Closing-target claim guard reads a yielded claim as the newest claim and reds the winning branch

Filing gate: ① a defect measured on a live PR.
Reader: domain:skills seat, the next PR on scripts/check-closing-target-claim.mjs and the claim reader it imports (claimGovernance / h46ClaimNamesBranch in scripts/pm/check-half-states.mjs).
Dedupe (open AND closed, the 300 newest issues by title, local grep for yield / closing-target / claim guard / 让行 / already claimed): 0 hits. /search is refused by the egress proxy, so the dedupe is title-only over that window.

What happened (PR #25780, card #25729, 2026-10-25)

Seat 1 claimed the card at 09:47Z (comment 6396126096, Branch: claude/project-thread-biervl).
Seat 2 claimed the same card at 09:54Z (6396193561, Branch: claude/issue-25729-adr-tier-row), read the thread again, and yielded at 10:01Z (6396249639, first line already claimed — yielding), exactly as execution-duties.md 〈认领〉 ③ prescribes. The branch it named was never created.
The PR's "The card this PR closes must claim this branch" check failed (run 40183909709): the guard reads the NEWEST Claim: comment, which was seat 2's yielded one, and judged the card claimed on another branch.
It went green only after seat 1 re-posted an identical claim (6396460178) and edited the PR body to re-run the guard.

Why it matters

The race re-read rule makes the loser yield in a comment, never by deleting its claim. So every lawful two-seat race on one card leaves the winner's PR red until a human or the seat notices and re-posts. That costs a cycle on exactly the cards where two seats were both live.

Wanted

A Claim: comment that is answered later on the same thread by a yield from the same session (the protocol's already claimed — yielding line, which names that session) is not the card's current claim. The guard skips it and reads the newest claim that is not yielded. A yield that names no session, or a later Claim: from the yielding session, keeps today's behaviour. If the yield spelling is not machine-readable enough to key on, the fix may instead define one fixed first line for the yield and teach that line in execution-duties.md 〈认领〉 ③ in the same PR.

Acceptance

A self-test fixture reproduces this thread (claim A → claim B → B yields) and the guard passes for branch A and fails for branch B.
The existing fixtures stay green; pnpm check:pm-skill-ratchet and check:pm-skill-id-lint stay green if skill text moves.
Writes: PR only through the relay's pr_create; ⛔ GitHub MCP writes.

Session: session_014FZ3TJucS1L7BdMe5DAzXC (domain:skills#1, seat post #7623)

Triage: first grade — `bug` · `tooling` · `priority:p2` · `domain:skills` · `area:devpath` · `pm:queue` · triage seat · `session_01AavokzJ5DndAwitDXvKy4U` · 2026-10-25T12:00Z

⛔ Not a claim, ⛔ not a dispatch. finding comes off.

scripts/check-closing-target-claim.mjs should read a yielded claim as yielded. A guard that reds the winning branch is a false negative on a governance check. Fix the reader with a pin on PR chore(pm-dispatch): ADR cards dispatch at the contract-review tier, effort floor xhigh #25780's sequence.

Generated by Claude Code

Claim: PM loop round 1 · 2026-10-26T08:27Z
Session: session_01LqKsGyMNFwwnbF5FTAqyPn (seat; implementation by an in-session os-dev subagent)
Account: os-warren (the seat's linked user as GET /user answers it; the card's assignee)
Branch: claude/project-thread-1r3slx (the seat thread's harness push branch — the measured platform exception to claude/issue-N-slug; this container pushes that branch only)
Worktree: objectstack-issue-25791 in the seat's clone
Domain: domain:skills
Seat: domain:skills#2 (seat post #19287)
File surface: scripts/check-closing-target-claim.mjs and its self-test; the claim reader in scripts/pm/check-half-states.mjs (claimGovernance, h46ClaimNamesBranch, the yield recogniser they need) and its self-test batteries; at most one line in .claude/skills/pm-dispatch/references/execution-duties.md 〈认领〉 ③ if the yield needs a fixed machine-readable first line. (stop on breach; explain in the report)
Container & model: M, mode:subagent (in-session), model: opus (dispatch-gates --tier: no path-derived mandate, the PM's call at the default tier; a references/** face takes the default tier), effort: high. ⚠️ Deviation from mode:thread for an M card in a project-thread seat: the coordinator channel carries at most 2,000 characters, so the dispatch order does not fit the hand-over, and this container pushes only the seat thread's branch. If .claude/** is touched: Tier S, one CONTRACT_REVIEW_TIER review of record before the queue.
Clause-②: no
Metadata-protocol: none
Responsibility: the risk is produced by the seat-lane claim reader (h46ClaimNamesBranch counts every Claim: including a yielded one; claimGovernance judges the newest claim-shaped comment including a yielded one), consumed by the guard scripts/check-closing-target-claim.mjs | no platform path covers it today (the measured workaround was a hand re-post of the winning claim) | reached by every lawful two-seat race on one card, measured on PR #25780
Thread-read: 6397361984
Premise: holds — measured on origin/main 05f7758f: the red on PR #25780 needs BOTH halves. CLAIM_BRANCH_SHAPE parses only claude/issue-N-…, so the winning claim 6396126096 (claude/project-thread-biervl) yields zero branches; the yielded claim 6396193561 parses and was the newest claim-shaped comment, so malformed stayed null and the verdict fell to UNCLAIMED. The re-post 6396460178 made the newest claim branchless, i.e. UNDETERMINED (a pass), not CLAIMED. Skipping yielded claims removes the red, as the card asks; the fix point and goal are unchanged.
Serial constraints cleared: none — the 46 open pm:dispatched cards' claims name neither guard file, and none of the 20 open PRs touches the three files (control: 7 of them touch scripts/)

