Phase 5 Step 7 (gated): registry becomes the single subagent store — do not merge until Steps 5-6 are stable - #44235
Closed
claw3649 wants to merge 64 commits into
Closed
Conversation
…res in the UI Model-resolution reliability and picker UX, including the corrections found in review of the in-flight batch: - model_switch: discover keyless local endpoints (ollama/llama.cpp/LM Studio) so `ollama rm`'d models stop lingering in the picker — but only when no explicit `models:` subset is pinned, preserving an intentional narrowing (parity with the custom-providers section; fixes 2 test regressions). - models: bypass the disk cache for custom/local providers (always live), falling through to the shared stale-beats-empty path so a transiently unreachable endpoint keeps its last-known-good list instead of going empty. - model_metadata: pin local Gemma 4 26B to its served 64k num_ctx ahead of the /api/show GGUF-max probe so the token budget can't over-run and silently truncate; reconcile a stale cache entry in place instead of re-probing every call (the previous guard never converged). - web/desktop picker: await the /model switch and surface a rejected switch (result.error) inline — keep the dialog open, roll back the optimistic model, toast benign advisories — instead of closing as though it worked. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- launchd_restart: try a graceful SIGUSR1 self-restart before terminate, with a hard-stop fallback; bind the force-kill to the freshly-resolved live PID (not the captured one) so PID reuse can't SIGKILL an unrelated process. - hermes launcher: re-exec into ./venv/bin/python on Python <3.11, but only after verifying that interpreter is itself >=3.11 — a stale venv no longer silently bypasses the version guard and crashes later with an opaque error. - tests: cover the SIGUSR1 graceful path and reconcile the terminate-drain fallback test with the new graceful-first behavior. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ontract Phase 2 of the Central Brain + Action Runtime design (docs/architecture): a single honest-status schema the gateway's exec handlers funnel through. - contract: ExecutionTask / ExecutionResult / Status / ErrorType / ExecError / SideEffect / NeedsInput / Constraints. `outputs` is a lossless payload bag so each handler's native shape round-trips back to its exact wire dict. - adapters: shell / cli / plugin / slash, each a pure (to_result, to_wire) pair with a byte-identical round-trip guarantee. - tests assert the round-trip identity (the byte-compat oracle). Pure dataclasses with no dependency on the gateway (server imports this, not the reverse). Nothing consumes the rich fields yet — that is Phase 4. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… Runtime - slash.exec: report a rejected live side-effect (/model, busy) and a raised plugin handler as a structured result.error (additive — TUI/web that read only output/warning are unaffected), keying off _SlashSideEffect.kind rather than prefix-matching the message text. - shell.exec / cli.exec / slash.exec / plugin path now build an ExecutionResult and render it back through the adapters — byte-identical wire, so the three frontends need no change. A non-zero exit is an honest FAILED result whose wire shape (result.code) is unchanged. - desktop typed `/model` path: surface result.error instead of rendering a rejected switch as a success line; add error? to SlashExecResponse. - add honest-status regression coverage for the plugin error field and for the already-honest shell.exec / cli.exec paths (none existed before). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A design doc for separating the Orchestration Core (reasoning/planning/memory) from the Action Runtime (execution/tool-use), grounded in the current Hermes gateway. Covers the decided terminology (OpenClaw = compat/reference only), the ExecutionTask/ExecutionResult contract, the phased migration (1a honest-status → 2 unified path → 3 SessionState + Brain-host → 4 intent/idempotency → 5 multi-runtime), cross-cutting rules (additive-first, test-first, observability, idempotency), and a decision log tracking what has landed. