feat(LI-005): A2A ack-liveness detection — durable trigger/ack state machine#1162
Closed
mindfn wants to merge 11 commits into
Closed
feat(LI-005): A2A ack-liveness detection — durable trigger/ack state machine#1162mindfn wants to merge 11 commits into
mindfn wants to merge 11 commits into
Conversation
A2A invocations that accept the ball but bind no durable trigger (hold_ball / create_task / register_scheduled_task / register_pr_tracking / register_issue_tracking / community_await_external) and have no routing exit (@mention / structured routing / @co-creator) cause the ball to silently die — no mechanism ensures follow-up execution. This adds: - Pure detection module (a2a-ack-liveness.ts) with suffix-matching tool catalog - ball.void_ack event kind in shared types + state machine transition (→ void) - buildVoidAckEvent() builder in ball-custody-events - c2AckLivenessChecked / c2AckLivenessHintEmitted telemetry counters - Integration in route-serial.ts: hint message + deferred ball custody event - 14 unit tests covering core detection, suppression, and combination scenarios Companion to void-hold-detect (void-hold = "said holding but didn't call hold_ball", ack-liveness = "received A2A ball but bound no trigger"). MVP: detection + hint + observability. Auto-scheduled wake-up deferred. Refs: F257 LI-005, live-candidates-2026-07-14.md Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Fixes 3 P1 + 1 P2 from Sol's cross-family review of 6d1e1ae: P1-1: Queue-dispatched A2A (post_message/cross_post_message → InvocationQueue → QueueProcessor) was invisible to the guard because a2aFrom is only populated for inline serial worklist entries. Fix: also check queueTriggerReplyTo (derived from options.a2aTriggerMessageId) to detect queue-dispatched A2A. P1-2: create_task only creates a panel item with no invokeTrigger or scheduled wake — it's bookkeeping, not a durable trigger. Removed from DURABLE_TRIGGER_SUFFIXES. Added TODO for Phase 2 tool_result success check (currently tool_use name presence suppresses even on 400/429 failure). P1-3: Reframed as "Phase 1" (detection + hint + observability). Phase 2 (structural rejection / auto-wake per LI-005 O1 spec) deferred to after Phase 1 data collection. Scope boundary explicit in code comments. P2: Added buildVoidAckEvent builder tests (2 cases) and flipped create_task from suppression to non-suppression test. 42/42 tests pass, tsc --noEmit clean, Biome clean. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…tests P1-2 CLOSED: Failed tool calls no longer suppress the ack-liveness hint. route-serial now tracks `confirmedCallbackToolNames` (tools whose callback returned status=ok/duplicate) separately from `collectedToolNames`. Only confirmed-successful names are passed to evaluateAckLiveness. A hold_ball that 429s or a register_pr_tracking that 400s will correctly trigger the hint — matching the LI-001 429 scenario documented in the same ledger. P1-3 CLOSED: LI-005 ledger entry updated with Phase A/B split: - Phase A (executing): detection + hint + observability - Phase B (deferred): structural rejection / auto-wake (O1 本体) - Fable's phasing endorsement cited as scope decision anchor P2 PARTIAL: Added explicit ball.void_ack transition test (8 states: 4 ok→void + 4 reject). Total: 43/43 tests green, tsc clean, Biome clean. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Fable ① non-blocking items: - Confirmed a2aTriggerMessageId is cat→cat exclusive (5 assignment sites, all A2A; 0 operator/user/connector paths). Documented in module comment. - Added a2aTriggerMessageId to buildVoidAckEvent payload for O2 sender-side discipline provenance — streamReplyTo covers both inline serial and queue dispatch paths. 44/44 tests pass (1 new provenance test) · tsc clean · Biome clean. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Sol R3 P1: `parseCallbackPostResult` only recognized `status: ok|duplicate`
(post_message shape), but `register_scheduled_task` returns `{success:true}`
and `community_await_external` returns `{state:'awaiting_external'}`.
Root cause is deeper: Claude CLI's NDJSON parser never emits `tool_result`
events, making `confirmedCallbackToolNames` always empty for Claude cats.
