(MOT-4184) feat(harness): add prompt evaluation support

.github/workflows/create-tag.yml

@@ -22,6 +22,7 @@ on:
           - cron
           - database
           - email
+          - eval
           - harness
           - hermes
           - http

.github/workflows/release.yml

@@ -17,6 +17,7 @@ on:
       - 'cron/v*'
       - 'database/v*'
       - 'email/v*'
+      - 'eval/v*'
       - 'harness/v*'
       - 'hermes/v*'
       - 'http/v*'

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. |

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"

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(