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…n sites Phase 3 Step 1 (Central Brain + Action Runtime design, §11). The per-session state was an untyped dict at _sessions[sid] (~58 refs, ~30 keys, locks floating as dict entries). Introduce a SessionState type to anchor the decomposition. SessionState subclasses `dict`, so every existing session["key"] / .get / .setdefault / "k" in session call keeps EXACT dict semantics — zero handler churn. A safety analysis confirmed the gateway has no membership test, bare pop, or iteration that relies on the session value dict, and that absent keys must stay absent (a dict subclass gives that for free, so no field defaults need tuning — notably agent_ready stays absent on the eager path and personality keeps its config fallback). Typed @Property accessors (history, history_lock, running, agent, …) let handlers migrate to attribute access incrementally in later steps; the locks stay dict items until the subscript sites are converted. Wraps the two real creation literals (_init_session, session.create); test-constructed plain dicts still work since SessionState IS a dict. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
SessionState type introduced and committed; note the next step (Step 2 lock-core methods). Also reflects the verified AIAgent instantiation-site count. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…robe Generalizes the root cause behind the Gemma4-26B 64k pin: Ollama's /api/show reports the GGUF training max, which over-states the window a LOCAL server actually serves (the user-set Modelfile num_ctx) and silently truncates conversations once they outgrow it. Steps 2b and 5e now resolve local endpoints num_ctx-first (query_ollama_num_ctx) while hosted servers keep the GGUF-first probe (their users can't set num_ctx); discovered values persist under the same lmstudio cache exclusion the sibling steps use. The explicit Gemma pin stays as a tested guard above these steps. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Opening the picker live-probes every custom/local endpoint (keyless-local discovery in sections 3/4 and the live "custom" branch of provider_model_ids); with a server down each open burns the full HTTP timeout. Add a 30s in-process memo around fetch_api_models — keyed by endpoint + credential fingerprint and pinned to the exact function object so monkeypatched tests stay hermetic (plus an autouse clear in tests/hermes_cli). Empty results and exceptions are never memoized, so a down server still retries on the next open and the stale-beats-empty disk-cache semantics are untouched. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…istry Pre-existing cross-file pollution: the dispatch-pool tests overwrite entries in the module-global _methods registry (slash.exec, session.compress, …) without restoring them, so four session.compress/snapshot tests in test_tui_gateway_server.py fail whenever both files run in one process. The server fixture now snapshots and restores _methods in place (same dict object) plus the directly-assigned module globals (_real_stdout, _hermes_home, _cfg_cache/_cfg_mtime/_cfg_path). Tests-only; both files also stay green alone. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Phase 3 Step 2 (Central Brain + Action Runtime design, §11). The compaction snapshot + compare-and-swap, the turn-release finally, and the one-shot agent-build guard now live as SessionState methods — snapshot_history, commit_compaction, begin_turn, end_turn, build_once — each a verbatim lift of the server.py block it replaced (same lock, same .get defaults, same order). Five gateway sites migrated; sites whose critical sections interleave extra work with the CAS (the prompt.submit busy-gate + truncate, the poller gates, bare running-flag cleanups, the mid-turn history CAS that logs under the lock) stay deliberately inline to preserve atomicity, each noted in place. Test helpers that feed the migrated paths now construct SessionState (a dict subclass, so semantics are unchanged); 9 unit tests pin the method contracts. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…esult wire
Phase 4 (Central Brain + Action Runtime design, §11): make the rich
ExecutionResult consumable.
- slash.exec accepts an optional task_id and echoes it additively (the wire is
byte-identical when absent — existing protocol tests pass unmodified).
- result_to_wire_rich renders the full contract: task_id, status, outputs,
error {type, retryable, message}, side_effects.
- New task.submit RPC: pilot scope intent="slash" only (reversible commands,
e.g. /model), other intents → 4030. The slash.exec body is extracted into a
shared _slash_exec_core returning the ExecutionResult; slash.exec renders the
legacy wire exactly as today, task.submit renders the rich wire.
