test(sdk): assert correlated retry terminal outcome

[probepark]

Agent

omo/cliproxy/gpt-6.1-sol via paseo on home2 (worktree g4-retry-end-corr-r2)

Task

repo: Yeachan-Heo/gajae-code (base: dev). PR must target dev.

What

Retain the public lifecycle boundary (lifecycleScope) across accepted auto-retries whose predecessor agent_end was suppressed, and keep session-initiated (unowned) lifecycle batches separate from queued SDK turns. The final agent_end of an auto-retried turn is now attributed to the owning invocation's correlation, so the ACP prompt settles at turn end instead of being dropped as incomplete_correlation and expiring at the 340s watchdog.

Why

hermes-ops gjc-acp-guard G4 (watchdog_expired = 0) failed on 2026-10-04: work:a896777e and work:94f7adc2 both hit a mid-turn Codex transient error (server_is_overloaded / request_timeout), auto-retried, finished with stopReason=stop, then logged acp_prompt_terminal_dropped reason=incomplete_correlation followed by acp_prompt_watchdog_expired. Layer: SDK session host lifecycle (session-runtime). paseo, relay, and reap are not involved.

Testing

• bun test src/sdk/host/session-runtime.test.ts test/agent-session-retry-busy-recovery.test.ts (packages/coding-agent, home2, head 46539b5): 187 pass, 0 fail (2980 expect calls, 248s)
• prepush: bun run --workspaces --if-present check:types rc=0; telegram-daemon-generation-guard rc=0

Acceptance

AC	Where	Local test
AC-1 final agent_end of an auto-retried turn carries the owning commandId/turnId	session-runtime.ts emitLifecycle agent_start (lifecycleScopeOwners / continuationBatch); agent-session.ts #lifecycleScopesByAttemptScope; types.ts AgentStartEvent.lifecycleScope	session-runtime.test.ts (retryable prompt terminal correlation)
AC-2 ACP prompt settles at turn end, not at the watchdog	same path; prompt terminal published from the retained batch	session-runtime.test.ts "assert correlated retry terminal outcome"
AC-3 session-initiated runs do not consume a queued SDK turn's receipt	unownedStart batches retained in openLifecycleBatches filters	session-runtime.test.ts; agent-session-retry-busy-recovery.test.ts
AC-4 production G4 back to 0	NOT DONE: needs merge plus rollout via gjc-stale-binary-cron, then DRY=1 gjc-acp-guard.py re-measure	—

Needs e2e

None in CI scope. Post-rollout verification is the hermes gjc-acp-guard G4 re-measure.

Risk classification

• low-risk
• regression-risk
• high-risk

Merge approval is expected to stay red on this agent-authored PR until a maintainer approves the exact head (human-review path, #6037); the contract check itself should be green.

Goal

After a mid-turn provider error that gjc auto-retries, the turn's final agent_end reaches the ACP layer without a complete correlation (commandId/turnId). ACP drops it as incomplete_correlation, so the externally submitted prompt never settles and acp_prompt_watchdog_expired fires 340s later. Fix it so that the final agent_end of an auto-retried turn carries the owning invocation's correlation and outcome. The ACP prompt has to settle at that moment, not at the watchdog.

Evidence (work mac, gjc 0.18.6 stack-run build cb60f55, 2026-10-04)

• session a896777e: ACP prompt commandId 924c51f1 / turnId 91d8f200.
  13:57:42Z assistant message stopReason=error "Codex error event: Our servers are currently overloaded (code=server_is_overloaded)" (thinking only)
  13:58:08Z the next assistant message (auto-retry) continues normally with toolUse, and so on
  14:03:06Z the final assistant has stopReason=stop
  14:03:06.228Z log acp_prompt_terminal_dropped terminalType=agent_end reason=incomplete_correlation expectedCommandId=924c51f1 expectedTurnId=91d8f200
  14:08:46Z acp_prompt_watchdog_expired cause="the SDK session host stopped producing frames" silenceMs=340003 lastFrameType=message_end
  .sdk-reconciliation/a896777e.json still has the prompt record status=in_flight.
• session 94f7adc2: the same sequence. The error was stream disconnected before completion (code=request_timeout) at 14:12:37Z, t

Local CI gate (prepush)