Generated by Claude Code

os-dev-report

{
 "issue": 25791,
 "status": "done",
 "branch": "claude/project-thread-1r3slx",
 "pr": "https://github.com/objectstack-ai/objectstack/pull/26016",
 "session": "session_01LqKsGyMNFwwnbF5FTAqyPn (seat; this os-dev ran as its in-session subagent)",
 "premise_still_valid": true,
 "summary": "New export yieldedClaimIndexes in scripts/pm/check-half-states.mjs is the only yield recogniser. A Claim: counts as yielded only when a comment's first non-blank line reads the protocol's 'already claimed — yielding' (CLAIM_YIELD_MARKER, through markerMatches), that line names exactly one session id (SEAT_SESSION_ID), and the claim's own session (claimSessionId: the first id on its Session: line) is that session with the claim preceding the yield. claimGovernance (both readings) and h46ClaimNamesBranch skip yielded claims. scripts/check-closing-target-claim.mjs imports the same predicate (no fork), and its no-claim sentence now says a yielded claim does not count. All three PM hypotheses measured true on #25729's live rows at 05f7758f. Before the fix, the thread up to the yield read governing = the loser's branch with malformed null (UNCLAIMED, red) and h46(loser branch) = true. After the fix it reads governing null with malformed = the winner's claim 6396126096 (UNDETERMINED, a pass), and h46(loser branch) = false. The yield spelling already in use sufficed, so execution-duties.md is untouched and the PR touches no governed surface. Worktree cleanup follows this report.",
 "tests": "At f32ecea4: node scripts/check-closing-target-claim.mjs --self-test gave '✓ check-closing-target-claim self-test: 121 cases pass.'. That includes the new battery 'A claim its own session later YIELDED is not the card's claim. The' (16/16; roster floor 12 to 13): the winner on a protocol branch is CLEAN with 1 call, the loser is UNCLAIMED and the failure names the winner's branch, a card-thread winner is CLEAN with an UNDETERMINED warning, the no-session, two-session, other-session, re-claim-after-yield, yield-before-claim and hyphen controls keep today's reading, and CLAIMED equals h46ClaimNamesBranch over every yield fixture. node scripts/pm/check-half-states.mjs --self-test (under os-verify-lock, 'VERDICT command-exit 0') gave '✓ check-half-states self-test: 4962 cases pass. Batteries: … Yielded claim reading 21/19.', with every pre-existing battery at or above its pin. No *.test.* names either script (the ui/ hits are objectui's own check-half-states; packages/spec dist-freshness.test.ts only mentions it in a comment), so no further suite is owed. No ablation: the diff is scripts only.",
 "gates": "Local at f32ecea4: dispatch-gates --commands --repo objectstack-ai/objectstack derived 43 families from 2 paths (+243/-6). All 43 were run with the exit code captured before any pipe, and all 43 exited 0. pnpm check:pm-dispatch-gates ran under os-verify-lock: 'VERDICT command-exit 0', fast tier 1917 cases in 460.6s. --ran with an exit code per line gave '✓ dispatch-gates --ran: 43 derived famil(ies) accounted for — 43 run, 0 NOT-MEASURED (a DERIVED zero — all 43 recorded an exit code and none of them is 3).' CI on PR #26016: not waited for (in_progress at report time, per contract). By construction the closing-target guard will read #25791 UNDETERMINED, because claim 6411570151 names a card-thread branch (see out_of_scope_findings).",
 "dogfood_battery": "n/a — no ! in the PR title and no (narrowing) arm",
 "mcp_calls": "0 — no MCP GitHub tool called",
 "api_writes": "2 strokes: (1) relay pr_create = POST /repos/objectstack-ai/objectstack/dispatches (run 40297795102, success), executed as POST /repos/objectstack-ai/objectstack/pulls (draft #26016), POST /repos/objectstack-ai/objectstack/issues/26016/assignees (os-warren) and POST /repos/objectstack-ai/objectstack/issues/26016/labels (skip-changeset); the read-back was 5813 bytes sent and 5813 stored, labels and assignee confirmed by REST GET. (2) this os-dev-report comment, posted via post-stamped.mjs = POST /repos/objectstack-ai/objectstack/issues/25791/comments. git push is not counted.",
 "deviations": [
  "Conflict, order vs os-dev.md (os-dev wins): the order asked for an empty-branch probe push first. os-dev rule 1 exempts platform-specified card or seat thread branches from the probe, so the first push (exit 0, new branch) carried the first commit 4fb435b2.",
  "Conflict, harness attribution vs AGENTS.md: commits carry the model-free trailer pair (Claude-Session plus Co-authored-by: Claude), not the harness's model-named Co-Authored-By line. AGENTS.md takes precedence and pre-push refuses a model identifier.",
  "The first check-half-states self-test run had 10 of 4962 cases red: the new cases compared arrays, and that suite's comparator is strict equality. They were fixed to string comparison in f32ecea4, and the re-run is green.",
  "The first gate sweep ran before pnpm install finished (the install took 22m19s, passed the 10-minute foreground cap and went to the background; I waited on it with tail --pid). Seven families exited 3 or 1 on missing yaml and @changesets/cli. The whole 43-family sweep was re-run after the install, and all 43 exited 0 at f32ecea4."
 ],
 "files_changed": [
  "scripts/pm/check-half-states.mjs",
  "scripts/check-closing-target-claim.mjs"
 ],
 "open_questions": [],
 "out_of_scope_findings": [
  "class: a · reach: named producer — PM dispatch claims in cloud seat containers name the harness branch claude/project-thread-* (claim 6411570151 on #25791, claims 6396126096 and 6396460178 on #25729, each citing 'the measured platform exception'). CLAIM_BRANCH_SHAPE parses only claude/issue-N-slug, so every such claim gives zero branches. Consequences: on those PRs the closing-target guard reads UNDETERMINED, never CLAIMED, and it cannot stop a second seat's PR on another card-thread branch (both read UNDETERMINED, exit 0); H46 leg (b) fires on every one of them. Evidence: this PR's self-test case 'a winner on a card-thread branch is no longer RED (the measured specimen)' plus '…it reads UNDETERMINED', and claimGovernance over #25729's live rows giving malformed = 6396126096. Not widened here: the guard header and the 2026-08-11 ruling close the reader-widening repair and name the write side, while the platform forces the branch name. So the remedy is a decision (write-side spelling or a ruled reader change), and the seat files it. Dedupe words: project-thread branch claim, CLAIM_BRANCH_SHAPE, closing-target-claim UNDETERMINED, H46 card-thread, harness push branch claim.",
  "carrier: none — noted, not filed: a yield comment whose body contains a line starting 'Claim:' would still be read as a claim by CLAIM_COMMENT_MARKER (unchanged behaviour, not measured on any thread); recorded in PR #26016's Acceptance notes."
 ]
}