- Idempotency: a locked in-process store (10-min TTL, 1024-entry cap) replays
the recorded result with replayed=true instead of re-executing; task_id
defaults to a fresh uuid4.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…-flight token Phase 4 race fix (from the PR review): selectModel applied optimistic switches with no in-flight correlation, so a slow switch A failing after a faster switch B succeeded rolled the store back to the pre-A model, clobbering B. Each selectModel call now captures a monotonically-increasing token; once a newer call starts, the older call is stale — a stale success commits nothing and a stale failure skips the store/cache rollback (the snapshot predates the newer switch) while keeping the failure toast so the user still learns their requested switch failed. refreshCurrentModel gains an optional staleness guard used by the persistGlobal paths; external callers are unaffected. The slash.exec request also carries an additive task_id for forward correlation; staleness is decided purely by the local token. Two race tests cover both interleavings. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…sion log Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…us surfacing Upstream moved the web model picker's chat path from slash.exec to config.set and added the expensive-model confirmation flow (pendingConfirm). Resolution keeps upstream's architecture and folds our honest-status work into it: all three modes (standalone onApply, gateway config.set, legacy onSubmit fallback) funnel through one awaited try/catch — a failed switch surfaces inline via submitError and the dialog stays open; a benign post-switch warning fires the new onWarning prop, which ChatSidebar toasts. The now-dead slash.exec onModelSubmit path is removed with upstream. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ttributes Phase 3 Step 3: 107 square-bracket sites in tui_gateway/server.py now use the typed accessors for the 12 hot fields (history_lock=25, session_key=22, agent=17, running=11, history_version=8, history=7, the rest 2-3 each). Mechanical and behavior-preserving by rule: only ["x"] reads/writes were converted — every .get()/.setdefault()/.pop() stays untouched because absent-key semantics differ (None/default vs KeyError), and non-session dicts (DB rows, payload/info dicts) are excluded. Tests that inject plain-dict sessions into the migrated paths now wrap them in SessionState (a dict subclass — no behavioral change to the tests themselves). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Phase 3 Brain-host proof-of-life (design §11 3c/3d): agent/brain_host.py hosts AgentSpec + a process-singleton BrainHost whose build_agent() constructs the SAME AIAgent the call site would (parity is asserted by test, not assumed). _make_agent now assembles its kwargs once and, only when HERMES_BRAIN_HOST=1, routes through the host (lazy import inside the branch — default-off adds zero import cost, proven via sys.modules in tests). V1 is deliberately the seam only; credential-pool/tool-schema/memory-session sharing arrives as the other ~20 construction sites migrate. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ision log Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ns through BrainHost Second and third construction sites behind the HERMES_BRAIN_HOST flag (default off, byte-equivalent when unset): gateway/run.py's per-turn LRU-miss site (intent "gateway-run") and the API server's _create_agent (intent "api-server"). Same assemble-kwargs-then-gate pattern as _make_agent; seven new gate tests (the api-server gate is unit-driven; the gateway-run gate is compile+source asserted — its 1200-line async host method is not unit-drivable). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…gistry) Grounded scoping of delegation.*/spawn_tree.*/delegate_tool state (registries, locks, persistence, interrupt trees) and the unified design: one AgentTaskRegistry keyed by task_id speaking ExecutionTask/ExecutionResult, with a 6-step dual-write migration, risks R1-R7 (notably: the AIAgent-internal _active_children interrupt chain must not be replaced), and the maintainer decisions Q1-Q6 required before implementation starts. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
build_agent(intent, **kwargs) is the compact form of the HERMES_BRAIN_HOST gate for migrating the remaining AIAgent construction sites. It lives in its own light module (imports only os) so the flag-off path never imports agent.brain_host — the zero-footprint invariant the gate tests assert — and run_agent is only imported at call time, matching the lazy-import style of the call sites it replaces. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…n_host_gate
Converts the 14 construction sites that were still calling AIAgent()
directly to build_agent("<intent>", ...) — exact same kwargs, callable
swap only — and collapses the three pre-existing inline HERMES_BRAIN_HOST
gate blocks (tui_gateway/_make_agent, gateway-run LRU-miss,
api_server._create_agent) into single build_agent calls with their
original intent strings.
New intents: prompt-size, oneshot, cli-background, cli, background-review,
curator, acp, cron, compress, history-hygiene, gateway-background,
feishu-comment, tui-background, preview-restart, batch, run-agent-cli,
delegate.
Deliberately untouched: cli.py's lazy AIAgent wrapper (re-export shim),
scripts/tool_search_livetest.py (dev script), and everything around
_active_subagents / spawn trees / interrupt handling in delegate_tool.
