fix(server): repeated capture timeouts no longer delay turn starts - #29196

Approvability

Verdict: Approved at b7208c0

Macroscope's review found this PR approvable — Repeated capture timeouts now suppress baseline retries for that workspace while post-turn captures continue checking for recovery. Healthy captures remain unchanged, and successful recovery clears the temporary backoff.

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

No actionable comments were generated in the recent review.

Merge Risk: Minimal · up to ea1e0

No identified issue needs resolution before merge; normal checks remain appropriate.

Pre-merge checks: 4

Check name	Status	Explanation
Title check		The title clearly summarizes the timeout backoff change and its effect on turn starts.
Description check		The description is mostly complete. It explains the problem, fix, trade-off, and verification results. It does not use the required section headings or establish whether the linked issue was triaged o…
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 is mostly complete. It explains the problem, fix, trade-off, and verification results. It does not use the required section headings or establish whether the linked issue was triaged or the scope was approved.

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

Usage-based review receipt

• Mode: Continue automatically
• Reviewed files: 2
• Charged: $0.50
• View usage details

Recent review info

Run configuration

• Source mode: Auto
• Repository: pingdotgg/t3code
• Branch: main
• File: .coderabbit.config.ts
• Commit: 83a3674
• Layers: Repository TS config
• Review profile: CHILL
• Plan: Advanced
• Run ID: 0d0a27ce-f24f-4f64-97e1-00b78d88ca57

Commits

Reviewing files that changed from the base of the PR and between b7208c0 and ea1e0ed.

Files selected for processing: 2

• apps/server/src/orchestration-v2/CheckpointService.test.ts
• apps/server/src/orchestration-v2/CheckpointService.ts

Limit details: You’ve used all 10 included reviews currently available.

Verified in a real dev environment

verified as-is at b7208c06beb1e4306f827fad807dae83303d89b8 on this linux boat vm. installed and ran github/main (30084f0018) in a separate worktree first, then the exact pr head. only one background dev server ran at a time. drove the actual web client with repo playwright and chrome, using real codex gpt-6.1-sol turns through the local proxy.

codex created a real 3 gib untracked binary in the app project. actual git staging exceeded the existing 30-second budget; no mocks, replay, or injected errors were used.

• main reproduced repeated baseline delays: startRootRun took 30,215.8 ms and 30,209.3 ms. codex started about 30,173 ms into each run.
• the pr initial baseline and post-turn captures still timed out. after those two failures, the next baseline was skipped: startRootRun took 53.8 ms, with codex starting at +32.9 ms and returning its answer in the client.
• the next post-turn capture still ran and timed out, proving recovery probes remain active.
• codex excluded the large binary through .git/info/exclude. post-turn capture then succeeded in 82.0 ms. a fresh thread captured a new baseline successfully in 80.6 ms, proving recovery cleared the backoff.
• healthy file changes produced the expected client diff. in an app-created isolated worktree, two codex turns changed note.txt; edit from here > revert files too restored the first snapshot and rewound the conversation. the shared-checkout restore guard also refused restoration and preserved the file.

63 focused checkpoint/capture/run-execution tests passed on blacksmith. vp run check:changed 47f2c8d29e981c7723467efe137ee19e91df924e passed format, lint, knip, graphql, affected typechecks, and adjacent tests after supplying missing worker git history. both independent scoped code reviews found no blockers. no code changes were needed.

main recording, elapsed time visible:

pr recording, same workspace and prompt after two timeouts:

runtime command transcript and measured spans. both recordings were inspected, played through, and checked for seeking after upload. they capture the web interaction at one frame per second; timing numbers above come from the real server spans.

Not verified

native electron, ios/android, windows/macos, the exact reporter workspace, other providers, and the original follow-up-steer symptom. this pr preserves the initial and post-turn timeout budgets; it does not eliminate those waits. the linux web/codex path, recovery, and neighboring checkpoint behavior were exercised directly.

Exact-head hosted checks and CodeRabbit review are green at ea1e0ed518d8f14d8437c080c7baaf33757c43dd, with no actionable code findings. The remaining description observation is accurate: the live body has prose plus Verification: rather than the repository's section headings.

@voltcrash A description-only adjustment can preserve all current content: put the repeated-delay explanation under ## Problem, the two-timeout skip/recovery policy and trade-off under ## Change, add ## Scope and approval, and change Verification: to ## Verification. A factual scope paragraph is:

Addresses #28996 through a focused correction to repeated baseline retries after demonstrated checkpoint timeouts. Initial/post-turn budgets, capture ordering, and restore validation remain unchanged. No maintainer direction approval is claimed; maintainers determine eligibility under the focused obvious-bug exception.

Keep the current model/harness attribution and evidence limits. No implementation or additional test change is requested by this description observation.