A delegated-task completion that queues behind another parent turn retains the model from queue time. If the thread's persisted model changes before the queue drains, delivery reverts that choice and starts the completion on the old model. Reproduced in the real Linux web dev app by delegating an async child, starting a second parent turn, and persisting Sonnet while the completion waited on Opus.
Delegated completions follow the thread's current selection and provider conversation through the existing monitor-update queue policy. User-authored queued messages retain their chosen selection. Focused regression cases cover model changes, provider-instance routing, current-provider usage limits and disabled-instance holds.
Fixes #22577 GitCafe: https://git.cafe/pingdotgg/t3code/issues/22577 Triaged bug via the automated issue flow: #22577, GitCafe mirror. One server queue-policy defect; no contract or client changes. Two isolated final reviewers returned MERGE; the final proof audit returned SUFFICIENT. This PR must not auto-land.
gpt-5.5 to gpt-5.4. The four final cases fail against the base implementation for model rollback, wrong-instance routing, wrong-provider usage-limit checks and disabled-instance execution. Original raw failure, four-case raw failures.vp test run for runtimeLayer, DelegatedCompletionDelivery, SteeringCompletion.integration, ClaudeAutomaticDelivery.integration, ProviderContinuationService, QueuedRunOrder, SelectionRestart.integration, ProviderSwitchService, ProjectionRecovery, RunCompletionReads, TurnStartReads and ThreadDeletion: 12 files, 175 tests passed. Output.tsc-rs --noEmit passed on the final test revision before merging main. Targeted test lint passed; orchestrator lint reports only an existing unused layerUnavailable warning. Formatting and diff whitespace checks passed.| Evidence | Before | After |
|---|---|---|
| Real app, Linux x64, collaborative Chrome 1280×800, isolated dev state; async child sleep 150s, second parent sleep 240s, persisted model change while completion queued | 70e5e384da: transcript; queue delivery writes Opus over Sonnet | d1736d5221: transcript, correlated native-model proof; the same completion starts and completes on Sonnet, and the thread remains on Sonnet |
The runtime-tested orchestrator is byte-identical at the final merged head; later changes add tests and merge upstream main. The picker alone changes the composer draft during an active turn, so the reproduction persisted the selection through the app's metadata command. Delivery timing remains unchanged. No render changes are claimed.
Not exercised: live cross-provider switching, native desktop/mobile, reconnects, or remote/relay/tunnel connections. Cross-instance routing and queue guards have deterministic server coverage; no platform-specific code changed. All reproduction servers were stopped.
Implementation and tests: Claude Opus 5.5 through Claude Code in T3 Code. Triage, runtime verification, evidence and proof: GPT-6.1-Sol through Codex in T3 Code; isolated reviews by GPT-6.1-Sol and Claude Opus 5.5.
Added 4 commits
Verdict: Needs Human Approval
Macroscope's review found this PR requires human approval — This is a focused server bug fix that keeps queued delegated completions aligned with the thread’s current model and provider, with targeted coverage for routing and queue guards. The PR explicitly requests that it not auto-land, so manual review is appropriate.
You can add or adjust custom eligibility rules. Learn more.
✅ 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 | +7 B (+0.1%) | 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 | 1.2 KiB | 1.2 KiB | +7 B (+0.6%) | 2.0 KiB | ✅ |
| Codex | Live turn WebSocket decoded | 20.9 KiB | 20.9 KiB | 0 B (0.0%) | 29.3 KiB | ✅ |
| Codex | Live turn messages | 2 | 2 | 0 (0.0%) | 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 | 21.2 KiB | 21.2 KiB | 0 B (0.0%) | 29.3 KiB | ✅ |
| Claude | Live turn messages | 2 | 2 | 0 (0.0%) | 8 | ✅ |
Baseline: 89fe509 · PR result: 6a8e8de · Source CI: success
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.
Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.
No actionable comments were generated in the recent review. 🎉
f16b1845-435c-4207-9eed-3742ab56fd93Reviewing files that changed from the base of the PR and between 89fe509232e787b51d0cf695861afe73316ce85b and 6a8e8de1eb474ae8d9c8542bbe3bd8e567eac1be.
apps/server/src/orchestration-v2/Orchestrator.tsapps/server/src/orchestration-v2/runtimeLayer.test.tsLimit details: You’ve used all 10 included reviews currently available.
Queued delegated completions now use the thread’s current model selection when they start. Runtime tests cover provider selection, usage-limit holds, and disabled provider instances.
Changes: Queued delegated completions
| Layer / File(s) | Summary |
|---|---|
Selection rule and queued completion validation apps/server/src/orchestration-v2/Orchestrator.ts, apps/server/src/orchestration-v2/runtimeLayer.test.ts | Delegated completions now follow the thread’s current model selection, as unbuffered monitor notifications do. Tests cover selected provider instances, provider-thread reuse when only the model changes, usage-limit holds, and disabled provider instances. |
Priority: ➖ Normal
Estimated code review effort: 2 (Simple) | ~10 minutes
Change: Bug fix
Possibly related issues
pingdotgg/t3code#21705 — The changes relate to queued selection behavior, but the reported cross-account hold release remains separate.Suggested reviewers: voltcrash
6a8e8Queued delegated completions follow the thread’s current model selection, while the inspected hold checks remain aligned with that selection. No issue identified here needs resolution before merge.
| Check name | Status | Explanation |
|---|---|---|
| Title check | Passed | The title clearly and concisely describes the main server fix: queued task completions now keep the current model. |
| Linked Issues check | Passed | Check skipped because no linked issues were found for this pull request. |
| Out of Scope Changes check | Passed | Check skipped because no linked issues were found for this pull request. |
| Description check | Passed | The pull request description includes all required sections: Problem, Change, Scope and approval, and Verification. It explains the defect, the implementation, issue and approval references, focused r… |
Explanation
The pull request description includes all required sections: Problem, Change, Scope and approval, and Verification. It explains the defect, the implementation, issue and approval references, focused regression coverage, test results, runtime evidence, limitations, and agent details.
Comment @coderabbitai help to get the list of available commands.
Closing in favour of #22579, which makes the same change.