Tests: 4 unit tests for the gate helper (flag off/zero/on + kwargs
parity) and a parametrized site-table source check covering all 16
files / 20 intents, replacing the old grep-based gate-exists test whose
marker strings the gate collapse removed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…scripts to typed attributes
Adds 10 property accessors to SessionState (attached_images,
pending_title, image_counter, show_reasoning, tool_progress_mode,
tool_started_at, personality, model_override, explicit_cwd,
edit_snapshots) with exact self["key"] read/write semantics, and
converts every server.py subscript on those keys to attribute access.
Two sites deliberately stay as subscripts: session["model_override"] in
_apply_model_switch (one caller passes a plain {"agent": None} dict on
the no-session global-switch path — the isinstance guard there exists
for exactly this) and session["_finalized"] (single private marker).
Accessor-sync and write-through tests extended to cover the 10 new
accessors.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…3 Step 4 in the decision log Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…stions Q1-Q2, Q4-Q6 Q3 (Status.RUNNING placement) stays open pending an impact sweep the maintainer requested: a consumer census of Status/ExecutionResult and proof that record-level TaskStatus loses no running-task visibility. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…sus, visibility map, lifecycle) Three-agent sweep: option A touches ~8 sites with one genuine breakage (the idempotency replay store would cache a non-terminal snapshot as the permanent answer) plus silent-success hazards on every legacy wire; option B touches zero existing sites and loses no running-task visibility, since every in-flight view (delegation.status, the subagent.*/tool.* event stream, bg counters) lives outside ExecutionResult. Recommendation recorded as option B, pending the maintainer's ruling. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… unblocked Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…itive) The in-process ledger for live and completed agent tasks, encoding the Q1-Q6 rulings: standalone module (not a BrainHost tenant), ephemeral records and replay store, a record-level TaskStatus enum carrying RUNNING while contract.Status stays terminal-only, and method-private storage so a shared backend can replace the dicts later. Nothing imports it yet — Step 2 dual-writes from delegate_tool. The interrupt path resolves the stored agent (weakref or direct) and calls agent.interrupt() outside the registry lock, the same mechanic as interrupt_subagent today. The replay store mirrors _TASK_RESULTS semantics (600s TTL, 1024 cap) ahead of the Step 5 fold-in. 17 unit tests, including two Q3-ruling pins: contract.Status has no RUNNING, and TaskStatus terminal values mirror contract.Status exactly. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…d the landed Step 1 path Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… Phase 5 Step 2 Each child registers an AgentTaskRecord (intent="delegate", weakref agent_ref) right after _register_subagent and completes it in the same finally block that unregisters the legacy dict entry, mapping the existing per-child entry to a terminal ExecutionResult: completed → SUCCEEDED, interrupted/timeout → FAILED + ErrorType.TRANSPORT (non-retryable, per the doc ruling), failed → PROVIDER_ERROR, escaped exception → INTERNAL, died-without-entry → complete(None) → FAILED. Both registry calls are try/except-guarded so a registry bug can never break a real delegate run. _active_subagents, _active_children, heartbeat and timeout paths are untouched — the registry is a parallel ledger until the Step 7 cutover. 8 new tests; delegate suites 172 green. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…sh refactors, keep honest-status surfacing Upstream restructured executeSlashCommand into a slash-action table with a generic runExec path (and then made slash commands registry-driven in #42351); /model-with-args now rides runExec, so our result.error honest-status check lives there — covering every exec-style command instead of just the typed /model path. The refreshCurrentModel guard keeps BOTH protections: upstream's live-session-owns-the-footer guard and our in-flight staleness token. The model-controls test file unions upstream's two refreshCurrentModel tests with our four selectModel rollback/race tests (6/6 green under --environment jsdom). Desktop src/ suite shows the byte-identical failure set to pure origin/main (6 pre-existing); python gateway suites green. Note: this replaces an earlier broken push of the same resolution — a mid-merge stash for a baseline measurement silently dropped MERGE_HEAD, so 46a668be2 landed as a single-parent commit with partial upstream content. Reset and redone as a true merge; baselines are measured in detached worktrees from now on. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…e up the stack Single conflict: config.set reasoning_effort gained _persist_live_session_runtime + a session.info emit upstream while this branch had converted the same lines to SessionState attribute access — resolution keeps both (attribute style per Phase 3 Step 3). Suites: 489 green (single known browser-hint env failure). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… closes review gap 1 delegate_task gains an optional registry_parent_task_id (default None = byte-identical engine default); when the gateway's intent="delegate" handler passes its task_id, child registry records link to the submitted parent record per the doc's Step 5 ruling. Only the AgentTaskRecord relabels — the legacy TUI dict keeps the engine's _parent_subagent_id, so interrupt propagation, depth display and the spawn-tree overlay are untouched. Non-string values degrade to the engine default. 4 new tests pin both ends of the chain (engine register site + gateway kwarg pass-through). 343 green. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Session-keyed tasks store mirroring the subagents store: seeded from task.list on every gateway open (doubling as reconnect reconciliation — running rows absent from the live list drop rather than getting an invented terminal status), updated live from task.started/ task.completed (snapshot session_id is authoritative; events without one are dropped, never attached to the focused chat). TaskBoard renders one row per task — status glyph, intent badge, label, tool progress, age — terminal rows linger 60s then drop. Mounted above the spawn tree in AgentsView. 9 store/render tests + canary suites green; tsc and eslint clean. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
After Step 7 the legacy-shaped list_active_subagents() is a thin wrapper over the registry, so a caller-supplied parent task id shows through it — the dual-ledger assertion (engine id retained in the TUI dict) only holds pre-cutover and stays pinned on the trace-events branch. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…op board, measured BrainHost cache Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…er merge Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…eview system Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…d client Mirrors the tools-card architecture exactly: seeded once via task.list on the JSON-RPC sidecar after session.create, live-updated from task.started/task.completed frames on the SAME /api/events socket the tools card uses. Rows: status glyph, intent badge, label/goal, live age, tool detail or error; terminal rows linger 60s. Pure reducer helpers extracted dependency-free (web/ has no test runner yet). Typecheck clean; web lint baseline untouched (44 pre-existing problems in unrelated files). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Author
🧪 Dogfooding dossier — evidence toward the "Steps 5-6 stable" merge gateDrove the full stack on this branch (real stdio JSON-RPC gateway,
9 warts catalogued in the exercise; the 4 substantive ones (child session-id namespace split, in-flight idempotency gap for key-only retries, child results missing This is one sustained synthetic session, not production mileage — posting it as the first stability data point for the gate, not the last. |
alt-glitch
added
type/refactor
…light key replay, child result ids
WART 1 — session-id namespace split: delegate batch children registered
under the agent's persistent session key while the parent carried the
gateway sid, splitting _tasks.jsonl across two spawn-trees dirs and
emitting child task.* events with a session_id matching no live gateway
session (dropped on per-session WS transports). delegate_task gains an
optional registry_session_id threaded to the child register site;
_task_submit_delegate passes the sid it already resolved. Default None
keeps non-gateway delegate runs unchanged.
WART 2 — in-flight idempotency gap: the duplicate guard only checked
the caller's task_id, so a broken-pipe retry reusing the
idempotency_key with a fresh task_id mid-batch would re-execute.
registry.find_running_by_key + the key stamped on the parent record;
retries now get the same 4034 rejection naming the running task_id.
The after-completion replayed:true path is untouched.
WART 3 — child ExecutionResults now embed the child's task_id (was
None, contradicting the always-carry rule parents follow).
WART 4 — delegate aggregate results entries gain an additive task_id
key so a Core caller can correlate entries to registry child records
without relying on the sa-{index}-{hex} naming convention (all existing
consumers use .get() access — verified by grep).
