval/tests/golden/schemas/eval.assert.normalized_text.json
new file mode 100644
index 000000000..b8cceef70
--- /dev/null
+++ b/eval/tests/golden/schemas/eval.assert.normalized_text.json
@@ -0,0 +1,370 @@
+{
+  "function_id": "eval::assert::normalized_text",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "definitions": {
+      "SessionMetricsResponseV1": {
+        "additionalProperties": false,
+        "properties": {
+          "by_session": {
+            "items": {
+              "$ref": "#/definitions/SessionUsageV1"
+            },
+            "type": "array"
+          },
+          "complete": {
+            "type": "boolean"
+          },
+          "root_session_id": {
+            "type": "string"
+          },
+          "totals": {
+            "$ref": "#/definitions/SessionUsageTotalsV1"
+          },
+          "traces": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/SessionTraceMetricsV1"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "description": "Trace/span aggregates when the engine's in-memory observability exporter is available. Usage metrics remain available when it is not."
+          }
+        },
+        "required": [
+          "by_session",
+          "complete",
+          "root_session_id",
+          "totals"
+        ],
+        "type": "object"
+      },
+      "SessionTraceMetricsV1": {
+        "additionalProperties": false,
+        "properties": {
+          "by_session": {
+            "items": {
+              "$ref": "#/definitions/SessionTraceUsageV1"
+            },
+            "type": "array"
+          },
+          "duration_ms": {
+            "description": "Elapsed window from the first observed span to the last observed span.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "error_span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "trace_count": {
+            "description": "Distinct traces across the root session and all descendants.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "by_session",
+          "duration_ms",
+          "error_span_count",
+          "span_count",
+          "trace_count"
+        ],
+        "type": "object"
+      },
+      "SessionTraceUsageV1": {
+        "additionalProperties": false,
+        "properties": {
+          "depth": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "duration_ms": {
+            "description": "Elapsed window from the session's first observed span to its last.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "error_span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "parent_session_id": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "session_id": {
+            "type": "string"
+          },
+          "span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "trace_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "depth",
+          "duration_ms",
+          "error_span_count",
+          "session_id",
+          "span_count",
+          "trace_count"
+        ],
+        "type": "object"
+      },
+      "SessionUsageTotalsV1": {
+        "additionalProperties": false,
+        "properties": {
+          "cache_read_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cache_write_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "function_call_errors": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "function_calls": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "input_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "output_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "reasoning_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "sessions": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "turns": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "function_call_errors",
+          "function_calls",
+          "sessions",
+          "turns"
+        ],
+        "type": "object"
+      },
+      "SessionUsageV1": {
+        "additionalProperties": false,
+        "properties": {
+          "cache_read_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cache_write_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "depth": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "function_call_errors": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "function_calls": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "input_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "output_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "parent_session_id": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "reasoning_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "session_id": {
+            "type": "string"
+          },
+          "turns": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "depth",
+          "function_call_errors",
+          "function_calls",
+          "session_id",
+          "turns"
+        ],
+        "type": "object"
+      },
+      "VariantRoleV1": {
+        "enum": [
+          "control",
+          "treatment"
+        ],
+        "type": "string"
+      }
+    },
+    "properties": {
+      "arguments": {
+        "default": {}
+      },
+      "evaluation_id": {
+        "type": "string"
+      },
+      "metrics": {
+        "$ref": "#/definitions/SessionMetricsResponseV1"
+      },
+      "output": true,
+      "role": {
+        "$ref": "#/definitions/VariantRoleV1"
+      },
+      "run_id": {
+        "type": "string"
+      },
+      "session_id": {
+        "type": "string"
+      }
+    },
+    "required": [
+      "evaluation_id",
+      "metrics",
+      "output",
+      "role",
+      "run_id",
+      "session_id"
+    ],
+    "title": "EvaluatorInputV1",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "additionalProperties": false,
+    "properties": {
+      "details": true,
+      "passed": {
+        "type": "boolean"
+      },
+      "reason": {
+        "type": [
+          "string",
+          "null"
+        ]
+      },
+      "score": {
+        "format": "double",
+        "type": [
+          "number",
+          "null"
+        ]
+      }
+    },
+    "required": [
+      "passed"
+    ],
+    "title": "EvaluatorResponseV1",
+    "type": "object"
+  }
+}
diff --git a/eval/tests/golden/schemas/eval.cancel.json b/eval/tests/golden/schemas/eval.cancel.json
new file mode 100644
index 000000000..3310e83ee
--- /dev/null
+++ b/eval/tests/golden/schemas/eval.cancel.json
@@ -0,0 +1,46 @@
+{
+  "function_id": "eval::cancel",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "properties": {
+      "evaluation_id": {
+        "type": "string"
+      }
+    },
+    "required": [
+      "evaluation_id"
+    ],
+    "title": "EvaluationIdRequestV1",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "additionalProperties": false,
+    "definitions": {
+      "EvalStatusV1": {
+        "enum": [
+          "queued",
+          "running",
+          "completed",
+          "failed",
+          "cancelled"
+        ],
+        "type": "string"
+      }
+    },
+    "properties": {
+      "cancelled": {
+        "type": "boolean"
+      },
+      "status": {
+        "$ref": "#/definitions/EvalStatusV1"
+      }
+    },
+    "required": [
+      "cancelled",
+      "status"
+    ],
+    "title": "EvalCancelResponseV1",
+    "type": "object"
+  }
+}
diff --git a/eval/tests/golden/schemas/eval.delete.json b/eval/tests/golden/schemas/eval.delete.json
new file mode 100644
index 000000000..c453a337c
--- /dev/null
+++ b/eval/tests/golden/schemas/eval.delete.json
@@ -0,0 +1,30 @@
+{
+  "function_id": "eval::delete",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "properties": {
+      "evaluation_id": {
+        "type": "string"
+      }
+    },
+    "required": [
+      "evaluation_id"
+    ],
+    "title": "EvaluationIdRequestV1",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "additionalProperties": false,
+    "properties": {
+      "deleted": {
+        "type": "boolean"
+      }
+    },
+    "required": [
+      "deleted"
+    ],
+    "title": "EvalDeleteResponseV1",
+    "type": "object"
+  }
+}
diff --git a/eval/tests/golden/schemas/eval.list.json b/eval/tests/golden/schemas/eval.list.json
new file mode 100644
index 000000000..8c2715f57
--- /dev/null
+++ b/eval/tests/golden/schemas/eval.list.json
@@ -0,0 +1,149 @@
+{
+  "function_id": "eval::list",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "properties": {
+      "limit": {
+        "format": "uint32",
+        "minimum": 0.0,
+        "type": [
+          "integer",
+          "null"
+        ]
+      }
+    },
+    "title": "EvalListRequestV1",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "additionalProperties": false,
+    "definitions": {
+      "ComparisonDimensionV1": {
+        "enum": [
+          "prompt",
+          "system_prompt"
+        ],
+        "type": "string"
+      },
+      "EvalStatusV1": {
+        "enum": [
+          "queued",
+          "running",
+          "completed",
+          "failed",
+          "cancelled"
+        ],
+        "type": "string"
+      },
+      "EvalSummaryV1": {
+        "additionalProperties": false,
+        "properties": {
+          "completed_at": {
+            "format": "int64",
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "control_label": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "created_at": {
+            "format": "int64",
+            "type": "integer"
+          },
+          "dimension": {
+            "$ref": "#/definitions/ComparisonDimensionV1"
+          },
+          "eligible": {
+            "type": [
+              "boolean",
+              "null"
+            ]
+          },
+          "error": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "evaluation_id": {
+            "type": "string"
+          },
+          "failed_runs": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "model": {
+            "type": "string"
+          },
+          "passed_runs": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "provider": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "status": {
+            "$ref": "#/definitions/EvalStatusV1"
+          },
+          "terminal_runs": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "total_runs": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "treatment_label": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "updated_at": {
+            "format": "int64",
+            "type": "integer"
+          }
+        },
+        "required": [
+          "created_at",
+          "dimension",
+          "evaluation_id",
+          "failed_runs",
+          "model",
+          "passed_runs",
+          "status",
+          "terminal_runs",
+          "total_runs",
+          "updated_at"
+        ],
+        "type": "object"
+      }
+    },
+    "properties": {
+      "evaluations": {
+        "items": {
+          "$ref": "#/definitions/EvalSummaryV1"
+        },
+        "type": "array"
+      }
+    },
+    "required": [
+      "evaluations"
+    ],
+    "title": "EvalListResponseV1",
+    "type": "object"
+  }
+}
diff --git a/eval/tests/golden/schemas/eval.on-turn-completed.json b/eval/tests/golden/schemas/eval.on-turn-completed.json
new file mode 100644
index 000000000..5a65ab0fa
--- /dev/null
+++ b/eval/tests/golden/schemas/eval.on-turn-completed.json
@@ -0,0 +1,32 @@
+{
+  "function_id": "eval::on-turn-completed",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "properties": {
+      "session_id": {
+        "default": "",
+        "type": "string"
+      },
+      "terminal": {
+        "default": false,
+        "type": "boolean"
+      }
+    },
+    "title": "WakeEventV1",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "additionalProperties": false,
+    "properties": {
+      "woke": {
+        "type": "boolean"
+      }
+    },
+    "required": [
+      "woke"
+    ],
+    "title": "WakeResponseV1",
+    "type": "object"
+  }
+}
diff --git a/eval/tests/golden/schemas/eval.rerun.json b/eval/tests/golden/schemas/eval.rerun.json
new file mode 100644
index 000000000..802fc92bf
--- /dev/null
+++ b/eval/tests/golden/schemas/eval.rerun.json
@@ -0,0 +1,50 @@
+{
+  "function_id": "eval::rerun",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "properties": {
+      "evaluation_id": {
+        "type": "string"
+      },
+      "reverse_order": {
+        "default": false,
+        "type": "boolean"
+      }
+    },
+    "required": [
+      "evaluation_id"
+    ],
+    "title": "EvalRerunRequestV1",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "additionalProperties": false,
+    "definitions": {
+      "EvalStatusV1": {
+        "enum": [
+          "queued",
+          "running",
+          "completed",
+          "failed",
+          "cancelled"
+        ],
+        "type": "string"
+      }
+    },
+    "properties": {
+      "evaluation_id": {
+        "type": "string"
+      },
+      "status": {
+        "$ref": "#/definitions/EvalStatusV1"
+      }
+    },
+    "required": [
+      "evaluation_id",
+      "status"
+    ],
+    "title": "EvalStartResponseV1",
+    "type": "object"
+  }
+}
diff --git a/eval/tests/golden/schemas/eval.result.json b/eval/tests/golden/schemas/eval.result.json
new file mode 100644
index 000000000..fd603052e
--- /dev/null
+++ b/eval/tests/golden/schemas/eval.result.json
@@ -0,0 +1,1557 @@
+{
+  "function_id": "eval::result",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "properties": {
+      "evaluation_id": {
+        "type": "string"
+      }
+    },
+    "required": [
+      "evaluation_id"
+    ],
+    "title": "EvaluationIdRequestV1",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "anyOf": [
+      {
+        "$ref": "#/definitions/EvalResultResponseV1"
+      },
+      {
+        "type": "null"
+      }
+    ],
+    "definitions": {
+      "AggregateDeltaV1": {
+        "additionalProperties": false,
+        "properties": {
+          "median_cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "median_error_span_count": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "median_function_call_errors": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "median_function_calls": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "median_reasoning_tokens": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "median_score": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "median_span_count": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "median_total_tokens": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "median_trace_count": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "median_trace_duration_ms": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "median_wall_time_ms": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "pass_rate": {
+            "format": "double",
+            "type": "number"
+          }
+        },
+        "required": [
+          "pass_rate"
+        ],
+        "type": "object"
+      },
+      "ComparisonDimensionV1": {
+        "enum": [
+          "prompt",
+          "system_prompt"
+        ],
+        "type": "string"
+      },
+      "EvalBenchmarkV1": {
+        "additionalProperties": false,
+        "properties": {
+          "cache_read_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cache_write_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "error_span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "function_call_errors": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "function_calls": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "input_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "output_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "reasoning_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "sessions": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "total_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "trace_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "trace_duration_ms": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "turns": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "wall_time_ms": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "function_call_errors",
+          "function_calls",
+          "sessions",
+          "turns",
+          "wall_time_ms"
+        ],
+        "type": "object"
+      },
+      "EvalFailureV1": {
+        "additionalProperties": false,
+        "properties": {
+          "code": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "function_id": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "message": {
+            "type": "string"
+          },
+          "phase": {
+            "$ref": "#/definitions/EvalPhaseV1"
+          }
+        },
+        "required": [
+          "message",
+          "phase"
+        ],
+        "type": "object"
+      },
+      "EvalLimitsV1": {
+        "additionalProperties": false,
+        "properties": {
+          "evaluation": {
+            "allOf": [
+              {
+                "$ref": "#/definitions/EvaluationLimitsV1"
+              }
+            ],
+            "default": {}
+          },
+          "execution": {
+            "allOf": [
+              {
+                "$ref": "#/definitions/ExecutionLimitsV1"
+              }
+            ],
+            "default": {
+              "invocation_timeout_seconds": 120,
+              "max_output_tokens_per_call": 8192,
+              "max_turns": 1000,
+              "scenario_timeout_seconds": 600
+            }
+          }
+        },
+        "type": "object"
+      },
+      "EvalModelConfigV1": {
+        "additionalProperties": false,
+        "properties": {
+          "mode": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/Mode"
+              },
+              {
+                "type": "null"
+              }
+            ]
+          },
+          "model": {
+            "type": "string"
+          },
+          "provider": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "provider_options": {
+            "additionalProperties": true,
+            "type": [
+              "object",
+              "null"
+            ]
+          },
+          "system_prompt_strategy": {
+            "allOf": [
+              {
+                "$ref": "#/definitions/SystemPromptStrategy"
+              }
+            ],
+            "default": "enrich"
+          },
+          "thinking_level": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/ThinkingLevel"
+              },
+              {
+                "type": "null"
+              }
+            ]
+          }
+        },
+        "required": [
+          "model"
+        ],
+        "type": "object"
+      },
+      "EvalPhaseV1": {
+        "enum": [
+          "setup",
+          "send",
+          "await",
+          "collect",
+          "evaluate",
+          "limit",
+          "cancel"
+        ],
+        "type": "string"
+      },
+      "EvalProgressV1": {
+        "additionalProperties": false,
+        "properties": {
+          "control_aggregate": {
+            "$ref": "#/definitions/VariantAggregateV1"
+          },
+          "delta": {
+            "$ref": "#/definitions/AggregateDeltaV1"
+          },
+          "effective_execution_order": {
+            "items": {
+              "type": "string"
+            },
+            "type": "array"
+          },
+          "order_sensitivity": {
+            "$ref": "#/definitions/OrderSensitivityV1"
+          },
+          "runs": {
+            "items": {
+              "$ref": "#/definitions/EvalRunReportV1"
+            },
+            "type": "array"
+          },
+          "treatment_aggregate": {
+            "$ref": "#/definitions/VariantAggregateV1"
+          }
+        },
+        "required": [
+          "control_aggregate",
+          "delta",
+          "effective_execution_order",
+          "order_sensitivity",
+          "runs",
+          "treatment_aggregate"
+        ],
+        "type": "object"
+      },
+      "EvalReportV1": {
+        "additionalProperties": false,
+        "properties": {
+          "completed_at": {
+            "format": "int64",
+            "type": "integer"
+          },
+          "control": {
+            "$ref": "#/definitions/VariantArtifactV1"
+          },
+          "control_aggregate": {
+            "$ref": "#/definitions/VariantAggregateV1"
+          },
+          "created_at": {
+            "format": "int64",
+            "type": "integer"
+          },
+          "delta": {
+            "allOf": [
+              {
+                "$ref": "#/definitions/AggregateDeltaV1"
+              }
+            ],
+            "description": "Treatment minus control. Negative efficiency deltas are improvements."
+          },
+          "dimension": {
+            "$ref": "#/definitions/ComparisonDimensionV1"
+          },
+          "effective_execution_order": {
+            "items": {
+              "type": "string"
+            },
+            "type": "array"
+          },
+          "eligible": {
+            "type": [
+              "boolean",
+              "null"
+            ]
+          },
+          "evaluation_id": {
+            "type": "string"
+          },
+          "evaluator": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/EvaluatorArtifactV1"
+              },
+              {
+                "type": "null"
+              }
+            ]
+          },
+          "execution_order_policy": {
+            "$ref": "#/definitions/ExecutionOrderV1"
+          },
+          "limits": {
+            "$ref": "#/definitions/EvalLimitsV1"
+          },
+          "model": {
+            "$ref": "#/definitions/EvalModelConfigV1"
+          },
+          "order_sensitivity": {
+            "$ref": "#/definitions/OrderSensitivityV1"
+          },
+          "runs": {
+            "items": {
+              "$ref": "#/definitions/EvalRunReportV1"
+            },
+            "type": "array"
+          },
+          "schema_version": {
+            "type": "string"
+          },
+          "shared_artifacts": {
+            "$ref": "#/definitions/SharedArtifactsV1"
+          },
+          "source_evaluation_id": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "treatment": {
+            "$ref": "#/definitions/VariantArtifactV1"
+          },
+          "treatment_aggregate": {
+            "$ref": "#/definitions/VariantAggregateV1"
+          }
+        },
+        "required": [
+          "completed_at",
+          "control",
+          "control_aggregate",
+          "created_at",
+          "delta",
+          "dimension",
+          "effective_execution_order",
+          "evaluation_id",
+          "execution_order_policy",
+          "limits",
+          "model",
+          "order_sensitivity",
+          "runs",
+          "schema_version",
+          "shared_artifacts",
+          "treatment",
+          "treatment_aggregate"
+        ],
+        "type": "object"
+      },
+      "EvalResultResponseV1": {
+        "additionalProperties": false,
+        "properties": {
+          "progress": {
+            "$ref": "#/definitions/EvalProgressV1"
+          },
+          "report": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/EvalReportV1"
+              },
+              {
+                "type": "null"
+              }
+            ]
+          },
+          "request": {
+            "$ref": "#/definitions/NormalizedEvalRequestV1"
+          },
+          "status": {
+            "$ref": "#/definitions/EvalStatusV1"
+          }
+        },
+        "required": [
+          "progress",
+          "request",
+          "status"
+        ],
+        "type": "object"
+      },
+      "EvalRunReportV1": {
+        "additionalProperties": false,
+        "properties": {
+          "benchmark": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/EvalBenchmarkV1"
+              },
+              {
+                "type": "null"
+              }
+            ]
+          },
+          "completed_at": {
+            "format": "int64",
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "evaluation": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/EvaluatorResponseV1"
+              },
+              {
+                "type": "null"
+              }
+            ]
+          },
+          "execution_position": {
+            "default": 0,
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "failures": {
+            "items": {
+              "$ref": "#/definitions/EvalFailureV1"
+            },
+            "type": "array"
+          },
+          "iteration": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "metrics": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/SessionMetricsResponseV1"
+              },
+              {
+                "type": "null"
+              }
+            ]
+          },
+          "output": true,
+          "pair_position": {
+            "default": 0,
+            "format": "uint8",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "passed": {
+            "type": [
+              "boolean",
+              "null"
+            ]
+          },
+          "role": {
+            "$ref": "#/definitions/VariantRoleV1"
+          },
+          "run_id": {
+            "type": "string"
+          },
+          "session_id": {
+            "type": "string"
+          },
+          "started_at": {
+            "format": "int64",
+            "type": "integer"
+          },
+          "status": {
+            "$ref": "#/definitions/EvalRunStatusV1"
+          },
+          "turn_id": {
+            "type": [
+              "string",
+              "null"
+            ]
+          }
+        },
+        "required": [
+          "iteration",
+          "role",
+          "run_id",
+          "session_id",
+          "started_at",
+          "status"
+        ],
+        "type": "object"
+      },
+      "EvalRunStatusV1": {
+        "enum": [
+          "pending",
+          "running",
+          "completed",
+          "failed",
+          "cancelled"
+        ],
+        "type": "string"
+      },
+      "EvalStatusV1": {
+        "enum": [
+          "queued",
+          "running",
+          "completed",
+          "failed",
+          "cancelled"
+        ],
+        "type": "string"
+      },
+      "EvalVariantV1": {
+        "additionalProperties": false,
+        "properties": {
+          "label": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "prompt": {
+            "type": "string"
+          },
+          "system_prompt": {
+            "description": "`null` disables the system prompt for this variant. An empty string keeps the harness behavior of resolving the provider or built-in prompt.",
+            "type": [
+              "string",
+              "null"
+            ]
+          }
+        },
+        "required": [
+          "prompt"
+        ],
+        "type": "object"
+      },
+      "EvaluationLimitsV1": {
+        "additionalProperties": false,
+        "properties": {
+          "max_error_spans": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "max_function_call_errors": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          }
+        },
+        "type": "object"
+      },
+      "EvaluatorArtifactV1": {
+        "additionalProperties": false,
+        "properties": {
+          "arguments_sha256": {
+            "type": "string"
+          },
+          "function_id": {
+            "type": "string"
+          }
+        },
+        "required": [
+          "arguments_sha256",
+          "function_id"
+        ],
+        "type": "object"
+      },
+      "EvaluatorResponseV1": {
+        "additionalProperties": false,
+        "properties": {
+          "details": true,
+          "passed": {
+            "type": "boolean"
+          },
+          "reason": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "score": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          }
+        },
+        "required": [
+          "passed"
+        ],
+        "type": "object"
+      },
+      "EvaluatorSpecV1": {
+        "additionalProperties": false,
+        "properties": {
+          "arguments": {
+            "default": {}
+          },
+          "function_id": {
+            "type": "string"
+          }
+        },
+        "required": [
+          "function_id"
+        ],
+        "type": "object"
+      },
+      "ExecutionLimitsV1": {
+        "additionalProperties": false,
+        "properties": {
+          "invocation_timeout_seconds": {
+            "default": 120,
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "max_cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "max_output_tokens_per_call": {
+            "default": 8192,
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "max_total_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "max_turns": {
+            "default": 1000,
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "scenario_timeout_seconds": {
+            "default": 600,
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "type": "object"
+      },
+      "ExecutionOrderV1": {
+        "enum": [
+          "balanced_control_first",
+          "balanced_treatment_first"
+        ],
+        "type": "string"
+      },
+      "ExposeMode": {
+        "description": "How allowed functions reach the model (harness.md § Exposure modes).",
+        "enum": [
+          "agent_trigger",
+          "native"
+        ],
+        "type": "string"
+      },
+      "FunctionPolicy": {
+        "description": "The fail-closed dispatch policy (harness.md § Functions). Absent on the send => every call denied (a plain chat loop).",
+        "properties": {
+          "allow": {
+            "default": [],
+            "items": {
+              "type": "string"
+            },
+            "type": "array"
+          },
+          "deny": {
+            "default": [],
+            "items": {
+              "type": "string"
+            },
+            "type": "array"
+          },
+          "expose": {
+            "allOf": [
+              {
+                "$ref": "#/definitions/ExposeMode"
+              }
+            ],
+            "default": "agent_trigger"
+          }
+        },
+        "type": "object"
+      },
+      "Mode": {
+        "description": "Console / send operating mode — prepends a short paragraph before the shared identity prompt. `ask` is also structural: the turn's dispatch policy is capped at the configured default policy, never widened.",
+        "enum": [
+          "ask",
+          "agent"
+        ],
+        "type": "string"
+      },
+      "NormalizedEvalRequestV1": {
+        "additionalProperties": false,
+        "properties": {
+          "control": {
+            "$ref": "#/definitions/EvalVariantV1"
+          },
+          "dimension": {
+            "$ref": "#/definitions/ComparisonDimensionV1"
+          },
+          "evaluator": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/EvaluatorSpecV1"
+              },
+              {
+                "type": "null"
+              }
+            ]
+          },
+          "execution_order": {
+            "allOf": [
+              {
+                "$ref": "#/definitions/ExecutionOrderV1"
+              }
+            ],
+            "default": "balanced_control_first"
+          },
+          "functions": {
+            "$ref": "#/definitions/FunctionPolicy"
+          },
+          "limits": {
+            "$ref": "#/definitions/EvalLimitsV1"
+          },
+          "metadata": true,
+          "model": {
+            "$ref": "#/definitions/EvalModelConfigV1"
+          },
+          "output": {
+            "$ref": "#/definitions/OutputContract"
+          },
+          "runs": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "source_evaluation_id": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "treatment": {
+            "$ref": "#/definitions/EvalVariantV1"
+          }
+        },
+        "required": [
+          "control",
+          "dimension",
+          "functions",
+          "limits",
+          "model",
+          "output",
+          "runs",
+          "treatment"
+        ],
+        "type": "object"
+      },
+      "OrderSensitivityV1": {
+        "additionalProperties": false,
+        "properties": {
+          "control": {
+            "$ref": "#/definitions/RoleOrderSensitivityV1"
+          },
+          "detected": {
+            "type": "boolean"
+          },
+          "treatment": {
+            "$ref": "#/definitions/RoleOrderSensitivityV1"
+          }
+        },
+        "required": [
+          "control",
+          "detected",
+          "treatment"
+        ],
+        "type": "object"
+      },
+      "OutputContract": {
+        "description": "Free text by default; `json` constrains the final answer to a JSON value, validated against `schema` when supplied.",
+        "oneOf": [
+          {
+            "properties": {
+              "type": {
+                "enum": [
+                  "text"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "schema": true,
+              "type": {
+                "enum": [
+                  "json"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "type"
+            ],
+            "type": "object"
+          }
+        ]
+      },
+      "RoleOrderSensitivityV1": {
+        "additionalProperties": false,
+        "properties": {
+          "differs": {
+            "type": "boolean"
+          },
+          "first_passed": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "first_runs": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "second_passed": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "second_runs": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "differs",
+          "first_passed",
+          "first_runs",
+          "second_passed",
+          "second_runs"
+        ],
+        "type": "object"
+      },
+      "SessionMetricsResponseV1": {
+        "additionalProperties": false,
+        "properties": {
+          "by_session": {
+            "items": {
+              "$ref": "#/definitions/SessionUsageV1"
+            },
+            "type": "array"
+          },
+          "complete": {
+            "type": "boolean"
+          },
+          "root_session_id": {
+            "type": "string"
+          },
+          "totals": {
+            "$ref": "#/definitions/SessionUsageTotalsV1"
+          },
+          "traces": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/SessionTraceMetricsV1"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "description": "Trace/span aggregates when the engine's in-memory observability exporter is available. Usage metrics remain available when it is not."
+          }
+        },
+        "required": [
+          "by_session",
+          "complete",
+          "root_session_id",
+          "totals"
+        ],
+        "type": "object"
+      },
+      "SessionTraceMetricsV1": {
+        "additionalProperties": false,
+        "properties": {
+          "by_session": {
+            "items": {
+              "$ref": "#/definitions/SessionTraceUsageV1"
+            },
+            "type": "array"
+          },
+          "duration_ms": {
+            "description": "Elapsed window from the first observed span to the last observed span.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "error_span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "trace_count": {
+            "description": "Distinct traces across the root session and all descendants.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "by_session",
+          "duration_ms",
+          "error_span_count",
+          "span_count",
+          "trace_count"
+        ],
+        "type": "object"
+      },
+      "SessionTraceUsageV1": {
+        "additionalProperties": false,
+        "properties": {
+          "depth": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "duration_ms": {
+            "description": "Elapsed window from the session's first observed span to its last.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "error_span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "parent_session_id": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "session_id": {
+            "type": "string"
+          },
+          "span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "trace_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "depth",
+          "duration_ms",
+          "error_span_count",
+          "session_id",
+          "span_count",
+          "trace_count"
+        ],
+        "type": "object"
+      },
+      "SessionUsageTotalsV1": {
+        "additionalProperties": false,
+        "properties": {
+          "cache_read_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cache_write_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "function_call_errors": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "function_calls": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "input_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "output_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "reasoning_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "sessions": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "turns": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "function_call_errors",
+          "function_calls",
+          "sessions",
+          "turns"
+        ],
+        "type": "object"
+      },
+      "SessionUsageV1": {
+        "additionalProperties": false,
+        "properties": {
+          "cache_read_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cache_write_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "depth": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "function_call_errors": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "function_calls": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "input_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "output_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "parent_session_id": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "reasoning_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "session_id": {
+            "type": "string"
+          },
+          "turns": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "depth",
+          "function_call_errors",
+          "function_calls",
+          "session_id",
+          "turns"
+        ],
+        "type": "object"
+      },
+      "SharedArtifactsV1": {
+        "additionalProperties": false,
+        "properties": {
+          "function_policy_sha256": {
+            "type": "string"
+          },
+          "limits_sha256": {
+            "type": "string"
+          },
+          "model_sha256": {
+            "type": "string"
+          },
+          "output_sha256": {
+            "type": "string"
+          }
+        },
+        "required": [
+          "function_policy_sha256",
+          "limits_sha256",
+          "model_sha256",
+          "output_sha256"
+        ],
+        "type": "object"
+      },
+      "SystemPromptStrategy": {
+        "description": "How a caller-supplied system prompt combines with the built-in identity prompt.",
+        "oneOf": [
+          {
+            "description": "Caller prompt replaces the built-in prompt verbatim.",
+            "enum": [
+              "override"
+            ],
+            "type": "string"
+          },
+          {
+            "description": "Caller prompt is appended to the built-in identity prompt.",
+            "enum": [
+              "enrich"
+            ],
+            "type": "string"
+          },
+          {
+            "description": "No system prompt is sent to the model.",
+            "enum": [
+              "disabled"
+            ],
+            "type": "string"
+          }
+        ]
+      },
+      "ThinkingLevel": {
+        "enum": [
+          "minimal",
+          "low",
+          "medium",
+          "high",
+          "xhigh"
+        ],
+        "type": "string"
+      },
+      "VariantAggregateV1": {
+        "additionalProperties": false,
+        "properties": {
+          "benchmarked_runs": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "evaluated_runs": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "median_cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "median_error_span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "median_function_call_errors": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "median_function_calls": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "median_input_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "median_output_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "median_reasoning_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "median_score": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "median_span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "median_total_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "median_trace_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "median_trace_duration_ms": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "median_wall_time_ms": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "pass_rate": {
+            "format": "double",
+            "type": "number"
+          },
+          "passed": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "runs": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "benchmarked_runs",
+          "evaluated_runs",
+          "pass_rate",
+          "passed",
+          "runs"
+        ],
+        "type": "object"
+      },
+      "VariantArtifactV1": {
+        "additionalProperties": false,
+        "properties": {
+          "label": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "prompt_sha256": {
+            "type": "string"
+          },
+          "system_prompt_sha256": {
+            "type": [
+              "string",
+              "null"
+            ]
+          }
+        },
+        "required": [
+          "prompt_sha256"
+        ],
+        "type": "object"
+      },
+      "VariantRoleV1": {
+        "enum": [
+          "control",
+          "treatment"
+        ],
+        "type": "string"
+      }
+    },
+    "title": "Nullable_EvalResultResponseV1"
+  }
+}
diff --git a/eval/tests/golden/schemas/eval.start.json b/eval/tests/golden/schemas/eval.start.json
new file mode 100644
index 000000000..b5f5b174f
--- /dev/null
+++ b/eval/tests/golden/schemas/eval.start.json
@@ -0,0 +1,444 @@
+{
+  "function_id": "eval::start",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "definitions": {
+      "ComparisonDimensionV1": {
+        "enum": [
+          "prompt",
+          "system_prompt"
+        ],
+        "type": "string"
+      },
+      "EvalLimitsV1": {
+        "additionalProperties": false,
+        "properties": {
+          "evaluation": {
+            "allOf": [
+              {
+                "$ref": "#/definitions/EvaluationLimitsV1"
+              }
+            ],
+            "default": {}
+          },
+          "execution": {
+            "allOf": [
+              {
+                "$ref": "#/definitions/ExecutionLimitsV1"
+              }
+            ],
+            "default": {
+              "invocation_timeout_seconds": 120,
+              "max_output_tokens_per_call": 8192,
+              "max_turns": 1000,
+              "scenario_timeout_seconds": 600
+            }
+          }
+        },
+        "type": "object"
+      },
+      "EvalModelConfigV1": {
+        "additionalProperties": false,
+        "properties": {
+          "mode": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/Mode"
+              },
+              {
+                "type": "null"
+              }
+            ]
+          },
+          "model": {
+            "type": "string"
+          },
+          "provider": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "provider_options": {
+            "additionalProperties": true,
+            "type": [
+              "object",
+              "null"
+            ]
+          },
+          "system_prompt_strategy": {
+            "allOf": [
+              {
+                "$ref": "#/definitions/SystemPromptStrategy"
+              }
+            ],
+            "default": "enrich"
+          },
+          "thinking_level": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/ThinkingLevel"
+              },
+              {
+                "type": "null"
+              }
+            ]
+          }
+        },
+        "required": [
+          "model"
+        ],
+        "type": "object"
+      },
+      "EvalVariantV1": {
+        "additionalProperties": false,
+        "properties": {
+          "label": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "prompt": {
+            "type": "string"
+          },
+          "system_prompt": {
+            "description": "`null` disables the system prompt for this variant. An empty string keeps the harness behavior of resolving the provider or built-in prompt.",
+            "type": [
+              "string",
+              "null"
+            ]
+          }
+        },
+        "required": [
+          "prompt"
+        ],
+        "type": "object"
+      },
+      "EvaluationLimitsV1": {
+        "additionalProperties": false,
+        "properties": {
+          "max_error_spans": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "max_function_call_errors": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          }
+        },
+        "type": "object"
+      },
+      "EvaluatorSpecV1": {
+        "additionalProperties": false,
+        "properties": {
+          "arguments": {
+            "default": {}
+          },
+          "function_id": {
+            "type": "string"
+          }
+        },
+        "required": [
+          "function_id"
+        ],
+        "type": "object"
+      },
+      "ExecutionLimitsV1": {
+        "additionalProperties": false,
+        "properties": {
+          "invocation_timeout_seconds": {
+            "default": 120,
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "max_cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "max_output_tokens_per_call": {
+            "default": 8192,
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "max_total_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "max_turns": {
+            "default": 1000,
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "scenario_timeout_seconds": {
+            "default": 600,
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "type": "object"
+      },
+      "ExecutionOrderV1": {
+        "enum": [
+          "balanced_control_first",
+          "balanced_treatment_first"
+        ],
+        "type": "string"
+      },
+      "ExposeMode": {
+        "description": "How allowed functions reach the model (harness.md § Exposure modes).",
+        "enum": [
+          "agent_trigger",
+          "native"
+        ],
+        "type": "string"
+      },
+      "FunctionPolicy": {
+        "description": "The fail-closed dispatch policy (harness.md § Functions). Absent on the send => every call denied (a plain chat loop).",
+        "properties": {
+          "allow": {
+            "default": [],
+            "items": {
+              "type": "string"
+            },
+            "type": "array"
+          },
+          "deny": {
+            "default": [],
+            "items": {
+              "type": "string"
+            },
+            "type": "array"
+          },
+          "expose": {
+            "allOf": [
+              {
+                "$ref": "#/definitions/ExposeMode"
+              }
+            ],
+            "default": "agent_trigger"
+          }
+        },
+        "type": "object"
+      },
+      "Mode": {
+        "description": "Console / send operating mode — prepends a short paragraph before the shared identity prompt. `ask` is also structural: the turn's dispatch policy is capped at the configured default policy, never widened.",
+        "enum": [
+          "ask",
+          "agent"
+        ],
+        "type": "string"
+      },
+      "OutputContract": {
+        "description": "Free text by default; `json` constrains the final answer to a JSON value, validated against `schema` when supplied.",
+        "oneOf": [
+          {
+            "properties": {
+              "type": {
+                "enum": [
+                  "text"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "schema": true,
+              "type": {
+                "enum": [
+                  "json"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "type"
+            ],
+            "type": "object"
+          }
+        ]
+      },
+      "SystemPromptStrategy": {
+        "description": "How a caller-supplied system prompt combines with the built-in identity prompt.",
+        "oneOf": [
+          {
+            "description": "Caller prompt replaces the built-in prompt verbatim.",
+            "enum": [
+              "override"
+            ],
+            "type": "string"
+          },
+          {
+            "description": "Caller prompt is appended to the built-in identity prompt.",
+            "enum": [
+              "enrich"
+            ],
+            "type": "string"
+          },
+          {
+            "description": "No system prompt is sent to the model.",
+            "enum": [
+              "disabled"
+            ],
+            "type": "string"
+          }
+        ]
+      },
+      "ThinkingLevel": {
+        "enum": [
+          "minimal",
+          "low",
+          "medium",
+          "high",
+          "xhigh"
+        ],
+        "type": "string"
+      }
+    },
+    "properties": {
+      "control": {
+        "$ref": "#/definitions/EvalVariantV1"
+      },
+      "dimension": {
+        "$ref": "#/definitions/ComparisonDimensionV1"
+      },
+      "evaluator": {
+        "anyOf": [
+          {
+            "$ref": "#/definitions/EvaluatorSpecV1"
+          },
+          {
+            "type": "null"
+          }
+        ]
+      },
+      "execution_order": {
+        "anyOf": [
+          {
+            "$ref": "#/definitions/ExecutionOrderV1"
+          },
+          {
+            "type": "null"
+          }
+        ]
+      },
+      "functions": {
+        "anyOf": [
+          {
+            "$ref": "#/definitions/FunctionPolicy"
+          },
+          {
+            "type": "null"
+          }
+        ],
+        "description": "Shared function policy for both variants. Omitted means deny all."
+      },
+      "limits": {
+        "allOf": [
+          {
+            "$ref": "#/definitions/EvalLimitsV1"
+          }
+        ],
+        "default": {
+          "evaluation": {},
+          "execution": {
+            "invocation_timeout_seconds": 120,
+            "max_output_tokens_per_call": 8192,
+            "max_turns": 1000,
+            "scenario_timeout_seconds": 600
+          }
+        }
+      },
+      "metadata": {
+        "description": "Shared harness metadata, including an optional filesystem scope."
+      },
+      "model": {
+        "$ref": "#/definitions/EvalModelConfigV1"
+      },
+      "output": {
+        "anyOf": [
+          {
+            "$ref": "#/definitions/OutputContract"
+          },
+          {
+            "type": "null"
+          }
+        ],
+        "description": "Shared output contract for both variants. Omitted means text."
+      },
+      "runs": {
+        "format": "uint32",
+        "minimum": 0.0,
+        "type": [
+          "integer",
+          "null"
+        ]
+      },
+      "treatment": {
+        "$ref": "#/definitions/EvalVariantV1"
+      }
+    },
+    "required": [
+      "control",
+      "dimension",
+      "model",
+      "treatment"
+    ],
+    "title": "EvalStartRequestV1",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "additionalProperties": false,
+    "definitions": {
+      "EvalStatusV1": {
+        "enum": [
+          "queued",
+          "running",
+          "completed",
+          "failed",
+          "cancelled"
+        ],
+        "type": "string"
+      }
+    },
+    "properties": {
+      "evaluation_id": {
+        "type": "string"
+      },
+      "status": {
+        "$ref": "#/definitions/EvalStatusV1"
+      }
+    },
+    "required": [
+      "evaluation_id",
+      "status"
+    ],
+    "title": "EvalStartResponseV1",
+    "type": "object"
+  }
+}
diff --git a/eval/tests/golden/schemas/eval.status.json b/eval/tests/golden/schemas/eval.status.json
new file mode 100644
index 000000000..5396d531b
--- /dev/null
+++ b/eval/tests/golden/schemas/eval.status.json
@@ -0,0 +1,157 @@
+{
+  "function_id": "eval::status",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "properties": {
+      "evaluation_id": {
+        "type": "string"
+      }
+    },
+    "required": [
+      "evaluation_id"
+    ],
+    "title": "EvaluationIdRequestV1",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "anyOf": [
+      {
+        "$ref": "#/definitions/EvalStatusResponseV1"
+      },
+      {
+        "type": "null"
+      }
+    ],
+    "definitions": {
+      "ActiveRunV1": {
+        "additionalProperties": false,
+        "properties": {
+          "iteration": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "role": {
+            "$ref": "#/definitions/VariantRoleV1"
+          },
+          "run_id": {
+            "type": "string"
+          },
+          "session_id": {
+            "type": "string"
+          },
+          "started_at": {
+            "format": "int64",
+            "type": "integer"
+          },
+          "turn_id": {
+            "type": [
+              "string",
+              "null"
+            ]
+          }
+        },
+        "required": [
+          "iteration",
+          "role",
+          "run_id",
+          "session_id",
+          "started_at"
+        ],
+        "type": "object"
+      },
+      "EvalStatusResponseV1": {
+        "additionalProperties": false,
+        "properties": {
+          "active": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/ActiveRunV1"
+              },
+              {
+                "type": "null"
+              }
+            ]
+          },
+          "completed_at": {
+            "format": "int64",
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "created_at": {
+            "format": "int64",
+            "type": "integer"
+          },
+          "error": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "evaluation_id": {
+            "type": "string"
+          },
+          "failed_runs": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "passed_runs": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "status": {
+            "$ref": "#/definitions/EvalStatusV1"
+          },
+          "terminal_runs": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "total_runs": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "updated_at": {
+            "format": "int64",
+            "type": "integer"
+          }
+        },
+        "required": [
+          "created_at",
+          "evaluation_id",
+          "failed_runs",
+          "passed_runs",
+          "status",
+          "terminal_runs",
+          "total_runs",
+          "updated_at"
+        ],
+        "type": "object"
+      },
+      "EvalStatusV1": {
+        "enum": [
+          "queued",
+          "running",
+          "completed",
+          "failed",
+          "cancelled"
+        ],
+        "type": "string"
+      },
+      "VariantRoleV1": {
+        "enum": [
+          "control",
+          "treatment"
+        ],
+        "type": "string"
+      }
+    },
+    "title": "Nullable_EvalStatusResponseV1"
+  }
+}
diff --git a/eval/tests/golden/schemas/eval.step.json b/eval/tests/golden/schemas/eval.step.json
new file mode 100644
index 000000000..a5a756076
--- /dev/null
+++ b/eval/tests/golden/schemas/eval.step.json
@@ -0,0 +1,52 @@
+{
+  "function_id": "eval::step",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "properties": {
+      "evaluation_id": {
+        "type": "string"
+      },
+      "step": {
+        "format": "uint64",
+        "minimum": 0.0,
+        "type": "integer"
+      }
+    },
+    "required": [
+      "evaluation_id",
+      "step"
+    ],
+    "title": "StepRequestV1",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "additionalProperties": false,
+    "definitions": {
+      "EvalStatusV1": {
+        "enum": [
+          "queued",
+          "running",
+          "completed",
+          "failed",
+          "cancelled"
+        ],
+        "type": "string"
+      }
+    },
+    "properties": {
+      "skipped": {
+        "type": "boolean"
+      },
+      "status": {
+        "$ref": "#/definitions/EvalStatusV1"
+      }
+    },
+    "required": [
+      "skipped",
+      "status"
+    ],
+    "title": "StepResponseV1",
+    "type": "object"
+  }
+}
diff --git a/eval/tests/golden/schemas/eval.sweep.json b/eval/tests/golden/schemas/eval.sweep.json
new file mode 100644
index 000000000..3feccdd87
--- /dev/null
+++ b/eval/tests/golden/schemas/eval.sweep.json
@@ -0,0 +1,41 @@
+{
+  "function_id": "eval::sweep",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "properties": {
+      "scheduled_at": {
+        "default": null,
+        "format": "int64",
+        "type": [
+          "integer",
+          "null"
+        ]
+      },
+      "scheduled_time": {
+        "default": null,
+        "type": [
+          "string",
+          "null"
+        ]
+      }
+    },
+    "title": "SweepEventV1",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "additionalProperties": false,
+    "properties": {
+      "swept": {
+        "format": "uint64",
+        "minimum": 0.0,
+        "type": "integer"
+      }
+    },
+    "required": [
+      "swept"
+    ],
+    "title": "SweepResponseV1",
+    "type": "object"
+  }
+}
diff --git a/eval/tests/manifest.rs b/eval/tests/manifest.rs
new file mode 100644
index 000000000..fca48d1c2
--- /dev/null
+++ b/eval/tests/manifest.rs
@@ -0,0 +1,42 @@
+use std::process::Command;
+
+use serde_json::Value;
+
+#[test]
+fn manifest_subcommand_emits_valid_json() {
+    let output = Command::new(env!("CARGO_BIN_EXE_eval"))
+        .arg("--manifest")
+        .output()
+        .expect("spawn eval --manifest");
+    assert!(
+        output.status.success(),
+        "stderr: {}",
+        String::from_utf8_lossy(&output.stderr)
+    );
+    let manifest: Value = serde_json::from_slice(&output.stdout).expect("valid manifest JSON");
+    assert_eq!(manifest["name"], "eval");
+    assert_eq!(manifest["version"], env!("CARGO_PKG_VERSION"));
+    assert!(manifest["description"]
+        .as_str()
+        .is_some_and(|value| !value.is_empty()));
+    assert!(manifest["default_config"].is_object());
+    assert!(manifest["supported_targets"]
+        .as_array()
+        .is_some_and(|targets| !targets.is_empty()));
+}
+
+#[test]
+fn worker_manifest_declares_runtime_dependencies() {
+    let source = std::fs::read_to_string(format!("{}/iii.worker.yaml", env!("CARGO_MANIFEST_DIR")))
+        .expect("read iii.worker.yaml");
+    let manifest: serde_yaml::Value = serde_yaml::from_str(&source).expect("parse worker manifest");
+    let dependencies = manifest["dependencies"]
+        .as_mapping()
+        .expect("dependencies map");
+    for dependency in ["harness", "state", "queue", "cron", "iii-observability"] {
+        assert!(
