[juliusmarminge]

Problem

Switching a delegated task's wake policy to always (for example after a task_status wait times out) plans a completion delivery from a snapshot of the parent run. If the parent's checkpoint capture commits completed between that read and the command's commit, the stale snapshot is written back: the run returns to running, loses completedAt and checkpointId, and queued work behind it stops advancing.

Change

delegated_task.wake-policy now hands its parent run id to the same updateDelegatedCompletionOnlyRunId path that acknowledge and dispose already use on main (32bf48f, #25672). EventSink.commitCommand rebuilds that run.updated from the stored run inside the transaction and applies only the planned delegatedCompletion. Every parent run the delivery planner returns differs from its input only in delegatedCompletion, so nothing else is dropped.

Refs #25757: the wake-policy path of the same stale-write race. https://git.cafe/pingdotgg/t3code/pulls/25667 fixes the accept path that stalled that thread.

Found while triaging #24322; it does not fix that issue's terminal negotiation timeout.

Takeover

Taken over from @juliusmarminge's automated triage run. While this PR waited, main landed the acknowledge/dispose half of the same fix in #25672, with a near-identical EventSink change. The merge with main resolves to main's EventSink, so this PR now adds only the wake-policy half, which the original covered without a test. The PR's acknowledge regression test passed on main unchanged and was dropped as a duplicate. Main's parametrised race test gains a wake-completed case.

Evidence

Server-only change; no client, contract or render path changes. The proof audit's capture matrix is empty, so the evidence is the regression test and a runtime transcript.

Case	main 3d369d5 Orchestrator.ts	head 38bed4a
acknowledge, parent fails mid-command	pass	pass
acknowledge, checkpoint completes mid-command	pass	pass
accept, checkpoint completes mid-command	pass	pass
wake-policy → always, checkpoint completes mid-command	fail: run back to running, completedAt and checkpointId null	pass: completed, completedAt and checkpointId kept, planned delegatedCompletion applied

Transcripts at 38bed4a: head.txt, control.txt (head tests with main's Orchestrator.ts). Proof audit: fixture matrix, runtime bundle.

Environment: fs-11 Linux VM, vp test run in apps/server, real orchestration and EventSink persistence, provider mocked. What else changed: nothing outside Orchestrator.ts (+5/−1) and DelegatedCompletionDelivery.test.ts.

Verification log

• Round 2 merged main 3d369d5, which reshaped the race test table in fix(server): preserve completed runs when accepting child deliveries #25666 (new accept row, delivery message). The wake-policy row is carried into main's table; Orchestrator.ts merged without conflict and is unchanged from the PR.
• Tests: DelegatedCompletionDelivery.test.ts 35 pass at 38bed4a; with main's Orchestrator.ts only the wake-policy row fails.
• vp run check:changed on 38bed4a: format, lint, knip 1/2 and 2/2, graphql, typecheck (3 packages), tests (4 files) pass.
• Review round 1 on 5abcaf3: reviewer A (GPT-6.1 Sol, Codex) MERGE; reviewer B (Claude Opus 5.5, Claude Code) MERGE.
• Review round 2 on 38bed4a: reviewer A (GPT-6.1 Sol, Codex) MERGE; reviewer B (Claude Opus 5.5, Claude Code) MERGE, head and control rerun independently. No confirmed findings.
• Proof audit pass 1 (GPT-6.1 Sol, Codex): SUFFICIENT, empty capture set; code unchanged since.
• Not reproduced: the terminal negotiation timeout from the issue, and live queue advancement behind a long-running provider.
• Unverified surfaces: none.
• Model and harness: Claude Opus 5.5 through Claude Code (lead, tests, implementation, review B); GPT-6.1 Sol through Codex (review A, proof audit).

Original description by @juliusmarminge

Problem

A child-result acknowledgement can read a waiting parent, then commit that stale snapshot after checkpoint capture has completed the run. This puts the completed parent back into waiting and prevents queued work from advancing. The maintainer observed a checkpointed parent stuck for over nine hours with more than 110 queued runs.

Change

Rebuild delegated-completion-only run updates from the stored lifecycle inside the command transaction. Acknowledgement, disposal and wake-policy updates retain their delegated-completion changes while preserving a checkpoint completion committed after their planning read.

Scope and approval

Originally linked to #24322

GitCafe: https://git.cafe/pingdotgg/t3code/issues/24322

Triaged bug via the automated issue flow; maintainer requested the retained checkpoint-race fix. This fixes the reproduced stale-acknowledgement race. The reported terminal negotiation timeout itself remains unverified; this change does not establish that all provider backlog failures share this cause.

Verification

• Retained red/green proof: the regression test fails before the implementation and passes with it; both isolated reviewers returned MERGE and proof audit returned SUFFICIENT in the prior run.
• Rechecked head e5dcae1e4e21a790d36dc5a4f54613b3df088d40, merged with main e2d3ded436f7: vp test run apps/server/src/orchestration-v2/DelegatedCompletionDelivery.test.ts — 29 passed.
• vp run check:changed — format, lint, both knip checks, GraphQL, affected typechecks and 42 tests in three files passed.
• Focused runtime sequence in the regression test: seed waiting parent with delivered child result → plan acknowledgement → commit checkpoint completion → commit acknowledgement → read projection. Result: status remains completed, completedAt and checkpointId survive, delivery is cleared.
• Server-only change; capture matrix is empty. No client render paths or contracts changed. Terminal negotiation timeout and live long-running provider backlog were not reproduced in this resumed run.