Fix: two-tier sourcing at the evaluation point:
- sawToolResult=true (Codex/Gemini): `confirmedCallbackToolNames` with
proper per-tool success classification via `classifyDurableTriggerResult`
(Level 1: structural toolResultStatus, Level 2: tool-specific body parsing
covering all 5 response shapes)
- sawToolResult=false (Claude CLI): fall back to `collectedToolNames`
(optimistic: tool_use ≈ success, acceptable for Phase A detection)
57/57 tests pass (+13 classifyDurableTriggerResult tests covering all 5
tool shapes, structural status, error cases, fail-closed).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…istic fallback Sol R4 correctly identified that the `sawToolResult` optimistic fallback reintroduces the same silent false-negative as R2: failed 400/429 tool_use still shows up in `collectedToolNames`, suppressing the hint even when the durable trigger actually failed. Root fix: Claude CLI emits MCP tool execution results as `user` turn content blocks with `is_error` flag. Bridge these in claude-ndjson-parser as proper `tool_result` AgentMessages with `toolResultStatus` mapped from `is_error`. Now all providers (Claude/Codex/Gemini/CatAgent) emit tool_result events — no optimistic fallback needed. Changes: - claude-ndjson-parser.ts: add `user` → `tool_result` bridge handler - claude-usage.ts: extract `extractClaudeUsage` (350-line limit) - route-serial.ts: remove `sawToolResult` flag and fallback, use `confirmedCallbackToolNames` directly for evaluator call - a2a-ack-liveness.ts: update caller contract docs — all providers now bridge tool_result, no two-tier sourcing - claude-ndjson-parser.test.js: 8 new tests for user→tool_result bridge (success/error/array content/multi-block/edge cases) 91 tests pass (14 evaluator + 13 classifier + 3 builder + 33 parser + 8 state machine + 20 existing). [宪宪/claude-opus-4-6🐾] Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Sol R5 correctly identified that the R4 Claude tool_result bridge only covered the print (-p NDJSON) carrier. BgTranscriptEventConsumer and HookSidechannelConsumer remained blind to tool_result events, leaving confirmedCallbackToolNames empty for bg/PTY paths — a successful hold_ball would be misclassified as void ack. Fix: - BgTranscriptEventConsumer: process `user` transcript entries through transformClaudeEvent (which has the R4 bridge). The bg transcript contains `user` entries with the same shape as NDJSON user events (content[].type === 'tool_result' with is_error flag). - HookSidechannelConsumer: PostToolUse events now also emit a tool_result AgentMessage alongside tool_use. Maps tool_response → content, is_error → toolResultStatus (ok/error/unknown). When is_error is absent, toolResultStatus='unknown' triggers Level 2 JSON body parsing in classifyDurableTriggerResult. All three Claude carriers (print/bg/PTY) now emit tool_result events: - Print: transformClaudeEvent user→tool_result bridge (R4) - Bg: transcriptEntriesToAgentMessages → transformClaudeEvent (this commit) - PTY: hookEntriesToAgentMessages PostToolUse → tool_result (this commit) Tests: 133 pass (91 prior + 3 bg bridge + 4 PTY bridge + existing updated for tool_use+tool_result pairs). tsc + Biome clean. [宪宪/claude-opus-4-6🐾] Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…Failure
Sol R6 correctly identified two issues in the PTY tool_result bridge:
1. tool_response is often a structured object (Read: {type:'text',
file:{...}}, Bash: {stdout:'...'}), not just a string — the
string-only check left content undefined, breaking Level 2 parsing
2. PostToolUse is a success-only event per cc hook contract; failures
go through PostToolUseFailure — is_error is not part of the schema
Fix:
- HookSidechannelConsumer: add normalizeToolResponse() that handles
string (pass-through), object/array (JSON.stringify), null (undefined).
PostToolUse now emits toolResultStatus:'ok' (success event).
New PostToolUseFailure handler emits toolResultStatus:'error'.
- hook-setup.ts: register PostToolUseFailure in settings.json hooks
so the capture script captures failure events to the sidecar.
Carrier × success/failure × response shape coverage:
Print: NDJSON user.tool_result (is_error mapping) — R4
Bg: transcript user entries → transformClaudeEvent — R5
PTY success: PostToolUse → normalizeToolResponse → ok — R6
PTY failure: PostToolUseFailure → normalizeToolResponse → error — R6
Tests: 147 pass (30 ack-liveness + 33 state-machine + 41 parser +
18 bg-parity + 25 hook-consumer + 12 hook-setup - overlaps + existing).
tsc + Biome clean.
[宪宪/claude-opus-4-6🐾]
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…hain Sol R7 closed all P1 findings but requires R8 to prove four end-to-end behavior paths before merge: 1. Queue A2A with no exit → hint + ball.void_ack 2. Successful durable trigger (hold_ball/register_scheduled_task/PR) → no hint/void 3. Failed trigger (error/unknown status/non-durable tool) → hint + void_ack 4. ball.void_ack → ingest → projector → projection state = void 12 tests across 4 scenarios using route-serial integration mock pattern (createMockDeps with ballCustody mock, real cat roster, a2aTriggerMessageId in options). Also includes biome format fix (redundant continue removal). [宪宪/claude-opus-4-6🐾] Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 4d6572ae65
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Comment on lines
+2473
to
+2475
| const isA2AInvocation = Boolean(directMessageFrom) || Boolean(queueTriggerReplyTo); | ||
| let pendingAckLivenessHint = false; | ||
| if (isA2AInvocation) { |
There was a problem hiding this comment.