+            dependencies.contains_key(serde_yaml::Value::String(dependency.into())),
+            "missing {dependency}"
+        );
+    }
+}
diff --git a/eval/tests/schemas.rs b/eval/tests/schemas.rs
new file mode 100644
index 000000000..d1b463797
--- /dev/null
+++ b/eval/tests/schemas.rs
@@ -0,0 +1,55 @@
+mod support;
+
+use eval::surface::catalog;
+
+fn golden_file_name(function_id: &str) -> String {
+    format!("schemas/{}.json", function_id.replace("::", "."))
+}
+
+#[test]
+fn catalog_matches_the_registered_surface() {
+    let ids: Vec<_> = catalog().iter().map(|spec| spec.function_id).collect();
+    assert_eq!(
+        ids,
+        [
+            "eval::start",
+            "eval::rerun",
+            "eval::list",
+            "eval::status",
+            "eval::result",
+            "eval::cancel",
+            "eval::delete",
+            "eval::assert::exact",
+            "eval::assert::normalized_text",
+            "eval::step",
+            "eval::on-turn-completed",
+            "eval::sweep",
+        ]
+    );
+}
+
+#[test]
+fn schemas_are_typed_and_match_goldens() {
+    let mut failures = Vec::new();
+    for spec in catalog() {
+        support::assert_typed_schema(
+            &format!("{} request", spec.function_id),
+            &spec.request_schema,
+        );
+        support::assert_typed_schema(
+            &format!("{} response", spec.function_id),
+            &spec.response_schema,
+        );
+        let value = serde_json::json!({
+            "function_id": spec.function_id,
+            "request_schema": spec.request_schema,
+            "response_schema": spec.response_schema,
+        });
+        let mut actual = serde_json::to_string_pretty(&value).unwrap();
+        actual.push('\n');
+        if let Err(error) = support::check_golden(&golden_file_name(spec.function_id), &actual) {
+            failures.push(error);
+        }
+    }
+    assert!(failures.is_empty(), "{}", failures.join("\n"));
+}
diff --git a/eval/tests/support/mod.rs b/eval/tests/support/mod.rs
new file mode 100644
index 000000000..ecc657d4b
--- /dev/null
+++ b/eval/tests/support/mod.rs
@@ -0,0 +1,52 @@
+#![allow(dead_code)]
+
+use std::fs;
+use std::path::PathBuf;
+
+pub fn golden_root() -> PathBuf {
+    PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("tests/golden")
+}
+
+pub fn check_golden(rel: &str, actual: &str) -> Result<(), String> {
+    let path = golden_root().join(rel);
+    if std::env::var("UPDATE_GOLDENS").as_deref() == Ok("1") {
+        if let Some(parent) = path.parent() {
+            fs::create_dir_all(parent)
+                .map_err(|error| format!("create {}: {error}", parent.display()))?;
+        }
+        fs::write(&path, actual).map_err(|error| format!("write {}: {error}", path.display()))?;
+        return Ok(());
+    }
+    let expected = fs::read_to_string(&path).map_err(|error| {
+        format!(
+            "golden file {} unreadable ({error}); run UPDATE_GOLDENS=1 cargo test",
+            path.display()
+        )
+    })?;
+    if expected == actual {
+        Ok(())
+    } else {
+        Err(format!("golden mismatch: tests/golden/{rel}"))
+    }
+}
+
+pub fn assert_typed_schema(label: &str, schema: &schemars::schema::RootSchema) {
+    let value = serde_json::to_value(schema).expect("schema serializes");
+    let object = value
+        .as_object()
+        .unwrap_or_else(|| panic!("{label}: schema is not an object"));
+    const DEFINING: [&str; 8] = [
+        "type",
+        "properties",
+        "$ref",
+        "allOf",
+        "anyOf",
+        "oneOf",
+        "enum",
+        "items",
+    ];
+    assert!(
+        DEFINING.iter().any(|key| object.contains_key(*key)),
+        "{label}: schema is untyped: {value}"
+    );
+}
diff --git a/eval/ui/build.mjs b/eval/ui/build.mjs
new file mode 100644
index 000000000..9ff59ba86
--- /dev/null
+++ b/eval/ui/build.mjs
@@ -0,0 +1,24 @@
+import esbuild from 'esbuild'
+
+const options = {
+  entryPoints: ['page.tsx', 'styles.css'],
+  bundle: true,
+  format: 'esm',
+  jsx: 'automatic',
+  outdir: 'dist',
+  external: [
+    'react',
+    'react-dom',
+    'react-dom/client',
+    'react/jsx-runtime',
+    '@iii-dev/console-ui',
+  ],
+  logLevel: 'info',
+}
+
+if (process.argv.includes('--watch')) {
+  const context = await esbuild.context(options)
+  await context.watch()
+} else {
+  await esbuild.build(options)
+}
diff --git a/eval/ui/package.json b/eval/ui/package.json
new file mode 100644
index 000000000..088ccaa93
--- /dev/null
+++ b/eval/ui/package.json
@@ -0,0 +1,18 @@
+{
+  "name": "@iii-workers/eval-ui",
+  "private": true,
+  "version": "0.0.0",
+  "type": "module",
+  "scripts": {
+    "build": "tsc --noEmit && node build.mjs",
+    "watch": "node build.mjs --watch"
+  },
+  "dependencies": {
+    "@iii-dev/console-ui": "workspace:*"
+  },
+  "devDependencies": {
+    "@types/react": "^19.2.14",
+    "esbuild": "^0.25.0",
+    "typescript": "^5.9.2"
+  }
+}
diff --git a/eval/ui/page.tsx b/eval/ui/page.tsx
new file mode 100644
index 000000000..a9c9a5ac5
--- /dev/null
+++ b/eval/ui/page.tsx
@@ -0,0 +1,10 @@
+import type { Host } from '@iii-dev/console-ui'
+import { EvalPage } from './src/page'
+
+export default function setup(host: Host) {
+  host.pages.register({
+    id: 'eval-benchmarks',
+    title: 'eval',
+    render: () => <EvalPage host={host} />,
+  })
+}
diff --git a/eval/ui/src/api.ts b/eval/ui/src/api.ts
new file mode 100644
index 000000000..9b64c9f1f
--- /dev/null
+++ b/eval/ui/src/api.ts
@@ -0,0 +1,99 @@
+import type { Host } from '@iii-dev/console-ui'
+import type {
+  CatalogModel,
+  EvalRequest,
+  EvalResultResponse,
+  EvalStatus,
+  EvalStatusResponse,
+  EvalSummary,
+} from './types'
+
+const TIMEOUT_MS = 30_000
+
+export interface EvalApi {
+  list(): Promise<EvalSummary[]>
+  start(request: EvalRequest): Promise<{
+    evaluation_id: string
+    status: EvalStatus
+  }>
+  status(evaluationId: string): Promise<EvalStatusResponse | null>
+  result(evaluationId: string): Promise<EvalResultResponse | null>
+  rerun(evaluationId: string, reverseOrder: boolean): Promise<{
+    evaluation_id: string
+    status: EvalStatus
+  }>
+  cancel(evaluationId: string): Promise<{ cancelled: boolean; status: EvalStatus }>
+  delete(evaluationId: string): Promise<{ deleted: boolean }>
+  models(): Promise<CatalogModel[]>
+  systemPrompt(provider?: string): Promise<string | null>
+}
+
+export function createEvalApi(host: Host): EvalApi {
+  const trigger = <T>(functionId: string, payload: Record<string, unknown>) =>
+    host.iii.trigger<T>(functionId, payload, { timeoutMs: TIMEOUT_MS })
+
+  return {
+    async list() {
+      const response = await trigger<{ evaluations: EvalSummary[] }>(
+        'eval::list',
+        { limit: 50 },
+      )
+      return response.evaluations
+    },
+    start(request) {
+      return trigger('eval::start', request as unknown as Record<string, unknown>)
+    },
+    status(evaluationId) {
+      return trigger('eval::status', { evaluation_id: evaluationId })
+    },
+    result(evaluationId) {
+      return trigger('eval::result', { evaluation_id: evaluationId })
+    },
+    rerun(evaluationId, reverseOrder) {
+      return trigger('eval::rerun', {
+        evaluation_id: evaluationId,
+        reverse_order: reverseOrder,
+      })
+    },
+    cancel(evaluationId) {
+      return trigger('eval::cancel', { evaluation_id: evaluationId })
+    },
+    delete(evaluationId) {
+      return trigger('eval::delete', { evaluation_id: evaluationId })
+    },
+    async models() {
+      const response = await trigger<{ models: CatalogModel[] }>(
+        'router::models::list',
+        {},
+      )
+      return response.models
+    },
+    async systemPrompt(provider) {
+      const response = await trigger<{
+        system_prompt?: string | null
+      }>(
+        'router::system_prompt::get',
+        provider ? { provider } : {},
+      )
+      return response.system_prompt?.trim() || null
+    },
+  }
+}
+
+export function errorMessage(error: unknown): string {
+  if (error instanceof Error) return error.message
+  if (typeof error === 'string') return error
+  try {
+    return JSON.stringify(error)
+  } catch {
+    return 'unknown error'
+  }
+}
+
+export function isTerminal(status: EvalStatus): boolean {
+  return (
+    status === 'completed' ||
+    status === 'failed' ||
+    status === 'cancelled'
+  )
+}
diff --git a/eval/ui/src/components.tsx b/eval/ui/src/components.tsx
new file mode 100644
index 000000000..34320374d
--- /dev/null
+++ b/eval/ui/src/components.tsx
@@ -0,0 +1,72 @@
+import type { ReactNode, TextareaHTMLAttributes } from 'react'
+import { Badge } from '@iii-dev/console-ui'
+import type { EvalStatus } from './types'
+
+export function Field({
+  label,
+  hint,
+  error,
+  children,
+  className = '',
+}: {
+  label?: string
+  hint?: string
+  error?: string
+  children: ReactNode
+  className?: string
+}) {
+  return (
+    <label className={`eval-ui-field ${className}`}>
+      {label ? <span className="eval-ui-label">{label}</span> : null}
+      {children}
+      {error ? (
+        <span className="eval-ui-field-error">{error}</span>
+      ) : hint ? (
+        <span className="eval-ui-hint">{hint}</span>
+      ) : null}
+    </label>
+  )
+}
+
+export function TextArea({
+  className = '',
+  ...props
+}: TextareaHTMLAttributes<HTMLTextAreaElement>) {
+  return <textarea className={`eval-ui-textarea ${className}`} {...props} />
+}
+
+export function StatusBadge({ status }: { status: EvalStatus }) {
+  const variant =
+    status === 'completed'
+      ? 'accent'
+      : status === 'failed'
+        ? 'alert'
+        : status === 'cancelled'
+          ? 'warn'
+          : 'default'
+  return <Badge variant={variant}>{status}</Badge>
+}
+
+export function formatDate(timestamp?: number): string {
+  if (!timestamp) return '—'
+  return new Date(timestamp).toLocaleString()
+}
+
+export function formatMetric(
+  value: number | undefined,
+  kind: 'number' | 'percent' | 'duration' | 'cost' = 'number',
+): string {
+  if (value === undefined || !Number.isFinite(value)) return '—'
+  if (kind === 'percent') return `${(value * 100).toFixed(1)}%`
+  if (kind === 'duration') {
+    return Math.abs(value) >= 1000
+      ? `${(value / 1000).toFixed(2)}s`
+      : `${value.toFixed(0)}ms`
+  }
+  if (kind === 'cost') return `$${value.toFixed(6)}`
+  return new Intl.NumberFormat().format(value)
+}
+
+export function shortId(value: string): string {
+  return value.length > 18 ? `${value.slice(0, 10)}…${value.slice(-6)}` : value
+}
diff --git a/eval/ui/src/events.ts b/eval/ui/src/events.ts
new file mode 100644
index 000000000..6f470b3c3
--- /dev/null
+++ b/eval/ui/src/events.ts
@@ -0,0 +1,33 @@
+import { useEffect } from 'react'
+import type { Host } from '@iii-dev/console-ui'
+import type { EvalStatus } from './types'
+
+const COMPLETED_FN = 'iii::eval-ui::completed'
+
+export interface EvalCompletedEvent {
+  evaluation_id: string
+  status: EvalStatus
+  eligible?: boolean
+  timestamp: number
+}
+
+export function useEvalCompleted(
+  host: Host,
+  onCompleted: (event: EvalCompletedEvent) => void,
+) {
+  useEffect(() => {
+    const offHandler = host.iii.on<EvalCompletedEvent>(
+      COMPLETED_FN,
+      onCompleted,
+    )
+    const offTrigger = host.iii.registerTrigger({
+      type: 'eval::completed',
+      function_id: `${COMPLETED_FN}::${host.iii.browserId}`,
+      config: {},
+    })
+    return () => {
+      offTrigger()
+      offHandler()
+    }
+  }, [host, onCompleted])
+}
diff --git a/eval/ui/src/form.ts b/eval/ui/src/form.ts
new file mode 100644
index 000000000..d93ad36c9
--- /dev/null
+++ b/eval/ui/src/form.ts
@@ -0,0 +1,272 @@
+import type { ComparisonDimension, EvalRequest, JsonValue } from './types'
+
+export type SystemPromptSource = 'default' | 'custom' | 'none'
+
+export interface EvalFormState {
+  dimension: ComparisonDimension
+  model: string
+  provider: string
+  runs: string
+  controlLabel: string
+  treatmentLabel: string
+  controlPrompt: string
+  treatmentPrompt: string
+  sharedSystemPromptSource: SystemPromptSource
+  sharedSystemPrompt: string
+  sharedUserPrompt: string
+  controlSystemPromptSource: SystemPromptSource
+  controlSystemPrompt: string
+  treatmentSystemPromptSource: SystemPromptSource
+  treatmentSystemPrompt: string
+  evaluatorMode: 'normalized_text' | 'exact' | 'custom'
+  expectedFormat: 'text' | 'json'
+  expectedValue: string
+  evaluatorFunctionId: string
+  evaluatorArguments: string
+  maxTotalTokens: string
+  maxCostUsd: string
+  invocationTimeoutSeconds: string
+  scenarioTimeoutSeconds: string
+  maxTurns: string
+  maxOutputTokensPerCall: string
+  maxFunctionCallErrors: string
+  maxErrorSpans: string
+  systemPromptStrategy: 'override' | 'enrich'
+  mode: '' | 'ask' | 'agent'
+  thinkingLevel: '' | 'minimal' | 'low' | 'medium' | 'high' | 'xhigh'
+  functionsJson: string
+  outputJson: string
+  metadataJson: string
+  providerOptionsJson: string
+}
+
+export const DEFAULT_FORM: EvalFormState = {
+  dimension: 'prompt',
+  model: '',
+  provider: '',
+  runs: '1',
+  controlLabel: 'baseline',
+  treatmentLabel: 'candidate',
+  controlPrompt: '',
+  treatmentPrompt: '',
+  sharedSystemPromptSource: 'default',
+  sharedSystemPrompt: '',
+  sharedUserPrompt: '',
+  controlSystemPromptSource: 'custom',
+  controlSystemPrompt: '',
+  treatmentSystemPromptSource: 'custom',
+  treatmentSystemPrompt: '',
+  evaluatorMode: 'normalized_text',
+  expectedFormat: 'text',
+  expectedValue: '',
+  evaluatorFunctionId: '',
+  evaluatorArguments: '{}',
+  maxTotalTokens: '',
+  maxCostUsd: '',
+  invocationTimeoutSeconds: '',
+  scenarioTimeoutSeconds: '',
+  maxTurns: '',
+  maxOutputTokensPerCall: '',
+  maxFunctionCallErrors: '',
+  maxErrorSpans: '',
+  systemPromptStrategy: 'override',
+  mode: 'agent',
+  thinkingLevel: '',
+  functionsJson: '{"allow":["*"],"deny":["approval::*","configuration::*"],"expose":"agent_trigger"}',
+  outputJson: '{"type":"text"}',
+  metadataJson: '',
+  providerOptionsJson: '',
+}
+
+export type BuildResult =
+  | { request: EvalRequest; errors: Record<string, never> }
+  | { request: null; errors: Record<string, string> }
+
+export function buildRequest(form: EvalFormState): BuildResult {
+  const errors: Record<string, string> = {}
+  const model = form.model.trim()
+  const provider = form.provider.trim()
+  if (!model) errors.model = 'model is required'
+
+  const runs = integer(form.runs, 'runs', errors, 1)
+  if (runs !== null && runs > 20) errors.runs = 'runs must be at most 20'
+  const maxTotalTokens = optionalInteger(form.maxTotalTokens, 'maxTotalTokens', errors, 1)
+  const invocationTimeoutSeconds = optionalInteger(
+    form.invocationTimeoutSeconds,
+    'invocationTimeoutSeconds',
+    errors,
+    1,
+  )
+  const scenarioTimeoutSeconds = optionalInteger(form.scenarioTimeoutSeconds, 'scenarioTimeoutSeconds', errors, 1)
+  const maxTurns = optionalInteger(form.maxTurns, 'maxTurns', errors, 1)
+  const maxOutputTokensPerCall = optionalInteger(
+    form.maxOutputTokensPerCall,
+    'maxOutputTokensPerCall',
+    errors,
+    1,
+  )
+  const maxFunctionCallErrors = optionalInteger(
+    form.maxFunctionCallErrors,
+    'maxFunctionCallErrors',
+    errors,
+    0,
+  )
+  const maxErrorSpans = optionalInteger(form.maxErrorSpans, 'maxErrorSpans', errors, 0)
+  const maxCostUsd = optionalNumber(form.maxCostUsd, 'maxCostUsd', errors, 0)
+
+  if (form.dimension === 'prompt') {
+    if (!form.controlPrompt.trim()) errors.controlPrompt = 'baseline prompt is required'
+    if (!form.treatmentPrompt.trim()) errors.treatmentPrompt = 'candidate prompt is required'
+    if (form.sharedSystemPromptSource === 'custom' && !form.sharedSystemPrompt.trim()) {
+      errors.sharedSystemPrompt = 'custom system prompt is required'
+    }
+    if (form.controlPrompt.trim() && form.treatmentPrompt.trim() && form.controlPrompt === form.treatmentPrompt) {
+      errors.treatmentPrompt = 'candidate prompt must differ from baseline'
+    }
+  } else {
+    if (form.controlSystemPromptSource === 'custom' && !form.controlSystemPrompt.trim()) {
+      errors.controlSystemPrompt = 'baseline system prompt is required'
+    }
+    if (form.treatmentSystemPromptSource === 'custom' && !form.treatmentSystemPrompt.trim()) {
+      errors.treatmentSystemPrompt = 'candidate system prompt is required'
+    }
+    if (!form.sharedUserPrompt.trim()) errors.sharedUserPrompt = 'shared user prompt is required'
+    const controlSystemPrompt = systemPromptValue(form.controlSystemPromptSource, form.controlSystemPrompt)
+    const treatmentSystemPrompt = systemPromptValue(form.treatmentSystemPromptSource, form.treatmentSystemPrompt)
+    if (!errors.controlSystemPrompt && !errors.treatmentSystemPrompt && controlSystemPrompt === treatmentSystemPrompt) {
+      errors.treatmentSystemPrompt = 'candidate system prompt must differ from baseline'
+    }
+  }
+
+  const evaluatorEnabled =
+    form.evaluatorMode === 'custom' ? Boolean(form.evaluatorFunctionId.trim()) : Boolean(form.expectedValue.trim())
+  let evaluator: EvalRequest['evaluator']
+  if (evaluatorEnabled && form.evaluatorMode !== 'custom') {
+    const expected =
+      form.evaluatorMode === 'exact' && form.expectedFormat === 'json'
+        ? parseJson(form.expectedValue, 'expectedValue', errors)
+        : form.expectedValue
+    evaluator = {
+      function_id: form.evaluatorMode === 'normalized_text' ? 'eval::assert::normalized_text' : 'eval::assert::exact',
+      arguments: { expected: expected ?? null },
+    }
+  } else if (evaluatorEnabled) {
+    evaluator = {
+      function_id: form.evaluatorFunctionId.trim(),
+      arguments: parseJson(form.evaluatorArguments, 'evaluatorArguments', errors) ?? {},
+    }
+  }
+
+  const functions = optionalObject(form.functionsJson, 'functionsJson', errors)
+  const output = optionalObject(form.outputJson, 'outputJson', errors)
+  const metadata = optionalObject(form.metadataJson, 'metadataJson', errors)
+  const providerOptions = optionalObject(form.providerOptionsJson, 'providerOptionsJson', errors)
+
+  if (Object.keys(errors).length > 0) return { request: null, errors }
+
+  const request: EvalRequest = {
+    dimension: form.dimension,
+    model: {
+      model,
+      ...(provider ? { provider } : {}),
+      system_prompt_strategy: form.systemPromptStrategy,
+      ...(form.mode ? { mode: form.mode } : {}),
+      ...(form.thinkingLevel ? { thinking_level: form.thinkingLevel } : {}),
+      ...(providerOptions ? { provider_options: providerOptions } : {}),
+    },
+    control: {
+      ...(form.controlLabel.trim() ? { label: form.controlLabel.trim() } : {}),
+      prompt: form.dimension === 'prompt' ? form.controlPrompt : form.sharedUserPrompt,
+      system_prompt:
+        form.dimension === 'prompt'
+          ? systemPromptValue(form.sharedSystemPromptSource, form.sharedSystemPrompt)
+          : systemPromptValue(form.controlSystemPromptSource, form.controlSystemPrompt),
+    },
+    treatment: {
+      ...(form.treatmentLabel.trim() ? { label: form.treatmentLabel.trim() } : {}),
+      prompt: form.dimension === 'prompt' ? form.treatmentPrompt : form.sharedUserPrompt,
+      system_prompt:
+        form.dimension === 'prompt'
+          ? systemPromptValue(form.sharedSystemPromptSource, form.sharedSystemPrompt)
+          : systemPromptValue(form.treatmentSystemPromptSource, form.treatmentSystemPrompt),
+    },
+    ...(evaluator ? { evaluator } : {}),
+    runs: runs as number,
+    execution_order: 'balanced_control_first',
+    limits: {
+      execution: {
+        ...(invocationTimeoutSeconds === null
+          ? {}
+          : { invocation_timeout_seconds: invocationTimeoutSeconds }),
+        ...(scenarioTimeoutSeconds === null ? {} : { scenario_timeout_seconds: scenarioTimeoutSeconds }),
+        ...(maxTurns === null ? {} : { max_turns: maxTurns }),
+        ...(maxOutputTokensPerCall === null
+          ? {}
+          : { max_output_tokens_per_call: maxOutputTokensPerCall }),
+        ...(maxTotalTokens === null ? {} : { max_total_tokens: maxTotalTokens }),
+        ...(maxCostUsd === null ? {} : { max_cost_usd: maxCostUsd }),
+      },
+      evaluation: {
+        ...(maxFunctionCallErrors === null ? {} : { max_function_call_errors: maxFunctionCallErrors }),
+        ...(maxErrorSpans === null ? {} : { max_error_spans: maxErrorSpans }),
+      },
+    },
+    ...(functions ? { functions: functions as unknown as EvalRequest['functions'] } : {}),
+    ...(output ? { output: output as unknown as EvalRequest['output'] } : {}),
+    ...(metadata ? { metadata } : {}),
+  }
+  return { request, errors: {} }
+}
+
+function systemPromptValue(source: SystemPromptSource, customPrompt: string): string | null {
+  if (source === 'none') return null
+  if (source === 'default') return ''
+  return customPrompt
+}
+
+function integer(raw: string, field: string, errors: Record<string, string>, minimum: number): number | null {
+  const value = Number(raw)
+  if (!Number.isInteger(value) || value < minimum) {
+    errors[field] = `must be an integer of at least ${minimum}`
+    return null
+  }
+  return value
+}
+
+function optionalInteger(raw: string, field: string, errors: Record<string, string>, minimum: number): number | null {
+  if (!raw.trim()) return null
+  return integer(raw, field, errors, minimum)
+}
+
+function optionalNumber(raw: string, field: string, errors: Record<string, string>, minimum: number): number | null {
+  if (!raw.trim()) return null
+  const value = Number(raw)
+  if (!Number.isFinite(value) || value < minimum) {
+    errors[field] = `must be a number of at least ${minimum}`
+    return null
+  }
+  return value
+}
+
+function parseJson(raw: string, field: string, errors: Record<string, string>): JsonValue | null {
+  try {
+    return JSON.parse(raw) as JsonValue
+  } catch (error) {
+    errors[field] = error instanceof Error ? `invalid JSON: ${error.message}` : 'invalid JSON'
+    return null
+  }
+}
+
+function parseObject(raw: string, field: string, errors: Record<string, string>): Record<string, JsonValue> | null {
+  const value = parseJson(raw, field, errors)
+  if (value === null || Array.isArray(value) || typeof value !== 'object') {
+    if (!errors[field]) errors[field] = 'must be a JSON object'
+    return null
+  }
+  return value
+}
+
+function optionalObject(raw: string, field: string, errors: Record<string, string>): Record<string, JsonValue> | null {
+  if (!raw.trim()) return null
+  return parseObject(raw, field, errors)
+}
diff --git a/eval/ui/src/page/EvaluationDetail.tsx b/eval/ui/src/page/EvaluationDetail.tsx
new file mode 100644
index 000000000..66c7c6e08
--- /dev/null
+++ b/eval/ui/src/page/EvaluationDetail.tsx
@@ -0,0 +1,661 @@
+import { useEffect, useMemo, useState } from 'react'
+import { Badge, Button, JsonHighlight, Markdown, Skeleton, StatusPanel } from '@iii-dev/console-ui'
+import { formatDate, formatMetric, shortId, StatusBadge } from '../components'
+import { isTerminal } from '../api'
+import type {
+  AggregateDelta,
+  EvalResultResponse,
+  EvalRun,
+  EvalStatusResponse,
+  EvalSummary,
+  VariantAggregate,
+} from '../types'
+
+export function EvaluationDetail({
+  summary,
+  status,
+  result,
+  loading,
+  error,
+  actionPending,
+  onCancel,
+  onDelete,
+  onRerun,
+}: {
+  summary: EvalSummary
+  status: EvalStatusResponse | null
+  result: EvalResultResponse | null
+  loading: boolean
+  error: string | null
+  actionPending: boolean
+  onCancel: () => void
+  onDelete: () => void
+  onRerun: (reverseOrder: boolean) => void
+}) {
+  const currentStatus = status?.status ?? result?.status ?? summary.status
+  const progress = status ?? summary
+  const report = result?.report
+  const request = result?.request
+  const [copied, setCopied] = useState(false)
+
+  useEffect(() => {
+    if (!copied) return
+    const timer = window.setTimeout(() => setCopied(false), 1_200)
+    return () => window.clearTimeout(timer)
+  }, [copied])
+
+  const copyEvaluationId = async () => {
+    try {
+      await copyText(summary.evaluation_id)
+      setCopied(true)
+    } catch {}
+  }
+
+  return (
+    <div className="eval-ui-detail">
+      <div className="eval-ui-detail-head">
+        <div>
+          <div className="eval-ui-title-row">
+            <h1>
+              {summary.control_label ?? 'control'} vs {summary.treatment_label ?? 'treatment'}
+            </h1>
+            <StatusBadge status={currentStatus} />
+            {report?.eligible !== undefined ? (
+              <Badge variant={report.eligible ? 'accent' : 'warn'}>
+                {report.eligible ? 'candidate eligible' : 'candidate not eligible'}
+              </Badge>
+            ) : null}
+          </div>
+          <p>
+            {summary.model}
+            {summary.provider ? ` · ${summary.provider}` : ''} · {summary.dimension.replace('_', ' ')}
+          </p>
+          <div className="eval-ui-id">
+            <span className="eval-ui-label">evaluation id</span>
+            <code title={summary.evaluation_id}>{shortId(summary.evaluation_id)}</code>
+            <button
+              type="button"
+              onClick={() => void copyEvaluationId()}
+              aria-label={copied ? 'copied' : 'copy evaluation id'}
+              title={copied ? 'copied' : 'copy'}
+            >
+              {copied ? <CheckIcon /> : <CopyIcon />}
+            </button>
+            <span className="eval-ui-sr-only" aria-live="polite">
+              {copied ? 'Evaluation ID copied to clipboard.' : ''}
+            </span>
+          </div>
+        </div>
+        <div className="eval-ui-actions">
+          {!isTerminal(currentStatus) ? (
+            <Button variant="ghost" size="sm" onClick={onCancel} disabled={actionPending}>
+              cancel
+            </Button>
+          ) : (
+            <>
+              <Button variant="ghost" size="sm" onClick={() => onRerun(false)} disabled={actionPending}>
+                run again
+              </Button>
+              <Button variant="ghost" size="sm" onClick={() => onRerun(true)} disabled={actionPending}>
+                run reversed
+              </Button>
+              <Button variant="ghost" size="sm" onClick={onDelete} disabled={actionPending}>
+                delete
+              </Button>
+            </>
+          )}
+        </div>
+      </div>
+
+      <Progress
+        terminal={progress.terminal_runs}
+        total={progress.total_runs}
+        passed={progress.passed_runs}
+        failed={progress.failed_runs}
+      />
+
+      {error ? (
+        <StatusPanel variant="alert" headline="evaluation details unavailable" detail={error} />
+      ) : loading && !result ? (
+        <div className="eval-ui-loading">
+          <Skeleton />
+          <Skeleton />
+          <Skeleton />
+        </div>
+      ) : null}
+
+      {!isTerminal(currentStatus) ? (
+        <StatusPanel
+          variant="info"
+          headline={`evaluation ${currentStatus}`}
+          detail={
+            status?.active
+              ? `${status.active.role} run ${status.active.iteration} is active (${shortId(status.active.session_id)}).`
+              : 'waiting for the next durable evaluation step.'
+          }
+        />
+      ) : currentStatus !== 'completed' && !report ? (
+        <StatusPanel
+          variant={currentStatus === 'failed' ? 'alert' : 'warn'}
+          headline={`evaluation ${currentStatus}`}
+          detail={status?.error ?? summary.error ?? 'no report was produced.'}
+        />
+      ) : null}
+
+      {request ? <Definition result={result} /> : null}
+      {result ? <Report result={result} activeRunId={status?.active?.run_id} /> : null}
+    </div>
+  )
+}
+
+function Progress({
+  terminal,
+  total,
+  passed,
+  failed,
+}: {
+  terminal: number
+  total: number
+  passed: number
+  failed: number
+}) {
+  const percentage = total > 0 ? Math.min(100, (terminal / total) * 100) : 0
+  const notEvaluated = Math.max(0, terminal - passed - failed)
+  return (
+    <div className="eval-ui-progress">
+      <div className="eval-ui-progress-meta">
+        <span>
+          {terminal}/{total} runs complete
+        </span>
+        <span>
+          {passed} met criteria · {failed} missed criteria
+          {notEvaluated > 0 ? ` · ${notEvaluated} not evaluated` : ''}
+        </span>
+      </div>
+      <div className="eval-ui-progress-track">
+        <span style={{ width: `${percentage}%` }} />
+      </div>
+    </div>
+  )
+}
+
+function Definition({ result }: { result: EvalResultResponse }) {
+  const { request } = result
+  const varied = request.dimension === 'prompt' ? 'prompt' : 'system prompt'
+  const shared = request.dimension === 'prompt' ? request.control.system_prompt : request.control.prompt
+  return (
+    <section className="eval-ui-panel">
+      <div className="eval-ui-panel-title">definition</div>
+      <div className="eval-ui-definition-meta">
+        <span>{request.runs} runs per variant</span>
+        <span>{orderLabel(request.execution_order)}</span>
+        {request.source_evaluation_id ? <span>rerun of {shortId(request.source_evaluation_id)}</span> : null}
+        <span>{request.evaluator?.function_id ?? 'no success criteria'}</span>
+        <span>
+          {request.limits.execution.max_total_tokens === undefined
+            ? 'no total-token limit'
+            : `max ${formatMetric(request.limits.execution.max_total_tokens)} tokens`}
+        </span>
+        {request.limits.execution.max_cost_usd !== undefined ? (
+          <span>max {formatMetric(request.limits.execution.max_cost_usd, 'cost')}</span>
+        ) : null}
+      </div>
+      <div className="eval-ui-order">
+        <span className="eval-ui-label">execution sequence</span>
+        <span className="eval-ui-hint">Runs execute from left to right. A is the baseline; B is the candidate.</span>
+        <div className="eval-ui-order-list">
+          {result.progress.effective_execution_order.map((value, index) => {
+            const step = executionStep(
+              value,
+              request.control.label ?? 'baseline',
+              request.treatment.label ?? 'candidate',
+            )
+            return (
+              <div className="eval-ui-order-item-wrap" key={`${value}-${index}`}>
+                {index > 0 ? <span className="eval-ui-order-arrow" aria-hidden="true">→</span> : null}
+                <div className="eval-ui-order-item" title={value}>
+                  <strong>{step.marker}</strong>
+                  <span>{step.label}</span>
+                </div>
+              </div>
+            )
+          })}
+        </div>
+      </div>
+      <div className="eval-ui-shared">
+        <span>{request.dimension === 'prompt' ? 'shared system prompt' : 'shared prompt'}</span>
+        <pre>
+          {request.dimension === 'prompt' && shared === null
+            ? '— no system prompt —'
+            : shared || '— built-in system prompt —'}
+        </pre>
+      </div>
+      <div className="eval-ui-variants">
+        {[
+          { marker: 'A', role: 'baseline', variant: request.control },
+          { marker: 'B', role: 'candidate', variant: request.treatment },
+        ].map(({ marker, role, variant }) => {
+          return (
+            <div className="eval-ui-variant readonly" key={role}>
+              <div className="eval-ui-variant-title">
+                {marker} · {variant.label ?? role}
+              </div>
+              <span className="eval-ui-label">{varied}</span>
+              <pre>
+                {request.dimension === 'prompt'
+                  ? variant.prompt
+                  : variant.system_prompt === null
+                    ? '— no system prompt —'
+                    : variant.system_prompt || '— built-in system prompt —'}
+              </pre>
+            </div>
+          )
+        })}
+      </div>
+      {result.report ? <ArtifactHashes report={result.report} /> : null}
+    </section>
+  )
+}
+
+function Report({ result, activeRunId }: { result: EvalResultResponse; activeRunId?: string }) {
+  const report = result.report
+  const progress = result.progress
+  return (
+    <section className="eval-ui-panel">
+      <div className="eval-ui-panel-title">report</div>
+      <div className="eval-ui-report-section">
+        <div className="eval-ui-label">runs</div>
+        <Runs
+          runs={progress.runs}
+          controlLabel={result.request.control.label ?? 'baseline'}
+          treatmentLabel={result.request.treatment.label ?? 'candidate'}
+          activeRunId={activeRunId}
+          orderSensitive={progress.order_sensitivity.detected}
+        />
+      </div>
+      <div className="eval-ui-report-section">
+        <div className="eval-ui-label">summary</div>
+        <div className="eval-ui-report-note">A is the baseline. B is the candidate. Delta is B − A.</div>
+        <AggregateTable
+          control={progress.control_aggregate}
+          treatment={progress.treatment_aggregate}
+          delta={progress.delta}
+          controlLabel={result.request.control.label ?? 'control'}
+          treatmentLabel={result.request.treatment.label ?? 'treatment'}
+        />
+        {progress.order_sensitivity.detected ? (
+          <StatusPanel
+            variant="warn"
+            headline="order-sensitive result"
+            detail="At least one variant has a different pass rate when it runs first versus second."
+          />
+        ) : null}
+      </div>
+      {report ? (
+        <div className="eval-ui-report-foot">
+          completed {formatDate(report.completed_at)} · efficiency metrics are descriptive and do not select a winner.
+        </div>
+      ) : (
+        <div className="eval-ui-report-foot">Summary includes completed runs only and updates while the evaluation runs.</div>
+      )}
+    </section>
+  )
+}
+
+interface MetricRow {
+  label: string
+  control: keyof VariantAggregate
+  delta: keyof AggregateDelta
+  format?: 'number' | 'percent' | 'duration' | 'cost'
+  higherIsBetter?: boolean
+}
+
+const METRICS: MetricRow[] = [
+  {
+    label: 'pass rate',
+    control: 'pass_rate',
+    delta: 'pass_rate',
+    format: 'percent',
+    higherIsBetter: true,
+  },
+  {
+    label: 'median score',
+    control: 'median_score',
+    delta: 'median_score',
+    higherIsBetter: true,
+  },
+  {
+    label: 'wall time',
+    control: 'median_wall_time_ms',
+    delta: 'median_wall_time_ms',
+    format: 'duration',
+  },
+  {
+    label: 'total tokens',
+    control: 'median_total_tokens',
+    delta: 'median_total_tokens',
+  },
+  {
+    label: 'reasoning tokens',
+    control: 'median_reasoning_tokens',
+    delta: 'median_reasoning_tokens',
+  },
+  {
+    label: 'cost',
+    control: 'median_cost_usd',
+    delta: 'median_cost_usd',
+    format: 'cost',
+  },
+  {
+    label: 'function calls',
+    control: 'median_function_calls',
+    delta: 'median_function_calls',
+  },
+  {
+    label: 'function-call errors',
+    control: 'median_function_call_errors',
+    delta: 'median_function_call_errors',
+  },
+  {
+    label: 'traces',
+    control: 'median_trace_count',
+    delta: 'median_trace_count',
+  },
+  {
+    label: 'spans',
+    control: 'median_span_count',
+    delta: 'median_span_count',
+  },
+  {
+    label: 'error spans',
+    control: 'median_error_span_count',
+    delta: 'median_error_span_count',
+  },
+  {
+    label: 'trace duration',
+    control: 'median_trace_duration_ms',
+    delta: 'median_trace_duration_ms',
+    format: 'duration',
+  },
+]
+
+function AggregateTable({
+  control,
+  treatment,
+  delta,
+  controlLabel,
+  treatmentLabel,
+}: {
+  control: VariantAggregate
+  treatment: VariantAggregate
+  delta: AggregateDelta
+  controlLabel: string
+  treatmentLabel: string
+}) {
+  return (
+    <div className="eval-ui-metrics">
+      <div className="eval-ui-success-counts">
+        <span>A · {control.passed}/{control.evaluated_runs} passed · {control.runs} completed</span>
+        <span>B · {treatment.passed}/{treatment.evaluated_runs} passed · {treatment.runs} completed</span>
+      </div>
+      <div className="eval-ui-metric header">
+        <span>metric</span>
+        <span>A · {controlLabel}</span>
+        <span>B · {treatmentLabel}</span>
+        <span>B − A</span>
+      </div>
+      {METRICS.map((metric) => {
+        const controlValue =
+          metric.control === 'pass_rate' && control.evaluated_runs === 0
+            ? undefined
+            : (control[metric.control] as number | undefined)
+        const treatmentValue =
+          metric.control === 'pass_rate' && treatment.evaluated_runs === 0
+            ? undefined
+            : (treatment[metric.control] as number | undefined)
+        const deltaValue =
+          metric.control === 'pass_rate' && (control.evaluated_runs === 0 || treatment.evaluated_runs === 0)
+            ? undefined
+            : (delta[metric.delta] as number | undefined)
+        const positive =
+          deltaValue !== undefined && deltaValue !== 0 && (metric.higherIsBetter ? deltaValue > 0 : deltaValue < 0)
+        const negative =
+          deltaValue !== undefined && deltaValue !== 0 && (metric.higherIsBetter ? deltaValue < 0 : deltaValue > 0)
+        return (
+          <div className="eval-ui-metric" key={metric.label}>
+            <span>{metric.label}</span>
+            <span>{formatMetric(controlValue, metric.format)}</span>
+            <span>{formatMetric(treatmentValue, metric.format)}</span>
+            <span className={positive ? 'positive' : negative ? 'negative' : undefined}>
+              {deltaValue !== undefined && deltaValue > 0 ? '+' : ''}
+              {formatMetric(deltaValue, metric.format)}
+            </span>
+          </div>
+        )
+      })}
+    </div>
+  )
+}
+
+function Runs({
+  runs,
+  controlLabel,
+  treatmentLabel,
+  activeRunId,
+  orderSensitive,
+}: {
+  runs: EvalRun[]
+  controlLabel: string
+  treatmentLabel: string
+  activeRunId?: string
+  orderSensitive: boolean
+}) {
+  const [selectedId, setSelectedId] = useState<string | null>(runs[0]?.run_id ?? null)
+  useEffect(() => {
+    if (!runs.some((run) => run.run_id === selectedId)) {
+      setSelectedId(runs[0]?.run_id ?? null)
+    }
+  }, [runs, selectedId])
+  const selected = useMemo(() => runs.find((run) => run.run_id === selectedId) ?? null, [runs, selectedId])
+  const iterations = useMemo(() => {
+    const values = new Map<number, { control?: EvalRun; treatment?: EvalRun }>()
+    for (const run of runs) {
+      const pair = values.get(run.iteration) ?? {}
+      pair[run.role] = run
+      values.set(run.iteration, pair)
+    }
+    return [...values.entries()].sort(([left], [right]) => left - right)
+  }, [runs])
+  return (
+    <div className="eval-ui-runs">
+      <div className="eval-ui-run-pairs">
+        {iterations.map(([iteration, pair]) => (
+          <div className={`eval-ui-run-pair${orderSensitive ? ' order-sensitive' : ''}`} key={iteration}>
+            <div className="eval-ui-run-pair-head">iteration {iteration}</div>
+            <div className="eval-ui-run-pair-grid">
+              {pair.control ? (
+                <RunCard
+                  run={pair.control}
+                  marker="A"
+                  label={controlLabel}
+                  active={pair.control.run_id === activeRunId}
+                  selected={pair.control.run_id === selectedId}
+                  onSelect={setSelectedId}
+                />
+              ) : null}
+              {pair.treatment ? (
+                <RunCard
+                  run={pair.treatment}
+                  marker="B"
+                  label={treatmentLabel}
+                  active={pair.treatment.run_id === activeRunId}
+                  selected={pair.treatment.run_id === selectedId}
+                  onSelect={setSelectedId}
+                />
+              ) : null}
+            </div>
+          </div>
+        ))}
+      </div>
+      {selected ? <RunDetail run={selected} /> : null}
+    </div>
+  )
+}
+
+function RunCard({
+  run,
+  marker,
+  label,
+  active,
+  selected,
+  onSelect,
+}: {
+  run: EvalRun
+  marker: string
+  label: string
+  active: boolean
+  selected: boolean
+  onSelect: (id: string) => void
+}) {
+  const [now, setNow] = useState(Date.now())
+  useEffect(() => {
+    if (!active) return
+    const timer = window.setInterval(() => setNow(Date.now()), 1_000)
+    return () => window.clearInterval(timer)
+  }, [active])
+  const elapsed = active ? Math.max(0, now - run.started_at) : run.benchmark?.wall_time_ms
+  return (
+    <button
+      type="button"
+      className={`eval-ui-run-card${active ? ' running' : ''}${selected ? ' selected' : ''}`}
+      onClick={() => onSelect(run.run_id)}
+    >
+      <span className="eval-ui-run-card-title">{marker} · {label}</span>
+      <span>{runStatus(run)}</span>
+      <span>{run.pair_position === 1 ? 'ran first' : 'ran second'} · #{run.execution_position}</span>
+      <span>{formatMetric(run.benchmark?.total_tokens)} tokens · {formatMetric(elapsed, 'duration')}</span>
+    </button>
+  )
+}
+
+function runStatus(run: EvalRun): string {
+  if (run.passed) return 'met criteria'
+  if (run.passed === false) return 'missed criteria'
+  if (run.status === 'completed') return 'not evaluated'
+  return run.status
+}
+
+function RunDetail({ run }: { run: EvalRun }) {
+  const failures = run.failures ?? []
+  return (
+    <div className="eval-ui-run-detail">
+      <div className="eval-ui-definition-meta">
+        <span>{run.session_id}</span>
+        <span>{run.evaluation?.reason ?? 'not evaluated'}</span>
+      </div>
+      {failures.length > 0 ? (
+        <div className="eval-ui-failures">
+          {failures.map((failure, index) => (
+            <div key={`${failure.phase}-${index}`}>
+              <strong>{failure.phase}</strong>
+              {failure.function_id ? ` · ${failure.function_id}` : ''} — {failure.message}
+            </div>
+          ))}
+        </div>
+      ) : null}
+      <div className="eval-ui-label">output</div>
+      <Markdown className="eval-ui-markdown-output">{outputAsMarkdown(run.output)}</Markdown>
+      {run.evaluation?.details !== undefined ? (
+        <>
+          <div className="eval-ui-label">evaluator details</div>
+          <JsonHighlight code={JSON.stringify(run.evaluation.details, null, 2)} wrap />
+        </>
+      ) : null}
+    </div>
+  )
+}
+
+function orderLabel(order: EvalResultResponse['request']['execution_order']): string {
+  return order === 'balanced_treatment_first' ? 'balanced · B first' : 'balanced · A first'
+}
+
+function executionStep(
+  value: string,
+  controlLabel: string,
+  treatmentLabel: string,
+): { marker: string; label: string } {
+  const [, role] = value.split(':')
+  if (role === 'control') {
+    return { marker: 'A', label: controlLabel }
+  }
+  if (role === 'treatment') {
+    return { marker: 'B', label: treatmentLabel }
+  }
+  return { marker: '?', label: role || value }
+}
+
+function CopyIcon() {
+  return (
+    <svg width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="2" aria-hidden>
+      <rect x="9" y="9" width="13" height="13" rx="2" ry="2" />
+      <path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1" />
+    </svg>
+  )
+}
+
+function CheckIcon() {
+  return (
+    <svg width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="2" aria-hidden>
+      <path d="m20 6-11 11-5-5" />
+    </svg>
+  )
+}
+
+async function copyText(value: string): Promise<void> {
+  if (navigator.clipboard?.writeText) {
+    await navigator.clipboard.writeText(value)
+    return
+  }
+  const input = document.createElement('textarea')
+  input.value = value
+  input.style.position = 'fixed'
+  input.style.opacity = '0'
+  document.body.appendChild(input)
+  input.select()
+  const copied = document.execCommand('copy')
+  input.remove()
+  if (!copied) throw new Error('copy command failed')
+}
+
+function ArtifactHashes({ report }: { report: NonNullable<EvalResultResponse['report']> }) {
+  const values = [
+    ['A prompt', report.control.prompt_sha256],
+    ['A system', report.control.system_prompt_sha256 ?? 'none'],
+    ['B prompt', report.treatment.prompt_sha256],
+    ['B system', report.treatment.system_prompt_sha256 ?? 'none'],
+    ['model', report.shared_artifacts.model_sha256],
+    ['functions', report.shared_artifacts.function_policy_sha256],
+    ['limits', report.shared_artifacts.limits_sha256],
+    ['output', report.shared_artifacts.output_sha256],
+    ['evaluator', report.evaluator?.arguments_sha256 ?? 'none'],
+  ]
+  return (
+    <details className="eval-ui-hashes">
+      <summary>comparison hashes</summary>
+      <div>
+        {values.map(([label, hash]) => (
+          <code key={label}>
+            <span>{label}</span>
+            {hash}
+          </code>
+        ))}
+      </div>
+    </details>
+  )
+}
+
+function outputAsMarkdown(output: EvalRun['output']): string {
+  if (typeof output === 'string') return output
+
+  return `\`\`\`json\n${JSON.stringify(output ?? null, null, 2)}\n\`\`\``
+}
diff --git a/eval/ui/src/page/History.tsx b/eval/ui/src/page/History.tsx
new file mode 100644
index 000000000..af765182e
--- /dev/null
+++ b/eval/ui/src/page/History.tsx
@@ -0,0 +1,76 @@
+import { Button, EmptyState } from '@iii-dev/console-ui'
+import { formatDate, shortId, StatusBadge } from '../components'
+import type { EvalSummary } from '../types'
+
+export function History({
+  evaluations,
+  selectedId,
+  loading,
+  onSelect,
+  onNew,
+  onRefresh,
+}: {
+  evaluations: EvalSummary[]
+  selectedId: string | null
+  loading: boolean
+  onSelect: (evaluationId: string) => void
+  onNew: () => void
+  onRefresh: () => void
+}) {
+  return (
+    <aside className="eval-ui-history">
+      <div className="eval-ui-history-head">
+        <span>recent evaluations</span>
+        <Button variant="pill" size="sm" onClick={onRefresh} disabled={loading}>
+          refresh
+        </Button>
+      </div>
+      <Button variant="primary" size="sm" onClick={onNew}>
+        new evaluation
+      </Button>
+      <div className="eval-ui-history-list">
+        {evaluations.length === 0 && !loading ? (
+          <EmptyState
+            title="no evaluations yet"
+            description="compare a baseline and candidate prompt to create the first report."
+            action={{ label: 'new evaluation', onClick: onNew }}
+          />
+        ) : (
+          evaluations.map((evaluation) => {
+            const control = evaluation.control_label ?? 'control'
+            const treatment = evaluation.treatment_label ?? 'treatment'
+            const progress =
+              evaluation.total_runs > 0
+                ? `${evaluation.terminal_runs}/${evaluation.total_runs}`
+                : '—'
+            return (
+              <button
+                key={evaluation.evaluation_id}
+                type="button"
+                className={`eval-ui-history-row${
+                  selectedId === evaluation.evaluation_id ? ' active' : ''
+                }`}
+                onClick={() => onSelect(evaluation.evaluation_id)}
+              >
+                <span className="eval-ui-history-row-top">
+                  <StatusBadge status={evaluation.status} />
+                  <span className="eval-ui-history-progress">{progress}</span>
+                </span>
+                <span className="eval-ui-history-title">
+                  {control} vs {treatment}
+                </span>
+                <span className="eval-ui-history-meta">
+                  {evaluation.model} · {evaluation.dimension.replace('_', ' ')}
+                </span>
+                <span className="eval-ui-history-meta">
+                  {formatDate(evaluation.created_at)} ·{' '}
+                  {shortId(evaluation.evaluation_id)}
+                </span>
+              </button>
+            )
+          })
+        )}
+      </div>
+    </aside>
+  )
+}
diff --git a/eval/ui/src/page/NewEvaluationForm.tsx b/eval/ui/src/page/NewEvaluationForm.tsx
new file mode 100644
index 000000000..3c93d8ff1
--- /dev/null
+++ b/eval/ui/src/page/NewEvaluationForm.tsx
@@ -0,0 +1,813 @@
+import { useEffect, useMemo, useRef, useState } from 'react'
+import { Button, Input, Select, StatusPanel, Tabs, TabsList, TabsTrigger } from '@iii-dev/console-ui'
+import { errorMessage } from '../api'
+import { Field, TextArea } from '../components'
+import { buildRequest, DEFAULT_FORM, type EvalFormState, type SystemPromptSource } from '../form'
+import type { CatalogModel, EvalRequest } from '../types'
+
+const MODEL_SEPARATOR = '\u0000'
+const SYSTEM_PROMPT_SOURCE_OPTIONS = [
+  { value: 'default', label: 'current default' },
+  { value: 'custom', label: 'custom' },
+  { value: 'none', label: 'none' },
+]
+
+export function NewEvaluationForm({
+  models,
+  modelCatalogError,
+  onLoadDefaultSystemPrompt,
+  onCreate,
+}: {
+  models: CatalogModel[]
+  modelCatalogError: string | null
+  onLoadDefaultSystemPrompt: (provider?: string) => Promise<string | null>
+  onCreate: (request: EvalRequest) => Promise<void>
+}) {
+  const formRef = useRef<HTMLFormElement>(null)
+  const [form, setForm] = useState<EvalFormState>({ ...DEFAULT_FORM })
+  const [errors, setErrors] = useState<Record<string, string>>({})
+  const [submitting, setSubmitting] = useState(false)
+  const [submitError, setSubmitError] = useState<string | null>(null)
+  const [manualModel, setManualModel] = useState(models.length === 0)
+  const [loadingDefaultFor, setLoadingDefaultFor] = useState<
+    'sharedSystemPrompt' | 'controlSystemPrompt' | 'treatmentSystemPrompt' | null
+  >(null)
+  const [defaultPromptError, setDefaultPromptError] = useState<string | null>(null)
+
+  useEffect(() => {
+    if (models.length > 0 && !form.model) setManualModel(false)
+  }, [form.model, models.length])
+
+  const groups = useMemo(() => {
+    const providers = new Map<string, Array<{ value: string; label: string; title?: string }>>()
+    for (const model of models) {
+      const options = providers.get(model.provider) ?? []
+      options.push({
+        value: `${model.provider}${MODEL_SEPARATOR}${model.id}`,
+        label: model.display_name ?? model.id,
+        title: `${model.id} · context ${model.context_window}`,
+      })
+      providers.set(model.provider, options)
+    }
+    return [...providers.entries()]
+      .sort(([left], [right]) => left.localeCompare(right))
+      .map(([label, options]) => ({
+        label,
+        options: options.sort((left, right) => left.label.localeCompare(right.label)),
+      }))
+  }, [models])
+
+  const selectedModel = form.model ? `${form.provider}${MODEL_SEPARATOR}${form.model}` : undefined
+  const update = <K extends keyof EvalFormState>(key: K, value: EvalFormState[K]) => {
+    setForm((current) => ({ ...current, [key]: value }))
+    if (key === 'model' || key === 'provider') setDefaultPromptError(null)
+    setErrors((current) => {
+      const relatedKey =
+        key === 'sharedSystemPromptSource'
+          ? 'sharedSystemPrompt'
+          : key === 'controlSystemPromptSource'
+            ? 'controlSystemPrompt'
+            : key === 'treatmentSystemPromptSource'
+              ? 'treatmentSystemPrompt'
+              : undefined
+      if (!current[key] && (!relatedKey || !current[relatedKey])) return current
+      const next = { ...current }
+      delete next[key]
+      if (relatedKey) delete next[relatedKey]
+      return next
+    })
+  }
+
+  const submit = async (event: React.FormEvent) => {
+    event.preventDefault()
+    const built = buildRequest(form)
+    if (!built.request) {
+      setErrors(built.errors)
+      window.requestAnimationFrame(() => {
+        const firstError = formRef.current?.querySelector<HTMLElement>('.eval-ui-field-error')
+        if (!firstError) return
+        const details = firstError.closest('details')
+        if (details) details.open = true
+        window.requestAnimationFrame(() => {
+          firstError.scrollIntoView({ behavior: 'smooth', block: 'center' })
+          firstError
+            .closest('label')
+            ?.querySelector<HTMLElement>('input, textarea, button, [tabindex]:not([tabindex="-1"])')
+            ?.focus()
+        })
+      })
+      return
+    }
+    setSubmitting(true)
+    setSubmitError(null)
+    try {
+      await onCreate(built.request)
+    } catch (error) {
+      setSubmitError(errorMessage(error))
+    } finally {
+      setSubmitting(false)
+    }
+  }
+
+  const loadDefaultSystemPrompt = async (
+    target: 'sharedSystemPrompt' | 'controlSystemPrompt' | 'treatmentSystemPrompt',
+  ) => {
+    if (form.systemPromptStrategy !== 'override') return
+    const provider = form.provider.trim()
+    if (!provider) {
+      setDefaultPromptError('Select a catalog model or enter its provider before loading the current default.')
+      return
+    }
+    setLoadingDefaultFor(target)
+    setDefaultPromptError(null)
+    try {
+      const prompt = await onLoadDefaultSystemPrompt(provider)
+      if (!prompt) {
+        setDefaultPromptError(
+          'No provider prompt is exposed by the router. Choose “current default” to use the prompt resolved by the harness.',
+        )
+        return
+      }
+      update(target, prompt)
+    } catch (error) {
+      setDefaultPromptError(errorMessage(error))
+    } finally {
+      setLoadingDefaultFor(null)
+    }
+  }
+
+  const comparison =
+    form.dimension === 'prompt'
+      ? {
+          title: 'prompt comparison',
+          changedLabel: 'prompt',
+          sharedLabel: 'system prompt',
+          sharedHint: 'The same system prompt is used for every A and B run. Leave empty to use the harness default.',
+          sharedPlaceholder: 'You are a precise assistant…',
+          baselineHint: 'The current prompt you want to measure.',
+          candidateHint: 'The proposed prompt you want to compare.',
+        }
+      : {
+          title: 'system prompt comparison',
+          changedLabel: 'system prompt',
+          sharedLabel: 'user prompt',
+          sharedHint: 'The same user prompt is used for every A and B run.',
+          sharedPlaceholder: 'Build a concise summary of…',
+          baselineHint: 'The current system prompt you want to measure.',
+          candidateHint: 'The proposed system prompt you want to compare.',
+        }
+  const usesCustomSystemPrompt =
+    form.dimension === 'prompt'
+      ? form.sharedSystemPromptSource === 'custom'
+      : form.controlSystemPromptSource === 'custom' || form.treatmentSystemPromptSource === 'custom'
+  const sharedSystemPromptSummary =
+    form.sharedSystemPromptSource === 'default'
+      ? 'using current default'
+      : form.sharedSystemPromptSource === 'none'
+        ? 'disabled'
+        : `custom · ${form.systemPromptStrategy}`
+
+  return (
+    <form ref={formRef} className="eval-ui-form" onSubmit={submit}>
+      <div className="eval-ui-section-head">
+        <div>
+          <h1>new {comparison.title}</h1>
+          <p>Change one thing. The model, success criteria, and run settings stay identical.</p>
+        </div>
+      </div>
+
+      {submitError ? <StatusPanel variant="alert" headline="evaluation could not start" detail={submitError} /> : null}
+      {Object.keys(errors).length > 0 ? (
+        <div className="eval-ui-error-summary" role="alert">
+          <strong>
+            Fix {Object.keys(errors).length} {Object.keys(errors).length === 1 ? 'field' : 'fields'} to run this
+            comparison.
+          </strong>
+          <span>{[...new Set(Object.values(errors))].join(' · ')}</span>
+        </div>
+      ) : null}
+
+      <section className="eval-ui-panel">
+        <div className="eval-ui-step-head">
+          <span className="eval-ui-step-number">1</span>
+          <div>
+            <div className="eval-ui-panel-title">choose what changes</div>
+            <p>Only the selected input differs between A and B.</p>
+          </div>
+        </div>
+        <Tabs
+          value={form.dimension}
+          onValueChange={(value) => {
+            update('dimension', value as EvalFormState['dimension'])
+            setErrors({})
+          }}
+        >
+          <TabsList>
+            <TabsTrigger value="prompt">prompt text</TabsTrigger>
+            <TabsTrigger value="system_prompt">system prompt</TabsTrigger>
+          </TabsList>
+        </Tabs>
+        <div className="eval-ui-comparison-rule">
+          <strong>{comparison.changedLabel}</strong> changes between A and B<span aria-hidden="true">·</span>
+          <strong>{comparison.sharedLabel}</strong> stays the same
+        </div>
+
+        <div className="eval-ui-variants">
+          <VariantEditor
+            marker="A"
+            role="baseline"
+            label={form.controlLabel}
+            value={form.dimension === 'prompt' ? form.controlPrompt : form.controlSystemPrompt}
+            valueLabel={comparison.changedLabel}
+            hint={comparison.baselineHint}
+            source={form.dimension === 'system_prompt' ? form.controlSystemPromptSource : undefined}
+            valueError={form.dimension === 'prompt' ? errors.controlPrompt : errors.controlSystemPrompt}
+            onLabel={(value) => update('controlLabel', value)}
+            onValue={(value) =>
+              form.dimension === 'prompt' ? update('controlPrompt', value) : update('controlSystemPrompt', value)
+            }
+            onSource={(source) => update('controlSystemPromptSource', source)}
+            onLoadDefault={
+              form.dimension === 'system_prompt' &&
+              form.controlSystemPromptSource === 'custom' &&
+              form.systemPromptStrategy === 'override'
+                ? () => void loadDefaultSystemPrompt('controlSystemPrompt')
+                : undefined
+            }
+            loadingDefault={loadingDefaultFor === 'controlSystemPrompt'}
+            defaultAvailable={Boolean(form.provider.trim())}
+          />
+          <VariantEditor
+            marker="B"
+            role="candidate"
+            label={form.treatmentLabel}
+            value={form.dimension === 'prompt' ? form.treatmentPrompt : form.treatmentSystemPrompt}
+            valueLabel={comparison.changedLabel}
+            hint={comparison.candidateHint}
+            source={form.dimension === 'system_prompt' ? form.treatmentSystemPromptSource : undefined}
+            valueError={form.dimension === 'prompt' ? errors.treatmentPrompt : errors.treatmentSystemPrompt}
+            onLabel={(value) => update('treatmentLabel', value)}
+            onValue={(value) =>
+              form.dimension === 'prompt' ? update('treatmentPrompt', value) : update('treatmentSystemPrompt', value)
+            }
+            onSource={(source) => update('treatmentSystemPromptSource', source)}
+            onLoadDefault={
+              form.dimension === 'system_prompt' &&
+              form.treatmentSystemPromptSource === 'custom' &&
+              form.systemPromptStrategy === 'override'