Found by live dogfooding (dossier on the Step 7 PR). 351 green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
claw3649
marked this pull request as ready for review
June 11, 2026 12:18
…over The wart fixes land canonically on the parent branch (as promised in the Step 7 PR dossier comment); this merge adapts the child-register site to the post-cutover single-ledger shape. Replaces a rogue direct commit (805c83407, never pushed) that an agent made here against instructions — that work was preserved and ported to the parent first. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ent wrapper Three CLI tests stubbed construction by patching cli.AIAgent — a pure pass-through wrapper the BrainHost migration no longer routes through — so the real AIAgent ran and each test failed on its environment (CI caught this; our targeted macOS runs never covered these files). Investigated each as a potential real regression first: the MCP wait-before-build ordering, the background-thread callback registration order, and the single resolve_runtime_provider call per init are all unchanged in production (the extra counted resolutions came from real AIAgent internals that the old patch merely prevented from running). Test-only fix: patch agent.brain_host_gate.build_agent — the seam production actually uses — with every behavioral assertion unchanged. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… idle reaper test_ws_disconnect_preserves_and_repoints_reconnectable_session seeds bare session dicts with no created_at/last_active. The module-level idle reaper reads missing timestamps as 0.0 — idle since the epoch — so the moment handle_ws's finally closes the transport, the session is evictable, and a reaper tick landing in the close-to-assert window pops it: KeyError 'plain'. The 300s tick makes this a once-in-many-runs flake that only bites long-lived pytest processes (CI's per-file subprocesses exit before the first tick), which is why it resisted plain-loop reproduction. A probe driving the production scan body at a fast tick reproduced the exact failure 199/200 before the fix and 0/200 after. Production is immune: every real creation site stamps both timestamps. Test-only: a _fresh_session helper stamps time.time() on both seeded sessions. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…tyle repoint works After Phase 3 Step 3 the WS-disconnect repoint path sets session.transport = _detached_ws_transport (attribute style); on a plain-dict seed that raises AttributeError, silently skipping the repoint, so test_ws_disconnect_preserves_and_repoints_reconnectable_session saw the session keep its live WSTransport (CI caught this — our targeted runs never ran this file's slice). _fresh_session now returns a SessionState, matching every real creation site. Folds the two reasons the seed must be a SessionState (timestamps + attribute access) into one docstring. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Author
Mileage dossier round 2 — flag-on BrainHost / Steps 5–6 stabilityRun context:
Pre-live regression pinsTargeted wart/parity pins with the flag on: Broader protocol/registry suite: Git tree after both test runs stayed unchanged except the pre-existing untracked Live JSON-RPC checksProtocol robustness:
Session/create + sandbox:
Delegate single + replay
Delegate parallel + in-flight duplicate + replay
{"code":4034,"message":"task already running: task-live-parallel-r2"}
Dual-write parity / live task boardDuring the live single and parallel delegate runs I polled
This directly exercises the Step 7 cutover risk: while tasks were running, the registry child set and legacy Wart 1 / session namespace + ledgerOnly one registry ledger was written: It contains 5 completed records, all with
So the child records did not split into the agent persistent session namespace; they landed in the gateway session namespace.
task.* push eventsObserved task events:
Parent task events carried the submitted trace ids:
Caveats / wart candidates to triage, not fixed live
Proposed measurable gate definitionTo make “do not merge until Steps 5–6 are stable” actionable, I suggest using a concrete gate such as:
Verdict for this run: core Steps 5–6 registry/delegate/idempotency/session-namespace/parity behavior held under a real flag-on live session; keep #44235 gated pending more mileage and triage of the child trace-id contract. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #44233 (→ #43422 → #43039). Draft with an explicit merge gate: per the design doc's own rule, this cutover must NOT merge until Phase 5 Steps 5-6 have been stable in production — it is staged here so the chain is visible and CI-checked, not to jump the queue.
Phase 5 Step 7 — the registry becomes the single subagent store
tools/delegate_tool.py(_active_subagents+ lock,_spawn_paused+ lock,_register_subagent/_unregister_subagent— repo-wide grep confirms zero live references; no test patched them).AgentTaskRegistry:list_active_subagents()maps records (filtered by the newis_subagentdiscriminator) onto the byte-identical legacy dict shape — including the last_tool-only-after-first-tool quirk;interrupt_subagentkeeps the legacy "Interrupted via TUI" reason;set/is_spawn_pauseddelegate to the registry flag.is_subagentexists because intent can't discriminate:task.submitdelegate batch parents are alsointent="delegate".AIAgent._active_childreninterrupt propagation.Verification: 637+ green across delegate/gateway/registry/server suites; cutover-vs-baseline failure sets diffed via detached worktree — identical (zero regressions).
🤖 Generated with Claude Code