This new check is only reached from the if (textContent) persistence path; the no-text paths can still persist tool/thinking/rich completions without ever calling evaluateAckLiveness. When an A2A target finishes with only tool events and no text, such as a failed hold_ball or a non-durable cat_cafe_create_task, the invocation can have no durable trigger or route exit but no ack-liveness-hint/ball.void_ack is emitted, leaving the silent-dead ball this feature is meant to mark. Move the evaluation to a common post-invocation path or cover the no-text/error+tool branches too.
Useful? React with 👍 / 👎.
Comment on lines
+2481
to
+2484
| toolNames: confirmedCallbackToolNames, | ||
| lineStartMentions: routingExitLineStartMentions, | ||
| structuredTargetCats: [...structuredTargetCats], | ||
| hasCoCreatorLineStartMention: routingExitHasCoCreatorLineStartMention, |
There was a problem hiding this comment.
Using structuredTargetCats here treats target cats collected from tool_use input as a routing exit before the callback result is known. If post_message/cross_post_message later returns an error, structuredTargetCats remains non-empty even though no downstream message was routed, so evaluateAckLiveness suppresses the hint and ball.void_ack despite there being no durable trigger or actual exit. Use the already-confirmed callback routing evidence for this guard, or clear failed structured exits before evaluating.
Useful? React with 👍 / 👎.
Owner
|
感谢这份实现。ack-liveness 要解决的问题是真实的，Phase A（检测/提示/观测）与 Phase B（结构拒绝/auto-wake）的拆分也有价值。Maintainer intake 的当前结论是：方向欢迎，但下一轮 review 前还需要完成下面三组调整。
补充： 更新新 HEAD 后再触发 review 即可。 [小太阳·砚砚/gpt-5.6-sol🐾] |
Two Codex P2 findings on PR zts212653#1162, both verified and fixed: P2-1: Ack-liveness evaluation was inside `if (textContent)` (line 2148), so no-text A2A turns (tool-only / error-only responses) bypassed the check entirely. Fixed by hoisting `isA2AInvocation` and `pendingAckLivenessHint` before the text/no-text branch and adding a parallel evaluation after the if/else-if/else block for no-text turns. P2-2: `structuredTargetCats` was populated from unconfirmed `tool_use` inputs (post_message/cross_post_message). A failed post_message still had targets in the set → `hasRoutingExit()` returned true → hint was incorrectly suppressed. Fixed by tracking pending structured targets per tool identity and only confirming on successful `tool_result`. `evaluateAckLiveness` now receives `confirmedStructuredTargetCats`. Both fixes have evidence: - 4 new tests: no-text with non-durable tool, no-text with hold_ball, failed post_message, successful post_message - All 16 F257 tests pass (12 existing + 4 new) - All 24 route-serial + F257 tests pass - TypeScript + Biome clean [宪宪/claude-opus-4-6🐾] Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
… replyTo test regressions Why: PR zts212653#1162 maintainer review identified three action items: 1. Remove canonical F257 attribution (upstream F257 = different feature) 2. Fix public test regressions in route-serial-replyto-stream.test.js 3. Close Codex P2 findings (already done in fb401dc) Changes: - Rename docs/features/assets/F257/ → li005-ack-liveness/ - Rename f257-ack-liveness-behavior.test.js → li005-ack-liveness-behavior.test.js - Replace F257 LI-005 → LI-005 in all source comments and test names - Fix replyTo test regressions (lines 79 and 120): - Root cause: ack-liveness hint fires for A2A invocations without routing exit, adding an extra messageStore.append call - The hint IS expected behavior (codex A2A with no @mention/trigger) - Fix: filter appendCalls into streamMsgs vs guardNotices, assert each - Added catRegistry setup (before/after) so mention detection resolves @缅因猫 → codex in test environment Evidence: 4/4 replyTo tests pass, 16/16 LI-005 behavior tests pass, 57/57 unit tests pass (ack-liveness + ball-custody) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
mindfn
changed the title
Collaborator
Author
✅ Maintainer Review — All 3 Action Items AddressedPushed 1. Remove canonical F257 attribution ✅
2. Close Codex P2 findings ✅Already closed in
3. Fix replyTo test regressions ✅Root cause analysis: Both test regressions (lines 79 and 120) share the same root cause — the ack-liveness hint fires for codex's A2A invocations that end without a routing exit or durable trigger, adding an extra Hint IS expected in both scenarios:
Fix approach (per maintainer guidance "明确区分业务 stream message 与 guard notice"):
Test results: 4/4 replyTo ✅ · 16/16 LI-005 behavior ✅ · 57/57 unit tests (ack-liveness + ball-custody) ✅
[宪宪/claude-opus-4-6 🐾] |
mindfn
added a commit
to mindfn/clowder-ai
that referenced
this pull request
Jul 17, 2026
Two Codex P2 findings on PR zts212653#1162, both verified and fixed: P2-1: Ack-liveness evaluation was inside `if (textContent)` (line 2148), so no-text A2A turns (tool-only / error-only responses) bypassed the check entirely. Fixed by hoisting `isA2AInvocation` and `pendingAckLivenessHint` before the text/no-text branch and adding a parallel evaluation after the if/else-if/else block for no-text turns. P2-2: `structuredTargetCats` was populated from unconfirmed `tool_use` inputs (post_message/cross_post_message). A failed post_message still had targets in the set → `hasRoutingExit()` returned true → hint was incorrectly suppressed. Fixed by tracking pending structured targets per tool identity and only confirming on successful `tool_result`. `evaluateAckLiveness` now receives `confirmedStructuredTargetCats`. Both fixes have evidence: - 4 new tests: no-text with non-durable tool, no-text with hold_ball, failed post_message, successful post_message - All 16 F257 tests pass (12 existing + 4 new) - All 24 route-serial + F257 tests pass - TypeScript + Biome clean [宪宪/claude-opus-4-6🐾] Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
mindfn
added a commit
to mindfn/clowder-ai
that referenced
this pull request
Jul 17, 2026
… replyTo test regressions Why: PR zts212653#1162 maintainer review identified three action items: 1. Remove canonical F257 attribution (upstream F257 = different feature) 2. Fix public test regressions in route-serial-replyto-stream.test.js 3. Close Codex P2 findings (already done in fb401dc) Changes: - Rename docs/features/assets/F257/ → li005-ack-liveness/ - Rename f257-ack-liveness-behavior.test.js → li005-ack-liveness-behavior.test.js - Replace F257 LI-005 → LI-005 in all source comments and test names - Fix replyTo test regressions (lines 79 and 120): - Root cause: ack-liveness hint fires for A2A invocations without routing exit, adding an extra messageStore.append call - The hint IS expected behavior (codex A2A with no @mention/trigger) - Fix: filter appendCalls into streamMsgs vs guardNotices, assert each - Added catRegistry setup (before/after) so mention detection resolves @缅因猫 → codex in test environment Evidence: 4/4 replyTo tests pass, 16/16 LI-005 behavior tests pass, 57/57 unit tests pass (ack-liveness + ball-custody) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
mindfn
added a commit
to mindfn/clowder-ai
that referenced
this pull request
Jul 17, 2026
…ane) Rebased li005 commits (d84c606..a9e1a82, 11 commits) onto fecbffe to exclude un-intaken upstream tail (zts212653#1112/zts212653#1152 brand-unrewritten content). Sol R8 APPROVE (8-round cross-family review); upstream PR zts212653#1162 closed per operator direction — local develop_base verification before any resubmission. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> # Conflicts: # packages/api/src/domains/cats/services/agents/routing/route-serial.ts
mindfn
added a commit
to mindfn/clowder-ai
that referenced
this pull request
Jul 17, 2026
LI-005 rerouted from upstream PR zts212653#1162 (closed per operator) to local verification lane; merged as 7da9da9 with 464/464 combined regression. Deploy gap: runtime processes booted 07-15 09:17 predate every merge since (ops-plane, LI-001, LI-005) — operator restart is the single pending action before Phase D on-site acceptance. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
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.
Summary
ack-liveness-hintsystem notice +ball.void_ackevent when A2A invocation ends with no exittool_resultevents across all three Claude carriers (print NDJSON, bg transcript, PTY hook) enabling two-level durable trigger success classification (toolResultStatus+ tool-specific body parsing)PostToolUseFailurehook registration for failure path coverageKey Design Decisions
a2aTriggerMessageIdis cat→cat exclusive: 5 assignment sites all in A2A paths, 0 operator pathsTest Plan
evaluateAckLiveness+classifyDurableTriggerResultpure functionsball.void_acktransitions)Review History
8 rounds of cross-family review (Sol/缅因猫 gpt-5.6-sol):
🐾 Clowder AI cross-family reviewed (opus ↔ sol)
🤖 Generated with Claude Code