+                ? () => void loadDefaultSystemPrompt('treatmentSystemPrompt')
+                : undefined
+            }
+            loadingDefault={loadingDefaultFor === 'treatmentSystemPrompt'}
+            defaultAvailable={Boolean(form.provider.trim())}
+          />
+        </div>
+
+        {form.dimension === 'prompt' ? (
+          <details className="eval-ui-secondary-input">
+            <summary>
+              <span>system prompt</span>
+              <span>optional · {sharedSystemPromptSummary}</span>
+            </summary>
+            <div className="eval-ui-secondary-input-body">
+              <Field label="system prompt source">
+                <Select
+                  value={form.sharedSystemPromptSource}
+                  options={SYSTEM_PROMPT_SOURCE_OPTIONS}
+                  onChange={(value) => update('sharedSystemPromptSource', value as SystemPromptSource)}
+                />
+              </Field>
+              {form.sharedSystemPromptSource === 'custom' ? (
+                <>
+                  <Field
+                    label="apply custom system prompt as"
+                    hint={
+                      form.systemPromptStrategy === 'override'
+                        ? 'Replaces the harness default.'
+                        : 'Appends this text to the harness default without loading it here.'
+                    }
+                  >
+                    <Select
+                      value={form.systemPromptStrategy}
+                      options={[
+                        { value: 'override', label: 'override' },
+                        { value: 'enrich', label: 'enrich' },
+                      ]}
+                      onChange={(value) =>
+                        update('systemPromptStrategy', value as EvalFormState['systemPromptStrategy'])
+                      }
+                    />
+                  </Field>
+                  {form.systemPromptStrategy === 'override' ? (
+                    <div className="eval-ui-load-default">
+                      <span className="eval-ui-label">starting point</span>
+                      <Button
+                        variant="ghost"
+                        size="sm"
+                        type="button"
+                        disabled={loadingDefaultFor !== null || !form.provider.trim()}
+                        onClick={() => void loadDefaultSystemPrompt('sharedSystemPrompt')}
+                      >
+                        {loadingDefaultFor === 'sharedSystemPrompt'
+                          ? 'loading…'
+                          : form.provider.trim()
+                            ? 'load current default'
+                            : 'select model to load default'}
+                      </Button>
+                      <span className="eval-ui-hint">Uses the selected provider, or the router default.</span>
+                    </div>
+                  ) : null}
+                  <Field label="custom system prompt" error={errors.sharedSystemPrompt} hint={comparison.sharedHint}>
+                    <TextArea
+                      value={form.sharedSystemPrompt}
+                      onChange={(event) => update('sharedSystemPrompt', event.target.value)}
+                      rows={6}
+                      placeholder={comparison.sharedPlaceholder}
+                    />
+                  </Field>
+                </>
+              ) : (
+                <SystemPromptSourceState source={form.sharedSystemPromptSource} />
+              )}
+            </div>
+          </details>
+        ) : (
+          <>
+            {usesCustomSystemPrompt ? (
+              <div className="eval-ui-system-prompt-mode">
+                <Field
+                  label="apply custom system prompts as"
+                  hint={
+                    form.systemPromptStrategy === 'override'
+                      ? 'Custom values replace the harness default.'
+                      : 'Custom values are appended to the harness default without loading it into either field.'
+                  }
+                >
+                  <Select
+                    value={form.systemPromptStrategy}
+                    options={[
+                      { value: 'override', label: 'override' },
+                      { value: 'enrich', label: 'enrich' },
+                    ]}
+                    onChange={(value) => update('systemPromptStrategy', value as EvalFormState['systemPromptStrategy'])}
+                  />
+                </Field>
+              </div>
+            ) : null}
+            <div className="eval-ui-shared-editor">
+              <div className="eval-ui-shared-editor-head">
+                <span>shared by A + B</span>
+                <strong>{comparison.sharedLabel}</strong>
+              </div>
+              <Field error={errors.sharedUserPrompt} hint={comparison.sharedHint}>
+                <TextArea
+                  value={form.sharedUserPrompt}
+                  onChange={(event) => update('sharedUserPrompt', event.target.value)}
+                  rows={4}
+                  placeholder={comparison.sharedPlaceholder}
+                />
+              </Field>
+            </div>
+          </>
+        )}
+
+        {defaultPromptError ? (
+          <StatusPanel variant="warn" headline="default system prompt unavailable" detail={defaultPromptError} />
+        ) : null}
+      </section>
+
+      <section className="eval-ui-panel">
+        <div className="eval-ui-step-head">
+          <span className="eval-ui-step-number">2</span>
+          <div>
+            <div className="eval-ui-panel-title">shared run settings</div>
+            <p>The same model and limits are applied to A and B.</p>
+          </div>
+        </div>
+        {models.length > 0 && !manualModel ? (
+          <Field label="model" error={errors.model}>
+            <Select
+              value={selectedModel}
+              groups={groups}
+              placeholder="select a catalog model"
+              onChange={(value) => {
+                const [provider, model] = value.split(MODEL_SEPARATOR)
+                setForm((current) => ({ ...current, provider, model }))
+                setDefaultPromptError(null)
+                setErrors((current) => {
+                  const next = { ...current }
+                  delete next.model
+                  return next
+                })
+              }}
+            />
+          </Field>
+        ) : (
+          <div className="eval-ui-grid-2">
+            <Field label="model" error={errors.model}>
+              <Input
+                value={form.model}
+                onChange={(value) => update('model', value)}
+                preserveCase
+                placeholder="codex/gpt-5.6-luna"
+              />
+            </Field>
+            <Field label="provider" hint="optional when model routing is unambiguous">
+              <Input
+                value={form.provider}
+                onChange={(value) => update('provider', value)}
+                preserveCase
+                placeholder="openai-codex"
+              />
+            </Field>
+          </div>
+        )}
+        <div className="eval-ui-inline-action">
+          {models.length > 0 ? (
+            <button type="button" className="eval-ui-link" onClick={() => setManualModel((current) => !current)}>
+              {manualModel ? 'use model catalog' : 'enter model manually'}
+            </button>
+          ) : modelCatalogError ? (
+            <span className="eval-ui-hint">model catalog unavailable — {modelCatalogError}</span>
+          ) : (
+            <span className="eval-ui-hint">model catalog is empty; enter a model manually.</span>
+          )}
+        </div>
+        <div className="eval-ui-grid-3">
+          <Field label="runs per variant" error={errors.runs}>
+            <Input
+              type="number"
+              min={1}
+              max={20}
+              value={form.runs}
+              onChange={(value) => update('runs', value)}
+              preserveCase
+            />
+            <button
+              type="button"
+              className="eval-ui-link"
+              onClick={() => update('runs', String(Math.max(2, Number(form.runs) || 1)))}
+            >
+              balanced A→B + B→A
+            </button>
+          </Field>
+          <Field
+            label="max total tokens"
+            error={errors.maxTotalTokens}
+            hint="optional; blank means no total-token limit"
+          >
+            <Input
+              type="number"
+              min={1}
+              value={form.maxTotalTokens}
+              onChange={(value) => update('maxTotalTokens', value)}
+              preserveCase
+              placeholder="no limit"
+            />
+          </Field>
+          <Field label="max cost (USD)" error={errors.maxCostUsd} hint="optional; requires catalog pricing">
+            <Input
+              type="number"
+              min={0}
+              step="0.000001"
+              value={form.maxCostUsd}
+              onChange={(value) => update('maxCostUsd', value)}
+              preserveCase
+              placeholder="no cost limit"
+            />
+          </Field>
+        </div>
+      </section>
+
+      <details className="eval-ui-advanced eval-ui-success-criteria">
+        <summary>success criteria (optional)</summary>
+        <div className="eval-ui-advanced-body">
+          <span className="eval-ui-hint">
+            Leave the expected value or custom function empty to collect outputs and metrics without scoring either
+            variant.
+          </span>
+          <Field label="judge each output with">
+            <Select
+              value={form.evaluatorMode}
+              options={[
+                {
+                  value: 'normalized_text',
+                  label: 'normalized text match',
+                },
+                { value: 'exact', label: 'exact value (strict)' },
+                { value: 'custom', label: 'custom iii function' },
+              ]}
+              onChange={(value) => update('evaluatorMode', value as EvalFormState['evaluatorMode'])}
+            />
+          </Field>
+          {form.evaluatorMode !== 'custom' ? (
+            <>
+              {form.evaluatorMode === 'exact' ? (
+                <Field label="expected format">
+                  <Select
+                    value={form.expectedFormat}
+                    options={[
+                      { value: 'text', label: 'text' },
+                      { value: 'json', label: 'JSON' },
+                    ]}
+                    onChange={(value) => update('expectedFormat', value as EvalFormState['expectedFormat'])}
+                  />
+                </Field>
+              ) : null}
+              <Field label="expected value" error={errors.expectedValue}>
+                <TextArea
+                  value={form.expectedValue}
+                  onChange={(event) => update('expectedValue', event.target.value)}
+                  rows={4}
+                  placeholder={form.evaluatorMode === 'exact' && form.expectedFormat === 'json' ? '{"ok":true}' : 'OK'}
+                />
+                {form.evaluatorMode === 'normalized_text' ? (
+                  <span className="eval-ui-hint">
+                    Ignores letter case, repeated whitespace, and surrounding punctuation.
+                  </span>
+                ) : null}
+              </Field>
+            </>
+          ) : (
+            <div className="eval-ui-grid-2">
+              <Field
+                label="evaluator function id"
+                error={errors.evaluatorFunctionId}
+                hint="receives output, metrics, run identity and arguments"
+              >
+                <Input
+                  value={form.evaluatorFunctionId}
+                  onChange={(value) => update('evaluatorFunctionId', value)}
+                  preserveCase
+                  placeholder="my-eval::assert"
+                />
+              </Field>
+              <Field label="arguments JSON" error={errors.evaluatorArguments}>
+                <TextArea
+                  className="code"
+                  value={form.evaluatorArguments}
+                  onChange={(event) => update('evaluatorArguments', event.target.value)}
+                  rows={5}
+                />
+              </Field>
+            </div>
+          )}
+        </div>
+      </details>
+
+      <details className="eval-ui-advanced">
+        <summary>advanced harness options</summary>
+        <div className="eval-ui-advanced-body">
+          <div className="eval-ui-grid-3">
+            <Field label="invocation timeout (s)" error={errors.invocationTimeoutSeconds} hint="optional">
+              <Input
+                type="number"
+                min={1}
+                value={form.invocationTimeoutSeconds}
+                onChange={(value) => update('invocationTimeoutSeconds', value)}
+                preserveCase
+                placeholder="default: 120"
+              />
+            </Field>
+            <Field label="scenario timeout (s)" error={errors.scenarioTimeoutSeconds} hint="optional">
+              <Input
+                type="number"
+                min={1}
+                value={form.scenarioTimeoutSeconds}
+                onChange={(value) => update('scenarioTimeoutSeconds', value)}
+                preserveCase
+                placeholder="default: 600"
+              />
+            </Field>
+            <Field label="max turns" error={errors.maxTurns} hint="optional">
+              <Input
+                type="number"
+                min={1}
+                value={form.maxTurns}
+                onChange={(value) => update('maxTurns', value)}
+                preserveCase
+                placeholder="default: 1000"
+              />
+            </Field>
+            <Field label="max output tokens / call" error={errors.maxOutputTokensPerCall} hint="optional">
+              <Input
+                type="number"
+                min={1}
+                value={form.maxOutputTokensPerCall}
+                onChange={(value) => update('maxOutputTokensPerCall', value)}
+                preserveCase
+                placeholder="default: 8192"
+              />
+            </Field>
+            <Field
+              label="max function-call errors"
+              error={errors.maxFunctionCallErrors}
+              hint="optional; blank means no limit"
+            >
+              <Input
+                type="number"
+                min={0}
+                value={form.maxFunctionCallErrors}
+                onChange={(value) => update('maxFunctionCallErrors', value)}
+                preserveCase
+                placeholder="no limit"
+              />
+            </Field>
+            <Field label="max error spans" error={errors.maxErrorSpans} hint="optional">
+              <Input
+                type="number"
+                min={0}
+                value={form.maxErrorSpans}
+                onChange={(value) => update('maxErrorSpans', value)}
+                preserveCase
+                placeholder="no span limit"
+              />
+            </Field>
+          </div>
+          <div className="eval-ui-grid-2">
+            <Field label="mode">
+              <Select
+                value={form.mode || undefined}
+                options={[
+                  { value: 'ask', label: 'ask' },
+                  { value: 'agent', label: 'agent' },
+                ]}
+                allowEmpty
+                emptyLabel="default"
+                onClear={() => update('mode', '')}
+                onChange={(value) => update('mode', value as EvalFormState['mode'])}
+              />
+            </Field>
+            <Field label="thinking level">
+              <Select
+                value={form.thinkingLevel || undefined}
+                options={['minimal', 'low', 'medium', 'high', 'xhigh'].map((value) => ({ value, label: value }))}
+                allowEmpty
+                emptyLabel="provider default"
+                onClear={() => update('thinkingLevel', '')}
+                onChange={(value) => update('thinkingLevel', value as EvalFormState['thinkingLevel'])}
+              />
+            </Field>
+          </div>
+          <JsonField
+            label="functions policy"
+            value={form.functionsJson}
+            error={errors.functionsJson}
+            hint="Optional. The preset matches chat; blank uses the harness default and denies function calls."
+            onChange={(value) => update('functionsJson', value)}
+          />
+          <JsonField
+            label="output contract"
+            value={form.outputJson}
+            error={errors.outputJson}
+            hint="Optional. Blank uses the harness text-output default."
+            onChange={(value) => update('outputJson', value)}
+          />
+          <JsonField
+            label="provider options"
+            value={form.providerOptionsJson}
+            error={errors.providerOptionsJson}
+            placeholder="{}"
+            onChange={(value) => update('providerOptionsJson', value)}
+          />
+          <JsonField
+            label="metadata"
+            value={form.metadataJson}
+            error={errors.metadataJson}
+            placeholder='{"fs_scope":{"root":"/workspace"}}'
+            onChange={(value) => update('metadataJson', value)}
+          />
+        </div>
+      </details>
+
+      <div className="eval-ui-submit">
+        <span>
+          Runs A {form.runs || '0'}× and B {form.runs || '0'}× in alternating order.
+        </span>
+        <Button variant="primary" size="md" type="submit" disabled={submitting}>
+          {submitting ? 'starting…' : 'run comparison'}
+        </Button>
+      </div>
+    </form>
+  )
+}
+
+function VariantEditor({
+  marker,
+  role,
+  label,
+  value,
+  valueLabel,
+  hint,
+  source,
+  valueError,
+  onLabel,
+  onValue,
+  onSource,
+  onLoadDefault,
+  loadingDefault = false,
+  defaultAvailable = true,
+}: {
+  marker: string
+  role: string
+  label: string
+  value: string
+  valueLabel: string
+  hint: string
+  source?: SystemPromptSource
+  valueError?: string
+  onLabel: (value: string) => void
+  onValue: (value: string) => void
+  onSource?: (source: SystemPromptSource) => void
+  onLoadDefault?: () => void
+  loadingDefault?: boolean
+  defaultAvailable?: boolean
+}) {
+  return (
+    <div className="eval-ui-variant">
+      <div className="eval-ui-variant-head">
+        <span className="eval-ui-variant-marker">{marker}</span>
+        <div>
+          <div className="eval-ui-variant-title">{role}</div>
+          <p>{hint}</p>
+        </div>
+      </div>
+      <Field label="name (optional)">
+        <Input value={label} onChange={onLabel} preserveCase />
+      </Field>
+      {source && onSource ? (
+        <Field label={`${valueLabel} source`} error={source === 'custom' ? undefined : valueError}>
+          <Select
+            value={source}
+            options={SYSTEM_PROMPT_SOURCE_OPTIONS}
+            onChange={(nextSource) => onSource(nextSource as SystemPromptSource)}
+          />
+        </Field>
+      ) : null}
+      {source === undefined || source === 'custom' ? (
+        <>
+          <Field label={source ? `custom ${valueLabel}` : valueLabel} error={valueError}>
+            <TextArea value={value} onChange={(event) => onValue(event.target.value)} rows={10} />
+          </Field>
+          {onLoadDefault ? (
+            <button
+              type="button"
+              className="eval-ui-link eval-ui-load-variant-default"
+              disabled={loadingDefault || !defaultAvailable}
+              onClick={onLoadDefault}
+            >
+              {loadingDefault
+                ? 'loading current default…'
+                : defaultAvailable
+                  ? 'load current default'
+                  : 'select model to load default'}
+            </button>
+          ) : null}
+        </>
+      ) : (
+        <SystemPromptSourceState source={source} />
+      )}
+    </div>
+  )
+}
+
+function SystemPromptSourceState({ source }: { source: Exclude<SystemPromptSource, 'custom'> }) {
+  return (
+    <div className="eval-ui-system-prompt-state">
+      <strong>{source === 'default' ? 'Current default' : 'No system prompt'}</strong>
+      <span>
+        {source === 'default'
+          ? 'Uses the system prompt resolved by the selected provider and harness.'
+          : 'Disables the provider and harness system prompt for this variant.'}
+      </span>
+    </div>
+  )
+}
+
+function JsonField({
+  label,
+  value,
+  error,
+  hint,
+  placeholder,
+  onChange,
+}: {
+  label: string
+  value: string
+  error?: string
+  hint?: string
+  placeholder?: string
+  onChange: (value: string) => void
+}) {
+  return (
+    <Field label={label} error={error} hint={hint}>
+      <TextArea
+        className="code"
+        value={value}
+        onChange={(event) => onChange(event.target.value)}
+        placeholder={placeholder}
+        rows={4}
+      />
+    </Field>
+  )
+}
diff --git a/eval/ui/src/page/index.tsx b/eval/ui/src/page/index.tsx
new file mode 100644
index 000000000..498605fb3
--- /dev/null
+++ b/eval/ui/src/page/index.tsx
@@ -0,0 +1,236 @@
+import { useCallback, useEffect, useMemo, useState } from 'react'
+import { EmptyState, Skeleton, StatusPanel, type Host } from '@iii-dev/console-ui'
+import { createEvalApi, errorMessage, isTerminal } from '../api'
+import { useEvalCompleted } from '../events'
+import type { CatalogModel, EvalRequest, EvalResultResponse, EvalStatusResponse, EvalSummary } from '../types'
+import { EvaluationDetail } from './EvaluationDetail'
+import { History } from './History'
+import { NewEvaluationForm } from './NewEvaluationForm'
+
+export function EvalPage({ host }: { host: Host }) {
+  const api = useMemo(() => createEvalApi(host), [host])
+  const [evaluations, setEvaluations] = useState<EvalSummary[]>([])
+  const [selectedId, setSelectedId] = useState<string | null>(null)
+  const [creating, setCreating] = useState(false)
+  const [historyLoading, setHistoryLoading] = useState(true)
+  const [historyError, setHistoryError] = useState<string | null>(null)
+  const [models, setModels] = useState<CatalogModel[]>([])
+  const [modelCatalogError, setModelCatalogError] = useState<string | null>(null)
+  const [status, setStatus] = useState<EvalStatusResponse | null>(null)
+  const [result, setResult] = useState<EvalResultResponse | null>(null)
+  const [detailLoading, setDetailLoading] = useState(false)
+  const [detailError, setDetailError] = useState<string | null>(null)
+  const [actionPending, setActionPending] = useState(false)
+
+  const loadHistory = useCallback(async () => {
+    setHistoryLoading(true)
+    try {
+      const next = await api.list()
+      setEvaluations(next)
+      setHistoryError(null)
+      setSelectedId((current) => {
+        if (current && next.some((item) => item.evaluation_id === current)) {
+          return current
+        }
+        return next[0]?.evaluation_id ?? null
+      })
+      if (next.length === 0) setCreating(true)
+    } catch (error) {
+      setHistoryError(errorMessage(error))
+    } finally {
+      setHistoryLoading(false)
+    }
+  }, [api])
+
+  const loadDetail = useCallback(
+    async (evaluationId: string, quiet = false) => {
+      if (!quiet) setDetailLoading(true)
+      try {
+        const [nextStatus, nextResult] = await Promise.all([api.status(evaluationId), api.result(evaluationId)])
+        setStatus(nextStatus)
+        setResult(nextResult)
+        setDetailError(null)
+        if (nextStatus) {
+          setEvaluations((current) =>
+            current.map((item) =>
+              item.evaluation_id === evaluationId
+                ? {
+                    ...item,
+                    status: nextStatus.status,
+                    total_runs: nextStatus.total_runs,
+                    terminal_runs: nextStatus.terminal_runs,
+                    passed_runs: nextStatus.passed_runs,
+                    failed_runs: nextStatus.failed_runs,
+                    updated_at: nextStatus.updated_at,
+                    completed_at: nextStatus.completed_at,
+                    error: nextStatus.error,
+                    eligible: nextResult?.report?.eligible ?? item.eligible,
+                  }
+                : item,
+            ),
+          )
+        }
+      } catch (error) {
+        setDetailError(errorMessage(error))
+      } finally {
+        if (!quiet) setDetailLoading(false)
+      }
+    },
+    [api],
+  )
+
+  useEffect(() => {
+    void loadHistory()
+    api
+      .models()
+      .then((catalog) => {
+        setModels(catalog)
+        setModelCatalogError(null)
+      })
+      .catch((error: unknown) => {
+        setModels([])
+        setModelCatalogError(errorMessage(error))
+      })
+  }, [api, loadHistory])
+
+  useEffect(() => {
+    if (!selectedId || creating) {
+      setStatus(null)
+      setResult(null)
+      return
+    }
+    void loadDetail(selectedId)
+  }, [creating, loadDetail, selectedId])
+
+  useEffect(() => {
+    if (!selectedId || creating || !status || isTerminal(status.status)) return
+    const timer = window.setInterval(() => {
+      void loadDetail(selectedId, true)
+    }, 2_000)
+    return () => window.clearInterval(timer)
+  }, [creating, loadDetail, selectedId, status])
+
+  const onCompleted = useCallback(
+    (event: { evaluation_id: string }) => {
+      void loadHistory()
+      if (event.evaluation_id === selectedId) {
+        void loadDetail(event.evaluation_id, true)
+      }
+    },
+    [loadDetail, loadHistory, selectedId],
+  )
+  useEvalCompleted(host, onCompleted)
+
+  const create = useCallback(
+    async (request: EvalRequest) => {
+      const started = await api.start(request)
+      setCreating(false)
+      setSelectedId(started.evaluation_id)
+      await loadHistory()
+      await loadDetail(started.evaluation_id)
+    },
+    [api, loadDetail, loadHistory],
+  )
+
+  const cancel = useCallback(async () => {
+    if (!selectedId) return
+    setActionPending(true)
+    try {
+      await api.cancel(selectedId)
+      await Promise.all([loadHistory(), loadDetail(selectedId)])
+    } catch (error) {
+      setDetailError(errorMessage(error))
+    } finally {
+      setActionPending(false)
+    }
+  }, [api, loadDetail, loadHistory, selectedId])
+
+  const remove = useCallback(async () => {
+    if (!selectedId) return
+    if (!window.confirm('Delete this evaluation and its stored report?')) return
+    setActionPending(true)
+    try {
+      await api.delete(selectedId)
+      setSelectedId(null)
+      setStatus(null)
+      setResult(null)
+      await loadHistory()
+    } catch (error) {
+      setDetailError(errorMessage(error))
+    } finally {
+      setActionPending(false)
+    }
+  }, [api, loadHistory, selectedId])
+
+  const rerun = useCallback(async (reverseOrder: boolean) => {
+    if (!selectedId) return
+    setActionPending(true)
+    try {
+      const started = await api.rerun(selectedId, reverseOrder)
+      setSelectedId(started.evaluation_id)
+      await loadHistory()
+      await loadDetail(started.evaluation_id)
+    } catch (error) {
+      setDetailError(errorMessage(error))
+    } finally {
+      setActionPending(false)
+    }
+  }, [api, loadDetail, loadHistory, selectedId])
+
+  const selected = evaluations.find((item) => item.evaluation_id === selectedId) ?? null
+
+  return (
+    <div className="eval-ui-container">
+      <div className={`eval-ui${creating ? ' creating' : ''}`}>
+        <History
+          evaluations={evaluations}
+          selectedId={creating ? null : selectedId}
+          loading={historyLoading}
+          onSelect={(evaluationId) => {
+            setCreating(false)
+            setSelectedId(evaluationId)
+          }}
+          onNew={() => setCreating(true)}
+          onRefresh={() => void loadHistory()}
+        />
+        <main className="eval-ui-main">
+          {historyError ? (
+            <StatusPanel variant="alert" headline="evaluation history unavailable" detail={historyError} />
+          ) : null}
+          {creating ? (
+            <NewEvaluationForm
+              models={models}
+              modelCatalogError={modelCatalogError}
+              onLoadDefaultSystemPrompt={(provider) => api.systemPrompt(provider)}
+              onCreate={create}
+            />
+          ) : selected ? (
+            <EvaluationDetail
+              summary={selected}
+              status={status}
+              result={result}
+              loading={detailLoading}
+              error={detailError}
+              actionPending={actionPending}
+              onCancel={() => void cancel()}
+              onDelete={() => void remove()}
+              onRerun={(reverseOrder) => void rerun(reverseOrder)}
+            />
+          ) : historyLoading ? (
+            <div className="eval-ui-loading">
+              <Skeleton />
+              <Skeleton />
+              <Skeleton />
+            </div>
+          ) : (
+            <EmptyState
+              title="select an evaluation"
+              description="choose a recent report or start a new prompt comparison."
+              action={{ label: 'new evaluation', onClick: () => setCreating(true) }}
+            />
+          )}
+        </main>
+      </div>
+    </div>
+  )
+}
diff --git a/eval/ui/src/types.ts b/eval/ui/src/types.ts
new file mode 100644
index 000000000..8f3bee2e2
--- /dev/null
+++ b/eval/ui/src/types.ts
@@ -0,0 +1,284 @@
+export type JsonValue =
+  | null
+  | boolean
+  | number
+  | string
+  | JsonValue[]
+  | { [key: string]: JsonValue }
+
+export type EvalStatus =
+  | 'queued'
+  | 'running'
+  | 'completed'
+  | 'failed'
+  | 'cancelled'
+
+export type EvalRunStatus =
+  | 'pending'
+  | 'running'
+  | 'completed'
+  | 'failed'
+  | 'cancelled'
+
+export type ComparisonDimension = 'prompt' | 'system_prompt'
+export type VariantRole = 'control' | 'treatment'
+export type ExecutionOrder = 'balanced_control_first' | 'balanced_treatment_first'
+
+export interface EvalVariant {
+  label?: string
+  prompt: string
+  system_prompt: string | null
+}
+
+export interface EvalModelConfig {
+  model: string
+  provider?: string
+  system_prompt_strategy: 'override' | 'enrich' | 'disabled'
+  mode?: 'ask' | 'agent'
+  thinking_level?: 'minimal' | 'low' | 'medium' | 'high' | 'xhigh'
+  provider_options?: Record<string, JsonValue>
+}
+
+export interface EvalLimits {
+  execution: {
+    invocation_timeout_seconds?: number
+    scenario_timeout_seconds?: number
+    max_turns?: number
+    max_output_tokens_per_call?: number
+    max_total_tokens?: number
+    max_cost_usd?: number
+  }
+  evaluation: {
+    max_function_call_errors?: number
+    max_error_spans?: number
+  }
+}
+
+export interface FunctionPolicy {
+  allow: string[]
+  deny: string[]
+  expose: 'agent_trigger' | 'native'
+}
+
+export type OutputContract =
+  | { type: 'text' }
+  | { type: 'json'; schema?: JsonValue }
+
+export interface EvalRequest {
+  dimension: ComparisonDimension
+  model: EvalModelConfig
+  control: EvalVariant
+  treatment: EvalVariant
+  evaluator?: {
+    function_id: string
+    arguments: JsonValue
+  }
+  runs: number
+  execution_order: ExecutionOrder
+  source_evaluation_id?: string
+  limits: EvalLimits
+  functions?: FunctionPolicy
+  output?: OutputContract
+  metadata?: JsonValue
+}
+
+export interface EvalSummary {
+  evaluation_id: string
+  status: EvalStatus
+  dimension: ComparisonDimension
+  model: string
+  provider?: string
+  control_label?: string
+  treatment_label?: string
+  total_runs: number
+  terminal_runs: number
+  passed_runs: number
+  failed_runs: number
+  eligible?: boolean
+  created_at: number
+  updated_at: number
+  completed_at?: number
+  error?: string
+}
+
+export interface EvalStatusResponse {
+  evaluation_id: string
+  status: EvalStatus
+  total_runs: number
+  terminal_runs: number
+  passed_runs: number
+  failed_runs: number
+  active?: {
+    run_id: string
+    role: VariantRole
+    iteration: number
+    session_id: string
+    turn_id?: string
+    started_at: number
+  }
+  created_at: number
+  updated_at: number
+  completed_at?: number
+  error?: string
+}
+
+export interface Benchmark {
+  wall_time_ms: number
+  sessions: number
+  turns: number
+  function_calls: number
+  function_call_errors: number
+  input_tokens?: number
+  output_tokens?: number
+  total_tokens?: number
+  cache_read_tokens?: number
+  cache_write_tokens?: number
+  reasoning_tokens?: number
+  cost_usd?: number
+  trace_count?: number
+  span_count?: number
+  error_span_count?: number
+  trace_duration_ms?: number
+}
+
+export interface VariantAggregate {
+  runs: number
+  evaluated_runs: number
+  passed: number
+  pass_rate: number
+  benchmarked_runs: number
+  median_score?: number
+  median_wall_time_ms?: number
+  median_input_tokens?: number
+  median_output_tokens?: number
+  median_total_tokens?: number
+  median_reasoning_tokens?: number
+  median_function_calls?: number
+  median_function_call_errors?: number
+  median_cost_usd?: number
+  median_trace_count?: number
+  median_span_count?: number
+  median_error_span_count?: number
+  median_trace_duration_ms?: number
+}
+
+export interface AggregateDelta {
+  pass_rate: number
+  median_score?: number
+  median_wall_time_ms?: number
+  median_total_tokens?: number
+  median_reasoning_tokens?: number
+  median_function_calls?: number
+  median_function_call_errors?: number
+  median_cost_usd?: number
+  median_trace_count?: number
+  median_span_count?: number
+  median_error_span_count?: number
+  median_trace_duration_ms?: number
+}
+
+export interface EvalRun {
+  run_id: string
+  role: VariantRole
+  iteration: number
+  execution_position: number
+  pair_position: number
+  status: EvalRunStatus
+  session_id: string
+  turn_id?: string
+  passed?: boolean
+  started_at: number
+  completed_at?: number
+  output?: JsonValue
+  metrics?: JsonValue
+  evaluation?: {
+    passed: boolean
+    score?: number
+    reason?: string
+    details?: JsonValue
+  }
+  benchmark?: Benchmark
+  failures?: Array<{
+    phase: string
+    message: string
+    function_id?: string
+  }>
+}
+
+export interface EvalReport {
+  schema_version: string
+  evaluation_id: string
+  source_evaluation_id?: string
+  execution_order_policy: ExecutionOrder
+  effective_execution_order: string[]
+  dimension: ComparisonDimension
+  model: EvalModelConfig
+  evaluator?: {
+    function_id: string
+    arguments_sha256: string
+  }
+  control: {
+    label?: string
+    prompt_sha256: string
+    system_prompt_sha256?: string
+  }
+  treatment: {
+    label?: string
+    prompt_sha256: string
+    system_prompt_sha256?: string
+  }
+  shared_artifacts: {
+    model_sha256: string
+    function_policy_sha256: string
+    limits_sha256: string
+    output_sha256: string
+  }
+  control_aggregate: VariantAggregate
+  treatment_aggregate: VariantAggregate
+  delta: AggregateDelta
+  order_sensitivity: OrderSensitivity
+  eligible?: boolean
+  runs: EvalRun[]
+  created_at: number
+  completed_at: number
+}
+
+export interface RoleOrderSensitivity {
+  first_runs: number
+  first_passed: number
+  second_runs: number
+  second_passed: number
+  differs: boolean
+}
+
+export interface OrderSensitivity {
+  detected: boolean
+  control: RoleOrderSensitivity
+  treatment: RoleOrderSensitivity
+}
+
+export interface EvalProgress {
+  effective_execution_order: string[]
+  control_aggregate: VariantAggregate
+  treatment_aggregate: VariantAggregate
+  delta: AggregateDelta
+  order_sensitivity: OrderSensitivity
+  runs: EvalRun[]
+}
+
+export interface EvalResultResponse {
+  status: EvalStatus
+  request: EvalRequest
+  progress: EvalProgress
+  report?: EvalReport
+}
+
+export interface CatalogModel {
+  id: string
+  provider: string
+  display_name?: string
+  context_window: number
+  max_output_tokens: number
+  supports_thinking?: boolean
+  supports_xhigh?: boolean
+}
diff --git a/eval/ui/styles.css b/eval/ui/styles.css
new file mode 100644
index 000000000..488345660
--- /dev/null
+++ b/eval/ui/styles.css
@@ -0,0 +1,986 @@
+[data-iii-ui="eval"] .eval-ui-container {
+  width: 100%;
+  min-width: 0;
+  container-name: eval-panel;
+  container-type: inline-size;
+}
+
+[data-iii-ui="eval"] .eval-ui {
+  display: grid;
+  grid-template-columns: 280px minmax(0, 1fr);
+  min-height: 100%;
+  font-family: var(--font-mono, ui-monospace, monospace);
+  color: var(--color-ink);
+  background: var(--color-bg);
+}
+
+[data-iii-ui="eval"] .eval-ui-history {
+  display: flex;
+  flex-direction: column;
+  gap: 12px;
+  min-width: 0;
+  padding: 16px;
+  border-right: 1px solid var(--color-rule);
+  background: var(--color-panel);
+}
+
+[data-iii-ui="eval"] .eval-ui-history-head {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 8px;
+  color: var(--color-ink-faint);
+  font-size: 10.5px;
+  font-weight: 600;
+  letter-spacing: 0.14em;
+  text-transform: uppercase;
+}
+
+[data-iii-ui="eval"] .eval-ui-history-list {
+  display: flex;
+  flex: 1;
+  flex-direction: column;
+  gap: 8px;
+  min-height: 0;
+  overflow-y: auto;
+}
+
+[data-iii-ui="eval"] .eval-ui-history-row {
+  display: flex;
+  flex-direction: column;
+  gap: 5px;
+  width: 100%;
+  padding: 10px;
+  border: 1px solid var(--color-rule);
+  background: var(--color-bg);
+  color: var(--color-ink);
+  font: inherit;
+  text-align: left;
+  cursor: pointer;
+}
+
+[data-iii-ui="eval"] .eval-ui-history-row:hover {
+  border-color: var(--color-ink-faint);
+  background: var(--color-paper-2);
+}
+
+[data-iii-ui="eval"] .eval-ui-history-row.active {
+  border-color: var(--color-accent);
+  box-shadow: inset 2px 0 0 var(--color-accent);
+}
+
+[data-iii-ui="eval"] .eval-ui-history-row-top {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 8px;
+}
+
+[data-iii-ui="eval"] .eval-ui-history-progress,
+[data-iii-ui="eval"] .eval-ui-history-meta {
+  overflow: hidden;
+  color: var(--color-ink-ghost);
+  font-size: 10.5px;
+  line-height: 1.35;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+}
+
+[data-iii-ui="eval"] .eval-ui-history-title {
+  overflow: hidden;
+  font-size: 12.5px;
+  font-weight: 600;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+}
+
+[data-iii-ui="eval"] .eval-ui-main {
+  min-width: 0;
+  padding: 24px;
+}
+
+[data-iii-ui="eval"] .eval-ui-form,
+[data-iii-ui="eval"] .eval-ui-detail {
+  display: flex;
+  flex-direction: column;
+  gap: 18px;
+  width: min(100%, 1120px);
+  margin: 0 auto;
+}
+
+[data-iii-ui="eval"] .eval-ui-section-head,
+[data-iii-ui="eval"] .eval-ui-detail-head {
+  display: flex;
+  align-items: flex-start;
+  justify-content: space-between;
+  gap: 16px;
+}
+
+[data-iii-ui="eval"] .eval-ui-section-head h1,
+[data-iii-ui="eval"] .eval-ui-detail-head h1 {
+  margin: 0;
+  color: var(--color-ink);
+  font-size: 18px;
+  font-weight: 650;
+  letter-spacing: -0.02em;
+}
+
+[data-iii-ui="eval"] .eval-ui-section-head p,
+[data-iii-ui="eval"] .eval-ui-detail-head p {
+  margin: 5px 0 0;
+  color: var(--color-ink-faint);
+  font-size: 12px;
+}
+
+[data-iii-ui="eval"] .eval-ui-id {
+  display: flex;
+  align-items: center;
+  gap: 7px;
+  margin-top: 7px;
+}
+
+[data-iii-ui="eval"] .eval-ui-id code {
+  color: var(--color-ink-faint);
+  font-size: 11px;
+}
+
+[data-iii-ui="eval"] .eval-ui-id button {
+  display: inline-flex;
+  flex: 0 0 auto;
+  align-items: center;
+  justify-content: center;
+  padding: 0;
+  border: 0;
+  background: transparent;
+  color: var(--color-ink-ghost);
+  cursor: pointer;
+  transition: color 120ms ease;
+}
+
+[data-iii-ui="eval"] .eval-ui-id button:hover {
+  color: var(--color-ink);
+}
+
+[data-iii-ui="eval"] .eval-ui-sr-only {
+  position: absolute;
+  width: 1px;
+  height: 1px;
+  padding: 0;
+  overflow: hidden;
+  clip: rect(0, 0, 0, 0);
+  white-space: nowrap;
+  border: 0;
+}
+
+[data-iii-ui="eval"] .eval-ui-title-row,
+[data-iii-ui="eval"] .eval-ui-actions,
+[data-iii-ui="eval"] .eval-ui-inline-action,
+[data-iii-ui="eval"] .eval-ui-definition-meta {
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  flex-wrap: wrap;
+}
+
+[data-iii-ui="eval"] .eval-ui-panel {
+  display: flex;
+  flex-direction: column;
+  gap: 16px;
+  padding: 16px;
+  border: 1px solid var(--color-rule);
+  background: var(--color-panel);
+}
+
+[data-iii-ui="eval"] .eval-ui-panel-title {
+  color: var(--color-ink-faint);
+  font-size: 10.5px;
+  font-weight: 600;
+  letter-spacing: 0.14em;
+  text-transform: uppercase;
+}
+
+[data-iii-ui="eval"] .eval-ui-step-head {
+  display: flex;
+  align-items: flex-start;
+  gap: 10px;
+}
+
+[data-iii-ui="eval"] .eval-ui-step-head p,
+[data-iii-ui="eval"] .eval-ui-variant-head p {
+  margin: 4px 0 0;
+  color: var(--color-ink-ghost);
+  font-size: 10.5px;
+  line-height: 1.45;
+}
+
+[data-iii-ui="eval"] .eval-ui-step-number,
+[data-iii-ui="eval"] .eval-ui-variant-marker {
+  display: inline-flex;
+  flex: 0 0 auto;
+  align-items: center;
+  justify-content: center;
+  width: 22px;
+  height: 22px;
+  border: 1px solid var(--color-rule-2);
+  background: var(--color-bg);
+  color: var(--color-ink);
+  font-size: 11px;
+  font-weight: 700;
+}
+
+[data-iii-ui="eval"] .eval-ui-comparison-rule {
+  display: flex;
+  align-items: center;
+  gap: 7px;
+  flex-wrap: wrap;
+  padding: 9px 10px;
+  border-left: 2px solid var(--color-accent);
+  background: var(--color-paper-2);
+  color: var(--color-ink-faint);
+  font-size: 11px;
+  line-height: 1.45;
+}
+
+[data-iii-ui="eval"] .eval-ui-comparison-rule strong {
+  color: var(--color-ink);
+  font-weight: 650;
+}
+
+[data-iii-ui="eval"] .eval-ui-shared-editor {
+  display: flex;
+  flex-direction: column;
+  gap: 9px;
+  padding: 12px;
+  border: 1px dashed var(--color-rule-2);
+  background: var(--color-paper-2);
+}
+
+[data-iii-ui="eval"] .eval-ui-shared-editor-head {
+  display: flex;
+  align-items: baseline;
+  justify-content: space-between;
+  gap: 12px;
+}
+
+[data-iii-ui="eval"] .eval-ui-shared-editor-head span {
+  color: var(--color-ink-ghost);
+  font-size: 9.5px;
+  font-weight: 600;
+  letter-spacing: 0.1em;
+  text-transform: uppercase;
+}
+
+[data-iii-ui="eval"] .eval-ui-shared-editor-head strong {
+  color: var(--color-ink);
+  font-size: 11.5px;
+  font-weight: 600;
+}
+
+[data-iii-ui="eval"] .eval-ui-secondary-input {
+  border: 1px dashed var(--color-rule-2);
+  background: var(--color-paper-2);
+}
+
+[data-iii-ui="eval"] .eval-ui-secondary-input summary {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 12px;
+  padding: 11px 12px;
+  color: var(--color-ink);
+  font-size: 11.5px;
+  cursor: pointer;
+}
+
+[data-iii-ui="eval"] .eval-ui-secondary-input summary span:last-child {
+  color: var(--color-ink-ghost);
+  font-size: 10px;
+}
+
+[data-iii-ui="eval"] .eval-ui-secondary-input-body {
+  display: flex;
+  flex-direction: column;
+  gap: 14px;
+  padding: 0 12px 12px;
+}
+
+[data-iii-ui="eval"] .eval-ui-load-default {
+  display: flex;
+  flex-direction: column;
+  align-items: flex-start;
+  gap: 6px;
+}
+
+[data-iii-ui="eval"] .eval-ui-system-prompt-mode {
+  width: min(100%, 360px);
+}
+
+[data-iii-ui="eval"] .eval-ui-load-variant-default {
+  align-self: flex-start;
+}
+
+[data-iii-ui="eval"] .eval-ui-grid-2,
+[data-iii-ui="eval"] .eval-ui-grid-3,
+[data-iii-ui="eval"] .eval-ui-variants {
+  display: grid;
+  gap: 14px;
+}
+
+[data-iii-ui="eval"] .eval-ui-grid-2,
+[data-iii-ui="eval"] .eval-ui-variants {
+  grid-template-columns: repeat(2, minmax(0, 1fr));
+}
+
+[data-iii-ui="eval"] .eval-ui-grid-3 {
+  grid-template-columns: repeat(3, minmax(0, 1fr));
+}
+
+[data-iii-ui="eval"] .eval-ui-variant {
+  display: flex;
+  flex-direction: column;
+  gap: 12px;
+  min-width: 0;
+  padding: 14px;
+  border: 1px solid var(--color-rule-2);
+  background: var(--color-bg);
+}
+
+[data-iii-ui="eval"] .eval-ui-variant:nth-child(2) {
+  border-color: var(--color-accent);
+}
+
+[data-iii-ui="eval"] .eval-ui-variant-head {
+  display: flex;
+  align-items: flex-start;
+  gap: 9px;
+}
+
+[data-iii-ui="eval"] .eval-ui-variant-marker {
+  border-color: var(--color-accent);
+  color: var(--color-accent);
+}
+
+[data-iii-ui="eval"] .eval-ui-variant-title {
+  color: var(--color-accent);
+  font-size: 10.5px;
+  font-weight: 600;
+  letter-spacing: 0.13em;
+  text-transform: uppercase;
+}
+
+[data-iii-ui="eval"] .eval-ui-variant.readonly pre,
+[data-iii-ui="eval"] .eval-ui-shared pre {
+  overflow: auto;
+  max-height: 280px;
+  margin: 0;
+  padding: 10px;
+  border: 1px solid var(--color-rule-2);
+  background: var(--color-paper-2);
+  color: var(--color-ink);
+  font: inherit;
+  font-size: 12px;
+  line-height: 1.55;
+  white-space: pre-wrap;
+  overflow-wrap: anywhere;
+}
+
+[data-iii-ui="eval"] .eval-ui-field {
+  display: flex;
+  flex-direction: column;
+  gap: 6px;
+  min-width: 0;
+}
+
+[data-iii-ui="eval"] .eval-ui-label {
+  color: var(--color-ink-faint);
+  font-size: 11.5px;
+}
+
+[data-iii-ui="eval"] .eval-ui-hint,
+[data-iii-ui="eval"] .eval-ui-report-foot {
+  color: var(--color-ink-ghost);
+  font-size: 10.5px;
+  line-height: 1.45;
+}
+
+[data-iii-ui="eval"] .eval-ui-field-error {
+  color: var(--color-alert);
+  font-size: 10.5px;
+  line-height: 1.45;
+}
+
+[data-iii-ui="eval"] .eval-ui-error-summary {
+  display: flex;
+  flex-direction: column;
+  gap: 5px;
+  padding: 10px 12px;
+  border: 1px solid var(--color-alert);
+  background: color-mix(in srgb, var(--color-alert) 7%, transparent);
+  color: var(--color-alert);
+  font-size: 10.5px;
+  line-height: 1.45;
+}
+
+[data-iii-ui="eval"] .eval-ui-error-summary strong {
+  font-size: 11.5px;
+}
+
+[data-iii-ui="eval"] .eval-ui-system-prompt-state {
+  display: flex;
+  flex-direction: column;
+  gap: 5px;
+  padding: 11px 12px;
+  border: 1px dashed var(--color-rule-2);
+  background: var(--color-paper-2);
+}
+
+[data-iii-ui="eval"] .eval-ui-system-prompt-state strong {
+  color: var(--color-ink);
+  font-size: 11.5px;
+  font-weight: 600;
+}
+
+[data-iii-ui="eval"] .eval-ui-system-prompt-state span {
+  color: var(--color-ink-ghost);
+  font-size: 10.5px;
+  line-height: 1.45;
+}
+
+[data-iii-ui="eval"] .eval-ui-textarea {
+  box-sizing: border-box;
+  width: 100%;
+  min-height: 88px;
+  padding: 10px;
+  resize: vertical;
+  border: 1px solid var(--color-rule);
+  border-radius: 0;
+  outline: none;
+  background: var(--color-bg);
+  color: var(--color-ink);
+  font: inherit;
+  font-size: 12.5px;
+  line-height: 1.55;
+}
+
+[data-iii-ui="eval"] .eval-ui-textarea.code {
+  font-size: 11.5px;
+}
+
+[data-iii-ui="eval"] .eval-ui-textarea:focus {
+  border-color: var(--color-accent);
+  box-shadow: 0 0 0 1px var(--color-accent);
+}
+
+[data-iii-ui="eval"] .eval-ui-link {
+  padding: 0;
+  border: 0;
+  background: transparent;
+  color: var(--color-accent);
+  font: inherit;
+  font-size: 11px;
+  text-decoration: underline;
+  cursor: pointer;
+}
+
+[data-iii-ui="eval"] .eval-ui-link:disabled {
+  cursor: wait;
+  opacity: 0.55;
+}
+
+[data-iii-ui="eval"] .eval-ui-advanced {
+  border: 1px solid var(--color-rule);
+  background: var(--color-panel);
+}
+
+[data-iii-ui="eval"] .eval-ui-advanced summary {
+  padding: 12px 14px;
+  color: var(--color-ink-faint);
+  font-size: 11px;
+  font-weight: 600;
+  letter-spacing: 0.1em;
+  text-transform: uppercase;
+  cursor: pointer;
+}
+
+[data-iii-ui="eval"] .eval-ui-advanced-body {
+  display: flex;
+  flex-direction: column;
+  gap: 16px;
+  padding: 0 14px 14px;
+}
+
+[data-iii-ui="eval"] .eval-ui-submit {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 16px;
+  padding-top: 4px;
+  color: var(--color-ink-faint);
+  font-size: 11px;
+}
+
+[data-iii-ui="eval"] .eval-ui-progress {
+  display: flex;
+  flex-direction: column;
+  gap: 7px;
+}
+
+[data-iii-ui="eval"] .eval-ui-progress-meta {
+  display: flex;
+  justify-content: space-between;
+  gap: 12px;
+  color: var(--color-ink-faint);
+  font-size: 11px;
+}
+
+[data-iii-ui="eval"] .eval-ui-progress-track {
+  height: 4px;
+  overflow: hidden;
+  background: var(--color-rule-2);
+}
+
+[data-iii-ui="eval"] .eval-ui-progress-track span {
+  display: block;
+  height: 100%;
+  background: var(--color-accent);
+  transition: width 180ms ease-out;
+}
+
+[data-iii-ui="eval"] .eval-ui-loading {
+  display: flex;
+  flex-direction: column;
+  gap: 10px;
+  width: min(100%, 760px);
+  margin: 30px auto;
+}
+
+[data-iii-ui="eval"] .eval-ui-loading > * {
+  min-height: 18px;
+}
+
+[data-iii-ui="eval"] .eval-ui-definition-meta {
+  color: var(--color-ink-ghost);
+  font-size: 10.5px;
+}
+
+[data-iii-ui="eval"] .eval-ui-definition-meta span {
+  padding: 2px 7px;
+  border: 1px solid var(--color-rule-2);
+}
+
+[data-iii-ui="eval"] .eval-ui-shared {
+  display: flex;
+  flex-direction: column;
+  gap: 6px;
+}
+
+[data-iii-ui="eval"] .eval-ui-shared > span {
+  color: var(--color-ink-faint);
+  font-size: 11.5px;
+}
+
+[data-iii-ui="eval"] .eval-ui-metrics {
+  overflow-x: auto;
+  border: 1px solid var(--color-rule-2);
+  background: var(--color-bg);
+}
+
+[data-iii-ui="eval"] .eval-ui-report-note {
+  margin-bottom: 10px;
+  color: var(--color-ink-faint);
+  font-size: 10.5px;
+  line-height: 1.45;
+}
+
+[data-iii-ui="eval"] .eval-ui-report-section {
+  display: flex;
+  flex-direction: column;
+  gap: 10px;
+}
+
+[data-iii-ui="eval"] .eval-ui-report-section + .eval-ui-report-section {
+  padding-top: 16px;
+  border-top: 1px solid var(--color-rule-2);
+}
+
+[data-iii-ui="eval"] .eval-ui-metric {
+  display: grid;
+  grid-template-columns: minmax(150px, 1.4fr) repeat(3, minmax(105px, 1fr));
+  min-width: 560px;
+  border-bottom: 1px solid var(--color-rule-2);
+}
+
+[data-iii-ui="eval"] .eval-ui-metric:last-child {
+  border-bottom: 0;
+}
+
+[data-iii-ui="eval"] .eval-ui-metric span {
+  padding: 8px 10px;
+  color: var(--color-ink);
+  font-size: 11.5px;
+  text-align: right;
+}
+
+[data-iii-ui="eval"] .eval-ui-metric span:first-child {
+  color: var(--color-ink-faint);
+  text-align: left;
+}
+
+[data-iii-ui="eval"] .eval-ui-metric.header {
+  background: var(--color-paper-2);
+}
+
+[data-iii-ui="eval"] .eval-ui-metric.header span {
+  color: var(--color-ink-ghost);
+  font-size: 10px;
+  font-weight: 600;
+  letter-spacing: 0.08em;
+  text-transform: uppercase;
+}
+
+[data-iii-ui="eval"] .eval-ui-metric .positive {
+  color: var(--color-ok);
+}
+
+[data-iii-ui="eval"] .eval-ui-metric .negative {
+  color: var(--color-alert);
+}
+
+[data-iii-ui="eval"] .eval-ui-runs {
+  display: grid;
+  grid-template-columns: minmax(300px, 0.9fr) minmax(0, 1.4fr);
+  gap: 14px;
+}
+
+[data-iii-ui="eval"] .eval-ui-run-pairs {
+  display: flex;
+  flex-direction: column;
+  gap: 10px;
+  min-width: 0;
+}
+
+[data-iii-ui="eval"] .eval-ui-run-pair {
+  padding: 9px;
+  border: 1px solid var(--color-rule-2);
+  background: var(--color-bg);
+}
+
+[data-iii-ui="eval"] .eval-ui-run-pair.order-sensitive {
+  border-left: 2px solid var(--color-warn, var(--color-alert));
+}
+
+[data-iii-ui="eval"] .eval-ui-run-pair-head {
+  margin-bottom: 7px;
+  color: var(--color-ink-ghost);
+  font-size: 9.5px;
+  font-weight: 600;
+  letter-spacing: 0.08em;
+  text-transform: uppercase;
+}
+
+[data-iii-ui="eval"] .eval-ui-run-pair-grid {
+  display: grid;
+  grid-template-columns: repeat(2, minmax(0, 1fr));
+  gap: 8px;
+}
+
+[data-iii-ui="eval"] .eval-ui-run-card {
+  display: flex;
+  flex-direction: column;
+  gap: 5px;
+  min-width: 0;
+  padding: 10px;
+  border: 1px solid var(--color-rule-2);
+  background: var(--color-panel);
+  color: var(--color-ink-faint);
+  font: inherit;
+  font-size: 10.5px;
+  text-align: left;
+  cursor: pointer;
+}
+
+[data-iii-ui="eval"] .eval-ui-run-card:hover,
+[data-iii-ui="eval"] .eval-ui-run-card.selected {
+  border-color: var(--color-accent);
+}
+
+[data-iii-ui="eval"] .eval-ui-run-card.running {
+  animation: eval-run-pulse 1.8s ease-in-out infinite;
+}
+
+[data-iii-ui="eval"] .eval-ui-run-card-title {
+  color: var(--color-ink);
+  font-size: 11.5px;
+  font-weight: 650;
+}
+
+@keyframes eval-run-pulse {
+  0%, 100% { box-shadow: 0 0 0 0 color-mix(in srgb, var(--color-accent) 0%, transparent); }
+  50% { box-shadow: 0 0 0 3px color-mix(in srgb, var(--color-accent) 18%, transparent); }
+}
+
+[data-iii-ui="eval"] .eval-ui-order {
+  display: flex;
+  flex-direction: column;
+  gap: 6px;
+}
+
+[data-iii-ui="eval"] .eval-ui-order-list {
+  display: flex;
+  align-items: center;
+  gap: 7px;
+  overflow-x: auto;
+  padding: 8px 10px;
+  border: 1px solid var(--color-rule-2);
+  background: var(--color-paper-2);
+}
+
+[data-iii-ui="eval"] .eval-ui-order-item-wrap {
+  display: flex;
+  flex: 0 0 auto;
+  align-items: center;
+  gap: 7px;
+}
+
+[data-iii-ui="eval"] .eval-ui-order-arrow {
+  color: var(--color-ink-ghost);
+  font-size: 11px;
+}
+
+[data-iii-ui="eval"] .eval-ui-order-item {
+  display: flex;
+  align-items: center;
+  gap: 6px;
+  padding: 6px 8px;
+  border: 1px solid var(--color-rule);
+  background: var(--color-bg);
+  color: var(--color-ink-faint);
+  font-size: 10.5px;
+  white-space: nowrap;
+}
+
+[data-iii-ui="eval"] .eval-ui-order-item strong {
+  color: var(--color-ink);
+}
+
+[data-iii-ui="eval"] .eval-ui-success-counts {
+  display: flex;
+  justify-content: space-between;
+  gap: 12px;
+  padding: 8px 10px;
+  border-bottom: 1px solid var(--color-rule-2);
+  color: var(--color-ink-faint);
+  font-size: 10.5px;
+}
+
+[data-iii-ui="eval"] .eval-ui-hashes {
+  border: 1px dashed var(--color-rule-2);
+}
+
+[data-iii-ui="eval"] .eval-ui-hashes summary {
+  padding: 9px 10px;
+  color: var(--color-ink-faint);
+  font-size: 10.5px;
+  cursor: pointer;
+}
+
+[data-iii-ui="eval"] .eval-ui-hashes > div {
+  display: grid;
+  gap: 5px;
+  padding: 0 10px 10px;
+}
+
+[data-iii-ui="eval"] .eval-ui-hashes code {
+  display: grid;
+  grid-template-columns: 90px minmax(0, 1fr);
+  gap: 8px;
+  overflow: hidden;
+  color: var(--color-ink-ghost);
+  font-size: 9.5px;
+  text-overflow: ellipsis;
+}
+
+[data-iii-ui="eval"] .eval-ui-hashes code span {
+  color: var(--color-ink-faint);
+}
+
+[data-iii-ui="eval"] .eval-ui-run-list {
+  min-width: 0;
+  overflow-x: auto;
+  border: 1px solid var(--color-rule-2);
+  background: var(--color-bg);
+}
+
+[data-iii-ui="eval"] .eval-ui-run-header,
+[data-iii-ui="eval"] .eval-ui-run-row {
+  display: grid;
+  grid-template-columns: minmax(155px, 1.5fr) 0.7fr 0.7fr 0.7fr;
+}
+
+[data-iii-ui="eval"] .eval-ui-run-header {
+  min-width: 480px;
+  padding: 7px 10px;
+  border-bottom: 1px solid var(--color-rule-2);
+  background: var(--color-paper-2);
+  color: var(--color-ink-ghost);
+  font-size: 9.5px;
+  font-weight: 600;
+  letter-spacing: 0.08em;
+  text-transform: uppercase;
+}
+
+[data-iii-ui="eval"] .eval-ui-run-row {
+  min-width: 480px;
+  width: 100%;
+  padding: 8px 10px;
+  border: 0;
+  border-bottom: 1px solid var(--color-rule-2);
+  background: transparent;
+  color: var(--color-ink);
+  font: inherit;
+  font-size: 11px;
+  text-align: left;
+  cursor: pointer;
+}
+
+[data-iii-ui="eval"] .eval-ui-run-header span:not(:first-child),
+[data-iii-ui="eval"] .eval-ui-run-row span:not(:first-child) {
+  text-align: right;
+}
+
+[data-iii-ui="eval"] .eval-ui-run-row:last-child {
+  border-bottom: 0;
+}
+
+[data-iii-ui="eval"] .eval-ui-run-row:hover,
+[data-iii-ui="eval"] .eval-ui-run-row.active {
+  background: var(--color-paper-2);
+}
+
+[data-iii-ui="eval"] .eval-ui-run-row.active {
+  box-shadow: inset 2px 0 0 var(--color-accent);
+}
+
+[data-iii-ui="eval"] .eval-ui-run-detail {
+  display: flex;
+  flex-direction: column;
+  gap: 10px;
+  min-width: 0;
+}
+
+[data-iii-ui="eval"] .eval-ui-markdown-output {
+  box-sizing: border-box;
+  height: 360px;
+  max-height: 360px;
+  overflow-x: hidden;
+  overflow-y: auto;
+  padding: 12px;
+  border: 1px solid var(--color-rule-2);
+  background: var(--color-bg);
+}
+
+[data-iii-ui="eval"] .eval-ui-failures {
+  padding: 10px;
+  border: 1px solid var(--color-alert);
+  color: var(--color-alert);
+  font-size: 11px;
+  line-height: 1.5;
+}
+
+[data-iii-ui="eval"] .eval-ui-report-foot {
+  padding-top: 12px;
+  border-top: 1px solid var(--color-rule-2);
+}
+
+@container eval-panel (max-width: 900px) {
+  [data-iii-ui="eval"] .eval-ui {
+    grid-template-columns: 1fr;
+  }
+
+  [data-iii-ui="eval"] .eval-ui.creating .eval-ui-history {
+    display: none;
+  }
+
+  [data-iii-ui="eval"] .eval-ui-history {
+    border-right: 0;
+    border-bottom: 1px solid var(--color-rule);
+  }
+
+  [data-iii-ui="eval"] .eval-ui-history-list {
+    flex: none;
+    flex-direction: row;
+    overflow-x: auto;
+    overflow-y: hidden;
+  }
+
+  [data-iii-ui="eval"] .eval-ui-history-row {
+    flex: 0 0 min(280px, 72%);
+  }
+
+  [data-iii-ui="eval"] .eval-ui-grid-3 {
+    grid-template-columns: 1fr;
+  }
+
+  [data-iii-ui="eval"] .eval-ui-runs {
+    grid-template-columns: 1fr;
+  }
+}
+
+@container eval-panel (max-width: 680px) {
+  [data-iii-ui="eval"] .eval-ui-main {
+    padding: 16px;
+  }
+
+  [data-iii-ui="eval"] .eval-ui-grid-2,
+  [data-iii-ui="eval"] .eval-ui-variants,
+  [data-iii-ui="eval"] .eval-ui-run-pair-grid {
+    grid-template-columns: 1fr;
+  }
+
+  [data-iii-ui="eval"] .eval-ui-submit,
+  [data-iii-ui="eval"] .eval-ui-section-head,
+  [data-iii-ui="eval"] .eval-ui-detail-head {
+    align-items: stretch;
+    flex-direction: column;
+  }
+}
+
+@container eval-panel (max-width: 520px) {
+  [data-iii-ui="eval"] .eval-ui-main {
+    padding: 12px;
+  }
+
+  [data-iii-ui="eval"] .eval-ui-panel {
+    padding: 12px;
+  }
+
+  [data-iii-ui="eval"] .eval-ui-run-header {
+    display: none;
+  }
+
+  [data-iii-ui="eval"] .eval-ui-run-row {
+    grid-template-columns: minmax(0, 1fr) auto;
+    gap: 5px 10px;
+    min-width: 0;
+  }
+
+  [data-iii-ui="eval"] .eval-ui-run-row span:first-child {
+    grid-column: 1 / -1;
+    font-weight: 600;
+  }
+
+  [data-iii-ui="eval"] .eval-ui-run-row span:nth-child(2) {
+    grid-column: 1 / -1;
+    text-align: left;
+  }
+
+  [data-iii-ui="eval"] .eval-ui-run-row span:nth-child(3)::before {
+    content: 'tokens ';
+    color: var(--color-ink-ghost);
+  }
+
+  [data-iii-ui="eval"] .eval-ui-run-row span:nth-child(4)::before {
+    content: 'time ';
+    color: var(--color-ink-ghost);
+  }
+}
diff --git a/eval/ui/tsconfig.json b/eval/ui/tsconfig.json
new file mode 100644
index 000000000..817d296b3
--- /dev/null
+++ b/eval/ui/tsconfig.json
@@ -0,0 +1,14 @@
+{
+  "compilerOptions": {
+    "target": "ES2022",
+    "lib": ["ES2022", "DOM", "DOM.Iterable"],
+    "module": "ESNext",
+    "moduleResolution": "Bundler",
+    "jsx": "react-jsx",
+    "strict": true,
+    "noEmit": true,
+    "skipLibCheck": true,
+    "allowSyntheticDefaultImports": true
+  },
+  "include": ["page.tsx", "src/**/*.ts", "src/**/*.tsx"]
+}
diff --git a/harness/Cargo.lock b/harness/Cargo.lock
index aab5745e6..9a0b8c098 100644
--- a/harness/Cargo.lock
+++ b/harness/Cargo.lock
@@ -502,7 +502,7 @@ dependencies = [
 