PREPUSH_CMD rc=0 bun run --workspaces --if-present check:types
PREPUSH_CMD rc=0 bun scripts/telegram-daemon-generation-guard.ts "$(git merge-base origin/dev HEAD)" "$(git rev-parse HEAD)"

Agent log (tail)

✓ post-acceptance invocation terminalization > a queued prompt stays non-terminal even if isIdle flips during the accept window [1409.18ms]
✓ post-acceptance invocation terminalization > a pre-acceptance failure rejects the submission without creating a record [1457.18ms]
✓ post-acceptance invocation terminalization > a later provider error never changes or re-opens an already terminal prompt [1431.71ms]
✓ accepted-control zero-execution bound (#4668) > an accepted prompt that never reaches agent_start remains recoverable [1977.55ms]
✓ accepted-control zero-execution bound (#4668) > a running prompt deadline stops its exact turn and keeps normal terminals working [2513.89ms]
✓ accepted-control zero-execution bound (#4668) > prompt agent_end waits for durable reconciliation before publication [3326.27ms]
[Final output truncated to the last 2000 chars; the full history is still peekable.]
[Thought] Running full package check
[Eval]
[Thought] Planning manual ACP settlement check

Preparing isolated baseline test environment
[Eval]
[Thought] Planning symlinking native addons for packaging
[Eval]
[Thought] Adding temporary symlink for native addon
[Eval]
[Thought] Investigating test false positive causes

Modifying test for three agent starts
[Apply patch]
[Eval]
[Thought] Running baseline temperature test
[Eval]
[Thought] Planning retry logic for agent_end errors

Inspecting test for prompt correlation adaptation
[Eval]
[Thought] Planning test with session switch and prompt correlation

Evaluating session switch impact on test failures
[Eval]
[Thought] Verifying cherry-pick commit differences

Preparing commit amendment
[Eval]
[Thought] Planning targeted test rerun and final checks

Confirming completion of focused tests and checks
[Eval]

[snowykr]

Verdict

APPROVED

Summary

This change retains lifecycle identity across accepted automatic retries and keeps session-initiated lifecycle batches separate from queued SDK turns. No actionable merge-blocking defect was identified at the reviewed head.

Findings / Required Changes

No blocking or actionable findings.

Non-blocking Observations

• The opt-in WSLv2/NTFS qualification was skipped; it was not required for these changed paths.

CI / Verification

• Dev CI run 37216544183 succeeded for head 46539b5d4620c9026263c48dd5e77b241e49d5a7 (base d9a517893150bf9dc10cbb749f3034c732ceb1fb). The affected-path test tasks for both changed test files, the SDK-host shard, isolated production-host task, and relevant aggregate/final validation completed successfully.
• A separate run for the same head was skipped after metadata-only changes; the workflow classifies it as non-code evidence. Other observed skips were conditional or opt-in, not failed required checks.
• I did not execute tests. CI artifacts exposed successful task conclusions, but not test-runner output or independently verifiable per-test counts.

Axis Coverage

Axis	Verdict	Coverage
A1 — Intent / Policy / Contract	APPROVED	Lifecycle correlation behavior matches the inspected SDK contract and changelog; no concrete intent or policy mismatch.
A2 — Architecture / Correctness / Failure	APPROVED	Retry and unowned-batch ownership, terminalization and cleanup preserve correlation across inspected failure paths.
A3 — Security / Privacy / Trust	APPROVED	The new lifecycle identity is internally derived; no client-controlled authorization path or privilege boundary change found.
A4 — Verification / Tests / CI	APPROVED	Changed regression suites and relevant exact-head CI validation succeeded; see CI limitations above.
A5 — Context / Compatibility / Platform	APPROVED	Consumers remain on the in-process lifecycle path; no persisted/generated/platform contract changes or materially harmful duplicate abstraction found.

Limitations

The review was static; tests were not run by the reviewer. CI runner artifacts were inaccessible, so successful task conclusions—not individual test output or counts—are the available verification evidence.

[Yeachan-Heo]

Merged into dev. snowykr approved the exact head 46539b5; CI on that head 42 passed / 0 failed, including check:@gajae-code/coding-agent. The branch was 13 commits behind dev and none of them touch its files; GitHub reported a clean merge. Thanks @probepark!

—
[repo owner's gaebal-gajae (clawdbot) 🦞]