Model: GPT-6.1 Sol through the Codex harness (resumption and verification); retained implementation from the earlier automated triage run.

🤖 Generated with Claude Code

[macroscopeapp]

Approvability

Verdict: Approved at 38bed4a

Macroscope's review found this PR approvable — Wake-policy updates now use the existing transactional merge that preserves the parent run’s checkpoint and lifecycle state. The change is narrowly scoped and leaves wake-policy defaults unchanged.

^{You can add or adjust custom eligibility rules. Learn more.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	5.0 KiB	5.0 KiB	−35 B (−0.7%)	6.8 KiB	✅
Codex	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Codex	Live turn WebSocket wire	1.3 KiB	1.2 KiB	−35 B (−2.7%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	21.6 KiB	21.5 KiB	−41 B (−0.2%)	24.4 KiB	✅
Codex	Live turn messages	2	1	−1 (−50.0%)	8	✅
Claude	Total thread wire	5.0 KiB	5.0 KiB	0 B (0.0%)	6.8 KiB	✅
Claude	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Claude	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Claude	Live turn WebSocket decoded	21.9 KiB	21.9 KiB	0 B (0.0%)	24.4 KiB	✅
Claude	Live turn messages	1	1	0 (0.0%)	8	✅

Baseline: 3d369d5 · PR result: 38bed4a · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 108.5 KiB
• Claude decoded thread snapshot: 108.8 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[coderabbitai]

No actionable comments were generated in the recent review.

Walkthrough

The wake-policy dispatcher now returns the affected parent run ID, and dispatchOnce records it for delegated-completion handling. The settlement-race test covers wake-policy, acknowledgement, and accept commands, and checks the resulting delivery state and settled run fields.

Changes: Delegated completion updates

Layer / File(s)	Summary
Identify the parent run for commit scope apps/server/src/orchestration-v2/Orchestrator.ts	The wake-policy handler returns the parent run ID, and dispatchOnce stores it in updateDelegatedCompletionOnlyRunId.
Test updates during run settlement apps/server/src/orchestration-v2/DelegatedCompletionDelivery.test.ts	The race test covers wake-policy, acknowledgement, and accept commands. It checks command-specific delivery states and verifies that settled run fields and planned delegated-completion state persist.

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Bug fix

Suggested reviewers: t3dotgg

Merge Risk: Minimal · up to 38bed

The wake-policy change preserves settled run state, with no merge-blocking issue identified.

Pre-merge checks: 4

Check name	Status	Explanation
Title check		The title clearly identifies the server-side wake-policy fix and its effect on checkpointed parent runs. It is specific, concise, and follows the repository's conventional commit style.
Description check		The description covers the problem, change, scope, approval context, verification steps, test results, limitations, and changed files. It also distinguishes the reproduced race from the unverified ter…
Linked Issues check		Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check		Check skipped because no linked issues were found for this pull request.

Full details: Description check

Explanation

The description covers the problem, change, scope, approval context, verification steps, test results, limitations, and changed files. It also distinguishes the reproduced race from the unverified terminal negotiation timeout.

Finishing Touches

Generate docstrings

• Commit to this branch
• Create a new PR

Generate unit tests (beta)

• Commit to this branch
• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

CodeRabbit Verify

• Run CodeRabbit Verify to test the changes in this pull request.

Recent review info

Run configuration

• Source mode: Auto
• Repository: pingdotgg/t3code
• Branch: fix/issue-24322
• File: .coderabbit.config.ts
• Commit: 38bed4a
• Layers: Repository TS config
• Review profile: CHILL
• Plan: Team
• Run ID: 52b9bb9d-8ca1-4030-b15a-645e77cca0aa

Commits

Reviewing files that changed from the base of the PR and between ceacd38 and 38bed4a.

Files selected for processing: 2

• apps/server/src/orchestration-v2/DelegatedCompletionDelivery.test.ts
• apps/server/src/orchestration-v2/Orchestrator.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 7 remain after this review.

[juliusmarminge]

The linked-issue and scope warnings were checked against the maintainer’s updated triage instructions. Those instructions explicitly identify the checkpointed parent stuck in waiting with 110+ queued runs, authorize a focused test of that event order when the live repro is unavailable, and request resuming this retained fix. This PR covers that reproduced checkpoint/acknowledgement race. It does not establish the cause of every terminal negotiation timeout; the body explicitly records that limitation. No handshake or provider-stdio change is justified by the reproduced race, so expanding this patch would exceed the requested fix. All 28 head checks pass; Macroscope approved and CodeRabbit reported no actionable code findings.

[juliusmarminge]

Taking over this one

Dismissing prior approval to re-evaluate ceacd38

[juliusmarminge]

Taking over this one

Dismissing prior approval to re-evaluate 38bed4a