 [[package]]
 name = "harness"
-version = "1.6.0"
+version = "1.6.1"
 dependencies = [
  "anyhow",
  "async-trait",
diff --git a/harness/Cargo.toml b/harness/Cargo.toml
index 5315561c2..0a740f21a 100644
--- a/harness/Cargo.toml
+++ b/harness/Cargo.toml
@@ -4,7 +4,7 @@ resolver = "2"
 
 [package]
 name = "harness"
-version = "1.6.0"
+version = "1.6.1"
 edition = "2021"
 publish = false
 
diff --git a/harness/Makefile b/harness/Makefile
index 47cf0a277..075211f77 100644
--- a/harness/Makefile
+++ b/harness/Makefile
@@ -17,8 +17,10 @@ SCRAPLING_ROOT := $(WORKTREE_ROOT_ABS)/scrapling
 SCRAPLING_VENV := $(SCRAPLING_ROOT)/.venv
 SCRAPLING_STAMP := $(SCRAPLING_VENV)/.installed
 
-STACK := iii-directory llm-router session-manager context-manager \
-         provider-anthropic provider-openai provider-openai-codex shell scrapling queue harness approval-gate console
+BASE_STACK := state queue iii-directory llm-router session-manager context-manager \
+              provider-anthropic provider-openai provider-openai-codex shell scrapling cron \
+              approval-gate harness console
+STACK := $(BASE_STACK) eval
 
 # scrapling is a Python worker (pyproject + venv, see prepare-scrapling);
 # every other STACK member is a Rust crate with <worker>/Cargo.toml. Cargo
@@ -26,8 +28,8 @@ STACK := iii-directory llm-router session-manager context-manager \
 PYTHON_WORKERS := scrapling
 RUST_WORKERS   := $(filter-out $(PYTHON_WORKERS),$(STACK))
 
-GOAL_ARGS        := $(filter-out help install-iii-next build install-local clean cargo-clean dev-up dev-down dev-restart \
-                    engine wait-engine restart restart-worker stop stop-worker stop-work stop-engine status smoke logs attach \
+GOAL_ARGS        := $(filter-out help install-iii-next build install-local clean cargo-clean dev-run dev-up dev-down dev-restart \
+                    engine wait-engine wait-base wait-stack restart restart-worker stop stop-worker stop-work stop-engine status smoke logs attach \
                     ensure-session prepare-scrapling require-engine where integration-e2e integration-playground integration-validate,$(MAKECMDGOALS))
 SELECTED_WORKERS := $(strip $(GOAL_ARGS))
 ACTIVE_WORKERS   := $(if $(SELECTED_WORKERS),$(SELECTED_WORKERS),$(STACK))
@@ -38,13 +40,14 @@ RUN_FLAG   := $(if $(filter release,$(RUN_PROFILE)),--release,)
 
 .DEFAULT_GOAL := help
 
-.PHONY: help install-iii-next build install-local clean cargo-clean dev-up dev-down dev-restart engine wait-engine \
+.PHONY: help install-iii-next build install-local clean cargo-clean dev-run dev-up dev-down dev-restart engine wait-engine wait-base wait-stack \
         restart restart-worker stop stop-worker stop-work stop-engine status smoke logs attach ensure-session prepare-scrapling \
         require-engine where integration-e2e integration-playground integration-validate $(STACK)
 
 help:
 	@printf '%s\n' \
 	  'Harness local development targets:' \
+	  '  make dev-run                        start engine + harness stack in tmux' \
 	  '  make dev-up                         start engine + harness stack in tmux' \
 	  '  make dev-down                       stop the tmux stack' \
 	  '  make dev-restart                    restart all workers in the stack' \
@@ -121,8 +124,14 @@ cargo-clean:
 	  echo "skipping $$w: no cargo artifacts"; \
 	done
 
+dev-run: dev-up
+
 dev-up: engine wait-engine
-	@$(MAKE) --no-print-directory restart $(STACK)
+	@$(MAKE) --no-print-directory build PROFILE=$(RUN_PROFILE) $(STACK)
+	@$(MAKE) --no-print-directory restart $(BASE_STACK)
+	@$(MAKE) --no-print-directory wait-base
+	@$(MAKE) --no-print-directory restart eval
+	@$(MAKE) --no-print-directory wait-stack
 	@$(MAKE) --no-print-directory status
 	@echo "console: http://127.0.0.1:$(CONSOLE_PORT)"
 	@echo "tmux:    tmux attach -t $(TMUX_SESSION)"
@@ -131,7 +140,11 @@ dev-down:
 	@tmux kill-session -t "$(TMUX_SESSION)" 2>/dev/null || true
 
 dev-restart:
-	@$(MAKE) --no-print-directory restart $(STACK)
+	@$(MAKE) --no-print-directory build PROFILE=$(RUN_PROFILE) $(STACK)
+	@$(MAKE) --no-print-directory restart $(BASE_STACK)
+	@$(MAKE) --no-print-directory wait-base
+	@$(MAKE) --no-print-directory restart eval
+	@$(MAKE) --no-print-directory wait-stack
 
 ensure-session:
 	@tmux has-session -t "$(TMUX_SESSION)" 2>/dev/null || \
@@ -170,6 +183,44 @@ wait-engine:
 	echo "engine did not become ready in $(TMUX_SESSION):engine on $(ENGINE_URL)"; \
 	exit 1
 
+wait-base:
+	@i=0; \
+	while [ $$i -lt 120 ]; do \
+	  workers=$$(iii trigger engine::workers::list --port "$(ENGINE_PORT)" --json '{}' 2>/dev/null || true); \
+	  missing=""; \
+	  for worker in $(BASE_STACK); do \
+	    echo "$$workers" | grep -Fq "\"name\": \"$$worker\"" || missing="$$missing $$worker"; \
+	  done; \
+	  if [ -z "$$missing" ] && \
+	     iii trigger harness::status --help --port "$(ENGINE_PORT)" >/dev/null 2>&1 && \
+	     iii trigger console::ui-manifest --port "$(ENGINE_PORT)" --json '{}' >/dev/null 2>&1; then \
+	    echo "base stack ready in $(TMUX_SESSION)"; \
+	    exit 0; \
+	  fi; \
+	  i=$$((i + 1)); \
+	  sleep 0.5; \
+	done; \
+	echo "base stack did not become ready; missing workers:$${missing:- none}"; \
+	echo "inspect logs with: tmux attach -t $(TMUX_SESSION)"; \
+	exit 1
+
+wait-stack:
+	@i=0; \
+	while [ $$i -lt 120 ]; do \
+	  workers=$$(iii trigger engine::workers::list --port "$(ENGINE_PORT)" --json '{}' 2>/dev/null || true); \
+	  if echo "$$workers" | grep -Fq '"name": "eval"' && \
+	     iii trigger eval::status --help --port "$(ENGINE_PORT)" >/dev/null 2>&1 && \
+	     iii trigger console::ui-manifest --port "$(ENGINE_PORT)" --json '{}' 2>/dev/null | grep -Fq '"worker": "eval"'; then \
+	    echo "stack ready in $(TMUX_SESSION)"; \
+	    exit 0; \
+	  fi; \
+	  i=$$((i + 1)); \
+	  sleep 0.5; \
+	done; \
+	echo "eval did not become ready"; \
+	echo "inspect logs with: tmux attach -t $(TMUX_SESSION)"; \
+	exit 1
+
 require-engine:
 	@engine_cmd=$$(tmux display-message -p -t "$(TMUX_SESSION):engine" '#{pane_current_command}' 2>/dev/null || true); \
 	if [ "$$engine_cmd" != "iii" ]; then \
diff --git a/harness/README.md b/harness/README.md
index fc4db0dae..72c322e11 100644
--- a/harness/README.md
+++ b/harness/README.md
@@ -172,11 +172,9 @@ escape hatch.
 An optional `mode` (`ask` | `agent`) prepends a short operating-mode
 paragraph; `ask` is also enforced structurally — the dispatch policy of an
 ask-mode send (a steer's inherited one, and an ask-mode spawned child's
-resolved one) is capped at the configured read-only baseline
-(`default_functions`), so reads stay available while writes, spawning, and
-trigger registration are denied regardless of the request. The cap applies to
-a NEW turn; a steer folded into an already-running turn keeps that turn's
-frozen policy until it finalises.
+resolved one) is capped at the configured default policy (`default_functions`).
+The cap applies to a NEW turn; a steer folded into an already-running turn
+keeps that turn's frozen policy until it finalises.
 
 A non-empty `options.system_prompt` is combined with the built-in
 prompt per `options.system_prompt_strategy`: `enrich` (default) appends it to
diff --git a/harness/architecture/README.md b/harness/architecture/README.md
index 5c83ce925..10466b042 100644
--- a/harness/architecture/README.md
+++ b/harness/architecture/README.md
@@ -72,7 +72,7 @@ flowchart LR
 | **Turn** | One run of the loop for a session: one or more generate steps until the model stops, with a coarse `TurnStatus` (`running` / `awaiting_functions` / `completed` / `cancelled` / `failed`). |
 | **Step** | One durable, enqueued `harness::turn` iteration: assemble → generate → dispatch. Re-enqueued until the turn finalises. |
 | **Steering / merge** | A `harness::send` for a session that already has a running turn folds the new message into it instead of starting a second turn (the response carries `merged: true`). |
-| **Dispatch policy** | The fail-closed `options.functions.allow` / `deny` globs deciding which functions the model may call. Absent or empty `allow` → a plain chat loop. Ask-mode turns are additionally capped at the operator's read-only baseline (`default_functions`). |
+| **Dispatch policy** | The fail-closed `options.functions.allow` / `deny` globs deciding which functions the model may call. Absent or empty `allow` → a plain chat loop. Ask-mode turns are additionally capped at the configured default policy (`default_functions`). |
 | **Exposure mode** | How allowed functions reach the model: one generic `agent_trigger` schema (default) or one schema per allowed function (`native`). |
 | **Output contract** | The turn's deliverable: free `text` (default) or `json` validated against a schema; the result rides `harness::turn-completed` and `harness::run`. |
 | **Hook** | A synchronous extension point (`harness::hook::*`) a sibling binds to veto / hold / mutate in-path. Hook *logic* lives in the sibling, never the harness. |
diff --git a/harness/iii.worker.yaml b/harness/iii.worker.yaml
index 4b383d5c3..2d5ab660a 100644
--- a/harness/iii.worker.yaml
+++ b/harness/iii.worker.yaml
@@ -8,7 +8,7 @@ tags: [agent, harness, loop, autonomous]
 description: Thin durable turn loop that wires session-manager, context-manager, and llm-router into an agent loop; spawns sub-agents as child sessions.
 
 dependencies:
-  state: "^0.21.2"
+  state: "^0.21.3"
   queue: "^0.2.0"
   cron: "^0.21.0"
   configuration: "^0.21.6"
diff --git a/harness/prompts/default.txt b/harness/prompts/default.txt
index 33a9e78cf..aa17760d5 100644
--- a/harness/prompts/default.txt
+++ b/harness/prompts/default.txt
@@ -152,6 +152,19 @@ failure-key respawn below; a flaky call inside the child uses the worker's own b
 `timeout`/`retries` arguments, not a hand-rolled loop. Independent spawns go in ONE
 message; each returns its child ids instantly and the children run in parallel.
 
+Before dispatching any task, audit every resolved resource selector the child must pass.
+Write literal selectors into the task — for example, `db: "primary"` — rather than
+asking the child to discover or guess them. Use database db: "<resolved name>" (and the
+equivalent literal selector for every other worker) whenever the task calls that resource.
+Do not dispatch a task until this audit passes. A discovery child ends a child immediately
+after discovery by writing the resolved selectors for its consumer; it does not keep
+working or leave the consumer to rediscover them.
+
+For every run, derive its variable suffix from the unique session id (plus a short random
+suffix when the session id is not already unique). Before creating a state scope, table,
+or other mutable namespace, confirm the namespace is absent; never reuse a prior run's
+scope or silently append to its data.
+
 Step 3. End your turn: reply with what you wired and started, then stop. Notifications
 wake you when the watched state changes; the engine owns the sequencing. Never poll,
 never wait, and never redo work a registered reaction owns.
@@ -161,9 +174,8 @@ job plus a few random characters for uniqueness, e.g. `fetch-headlines-b4k9`. Ne
 with your own session id. Omitted, the engine mints an opaque UUID row in the console; a slug
 without the random suffix can collide with an earlier run and silently resume that session,
 old transcript and all. A react trigger's `metadata.session_id` has its own rule (Reacting
-section below): pin a fresh unique id for each stage that must run as its own child; omit it
-only when the reaction should deliver back into the registering chat — and never pin a fixed
-id on a recurring trigger, which funnels every firing into one session.
+section below): ordinary event reactions omit it so each event gets a fresh child; set it only
+when deliberate session reuse is required, and never pin a fixed id on a recurring trigger.
 
 By default an in-turn child INHERITS your full policy — the same permissions you hold. You
 can never grant a child MORE than you have (spawns never escalate), and you may narrow one
@@ -194,16 +206,11 @@ to `harness::react` and put the sub-agent you want in the trigger's `metadata`:
                   parent_session_id: "<your session id — the console root for the child>" }
     }
 
-`metadata.session_id` picks WHICH session the reacting sub-agent runs in — omitting it does
-NOT create a fresh distinct child. It falls back to the session that REGISTERED this trigger,
-so the reaction fires back into THAT chat every time (fine for a pipeline's last stage,
-deliberately delivering the final result "back here" — wrong for any EARLIER stage meant to
-run as its own child). Any stage you want spawned as a distinct sub-agent — including each
-branch of a fan-out like "two parallel analysts" — needs its OWN explicit `session_id`, picked
-by you, unique to this run (same discipline as a direct `harness::spawn` call and as the Join
-section's predecessor ids below): a readable slug plus a few random characters, e.g.
-`summarizer-b4k9`.
-Reusing an id from an earlier run silently RESUMES that old session instead of starting fresh.
+For an ordinary event reaction, you MUST omit `metadata.session_id`: every event creates a
+fresh distinct child under the resolved console root. Set `metadata.session_id` only when a
+stage deliberately needs to reuse a known session; an id from an earlier run silently
+resumes that old session instead of starting fresh. `metadata.parent_session_id` controls
+where the fresh child appears in the console tree and does not change its delivery session.
 
 `metadata.parent_session_id` pins where the reacting sub-agent nests in the console tree
 (unrelated to which session it runs in). It MUST be a REAL session id — normally your own. An
@@ -238,8 +245,10 @@ ALWAYS a plain notify (`engine::register_trigger` with NO `function_id`); bindin
 wake to a react/call edge reads into the void and strands the run.
 
 `harness::react` simple edges are one-shot by default (except cron, which is recurring).
-Set `once: false` only for a deliberate standing watcher. Join predecessors ignore `once`;
-the join owns their lifecycle.
+Join predecessor edges are standing by default; the join owns their lifecycle. Set `once: true`
+only for a deliberate one-shot reaction on a cron edge. `once` is a TOP-LEVEL
+registration field — putting it inside `metadata` does nothing. Set `once: false` only for
+a deliberate standing non-join watcher.
 
 `harness::react` spawns a sub-agent (`harness::spawn`) with your `task` (the event JSON appended
 so it sees what fired — a `turn-completed` event carries the turn's `status` and, when it
@@ -375,10 +384,10 @@ Pick the validator KIND by the rule:
   it at a scratch subfield (`into: "/value/_seen"`) or the literal `value` you meant to
   write is silently overwritten.
 - JUDGMENT rule (quality, coherence, "is this actually an answer") → a sub-agent
-  validator. Pin it into its OWN session — `metadata.session_id:
-  "validator-<run>-<rand>"`, with `metadata.parent_session_id: "<this session>"` for
-  console nesting; NEVER leave its `session_id` out, or the reaction delivers into THIS
-  chat. Grant it `metadata.options: { functions: { allow: ["state::get", "state::set"] } }`.
+  validator. Omit `metadata.session_id` so each validation event gets a fresh child, and
+  set `metadata.parent_session_id: "<this session>"` when you want console nesting. Pin
+  `metadata.session_id` only when a validator must deliberately reuse a known session.
+  Grant it `metadata.options: { functions: { allow: ["state::get", "state::set"] } }`.
   Its task is self-contained, e.g.: "For each of <scope>/<key1>, <scope>/<key2>:
   `state::get` it, expecting `{ ok, value | error }`. All present with ok true →
   `state::set` <scope>/turn_complete = { done: true, keys: [...], summary: '<one line>' }.
diff --git a/harness/src/budget.rs b/harness/src/budget.rs
new file mode 100644
index 000000000..f69f9281b
--- /dev/null
+++ b/harness/src/budget.rs
@@ -0,0 +1,519 @@
+//! Durable hard-budget ledger shared by a root harness session and its
+//! in-turn sub-agents.
+//!
+//! Each generation reserves its worst-case input/output usage before calling
+//! the router, then reconciles the reservation with the provider's actual
+//! usage. The existing per-key in-process lock serializes sibling sessions
+//! against the same root ledger; the state worker makes the counters durable.
+
+use serde::{Deserialize, Serialize};
+
+use crate::deps::Deps;
+use crate::error::HarnessError;
+use crate::types::event::Usage;
+use crate::types::model::Pricing;
+use crate::types::turn::{TurnOptions, TurnRecord};
+
+const BUDGET_SCOPE: &str = "harness_budget";
+const MICRO_USD_PER_USD: f64 = 1_000_000.0;
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+struct BudgetState {
+    root_session_id: String,
+    max_total_tokens: Option<u64>,
+    used_tokens: u64,
+    reserved_tokens: u64,
+    max_cost_microusd: Option<u64>,
+    used_cost_microusd: u64,
+    reserved_cost_microusd: u64,
+}
+
+impl BudgetState {
+    fn new(root_session_id: &str, options: &TurnOptions) -> Result<Self, HarnessError> {
+        Ok(Self {
+            root_session_id: root_session_id.to_string(),
+            max_total_tokens: options.max_total_tokens,
+            used_tokens: 0,
+            reserved_tokens: 0,
+            max_cost_microusd: options.max_cost_usd.map(cost_limit_microusd).transpose()?,
+            used_cost_microusd: 0,
+            reserved_cost_microusd: 0,
+        })
+    }
+
+    fn verify_limits(&self, options: &TurnOptions) -> Result<(), HarnessError> {
+        let requested_cost = options.max_cost_usd.map(cost_limit_microusd).transpose()?;
+        if self.max_total_tokens != options.max_total_tokens
+            || self.max_cost_microusd != requested_cost
+        {
+            return Err(HarnessError::InvalidRequest(
+                "execution budgets are frozen when a root session starts and cannot be changed"
+                    .into(),
+            ));
+        }
+        Ok(())
+    }
+
+    fn reserve(
+        &mut self,
+        requested_tokens: u64,
+        requested_cost_microusd: u64,
+    ) -> Result<(), BudgetRejection> {
+        if let Some(limit) = self.max_total_tokens {
+            let committed = self.used_tokens.saturating_add(self.reserved_tokens);
+            if committed.saturating_add(requested_tokens) > limit {
+                return Err(BudgetRejection::Exceeded(format!(
+                    "generation requires up to {requested_tokens} tokens, but only {} remain in the shared {limit}-token budget",
+                    limit.saturating_sub(committed)
+                )));
+            }
+        }
+        if let Some(limit) = self.max_cost_microusd {
+            let committed = self
+                .used_cost_microusd
+                .saturating_add(self.reserved_cost_microusd);
+            if committed.saturating_add(requested_cost_microusd) > limit {
+                return Err(BudgetRejection::Exceeded(format!(
+                    "generation requires up to {}, but only {} remain in the shared {} budget",
+                    display_microusd(requested_cost_microusd),
+                    display_microusd(limit.saturating_sub(committed)),
+                    display_microusd(limit)
+                )));
+            }
+        }
+        self.reserved_tokens = self.reserved_tokens.saturating_add(requested_tokens);
+        self.reserved_cost_microusd = self
+            .reserved_cost_microusd
+            .saturating_add(requested_cost_microusd);
+        Ok(())
+    }
+
+    fn reconcile(
+        &mut self,
+        reservation: &BudgetReservation,
+        actual_tokens: u64,
+        actual_cost_microusd: u64,
+    ) {
+        self.reserved_tokens = self
+            .reserved_tokens
+            .saturating_sub(reservation.reserved_tokens);
+        self.reserved_cost_microusd = self
+            .reserved_cost_microusd
+            .saturating_sub(reservation.reserved_cost_microusd);
+        self.used_tokens = self.used_tokens.saturating_add(actual_tokens);
+        self.used_cost_microusd = self.used_cost_microusd.saturating_add(actual_cost_microusd);
+    }
+
+    fn release(&mut self, reservation: &BudgetReservation) {
+        self.reserved_tokens = self
+            .reserved_tokens
+            .saturating_sub(reservation.reserved_tokens);
+        self.reserved_cost_microusd = self
+            .reserved_cost_microusd
+            .saturating_sub(reservation.reserved_cost_microusd);
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct BudgetReservation {
+    root_session_id: String,
+    reserved_tokens: u64,
+    reserved_cost_microusd: u64,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum BudgetRejection {
+    Exceeded(String),
+    Unavailable(String),
+}
+
+impl BudgetRejection {
+    pub fn reason(&self) -> &str {
+        match self {
+            Self::Exceeded(reason) | Self::Unavailable(reason) => reason,
+        }
+    }
+}
+
+#[derive(Debug, Clone)]
+pub enum ReserveOutcome {
+    Unlimited,
+    Reserved(BudgetReservation),
+    Rejected(BudgetRejection),
+}
+
+/// Validate and freeze a root session's requested budget. Existing sessions
+/// inherit their first turn's ledger; they cannot acquire a budget after
+/// unmetered work has already happened or replace one mid-session.
+pub async fn prepare_root(
+    deps: &Deps,
+    session_id: &str,
+    options: &mut TurnOptions,
+    previous: Option<&TurnRecord>,
+) -> Result<(), HarnessError> {
+    validate_options(options)?;
+
+    if let Some(previous) = previous {
+        let previous_budgeted = previous.options.budget_root_session_id.is_some();
+        let requested_budgeted =
+            options.max_total_tokens.is_some() || options.max_cost_usd.is_some();
+
+        if !previous_budgeted {
+            if requested_budgeted {
+                return Err(HarnessError::InvalidRequest(
+                    "cannot add an execution budget to an existing unmetered session; start a new session"
+                        .into(),
+                ));
+            }
+            return Ok(());
+        }
+
+        if options
+            .max_total_tokens
+            .is_some_and(|value| Some(value) != previous.options.max_total_tokens)
+            || options.max_cost_usd.is_some_and(|value| {
+                cost_limit_microusd(value).ok()
+                    != previous
+                        .options
+                        .max_cost_usd
+                        .and_then(|prior| cost_limit_microusd(prior).ok())
+            })
+        {
+            return Err(HarnessError::InvalidRequest(
+                "execution budgets are frozen when a root session starts and cannot be changed"
+                    .into(),
+            ));
+        }
+
+        options.max_total_tokens = previous.options.max_total_tokens;
+        options.max_cost_usd = previous.options.max_cost_usd;
+        options.budget_root_session_id = previous.options.budget_root_session_id.clone();
+        let root = options.budget_root_session_id.as_deref().ok_or_else(|| {
+            HarnessError::State("budgeted turn is missing its root session id".into())
+        })?;
+        let _guard = deps.locks.guard(&budget_lock_key(root)).await;
+        let state = load_state(deps, root).await?.ok_or_else(|| {
+            HarnessError::State(format!("shared budget ledger {root} is missing"))
+        })?;
+        state.verify_limits(options)?;
+        return Ok(());
+    }
+
+    if options.max_total_tokens.is_none() && options.max_cost_usd.is_none() {
+        return Ok(());
+    }
+
+    options.budget_root_session_id = Some(session_id.to_string());
+    let _guard = deps.locks.guard(&budget_lock_key(session_id)).await;
+    match load_state(deps, session_id).await? {
+        Some(state) => state.verify_limits(options),
+        None => save_state(deps, &BudgetState::new(session_id, options)?).await,
+    }
+}
+
+/// Reserve one generation's maximum possible charge. Cost budgets fail closed
+/// when the selected model lacks usable catalog pricing.
+pub async fn reserve(
+    deps: &Deps,
+    record: &TurnRecord,
+    input_tokens: u64,
+    output_tokens: u64,
+) -> Result<ReserveOutcome, HarnessError> {
+    let Some(root) = record.options.budget_root_session_id.as_deref() else {
+        return Ok(ReserveOutcome::Unlimited);
+    };
+    let requested_tokens = input_tokens.saturating_add(output_tokens);
+    let requested_cost_microusd = if record.options.max_cost_usd.is_some() {
+        let model = deps
+            .router()
+            .await
+            .models_get(record.options.provider.as_deref(), &record.options.model)
+            .await;
+        let Some(pricing) = model.and_then(|model| model.pricing) else {
+            return Ok(ReserveOutcome::Rejected(BudgetRejection::Unavailable(
+                format!(
+                    "cannot enforce max_cost_usd for model {} because no pricing is configured in \
+                     the model catalog; configure input/output pricing or remove max_cost_usd",
+                    record.options.model
+                ),
+            )));
+        };
+        match maximum_cost_microusd(input_tokens, output_tokens, &pricing) {
+            Ok(cost) => cost,
+            Err(reason) => {
+                return Ok(ReserveOutcome::Rejected(BudgetRejection::Unavailable(
+                    reason,
+                )))
+            }
+        }
+    } else {
+        0
+    };
+
+    let _guard = deps.locks.guard(&budget_lock_key(root)).await;
+    let mut state = load_state(deps, root)
+        .await?
+        .ok_or_else(|| HarnessError::State(format!("shared budget ledger {root} is missing")))?;
+    state.verify_limits(&record.options)?;
+    match state.reserve(requested_tokens, requested_cost_microusd) {
+        Ok(()) => {
+            save_state(deps, &state).await?;
+            Ok(ReserveOutcome::Reserved(BudgetReservation {
+                root_session_id: root.to_string(),
+                reserved_tokens: requested_tokens,
+                reserved_cost_microusd: requested_cost_microusd,
+            }))
+        }
+        Err(rejection) => Ok(ReserveOutcome::Rejected(rejection)),
+    }
+}
+
+/// Release a reservation when generation never reached the provider.
+pub async fn release(deps: &Deps, reservation: &BudgetReservation) -> Result<(), HarnessError> {
+    let root = &reservation.root_session_id;
+    let _guard = deps.locks.guard(&budget_lock_key(root)).await;
+    let mut state = load_state(deps, root)
+        .await?
+        .ok_or_else(|| HarnessError::State(format!("shared budget ledger {root} is missing")))?;
+    state.release(reservation);
+    save_state(deps, &state).await
+}
+
+/// Replace a worst-case reservation with actual provider usage. Missing or
+/// partial usage consumes the full reservation so a provider cannot silently
+/// bypass the hard limit.
+pub async fn reconcile(
+    deps: &Deps,
+    reservation: &BudgetReservation,
+    usage: Option<&Usage>,
+) -> Result<(), HarnessError> {
+    let actual_tokens = usage
+        .and_then(|usage| match (usage.input, usage.output) {
+            (Some(input), Some(output)) => Some(input.saturating_add(output)),
+            _ => None,
+        })
+        .unwrap_or(reservation.reserved_tokens);
+    let actual_cost_microusd = usage
+        .and_then(|usage| usage.cost_usd)
+        .and_then(|cost| observed_cost_microusd(cost).ok())
+        .unwrap_or(reservation.reserved_cost_microusd);
+
+    let root = &reservation.root_session_id;
+    let _guard = deps.locks.guard(&budget_lock_key(root)).await;
+    let mut state = load_state(deps, root)
+        .await?
+        .ok_or_else(|| HarnessError::State(format!("shared budget ledger {root} is missing")))?;
+    state.reconcile(reservation, actual_tokens, actual_cost_microusd);
+    save_state(deps, &state).await
+}
+
+pub async fn purge(
+    deps: &Deps,
+    root_session_id: &str,
+    timeout_ms: u64,
+) -> Result<(), HarnessError> {
+    let _guard = deps.locks.guard(&budget_lock_key(root_session_id)).await;
+    crate::state::state_delete(&deps.iii, BUDGET_SCOPE, root_session_id, timeout_ms).await
+}
+
+fn validate_options(options: &TurnOptions) -> Result<(), HarnessError> {
+    if options.max_total_tokens == Some(0) {
+        return Err(HarnessError::InvalidRequest(
+            "max_total_tokens must be greater than zero".into(),
+        ));
+    }
+    if let Some(cost) = options.max_cost_usd {
+        cost_limit_microusd(cost)?;
+    }
+    Ok(())
+}
+
+fn cost_limit_microusd(cost_usd: f64) -> Result<u64, HarnessError> {
+    if !cost_usd.is_finite() || cost_usd < 0.0 {
+        return Err(HarnessError::InvalidRequest(
+            "max_cost_usd must be a finite non-negative number".into(),
+        ));
+    }
+    let scaled = cost_usd * MICRO_USD_PER_USD;
+    if scaled > u64::MAX as f64 {
+        return Err(HarnessError::InvalidRequest(
+            "max_cost_usd is too large".into(),
+        ));
+    }
+    Ok(scaled.floor() as u64)
+}
+
+fn observed_cost_microusd(cost_usd: f64) -> Result<u64, HarnessError> {
+    if !cost_usd.is_finite() || cost_usd < 0.0 {
+        return Err(HarnessError::State(
+            "provider returned an invalid cost_usd".into(),
+        ));
+    }
+    let scaled = cost_usd * MICRO_USD_PER_USD;
+    if scaled > u64::MAX as f64 {
+        return Err(HarnessError::State("provider cost_usd is too large".into()));
+    }
+    Ok(scaled.ceil() as u64)
+}
+
+fn maximum_cost_microusd(
+    input_tokens: u64,
+    output_tokens: u64,
+    pricing: &Pricing,
+) -> Result<u64, String> {
+    let base_input_rate = pricing
+        .input
+        .filter(|rate| rate.is_finite() && *rate >= 0.0)
+        .ok_or_else(|| "model catalog has no valid input pricing".to_string())?;
+    let input_rate = [pricing.cache_read, pricing.cache_write]
+        .into_iter()
+        .flatten()
+        .try_fold(None::<f64>, |maximum, rate| {
+            if !rate.is_finite() || rate < 0.0 {
+                return Err("model catalog contains invalid input pricing".to_string());
+            }
+            Ok(Some(maximum.map_or(rate, |current| current.max(rate))))
+        })?
+        .map_or(base_input_rate, |cached_rate| {
+            base_input_rate.max(cached_rate)
+        });
+    let output_rate = pricing
+        .output
+        .filter(|rate| rate.is_finite() && *rate >= 0.0)
+        .ok_or_else(|| "model catalog has no valid output pricing".to_string())?;
+
+    // Catalog rates are USD per million tokens. Numerically, token * rate is
+    // therefore already micro-USD.
+    let cost = input_tokens as f64 * input_rate + output_tokens as f64 * output_rate;
+    if !cost.is_finite() || cost > u64::MAX as f64 {
+        return Err("model pricing produces an unrepresentable cost".into());
+    }
+    Ok(cost.ceil() as u64)
+}
+
+fn display_microusd(value: u64) -> String {
+    format!("${:.6}", value as f64 / MICRO_USD_PER_USD)
+}
+
+fn budget_lock_key(root_session_id: &str) -> String {
+    format!("budget:{root_session_id}")
+}
+
+async fn load_state(
+    deps: &Deps,
+    root_session_id: &str,
+) -> Result<Option<BudgetState>, HarnessError> {
+    let cfg = deps.cfg().await;
+    let value = crate::state::state_get(
+        &deps.iii,
+        BUDGET_SCOPE,
+        root_session_id,
+        cfg.session_timeout_ms,
+    )
+    .await?;
+    if value.is_null() {
+        return Ok(None);
+    }
+    serde_json::from_value(value)
+        .map(Some)
+        .map_err(|error| HarnessError::State(format!("budget ledger parse: {error}")))
+}
+
+async fn save_state(deps: &Deps, state: &BudgetState) -> Result<(), HarnessError> {
+    let cfg = deps.cfg().await;
+    let value = serde_json::to_value(state)
+        .map_err(|error| HarnessError::State(format!("budget ledger serialize: {error}")))?;
+    crate::state::state_set(
+        &deps.iii,
+        BUDGET_SCOPE,
+        &state.root_session_id,
+        value,
+        cfg.session_timeout_ms,
+    )
+    .await
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn state(tokens: Option<u64>, cost_microusd: Option<u64>) -> BudgetState {
+        BudgetState {
+            root_session_id: "root".into(),
+            max_total_tokens: tokens,
+            used_tokens: 0,
+            reserved_tokens: 0,
+            max_cost_microusd: cost_microusd,
+            used_cost_microusd: 0,
+            reserved_cost_microusd: 0,
+        }
+    }
+
+    #[test]
+    fn concurrent_reservations_cannot_exceed_the_shared_limit() {
+        let mut ledger = state(Some(100), Some(1_000));
+        assert!(ledger.reserve(60, 700).is_ok());
+        assert!(matches!(
+            ledger.reserve(41, 100),
+            Err(BudgetRejection::Exceeded(_))
+        ));
+        assert!(matches!(
+            ledger.reserve(20, 301),
+            Err(BudgetRejection::Exceeded(_))
+        ));
+    }
+
+    #[test]
+    fn reconciliation_replaces_reserved_with_actual_usage() {
+        let mut ledger = state(Some(100), Some(1_000));
+        ledger.reserve(60, 700).unwrap();
+        let reservation = BudgetReservation {
+            root_session_id: "root".into(),
+            reserved_tokens: 60,
+            reserved_cost_microusd: 700,
+        };
+        ledger.reconcile(&reservation, 25, 300);
+        assert_eq!(ledger.reserved_tokens, 0);
+        assert_eq!(ledger.reserved_cost_microusd, 0);
+        assert_eq!(ledger.used_tokens, 25);
+        assert_eq!(ledger.used_cost_microusd, 300);
+        assert!(ledger.reserve(75, 700).is_ok());
+    }
+
+    #[test]
+    fn pricing_reserves_the_most_expensive_input_class() {
+        let pricing = Pricing {
+            input: Some(2.0),
+            output: Some(10.0),
+            cache_read: Some(0.2),
+            cache_write: Some(2.5),
+        };
+        assert_eq!(maximum_cost_microusd(1_000, 100, &pricing), Ok(3_500));
+    }
+
+    #[test]
+    fn pricing_fails_closed_without_uncached_input_or_output_rates() {
+        let no_input = Pricing {
+            cache_read: Some(0.2),
+            output: Some(10.0),
+            ..Pricing::default()
+        };
+        assert!(maximum_cost_microusd(1_000, 100, &no_input)
+            .unwrap_err()
+            .contains("input pricing"));
+
+        let no_output = Pricing {
+            input: Some(2.0),
+            ..Pricing::default()
+        };
+        assert!(maximum_cost_microusd(1_000, 100, &no_output)
+            .unwrap_err()
+            .contains("output pricing"));
+    }
+
+    #[test]
+    fn configured_cost_rounds_down_and_observed_cost_rounds_up() {
+        assert_eq!(cost_limit_microusd(0.000_001_9), Ok(1));
+        assert_eq!(observed_cost_microusd(0.000_001_1), Ok(2));
+    }
+}
diff --git a/harness/src/clients/router.rs b/harness/src/clients/router.rs
index d49bcf67b..3344a9c9c 100644
--- a/harness/src/clients/router.rs
+++ b/harness/src/clients/router.rs
@@ -41,6 +41,7 @@ pub struct ChatParams {
     pub messages: Vec<Value>,
     pub tools: Vec<AgentFunction>,
     pub response_format: Option<Value>,
+    pub max_output_tokens: Option<u64>,
     pub thinking_level: Option<ThinkingLevel>,
     pub provider_options: Option<Value>,
 }
@@ -137,6 +138,9 @@ impl RouterClient {
         if let Some(rf) = &params.response_format {
             payload["response_format"] = rf.clone();
         }
+        if let Some(max_output_tokens) = params.max_output_tokens {
+            payload["max_output_tokens"] = json!(max_output_tokens);
+        }
         if let Some(tl) = &params.thinking_level {
             payload["thinking_level"] = serde_json::to_value(tl).unwrap_or(Value::Null);
         }
diff --git a/harness/src/clients/session.rs b/harness/src/clients/session.rs
index 5f9239f26..e9f7f848c 100644
--- a/harness/src/clients/session.rs
+++ b/harness/src/clients/session.rs
@@ -1,6 +1,7 @@
 //! `session-manager` client: ensure/create a session, append and update
 //! messages, set status, and load the active path. Required dependency.
 
