Repository navigation
Conversation
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
First Principles Review (Fable 5, fork) — 🔴 BLOCKPremise-level review of First-Principles-Verdict: BLOCK Retiming a Go-All stage from a 30-minute stop to a 4-hour ceiling is a product-shape change; base pins say "stops," and no RFC covers it. Not justified as shipped
What this change shipsInventory (10 items) — 5 justifiedIntent: stop Autopilot Go-All from cutting productive stages at 30 minutes, and make a dashboard timeout actually stop the native agent turn — declared a FIX; the timeout-semantics half is a shape change.
Blockersproduct-shape change without accepted RFC — the Autopilot loop's default stage stop is replaced. The description claims "the bundled Autopilot contract says it gates whether another turn may start and an admitted stage may outlast it," but the base carries the opposite as pinned decisions this diff deletes: the config table row "Wall-clock budget per stage before auto-run stops" (autopilot.md:501), the tracker docstring naming "the orchestrator's [FIRST-PRINCIPLES-REVIEWED] afbb210 |
GPT 5.6 Review (fork) — 🔴 changes requested (blocking)Reviewed 1 of 1 blocking finding(s) are security-class and were withheld from adjudication, so the blocking verdict stands. BLOCKING -- src/kiro_crew/dashboard/chat_runner.py:17326 -- Native cancellation occurs after the ACP stream marks the turn inactive [BLOCK-MERGE] afbb210 Adjudication (Opus 5) — is blocking on each finding proportionate?F1 — harm: UNBOUNDED-class residual per the fence anchor (silent loss of a turn's work while a native tool keeps running). [ADJUDICATION] afbb210 total=0 uphold=0 downgrade=0 |
Design Review (Fable 5, fork) — 🔴 BLOCK (blocking)Design-level review of Design-Verdict: BLOCK Redefines BlockersRewrites the pins for the opposite behavior and treats them as a gap, not a decision Suggestions
[DESIGN-REVIEWED] afbb210 |
Opus 5 Review (fork) — ✅ no blocking findingsReviewed Review detailsNothing blocks; the new native-ACP cancel handoff is inert on the exact path it was added for. FINDING — src/kiro_crew/dashboard/chat_runner.py:17321 — when FINDING — src/kiro_crew/dashboard/chat_orchestrator.py:1476 — the start-gate's SEL record still writes [OPUS-REVIEWED] afbb210 |
Treat the 30-minute stage budget as an admission gate between automatic turns instead of cancelling productive work mid-stage. Keep the independent chat-turn ceiling, cancel only a native turn whose lease this runner owns, surface unacknowledged cancellation, and merge current upstream main.
|
Status: returned to draft pending RFC #14043. The RFC records six distinct 30-minute timeout events and two confirmed dashboard/native split-brain continuations. Current-head review also proved the prototype's native cancel runs too late: ACP stream unwind can clear active-turn state before cancel_current, causing no_turn without sending session/cancel. Next revision will split the urgent pre-unwind, lease-owned cancellation fix from the admission-only stage-budget change. The latter remains blocked until the RFC is accepted on main. |
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
3 similar comments
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
2 similar comments
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
4 similar comments
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
5 similar comments
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
1 similar comment
|
👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review. When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically. |
|
PR triage (automated, 2026-10-05) — stale against main + overlaps #16301 Checked against
Recommendation: this PR is not being closed, because its cancel goal is still open on main. Please carry the native-cancel work only in #16301, then close this draft. If you want to keep it, rebase it onto main and drop the Autopilot/stage-budget changes. |
|
Thanks for this, @junjiequ. We're not building a replacement here: the Autopilot stage loop this PR fixes was removed from main (#15362, #15359, #15682), so the files and the |
|
Thanks for this fix, @junjiequ. Closing because the Autopilot plan/stage loop it adjusts has been removed from main: #15362 (backend, merged 2026-09-30) and #15359 (frontend), per |
Problem / Motivation
Four independent Autopilot Go-All sessions hit the same 30-minute card while actively editing, testing, or monitoring work. The implementation used
orchestrator.stage_timeout_secondsas a hard deadline around the running agent turn, even though the bundled Autopilot contract says it gates whether another turn may start and an admitted stage may outlast it.The timeout aftermath also split ownership: KiroCrew cancelled the dashboard coroutine without reliably cancelling the native ACP turn, preserved a stage boundary that trapped ordinary guidance, and exposed no clear recovery controls.
Why it matters
Normal productive stages were classified as runaway work and stopped at exactly 30 minutes. In two traced reproductions, the native ACP turn continued after the dashboard claimed it stopped: one completed a build and another began additional file writes. Late output became detached from the dashboard, temporary evidence could be reclaimed, and user guidance queued behind the preserved boundary.
What changed (motivation → approach → change)
The timeout model now has separate responsibilities:
orchestrator.stage_timeout_secondsis an unattended start gate. A productive stage may finish after 30 minutes; Go-All pauses before admitting the next stage and accurately names both the stage that spent the budget and the pending stage.agent.chat_turn_timeout_secsceiling (four hours by default), while existing ACP liveness controls continue to handle model silence and in-flight tool stalls.session/cancel, waits up toagent.soft_stop_budget_secsfor acknowledgement, and retains the cancellation task across a repeated outer cancellation.This restores the documented Autopilot contract instead of increasing the 30-minute value or removing runaway protection.
Backwards compatibility
Compatible. Existing configuration keys and defaults remain unchanged. Go-All still stops before starting more work once a stage spends its budget. Explicit Go, failure limits, round limits, whole-plan limits, session leases, and provider liveness controls remain in force.
Tests
Added or updated regressions that prove:
Validation on
afbb21042:git diff --check, and mypy across 1,725 source files passed;mainwithout rewriting the published first commit.The diff-scoped backend gate collected 34,962 tests in its first batch. It found the config-baseline drift introduced by the updated setting help text, which is fixed and passes in isolation. Its other failure is the unchanged Code Review Sage symlink-race test: the fixture requires a pytest temporary directory outside every managed root, while this environment sets
TMPDIRunder KiroCrew's managed scratch root. The same isolated node fails without this feature diff.The first published CI round also found a synchronous
cancel_currenttest double on Linux and Windows; that fixture now uses the async contract and passes locally. One unrelated fairness timing assertion missed its<0.1sthreshold at0.198s.Manual verification
Observed the premature timeout signature in four independent local dashboard sessions. Two were traced through both dashboard and native ACP records, proving the native turn continued after the dashboard timeout. The two additional screenshots show productive file edits immediately before the same Stage 1 timeout card.
A gateway running this revision is still needed for the final visual capture of the corrected boundary card.
Screenshots / video
The reported screenshots show the old failure state. The fixed live card will be captured once a gateway runs this revision; the backend and transcript behavior are covered by the regressions above.
Related Issues
Pattern harvest
Rule candidate: review-prompt
Pattern: a workflow budget that governs admission must be checked between units of work, not reused as the cancellation deadline for an already-running unit. Native cancellation must be scoped to the lease owner and its acknowledgement must be visible.
Checklist
Contribution License Agreement
N/A — repository placeholder retained; no CLA text invented.