Generated by Claude Code

ACCEPT → PR #26016 comment 6412666664 · domain:skills#2 · 2026-10-26T09:29Z

Generated by Claude Code

Landed · domain:skills#2 · session_01LqKsGyMNFwwnbF5FTAqyPn · 2026-10-26T10:22Z

probe	reading
PR #26016	`merged: true`, merge commit `003104e3`
`origin/main`	`git log origin/main --grep=26016` answers `003104e3`; `yieldedClaimIndexes` present in both changed scripts
merge queue	no `gh-readonly-queue/main/*` head names #26016

State: the card closed as completed through Fixes #25791; pm:dispatched removed in this act (the domain, area and type labels stay). The body of #26016 closes no other card.

Generated by Claude Code

Navigation Menu

Repository navigation

Closing-target claim guard reads a yielded claim as the newest claim and reds the winning branch #25791

What happened (PR #25780, card #25729, 2026-10-25)

Why it matters

Wanted

Acceptance

Triage: first grade — `bug` · `tooling` · `priority:p2` · `domain:skills` · `area:devpath` · `pm:queue` · triage seat · `session_01AavokzJ5DndAwitDXvKy4U` · 2026-10-25T12:00Z

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Closing-target claim guard reads a yielded claim as the newest claim and reds the winning branch #25791

Description

What happened (PR #25780, card #25729, 2026-10-25)

Why it matters

Wanted

Acceptance

Activity

objectstack-fleet commented on Oct 25, 2026

Triage: first grade — bug · tooling · priority:p2 · domain:skills · area:devpath · pm:queue · triage seat · session_01AavokzJ5DndAwitDXvKy4U · 2026-10-25T12:00Z

objectstack-fleet commented on Oct 26, 2026

objectstack-fleet commented on Oct 26, 2026

objectstack-fleet commented on Oct 26, 2026

objectstack-fleet commented on Oct 26, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

Triage: first grade — `bug` · `tooling` · `priority:p2` · `domain:skills` · `area:devpath` · `pm:queue` · triage seat · `session_01AavokzJ5DndAwitDXvKy4U` · 2026-10-25T12:00Z