+use std::collections::BTreeSet;
 use std::sync::Arc;
 
 use iii_sdk::protocol::TriggerRequest;
@@ -29,6 +30,35 @@ pub struct LoadedCustom {
     pub data: Value,
 }
 
+/// The durable parent linkage stored in `SessionMeta.metadata` when the
+/// harness creates a child session.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct SessionLink {
+    pub session_id: String,
+    pub parent_session_id: String,
+    pub parent_turn_id: Option<String>,
+}
+
+#[derive(Debug, Clone, Deserialize)]
+struct LoadedSessionMeta {
+    session_id: String,
+    #[serde(default)]
+    metadata: Option<serde_json::Map<String, Value>>,
+}
+
+#[derive(Debug, Clone, Deserialize)]
+struct GetSessionResponse {
+    #[serde(rename = "meta")]
+    _meta: LoadedSessionMeta,
+}
+
+#[derive(Debug, Clone, Deserialize)]
+struct ListSessionsResponse {
+    sessions: Vec<LoadedSessionMeta>,
+    #[serde(default)]
+    next_cursor: Option<String>,
+}
+
 #[derive(Clone)]
 pub struct SessionClient {
     iii: Arc<IIIClient>,
@@ -64,6 +94,84 @@ impl SessionClient {
         (!title.is_empty()).then_some(title)
     }
 
+    /// Whether a durable session metadata record exists.
+    pub async fn exists(&self, session_id: &str) -> Result<bool, HarnessError> {
+        let response = self
+            .call("session::get", json!({ "session_id": session_id }))
+            .await?;
+        if response.is_null() {
+            return Ok(false);
+        }
+        serde_json::from_value::<GetSessionResponse>(response)
+            .map(|_| true)
+            .map_err(|error| {
+                HarnessError::Dependency(format!(
+                    "session::get returned malformed metadata for {session_id}: {error}"
+                ))
+            })
+    }
+
+    /// Load every durable session whose metadata points at `parent_session_id`.
+    /// Pagination is strict: a repeated cursor is treated as incomplete
+    /// evidence instead of looping or silently truncating the tree.
+    pub async fn children(
+        &self,
+        parent_session_id: &str,
+    ) -> Result<Vec<SessionLink>, HarnessError> {
+        const PAGE_LIMIT: u64 = 500;
+        let mut out = Vec::new();
+        let mut cursor: Option<String> = None;
+        let mut seen_cursors = BTreeSet::new();
+
+        loop {
+            let mut payload = json!({
+                "limit": PAGE_LIMIT,
+                "order": "created_asc",
+                "metadata": { "parent_session_id": parent_session_id },
+            });
+            if let Some(value) = &cursor {
+                payload["cursor"] = json!(value);
+            }
+            let response = self.call("session::list", payload).await?;
+            let page: ListSessionsResponse = serde_json::from_value(response).map_err(|error| {
+                HarnessError::Dependency(format!(
+                    "session::list returned malformed child metadata for {parent_session_id}: {error}"
+                ))
+            })?;
+            for meta in page.sessions {
+                let Some(metadata) = meta.metadata else {
+                    continue;
+                };
+                let Some(parent) = metadata.get("parent_session_id").and_then(Value::as_str) else {
+                    continue;
+                };
+                out.push(SessionLink {
+                    session_id: meta.session_id,
+                    parent_session_id: parent.to_string(),
+                    parent_turn_id: metadata
+                        .get("parent_turn_id")
+                        .and_then(Value::as_str)
+                        .map(str::to_string),
+                });
+            }
+
+            match page.next_cursor.filter(|value| !value.is_empty()) {
+                Some(next) => {
+                    if !seen_cursors.insert(next.clone()) {
+                        return Err(HarnessError::Dependency(format!(
+                            "session::list repeated cursor while loading children of {parent_session_id}"
+                        )));
+                    }
+                    cursor = Some(next);
+                }
+                None => break,
+            }
+        }
+
+        out.sort_by(|left, right| left.session_id.cmp(&right.session_id));
+        Ok(out)
+    }
+
     /// Idempotently ensure a session exists, applying `metadata` on creation.
     /// Returns whether this call CREATED the session — `false` means it already
     /// existed and the supplied metadata (e.g. parent linkage) was NOT applied.
@@ -176,6 +284,7 @@ impl SessionClient {
         session_id: &str,
         entry_id: &str,
         content: &[ContentBlock],
+        usage: Option<&crate::types::event::Usage>,
         details: Option<&Value>,
         origin: Option<&Value>,
     ) -> Result<u64, HarnessError> {
@@ -187,6 +296,11 @@ impl SessionClient {
         if let Some(d) = details {
             payload["details"] = d.clone();
         }
+        if let Some(u) = usage {
+            payload["usage"] = serde_json::to_value(u).map_err(|error| {
+                HarnessError::Internal(format!("serialize assistant usage: {error}"))
+            })?;
+        }
         if let Some(o) = origin {
             payload["origin"] = o.clone();
         }
@@ -247,6 +361,65 @@ impl SessionClient {
         Ok(out)
     }
 
+    /// Strict evidence reader for evaluation metrics. Unlike [`Self::messages`],
+    /// malformed entries fail the read instead of being skipped.
+    pub async fn messages_strict(
+        &self,
+        session_id: &str,
+    ) -> Result<Vec<LoadedEntry>, HarnessError> {
+        const PAGE_LIMIT: u64 = 500;
+        let mut out = Vec::new();
+        let mut cursor: Option<String> = None;
+        let mut seen_cursors = BTreeSet::new();
+
+        loop {
+            let mut payload = json!({
+                "session_id": session_id,
+                "include_custom": false,
+                "limit": PAGE_LIMIT,
+            });
+            if let Some(value) = &cursor {
+                payload["cursor"] = json!(value);
+            }
+            let response = self.call("session::messages", payload).await?;
+            let messages = response
+                .get("messages")
+                .and_then(Value::as_array)
+                .ok_or_else(|| {
+                    HarnessError::Dependency(format!(
+                        "session::messages returned no messages array for {session_id}"
+                    ))
+                })?;
+            for item in messages {
+                out.push(serde_json::from_value::<LoadedEntry>(item.clone()).map_err(
+                    |error| {
+                        HarnessError::Dependency(format!(
+                            "session::messages returned malformed evidence for {session_id}: {error}"
+                        ))
+                    },
+                )?);
+            }
+
+            match response
+                .get("next_cursor")
+                .and_then(Value::as_str)
+                .filter(|value| !value.is_empty())
+            {
+                Some(next) => {
+                    if !seen_cursors.insert(next.to_string()) {
+                        return Err(HarnessError::Dependency(format!(
+                            "session::messages repeated cursor while loading {session_id}"
+                        )));
+                    }
+                    cursor = Some(next.to_string());
+                }
+                None => break,
+            }
+        }
+
+        Ok(out)
+    }
+
     /// Load only the entries strictly after `after_entry_id` on the active
     /// path — the incremental form of [`Self::messages`] for callers holding
     /// a watermark. `Ok(None)` means the watermark left the active path
diff --git a/harness/src/config.rs b/harness/src/config.rs
index eaf3d314f..63db84270 100644
--- a/harness/src/config.rs
+++ b/harness/src/config.rs
@@ -81,11 +81,10 @@ pub struct WorkerConfig {
     /// Dispatch policy for a PARENTLESS spawn (a direct/CLI `harness::spawn`
     /// or a trigger-fired `harness::react` spawn) whose request carries no
     /// `options.functions`. Children of live turns still inherit/subset the
-    /// parent policy, and explicit options always win. Defaults to a
-    /// read-only baseline (discovery, reads, subscription management — no
-    /// writes, no spend, no spawning); set to `null` explicitly to restore
-    /// deny-all. Doubles as the ceiling every ask-mode turn's policy is
-    /// clamped to — `null` makes ask mode a plain chat loop.
+    /// parent policy, and explicit options always win. Defaults to allowing
+    /// every function; set to `null` explicitly to restore deny-all. Doubles
+    /// as the ceiling every ask-mode turn's policy is clamped to — `null`
+    /// makes ask mode a plain chat loop.
     #[serde(default = "default_functions")]
     pub default_functions: Option<FunctionPolicy>,
 
@@ -243,48 +242,16 @@ fn default_sweep_expression() -> String {
     "0 0 0 * * *".to_string()
 }
 fn default_functions() -> Option<FunctionPolicy> {
-    // Read-only baseline for PARENTLESS spawns (direct/CLI/trigger-fired) —
-    // the only children with no parent policy to inherit. Discovery and reads
-    // only: excludes every write surface, router spend, spawning, and trigger
-    // registration; a caller grants anything more explicitly via options /
-    // react metadata.options. (In-turn children instead inherit their parent's
-    // full policy — see `subagent::seed_child`.)
+    // Default policy for PARENTLESS spawns that carry no inherited policy
+    // (direct/CLI/raw trigger calls). In-turn children inherit their parent
+    // directly; harness::react persists the registering turn's policy into
+    // its parentless spawn options.
     //
     // DOUBLES AS THE ASK-MODE CEILING: every ask-mode send / steer / child has
-    // its dispatch policy clamped to this list (see `functions::send::
-    // clamp_for_mode`). Editing the allow set below widens or narrows what
-    // EVERY ask-mode turn may call, not just parentless spawns — keep it to
-    // non-mutating ids. Prefer exact ids over globs here: `clamp_policy`
-    // under-approximates against glob baseline entries (see its doc).
+    // its dispatch policy clamped to this policy (see
+    // `functions::send::clamp_for_mode`).
     Some(FunctionPolicy {
-        allow: [
-            "engine::functions::list",
-            "engine::functions::info",
-            "engine::triggers::list",
-            "engine::triggers::info",
-            "engine::workers::list",
-            "engine::workers::info",
-            "engine::registered-triggers::list",
-            "engine::registered-triggers::info",
-            "state::get",
-            "state::list",
-            "router::models::list",
-            "router::models::get",
-            "router::models::supports",
-            "harness::status",
-            "worker::list",
-            "directory::registry::workers::list",
-            "directory::registry::workers::info",
-            "coder::info",
-            "coder::read-file",
-            "coder::search",
-            "coder::list-folder",
-            "coder::tree",
-            "scrapling::fetch",
-        ]
-        .into_iter()
-        .map(String::from)
-        .collect(),
+        allow: vec!["*".to_string()],
         deny: vec![],
         expose: Default::default(),
     })
@@ -353,28 +320,11 @@ mod tests {
     }
 
     #[test]
-    fn default_functions_is_read_only_baseline_and_nullable() {
+    fn default_functions_allows_everything_and_is_nullable() {
         let cfg = WorkerConfig::from_json(&serde_json::json!({})).unwrap();
-        let policy = cfg.default_functions.expect("baseline present by default");
-        assert!(policy
-            .allow
-            .contains(&"engine::functions::list".to_string()));
-        assert!(policy.allow.contains(&"state::get".to_string()));
-        // No write surface, no spend, no spawning, no trigger registration in
-        // the baseline — children are leaves.
-        for denied in [
-            "state::set",
-            "harness::spawn",
-            "engine::register_trigger",
-            "engine::unregister_trigger",
-            "router::chat",
-            "shell::exec",
-        ] {
-            assert!(
-                !policy.allow.contains(&denied.to_string()),
-                "{denied} must not be in the baseline"
-            );
-        }
+        let policy = cfg.default_functions.expect("policy present by default");
+        assert_eq!(policy.allow, vec!["*"]);
+        assert!(policy.deny.is_empty());
         // Explicit null restores deny-all.
         let cfg =
             WorkerConfig::from_json(&serde_json::json!({ "default_functions": null })).unwrap();
diff --git a/harness/src/functions/metrics.rs b/harness/src/functions/metrics.rs
new file mode 100644
index 000000000..a7e047897
--- /dev/null
+++ b/harness/src/functions/metrics.rs
@@ -0,0 +1,567 @@
+//! `harness::metrics` — aggregate durable model usage and function outcomes
+//! plus trace/span observability over one complete root-and-descendant session
+//! tree.
+
+use std::collections::{BTreeSet, HashMap};
+
+use iii_sdk::protocol::TriggerRequest;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+use serde_json::json;
+
+use crate::deps::Deps;
+use crate::error::HarnessError;
+use crate::types::content::ContentBlock;
+use crate::types::message::AgentMessage;
+
+use super::session_tree::{self, SessionTreeNodeV1};
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+pub struct SessionMetricsRequestV1 {
+    pub root_session_id: String,
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct SessionMetricsResponseV1 {
+    pub root_session_id: String,
+    pub complete: bool,
+    pub totals: SessionUsageTotalsV1,
+    pub by_session: Vec<SessionUsageV1>,
+    /// Trace/span aggregates when the engine's in-memory observability exporter
+    /// is available. Usage metrics remain available when it is not.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub traces: Option<SessionTraceMetricsV1>,
+}
+
+#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct SessionUsageTotalsV1 {
+    pub sessions: u64,
+    pub turns: u64,
+    pub function_calls: u64,
+    pub function_call_errors: u64,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub input_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub output_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub cache_read_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub cache_write_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub reasoning_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub cost_usd: Option<f64>,
+}
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct SessionUsageV1 {
+    pub session_id: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub parent_session_id: Option<String>,
+    pub depth: u32,
+    pub turns: u64,
+    pub function_calls: u64,
+    pub function_call_errors: u64,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub input_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub output_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub cache_read_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub cache_write_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub reasoning_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub cost_usd: Option<f64>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct SessionTraceMetricsV1 {
+    /// Distinct traces across the root session and all descendants.
+    pub trace_count: u64,
+    pub span_count: u64,
+    pub error_span_count: u64,
+    /// Elapsed window from the first observed span to the last observed span.
+    pub duration_ms: u64,
+    pub by_session: Vec<SessionTraceUsageV1>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct SessionTraceUsageV1 {
+    pub session_id: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub parent_session_id: Option<String>,
+    pub depth: u32,
+    pub trace_count: u64,
+    pub span_count: u64,
+    pub error_span_count: u64,
+    /// Elapsed window from the session's first observed span to its last.
+    pub duration_ms: u64,
+}
+
+#[derive(Debug, Deserialize)]
+struct TraceGroupByResponse {
+    groups: Vec<TraceGroup>,
+}
+
+#[derive(Debug, Deserialize)]
+struct TraceGroup {
+    value: String,
+    trace_ids: Vec<String>,
+    span_count: u64,
+    first_seen_ms: u64,
+    last_seen_ms: u64,
+    duration_ms: u64,
+    error_count: u64,
+}
+
+pub async fn handle(
+    deps: &Deps,
+    req: SessionMetricsRequestV1,
+) -> Result<SessionMetricsResponseV1, HarnessError> {
+    let tree = session_tree::collect(deps, &req.root_session_id).await?;
+    if !tree.complete {
+        return Ok(incomplete(&req.root_session_id));
+    }
+
+    let session = deps.session().await;
+    let cfg = deps.cfg().await;
+    let mut total = UsageAccumulator::default();
+    let mut by_session = Vec::with_capacity(tree.sessions.len());
+    for node in &tree.sessions {
+        if !session.exists(&node.session_id).await? {
+            return Ok(incomplete(&req.root_session_id));
+        }
+        let Some(turn) =
+            crate::state::get_turn(&deps.iii, &node.session_id, cfg.session_timeout_ms).await?
+        else {
+            return Ok(incomplete(&req.root_session_id));
+        };
+        if !terminal_status(Some(turn.status)) {
+            return Ok(incomplete(&req.root_session_id));
+        }
+        let entries = session.messages_strict(&node.session_id).await?;
+        let mut current = UsageAccumulator::default();
+        for entry in entries {
+            if let Some(message) = entry.message.as_ref() {
+                current.observe(message);
+                total.observe(message);
+            }
+        }
+        by_session.push(current.finish_session(node));
+    }
+    let traces = collect_trace_metrics(deps, &tree.sessions, &req.root_session_id).await;
+
+    Ok(SessionMetricsResponseV1 {
+        root_session_id: req.root_session_id,
+        complete: true,
+        totals: total.finish_totals(tree.sessions.len() as u64),
+        by_session,
+        traces,
+    })
+}
+
+fn terminal_status(status: Option<crate::types::turn::TurnStatus>) -> bool {
+    status.is_some_and(crate::types::turn::TurnStatus::is_terminal)
+}
+
+fn incomplete(root_session_id: &str) -> SessionMetricsResponseV1 {
+    SessionMetricsResponseV1 {
+        root_session_id: root_session_id.to_string(),
+        complete: false,
+        totals: SessionUsageTotalsV1::default(),
+        by_session: Vec::new(),
+        traces: None,
+    }
+}
+
+async fn collect_trace_metrics(
+    deps: &Deps,
+    sessions: &[SessionTreeNodeV1],
+    root_session_id: &str,
+) -> Option<SessionTraceMetricsV1> {
+    let cfg = deps.cfg().await;
+    let response = deps
+        .iii
+        .trigger(TriggerRequest {
+            function_id: "engine::traces::group_by".into(),
+            payload: json!({
+                "attribute": "iii.session.id",
+                "limit": u32::MAX,
+                "include_internal": true,
+            }),
+            action: None,
+            timeout_ms: Some(cfg.dispatch_timeout_ms),
+        })
+        .await;
+    let value = match response {
+        Ok(value) => value,
+        Err(error) => {
+            tracing::warn!(
+                %root_session_id,
+                %error,
+                "harness::metrics: trace metrics unavailable"
+            );
+            return None;
+        }
+    };
+    let response: TraceGroupByResponse = match serde_json::from_value(value) {
+        Ok(response) => response,
+        Err(error) => {
+            tracing::warn!(
+                %root_session_id,
+                %error,
+                "harness::metrics: invalid trace metrics response"
+            );
+            return None;
+        }
+    };
+    Some(summarize_trace_groups(sessions, response.groups))
+}
+
+fn summarize_trace_groups(
+    sessions: &[SessionTreeNodeV1],
+    groups: Vec<TraceGroup>,
+) -> SessionTraceMetricsV1 {
+    let mut groups: HashMap<_, _> = groups
+        .into_iter()
+        .map(|group| (group.value.clone(), group))
+        .collect();
+    let mut trace_ids = BTreeSet::new();
+    let mut span_count = 0u64;
+    let mut error_span_count = 0u64;
+    let mut first_seen_ms = None::<u64>;
+    let mut last_seen_ms = None::<u64>;
+    let mut by_session = Vec::with_capacity(sessions.len());
+
+    for session in sessions {
+        let group = groups.remove(&session.session_id);
+        let (session_trace_count, session_span_count, session_errors, duration_ms) = match group {
+            Some(group) => {
+                trace_ids.extend(group.trace_ids.iter().cloned());
+                span_count = span_count.saturating_add(group.span_count);
+                error_span_count = error_span_count.saturating_add(group.error_count);
+                first_seen_ms = Some(first_seen_ms.map_or(group.first_seen_ms, |current| {
+                    current.min(group.first_seen_ms)
+                }));
+                last_seen_ms = Some(last_seen_ms.map_or(group.last_seen_ms, |current| {
+                    current.max(group.last_seen_ms)
+                }));
+                (
+                    group.trace_ids.len() as u64,
+                    group.span_count,
+                    group.error_count,
+                    group.duration_ms,
+                )
+            }
+            None => (0, 0, 0, 0),
+        };
+        by_session.push(SessionTraceUsageV1 {
+            session_id: session.session_id.clone(),
+            parent_session_id: session.parent_session_id.clone(),
+            depth: session.depth,
+            trace_count: session_trace_count,
+            span_count: session_span_count,
+            error_span_count: session_errors,
+            duration_ms,
+        });
+    }
+
+    SessionTraceMetricsV1 {
+        trace_count: trace_ids.len() as u64,
+        span_count,
+        error_span_count,
+        duration_ms: match (first_seen_ms, last_seen_ms) {
+            (Some(first), Some(last)) => last.saturating_sub(first),
+            _ => 0,
+        },
+        by_session,
+    }
+}
+
+#[derive(Debug, Default)]
+struct UsageAccumulator {
+    turns: u64,
+    function_calls: u64,
+    function_call_errors: u64,
+    input: OptionalU64Sum,
+    output: OptionalU64Sum,
+    cache_read: OptionalU64Sum,
+    cache_write: OptionalU64Sum,
+    reasoning: OptionalU64Sum,
+    cost_usd: OptionalF64Sum,
+}
+
+impl UsageAccumulator {
+    fn observe(&mut self, message: &AgentMessage) {
+        match message {
+            AgentMessage::Assistant(assistant) => {
+                self.turns += 1;
+                self.function_calls += assistant
+                    .content
+                    .iter()
+                    .filter(|block| matches!(block, ContentBlock::FunctionCall { .. }))
+                    .count() as u64;
+                let usage = assistant.usage.as_ref();
+                self.input.observe(usage.and_then(|value| value.input));
+                self.output.observe(usage.and_then(|value| value.output));
+                self.cache_read
+                    .observe(usage.and_then(|value| value.cache_read));
+                self.cache_write
+                    .observe(usage.and_then(|value| value.cache_write));
+                self.reasoning
+                    .observe(usage.and_then(|value| value.reasoning));
+                self.cost_usd
+                    .observe(usage.and_then(|value| value.cost_usd));
+            }
+            AgentMessage::FunctionResult(result) if result.is_error => {
+                self.function_call_errors += 1;
+            }
+            _ => {}
+        }
+    }
+
+    fn finish_session(self, node: &SessionTreeNodeV1) -> SessionUsageV1 {
+        SessionUsageV1 {
+            session_id: node.session_id.clone(),
+            parent_session_id: node.parent_session_id.clone(),
+            depth: node.depth,
+            turns: self.turns,
+            function_calls: self.function_calls,
+            function_call_errors: self.function_call_errors,
+            input_tokens: self.input.finish(),
+            output_tokens: self.output.finish(),
+            cache_read_tokens: self.cache_read.finish(),
+            cache_write_tokens: self.cache_write.finish(),
+            reasoning_tokens: self.reasoning.finish(),
+            cost_usd: self.cost_usd.finish(),
+        }
+    }
+
+    fn finish_totals(self, sessions: u64) -> SessionUsageTotalsV1 {
+        SessionUsageTotalsV1 {
+            sessions,
+            turns: self.turns,
+            function_calls: self.function_calls,
+            function_call_errors: self.function_call_errors,
+            input_tokens: self.input.finish(),
+            output_tokens: self.output.finish(),
+            cache_read_tokens: self.cache_read.finish(),
+            cache_write_tokens: self.cache_write.finish(),
+            reasoning_tokens: self.reasoning.finish(),
+            cost_usd: self.cost_usd.finish(),
+        }
+    }
+}
+
+#[derive(Debug, Default)]
+struct OptionalU64Sum {
+    value: u64,
+    observations: u64,
+    all_present: bool,
+}
+
+impl OptionalU64Sum {
+    fn observe(&mut self, value: Option<u64>) {
+        if self.observations == 0 {
+            self.all_present = true;
+        }
+        self.observations += 1;
+        match value {
+            Some(value) => self.value = self.value.saturating_add(value),
+            None => self.all_present = false,
+        }
+    }
+
+    fn finish(self) -> Option<u64> {
+        (self.observations > 0 && self.all_present).then_some(self.value)
+    }
+}
+
+#[derive(Debug, Default)]
+struct OptionalF64Sum {
+    value: f64,
+    observations: u64,
+    all_present: bool,
+}
+
+impl OptionalF64Sum {
+    fn observe(&mut self, value: Option<f64>) {
+        if self.observations == 0 {
+            self.all_present = true;
+        }
+        self.observations += 1;
+        match value {
+            Some(value) => self.value += value,
+            None => self.all_present = false,
+        }
+    }
+
+    fn finish(self) -> Option<f64> {
+        (self.observations > 0 && self.all_present).then_some(self.value)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use serde_json::json;
+
+    use super::*;
+    use crate::types::turn::TurnStatus;
+
+    fn message(value: serde_json::Value) -> AgentMessage {
+        serde_json::from_value(value).unwrap()
+    }
+
+    fn session(id: &str, parent: Option<&str>, depth: u32) -> SessionTreeNodeV1 {
+        SessionTreeNodeV1 {
+            session_id: id.into(),
+            parent_session_id: parent.map(str::to_string),
+            parent_turn_id: None,
+            depth,
+        }
+    }
+
+    #[test]
+    fn counts_generations_calls_errors_and_usage() {
+        let mut usage = UsageAccumulator::default();
+        usage.observe(&message(json!({
+            "role": "assistant",
+            "content": [{
+                "type": "function_call",
+                "id": "call-1",
+                "function_id": "orders::refund",
+                "arguments": {"order_id": 4512}
+            }],
+            "stop_reason": "function_call",
+            "usage": {"input": 10, "output": 2, "cost_usd": 0.1},
+            "model": "model",
+            "provider": "provider",
+            "timestamp": 1
+        })));
+        usage.observe(&message(json!({
+            "role": "function_result",
+            "function_call_id": "call-1",
+            "function_id": "orders::refund",
+            "content": [],
+            "details": {},
+            "is_error": true,
+            "timestamp": 2
+        })));
+        let totals = usage.finish_totals(1);
+        assert_eq!(totals.sessions, 1);
+        assert_eq!(totals.turns, 1);
+        assert_eq!(totals.function_calls, 1);
+        assert_eq!(totals.function_call_errors, 1);
+        assert_eq!(totals.input_tokens, Some(10));
+        assert_eq!(totals.output_tokens, Some(2));
+        assert_eq!(totals.cost_usd, Some(0.1));
+        assert_eq!(totals.cache_read_tokens, None);
+    }
+
+    #[test]
+    fn optional_usage_never_reports_partial_sums() {
+        let mut usage = UsageAccumulator::default();
+        for input in [Some(10), None] {
+            usage.observe(&message(json!({
+                "role": "assistant",
+                "content": [],
+                "stop_reason": "end",
+                "usage": {"input": input, "output": 1},
+                "model": "model",
+                "provider": "provider",
+                "timestamp": 1
+            })));
+        }
+        let totals = usage.finish_totals(1);
+        assert_eq!(totals.turns, 2);
+        assert_eq!(totals.input_tokens, None);
+        assert_eq!(totals.output_tokens, Some(2));
+    }
+
+    #[test]
+    fn incomplete_metrics_never_contain_partial_values() {
+        let response = incomplete("root");
+        assert!(!response.complete);
+        assert!(response.by_session.is_empty());
+        assert_eq!(response.totals, SessionUsageTotalsV1::default());
+        assert_eq!(response.traces, None);
+    }
+
+    #[test]
+    fn metrics_complete_only_for_present_terminal_turns() {
+        assert!(!terminal_status(None));
+        assert!(!terminal_status(Some(TurnStatus::Running)));
+        assert!(!terminal_status(Some(TurnStatus::AwaitingFunctions)));
+        assert!(terminal_status(Some(TurnStatus::Completed)));
+        assert!(terminal_status(Some(TurnStatus::Cancelled)));
+        assert!(terminal_status(Some(TurnStatus::Failed)));
+    }
+
+    #[test]
+    fn trace_summary_covers_only_the_session_tree_and_deduplicates_traces() {
+        let sessions = [
+            session("root", None, 0),
+            session("child", Some("root"), 1),
+            session("without-spans", Some("root"), 1),
+        ];
+        let summary = summarize_trace_groups(
+            &sessions,
+            vec![
+                TraceGroup {
+                    value: "root".into(),
+                    trace_ids: vec!["trace-a".into(), "trace-shared".into()],
+                    span_count: 5,
+                    first_seen_ms: 100,
+                    last_seen_ms: 300,
+                    duration_ms: 200,
+                    error_count: 1,
+                },
+                TraceGroup {
+                    value: "child".into(),
+                    trace_ids: vec!["trace-shared".into(), "trace-b".into()],
+                    span_count: 7,
+                    first_seen_ms: 250,
+                    last_seen_ms: 600,
+                    duration_ms: 350,
+                    error_count: 2,
+                },
+                TraceGroup {
+                    value: "unrelated".into(),
+                    trace_ids: vec!["trace-other".into()],
+                    span_count: 99,
+                    first_seen_ms: 1,
+                    last_seen_ms: 999,
+                    duration_ms: 998,
+                    error_count: 9,
+                },
+            ],
+        );
+
+        assert_eq!(summary.trace_count, 3);
+        assert_eq!(summary.span_count, 12);
+        assert_eq!(summary.error_span_count, 3);
+        assert_eq!(summary.duration_ms, 500);
+        assert_eq!(summary.by_session.len(), 3);
+        assert_eq!(summary.by_session[0].trace_count, 2);
+        assert_eq!(summary.by_session[1].span_count, 7);
+        assert_eq!(
+            summary.by_session[2],
+            SessionTraceUsageV1 {
+                session_id: "without-spans".into(),
+                parent_session_id: Some("root".into()),
+                depth: 1,
+                trace_count: 0,
+                span_count: 0,
+                error_span_count: 0,
+                duration_ms: 0,
+            }
+        );
+    }
+}
diff --git a/harness/src/functions/mod.rs b/harness/src/functions/mod.rs
index f9d2454c2..4f47f0fb0 100644
--- a/harness/src/functions/mod.rs
+++ b/harness/src/functions/mod.rs
@@ -6,14 +6,17 @@
 pub mod filesystem;
 pub mod function_resolve;
 pub mod function_trigger;
+pub mod metrics;
 pub mod on_session_deleted;
 pub mod react;
 pub mod send;
+pub mod session_tree;
 pub mod spawn;
 pub mod status;
 pub mod stop;
 pub mod subscribe;
 pub mod sweep_pending;
+pub mod teardown;
 pub mod turn;
 
 use std::future::Future;
@@ -39,7 +42,9 @@ pub const SPAWN_DESC: &str =
     "Spawn a sub-agent fire-and-forget: seeds a child session/turn and returns \
      {child_session_id, child_turn_id} immediately. The child's result is NEVER delivered back \
      to the caller — register consumers (state triggers, or harness::turn-completed -> \
-     harness::react) BEFORE spawning.";
+     harness::react) BEFORE spawning. The task must include literal values for every required \
+     resource selector (for example `db: \"primary\"`). Omit child `max_turns` unless its budget \
+     covers discovery, contract lookup, work, and the deliverable.";
 
 pub const TURN_ID: &str = "harness::turn";
 pub const TURN_DESC: &str =
@@ -61,6 +66,19 @@ pub const STOP_DESC: &str =
 pub const STATUS_ID: &str = "harness::status";
 pub const STATUS_DESC: &str = "Read the current turn status for a session.";
 
+pub const SESSION_TREE_ID: &str = "harness::session-tree";
+pub const SESSION_TREE_DESC: &str =
+    "Read the durable root-and-descendant session tree for one harness run.";
+
+pub const METRICS_ID: &str = "harness::metrics";
+pub const METRICS_DESC: &str =
+    "Aggregate durable model usage, function outcomes, and available trace/span observability. \
+     `complete` is true only after every session in the durable tree has reached a terminal turn.";
+
+pub const TEARDOWN_ID: &str = "harness::teardown";
+pub const TEARDOWN_DESC: &str =
+    "Internal control-plane: remove trigger bindings owned by a root harness session tree.";
+
 pub const UNQUEUE_ID: &str = "harness::unqueue";
 pub const UNQUEUE_DESC: &str =
     "Internal control-plane: remove a still-parked queued message by entry_id (the console's edit-queued path).";
@@ -236,7 +254,19 @@ pub fn register_all(iii: &Arc<IIIClient>, deps: &Arc<Deps>) {
     register(iii, deps, STATUS_ID, STATUS_DESC, |d, r| async move {
         status::handle(&d, r).await
     });
-
+    register(
+        iii,
+        deps,
+        SESSION_TREE_ID,
+        SESSION_TREE_DESC,
+        |d, r| async move { session_tree::handle(&d, r).await },
+    );
+    register(iii, deps, METRICS_ID, METRICS_DESC, |d, r| async move {
+        metrics::handle(&d, r).await
+    });
+    register_internal(iii, deps, TEARDOWN_ID, TEARDOWN_DESC, |d, r| async move {
+        teardown::handle(&d, r).await
+    });
     // Trusted control-plane (console) — registered, kept off the agent catalog.
     register_internal(iii, deps, UNQUEUE_ID, UNQUEUE_DESC, |d, r| async move {
         send::unqueue(&d, r).await
diff --git a/harness/src/functions/on_session_deleted.rs b/harness/src/functions/on_session_deleted.rs
index b716e86fc..9221be529 100644
--- a/harness/src/functions/on_session_deleted.rs
+++ b/harness/src/functions/on_session_deleted.rs
@@ -23,29 +23,10 @@ pub async fn handle(
     deps: &Deps,
     event: SessionDeletedEvent,
 ) -> Result<SessionDeletedAck, HarnessError> {
-    let dropped = deps.subscriptions.take_session(&event.session_id);
     deps.subscriptions.forget_lineage(&event.session_id);
-    let tracked = dropped.len() as u64;
-    if tracked > 0 {
-        for (_sub_id, trigger_id) in dropped {
-            if let Some(trigger_id) = trigger_id {
-                crate::functions::subscribe::unregister_engine_trigger(deps, &trigger_id).await;
-            }
-        }
-        tracing::info!(
-            session_id = %event.session_id,
-            removed = tracked,
-            "session deleted: ephemeral subscriptions dropped"
-        );
-    }
+    let removed = super::teardown::cleanup_session(deps, &event.session_id).await;
     let cfg = deps.cfg().await;
     crate::filesystem_grants::purge(&deps.iii, &event.session_id, cfg.session_timeout_ms).await?;
-    // Durable complement: bindings registered before a harness restart are no
-    // longer in the in-memory registry but still fire engine-side — sweep them
-    // by their owner stamp so deleting a chat fully tears down its wiring.
-    let swept = crate::subscriptions::reconcile::sweep_owner(deps, &event.session_id).await;
-    Ok(SessionDeletedAck {
-        ok: true,
-        removed: tracked + swept as u64,
-    })
+    crate::budget::purge(deps, &event.session_id, cfg.session_timeout_ms).await?;
+    Ok(SessionDeletedAck { ok: true, removed })
 }
diff --git a/harness/src/functions/react.rs b/harness/src/functions/react.rs
index 91cce9680..54fb18e51 100644
--- a/harness/src/functions/react.rs
+++ b/harness/src/functions/react.rs
@@ -105,7 +105,9 @@ const JOIN_SCOPE: &str = "harness::react_join";
 /// breaker #2) only guards chains that stay on turn events — a cycle routed
 /// through an agent's `state::set` re-enters at depth 0, so a runaway there
 /// shows up as raw fire RATE instead of depth. Cap fires per subscription.
-pub const MAX_FIRES_PER_WINDOW: usize = 10;
+/// Allows a moderate event fan-out (for example, one reaction per item in a
+/// small parallel batch) while still stopping an accidental tight loop.
+pub const MAX_FIRES_PER_WINDOW: usize = 32;
 pub const FIRE_WINDOW_MS: i64 = 60_000;
 
 /// Effective gate limits. Overridable via `III_HARNESS_MAX_FIRES_PER_WINDOW`
@@ -304,10 +306,9 @@ pub struct ReactSpec {
     /// with `task`.
     #[serde(default)]
     pub call: Option<CallSpec>,
-    /// Spawn into this session (e.g. a fork); when omitted, defaults to the
-    /// registering session (the pipeline's owner) so the result lands back
-    /// as a turn in that chat — a fresh detached child only for raw
-    /// registrations that carry no owner stamp.
+    /// Spawn into this existing session (e.g. a fork). When omitted, each
+    /// event creates a fresh child session under `parent_session_id` or the
+    /// registering session's root.
     #[serde(default)]
     pub session_id: Option<String>,
     #[serde(default)]
@@ -372,9 +373,9 @@ pub struct ReactResult {
     /// already fired, error). Present iff `!spawned && !called`.
     #[serde(skip_serializing_if = "Option::is_none")]
     pub note: Option<String>,
-    /// The spawned turn id — per-fire unique even when delivery reuses the
-    /// pinned/owner session (the default). Local bookkeeping for fired-record
-    /// entry ids only; kept off the wire.
+    /// The spawned turn id — per-fire unique even when delivery uses an
+    /// explicitly pinned session. Local bookkeeping for fired-record entry ids
+    /// only; kept off the wire.
     #[serde(skip)]
     #[schemars(skip)]
     pub child_turn_id: Option<String>,
@@ -725,10 +726,9 @@ pub async fn handle(
         },
     };
 
-    // Delivery session, resolved once for BOTH dispatch paths below: an
-    // explicit pin wins; otherwise the registering (owner) session, so a
-    // reaction with no pin lands as a turn in the chat that wired it instead
-    // of a detached child nobody watches.
+    // Delivery session, resolved once for BOTH dispatch paths below. Only an
+    // explicit pin reuses a session; otherwise spawn creates a fresh child
+    // under the console-tree parent resolved above.
     let mut spec = spec;
     spec.session_id = reaction_delivery_session(&spec);
 
@@ -763,8 +763,8 @@ pub async fn handle(
                 if r.spawned {
                     let sub = spec.subscription_id.as_deref().unwrap_or("sub");
                     // Per-fire suffix: the spawned turn id is unique even when
-                    // delivery reuses the pinned/owner session (the default),
-                    // where the child session id repeats and would dedup every
+                    // delivery reuses an explicitly pinned session, where the
+                    // child session id repeats and would dedup every
                     // recurring fire after the first into one record.
                     // ponytail: `spawn` fallback when the spawn returned no
                     // ids — such fires dedup to one record, bounded by the
@@ -947,8 +947,8 @@ async fn join_edge(
     }
 
     // Fire the downstream fed ALL predecessors' results, then GC the
-    // accumulator record. The delivery session (owner fallback when unpinned)
-    // is already resolved by the caller. Joins fire once, so this cannot spam.
+    // accumulator record. The delivery session is already resolved by the
+    // caller. Joins fire once, so this cannot spam.
     let res = match spec.call.clone() {
         // Call downstream: the accumulated results map replaces the single
         // fired event as the injected value.
@@ -1285,14 +1285,10 @@ async fn state_update(deps: &Deps, key: &str, ops: Vec<Value>) -> Result<Value,
 }
 
 /// The session a reaction's spawn delivers into — simple edge or a completed
-/// join's downstream alike: an explicit spec pin wins; otherwise the
-/// registering session (the pipeline's owner), so the result lands as a turn
-/// in the chat that wired it. `None` (a fresh detached child) only for raw
-/// registrations that carry no owner stamp.
+/// join's downstream alike. An explicit pin is preserved; otherwise `None`
+/// tells `harness::spawn` to create a fresh child session.
 fn reaction_delivery_session(spec: &ReactSpec) -> Option<String> {
-    spec.session_id
-        .clone()
-        .or_else(|| spec.owner_session_id.clone())
+    spec.session_id.clone()
 }
 
 async fn record_reaction_outcome(
@@ -2216,21 +2212,18 @@ mod tests {
     }
 
     #[test]
-    fn reaction_delivery_session_prefers_pin_then_owner_stamp() {
+    fn reaction_delivery_session_only_reuses_explicit_pin() {
         // Applies uniformly to both dispatch paths: a simple (non-join) edge
         // and a join's downstream spawn both resolve through this function.
         let mut s = spec();
         s.owner_session_id = Some("console-owner".into());
         // Explicit pin wins.
         assert_eq!(reaction_delivery_session(&s).as_deref(), Some("s_run"));
-        // No pin: a non-join reaction lands in the chat that wired it, same
-        // as a join's fan-in result would.
+        // No pin: both simple and join reactions create a fresh child. The
+        // owner remains the tree/lifecycle anchor, not the delivery session.
         s.session_id = None;
-        assert_eq!(
-            reaction_delivery_session(&s).as_deref(),
-            Some("console-owner")
-        );
-        // Raw registration without an owner stamp: fresh child stands.
+        assert_eq!(reaction_delivery_session(&s), None);
+        // The owner stamp does not change delivery semantics.
         s.owner_session_id = None;
         assert_eq!(reaction_delivery_session(&s), None);
     }
diff --git a/harness/src/functions/send.rs b/harness/src/functions/send.rs
index 16e12aa45..ea62630dc 100644
--- a/harness/src/functions/send.rs
+++ b/harness/src/functions/send.rs
@@ -35,13 +35,24 @@ pub struct SendOptions {
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub system_prompt: Option<String>,
     /// How `system_prompt` combines with the built-in prompt: `override`
-    /// replaces it; `enrich` (default) appends to it.
+    /// replaces it; `enrich` (default) appends to it; `disabled` omits it.
     #[serde(default)]
     pub system_prompt_strategy: SystemPromptStrategy,
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub mode: Option<Mode>,
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_turns: Option<u32>,
+    /// Per-generation output-token ceiling forwarded to the router.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub max_output_tokens: Option<u64>,
+    /// Hard input-plus-output token budget for the complete root-and-subagent
+    /// session tree.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub max_total_tokens: Option<u64>,
+    /// Hard USD budget for the complete root-and-subagent session tree.
+    /// Every model used by the tree must advertise catalog pricing.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub max_cost_usd: Option<f64>,
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub thinking_level: Option<ThinkingLevel>,
     /// Provider-native per-call options, namespaced by provider id.
@@ -54,7 +65,7 @@ pub struct SendOptions {
     /// call; omitted when steering an EXISTING session → inherit the prior
     /// turn's policy (a nudge must not disarm a live run). Pass
     /// `{ allow: [] }` to strip explicitly. On a NEW `ask`-mode turn the
-    /// effective policy is capped at the operator's read-only baseline; a
+    /// effective policy is capped at the configured default policy; a
     /// steer folded into an already-running turn keeps that turn's frozen
     /// policy until it finalises.
     #[serde(default, skip_serializing_if = "Option::is_none")]
@@ -222,6 +233,7 @@ pub async fn start(deps: &Deps, req: SendRequest) -> Result<StartOutcome, Harnes
         }
         None => session.create(title.as_deref(), metadata.as_ref()).await?,
     };
+    crate::budget::prepare_root(deps, &session_id, &mut options, prev.as_ref()).await?;
 
     // Entry id: idempotent when a dedupe key is set.
     let entry_id = req
@@ -512,6 +524,10 @@ fn build_options(
         ),
         mode: opts.mode,
         max_turns: opts.max_turns.unwrap_or(cfg.default_max_turns),
+        max_output_tokens: opts.max_output_tokens,
+        max_total_tokens: opts.max_total_tokens,
+        max_cost_usd: opts.max_cost_usd,
+        budget_root_session_id: None,
         thinking_level: opts.thinking_level,
         provider_options: opts.provider_options,
         output: opts.output.unwrap_or_default(),
@@ -522,10 +538,10 @@ fn build_options(
     }
 }
 
-/// The single chokepoint for the "ask is structurally read-only" invariant:
+/// The single chokepoint for the ask-mode policy cap:
 /// every turn-seeding path (a fresh send, an inherited steer, a spawned child)
 /// routes its resolved dispatch policy through here, so ask mode is capped at
-/// the operator's read-only baseline no matter how the policy was assembled.
+/// the configured default policy no matter how the policy was assembled.
 /// A non-ask turn passes through untouched.
 pub(crate) fn clamp_for_mode(
     cfg: &WorkerConfig,
@@ -854,6 +870,10 @@ mod tests {
             system_prompt: None,
             mode,
             max_turns: 16,
+            max_output_tokens: None,
+            max_total_tokens: None,
+            max_cost_usd: None,
+            budget_root_session_id: None,
             thinking_level: None,
             provider_options: None,
             output: OutputContract::Text,
@@ -873,12 +893,13 @@ mod tests {
             expose: Default::default(),
         };
 
-        // An ask-mode steer keeps the run armed but capped at the baseline.
+        // An ask-mode steer keeps the run armed and is capped at the wildcard
+        // default policy.
         let mut options = options_with(Some(Mode::Ask), None);
         inherit_prior_functions(&cfg, &mut options, Some(&broad));
         let compiled = policy::CompiledPolicy::from(options.functions.as_ref());
         assert!(compiled.allows("state::get"));
-        assert!(!compiled.allows("state::set"));
+        assert!(compiled.allows("state::set"));
 
         // Outside ask mode the prior policy is inherited whole.
         let mut options = options_with(Some(Mode::Agent), None);
@@ -897,7 +918,7 @@ mod tests {
     }
 
     #[test]
-    fn build_options_clamps_functions_to_the_read_only_baseline_in_ask_mode() {
+    fn build_options_clamps_functions_to_the_default_policy_in_ask_mode() {
         let cfg = WorkerConfig::default();
         let req = SendRequest {
             session_id: None,
@@ -918,21 +939,13 @@ mod tests {
         };
         let opts = build_options(&cfg, &req, "m".into(), req.provider.clone(), None);
         let compiled = policy::CompiledPolicy::from(opts.functions.as_ref());
-        // Baseline reads survive the clamp…
+        // The wildcard default preserves the requested policy…
         assert!(compiled.allows("state::get"));
         assert!(compiled.allows("engine::functions::list"));
-        // …every write/orchestration surface is out, wildcard request or not.
-        for denied in [
-            "state::set",
-            "harness::spawn",
-            "engine::register_trigger",
-            "shell::run",
-        ] {
-            assert!(
-                !compiled.allows(denied),
-                "{denied} must be denied in ask mode"
-            );
-        }
+        assert!(compiled.allows("state::set"));
+        assert!(compiled.allows("harness::spawn"));
+        assert!(compiled.allows("engine::register_trigger"));
+        assert!(compiled.allows("shell::run"));
     }
 
     #[test]
diff --git a/harness/src/functions/session_tree.rs b/harness/src/functions/session_tree.rs
new file mode 100644
index 000000000..7b8784d6b
--- /dev/null
+++ b/harness/src/functions/session_tree.rs
@@ -0,0 +1,193 @@
+//! `harness::session-tree` — resolve the durable root-and-descendant session
+//! membership recorded by the harness in session metadata.
+
+use std::collections::{BTreeSet, VecDeque};
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::clients::session::SessionLink;
+use crate::deps::Deps;
+use crate::error::HarnessError;
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+pub struct SessionTreeRequestV1 {
+    pub root_session_id: String,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct SessionTreeNodeV1 {
+    pub session_id: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub parent_session_id: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub parent_turn_id: Option<String>,
+    pub depth: u32,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct SessionTreeResponseV1 {
+    pub root_session_id: String,
+    pub sessions: Vec<SessionTreeNodeV1>,
+    pub complete: bool,
+}
+
+pub async fn handle(
+    deps: &Deps,
+    req: SessionTreeRequestV1,
+) -> Result<SessionTreeResponseV1, HarnessError> {
+    collect(deps, &req.root_session_id).await
+}
+
+pub(crate) async fn collect(
+    deps: &Deps,
+    root_session_id: &str,
+) -> Result<SessionTreeResponseV1, HarnessError> {
+    let session = deps.session().await;
+    if !session.exists(root_session_id).await? {
+        return Ok(SessionTreeResponseV1 {
+            root_session_id: root_session_id.to_string(),
+            sessions: Vec::new(),
+            complete: false,
+        });
+    }
+
+    let mut builder = TreeBuilder::new(root_session_id);
+    while let Some((parent_session_id, depth)) = builder.next_parent() {
+        let children = session.children(&parent_session_id).await?;
+        builder.add_children(&parent_session_id, depth + 1, children);
+    }
+    Ok(builder.finish())
+}
+
+struct TreeBuilder {
+    root_session_id: String,
+    sessions: Vec<SessionTreeNodeV1>,
+    pending: VecDeque<(String, u32)>,
+    seen: BTreeSet<String>,
+    complete: bool,
+}
+
+impl TreeBuilder {
+    fn new(root_session_id: &str) -> Self {
+        let root_session_id = root_session_id.to_string();
+        Self {
+            sessions: vec![SessionTreeNodeV1 {
+                session_id: root_session_id.clone(),
+                parent_session_id: None,
+                parent_turn_id: None,
+                depth: 0,
+            }],
+            pending: VecDeque::from([(root_session_id.clone(), 0)]),
+            seen: BTreeSet::from([root_session_id.clone()]),
+            root_session_id,
+            complete: true,
+        }
+    }
+
+    fn next_parent(&mut self) -> Option<(String, u32)> {
+        self.pending.pop_front()
+    }
+
+    fn add_children(&mut self, parent_session_id: &str, depth: u32, children: Vec<SessionLink>) {
+        for child in children {
+            if child.parent_session_id != parent_session_id
+                || !self.seen.insert(child.session_id.clone())
+            {
+                self.complete = false;
+                continue;
+            }
+            self.pending.push_back((child.session_id.clone(), depth));
+            self.sessions.push(SessionTreeNodeV1 {
+                session_id: child.session_id,
+                parent_session_id: Some(parent_session_id.to_string()),
+                parent_turn_id: child.parent_turn_id,
+                depth,
+            });
+        }
+    }
+
+    fn finish(mut self) -> SessionTreeResponseV1 {
+        self.sessions.sort_by(|left, right| {
+            left.depth
+                .cmp(&right.depth)
+                .then_with(|| left.session_id.cmp(&right.session_id))
+        });
+        SessionTreeResponseV1 {
+            root_session_id: self.root_session_id,
+            sessions: self.sessions,
+            complete: self.complete,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn link(session: &str, parent: &str, turn: Option<&str>) -> SessionLink {
+        SessionLink {
+            session_id: session.into(),
+            parent_session_id: parent.into(),
+            parent_turn_id: turn.map(str::to_string),
+        }
+    }
+
+    #[test]
+    fn builder_orders_root_then_depth_then_session_id() {
+        let mut tree = TreeBuilder::new("root");
+        let (root, depth) = tree.next_parent().unwrap();
+        tree.add_children(
+            &root,
+            depth + 1,
+            vec![
+                link("child-b", "root", None),
+                link("child-a", "root", Some("turn-1")),
+            ],
+        );
+        while let Some((parent, parent_depth)) = tree.next_parent() {
+            if parent == "child-a" {
+                tree.add_children(
+                    &parent,
+                    parent_depth + 1,
+                    vec![link("grandchild", "child-a", Some("turn-2"))],
+                );
+            }
+        }
+        let tree = tree.finish();
+        assert!(tree.complete);
+        assert_eq!(
+            tree.sessions
+                .iter()
+                .map(|node| (node.session_id.as_str(), node.depth))
+                .collect::<Vec<_>>(),
+            vec![
+                ("root", 0),
+                ("child-a", 1),
+                ("child-b", 1),
+                ("grandchild", 2)
+            ]
+        );
+        assert_eq!(tree.sessions[1].parent_turn_id.as_deref(), Some("turn-1"));
+    }
+
+    #[test]
+    fn duplicate_or_mismatched_links_make_tree_incomplete() {
+        let mut tree = TreeBuilder::new("root");
+        tree.add_children(
+            "root",
+            1,
+            vec![
+                link("child", "root", None),
+                link("child", "root", None),
+                link("other", "wrong-parent", None),
+                link("root", "root", None),
+            ],
+        );
+        let tree = tree.finish();
+        assert!(!tree.complete);
+        assert_eq!(tree.sessions.len(), 2);
+    }
+}
diff --git a/harness/src/functions/spawn.rs b/harness/src/functions/spawn.rs
index 768a41dd5..b8adf17e7 100644
--- a/harness/src/functions/spawn.rs
+++ b/harness/src/functions/spawn.rs
@@ -20,21 +20,27 @@ pub struct SpawnOptions {
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub system_prompt: Option<String>,
     /// How `system_prompt` combines with the built-in prompt: `override`
-    /// replaces it; `enrich` (default) appends to it.
+    /// replaces it; `enrich` (default) appends to it; `disabled` omits it.
     #[serde(default)]
     pub system_prompt_strategy: SystemPromptStrategy,
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub mode: Option<Mode>,
-    /// Capped at the parent's remaining turn budget.
+    /// Capped at the parent's remaining turn budget. Omit unless a strict
+    /// child-specific cap is required. It must cover discovery/contract calls
+    /// plus every work call; very small values (for example 2-3) commonly
+    /// strand the child before it can produce its deliverable.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_turns: Option<u32>,
+    /// Inherits the parent's ceiling unless explicitly narrowed/overridden.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub max_output_tokens: Option<u64>,
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub thinking_level: Option<ThinkingLevel>,
     /// The child's deliverable: text / json / json+schema.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub output: Option<OutputContract>,
     /// Intersected with the parent policy — narrow, never escalate. An
-    /// `ask`-mode child is further capped at the read-only baseline.
+    /// `ask`-mode child is further capped at the configured default policy.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub functions: Option<FunctionPolicy>,
     /// Fan-out guard for the child's own spawns.
@@ -51,7 +57,9 @@ pub struct SpawnOptions {
 
 #[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
 pub struct SpawnRequest {
-    /// The child's goal — its opening user message.
+    /// The child's self-contained goal — its opening user message. Include
+    /// every resolved required selector literally (for example `Use database
+    /// db: "primary"`); the child cannot infer resources from the parent.
     pub task: MessageInput,
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub model: Option<String>,
diff --git a/harness/src/functions/status.rs b/harness/src/functions/status.rs
index 0f8d05156..56a835656 100644
--- a/harness/src/functions/status.rs
+++ b/harness/src/functions/status.rs
@@ -9,7 +9,7 @@ use crate::deps::Deps;
 use crate::error::HarnessError;
 use crate::types::turn::TurnStatus;
 
