Repository navigation
ship finish: an owner-ruled note closes its finding, and every ruling is listed (GST-1537) - #4050
Conversation
… is listed (GST-1537) Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
|
😎 This pull request was merged. |
… like its ship-bin-fixture siblings (GST-1537) Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
…ng Review lists ruled notes (GST-1537) Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
Review-Finding: p1/specialist-batched#1 reader=state-machines Review-Finding: p1/red-team#2 reader=red-team Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
Review-Finding: p2/red-team#1 reader=red-team Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
Review-Finding: p2/red-team#1 reader=red-team Review-Finding: s11/outside-voice#1 reader=outside-voice Review-Finding: p1/specialist-batched#2 reader=testing Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
E2E Evals: ✅ PASS25 executed, 0 reused rule/judge records | 0 manual accepted (unscored; no score-cache credit) | $11.98 rule/judge cost | reconcile exit: 0 Profile: pr / mode Rule and judge shards
Fail-closed reconciliationSliced lane: planner → duration-packed executors → fail-closed report. Behavior cases run a pre-registered 3-trial panel (PASS at 2/3 with no contract violation); a PASS 2/3 is shown with its failed trial, never as a clean pass. Reused results retain their original provenance and expiry. Dispatched narrow screen (gstack-ship-queue ready --screen) of merge commit |
What changed
An owner can now rule a review finding a disclosed note, and
finishtreats that finding as closed on a granted fix-check. The ruling is an approval row bound to the stop head the owner ruled at, never a field on the lane's own review-log row, and every ruling is listed in finish's output, the PR body and the receipt.gstack-approvals rule --finding <fingerprint> --head <stop head> --reason <why> --words <exact words> --by <who ruled>writes an approved row with scoperuled_note <fingerprint> <head40>and askruled_note <fingerprint> at <head40>: <reason>.--headmust resolve to a commit (git rev-parse --verify <sha>^{commit}), the same commit areview_loop_grantnames.rulerefuses a reason with a newline or control character, a--bynaming gstack, and any ruling inside the checkout's open /ship run.addrefuses theruled_notescope, soruleis the only path.gstack-approvals rulingsprints oneRuled note: …line per ruling.lib/ship-release.tsroute) prints every ruling asRULED_NOTE: <fingerprint> at <head12> ruled by <who> at <time>: <reason> (approval <id>). A ruling closes a finding on the newest fix-check row only when its head lies inside that row'srange_base..commit, the same ancestry test the grant passes. So the forward order (the owner rules at the stop, then the granted fix-check runs) and the qm#3122 order (the fix-check runs, then the owner rules) both bind, and a ruling made at an earlier /ship's head closes nothing and is printed with(outside this fix-check's range; not applied). A skipped CRITICAL with no ruling stays open. Aruledfield oraction: "ruled"the lane writes on its own row is never read.ruled_notesfield ({finding, head, by, reason, at, approval}, head required).gstack-ship-receipt write --bodyrefuses a body missing anyRuled note:line word for word. The READY tail prints each asRULED_NOTE:, and the prose check skips those lines.Why
On qm#3122 at 532cc5c (CAPY-1, gstack 1.96.0.0), the owner ruled two CRITICALs notes, not fixes. The runner logged them as
skipped, and a skipped CRITICAL counts as open. So the granted fix-check row, whose own findings were all below the floor, never read CURRENT, and finish stopped withFIX_CHECK: not counted (review has unresolved findings or did not finish clean). A green, reviewed head couldn't get a receipt.Proof
test/ship-finish-ruled-notes.test.ts, written failing first at each step:RULED_NOTE:lines.ruled: "note"oraction: "ruled",addwith the ruling scope,ruleinside an open /ship run, andrule --by gstack ….rulerefuses a missing--head, an unknown sha, a tree sha and a multi-line reason.Blast radius
Merge-check code: finish's fix-check credit,
reviewFreshness(finish is the only caller that passes rulings; gstack-review-read and Step 11.5 are unchanged), receipt write and its schema, andgstack-approvals. With no ruling rows, every path behaves as before.Security read
No network calls, no new environment reads, no permission or workflow changes, no new dependencies. The trust anchor is the existing approval record. As with
review_loop_grant, anyone who can rungstack-approvalson the machine can write a row, so the lane-side guards are the open-run refusal, the rule that the review log never carries a ruling, and the head binding that limits each ruling to one fix-check range.Summary
The owner's "this ships as a note" ruling now lets a granted fix-check count, without letting the lane rule on its own findings, and without letting an old ruling close a new finding.
Approvals
Approved: yes (record r-ac892057cdb0). The full approval record stays private: this repository is public or its visibility is unknown.
Pre-Landing Review
Pre-Landing Review: 10 issues — 3 auto-fixed (two CRITICALs in round 1, one in round 2), 0 asked, 0 deferred, 7 noted. Full panel (specialists, red team, native adversarial voice on Opus 5.5, outside voice on openai/gpt-6.1-sol), two rounds plus one scoped fix-check under the Captain's grant a-608eba2f6cc4.
REVIEW_LOOP: passes=2 blocking_found=7 deferred=0 noted=4 fixed=5 scope=frozen unrecorded=3 ledgers_read=5 fix_checks=1 dispatches=8
Review-loop exception: accepted by GStack Minder (GSTA-36), via GSTA-321 2026-11-06 via approval a-608eba2f6cc4 (review_loop_grant b8219ba): the full panel applies (at most 2 rounds plus one scoped fix-check, then park or split) (fix-check at f6bb1c3, 2 findings)
Gate edits: none detected (GATE_SCAN tagged=0 at f6bb1c3)
Fixed:
rulerefuses a multi-line reason.Review notes
rule --branchfrom a sibling worktree, or a direct append to the record, is not blocked. Disclosure in the PR body and receipt is what limits it, as withreview_loop_grant.writegets--body; receipt acceptance in the queue doesn't re-check it (Step 11 outside voice).ask[2] !== scope[2]guard is untested (fix-check).reviewFreshnesswithout rulings, so the review dashboard shows a ruled fix-check as UNVERIFIED while finish counts it (Step 11 native voice).Documentation
Scope: the 16 committed paths in a339b93..f6bb1c3 (GST-1537 ruled notes), plus the docs that mention gstack-approvals or the review-loop cap. Result: current. docs/ship-policy.md#ruled-notes documents
gstack-approvals ruleandrulings, whatrulerefuses, how a ruling's head must fall inside the fix-check's range, theRULED_NOTE:output and theruled_notesreceipt field. docs/review-policy.md links to that section, and the PR-body template asks for eachRuled note:line. Nothing was edited. Debt: none found.Test plan
bun test test/ship-finish-ruled-notes.test.ts(9 pass) plus approval-record, headless-artifacts, ship-receipt-prose, parity-suite, skill-md-size-guard, module-size-ratchet, test-free-shards, review-packets-removed and ship-release: 196 tests, 0 fail, at f6bb1c3.bun run typecheckclean; push guards passed (typecheck, skill-size, parity-suite, context-budget, module-size, goldens, conflict-markers).