-#[derive(Debug, Clone, Default, Deserialize, JsonSchema)]
+#[derive(Debug, Clone, Default, Serialize, Deserialize, JsonSchema)]
 pub struct StatusRequest {
     pub session_id: String,
 }
diff --git a/harness/src/functions/stop.rs b/harness/src/functions/stop.rs
index e8ccf85dd..6a3c2b0d2 100644
--- a/harness/src/functions/stop.rs
+++ b/harness/src/functions/stop.rs
@@ -9,7 +9,7 @@ use serde::{Deserialize, Serialize};
 use crate::deps::Deps;
 use crate::error::HarnessError;
 
-#[derive(Debug, Clone, Default, Deserialize, JsonSchema)]
+#[derive(Debug, Clone, Default, Serialize, Deserialize, JsonSchema)]
 pub struct StopRequest {
     pub session_id: String,
     /// Omit to stop the current turn.
diff --git a/harness/src/functions/subscribe.rs b/harness/src/functions/subscribe.rs
index 72345691a..74c1114a6 100644
--- a/harness/src/functions/subscribe.rs
+++ b/harness/src/functions/subscribe.rs
@@ -20,6 +20,7 @@ use crate::policy::CompiledPolicy;
 use crate::subscriptions::{self, NOTIFY_AGENT_ID};
 use crate::trigger::{self, ResultData};
 use crate::types::content::ContentBlock;
+use crate::types::model::AgentFunction;
 
 /// The engine function the agent calls to subscribe. The harness intercepts it
 /// (the agent never reaches the raw engine registrar) so it can stamp the
@@ -92,6 +93,48 @@ pub struct SubscribeResponse {
     pub note: Option<String>,
 }
 
+/// Agent-facing unsubscribe contract. The id is the harness subscription id
+/// returned by [`SubscribeResponse`], not the underlying engine binding id.
+#[derive(Debug, Clone, Deserialize, JsonSchema)]
+#[schemars(rename = "UnsubscribeArgs")]
+pub struct UnsubscribeRequest {
+    pub id: String,
+}
+
+#[derive(Debug, Clone, Serialize, JsonSchema)]
+pub struct UnsubscribeResponse {
+    pub removed: bool,
+}
+
+/// Subscription controls are intercepted by the harness and therefore do not
+/// appear in the engine's public function registry. Native exposure still has
+/// to publish their real contracts when the turn policy allows them.
+pub fn native_control_tools(policy: &CompiledPolicy) -> Vec<AgentFunction> {
+    [
+        (
+            REGISTER_TRIGGER_ID,
+            "Subscribe the current harness session to a iii trigger. Omit function_id to wake \
+             this session, or use harness::react with a reaction spec in metadata.",
+            crate::surface::schema_value::<SubscribeRequest>(),
+        ),
+        (
+            UNREGISTER_TRIGGER_ID,
+            "Remove a trigger subscription owned by the current harness session.",
+            crate::surface::schema_value::<UnsubscribeRequest>(),
+        ),
+    ]
+    .into_iter()
+    .filter(|(function_id, _, _)| policy.allows(function_id))
+    .map(|(function_id, description, parameters)| AgentFunction {
+        name: function_id.to_string(),
+        description: description.to_string(),
+        parameters,
+        label: None,
+        execution_mode: Some("sequential".to_string()),
+    })
+    .collect()
+}
+
 /// The `session_id` a turn-event registration filters on, if any — the only
 /// filter shape whose typo/mismatch starves silently.
 fn turn_event_session_filter(req: &SubscribeRequest) -> Option<&str> {
@@ -338,10 +381,15 @@ fn inherit_model(metadata: &mut Option<Value>, caller: Option<CallerModel<'_>>)
 }
 
 async fn intercept_unregister(deps: &Deps, args: &Value, session_id: &str) -> ResultData {
-    let id = match unregister_subscription_id(args) {
-        Ok(id) => id,
-        Err(e) => return error_result(e),
+    let req: UnsubscribeRequest = match serde_json::from_value(args.clone()) {
+        Ok(req) => req,
+        Err(error) => {
+            return error_result(format!(
+                "{UNREGISTER_TRIGGER_ID} requires a string `id`: {error}"
+            ))
+        }
     };
+    let id = req.id.as_str();
 
     if let Some(owner) = deps.subscriptions.session_of(id) {
         // The registrant itself, or any session in its reaction lineage: a
@@ -361,13 +409,7 @@ async fn intercept_unregister(deps: &Deps, args: &Value, session_id: &str) -> Re
         }
         None => false,
     };
-    ok_result(&json!({ "removed": removed }))
-}
-
-fn unregister_subscription_id(args: &Value) -> Result<&str, String> {
-    args.get("id")
-        .and_then(Value::as_str)
-        .ok_or_else(|| "engine::unregister_trigger requires an `id`".to_string())
+    ok_result(&UnsubscribeResponse { removed })
 }
 
 async fn handle(
@@ -1465,17 +1507,29 @@ mod tests {
     }
 
     #[test]
-    fn unregister_requires_string_subscription_id() {
-        assert_eq!(
-            unregister_subscription_id(&json!({})).unwrap_err(),
-            "engine::unregister_trigger requires an `id`"
-        );
-        assert_eq!(
-            unregister_subscription_id(&json!({ "id": 42 })).unwrap_err(),
-            "engine::unregister_trigger requires an `id`"
-        );
+    fn native_subscription_controls_follow_dispatch_policy() {
+        let policy = crate::types::turn::FunctionPolicy {
+            allow: vec![REGISTER_TRIGGER_ID.into(), UNREGISTER_TRIGGER_ID.into()],
+            deny: vec![UNREGISTER_TRIGGER_ID.into()],
+            expose: crate::types::turn::ExposeMode::Native,
+        };
+        let tools = native_control_tools(&CompiledPolicy::from(Some(&policy)));
+        assert_eq!(tools.len(), 1);
+        assert_eq!(tools[0].name, REGISTER_TRIGGER_ID);
+        assert_eq!(tools[0].parameters["required"], json!(["trigger_type"]));
+        assert!(tools[0].parameters["properties"]
+            .get("session_id")
+            .is_none());
+    }
+
+    #[test]
+    fn unregister_contract_requires_string_subscription_id() {
+        assert!(serde_json::from_value::<UnsubscribeRequest>(json!({})).is_err());
+        assert!(serde_json::from_value::<UnsubscribeRequest>(json!({ "id": 42 })).is_err());
         assert_eq!(
-            unregister_subscription_id(&json!({ "id": "sub_1" })).unwrap(),
+            serde_json::from_value::<UnsubscribeRequest>(json!({ "id": "sub_1" }))
+                .unwrap()
+                .id,
             "sub_1"
         );
     }
diff --git a/harness/src/functions/teardown.rs b/harness/src/functions/teardown.rs
new file mode 100644
index 000000000..dc527817d
--- /dev/null
+++ b/harness/src/functions/teardown.rs
@@ -0,0 +1,55 @@
+//! `harness::teardown` — remove trigger bindings owned by a complete harness
+//! session tree. This is trusted control-plane cleanup for evals and other
+//! lifecycle owners; it is not exposed to models.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::deps::Deps;
+use crate::error::HarnessError;
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+pub struct TeardownRequestV1 {
+    pub root_session_id: String,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct TeardownResponseV1 {
+    pub removed: u64,
+}
+
+pub async fn handle(
+    deps: &Deps,
+    req: TeardownRequestV1,
+) -> Result<TeardownResponseV1, HarnessError> {
+    let tree = super::session_tree::collect(deps, &req.root_session_id).await?;
+    let mut removed = 0u64;
+    for node in tree.sessions {
+        removed = removed.saturating_add(cleanup_session(deps, &node.session_id).await);
+    }
+    Ok(TeardownResponseV1 { removed })
+}
+
+pub(crate) async fn cleanup_session(deps: &Deps, session_id: &str) -> u64 {
+    let dropped = deps.subscriptions.take_session(session_id);
+    let tracked = dropped.len() as u64;
+    for (_subscription_id, trigger_id) in dropped {
+        if let Some(trigger_id) = trigger_id {
+            super::subscribe::unregister_engine_trigger(deps, &trigger_id).await;
+        }
+    }
+
+    // Durable complement: after a harness restart the in-memory registry no
+    // longer knows the binding, but the engine still carries its owner stamp.
+    let swept = crate::subscriptions::reconcile::sweep_owner(deps, session_id).await;
+    let removed = tracked.saturating_add(swept as u64);
+    if removed > 0 {
+        tracing::info!(
+            session_id,
+            removed,
+            "harness session trigger bindings removed"
+        );
+    }
+    removed
+}
diff --git a/harness/src/hooks/runner.rs b/harness/src/hooks/runner.rs
index 12be606d2..cba626acb 100644
--- a/harness/src/hooks/runner.rs
+++ b/harness/src/hooks/runner.rs
@@ -263,12 +263,24 @@ impl HookRegistry {
     /// Invoke one bound hook function, bounded by `timeout_ms` and resolved by
     /// `on_error`. A redelivered step re-runs hooks — they must be idempotent.
     async fn invoke(&self, binding: &HookBinding, input: Value) -> HookOutcome {
-        let fut = self.iii.trigger(TriggerRequest {
-            function_id: binding.function_id.clone(),
-            payload: input,
-            action: None,
-            timeout_ms: Some(binding.timeout_ms),
-        });
+        let point = input
+            .get("point")
+            .and_then(Value::as_str)
+            .unwrap_or("unknown")
+            .to_string();
+        let baggage = [
+            ("iii.tag.kind", "harness.hook"),
+            ("iii.hook.point", point.as_str()),
+        ];
+        let fut = iii_helpers::observability::run_with_baggage(
+            &baggage,
+            self.iii.trigger(TriggerRequest {
+                function_id: binding.function_id.clone(),
+                payload: input,
+                action: None,
+                timeout_ms: Some(binding.timeout_ms),
+            }),
+        );
         let bounded = tokio::time::timeout(
             Duration::from_millis(binding.timeout_ms.saturating_add(1_000)),
             fut,
diff --git a/harness/src/lib.rs b/harness/src/lib.rs
index dc25f94f8..6342c0f57 100644
--- a/harness/src/lib.rs
+++ b/harness/src/lib.rs
@@ -7,6 +7,7 @@
 //! function calls, and repeat until the turn stops — as durable enqueued steps
 //! so a crash resumes mid-turn.
 
+pub mod budget;
 pub mod clients;
 pub mod config;
 pub mod configuration;
diff --git a/harness/src/policy.rs b/harness/src/policy.rs
index 6f76f7564..b32758693 100644
--- a/harness/src/policy.rs
+++ b/harness/src/policy.rs
@@ -82,19 +82,16 @@ pub fn subset_policy(
     })
 }
 
-/// Cap a requested policy at the operator's read-only baseline (ask mode):
+/// Cap a requested policy at the operator's configured baseline (ask mode):
 /// the effective allow is the baseline kept only where the REQUEST covers it,
 /// denies are unioned, and the request's exposure choice is kept. Never
 /// widens — an absent or deny-all request stays deny-all, and an absent or
 /// hollow (empty-allow) baseline denies everything.
 ///
 /// Coverage is [`glob_covered`] (exact id, `*`, or `<prefix>::*`), so the
-/// result is a conservative UNDER-approximation of a set intersection, not a
-/// true one: a glob baseline entry (e.g. `coder::*`) survives only when a
-/// request glob covers it, so a narrower exact request (`coder::read-file`)
-/// against that baseline yields deny-all. That fails closed, never open — the
-/// shipped baseline (`config::default_functions`) is all exact ids, where the
-/// distinction does not arise.
+/// result is a conservative UNDER-approximation of a set intersection for
+/// prefix globs. A wildcard baseline is handled as an unrestricted ceiling,
+/// so a narrower requested allow-list remains effective.
 pub fn clamp_policy(
     baseline: Option<&FunctionPolicy>,
     requested: Option<&FunctionPolicy>,
@@ -110,6 +107,13 @@ pub fn clamp_policy(
         !baseline.allow.is_empty(),
         "clamp baseline must be non-empty here or the flip widens"
     );
+    // A wildcard baseline covers every requested function. Preserve the
+    // request's allow-list and add any baseline denies.
+    if baseline.allow.iter().any(|pattern| pattern == "*") {
+        let mut unrestricted = requested.clone();
+        unrestricted.deny.extend(baseline.deny.iter().cloned());
+        return Some(unrestricted);
+    }
     // Subset with the roles flipped — keep the baseline's ids where the
     // REQUEST covers them — so broad requests (`*`, `<prefix>::*`) intersect
     // down to the baseline instead of collapsing to deny-all.
@@ -419,6 +423,16 @@ mod tests {
         assert!(!compiled.allows("state::list")); // baseline deny bites through the clamp
     }
 
+    #[test]
+    fn ask_clamp_wildcard_baseline_preserves_narrow_request() {
+        let baseline = policy(&["*"], &[]);
+        let requested = policy(&["state::get"], &[]);
+        let compiled =
+            CompiledPolicy::from(clamp_policy(Some(&baseline), Some(&requested)).as_ref());
+        assert!(compiled.allows("state::get"));
+        assert!(!compiled.allows("state::set"));
+    }
+
     #[test]
     fn ask_clamp_under_approximates_a_glob_baseline_fail_closed() {
         // Documented conservative behavior: a glob baseline entry survives only
diff --git a/harness/src/prompt/mod.rs b/harness/src/prompt/mod.rs
index c3de744cf..6f5106136 100644
--- a/harness/src/prompt/mod.rs
+++ b/harness/src/prompt/mod.rs
@@ -24,6 +24,8 @@ pub enum SystemPromptStrategy {
     /// Caller prompt is appended to the built-in identity prompt.
     #[default]
     Enrich,
+    /// No system prompt is sent to the model.
+    Disabled,
 }
 
 #[derive(Debug, Clone, Copy, Default)]
@@ -45,8 +47,8 @@ pub fn build_system_prompt(opts: SystemPromptOpts<'_>) -> String {
 
 /// Resolve the system prompt for a turn. With `Override`, a non-empty caller
 /// prompt wins verbatim; with `Enrich`, it is appended to the built-in prompt.
-/// An absent or empty caller prompt yields the built-in prompt under either
-/// strategy.
+/// An absent or empty caller prompt yields the built-in prompt under
+/// `Override` and `Enrich`. `Disabled` always yields no system prompt.
 pub fn resolve_system_prompt(
     override_prompt: Option<String>,
     strategy: SystemPromptStrategy,
@@ -56,6 +58,7 @@ pub fn resolve_system_prompt(
     let built_in = || build_system_prompt(SystemPromptOpts { mode, identity });
     let custom = override_prompt.as_deref().filter(|s| !s.is_empty());
     match (strategy, custom) {
+        (SystemPromptStrategy::Disabled, _) => None,
         (SystemPromptStrategy::Override, Some(s)) => Some(s.to_string()),
         (SystemPromptStrategy::Enrich, Some(s)) => Some(format!("{}\n\n{}", built_in(), s)),
         (_, None) => Some(built_in()),
diff --git a/harness/src/prompt/mode.rs b/harness/src/prompt/mode.rs
index 50c149631..ba5ea96dd 100644
--- a/harness/src/prompt/mode.rs
+++ b/harness/src/prompt/mode.rs
@@ -5,7 +5,7 @@ use serde::{Deserialize, Serialize};
 
 /// Console / send operating mode — prepends a short paragraph before the
 /// shared identity prompt. `ask` is also structural: the turn's dispatch
-/// policy is capped at the operator's read-only baseline, never widened.
+/// policy is capped at the configured default policy, never widened.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
 #[serde(rename_all = "snake_case")]
 pub enum Mode {
diff --git a/harness/src/prompt/tests.rs b/harness/src/prompt/tests.rs
index 1271c2cda..a39589fb5 100644
--- a/harness/src/prompt/tests.rs
+++ b/harness/src/prompt/tests.rs
@@ -84,6 +84,19 @@ fn resolve_enrich_with_empty_custom_falls_through_to_builtin() {
     assert_eq!(out, built_in);
 }
 
+#[test]
+fn resolve_disabled_omits_system_prompt() {
+    assert_eq!(
+        resolve_system_prompt(
+            Some("ignored".into()),
+            SystemPromptStrategy::Disabled,
+            Some(Mode::Agent),
+            Some(IDENTITY),
+        ),
+        None
+    );
+}
+
 #[test]
 fn identity_line_and_agent_trigger_preserved() {
     let out = default_prompt();
@@ -109,6 +122,34 @@ fn mesh_model() {
     assert!(out.contains("register a trigger; do not poll"));
 }
 
+#[test]
+fn delegation_carries_resolved_resource_selectors() {
+    let out = default_prompt().replace('\n', " ");
+    assert!(out.contains("every resolved resource selector the child must pass"));
+    assert!(out.contains("`db: \"primary\"`"));
+    assert!(out.contains("Use database db: \"<resolved name>\""));
+    assert!(out.contains("Do not dispatch a task until this audit passes."));
+    assert!(out.contains("ends a child immediately after discovery"));
+}
+
+#[test]
+fn reaction_defaults_match_runtime() {
+    let out = default_prompt().replace('\n', " ");
+    assert!(out.contains("every event creates a fresh distinct child"));
+    assert!(out.contains("you MUST omit `metadata.session_id`"));
+    assert!(out.contains("edges are standing by default"));
+    assert!(out.contains("Set `once: true` only for a deliberate one-shot reaction"));
+    assert!(out.contains("`once` is a TOP-LEVEL"));
+    assert!(out.contains("putting it inside `metadata` does nothing"));
+}
+
+#[test]
+fn fresh_namespaces_are_derived_and_checked() {
+    let out = default_prompt().replace('\n', " ");
+    assert!(out.contains("derive its variable suffix from the unique session id"));
+    assert!(out.contains("confirm the namespace is absent"));
+}
+
 #[test]
 fn registry_allowlist_invariant() {
     let out = default_prompt();
diff --git a/harness/src/subagent.rs b/harness/src/subagent.rs
index 6dbc5b774..1cbda0839 100644
--- a/harness/src/subagent.rs
+++ b/harness/src/subagent.rs
@@ -133,8 +133,8 @@ pub async fn spawn_child(
 /// request. Dumbness is a prompt/data-flow property, not a capability wall — a
 /// child CAN spawn/register, it just shouldn't (subagent.txt). A parentless
 /// (direct/CLI/trigger-fired) spawn has no parent to mirror: explicit options
-/// win, else the read-only baseline. Whatever was resolved, an ask-mode child
-/// is then capped at the baseline — ask is structurally read-only everywhere.
+/// win, else the configured default policy. Whatever was resolved, an ask-mode
+/// child is then capped at that policy.
 fn child_functions(
     cfg: &WorkerConfig,
     parent_record: Option<&TurnRecord>,
@@ -301,6 +301,17 @@ async fn seed_child(
             ),
             mode: req.options.as_ref().and_then(|o| o.mode),
             max_turns,
+            max_output_tokens: req
+                .options
+                .as_ref()
+                .and_then(|o| o.max_output_tokens)
+                .or_else(|| parent_record.and_then(|p| p.options.max_output_tokens)),
+            // Children cannot widen or replace their root turn's execution
+            // budget; every generation charges the same durable ledger.
+            max_total_tokens: parent_record.and_then(|p| p.options.max_total_tokens),
+            max_cost_usd: parent_record.and_then(|p| p.options.max_cost_usd),
+            budget_root_session_id: parent_record
+                .and_then(|p| p.options.budget_root_session_id.clone()),
             thinking_level: req.options.as_ref().and_then(|o| o.thinking_level),
             provider_options: None,
             output: req
@@ -435,6 +446,10 @@ mod tests {
                 system_prompt: None,
                 mode: None,
                 max_turns: 16,
+                max_output_tokens: None,
+                max_total_tokens: None,
+                max_cost_usd: None,
+                budget_root_session_id: None,
                 thinking_level: None,
                 provider_options: None,
                 output: OutputContract::Text,
@@ -577,7 +592,7 @@ mod tests {
     }
 
     #[test]
-    fn ask_mode_child_is_capped_at_the_read_only_baseline() {
+    fn ask_mode_child_is_capped_at_the_default_policy() {
         let cfg = WorkerConfig::default();
         let mut parent = parent_record(None);
         parent.options.functions = Some(broad_policy());
@@ -586,12 +601,12 @@ mod tests {
         let inherited = child_functions(&cfg, Some(&parent), None, Some(Mode::Agent));
         assert!(policy::CompiledPolicy::from(inherited.as_ref()).allows("state::set"));
 
-        // An ask-mode child is capped at the baseline, whatever it inherited.
+        // An ask-mode child is capped at the wildcard default, whatever it inherited.
         let asked = child_functions(&cfg, Some(&parent), None, Some(Mode::Ask));
         let compiled = policy::CompiledPolicy::from(asked.as_ref());
         assert!(compiled.allows("state::get"));
-        assert!(!compiled.allows("state::set"));
-        assert!(!compiled.allows("harness::spawn"));
+        assert!(compiled.allows("state::set"));
+        assert!(compiled.allows("harness::spawn"));
     }
 
     #[test]
@@ -611,18 +626,17 @@ mod tests {
     }
 
     #[test]
-    fn parentless_ask_child_keeps_the_baseline() {
+    fn parentless_ask_child_keeps_the_default_policy() {
         let cfg = WorkerConfig::default();
         let f = child_functions(&cfg, None, None, Some(Mode::Ask));
         let compiled = policy::CompiledPolicy::from(f.as_ref());
         assert!(compiled.allows("state::get"));
-        assert!(!compiled.allows("state::set"));
+        assert!(compiled.allows("state::set"));
     }
 
     #[test]
     fn ask_mode_child_clamps_an_explicit_broad_request() {
-        // The direct escalation vector: a spawn that explicitly asks for `*`.
-        // Both a parented and a parentless ask child must still be capped.
+        // A wildcard request remains unrestricted under the wildcard default.
         let cfg = WorkerConfig::default();
         let mut parent = parent_record(None);
         parent.options.functions = Some(broad_policy());
@@ -630,8 +644,8 @@ mod tests {
             let f = child_functions(&cfg, parent_rec, Some(&broad_policy()), Some(Mode::Ask));
             let compiled = policy::CompiledPolicy::from(f.as_ref());
             assert!(compiled.allows("state::get"));
-            assert!(!compiled.allows("state::set"));
-            assert!(!compiled.allows("harness::spawn"));
+            assert!(compiled.allows("state::set"));
+            assert!(compiled.allows("harness::spawn"));
         }
     }
 
@@ -645,6 +659,6 @@ mod tests {
         let defaulted = child_functions(&cfg, None, None, None);
         let compiled = policy::CompiledPolicy::from(defaulted.as_ref());
         assert!(compiled.allows("state::get"));
-        assert!(!compiled.allows("state::set"));
+        assert!(compiled.allows("state::set"));
     }
 }
diff --git a/harness/src/surface.rs b/harness/src/surface.rs
index 121205453..27726daba 100644
--- a/harness/src/surface.rs
+++ b/harness/src/surface.rs
@@ -10,14 +10,17 @@ use crate::functions::react::REACT_ID;
 use crate::functions::{
     function_resolve::{FunctionResolveRequest, FunctionResolveResponse},
     function_trigger::{FunctionTriggerRequest, FunctionTriggerResponse},
+    metrics::{SessionMetricsRequestV1, SessionMetricsResponseV1},
     react::{ReactEvent, ReactResult},
     send::{SendRequest, SendResponse},
+    session_tree::{SessionTreeRequestV1, SessionTreeResponseV1},
     spawn::{SpawnRequest, SpawnResponse},
     status::{StatusReport, StatusRequest},
     stop::{StopRequest, StopResponse},
 };
 use crate::functions::{
-    FUNCTION_RESOLVE_ID, FUNCTION_TRIGGER_ID, SEND_ID, SPAWN_ID, STATUS_ID, STOP_ID, TURN_ID,
+    FUNCTION_RESOLVE_ID, FUNCTION_TRIGGER_ID, METRICS_ID, SEND_ID, SESSION_TREE_ID, SPAWN_ID,
+    STATUS_ID, STOP_ID, TURN_ID,
 };
 use crate::turn_loop::{TurnStepPayload, TurnStepResult};
 
@@ -64,6 +67,8 @@ pub fn catalog() -> Vec<FunctionSpec> {
         spec::<FunctionResolveRequest, FunctionResolveResponse>(FUNCTION_RESOLVE_ID),
         spec::<StopRequest, StopResponse>(STOP_ID),
         spec::<StatusRequest, Option<StatusReport>>(STATUS_ID),
+        spec::<SessionTreeRequestV1, SessionTreeResponseV1>(SESSION_TREE_ID),
+        spec::<SessionMetricsRequestV1, SessionMetricsResponseV1>(METRICS_ID),
         spec::<ReactEvent, ReactResult>(REACT_ID),
     ]
 }
diff --git a/harness/src/turn_loop.rs b/harness/src/turn_loop.rs
index 3165bc71e..3989db397 100644
--- a/harness/src/turn_loop.rs
+++ b/harness/src/turn_loop.rs
@@ -47,6 +47,18 @@ const CONTEXT_OVERFLOW_FAILURE: FailureInfo = FailureInfo {
     retryable: false,
 };
 
+const BUDGET_EXCEEDED_FAILURE: FailureInfo = FailureInfo {
+    code: "harness.budget_exceeded",
+    phase: "budget_preflight",
+    retryable: false,
+};
+
+const BUDGET_UNAVAILABLE_FAILURE: FailureInfo = FailureInfo {
+    code: "harness.budget_unavailable",
+    phase: "budget_preflight",
+    retryable: false,
+};
+
 /// Provider adapters add framing outside the fields visible to the harness.
 /// Keep a small fixed reserve in addition to the deterministic JSON estimate.
 const PROVIDER_FRAMING_ALLOWANCE_TOKENS: u64 = 64;
@@ -315,7 +327,13 @@ pub async fn run_step(
         PRE_GENERATE_HOOK_ALLOWANCE_TOKENS
     };
     let mut reassembled = false;
-    let (gen_system_prompt, gen_annotations, gen_messages) = loop {
+    let (
+        gen_system_prompt,
+        gen_annotations,
+        gen_messages,
+        generation_input_tokens,
+        generation_max_output_tokens,
+    ) = loop {
         let assembled = match assemble_context(
             deps,
             &session,
@@ -448,7 +466,18 @@ pub async fn run_step(
             final_count.tokens.saturating_add(request_overhead_tokens)
         };
         if final_request_tokens <= assembled.usable {
-            break (gen_system_prompt, gen_annotations, gen_messages);
+            let max_output_tokens = record
+                .options
+                .max_output_tokens
+                .unwrap_or(assembled.effective_max_output_tokens)
+                .min(assembled.effective_max_output_tokens);
+            break (
+                gen_system_prompt,
+                gen_annotations,
+                gen_messages,
+                final_request_tokens,
+                max_output_tokens,
+            );
         }
         if reassembled {
             let reason = format!(
@@ -483,6 +512,33 @@ pub async fn run_step(
         );
     };
 
+    // A stop that landed during context assembly is visible through the
+    // in-process signal even though its durable write is blocked by this
+    // session lock. Do not reserve budget for a call that will never start.
+    if deps.cancels.is_fired(&record.turn_id) {
+        record.abort = true;
+        return finalize_cancelled(deps, &session, &mut record, "cancelled").await;
+    }
+
+    let budget_reservation = match crate::budget::reserve(
+        deps,
+        &record,
+        generation_input_tokens,
+        generation_max_output_tokens,
+    )
+    .await?
+    {
+        crate::budget::ReserveOutcome::Unlimited => None,
+        crate::budget::ReserveOutcome::Reserved(reservation) => Some(reservation),
+        crate::budget::ReserveOutcome::Rejected(rejection) => {
+            let failure = match &rejection {
+                crate::budget::BudgetRejection::Exceeded(_) => BUDGET_EXCEEDED_FAILURE,
+                crate::budget::BudgetRejection::Unavailable(_) => BUDGET_UNAVAILABLE_FAILURE,
+            };
+            return finalize_failed(deps, &session, &mut record, rejection.reason(), failure).await;
+        }
+    };
+
     let assistant_origin = origin_with(&record.turn_id, &gen_annotations);
 
     // Generate: append an empty assistant under a deterministic id, stream
@@ -490,7 +546,7 @@ pub async fn run_step(
     let assistant_id = ids::assistant_entry_id(&record.turn_id, payload.step);
     let provider = record.options.provider.clone().unwrap_or_default();
     let empty = empty_assistant(&provider, &record.options.model);
-    let _ = session
+    if let Err(error) = session
         .append(
             &record.session_id,
             &AgentMessage::Assistant(empty),
@@ -498,7 +554,13 @@ pub async fn run_step(
             None,
             Some(&assistant_origin),
         )
-        .await?;
+        .await
+    {
+        if let Some(reservation) = budget_reservation.as_ref() {
+            crate::budget::release(deps, reservation).await?;
+        }
+        return Err(error);
+    }
 
     let sink = SessionStreamSink {
         session: session.clone(),
@@ -514,17 +576,34 @@ pub async fn run_step(
         messages: gen_messages,
         tools,
         response_format,
+        // Forward a cap only when the caller set one. `generation_max_output_tokens`
+        // is the internal reservation context assembly budgets against; sending it
+        // unasked would put the model's own ceiling on every request and let a
+        // provider apply a different policy than its default. When the caller DID
+        // ask, send the reservation rather than the raw request — it is their value
+        // already clamped to the model's effective limit, so the provider is never
+        // told it may emit more than we reserved.
+        max_output_tokens: record
+            .options
+            .max_output_tokens
+            .map(|_| generation_max_output_tokens),
         thinking_level: record.options.thinking_level,
         provider_options,
     };
     record.stream_request_id = Some(params.request_id.clone());
-    crate::state::put_turn(&deps.iii, &record, cfg.session_timeout_ms).await?;
+    if let Err(error) = crate::state::put_turn(&deps.iii, &record, cfg.session_timeout_ms).await {
+        if let Some(reservation) = budget_reservation.as_ref() {
+            crate::budget::release(deps, reservation).await?;
+        }
+        return Err(error);
+    }
 
-    // A stop that landed during context assembly (its durable write blocked on
-    // the guard held here) is only visible via the in-process signal. Bail
-    // before dispatching generation at all — otherwise the provider starts a
-    // stream that is dead on arrival.
+    // A stop may land after reservation but before dispatch. Release the
+    // unused amount before finalising the cancelled turn.
     if deps.cancels.is_fired(&record.turn_id) {
+        if let Some(reservation) = budget_reservation.as_ref() {
+            crate::budget::release(deps, reservation).await?;
+        }
         record.abort = true;
         return finalize_cancelled(deps, &session, &mut record, "cancelled").await;
     }
@@ -549,7 +628,21 @@ pub async fn run_step(
     // was a no-op (registration race / router restart). Level-triggered, so a
     // stop fired before this subscribe is still observed.
     let abort_rx = deps.cancels.watch(&record.turn_id);
-    let mut outcome = router.chat(params, &sink, abort_rx).await?;
+    let mut outcome = match router.chat(params, &sink, abort_rx).await {
+        Ok(outcome) => outcome,
+        Err(error) => {
+            // The provider may have consumed the request without returning
+            // usage. Charge the full reservation so transport failures cannot
+            // bypass a hard budget.
+            if let Some(reservation) = budget_reservation.as_ref() {
+                crate::budget::reconcile(deps, reservation, None).await?;
+            }
+            return Err(error);
+        }
+    };
+    if let Some(reservation) = budget_reservation.as_ref() {
+        crate::budget::reconcile(deps, reservation, outcome.message.usage.as_ref()).await?;
+    }
 
     // A stop mid-generation truncates tool-call blocks before their arguments
     // stream. An argument-less call can never execute (the turn is cancelled)
@@ -575,6 +668,7 @@ pub async fn run_step(
             &record.session_id,
             &assistant_id,
             &outcome.message.content,
+            outcome.message.usage.as_ref(),
             None,
             Some(&assistant_origin),
         )
@@ -1996,6 +2090,7 @@ async fn assemble_context(
         messages,
         usable: out.usable,
         token_count: out.token_count,
+        effective_max_output_tokens: out.effective_max_output_tokens,
     })
 }
 
@@ -2072,6 +2167,8 @@ struct Assembled {
     /// prompt + tools + request overhead). Reused as the final request
     /// count when nothing mutates the request after assembly.
     token_count: u64,
+    /// Model/output ceiling resolved by context-manager for this request.
+    effective_max_output_tokens: u64,
 }
 
 struct ContextAssemblyInputs<'a> {
@@ -2250,11 +2347,18 @@ async fn build_tools(deps: &Deps, record: &TurnRecord) -> Vec<crate::types::mode
         ExposeMode::Native => {
             let policy = CompiledPolicy::from(record.options.functions.as_ref());
             let snapshot = deps.functions().await;
-            let mut tools = Vec::new();
+            // Subscription controls are harness-intercepted virtual functions,
+            // so the engine's public registry intentionally does not list
+            // them. Publish their real schemas alongside registry functions
+            // whenever this turn's dispatch policy allows them.
+            let mut tools = crate::functions::subscribe::native_control_tools(&policy);
             for descriptor in snapshot.functions.iter() {
                 if !policy.allows(&descriptor.function_id) {
                     continue;
                 }
+                if tools.iter().any(|tool| tool.name == descriptor.function_id) {
+                    continue;
+                }
                 tools.push(crate::types::model::AgentFunction {
                     name: descriptor.function_id.clone(),
                     description: descriptor.description.clone().unwrap_or_default(),
@@ -2296,6 +2400,7 @@ impl StreamSink for SessionStreamSink {
                 &self.session_id,
                 &self.entry_id,
                 &message.content,
+                message.usage.as_ref(),
                 None,
                 Some(&origin),
             )
diff --git a/harness/src/types/turn.rs b/harness/src/types/turn.rs
index 70d6b0820..ead505eba 100644
--- a/harness/src/types/turn.rs
+++ b/harness/src/types/turn.rs
@@ -75,6 +75,20 @@ pub struct TurnOptions {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub mode: Option<Mode>,
     pub max_turns: u32,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub max_output_tokens: Option<u64>,
+    /// Hard input-plus-output token budget shared by this root turn and every
+    /// in-turn sub-agent it spawns.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub max_total_tokens: Option<u64>,
+    /// Hard USD budget shared by this root turn and every in-turn sub-agent it
+    /// spawns. Requires catalog pricing for every model used by the tree.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub max_cost_usd: Option<f64>,
+    /// Root session whose durable budget ledger this turn charges. Internal
+    /// bookkeeping populated by `harness::send` and inherited by sub-agents.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub budget_root_session_id: Option<String>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub thinking_level: Option<ThinkingLevel>,
     #[serde(skip_serializing_if = "Option::is_none")]
@@ -312,6 +326,10 @@ mod tests {
                 system_prompt: None,
                 mode: None,
                 max_turns: 16,
+                max_output_tokens: None,
+                max_total_tokens: None,
+                max_cost_usd: None,
+                budget_root_session_id: None,
                 thinking_level: None,
                 provider_options: None,
                 output: OutputContract::Text,
diff --git a/harness/tests/agent-quality/Cargo.toml b/harness/tests/agent-quality/Cargo.toml
new file mode 100644
index 000000000..32b5383e0
--- /dev/null
+++ b/harness/tests/agent-quality/Cargo.toml
@@ -0,0 +1,36 @@
+[package]
+name = "harness-agent-quality"
+version = "0.1.0"
+edition = "2021"
+publish = false
+
+[[bin]]
+name = "harness-agent-quality"
+path = "src/bin/agent-quality.rs"
+
+[[bin]]
+name = "harness-prompt-eval"
+path = "src/bin/prompt-eval.rs"
+
+[lib]
+name = "harness_agent_quality"
+path = "src/lib.rs"
+
+[dependencies]
+harness = { path = "../.." }
+eval = { path = "../../../eval" }
+iii-sdk = "=0.21.6"
+tokio = { version = "1", features = ["rt-multi-thread", "macros", "sync", "time"] }
+serde = { version = "1", features = ["derive"] }
+serde_json = "1"
+anyhow = "1"
+thiserror = "2"
+clap = { version = "4", features = ["derive", "env"] }
+schemars = "0.8"
+sha2 = "0.10"
+uuid = { version = "1", features = ["v4"] }
+tracing = "0.1"
+tracing-subscriber = { version = "0.3", features = ["fmt", "env-filter"] }
+
+[dev-dependencies]
+tempfile = "3"
diff --git a/harness/tests/agent-quality/README.md b/harness/tests/agent-quality/README.md
new file mode 100644
index 000000000..2b5d1a8d1
--- /dev/null
+++ b/harness/tests/agent-quality/README.md
@@ -0,0 +1,138 @@
+# Harness prompt evaluation
+
+This package compares two harness subjects against the same real-model
+scenarios. It supports exactly two experiments:
+
+- Same system prompt with two different models.
+- Same model with two different system prompts.
+
+Every scenario follows one execution path:
+
+1. Build the scenario prompt.
+2. Call `harness::send` once.
+3. Wait for `harness::turn-completed`.
+4. Read `harness::metrics` and the root transcript.
+5. Evaluate whether the requested result works.
+
+The common runner owns steps 2–4. A scenario defines only its prompt and its
+result evaluation. Evaluations may inspect the transcript or call existing iii
+functions to test durable output.
+
+## Compare subjects
+
+```bash
+cargo run -p harness-agent-quality --bin harness-prompt-eval -- \
+  --control subjects/luna.json \
+  --treatment subjects/sol.json \
+  --runs 3 \
+  --output target/prompt-eval \
+  --max-total-tokens 100000
+```
+
+The comparison type is inferred from the subject files. Every setting other
+than the model or system-prompt contents must remain equal. Ambiguous,
+identical, or multi-variable comparisons fail before a model call.
+
+Use repeatable `--scenario` arguments for a focused comparison:
+
+```bash
+cargo run -p harness-agent-quality --bin harness-prompt-eval -- \
+  --control path/to/control.json \
+  --treatment path/to/treatment.json \
+  --runs 3 \
+  --output target/prompt-eval \
+  --scenario plain_response \
+  --scenario security_review
+```
+
+The output contains `comparison.json` and one self-contained `results.json`
+for every control and treatment run. The comparison reports pass counts and
+median execution, token, function-call, error, cost, trace, span, and trace
+duration metrics per scenario when the engine's in-memory observability
+exporter is available.
+The treatment passes the correctness gate only when every treatment run passes
+and no scenario regresses relative to the control.
+
+## Limits
+
+Every control and treatment run uses one shared limit policy. Limits are not
+part of a subject, so a comparison still changes only the model or system
+prompt.
+
+Execution limits constrain the run while it is active:
+
+- `--scenario-timeout-seconds` defaults to `300`; expiry calls `harness::stop`.
+- `--invocation-timeout-seconds` defaults to `120` for each iii invocation.
+- `--max-turns` defaults to `20`.
+- `--max-output-tokens-per-call` defaults to `8192`.
+- `--max-total-tokens` defaults to `100000` and means input plus output tokens
+  over the complete root-and-descendant session tree. The harness reserves
+  each call before dispatch and stops the turn before this budget can be
+  exceeded.
+- `--max-cost-usd` is optional. The harness reserves cost before dispatch and
+  fails closed when any selected model lacks catalog pricing.
+
+Evaluation limits use `harness::metrics` after the requested result has been
+evaluated:
+
+- `--max-function-call-errors` defaults to `0`.
+- `--max-error-spans` is optional because it requires the engine's in-memory
+  observability exporter.
+
+Token and cost limits are also checked against `harness::metrics` after the
+run, while the remaining evaluation limits fail the completed scenario.
+Configuring an optional gate without its required metric fails closed as an
+evidence error. Every `results.json` and `comparison.json` records the
+effective limit policy.
+
+## Scenarios
+
+The current scenarios are:
+
+- `plain_response`: checks an exact text response without tools.
+- `single_function`: checks one durable state mutation.
+- `security_review`: checks structured reasoning from the final response.
+- `triggered_work`: checks the durable result of reactive work.
+
+`single_function` and `triggered_work` exercise multiple internal model turns,
+but each scenario still starts with exactly one runner call to `harness::send`.
+
+To add a scenario, add a module that returns a `ScenarioSpec` containing:
+
+- the prompt;
+- private data needed by the evaluation;
+- an async evaluation function.
+
+Do not reproduce send, completion, metrics, transcript, timeout, or report
+handling in scenario modules.
+
+## Subjects
+
+Subject files resolve `system_prompt_path` relative to the subject file:
+
+```json
+{
+  "schema_version": "1",
+  "subject_id": "baseline",
+  "model": "resolved-model-id",
+  "provider": "provider-route",
+  "system_prompt_path": "system-prompt.md",
+  "system_prompt_strategy": "override",
+  "thinking_level": "low"
+}
+```
+
+For a model comparison, both subjects use the same provider and exact prompt
+contents but different `model` values. For a system-prompt comparison, they use
+the same model and provider but different prompt contents.
+
+## Run one subject
+
+The lower-level runner is useful for developing or diagnosing a scenario:
+
+```bash
+cargo run -p harness-agent-quality -- \
+  --subject subjects/luna.json \
+  --output target/agent-quality \
+  --scenario single_function
+```
diff --git a/harness/tests/agent-quality/src/bin/agent-quality.rs b/harness/tests/agent-quality/src/bin/agent-quality.rs
new file mode 100644
index 000000000..e59414990
--- /dev/null
+++ b/harness/tests/agent-quality/src/bin/agent-quality.rs
@@ -0,0 +1,61 @@
+use std::path::PathBuf;
+
+use anyhow::{bail, Context, Result};
+use clap::Parser;
+use harness_agent_quality::scenarios::{self, ScenarioId};
+use harness_agent_quality::{run_suite, subject, LimitsArgs, SuiteRunConfig};
+
+#[derive(Debug, Parser)]
+#[command(
+    name = "harness-agent-quality",
+    about = "Run real-model harness E2E scenarios"
+)]
+struct Cli {
+    #[arg(long)]
+    subject: PathBuf,
+
+    #[arg(long)]
+    output: PathBuf,
+
+    #[arg(long, env = "III_URL", default_value = "ws://127.0.0.1:49134")]
+    url: String,
+
+    #[command(flatten)]
+    limits: LimitsArgs,
+
+    /// Run only the selected scenario. Repeat to select more than one.
+    #[arg(long, value_enum)]
+    scenario: Vec<ScenarioId>,
+}
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    tracing_subscriber::fmt()
+        .with_env_filter(
+            tracing_subscriber::EnvFilter::try_from_default_env()
+                .unwrap_or_else(|_| tracing_subscriber::EnvFilter::new("info")),
+        )
+        .init();
+    let cli = Cli::parse();
+    let subject = subject::load(&cli.subject).map_err(anyhow::Error::msg)?;
+    let limits = cli.limits.resolve().map_err(anyhow::Error::msg)?;
+    let outcome = run_suite(SuiteRunConfig {
+        url: cli.url,
+        subject,
+        output: cli.output,
+        scenarios: scenarios::selected(&cli.scenario),
+        limits,
+    })
+    .await
+    .map_err(anyhow::Error::msg)
+    .context("run agent-quality suite")?;
+
+    if !outcome.report.passed {
+        bail!(
+            "agent-quality scenario failed; see {}/results.json",
+            outcome.run_dir.display()
+        );
+    }
+    tracing::info!(path = %outcome.run_dir.display(), "agent-quality run passed");
+    Ok(())
+}
diff --git a/harness/tests/agent-quality/src/bin/prompt-eval.rs b/harness/tests/agent-quality/src/bin/prompt-eval.rs
new file mode 100644
index 000000000..81ebcbd29
--- /dev/null
+++ b/harness/tests/agent-quality/src/bin/prompt-eval.rs
@@ -0,0 +1,126 @@
+use std::path::PathBuf;
+
+use anyhow::{bail, Context, Result};
+use clap::Parser;
+use harness_agent_quality::scenarios::{self, ScenarioId};
+use harness_agent_quality::{
+    compare_runs, comparison_dimension, run_suite, subject, AgentQualityLimitsV1,
+    AgentQualityRunReportV1, LimitsArgs, ResolvedAgentQualitySubjectV1, SuiteRunConfig,
+};
+
+#[derive(Debug, Parser)]
+#[command(
+    name = "harness-prompt-eval",
+    about = "Compare two harness subjects while changing only the model or system prompt"
+)]
+struct Cli {
+    #[arg(long)]
+    control: PathBuf,
+    #[arg(long)]
+    treatment: PathBuf,
+    #[arg(long, default_value_t = 3)]
+    runs: u32,
+    #[arg(long)]
+    output: PathBuf,
+    #[arg(long, env = "III_URL", default_value = "ws://127.0.0.1:49134")]
+    url: String,
+    #[command(flatten)]
+    limits: LimitsArgs,
+    #[arg(long, value_enum)]
+    scenario: Vec<ScenarioId>,
+}
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    tracing_subscriber::fmt()
+        .with_env_filter(
+            tracing_subscriber::EnvFilter::try_from_default_env()
+                .unwrap_or_else(|_| tracing_subscriber::EnvFilter::new("info")),
+        )
+        .init();
+    let cli = Cli::parse();
+    if cli.runs == 0 {
+        bail!("--runs must be greater than zero");
+    }
+    let limits = cli.limits.resolve().map_err(anyhow::Error::msg)?;
+
+    let control = subject::load(&cli.control).map_err(anyhow::Error::msg)?;
+    let treatment = subject::load(&cli.treatment).map_err(anyhow::Error::msg)?;
+    let dimension = comparison_dimension(&control, &treatment).map_err(anyhow::Error::msg)?;
+
+    let scenarios = scenarios::selected(&cli.scenario);
+    let mut control_runs = Vec::with_capacity(cli.runs as usize);
+    let mut treatment_runs = Vec::with_capacity(cli.runs as usize);
+
+    for repetition in 0..cli.runs {
+        let order = if repetition % 2 == 0 {
+            [true, false]
+        } else {
+            [false, true]
+        };
+        for is_control in order {
+            let (variant, subject) = if is_control {
+                ("control", control.clone())
+            } else {
+                ("treatment", treatment.clone())
+            };
+            let report =
+                run_variant(&cli, variant, repetition, subject, &scenarios, limits).await?;
+            if is_control {
+                control_runs.push(report);
+            } else {
+                treatment_runs.push(report);
+            }
+        }
+    }
+
+    let comparison = compare_runs(
+        dimension,
+        control.artifact(),
+        treatment.artifact(),
+        &scenarios,
+        &control_runs,
+        &treatment_runs,
+    )
+    .map_err(anyhow::Error::msg)?;
+    let path = comparison
+        .write_to(&cli.output)
+        .map_err(anyhow::Error::msg)
+        .context("write prompt comparison")?;
+
+    if !comparison.eligible {
+        bail!(
+            "treatment failed the correctness gate; see {}",
+            path.display()
+        );
+    }
+    tracing::info!(path = %path.display(), "prompt comparison passed");
+    Ok(())
+}
+
+async fn run_variant(
+    cli: &Cli,
+    variant: &str,
+    repetition: u32,
+    subject: ResolvedAgentQualitySubjectV1,
+    scenarios: &[ScenarioId],
+    limits: AgentQualityLimitsV1,
+) -> Result<AgentQualityRunReportV1> {
+    tracing::info!(
+        variant,
+        run = repetition + 1,
+        total_runs = cli.runs,
+        "running prompt evaluation variant"
+    );
+    let outcome = run_suite(SuiteRunConfig {
+        url: cli.url.clone(),
+        subject,
+        output: cli.output.join("runs").join(variant),
+        scenarios: scenarios.to_vec(),
+        limits,
+    })
+    .await
+    .map_err(anyhow::Error::msg)
+    .with_context(|| format!("run {variant} repetition {}", repetition + 1))?;
+    Ok(outcome.report)
+}
diff --git a/harness/tests/agent-quality/src/comparison.rs b/harness/tests/agent-quality/src/comparison.rs
new file mode 100644
index 000000000..5c9047b01
--- /dev/null
+++ b/harness/tests/agent-quality/src/comparison.rs
@@ -0,0 +1,391 @@
+use std::fs;
+use std::path::{Path, PathBuf};
+
+use schemars::JsonSchema;
+use serde::Serialize;
+
+use crate::error::EvalError;
+use crate::limits::AgentQualityLimitsV1;
+use crate::report::AgentQualityRunReportV1;
+use crate::scenarios::ScenarioId;
+use crate::subject::{ResolvedAgentQualitySubjectV1, SubjectArtifactV1};
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum ComparisonDimension {
+    Model,
+    SystemPrompt,
+}
+
+#[derive(Debug, Clone, Serialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct ScenarioAggregateV1 {
+    pub runs: u32,
+    pub passed: u32,
+    pub pass_rate: f64,
+    pub benchmarked_runs: u32,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_wall_time_ms: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_input_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_output_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_total_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_reasoning_tokens: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_function_calls: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_function_call_errors: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_cost_usd: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_trace_count: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_span_count: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_error_span_count: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub median_trace_duration_ms: Option<u64>,
+}
+
+#[derive(Debug, Clone, Serialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct ScenarioComparisonV1 {
+    pub scenario_id: String,
+    pub control: ScenarioAggregateV1,
+    pub treatment: ScenarioAggregateV1,
+}
+
+#[derive(Debug, Clone, Serialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct PromptComparisonReportV1 {
+    pub schema_version: String,
+    pub dimension: ComparisonDimension,
+    pub control: SubjectArtifactV1,
+    pub treatment: SubjectArtifactV1,
+    pub limits: AgentQualityLimitsV1,
+    pub eligible: bool,
+    pub regressed_scenarios: Vec<String>,
+    pub scenarios: Vec<ScenarioComparisonV1>,
+}
+
+impl PromptComparisonReportV1 {
+    pub fn write_to(&self, output: &Path) -> Result<PathBuf, EvalError> {
+        fs::create_dir_all(output).map_err(|error| {
+            EvalError::setup(format!(
+                "create comparison directory {}: {error}",
+                output.display()
+            ))
+        })?;
+        let path = output.join("comparison.json");
+        let mut bytes = serde_json::to_vec_pretty(self)
+            .map_err(|error| EvalError::setup(format!("serialize {}: {error}", path.display())))?;
+        bytes.push(b'\n');
+        fs::write(&path, bytes)
+            .map_err(|error| EvalError::setup(format!("write {}: {error}", path.display())))?;
+        Ok(path)
+    }
+}
+
+pub fn comparison_dimension(
+    control: &ResolvedAgentQualitySubjectV1,
+    treatment: &ResolvedAgentQualitySubjectV1,
+) -> Result<ComparisonDimension, EvalError> {
+    require_equal(
+        "system_prompt_strategy",
+        control.system_prompt_strategy,
+        treatment.system_prompt_strategy,
+    )?;
+    require_equal(
+        "thinking_level",
+        control.thinking_level,
+        treatment.thinking_level,
+    )?;
+    require_equal(
+        "provider_options",
+        &control.provider_options,
+        &treatment.provider_options,
+    )?;
+    require_equal("provider", &control.provider, &treatment.provider)?;
+
+    let prompt_changed = control.system_prompt_sha256 != treatment.system_prompt_sha256;
+    let model_changed = control.model != treatment.model;
+
+    if !prompt_changed && model_changed {
+        return Ok(ComparisonDimension::Model);
+    }
+    if prompt_changed && !model_changed {
+        return Ok(ComparisonDimension::SystemPrompt);
+    }
+    Err(EvalError::setup(
+        "comparison must change either the model or the system prompt, but not both",
+    ))
+}
+
+pub fn compare_runs(
+    dimension: ComparisonDimension,
+    control: SubjectArtifactV1,
+    treatment: SubjectArtifactV1,
+    scenarios: &[ScenarioId],
+    control_runs: &[AgentQualityRunReportV1],
+    treatment_runs: &[AgentQualityRunReportV1],
+) -> Result<PromptComparisonReportV1, EvalError> {
+    if control_runs.is_empty() || control_runs.len() != treatment_runs.len() {
+        return Err(EvalError::setup(
+            "comparison requires the same non-zero number of control and treatment runs",
+        ));
+    }
+    let limits = control_runs[0].limits;
+    if control_runs
+        .iter()
+        .chain(treatment_runs)
+        .any(|report| report.limits != limits)
+    {
+        return Err(EvalError::setup(
+            "every control and treatment run must use the same limits",
+        ));
+    }
+
+    let comparisons: Vec<_> = scenarios
+        .iter()
+        .map(|scenario| {
+            let scenario_id = scenario.as_str();
+            Ok(ScenarioComparisonV1 {
+                scenario_id: scenario_id.to_string(),
+                control: aggregate(control_runs, scenario_id)?,
+                treatment: aggregate(treatment_runs, scenario_id)?,
+            })
+        })
+        .collect::<Result<_, EvalError>>()?;
+    let regressed_scenarios: Vec<_> = comparisons
+        .iter()
+        .filter(|comparison| comparison.treatment.passed < comparison.control.passed)
+        .map(|comparison| comparison.scenario_id.clone())
+        .collect();
+    let treatment_passed_all = comparisons
+        .iter()
+        .all(|comparison| comparison.treatment.passed == comparison.treatment.runs);
+
+    Ok(PromptComparisonReportV1 {
+        schema_version: "1".into(),
+        dimension,
+        control,
+        treatment,
+        limits,
+        eligible: regressed_scenarios.is_empty() && treatment_passed_all,
+        regressed_scenarios,
+        scenarios: comparisons,
+    })
+}
+
+fn aggregate(
+    reports: &[AgentQualityRunReportV1],
+    scenario_id: &str,
+) -> Result<ScenarioAggregateV1, EvalError> {
+    let scenarios: Vec<_> = reports
+        .iter()
+        .map(|report| {
+            report
+                .scenarios
+                .iter()
+                .find(|scenario| scenario.scenario_id == scenario_id)
+                .ok_or_else(|| {
+                    EvalError::setup(format!(
+                        "run {} has no scenario {scenario_id}",
+                        report.run_id
+                    ))
+                })
+        })
+        .collect::<Result<_, _>>()?;
+    let benchmarks: Vec<_> = scenarios
+        .iter()
+        .filter_map(|scenario| scenario.benchmark())
+        .collect();
+    let runs = reports.len() as u32;
+    let passed = scenarios.iter().filter(|scenario| scenario.passed).count() as u32;
+
+    Ok(ScenarioAggregateV1 {
+        runs,
+        passed,
+        pass_rate: passed as f64 / runs as f64,
+        benchmarked_runs: benchmarks.len() as u32,
+        median_wall_time_ms: median_u64(benchmarks.iter().map(|value| value.wall_time_ms)),
+        median_input_tokens: median_u64(benchmarks.iter().filter_map(|value| value.input_tokens)),
+        median_output_tokens: median_u64(benchmarks.iter().filter_map(|value| value.output_tokens)),
+        median_total_tokens: median_u64(benchmarks.iter().filter_map(|value| value.total_tokens)),
+        median_reasoning_tokens: median_u64(
+            benchmarks.iter().filter_map(|value| value.reasoning_tokens),
+        ),
+        median_function_calls: median_u64(benchmarks.iter().map(|value| value.function_calls)),
+        median_function_call_errors: median_u64(
+            benchmarks.iter().map(|value| value.function_call_errors),
+        ),
+        median_cost_usd: median_f64(benchmarks.iter().filter_map(|value| value.cost_usd)),
+        median_trace_count: median_u64(benchmarks.iter().filter_map(|value| value.trace_count)),
+        median_span_count: median_u64(benchmarks.iter().filter_map(|value| value.span_count)),
+        median_error_span_count: median_u64(
+            benchmarks.iter().filter_map(|value| value.error_span_count),
+        ),
+        median_trace_duration_ms: median_u64(
+            benchmarks
+                .iter()
+                .filter_map(|value| value.trace_duration_ms),
+        ),
+    })
+}
+
+fn median_u64(values: impl IntoIterator<Item = u64>) -> Option<u64> {
+    let mut values: Vec<_> = values.into_iter().collect();
+    if values.is_empty() {
+        return None;
+    }
+    values.sort_unstable();
+    let middle = values.len() / 2;
+    if values.len() % 2 == 1 {
+        Some(values[middle])
+    } else {
+        Some(values[middle - 1].saturating_add(values[middle]) / 2)
+    }
+}
+
+fn median_f64(values: impl IntoIterator<Item = f64>) -> Option<f64> {
+    let mut values: Vec<_> = values
+        .into_iter()
+        .filter(|value| value.is_finite())
+        .collect();
+    if values.is_empty() {
+        return None;
+    }
+    values.sort_by(f64::total_cmp);
+    let middle = values.len() / 2;
+    if values.len() % 2 == 1 {
+        Some(values[middle])
+    } else {
+        Some((values[middle - 1] + values[middle]) / 2.0)
+    }
+}
+
+fn require_equal<T: PartialEq>(name: &str, control: T, treatment: T) -> Result<(), EvalError> {
+    if control == treatment {
+        Ok(())
+    } else {
+        Err(EvalError::setup(format!(
+            "{name} must remain equal between control and treatment"
+        )))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::BTreeMap;
+
+    use harness::prompt::SystemPromptStrategy;
+    use harness::types::model::ThinkingLevel;
+
+    use super::*;
+    use crate::report::AgentQualityScenarioReportV1;
+
+    fn subject(model: &str, provider: &str, prompt_hash: &str) -> ResolvedAgentQualitySubjectV1 {
+        ResolvedAgentQualitySubjectV1 {
+            subject_id: format!("{model}-{prompt_hash}"),
+            subject_sha256: "a".repeat(64),
+            system_prompt_sha256: prompt_hash.into(),
+            model: model.into(),
+            provider: provider.into(),
+            system_prompt: "prompt".into(),
+            system_prompt_strategy: SystemPromptStrategy::Override,
+            thinking_level: Some(ThinkingLevel::Low),
+            provider_options: Some(BTreeMap::new()),
+        }
+    }
+
+    #[test]
+    fn model_comparison_changes_only_the_model_route() {
+        let control = subject("model-a", "provider", "prompt-a");
+        let treatment = subject("model-b", "provider", "prompt-a");
+        assert_eq!(
+            comparison_dimension(&control, &treatment).unwrap(),
+            ComparisonDimension::Model
+        );
+    }
+
+    #[test]
+    fn system_prompt_comparison_changes_only_the_prompt() {
+        let control = subject("model-a", "provider", "prompt-a");
+        let treatment = subject("model-a", "provider", "prompt-b");
+        assert_eq!(
+            comparison_dimension(&control, &treatment).unwrap(),
+            ComparisonDimension::SystemPrompt
+        );
+    }
+
+    #[test]
+    fn rejects_ambiguous_or_identical_comparisons() {
+        let baseline = subject("model-a", "provider", "prompt-a");
+        assert!(comparison_dimension(&baseline, &baseline).is_err());
+        assert!(
+            comparison_dimension(&baseline, &subject("model-b", "provider", "prompt-b")).is_err()
+        );
+        assert!(
+            comparison_dimension(&baseline, &subject("model-a", "other-provider", "prompt-a"))
+                .is_err()
+        );
+    }
+
+    #[test]
+    fn median_handles_odd_even_and_missing_values() {
+        assert_eq!(median_u64([]), None);
+        assert_eq!(median_u64([3]), Some(3));
+        assert_eq!(median_u64([9, 1, 5]), Some(5));
+        assert_eq!(median_u64([10, 2]), Some(6));
+        assert_eq!(median_f64([]), None);
+        assert_eq!(median_f64([1.0, 3.0]), Some(2.0));
+    }
+
+    #[test]
+    fn treatment_must_pass_every_run_without_regressing() {
+        let control_subject = subject("model-a", "provider", "prompt").artifact();
+        let treatment_subject = subject("model-b", "provider", "prompt").artifact();
+        let control = [
+            report("control-1", true, control_subject.clone()),
+            report("control-2", true, control_subject.clone()),
+        ];
+        let treatment = [
+            report("treatment-1", true, treatment_subject.clone()),
+            report("treatment-2", false, treatment_subject.clone()),
+        ];
+
+        let comparison = compare_runs(
+            ComparisonDimension::Model,
+            control_subject,
+            treatment_subject,
+            &[ScenarioId::PlainResponse],
+            &control,
+            &treatment,
+        )
+        .unwrap();
+        assert!(!comparison.eligible);
+        assert_eq!(comparison.regressed_scenarios, ["plain_response"]);
+        assert_eq!(comparison.scenarios[0].control.passed, 2);
+        assert_eq!(comparison.scenarios[0].treatment.passed, 1);
+    }
+
+    fn report(run_id: &str, passed: bool, subject: SubjectArtifactV1) -> AgentQualityRunReportV1 {
+        let mut scenario = AgentQualityScenarioReportV1::new(
+            "plain_response",
+            format!("{run_id}-session"),
+            "prompt".into(),
+        );
+        scenario.wall_time_ms = 10;
+        scenario.passed = passed;
+        scenario.evaluation = passed.then(|| serde_json::json!({ "ok": true }));
+        AgentQualityRunReportV1::new(
+            run_id.into(),
+            subject,
+            AgentQualityLimitsV1::default(),
+            vec![scenario],
+        )
+    }
+}
diff --git a/harness/tests/agent-quality/src/completion.rs b/harness/tests/agent-quality/src/completion.rs
new file mode 100644
index 000000000..1be8f57a2
--- /dev/null
+++ b/harness/tests/agent-quality/src/completion.rs
@@ -0,0 +1,171 @@
+use std::collections::HashMap;
+use std::sync::{Arc, Mutex};
+use std::time::Duration;
+
+use iii_sdk::trigger::Trigger;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+use serde_json::Value;
+use tokio::sync::Notify;
+
+use crate::error::{EvalError, FailureClass, Phase};
+
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+pub struct CompletionEventV1 {
+    pub session_id: String,
+    pub turn_id: String,
+    pub status: String,
+    pub terminal: bool,
+    pub timestamp: i64,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub result: Option<Value>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub result_error: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub reason: Option<String>,
+}
+
+impl CompletionEventV1 {
+    fn successful(&self) -> bool {
+        self.terminal && self.status == "completed" && self.result_error.is_none()
+    }
+}
+
+#[derive(Clone, Default)]
+pub struct CompletionInbox {
+    events: Arc<Mutex<HashMap<(String, String), CompletionEventV1>>>,
+    notify: Arc<Notify>,
+}
+
+impl CompletionInbox {
+    pub fn push(&self, event: CompletionEventV1) {
+        if !event.terminal {
+            return;
+        }
+        self.events
+            .lock()
+            .unwrap_or_else(|poisoned| poisoned.into_inner())
+            .entry((event.session_id.clone(), event.turn_id.clone()))
+            .or_insert(event);
+        self.notify.notify_waiters();
+    }
+
+    pub async fn wait_terminal(
+        &self,
+        session_id: &str,
+        turn_id: &str,
+        timeout: Duration,
+    ) -> Result<CompletionEventV1, EvalError> {
+        let key = (session_id.to_string(), turn_id.to_string());
+        let deadline = tokio::time::Instant::now() + timeout;
+        loop {
+            let notified = self.notify.notified();
+            if let Some(event) = self
+                .events
+                .lock()
+                .unwrap_or_else(|poisoned| poisoned.into_inner())
+                .remove(&key)
+            {
+                return Ok(event);
+            }
+            if tokio::time::timeout_at(deadline, notified).await.is_err() {
+                return Err(EvalError::timeout(
+                    Phase::Await,
+                    format!(
+                        "timed out waiting for harness::turn-completed for session {session_id}, turn {turn_id}"
+                    ),
+                ));
+            }
+        }
+    }
+}
+
+pub fn validate_success(event: &CompletionEventV1) -> Result<(), EvalError> {
+    if event.successful() {
+        return Ok(());
+    }
+    Err(EvalError::new(
+        FailureClass::SubjectError,
+        Phase::Await,
+        None,
+        Some("harness::turn-completed".into()),
+        format!(
+            "subject turn ended with status {:?}: {}",
+            event.status,
+            event
+                .result_error
+                .as_deref()
+                .or(event.reason.as_deref())
+                .unwrap_or("no reason supplied")
+        ),
+    ))
+}
+
+pub struct CompletionBinding(Option<Trigger>);
+
+impl CompletionBinding {
+    pub fn new(trigger: Trigger) -> Self {
+        Self(Some(trigger))
+    }
+
+    pub fn unregister(mut self) {
+        if let Some(trigger) = self.0.take() {
+            trigger.unregister();
+        }
+    }
+}
+
+impl Drop for CompletionBinding {
+    fn drop(&mut self) {
+        if let Some(trigger) = self.0.take() {
+            trigger.unregister();
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn event(session: &str, turn: &str, status: &str, terminal: bool) -> CompletionEventV1 {
+        CompletionEventV1 {
+            session_id: session.into(),
+            turn_id: turn.into(),
+            status: status.into(),
+            terminal,
+            timestamp: 1,
+            result: None,
+            result_error: None,
+            reason: None,
+        }
+    }
+
+    #[tokio::test]
+    async fn receives_terminal_event_by_session_and_turn() {
+        let inbox = CompletionInbox::default();
+        inbox.push(event("s1", "t1", "completed", false));
+        inbox.push(event("s1", "t1", "completed", true));
+        let received = inbox
+            .wait_terminal("s1", "t1", Duration::from_millis(10))
+            .await
+            .unwrap();
+        assert!(received.successful());
+    }
+
+    #[tokio::test]
+    async fn missing_event_times_out() {
+        let error = CompletionInbox::default()
+            .wait_terminal("s1", "t1", Duration::from_millis(1))
+            .await
+            .unwrap_err();
+        assert_eq!(error.record.class, FailureClass::Timeout);
+    }
+
+    #[test]
+    fn failed_and_cancelled_events_are_subject_errors() {
+        for status in ["failed", "cancelled"] {
+            let error = validate_success(&event("s1", "t1", status, true)).unwrap_err();
+            assert_eq!(error.record.class, FailureClass::SubjectError);
+        }
+    }
+}
diff --git a/harness/tests/agent-quality/src/context.rs b/harness/tests/agent-quality/src/context.rs
new file mode 100644
index 000000000..317e8e433
--- /dev/null
+++ b/harness/tests/agent-quality/src/context.rs
@@ -0,0 +1,316 @@
+use std::sync::Arc;
+use std::time::Duration;
+
+use iii_sdk::errors::Error as SdkError;
+use iii_sdk::protocol::{RegisterTriggerInput, TriggerRequest};
+use iii_sdk::runtime::WorkerMetadata;
+use iii_sdk::{register_worker, IIIClient, InitOptions, RegisterFunction};
+use serde::de::DeserializeOwned;
+use serde::Serialize;
+use serde_json::{json, Value};
+
+use crate::completion::{validate_success, CompletionBinding, CompletionEventV1, CompletionInbox};
+use crate::error::{EvalError, Phase};
+use crate::subject::ResolvedAgentQualitySubjectV1;
+
+const TURN_COMPLETED: &str = "harness::turn-completed";
+
+pub struct ScenarioContext {
+    client: Arc<IIIClient>,
+    subject: Arc<ResolvedAgentQualitySubjectV1>,
+    inbox: CompletionInbox,
+    invocation_timeout: Duration,
+    completion_timeout: Duration,
+    callback_id: String,
+}
+
+impl ScenarioContext {
+    pub async fn connect(
+        url: &str,
+        run_id: &str,
+        subject: ResolvedAgentQualitySubjectV1,
+        invocation_timeout: Duration,
+        completion_timeout: Duration,
+    ) -> Result<Self, EvalError> {
+        let client = Arc::new(register_worker(
+            url,
+            InitOptions {
+                metadata: Some(WorkerMetadata {
+                    runtime: "rust".to_string(),
+                    version: env!("CARGO_PKG_VERSION").to_string(),
+                    name: "harness-agent-quality".to_string(),
+                    os: std::env::consts::OS.to_string(),
+                    pid: Some(std::process::id()),
+                    ..WorkerMetadata::default()
+                }),
+                ..InitOptions::default()
+            },
+        ));
+        let callback_id = format!("harness-agent-quality::{run_id}::on-turn-completed");
+        let inbox = CompletionInbox::default();
+        let callback_inbox = inbox.clone();
+        client.register_function(
+            &callback_id,
+            RegisterFunction::new_async(move |event: CompletionEventV1| {
+                let inbox = callback_inbox.clone();
+                async move {
+                    inbox.push(event);
+                    Ok::<Value, SdkError>(json!({ "accepted": true }))
+                }
+            })
+            .description("Agent-quality runner sink for terminal harness turn events")
+            .metadata(json!({ "internal": true })),
+        );
+
+        let context = Self {
+            client,
+            subject: Arc::new(subject),
+            inbox,
+            invocation_timeout,
+            completion_timeout,
+            callback_id,
+        };
+        context.wait_for_callback_registration().await?;
+        Ok(context)
+    }
+
+    pub fn subject(&self) -> &ResolvedAgentQualitySubjectV1 {
+        &self.subject
+    }
+
+    pub async fn trigger<I, O>(&self, function_id: &str, payload: I) -> Result<O, EvalError>
+    where
+        I: Serialize + Send,
+        O: DeserializeOwned,
+    {
+        self.trigger_phase(Phase::Assert, function_id, payload)
+            .await
+    }
+
+    pub(crate) async fn trigger_phase<I, O>(
+        &self,
+        phase: Phase,
+        function_id: &str,
+        payload: I,
+    ) -> Result<O, EvalError>
+    where
+        I: Serialize + Send,
+        O: DeserializeOwned,
+    {
+        let payload = serde_json::to_value(payload).map_err(|error| {
+            EvalError::serialization(phase, function_id, format!("serialize request: {error}"))
+        })?;
+        let request = TriggerRequest {
+            function_id: function_id.to_string(),
+            payload,
+            action: None,
+            timeout_ms: Some(self.invocation_timeout.as_millis().min(u64::MAX as u128) as u64),
+        };
+        let value =
+            match tokio::time::timeout(self.invocation_timeout, self.client.trigger(request)).await
+            {
+                Err(_) => {
+                    return Err(EvalError::timeout(
+                        phase,
+                        format!("local deadline invoking {function_id}"),
+                    ));
+                }
+                Ok(Err(error)) => return Err(EvalError::from_sdk(phase, function_id, error)),
+                Ok(Ok(value)) => value,
+            };
+        serde_json::from_value(value).map_err(|error| {
+            EvalError::serialization(phase, function_id, format!("decode response: {error}"))
+        })
+    }
+
+    pub async fn bind_completion(&self, session_id: &str) -> Result<CompletionBinding, EvalError> {
+        let config = json!({ "session_id": session_id });
+        let trigger = self
+            .client
+            .register_trigger(RegisterTriggerInput {
+                trigger_type: TURN_COMPLETED.to_string(),
+                function_id: self.callback_id.clone(),
+                config: config.clone(),
+                metadata: None,
+            })
+            .map_err(|error| {
+                EvalError::from_sdk(Phase::Setup, "engine::register_trigger", error)
+            })?;
+        let binding = CompletionBinding::new(trigger);
+        if let Err(error) = self.wait_for_completion_binding(&config).await {
+            binding.unregister();
+            return Err(error);
+        }
+        Ok(binding)
+    }
+
+    pub async fn await_completion(
+        &self,
+        session_id: &str,
+        turn_id: &str,
+    ) -> Result<CompletionEventV1, EvalError> {
+        let event = self
+            .inbox
+            .wait_terminal(session_id, turn_id, self.completion_timeout)
+            .await?;
+        validate_success(&event)?;
+        Ok(event)
+    }
+
+    pub async fn metrics(
+        &self,
+        root_session_id: &str,
+    ) -> Result<harness::functions::metrics::SessionMetricsResponseV1, EvalError> {
+        self.trigger_phase(
+            Phase::Collect,
+            "harness::metrics",
+            json!({ "root_session_id": root_session_id }),
+        )
+        .await
+    }
+
+    pub async fn transcript(&self, session_id: &str) -> Result<Value, EvalError> {
+        let mut cursor: Option<String> = None;
+        let mut messages = Vec::new();
+        loop {
+            let response: Value = self
+                .trigger_phase(
+                    Phase::Collect,
+                    "session::messages",
+                    json!({
+                        "session_id": session_id,
+                        "limit": 500,
+                        "cursor": cursor,
+                        "include_custom": true,
+                    }),
+                )
+                .await?;
+            let page = response
+                .get("messages")
+                .and_then(Value::as_array)
+                .ok_or_else(|| {
+                    EvalError::evidence(
+                        "session::messages",
+                        format!("malformed transcript page for {session_id}"),
+                    )
+                })?;
+            messages.extend(page.iter().cloned());
+            let next = response
+                .get("next_cursor")
+                .and_then(Value::as_str)
+                .map(str::to_string);
+            if next.is_none() {
+                break;
+            }
+            if next == cursor {
+                return Err(EvalError::evidence(
+                    "session::messages",
+                    format!("repeated transcript cursor for {session_id}"),
+                ));
+            }
+            cursor = next;
+        }
+        Ok(json!({ "messages": messages }))
+    }
+
+    pub async fn stop_session(
+        &self,
+        session_id: &str,
+        turn_id: Option<&str>,
+    ) -> Result<(), EvalError> {
+        let _: harness::functions::stop::StopResponse = self
+            .trigger_phase(
+                Phase::Cleanup,
+                "harness::stop",
+                json!({ "session_id": session_id, "turn_id": turn_id }),
+            )
+            .await?;
+        Ok(())
+    }
+
+    pub async fn shutdown(&self) {
+        self.client.shutdown_async().await;
+    }
+
+    async fn wait_for_callback_registration(&self) -> Result<(), EvalError> {
+        self.wait_until("runner function registration", || async {
+            let response = self
+                .raw_trigger(
+                    "engine::functions::list",
+                    json!({ "include_internal": true }),
+                )
+                .await
+                .ok()?;
+            response
+                .get("functions")?
+                .as_array()?
+                .iter()
+                .any(|item| {
+                    item.get("function_id").and_then(Value::as_str)
+                        == Some(self.callback_id.as_str())
+                })
+                .then_some(())
+        })
+        .await
+    }
+
+    async fn wait_for_completion_binding(&self, config: &Value) -> Result<(), EvalError> {
+        self.wait_until("completion trigger registration", || async {
+            let response = self
+                .raw_trigger(
+                    "engine::registered-triggers::list",
+                    json!({
+                        "include_internal": true,
+                        "function_id": self.callback_id,
+                        "trigger_type": TURN_COMPLETED,
+                    }),
+                )
+                .await
+                .ok()?;
+            response
+                .get("registered_triggers")?
+                .as_array()?
+                .iter()
+                .any(|item| {
+                    item.get("trigger_type").and_then(Value::as_str) == Some(TURN_COMPLETED)
+                        && item.get("function_id").and_then(Value::as_str)
+                            == Some(self.callback_id.as_str())
+                        && item.get("config") == Some(config)
+                })
+                .then_some(())
+        })
+        .await
+    }
+
+    async fn wait_until<F, Fut>(&self, label: &str, mut check: F) -> Result<(), EvalError>
+    where
+        F: FnMut() -> Fut,
+        Fut: std::future::Future<Output = Option<()>>,
+    {
+        let deadline = tokio::time::Instant::now() + Duration::from_secs(15);
+        loop {
+            if check().await.is_some() {
+                return Ok(());
+            }
+            if tokio::time::Instant::now() >= deadline {
+                return Err(EvalError::setup(format!("timed out waiting for {label}")));
+            }
+            tokio::time::sleep(Duration::from_millis(50)).await;
+        }
+    }
+
+    async fn raw_trigger(&self, function_id: &str, payload: Value) -> Result<Value, SdkError> {
+        let timeout_ms = self.invocation_timeout.as_millis().min(u64::MAX as u128) as u64;
+        tokio::time::timeout(
+            self.invocation_timeout,
+            self.client.trigger(TriggerRequest {
+                function_id: function_id.to_string(),
+                payload,
+                action: None,
+                timeout_ms: Some(timeout_ms),
+            }),
+        )
+        .await
+        .map_err(|_| SdkError::Timeout)?
+    }
+}
diff --git a/harness/tests/agent-quality/src/error.rs b/harness/tests/agent-quality/src/error.rs
new file mode 100644
index 000000000..6ccff51e9
--- /dev/null
+++ b/harness/tests/agent-quality/src/error.rs
@@ -0,0 +1,140 @@
+use iii_sdk::errors::Error as SdkError;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum FailureClass {
+    SetupError,
+    SubjectError,
+    AssertionFailure,
+    EvidenceError,
+    Timeout,
+    CleanupError,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum Phase {
+    Setup,
+    Send,
+    Await,
+    Collect,
+    Assert,
+    Cleanup,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct FailureRecord {
+    pub class: FailureClass,
+    pub phase: Phase,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub code: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub function_id: Option<String>,
+    pub message: String,
+}
+
+#[derive(Debug, thiserror::Error)]
+#[error("{message}")]
+pub struct EvalError {
+    pub record: FailureRecord,
+    message: String,
+}
+
+impl EvalError {
+    pub fn new(
+        class: FailureClass,
+        phase: Phase,
+        code: Option<String>,
+        function_id: Option<String>,
+        message: impl Into<String>,
+    ) -> Self {
+        let message = message.into();
+        Self {
+            record: FailureRecord {
+                class,
+                phase,
+                code,
+                function_id,
+                message: message.clone(),
+            },
+            message,
+        }
+    }
+
+    pub fn setup(message: impl Into<String>) -> Self {
+        Self::new(FailureClass::SetupError, Phase::Setup, None, None, message)
+    }
+
+    pub fn timeout(phase: Phase, message: impl Into<String>) -> Self {
+        Self::new(FailureClass::Timeout, phase, None, None, message)
+    }
+
+    pub fn assertion(message: impl Into<String>) -> Self {
+        Self::new(
+            FailureClass::AssertionFailure,
+            Phase::Assert,
+            None,
+            None,
+            message,
+        )
+    }
+
+    pub fn limit(message: impl Into<String>) -> Self {
+        Self::new(
+            FailureClass::AssertionFailure,
+            Phase::Assert,
+            Some("agent_quality/limit_exceeded".to_string()),
+            Some("harness::metrics".to_string()),
+            message,
+        )
+    }
+
+    pub fn evidence(function_id: &str, message: impl Into<String>) -> Self {
+        Self::new(
+            FailureClass::EvidenceError,
+            Phase::Collect,
+            None,
+            Some(function_id.to_string()),
+            message,
+        )
+    }
+
+    pub fn from_sdk(phase: Phase, function_id: &str, error: SdkError) -> Self {
+        let code = error.invocation_error().map(|value| value.code);
+        let class = if matches!(error, SdkError::Timeout) {
+            FailureClass::Timeout
+        } else {
+            class_for_phase(phase)
+        };
+        Self::new(
+            class,
+            phase,
+            code,
+            Some(function_id.to_string()),
+            error.to_string(),
+        )
+    }
+
+    pub fn serialization(phase: Phase, function_id: &str, message: impl Into<String>) -> Self {
+        Self::new(
+            class_for_phase(phase),
+            phase,
+            Some("SERDE".to_string()),
+            Some(function_id.to_string()),
+            message,
+        )
+    }
+}
+
+pub fn class_for_phase(phase: Phase) -> FailureClass {
+    match phase {
+        Phase::Setup => FailureClass::SetupError,
+        Phase::Send => FailureClass::SubjectError,
+        Phase::Await | Phase::Collect => FailureClass::EvidenceError,
+        Phase::Assert => FailureClass::AssertionFailure,
+        Phase::Cleanup => FailureClass::CleanupError,
+    }
+}
diff --git a/harness/tests/agent-quality/src/lib.rs b/harness/tests/agent-quality/src/lib.rs
new file mode 100644
index 000000000..527a8d1ff
--- /dev/null
+++ b/harness/tests/agent-quality/src/lib.rs
@@ -0,0 +1,19 @@
+pub mod comparison;
+pub mod completion;
+pub mod context;
+pub mod error;
+pub mod limits;
+pub mod report;
+pub mod scenarios;
+pub mod subject;
+pub mod suite;
+
+pub use comparison::{
+    compare_runs, comparison_dimension, ComparisonDimension, PromptComparisonReportV1,
+};
+pub use context::ScenarioContext;
+pub use error::{EvalError, FailureClass, FailureRecord, Phase};
+pub use limits::{AgentQualityLimitsV1, EvaluationLimitsV1, ExecutionLimitsV1, LimitsArgs};
+pub use report::{AgentQualityRunReportV1, AgentQualityScenarioReportV1, ScenarioObservationV1};
+pub use subject::{ResolvedAgentQualitySubjectV1, SubjectArtifactV1};
+pub use suite::{run_suite, SuiteRunConfig, SuiteRunOutcome};
diff --git a/harness/tests/agent-quality/src/limits.rs b/harness/tests/agent-quality/src/limits.rs
new file mode 100644
index 000000000..6aed77cb9
--- /dev/null
+++ b/harness/tests/agent-quality/src/limits.rs
@@ -0,0 +1,183 @@
+use clap::Args;
+use harness::functions::metrics::SessionMetricsResponseV1;
+
+use crate::error::EvalError;
+
+pub use eval::limits::{
+    EvalLimitsV1 as AgentQualityLimitsV1, EvaluationLimitsV1, ExecutionLimitsV1,
+    AGENT_QUALITY_DEFAULT_MAX_TOTAL_TOKENS, DEFAULT_INVOCATION_TIMEOUT_SECONDS,
+    DEFAULT_MAX_OUTPUT_TOKENS_PER_CALL, DEFAULT_MAX_TURNS, DEFAULT_SCENARIO_TIMEOUT_SECONDS,
+};
+
+#[derive(Debug, Clone, Args)]
+pub struct LimitsArgs {
+    /// Timeout for one iii function invocation.
+    #[arg(long, default_value_t = DEFAULT_INVOCATION_TIMEOUT_SECONDS)]
+    invocation_timeout_seconds: u64,
+
+    /// Hard wall-clock limit for one scenario; expiry calls harness::stop.
+    #[arg(long, default_value_t = DEFAULT_SCENARIO_TIMEOUT_SECONDS)]
+    scenario_timeout_seconds: u64,
+
+    /// Maximum model generation steps in a harness turn.
+    #[arg(long, default_value_t = DEFAULT_MAX_TURNS)]
+    max_turns: u32,
+
+    /// Maximum output tokens forwarded to each model call.
+    #[arg(long, default_value_t = DEFAULT_MAX_OUTPUT_TOKENS_PER_CALL)]
+    max_output_tokens_per_call: u64,
+
+    /// Hard input-plus-output budget over the complete session tree.
+    #[arg(long, default_value_t = AGENT_QUALITY_DEFAULT_MAX_TOTAL_TOKENS)]
+    max_total_tokens: u64,
+
+    /// Optional hard cost budget. Requires model catalog pricing.
+    #[arg(long)]
+    max_cost_usd: Option<f64>,
+
+    /// Post-execution ceiling for failed function calls.
+    #[arg(long, default_value_t = 0)]
+    max_function_call_errors: u64,
+
+    /// Optional post-execution ceiling for error spans. Requires in-memory traces.
+    #[arg(long)]
+    max_error_spans: Option<u64>,
+}
+
+impl LimitsArgs {
+    pub fn resolve(&self) -> Result<AgentQualityLimitsV1, EvalError> {
+        let limits = AgentQualityLimitsV1 {
+            execution: ExecutionLimitsV1 {
+                invocation_timeout_seconds: self.invocation_timeout_seconds,
+                scenario_timeout_seconds: self.scenario_timeout_seconds,
+                max_turns: self.max_turns,
+                max_output_tokens_per_call: self.max_output_tokens_per_call,
+                max_total_tokens: Some(self.max_total_tokens),
+                max_cost_usd: self.max_cost_usd,
+            },
+            evaluation: EvaluationLimitsV1 {
+                max_function_call_errors: self.max_function_call_errors,
+                max_error_spans: self.max_error_spans,
+            },
+        };
+        limits
+            .validate()
+            .map_err(|error| EvalError::setup(error.to_string()))?;
+        Ok(limits)
+    }
+}
+
+pub fn limit_failures(
+    limits: AgentQualityLimitsV1,
+    metrics: &SessionMetricsResponseV1,
+) -> Vec<EvalError> {
+    limits
+        .failures(metrics)
+        .into_iter()
+        .map(|failure| match failure.phase {
+            eval::error::EvalPhaseV1::Collect => EvalError::evidence(
+                failure.function_id.as_deref().unwrap_or("harness::metrics"),
+                failure.message,
+            ),
+            _ => EvalError::limit(failure.message),
+        })
+        .collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use harness::functions::metrics::{SessionTraceMetricsV1, SessionUsageTotalsV1};
+
+    use super::*;
+
+    fn metrics() -> SessionMetricsResponseV1 {
+        SessionMetricsResponseV1 {
+            root_session_id: "root".into(),
+            complete: true,
+            totals: SessionUsageTotalsV1 {
+                input_tokens: Some(70),
+                output_tokens: Some(20),
+                cost_usd: Some(0.25),
+                ..SessionUsageTotalsV1::default()
+            },
+            by_session: Vec::new(),
+            traces: Some(SessionTraceMetricsV1 {
+                trace_count: 1,
+                span_count: 5,
+                error_span_count: 0,
+                duration_ms: 10,
+                by_session: Vec::new(),
+            }),
+        }
+    }
+
+    #[test]
+    fn defaults_define_both_execution_and_evaluation_limits() {
+        let limits = AgentQualityLimitsV1::default();
+        assert_eq!(limits.execution.max_turns, 20);
+        assert_eq!(limits.execution.max_output_tokens_per_call, 8_192);
+        assert_eq!(limits.execution.max_total_tokens, None);
+        assert_eq!(limits.evaluation.max_function_call_errors, 0);
+    }
+
+    #[test]
+    fn evaluation_limits_accept_metrics_within_budget() {
+        assert!(AgentQualityLimitsV1::default()
+            .failures(&metrics())
+            .is_empty());
+    }
+
+    #[test]
+    fn evaluation_limits_report_each_exceeded_budget() {
+        let mut value = metrics();
+        value.totals.function_call_errors = 2;
+        value.traces.as_mut().unwrap().error_span_count = 3;
+        let limits = AgentQualityLimitsV1 {
+            execution: ExecutionLimitsV1 {
+                max_total_tokens: Some(80),
+                max_cost_usd: Some(0.2),
+                ..AgentQualityLimitsV1::default().execution
+            },
+            evaluation: EvaluationLimitsV1 {
+                max_function_call_errors: 1,
+                max_error_spans: Some(2),
+            },
+        };
+        let messages: Vec<_> = limit_failures(limits, &value)
+            .into_iter()
+            .map(|error| error.to_string())
+            .collect();
+        assert_eq!(messages.len(), 4);
+        assert!(messages.iter().any(|message| message.contains("tokens")));
+        assert!(messages.iter().any(|message| message.contains("cost")));
+        assert!(messages
+            .iter()
+            .any(|message| message.contains("function call errors")));
+        assert!(messages
+            .iter()
+            .any(|message| message.contains("error spans")));
+    }
+
+    #[test]
+    fn optional_limits_fail_closed_when_metrics_are_missing() {
+        let mut value = metrics();
+        value.totals.cost_usd = None;
+        value.traces = None;
+        let limits = AgentQualityLimitsV1 {
+            execution: ExecutionLimitsV1 {
+                max_total_tokens: Some(100),
+                max_cost_usd: Some(1.0),
+                ..AgentQualityLimitsV1::default().execution
+            },
+            evaluation: EvaluationLimitsV1 {
+                max_function_call_errors: 0,
+                max_error_spans: Some(0),
+            },
+        };
+        let failures = limit_failures(limits, &value);
+        assert_eq!(failures.len(), 2);
+        assert!(failures
+            .iter()
+            .all(|error| error.to_string().contains("requires")));
+    }
+}
diff --git a/harness/tests/agent-quality/src/report.rs b/harness/tests/agent-quality/src/report.rs
new file mode 100644
index 000000000..6161f601e
--- /dev/null
+++ b/harness/tests/agent-quality/src/report.rs
@@ -0,0 +1,162 @@
+use std::fs;
+use std::path::{Path, PathBuf};
+
+use harness::functions::metrics::SessionMetricsResponseV1;
+use serde::Serialize;
+use serde_json::Value;
+
+use crate::error::{EvalError, FailureRecord};
+use crate::limits::AgentQualityLimitsV1;
+use crate::subject::SubjectArtifactV1;
+
+pub use eval::report::EvalBenchmarkV1 as AgentQualityBenchmarkV1;
+
+#[derive(Debug, Serialize)]
+pub struct ScenarioObservationV1 {
+    pub metrics: SessionMetricsResponseV1,
+    pub transcript: Value,
+}
+
+#[derive(Debug, Serialize)]
+pub struct AgentQualityScenarioReportV1 {
+    pub scenario_id: &'static str,
+    pub prompt: String,
+    pub session_id: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub turn_id: Option<String>,
+    pub wall_time_ms: u64,
+    pub passed: bool,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub observation: Option<ScenarioObservationV1>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub evaluation: Option<Value>,
+    #[serde(skip_serializing_if = "Vec::is_empty")]
+    pub failures: Vec<FailureRecord>,
+}
+
+impl AgentQualityScenarioReportV1 {
+    pub fn new(scenario_id: &'static str, session_id: String, prompt: String) -> Self {
+        Self {
+            scenario_id,
+            prompt,
+            session_id,
+            turn_id: None,
+            wall_time_ms: 0,
+            passed: false,
+            observation: None,
+            evaluation: None,
+            failures: Vec::new(),
+        }
+    }
+
+    pub fn push_failure(&mut self, error: EvalError) {
+        self.failures.push(error.record);
+        self.passed = false;
+    }
+
+    pub fn finish(&mut self, wall_time_ms: u64) {
+        self.wall_time_ms = wall_time_ms;
+        self.passed = self.failures.is_empty() && self.evaluation.is_some();
+    }
+
+    pub fn benchmark(&self) -> Option<AgentQualityBenchmarkV1> {
+        let metrics = &self.observation.as_ref()?.metrics;
+        AgentQualityBenchmarkV1::from_metrics(metrics, self.wall_time_ms)
+    }
+}
+
+#[derive(Debug, Serialize)]
+pub struct AgentQualityRunReportV1 {
+    pub schema_version: &'static str,
+    pub run_id: String,
+    pub subject: SubjectArtifactV1,
+    pub limits: AgentQualityLimitsV1,
+    pub scenarios: Vec<AgentQualityScenarioReportV1>,
+    pub passed: bool,
+}
+
+impl AgentQualityRunReportV1 {
+    pub fn new(
+        run_id: String,
+        subject: SubjectArtifactV1,
+        limits: AgentQualityLimitsV1,
+        scenarios: Vec<AgentQualityScenarioReportV1>,
+    ) -> Self {
+        let passed = scenarios.iter().all(|scenario| scenario.passed);
+        Self {
+            schema_version: "1",
+            run_id,
+            subject,
+            limits,
+            scenarios,
+            passed,
+        }
+    }
+
+    pub fn write_to(&self, output_root: &Path) -> Result<PathBuf, EvalError> {
+        let run_dir = output_root.join(&self.run_id);
+        fs::create_dir_all(&run_dir).map_err(|error| {
+            EvalError::setup(format!(
+                "create report directory {}: {error}",
+                run_dir.display()
+            ))
+        })?;
+        let path = run_dir.join("results.json");
+        let mut bytes = serde_json::to_vec_pretty(self)
+            .map_err(|error| EvalError::setup(format!("serialize {}: {error}", path.display())))?;
+        bytes.push(b'\n');
+        fs::write(&path, bytes)
+            .map_err(|error| EvalError::setup(format!("write {}: {error}", path.display())))?;
+        Ok(run_dir)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use harness::prompt::SystemPromptStrategy;
+    use serde_json::json;
+    use tempfile::tempdir;
+
+    use super::*;
+
+    fn subject() -> SubjectArtifactV1 {
+        SubjectArtifactV1 {
+            schema_version: "1",
+            subject_id: "subject".into(),
+            subject_sha256: "a".repeat(64),
+            system_prompt_sha256: "b".repeat(64),
+            model: "model".into(),
+            provider: "provider".into(),
+            system_prompt_strategy: SystemPromptStrategy::Override,
+            thinking_level: None,
+            provider_options: None,
+        }
+    }
+
+    #[test]
+    fn writes_one_self_contained_run_report() {
+        let mut scenario =
+            AgentQualityScenarioReportV1::new("example", "session-1".into(), "prompt".into());
+        scenario.push_failure(EvalError::assertion("wrong result"));
+        scenario.finish(12);
+        let report = AgentQualityRunReportV1::new(
+            "run-1".into(),
+            subject(),
+            AgentQualityLimitsV1::default(),
+            vec![scenario],
+        );
+        let dir = tempdir().unwrap();
+        let run = report.write_to(dir.path()).unwrap();
+        let results: Value =
+            serde_json::from_slice(&fs::read(run.join("results.json")).unwrap()).unwrap();
+        assert_eq!(results["schema_version"], "1");
+        assert_eq!(results["passed"], false);
+        assert_eq!(results["limits"]["execution"]["max_turns"], 20);
+        assert_eq!(results["scenarios"][0]["prompt"], "prompt");
+        assert_eq!(
+            results["scenarios"][0]["failures"][0]["message"],
+            json!("wrong result")
+        );
+        assert_eq!(fs::read_dir(run).unwrap().count(), 1);
+    }
+}
diff --git a/harness/tests/agent-quality/src/scenarios/common.rs b/harness/tests/agent-quality/src/scenarios/common.rs
new file mode 100644
index 000000000..35bc372f9
--- /dev/null
+++ b/harness/tests/agent-quality/src/scenarios/common.rs
@@ -0,0 +1,231 @@
+use std::time::{Duration, Instant};
+
+use harness::functions::send::{MessageInput, SendOptions, SendRequest, SendResponse, SessionInit};
+use harness::types::turn::FunctionPolicy;
+use serde_json::{json, Value};
+use uuid::Uuid;
+
+use crate::context::ScenarioContext;
+use crate::error::{EvalError, FailureClass, Phase};
+use crate::limits::{limit_failures, AgentQualityLimitsV1};
+use crate::report::{AgentQualityScenarioReportV1, ScenarioObservationV1};
+
+use super::ScenarioSpec;
+
+pub async fn run(
+    context: &ScenarioContext,
+    run_id: &str,
+    limits: &AgentQualityLimitsV1,
+    spec: ScenarioSpec,
+) -> AgentQualityScenarioReportV1 {
+    let started = Instant::now();
+    let session_id = format!("aq_{}", Uuid::new_v4().simple());
+    let mut report =
+        AgentQualityScenarioReportV1::new(spec.id, session_id.clone(), spec.prompt.clone());
+    let binding = match context.bind_completion(&session_id).await {
+        Ok(binding) => binding,
+        Err(error) => {
+            report.push_failure(error);
+            report.finish(elapsed_ms(started));
+            return report;
+        }
+    };
+
+    let execution = async {
+        let response: SendResponse = context
+            .trigger_phase(
+                Phase::Send,
+                "harness::send",
+                send_request(context, run_id, &session_id, limits, &spec),
+            )
+            .await?;
+        validate_send_response(&response, &session_id)?;
+        report.turn_id = Some(response.turn_id.clone());
+        context
+            .await_completion(&session_id, &response.turn_id)
+            .await?;
+
+        let metrics = context.metrics(&session_id).await?;
+        validate_metrics(&metrics)?;
+        let transcript = context.transcript(&session_id).await?;
+        report.observation = Some(ScenarioObservationV1 {
+            metrics,
+            transcript,
+        });
+        let evaluation = (spec.evaluate)(
+            context,
+            report.observation.as_ref().unwrap(),
+            &spec.evaluation_context,
+        )
+        .await?;
+        report.evaluation = Some(evaluation);
+        Ok(())
+    };
+
+    if let Err(error) = bounded(limits.execution.scenario_timeout(), spec.id, execution).await {
+        report.push_failure(error);
+        if let Err(stop_error) = context
+            .stop_session(&session_id, report.turn_id.as_deref())
+            .await
+        {
+            report.push_failure(stop_error);
+        }
+    }
+    let limit_failures = report
+        .observation
+        .as_ref()
+        .map(|observation| limit_failures(*limits, &observation.metrics))
+        .unwrap_or_default();
+    for failure in limit_failures {
+        report.push_failure(failure);
+    }
+    binding.unregister();
+    report.finish(elapsed_ms(started));
+    report
+}
+
+fn send_request(
+    context: &ScenarioContext,
+    run_id: &str,
+    session_id: &str,
+    limits: &AgentQualityLimitsV1,
+    spec: &ScenarioSpec,
+) -> SendRequest {
+    let subject = context.subject();
+    SendRequest {
+        session_id: Some(session_id.to_string()),
+        message: MessageInput::Text(spec.prompt.clone()),
+        model: Some(subject.model.clone()),
+        provider: Some(subject.provider.clone()),
+        idempotency_key: Some(format!("agent-quality:{run_id}:{}:send", spec.id)),
+        session: Some(SessionInit {
+            title: Some(format!("Agent quality: {}", spec.id)),
+            metadata: Some(json!({
+                "agent_quality_run_id": run_id,
+                "agent_quality_scenario": spec.id,
+            })),
+        }),
+        options: Some(SendOptions {
+            system_prompt: Some(subject.system_prompt.clone()),
+            system_prompt_strategy: subject.system_prompt_strategy,
+            max_turns: Some(limits.execution.max_turns),
+            max_output_tokens: Some(limits.execution.max_output_tokens_per_call),
+            max_total_tokens: limits.execution.max_total_tokens,
+            max_cost_usd: limits.execution.max_cost_usd,
+            thinking_level: subject.thinking_level,
+            provider_options: subject.provider_options.clone(),
+            functions: Some(FunctionPolicy {
+                allow: vec!["*".into()],
+                ..FunctionPolicy::default()
+            }),
+            ..SendOptions::default()
+        }),
+    }
+}
+
+fn validate_send_response(
+    response: &SendResponse,
+    expected_session_id: &str,
+) -> Result<(), EvalError> {
+    if response.session_id != expected_session_id {
+        return Err(EvalError::new(
+            FailureClass::SubjectError,
+            Phase::Send,
+            None,
+            Some("harness::send".into()),
+            format!(
+                "harness::send returned session {} instead of {expected_session_id}",
+                response.session_id
+            ),
+        ));
+    }
+    if !response.accepted {
+        return Err(EvalError::new(
+            FailureClass::SubjectError,
+            Phase::Send,
+            None,
+            Some("harness::send".into()),
+            format!("harness::send did not accept the turn: {response:?}"),
+        ));
+    }
+    Ok(())
+}
+
+fn validate_metrics(
+    metrics: &harness::functions::metrics::SessionMetricsResponseV1,
+) -> Result<(), EvalError> {
+    if !metrics.complete {
+        return Err(EvalError::evidence(
+            "harness::metrics",
+            "session metrics are incomplete",
+        ));
+    }
+    for (name, value) in [
+        ("input_tokens", metrics.totals.input_tokens),
+        ("output_tokens", metrics.totals.output_tokens),
+    ] {
+        if value == Some(0) {
+            return Err(EvalError::evidence(
+                "harness::metrics",
+                format!("{name} was reported as zero"),
+            ));
+        }
+    }
+    Ok(())
+}
+
+pub fn assistant_texts(transcript: &Value) -> Vec<String> {
+    transcript
+        .get("messages")
+        .and_then(Value::as_array)
+        .into_iter()
+        .flatten()
+        .filter_map(|entry| entry.get("message"))
+        .filter(|message| message.get("role").and_then(Value::as_str) == Some("assistant"))
+        .map(message_text)
+        .collect()
+}
+
+pub fn state_value(response: Value) -> Value {
+    match response {
+        Value::Object(mut object)
+            if object.get("ok").and_then(Value::as_bool) == Some(true)
+                && object.contains_key("value") =>
+        {
+            object.remove("value").unwrap_or(Value::Null)
+        }
+        response => response,
+    }
+}
+
+async fn bounded<F>(timeout: Duration, scenario_id: &str, future: F) -> Result<(), EvalError>
+where
+    F: std::future::Future<Output = Result<(), EvalError>>,
+{
+    match tokio::time::timeout(timeout, future).await {
+        Ok(result) => result,
+        Err(_) => Err(EvalError::timeout(
+            Phase::Await,
+            format!("scenario {scenario_id} exceeded {}s", timeout.as_secs()),
+        )),
+    }
+}
+
+fn message_text(message: &Value) -> String {
+    message
+        .get("content")
+        .and_then(Value::as_array)
+        .into_iter()
+        .flatten()
+        .filter_map(|block| {
+            (block.get("type").and_then(Value::as_str) == Some("text"))
+                .then(|| block.get("text").and_then(Value::as_str))
+                .flatten()
+        })
+        .collect::<Vec<_>>()
+        .join("")
+}
+
+fn elapsed_ms(started: Instant) -> u64 {
+    started.elapsed().as_millis().min(u64::MAX as u128) as u64
+}
diff --git a/harness/tests/agent-quality/src/scenarios/mod.rs b/harness/tests/agent-quality/src/scenarios/mod.rs
new file mode 100644
index 000000000..19fc0b923
--- /dev/null
+++ b/harness/tests/agent-quality/src/scenarios/mod.rs
@@ -0,0 +1,107 @@
+use std::future::Future;
+use std::pin::Pin;
+
+use clap::ValueEnum;
+use serde_json::Value;
+
+use crate::context::ScenarioContext;
+use crate::error::EvalError;
+use crate::limits::AgentQualityLimitsV1;
+use crate::report::{AgentQualityScenarioReportV1, ScenarioObservationV1};
+
+pub mod common;
+pub mod plain_response;
+pub mod security_review;
+pub mod single_function;
+pub mod triggered_work;
+
+pub type EvaluationFuture<'a> = Pin<Box<dyn Future<Output = Result<Value, EvalError>> + Send + 'a>>;
+pub type ScenarioEvaluator =
+    for<'a> fn(&'a ScenarioContext, &'a ScenarioObservationV1, &'a Value) -> EvaluationFuture<'a>;
+
+pub struct ScenarioSpec {
+    pub id: &'static str,
+    pub prompt: String,
+    pub evaluation_context: Value,
+    pub evaluate: ScenarioEvaluator,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, ValueEnum)]
+pub enum ScenarioId {
+    #[value(name = "plain_response")]
+    PlainResponse,
+    #[value(name = "single_function")]
+    SingleFunction,
+    #[value(name = "security_review")]
+    SecurityReview,
+    #[value(name = "triggered_work")]
+    TriggeredWork,
+}
+
+impl ScenarioId {
+    pub const ALL: [Self; 4] = [
+        Self::PlainResponse,
+        Self::SingleFunction,
+        Self::SecurityReview,
+        Self::TriggeredWork,
+    ];
+
+    pub fn as_str(self) -> &'static str {
+        match self {
+            Self::PlainResponse => plain_response::ID,
+            Self::SingleFunction => single_function::ID,
+            Self::SecurityReview => security_review::ID,
+            Self::TriggeredWork => triggered_work::ID,
+        }
+    }
+
+    pub async fn run(
+        self,
+        context: &ScenarioContext,
+        run_id: &str,
+        limits: &AgentQualityLimitsV1,
+    ) -> AgentQualityScenarioReportV1 {
+        let spec = match self {
+            Self::PlainResponse => plain_response::scenario(run_id),
+            Self::SingleFunction => single_function::scenario(run_id),
+            Self::SecurityReview => security_review::scenario(run_id),
+            Self::TriggeredWork => triggered_work::scenario(run_id),
+        };
+        common::run(context, run_id, limits, spec).await
+    }
+}
+
+pub fn selected(requested: &[ScenarioId]) -> Vec<ScenarioId> {
+    if requested.is_empty() {
+        return ScenarioId::ALL.to_vec();
+    }
+    requested.iter().copied().fold(Vec::new(), |mut ids, id| {
+        if !ids.contains(&id) {
+            ids.push(id);
+        }
+        ids
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn defaults_to_the_prompt_evaluation_scenarios() {
+        assert_eq!(selected(&[]), ScenarioId::ALL);
+        assert_eq!(ScenarioId::ALL.len(), 4);
+    }
+
+    #[test]
+    fn explicit_selection_preserves_order_and_deduplicates() {
+        assert_eq!(
+            selected(&[
+                ScenarioId::TriggeredWork,
+                ScenarioId::PlainResponse,
+                ScenarioId::TriggeredWork,
+            ]),
+            vec![ScenarioId::TriggeredWork, ScenarioId::PlainResponse]
+        );
+    }
+}
diff --git a/harness/tests/agent-quality/src/scenarios/plain_response.rs b/harness/tests/agent-quality/src/scenarios/plain_response.rs
new file mode 100644
index 000000000..1c87bae0a
--- /dev/null
+++ b/harness/tests/agent-quality/src/scenarios/plain_response.rs
@@ -0,0 +1,44 @@
+use serde_json::{json, Value};
+
+use crate::context::ScenarioContext;
+use crate::error::EvalError;
+use crate::report::ScenarioObservationV1;
+
+use super::{common, EvaluationFuture, ScenarioSpec};
+
+pub const ID: &str = "plain_response";
+const EXPECTED: &str = "AGENT_QUALITY_PLAIN_OK";
+
+pub fn scenario(_run_id: &str) -> ScenarioSpec {
+    ScenarioSpec {
+        id: ID,
+        prompt: format!(
+            "Reply with exactly `{EXPECTED}` and nothing else. Do not call any function."
+        ),
+        evaluation_context: Value::Null,
+        evaluate,
+    }
+}
+
+fn evaluate<'a>(
+    _context: &'a ScenarioContext,
+    observation: &'a ScenarioObservationV1,
+    _evaluation_context: &'a Value,
+) -> EvaluationFuture<'a> {
+    Box::pin(async move {
+        let metrics = &observation.metrics;
+        if metrics.totals.function_calls != 0 || metrics.totals.turns != 1 {
+            return Err(EvalError::assertion(format!(
+                "expected one assistant turn and no function calls, observed {} turns and {} calls",
+                metrics.totals.turns, metrics.totals.function_calls
+            )));
+        }
+        let texts = common::assistant_texts(&observation.transcript);
+        if texts != [EXPECTED] {
+            return Err(EvalError::assertion(format!(
+                "expected one assistant entry containing {EXPECTED:?}, observed {texts:?}"
+            )));
+        }
+        Ok(json!({ "expected": EXPECTED, "actual": texts[0] }))
+    })
+}
diff --git a/harness/tests/agent-quality/src/scenarios/security_review.rs b/harness/tests/agent-quality/src/scenarios/security_review.rs
new file mode 100644
index 000000000..3df0da4d5
--- /dev/null
+++ b/harness/tests/agent-quality/src/scenarios/security_review.rs
@@ -0,0 +1,67 @@
+use serde_json::{json, Value};
+
+use crate::context::ScenarioContext;
+use crate::error::EvalError;
+use crate::report::ScenarioObservationV1;
+
+use super::{common, EvaluationFuture, ScenarioSpec};
+
+pub const ID: &str = "security_review";
+
+pub fn scenario(_run_id: &str) -> ScenarioSpec {
+    let expected = json!({
+        "findings": [
+            { "id": 1, "issue": "hardcoded_credential" },
+            { "id": 2, "issue": "sql_injection" },
+            { "id": 3, "issue": "open_redirect" }
+        ]
+    });
+    ScenarioSpec {
+        id: ID,
+        prompt: r#"Review these three independent snippets:
+1. `const API_KEY = "sk-live-example";`
+2. `db.execute("SELECT * FROM users WHERE id = " + request.id);`
+3. `response.redirect(request.query.next);`
+
+Return only valid JSON with this shape:
+{"findings":[{"id":1,"issue":"..."},{"id":2,"issue":"..."},{"id":3,"issue":"..."}]}
+
+Use exactly these issue names when applicable: hardcoded_credential, sql_injection, open_redirect. Do not call a function."#
+            .into(),
+        evaluation_context: expected,
+        evaluate,
+    }
+}
+
+fn evaluate<'a>(
+    _context: &'a ScenarioContext,
+    observation: &'a ScenarioObservationV1,
+    expected: &'a Value,
+) -> EvaluationFuture<'a> {
+    Box::pin(async move {
+        let metrics = &observation.metrics;
+        if metrics.totals.function_calls != 0 || metrics.totals.turns != 1 {
+            return Err(EvalError::assertion(format!(
+                "expected one reasoning turn and no function calls, observed {} turns and {} calls",
+                metrics.totals.turns, metrics.totals.function_calls
+            )));
+        }
+        let texts = common::assistant_texts(&observation.transcript);
+        if texts.len() != 1 {
+            return Err(EvalError::assertion(format!(
+                "expected one assistant response, observed {texts:?}"
+            )));
+        }
+        let actual: Value = serde_json::from_str(texts[0].trim()).map_err(|error| {
+            EvalError::assertion(format!(
+                "security review did not return valid JSON: {error}"
+            ))
+        })?;
+        if actual != *expected {
+            return Err(EvalError::assertion(format!(
+                "security findings mismatch: expected {expected}, observed {actual}"
+            )));
+        }
+        Ok(json!({ "expected": expected, "actual": actual }))
+    })
+}
diff --git a/harness/tests/agent-quality/src/scenarios/single_function.rs b/harness/tests/agent-quality/src/scenarios/single_function.rs
new file mode 100644
index 000000000..5c8e48385
--- /dev/null
+++ b/harness/tests/agent-quality/src/scenarios/single_function.rs
@@ -0,0 +1,63 @@
+use serde_json::{json, Value};
+
+use crate::context::ScenarioContext;
+use crate::error::EvalError;
+use crate::report::ScenarioObservationV1;
+
+use super::{common, EvaluationFuture, ScenarioSpec};
+
+pub const ID: &str = "single_function";
+
+pub fn scenario(run_id: &str) -> ScenarioSpec {
+    let scope = format!("agent-quality:{run_id}");
+    let key = ID;
+    let expected = json!({ "run_id": run_id, "scenario": ID, "status": "written" });
+    let arguments = json!({ "scope": scope, "key": key, "value": expected });
+    ScenarioSpec {
+        id: ID,
+        prompt: format!(
+            "Call state::set exactly once with these exact arguments: {}. After the function succeeds, finish the task.",
+            serde_json::to_string(&arguments).unwrap()
+        ),
+        evaluation_context: arguments,
+        evaluate,
+    }
+}
+
+fn evaluate<'a>(
+    context: &'a ScenarioContext,
+    observation: &'a ScenarioObservationV1,
+    evaluation_context: &'a Value,
+) -> EvaluationFuture<'a> {
+    Box::pin(async move {
+        let scope = evaluation_context["scope"]
+            .as_str()
+            .ok_or_else(|| EvalError::assertion("missing evaluation scope"))?;
+        let key = evaluation_context["key"]
+            .as_str()
+            .ok_or_else(|| EvalError::assertion("missing evaluation key"))?;
+        let expected = &evaluation_context["value"];
+        let observed = common::state_value(
+            context
+                .trigger("state::get", json!({ "scope": scope, "key": key }))
+                .await?,
+        );
+        context
+            .trigger::<_, Value>("state::delete", json!({ "scope": scope, "key": key }))
+            .await?;
+
+        if observed != *expected {
+            return Err(EvalError::assertion(format!(
+                "state value mismatch: expected {expected}, observed {observed}"
+            )));
+        }
+        let metrics = &observation.metrics;
+        if metrics.totals.function_calls == 0 || metrics.totals.function_call_errors != 0 {
+            return Err(EvalError::assertion(format!(
+                "expected successful function execution, observed {} calls and {} errors",
+                metrics.totals.function_calls, metrics.totals.function_call_errors
+            )));
+        }
+        Ok(json!({ "expected": expected, "actual": observed }))
+    })
+}
diff --git a/harness/tests/agent-quality/src/scenarios/triggered_work.rs b/harness/tests/agent-quality/src/scenarios/triggered_work.rs
new file mode 100644
index 000000000..717b09287
--- /dev/null
+++ b/harness/tests/agent-quality/src/scenarios/triggered_work.rs
@@ -0,0 +1,116 @@
+use std::time::Duration;
+
+use serde_json::{json, Value};
+
+use crate::context::ScenarioContext;
+use crate::error::EvalError;
+use crate::report::ScenarioObservationV1;
+
+use super::{common, EvaluationFuture, ScenarioSpec};
+
+pub const ID: &str = "triggered_work";
+
+pub fn scenario(run_id: &str) -> ScenarioSpec {
+    let scope = format!("agent-quality:{run_id}");
+    let signal_key = format!("{ID}:signal");
+    let result_key = format!("{ID}:result");
+    let expected = json!({ "status": "triggered", "run_id": run_id });
+    let registration = json!({
+        "trigger_type": "state",
+        "config": { "scope": scope, "key": signal_key },
+        "once": true,
+        "function_id": "harness::react",
+        "metadata": {
+            "call": {
+                "function_id": "state::set",
+                "payload": {
+                    "scope": scope,
+                    "key": result_key,
+                    "value": expected
+                }
+            }
+        }
+    });
+    let signal = json!({
+        "scope": scope,
+        "key": signal_key,
+        "value": { "status": "fire" }
+    });
+    ScenarioSpec {
+        id: ID,
+        prompt: format!(
+            "First call engine::register_trigger exactly once with these arguments: {}. After it succeeds, call state::set exactly once with these arguments: {}. Do not set the result key directly. Finish after the signal succeeds.",
+            serde_json::to_string(&registration).unwrap(),
+            serde_json::to_string(&signal).unwrap()
+        ),
+        evaluation_context: json!({
+            "scope": scope,
+            "signal_key": signal_key,
+            "result_key": result_key,
+            "expected": expected,
+        }),
+        evaluate,
+    }
+}
+
+fn evaluate<'a>(
+    context: &'a ScenarioContext,
+    observation: &'a ScenarioObservationV1,
+    evaluation_context: &'a Value,
+) -> EvaluationFuture<'a> {
+    Box::pin(async move {
+        let scope = string_field(evaluation_context, "scope")?;
+        let signal_key = string_field(evaluation_context, "signal_key")?;
+        let result_key = string_field(evaluation_context, "result_key")?;
+        let expected = &evaluation_context["expected"];
+        let observed = wait_for_state(context, scope, result_key).await?;
+
+        for key in [signal_key, result_key] {
+            context
+                .trigger::<_, Value>("state::delete", json!({ "scope": scope, "key": key }))
+                .await?;
+        }
+        if observed != *expected {
+            return Err(EvalError::assertion(format!(
+                "triggered result mismatch: expected {expected}, observed {observed}"
+            )));
+        }
+        if observation.metrics.totals.function_call_errors != 0 {
+            return Err(EvalError::assertion(format!(
+                "expected no function-call errors, observed {}",
+                observation.metrics.totals.function_call_errors
+            )));
+        }
+        Ok(json!({ "expected": expected, "actual": observed }))
+    })
+}
+
+async fn wait_for_state(
+    context: &ScenarioContext,
+    scope: &str,
+    key: &str,
+) -> Result<Value, EvalError> {
+    let deadline = tokio::time::Instant::now() + Duration::from_secs(10);
+    loop {
+        let value = common::state_value(
+            context
+                .trigger("state::get", json!({ "scope": scope, "key": key }))
+                .await?,
+        );
+        if !value.is_null() {
+            return Ok(value);
+        }
+        if tokio::time::Instant::now() >= deadline {
+            return Err(EvalError::assertion(
+                "triggered state result did not arrive",
+            ));
+        }
+        tokio::time::sleep(Duration::from_millis(50)).await;
+    }
+}
+
+fn string_field<'a>(value: &'a Value, field: &str) -> Result<&'a str, EvalError> {
+    value[field]
+        .as_str()
+        .ok_or_else(|| EvalError::assertion(format!("missing evaluation field {field}")))
+}
diff --git a/harness/tests/agent-quality/src/subject.rs b/harness/tests/agent-quality/src/subject.rs
new file mode 100644
index 000000000..fb5cb5006
--- /dev/null
+++ b/harness/tests/agent-quality/src/subject.rs
@@ -0,0 +1,182 @@
+use std::collections::BTreeMap;
+use std::path::{Path, PathBuf};
+
+use harness::prompt::SystemPromptStrategy;
+use harness::types::model::ThinkingLevel;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+use serde_json::Value;
+use sha2::{Digest, Sha256};
+
+use crate::error::EvalError;
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct AgentQualitySubjectV1 {
+    pub schema_version: String,
+    pub subject_id: String,
+    pub model: String,
+    pub provider: String,
+    pub system_prompt_path: PathBuf,
+    pub system_prompt_strategy: SystemPromptStrategy,
+    #[serde(default)]
+    pub thinking_level: Option<ThinkingLevel>,
+    #[serde(default)]
+    pub provider_options: Option<BTreeMap<String, Value>>,
+}
+
+#[derive(Debug, Clone)]
+pub struct ResolvedAgentQualitySubjectV1 {
+    pub subject_id: String,
+    pub subject_sha256: String,
+    pub system_prompt_sha256: String,
+    pub model: String,
+    pub provider: String,
+    pub system_prompt: String,
+    pub system_prompt_strategy: SystemPromptStrategy,
+    pub thinking_level: Option<ThinkingLevel>,
+    pub provider_options: Option<BTreeMap<String, Value>>,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct SubjectArtifactV1 {
+    pub schema_version: &'static str,
+    pub subject_id: String,
+    pub subject_sha256: String,
+    pub system_prompt_sha256: String,
+    pub model: String,
+    pub provider: String,
+    pub system_prompt_strategy: SystemPromptStrategy,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub thinking_level: Option<ThinkingLevel>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub provider_options: Option<BTreeMap<String, Value>>,
+}
+
+impl ResolvedAgentQualitySubjectV1 {
+    pub fn artifact(&self) -> SubjectArtifactV1 {
+        SubjectArtifactV1 {
+            schema_version: "1",
+            subject_id: self.subject_id.clone(),
+            subject_sha256: self.subject_sha256.clone(),
+            system_prompt_sha256: self.system_prompt_sha256.clone(),
+            model: self.model.clone(),
+            provider: self.provider.clone(),
+            system_prompt_strategy: self.system_prompt_strategy,
+            thinking_level: self.thinking_level,
+            provider_options: self.provider_options.clone(),
+        }
+    }
+}
+
+pub fn load(path: &Path) -> Result<ResolvedAgentQualitySubjectV1, EvalError> {
+    let subject_bytes = std::fs::read(path)
+        .map_err(|error| EvalError::setup(format!("read subject {}: {error}", path.display())))?;
+    let subject: AgentQualitySubjectV1 = serde_json::from_slice(&subject_bytes)
+        .map_err(|error| EvalError::setup(format!("parse subject {}: {error}", path.display())))?;
+    if subject.schema_version != "1" {
+        return Err(EvalError::setup(format!(
+            "unsupported subject schema_version {:?}; expected \"1\"",
+            subject.schema_version
+        )));
+    }
+    for (name, value) in [
+        ("subject_id", subject.subject_id.as_str()),
+        ("model", subject.model.as_str()),
+        ("provider", subject.provider.as_str()),
+    ] {
+        if value.trim().is_empty() {
+            return Err(EvalError::setup(format!("subject {name} cannot be empty")));
+        }
+    }
+
+    let base = path.parent().unwrap_or_else(|| Path::new("."));
+    let prompt_path = if subject.system_prompt_path.is_absolute() {
+        subject.system_prompt_path.clone()
+    } else {
+        base.join(&subject.system_prompt_path)
+    };
+    let prompt_bytes = std::fs::read(&prompt_path).map_err(|error| {
+        EvalError::setup(format!(
+            "read system prompt {}: {error}",
+            prompt_path.display()
+        ))
+    })?;
+    let system_prompt = String::from_utf8(prompt_bytes.clone()).map_err(|error| {
+        EvalError::setup(format!(
+            "system prompt {} is not UTF-8: {error}",
+            prompt_path.display()
+        ))
+    })?;
+
+    Ok(ResolvedAgentQualitySubjectV1 {
+        subject_id: subject.subject_id,
+        subject_sha256: sha256(&subject_bytes),
+        system_prompt_sha256: sha256(&prompt_bytes),
+        model: subject.model,
+        provider: subject.provider,
+        system_prompt,
+        system_prompt_strategy: subject.system_prompt_strategy,
+        thinking_level: subject.thinking_level,
+        provider_options: subject.provider_options,
+    })
+}
+
+fn sha256(bytes: &[u8]) -> String {
+    format!("{:x}", Sha256::digest(bytes))
+}
+
+#[cfg(test)]
+mod tests {
+    use std::fs;
+
+    use tempfile::tempdir;
+
+    use super::*;
+
+    #[test]
+    fn resolves_prompt_relative_to_subject_and_hashes_exact_bytes() {
+        let dir = tempdir().unwrap();
+        fs::write(dir.path().join("prompt.md"), b"exact prompt\n").unwrap();
+        fs::write(
+            dir.path().join("subject.json"),
+            br#"{
+              "schema_version":"1",
+              "subject_id":"baseline",
+              "model":"model-v1",
+              "provider":"provider",
+              "system_prompt_path":"prompt.md",
+              "system_prompt_strategy":"override"
+            }"#,
+        )
+        .unwrap();
+
+        let resolved = load(&dir.path().join("subject.json")).unwrap();
+        assert_eq!(resolved.system_prompt, "exact prompt\n");
+        assert_eq!(resolved.system_prompt_sha256, sha256(b"exact prompt\n"));
+    }
+
+    #[test]
+    fn rejects_execution_limits_and_unknown_subject_fields() {
+        let dir = tempdir().unwrap();
+        fs::write(dir.path().join("prompt.md"), "prompt").unwrap();
+        fs::write(
+            dir.path().join("subject.json"),
+            r#"{
+              "schema_version":"1",
+              "subject_id":"baseline",
+              "model":"model-v1",
+              "provider":"provider",
+              "system_prompt_path":"prompt.md",
+              "system_prompt_strategy":"override",
+              "max_output_tokens":42,
+              "unexpected":true
+            }"#,
+        )
+        .unwrap();
+
+        let error = load(&dir.path().join("subject.json")).unwrap_err();
+        assert!(error.to_string().contains("unknown field"));
+    }
+}
diff --git a/harness/tests/agent-quality/src/suite.rs b/harness/tests/agent-quality/src/suite.rs
new file mode 100644
index 000000000..f5675eadf
--- /dev/null
+++ b/harness/tests/agent-quality/src/suite.rs
@@ -0,0 +1,57 @@
+use std::path::PathBuf;
+
+use uuid::Uuid;
+
+use crate::context::ScenarioContext;
+use crate::error::EvalError;
+use crate::limits::AgentQualityLimitsV1;
+use crate::report::AgentQualityRunReportV1;
+use crate::scenarios::ScenarioId;
+use crate::subject::ResolvedAgentQualitySubjectV1;
+
+pub struct SuiteRunConfig {
+    pub url: String,
+    pub subject: ResolvedAgentQualitySubjectV1,
+    pub output: PathBuf,
+    pub scenarios: Vec<ScenarioId>,
+    pub limits: AgentQualityLimitsV1,
+}
+
+pub struct SuiteRunOutcome {
+    pub report: AgentQualityRunReportV1,
+    pub run_dir: PathBuf,
+}
+
+pub async fn run_suite(config: SuiteRunConfig) -> Result<SuiteRunOutcome, EvalError> {
+    config
+        .limits
+        .validate()
+        .map_err(|error| EvalError::setup(error.to_string()))?;
+    let run_id = Uuid::new_v4().simple().to_string();
+    let subject_artifact = config.subject.artifact();
+    let context = ScenarioContext::connect(
+        &config.url,
+        &run_id,
+        config.subject,
+        config.limits.execution.invocation_timeout(),
+        config.limits.execution.scenario_timeout(),
+    )
+    .await?;
+
+    let mut scenario_reports = Vec::with_capacity(config.scenarios.len());
+    for scenario_id in config.scenarios {
+        tracing::info!(
+            scenario = scenario_id.as_str(),
+            subject = %subject_artifact.subject_id,
+            "running agent-quality scenario"
+        );
+        scenario_reports.push(scenario_id.run(&context, &run_id, &config.limits).await);
+    }
+
+    let report =
+        AgentQualityRunReportV1::new(run_id, subject_artifact, config.limits, scenario_reports);
+    let write_result = report.write_to(&config.output);
+    context.shutdown().await;
+    let run_dir = write_result?;
+    Ok(SuiteRunOutcome { report, run_dir })
+}
diff --git a/harness/tests/agent-quality/subjects/evaluation.md b/harness/tests/agent-quality/subjects/evaluation.md
new file mode 100644
index 000000000..d4f5f6d2e
--- /dev/null
+++ b/harness/tests/agent-quality/subjects/evaluation.md
@@ -0,0 +1,6 @@
+You are running a deterministic harness evaluation.
+
+Follow the scenario instructions exactly. Use only the functions exposed for the
+current turn, preserve every requested literal value, and do not perform extra
+work. Treat function results as authoritative. When the requested durable state
+has been produced, return the exact completion text requested by the scenario.
diff --git a/harness/tests/agent-quality/subjects/luna.json b/harness/tests/agent-quality/subjects/luna.json
new file mode 100644
index 000000000..251b675a2
--- /dev/null
+++ b/harness/tests/agent-quality/subjects/luna.json
@@ -0,0 +1,9 @@
+{
+  "schema_version": "1",
+  "subject_id": "codex-luna-baseline",
+  "model": "codex/gpt-5.6-luna",
+  "provider": "openai-codex",
+  "system_prompt_path": "evaluation.md",
+  "system_prompt_strategy": "enrich",
+  "thinking_level": "low"
+}
diff --git a/harness/tests/agent-quality/subjects/sol.json b/harness/tests/agent-quality/subjects/sol.json
new file mode 100644
index 000000000..7bfb0cb1b
--- /dev/null
+++ b/harness/tests/agent-quality/subjects/sol.json
@@ -0,0 +1,9 @@
+{
+  "schema_version": "1",
+  "subject_id": "codex-sol-baseline",
+  "model": "codex/gpt-5.6-sol",
+  "provider": "openai-codex",
+  "system_prompt_path": "evaluation.md",
+  "system_prompt_strategy": "enrich",
+  "thinking_level": "low"
+}
diff --git a/harness/tests/e2e/src/scenarios/coalesced_fire.rs b/harness/tests/e2e/src/scenarios/coalesced_fire.rs
index a59cfaf7e..a8dab8b84 100644
--- a/harness/tests/e2e/src/scenarios/coalesced_fire.rs
+++ b/harness/tests/e2e/src/scenarios/coalesced_fire.rs
@@ -84,7 +84,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                     .turn_request()
                     .system_prompt_sha256("{{system_prompt_sha256}}")
                     .messages_exact([Message::user(MESSAGE)])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER], [record.tool()]),
             )
             .respond(Response::function_call_raw(
                 "call-arm",
@@ -108,7 +108,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                         json!({ "role": "function_result", "function_call_id": "call-arm",
                                 "is_error": false }),
                     ])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER], [record.tool()]),
             )
             .respond(Response::text("armed", 10, 2)),
     );
diff --git a/harness/tests/e2e/src/scenarios/dsl.rs b/harness/tests/e2e/src/scenarios/dsl.rs
index 32d9bc673..b49c132dc 100644
--- a/harness/tests/e2e/src/scenarios/dsl.rs
+++ b/harness/tests/e2e/src/scenarios/dsl.rs
@@ -424,6 +424,39 @@ impl Request {
         self
     }
 
+    /// Native exposure publishes the harness's intercepted subscription
+    /// controls (`subscribe::native_control_tools`) AHEAD of the registry
+    /// functions this turn's policy allows. Their descriptions and schemas
+    /// live in the harness crate, which fixtures cannot import, so the leading
+    /// control tools are pinned by name while the registry tools that follow
+    /// stay byte-exact. Still an exact match on the list as a whole: an
+    /// unexpected extra tool keeps its prose and fails the comparison.
+    pub(super) fn tools_exact_after_controls(
+        mut self,
+        controls: impl IntoIterator<Item = &'static str>,
+        tools: impl IntoIterator<Item = Value>,
+    ) -> Self {
+        let controls: Vec<&str> = controls.into_iter().collect();
+        let normalize = (0..controls.len())
+            .flat_map(|index| {
+                ["description", "parameters"].map(|field| JsonNormalizerV1 {
+                    pointer: format!("/{index}/{field}"),
+                    operation: NormalizerOperation::Delete,
+                })
+            })
+            .collect::<Vec<_>>();
+        let expected = controls
+            .iter()
+            .map(|name| json!({ "name": name, "execution_mode": "sequential" }))
+            .chain(tools)
+            .collect();
+        self.tools = Some(JsonMatcherV1::Exact {
+            expected: Value::Array(expected),
+            normalize: Some(normalize),
+        });
+        self
+    }
+
     pub(super) fn tools_subset(mut self, tools: impl IntoIterator<Item = Value>) -> Self {
         self.tools = Some(JsonMatcherV1::Subset {
             expected: Value::Array(tools.into_iter().collect()),
diff --git a/harness/tests/e2e/src/scenarios/join_spec_mismatch.rs b/harness/tests/e2e/src/scenarios/join_spec_mismatch.rs
index d0c451787..0d3088de2 100644
--- a/harness/tests/e2e/src/scenarios/join_spec_mismatch.rs
+++ b/harness/tests/e2e/src/scenarios/join_spec_mismatch.rs
@@ -62,7 +62,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                     .turn_request()
                     .system_prompt_sha256("{{system_prompt_sha256}}")
                     .messages_exact([Message::user(MESSAGE)])
-                    .without_tools(),
+                    .tools_exact_after_controls([REGISTER], []),
             )
             .respond(Response::function_call_raw(
                 "call-w1",
@@ -92,7 +92,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                         json!({ "role": "function_result", "function_call_id": "call-w1",
                                 "is_error": false }),
                     ])
-                    .without_tools(),
+                    .tools_exact_after_controls([REGISTER], []),
             )
             .respond(Response::function_call_raw(
                 "call-w2",
@@ -121,7 +121,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                         json!({ "role": "function_result", "function_call_id": "call-w2",
                                 "is_error": true }),
                     ])
-                    .without_tools(),
+                    .tools_exact_after_controls([REGISTER], []),
             )
             .respond(Response::text("join mismatch rejected", 12, 3)),
     )
diff --git a/harness/tests/e2e/src/scenarios/late_join_replay.rs b/harness/tests/e2e/src/scenarios/late_join_replay.rs
index df4f2a918..c9469bac1 100644
--- a/harness/tests/e2e/src/scenarios/late_join_replay.rs
+++ b/harness/tests/e2e/src/scenarios/late_join_replay.rs
@@ -121,7 +121,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                     .turn_request_step(0)
                     .system_prompt_sha256("{{system_prompt_sha256}}")
                     .messages_exact([Message::user(MESSAGE)])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER], [record.tool()]),
             )
             .respond(Response::function_call_raw(
                 "call-spawn",
@@ -143,7 +143,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                         json!({ "role": "function_result", "function_call_id": "call-spawn",
                                 "is_error": false }),
                     ])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER], [record.tool()]),
             )
             .respond(Response::function_call_raw(
                 "call-witness",
@@ -167,7 +167,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                         json!({ "role": "function_result", "function_call_id": "call-witness",
                                 "is_error": false }),
                     ])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER], [record.tool()]),
             )
             .respond(Response::text("armed", 10, 2)),
     )
@@ -179,7 +179,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                     .turn_request_step(0)
                     .system_prompt_regex(".")
                     .messages_subset([json!({ "role": "user" })])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER], [record.tool()]),
             )
             .respond(Response::text("work done", 8, 2)),
     )
@@ -201,7 +201,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                             { "type": "text", "text": MESSAGE_2 }
                         ] }),
                     ])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER], [record.tool()]),
             )
             .respond(Response::function_call_raw(
                 "call-late-join",
@@ -229,7 +229,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                         json!({ "role": "function_result", "function_call_id": "call-late-join",
                                 "is_error": false }),
                     ])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER], [record.tool()]),
             )
             .respond(Response::text("late join armed", 10, 2)),
     )
diff --git a/harness/tests/e2e/src/scenarios/reaction_policy_inheritance.rs b/harness/tests/e2e/src/scenarios/reaction_policy_inheritance.rs
index 9e8e66f9c..a6ca504e7 100644
--- a/harness/tests/e2e/src/scenarios/reaction_policy_inheritance.rs
+++ b/harness/tests/e2e/src/scenarios/reaction_policy_inheritance.rs
@@ -2,13 +2,13 @@
 //! turn's dispatch policy.
 //!
 //! The rctest-k7m3 run-3 failure shape: an agent registers a wrap-up reaction
-//! (no `options.functions`) delivered back into its own session; the reaction
-//! turn used to spawn parentless onto the read-only baseline and was denied
-//! the very functions every earlier turn in that chat could call. With the
-//! registrant-policy stamp the reaction turn inherits the registering turn's
-//! policy, so its scripted call to the controlled function must now dispatch
+//! (no `options.functions`) explicitly delivered back into its own session;
+//! the reaction used to fall back to the read-only baseline and was denied the
+//! functions every earlier turn in that chat could call. With the
+//! registrant-policy stamp the reaction inherits the registering turn's
+//! policy, so its scripted call to the controlled function must dispatch
 //! successfully — before the fix that call errors with "not permitted by this
-//! agent's dispatch policy" and the second terminal turn's history pins fail.
+//! agent's dispatch policy".
 //!
 //! The coalescing half of the same PR is pinned at the unit level
 //! (`fire_gate_coalesces_capped_fires_newest_wins`): its trailing fire lands
@@ -55,8 +55,8 @@ pub(super) fn scenario() -> ScenarioFixture {
             .allow_id(REGISTER)
             .allow_function(&record),
     )
-    // Turn 1 completes, the turn-completed binding fires, and the reaction
-    // seeds a SECOND terminal turn in the same session.
+    // Turn 1 completes, the turn-completed binding fires, and the explicitly
+    // pinned reaction seeds a SECOND terminal turn in the same session.
     .terminal_turn_statuses(["completed", "completed"])
     .generation(
         Generation::new(1)
@@ -65,7 +65,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                     .turn_request()
                     .system_prompt_sha256("{{system_prompt_sha256}}")
                     .messages_exact([Message::user(MESSAGE)])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER], [record.tool()]),
             )
             .respond(Response::function_call_raw(
                 "call-arm",
@@ -77,6 +77,9 @@ pub(super) fn scenario() -> ScenarioFixture {
                     "metadata": {
                         // Deliberately NO `options`: the reaction must inherit
                         // the registering turn's policy to call the recorder.
+                        // Pin the delivery session explicitly; unpinned
+                        // reactions intentionally create fresh children.
+                        "session_id": "{{session_id}}",
                         "task": "The wrap-up fired. Call the recorder exactly once \
                                  with value \"from-reaction\", then stop."
                     }
@@ -99,7 +102,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                         json!({ "role": "function_result", "function_call_id": "call-arm",
                                 "is_error": false }),
                     ])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER], [record.tool()]),
             )
             .respond(Response::text("armed", 10, 2)),
     )
@@ -121,7 +124,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                         // The reaction seed, fenced with the fired event.
                         json!({ "role": "user" }),
                     ])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER], [record.tool()]),
             )
             .respond(Response::function_call(
                 "call-record",
@@ -151,7 +154,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                         json!({ "role": "function_result", "function_call_id": "call-record",
                                 "is_error": false }),
                     ])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER], [record.tool()]),
             )
             .respond(Response::text("reaction recorded", 12, 3)),
     )
@@ -196,5 +199,7 @@ mod tests {
         let fixture = scenario();
         fixture.validate().unwrap();
         assert_eq!(fixture.expected_terminal_turns, 2);
+        assert_eq!(fixture.await_target_calls, None);
+        assert_eq!(fixture.expected_traces(), 1);
     }
 }
diff --git a/harness/tests/e2e/src/scenarios/reaction_unregisters_run.rs b/harness/tests/e2e/src/scenarios/reaction_unregisters_run.rs
index cac1bb554..84e92cec5 100644
--- a/harness/tests/e2e/src/scenarios/reaction_unregisters_run.rs
+++ b/harness/tests/e2e/src/scenarios/reaction_unregisters_run.rs
@@ -99,7 +99,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                     .turn_request()
                     .system_prompt_sha256("{{system_prompt_sha256}}")
                     .messages_exact([Message::user(MESSAGE)])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER, UNREGISTER], [record.tool()]),
             )
             .respond(Response::function_call_raw(
                 "call-arm",
@@ -123,7 +123,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                         json!({ "role": "function_result", "function_call_id": "call-arm",
                                 "is_error": false }),
                     ])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER, UNREGISTER], [record.tool()]),
             )
             // The registration function_result in THIS request carries the
             // runtime subscription id; capture it for the cleaner's call.
@@ -139,7 +139,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                     .turn_request_step(0)
                     .system_prompt_regex(".")
                     .messages_subset([json!({ "role": "user" })])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER, UNREGISTER], [record.tool()]),
             )
             .respond(Response::function_call_raw(
                 "call-unreg",
@@ -169,7 +169,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                                 "is_error": false,
                                 "content": [{ "type": "text", "text": "{\"removed\":true}" }] }),
                     ])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER, UNREGISTER], [record.tool()]),
             )
             .respond(Response::function_call(
                 "call-record",
@@ -186,7 +186,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                     .turn_request_step(2)
                     .system_prompt_regex(".")
                     .messages_subset([json!({ "role": "user" })])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER, UNREGISTER], [record.tool()]),
             )
             .respond(Response::text("cleaned up", 12, 3)),
     )
diff --git a/harness/tests/e2e/src/scenarios/state_worker_sidecar.rs b/harness/tests/e2e/src/scenarios/state_worker_sidecar.rs
index cd4707212..b09864a3c 100644
--- a/harness/tests/e2e/src/scenarios/state_worker_sidecar.rs
+++ b/harness/tests/e2e/src/scenarios/state_worker_sidecar.rs
@@ -66,8 +66,11 @@ pub(super) fn scenario() -> ScenarioFixture {
             .allow_id(REGISTER)
             .allow_function(&record),
     )
-    // Main turn (1) + the state-fired reaction turn (2).
-    .terminal_turn_statuses(["completed", "completed"])
+    // The state-fired reaction runs in a fresh, untracked child. Its recorder
+    // call proves the child ran; only the registering session's terminal turn
+    // and traces are tracked directly.
+    .await_target_calls(1)
+    .expect_traces(1)
     // The probe trips the key AFTER the main turn completes, so the reaction
     // turn runs alone. `state::set` on the standalone worker fans out to the
     // react binding — the seam under test.
@@ -84,7 +87,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                     .turn_request()
                     .system_prompt_sha256("{{system_prompt_sha256}}")
                     .messages_exact([Message::user(MESSAGE)])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER], [record.tool()]),
             )
             .respond(Response::function_call_raw(
                 "call-arm",
@@ -108,7 +111,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                         json!({ "role": "function_result", "function_call_id": "call-arm",
                                 "is_error": false }),
                     ])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER], [record.tool()]),
             )
             .respond(Response::text("armed", 10, 2)),
     )
@@ -121,7 +124,7 @@ pub(super) fn scenario() -> ScenarioFixture {
                     .turn_request_step(0)
                     .system_prompt_regex(".")
                     .messages_subset([json!({ "role": "user" })])
-                    .tools_exact([record.tool()]),
+                    .tools_exact_after_controls([REGISTER], [record.tool()]),
             )
             .respond(Response::function_call(
                 "call-record",
@@ -137,21 +140,28 @@ pub(super) fn scenario() -> ScenarioFixture {
                 Request::new()
                     .turn_request_step(1)
                     .system_prompt_regex(".")
-                    // The reaction runs in the registering session, so its
-                    // history begins with the original user message; the
-                    // strict-ordinal router + step already sequence gen3→gen4.
-                    // The recorder call itself is asserted in `verify`.
-                    .messages_subset([json!({ "role": "user" })])
-                    .tools_exact([record.tool()]),
+                    .messages_subset([
+                        json!({ "role": "user" }),
+                        json!({ "role": "assistant", "content": [
+                            { "type": "function_call", "id": "call-record" }
+                        ] }),
+                        json!({ "role": "function_result", "function_call_id": "call-record",
+                                "is_error": false }),
+                    ])
+                    .tools_exact_after_controls([REGISTER], [record.tool()]),
             )
             .respond(Response::text("state reaction recorded", 12, 3)),
     )
     .verify(|run| {
-        // The second terminal turn exists ONLY if the state worker delivered
-        // the metadata sidecar so harness::react could resolve the task.
-        run.expect_assistant_texts(["armed", "state reaction recorded"])?;
-        run.expect_function_calls("record", 1)?;
-        run.expect_call_payload("record", json!({ "value": "from-state-reaction" }))?;
+        // The recorder is reachable only if the state worker delivered the
+        // metadata sidecar so harness::react could resolve and spawn the task.
+        run.expect_assistant_texts(["armed"])?;
+        run.expect_target_calls(1)?;
+        anyhow::ensure!(
+            run.target_calls[0] == json!({ "value": "from-state-reaction" }),
+            "recorder payload {:?} != from-state-reaction",
+            run.target_calls[0]
+        );
 
         for item in &run.transcript {
             let msg = item.get("message").cloned().unwrap_or(Value::Null);
@@ -183,10 +193,12 @@ mod tests {
     use super::*;
 
     #[test]
-    fn fixture_is_valid_and_awaits_the_reaction_turn() {
+    fn fixture_is_valid_and_awaits_the_untracked_reaction() {
         let fixture = scenario();
         fixture.validate().unwrap();
-        assert_eq!(fixture.expected_terminal_turns, 2);
+        assert_eq!(fixture.expected_terminal_turns, 1);
+        assert_eq!(fixture.await_target_calls, Some(1));
+        assert_eq!(fixture.expected_traces(), 1);
         assert_eq!(fixture.probe_actions.len(), 1);
         assert_eq!(fixture.probe_actions[0].after_turns, 1);
     }
diff --git a/harness/tests/golden/schemas/harness.metrics.json b/harness/tests/golden/schemas/harness.metrics.json
new file mode 100644
index 000000000..b01c6cd8c
--- /dev/null
+++ b/harness/tests/golden/schemas/harness.metrics.json
@@ -0,0 +1,316 @@
+{
+  "function_id": "harness::metrics",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "properties": {
+      "root_session_id": {
+        "type": "string"
+      }
+    },
+    "required": [
+      "root_session_id"
+    ],
+    "title": "SessionMetricsRequestV1",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "additionalProperties": false,
+    "definitions": {
+      "SessionTraceMetricsV1": {
+        "additionalProperties": false,
+        "properties": {
+          "by_session": {
+            "items": {
+              "$ref": "#/definitions/SessionTraceUsageV1"
+            },
+            "type": "array"
+          },
+          "duration_ms": {
+            "description": "Elapsed window from the first observed span to the last observed span.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "error_span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "trace_count": {
+            "description": "Distinct traces across the root session and all descendants.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "by_session",
+          "duration_ms",
+          "error_span_count",
+          "span_count",
+          "trace_count"
+        ],
+        "type": "object"
+      },
+      "SessionTraceUsageV1": {
+        "additionalProperties": false,
+        "properties": {
+          "depth": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "duration_ms": {
+            "description": "Elapsed window from the session's first observed span to its last.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "error_span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "parent_session_id": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "session_id": {
+            "type": "string"
+          },
+          "span_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "trace_count": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "depth",
+          "duration_ms",
+          "error_span_count",
+          "session_id",
+          "span_count",
+          "trace_count"
+        ],
+        "type": "object"
+      },
+      "SessionUsageTotalsV1": {
+        "additionalProperties": false,
+        "properties": {
+          "cache_read_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cache_write_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "function_call_errors": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "function_calls": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "input_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "output_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "reasoning_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "sessions": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "turns": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "function_call_errors",
+          "function_calls",
+          "sessions",
+          "turns"
+        ],
+        "type": "object"
+      },
+      "SessionUsageV1": {
+        "additionalProperties": false,
+        "properties": {
+          "cache_read_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cache_write_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "depth": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "function_call_errors": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "function_calls": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "input_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "output_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "parent_session_id": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "reasoning_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "session_id": {
+            "type": "string"
+          },
+          "turns": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "depth",
+          "function_call_errors",
+          "function_calls",
+          "session_id",
+          "turns"
+        ],
+        "type": "object"
+      }
+    },
+    "properties": {
+      "by_session": {
+        "items": {
+          "$ref": "#/definitions/SessionUsageV1"
+        },
+        "type": "array"
+      },
+      "complete": {
+        "type": "boolean"
+      },
+      "root_session_id": {
+        "type": "string"
+      },
+      "totals": {
+        "$ref": "#/definitions/SessionUsageTotalsV1"
+      },
+      "traces": {
+        "anyOf": [
+          {
+            "$ref": "#/definitions/SessionTraceMetricsV1"
+          },
+          {
+            "type": "null"
+          }
+        ],
+        "description": "Trace/span aggregates when the engine's in-memory observability exporter is available. Usage metrics remain available when it is not."
+      }
+    },
+    "required": [
+      "by_session",
+      "complete",
+      "root_session_id",
+      "totals"
+    ],
+    "title": "SessionMetricsResponseV1",
+    "type": "object"
+  }
+}
diff --git a/harness/tests/golden/schemas/harness.send.json b/harness/tests/golden/schemas/harness.send.json
index dbc02a971..eab964f00 100644
--- a/harness/tests/golden/schemas/harness.send.json
+++ b/harness/tests/golden/schemas/harness.send.json
@@ -385,7 +385,7 @@
         "description": "`message` is either a plain string (sugar for a user text message) or a full `AgentMessage`."
       },
       "Mode": {
-        "description": "Console / send operating mode — prepends a short paragraph before the shared identity prompt. `ask` is also structural: the turn's dispatch policy is capped at the operator's read-only baseline, never widened.",
+        "description": "Console / send operating mode — prepends a short paragraph before the shared identity prompt. `ask` is also structural: the turn's dispatch policy is capped at the configured default policy, never widened.",
         "enum": [
           "ask",
           "agent"
@@ -438,7 +438,33 @@
                 "type": "null"
               }
             ],
-            "description": "The fail-closed dispatch policy. Omitted on a NEW session → deny every call; omitted when steering an EXISTING session → inherit the prior turn's policy (a nudge must not disarm a live run). Pass `{ allow: [] }` to strip explicitly. On a NEW `ask`-mode turn the effective policy is capped at the operator's read-only baseline; a steer folded into an already-running turn keeps that turn's frozen policy until it finalises."
+            "description": "The fail-closed dispatch policy. Omitted on a NEW session → deny every call; omitted when steering an EXISTING session → inherit the prior turn's policy (a nudge must not disarm a live run). Pass `{ allow: [] }` to strip explicitly. On a NEW `ask`-mode turn the effective policy is capped at the configured default policy; a steer folded into an already-running turn keeps that turn's frozen policy until it finalises."
+          },
+          "max_cost_usd": {
+            "description": "Hard USD budget for the complete root-and-subagent session tree. Every model used by the tree must advertise catalog pricing.",
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "max_output_tokens": {
+            "description": "Per-generation output-token ceiling forwarded to the router.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "max_total_tokens": {
+            "description": "Hard input-plus-output token budget for the complete root-and-subagent session tree.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
           },
           "max_turns": {
             "format": "uint32",
@@ -493,7 +519,7 @@
               }
             ],
             "default": "enrich",
-            "description": "How `system_prompt` combines with the built-in prompt: `override` replaces it; `enrich` (default) appends to it."
+            "description": "How `system_prompt` combines with the built-in prompt: `override` replaces it; `enrich` (default) appends to it; `disabled` omits it."
           },
           "thinking_level": {
             "anyOf": [
@@ -547,6 +573,13 @@
               "enrich"
             ],
             "type": "string"
+          },
+          {
+            "description": "No system prompt is sent to the model.",
+            "enum": [
+              "disabled"
+            ],
+            "type": "string"
           }
         ]
       },
diff --git a/harness/tests/golden/schemas/harness.session-tree.json b/harness/tests/golden/schemas/harness.session-tree.json
new file mode 100644
index 000000000..bf02e965e
--- /dev/null
+++ b/harness/tests/golden/schemas/harness.session-tree.json
@@ -0,0 +1,73 @@
+{
+  "function_id": "harness::session-tree",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "properties": {
+      "root_session_id": {
+        "type": "string"
+      }
+    },
+    "required": [
+      "root_session_id"
+    ],
+    "title": "SessionTreeRequestV1",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "additionalProperties": false,
+    "definitions": {
+      "SessionTreeNodeV1": {
+        "additionalProperties": false,
+        "properties": {
+          "depth": {
+            "format": "uint32",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "parent_session_id": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "parent_turn_id": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "session_id": {
+            "type": "string"
+          }
+        },
+        "required": [
+          "depth",
+          "session_id"
+        ],
+        "type": "object"
+      }
+    },
+    "properties": {
+      "complete": {
+        "type": "boolean"
+      },
+      "root_session_id": {
+        "type": "string"
+      },
+      "sessions": {
+        "items": {
+          "$ref": "#/definitions/SessionTreeNodeV1"
+        },
+        "type": "array"
+      }
+    },
+    "required": [
+      "complete",
+      "root_session_id",
+      "sessions"
+    ],
+    "title": "SessionTreeResponseV1",
+    "type": "object"
+  }
+}
diff --git a/harness/tests/golden/schemas/harness.spawn.json b/harness/tests/golden/schemas/harness.spawn.json
index ddbdc0bd0..539d37109 100644
--- a/harness/tests/golden/schemas/harness.spawn.json
+++ b/harness/tests/golden/schemas/harness.spawn.json
@@ -385,7 +385,7 @@
         "description": "`message` is either a plain string (sugar for a user text message) or a full `AgentMessage`."
       },
       "Mode": {
-        "description": "Console / send operating mode — prepends a short paragraph before the shared identity prompt. `ask` is also structural: the turn's dispatch policy is capped at the operator's read-only baseline, never widened.",
+        "description": "Console / send operating mode — prepends a short paragraph before the shared identity prompt. `ask` is also structural: the turn's dispatch policy is capped at the configured default policy, never widened.",
         "enum": [
           "ask",
           "agent"
@@ -444,7 +444,7 @@
                 "type": "null"
               }
             ],
-            "description": "Intersected with the parent policy — narrow, never escalate. An `ask`-mode child is further capped at the read-only baseline."
+            "description": "Intersected with the parent policy — narrow, never escalate. An `ask`-mode child is further capped at the configured default policy."
           },
           "max_children": {
             "description": "Fan-out guard for the child's own spawns.",
@@ -455,8 +455,17 @@
               "null"
             ]
           },
+          "max_output_tokens": {
+            "description": "Inherits the parent's ceiling unless explicitly narrowed/overridden.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
           "max_turns": {
-            "description": "Capped at the parent's remaining turn budget.",
+            "description": "Capped at the parent's remaining turn budget. Omit unless a strict child-specific cap is required. It must cover discovery/contract calls plus every work call; very small values (for example 2-3) commonly strand the child before it can produce its deliverable.",
             "format": "uint32",
             "minimum": 0.0,
             "type": [
@@ -498,7 +507,7 @@
               }
             ],
             "default": "enrich",
-            "description": "How `system_prompt` combines with the built-in prompt: `override` replaces it; `enrich` (default) appends to it."
+            "description": "How `system_prompt` combines with the built-in prompt: `override` replaces it; `enrich` (default) appends to it; `disabled` omits it."
           },
           "thinking_level": {
             "anyOf": [
@@ -539,6 +548,13 @@
               "enrich"
             ],
             "type": "string"
+          },
+          {
+            "description": "No system prompt is sent to the model.",
+            "enum": [
+              "disabled"
+            ],
+            "type": "string"
           }
         ]
       },
@@ -693,7 +709,7 @@
             "$ref": "#/definitions/MessageInput"
           }
         ],
-        "description": "The child's goal — its opening user message."
+        "description": "The child's self-contained goal — its opening user message. Include every resolved required selector literally (for example `Use database db: \"primary\"`); the child cannot infer resources from the parent."
       }
     },
     "required": [
diff --git a/harness/tests/manifest.rs b/harness/tests/manifest.rs
index 2910e6a2c..54a04e5d5 100644
--- a/harness/tests/manifest.rs
+++ b/harness/tests/manifest.rs
@@ -70,7 +70,7 @@ fn worker_manifest_uses_the_standalone_state_worker() {
 
     assert_eq!(
         dependencies.get(serde_yaml::Value::String("state".into())),
-        Some(&serde_yaml::Value::String("^0.21.2".into()))
+        Some(&serde_yaml::Value::String("^0.21.3".into()))
     );
     assert!(!dependencies.contains_key(serde_yaml::Value::String("iii-state".into())));
 }
diff --git a/harness/tests/schemas.rs b/harness/tests/schemas.rs
index 8c9016df4..5e08b10cf 100644
--- a/harness/tests/schemas.rs
+++ b/harness/tests/schemas.rs
@@ -40,6 +40,8 @@ fn catalog_lists_all_functions_in_registration_order() {
             "harness::function::resolve",
             "harness::stop",
             "harness::status",
+            "harness::session-tree",
+            "harness::metrics",
             "harness::react",
         ]
     );
diff --git a/iii-permissions.yaml b/iii-permissions.yaml
index 6dae851f6..d59e2a674 100644
--- a/iii-permissions.yaml
+++ b/iii-permissions.yaml
@@ -101,6 +101,12 @@ rules:
   - '!harness::stop'
   - '!harness::on-config-change'
   - '!harness::sweep-pending'
+  # eval: durable orchestration internals are trigger/queue targets only.
+  # Starting, cancelling, and deleting evaluations remain approval-gated by
+  # the default. Status/result are read-only and the exact evaluator is pure.
+  - '!eval::step'
+  - '!eval::on-turn-completed'
+  - '!eval::sweep'
   # Trigger-bridge target: fired only by subscriptions (engine-side, bypassing
   # this gate). Never agent-callable directly — the model binds it via
   # engine::register_trigger and names its id from the system prompt.
@@ -166,6 +172,9 @@ rules:
   - router::provider::list
   # harness::status is read-only (current turn status for a session).
   - harness::status
+  - eval::status
+  - eval::result
+  - eval::assert::exact
   - oauth::anthropic::status
   - oauth::openai-codex::status
   - engine::functions::list
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index 4fdb7b312..4a4f24248 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -195,6 +195,22 @@ importers:
         specifier: ^5.9.2
         version: 5.9.3
 
+  eval/ui:
+    dependencies:
+      '@iii-dev/console-ui':
+        specifier: workspace:*
+        version: link:../../packages/console-ui
+    devDependencies:
+      '@types/react':
+        specifier: ^19.2.14
+        version: 19.2.17
+      esbuild:
+        specifier: ^0.25.0
+        version: 0.25.12
+      typescript:
+        specifier: ^5.9.2
+        version: 5.9.3
+
   packages/console-ui:
     devDependencies:
       '@types/react':
diff --git a/pnpm-workspace.yaml b/pnpm-workspace.yaml
index e6430eca5..83176624d 100644
--- a/pnpm-workspace.yaml
+++ b/pnpm-workspace.yaml
@@ -10,6 +10,7 @@ packages:
   - console/web
   - console/ui
   - database/ui
+  - eval/ui
   - state/ui
   - iii-directory/ui