claw: the issue-shape gate judges a stale payload -- #10644 stayed closed 1246 s with a passing body #10662

What happened

.github/workflows/needs-boundary.yml judges an issue from the event payload: BODY: ${{ github.event.issue.body }}, REFUSED: contains(labels, 'shape-refused'), STATE: ${{ github.event.issue.state }}. GitHub fixes that payload when the event is created. On 2026-10-18 the job waited about twenty minutes for a runner, so the gate judged bodies and labels that were no longer true.

#10644, times from the run and job API (UTC):

The issue sat closed for 1246 s with a passing body, and only a third edit undid it. Five other lane issues were closed the same night (#10630, #10645, #10600, #10617, #10608). Each needed a manual reopen or another edit, and every edit queues another gate run.

specs/claw/shape_race.t27 replays a gate timeline under two readings. The payload reading is what the gate does today. The live reading reads the body, labels and state when the job runs. With the #10644 fixture, the payload reading leaves the issue wrong after 2 of 3 jobs. The live reading leaves it wrong after none. Without the 20:56 edit, the payload reading ends stuck (closed with a passing body). Mergify, bors and Kodiak make the same choice: they re-read live state when they act, not the event that woke them.

Proposed fix (owner's decision; this lane does not edit workflow files)

In the shape job, before judging, read the live issue: gh issue view "$N" --json body,labels,state. Then judge that body, and decide close or reopen from those labels and that state. Keep the payload only for the issue number.

User Scenarios

• Given an issue opened with a failing body and edited to a passing body before its first gate job runs, when the gate jobs run, then the issue ends open without the shape-refused label.
• Given an issue the gate closed, when an edit's job runs whose payload predates the close, then the gate still sees the live shape-refused label and reopens the issue if the live body passes.
• Given the claw: the merger waits about an hour for a runner -- 34 merge runs queued behind PR jobs (measured 2026-10-18) #10644 timeline as rows created TAB ran TAB ok, when race_plan in specs/claw/shape_race.t27 replays it, then it prints F shape-race race, with the payload reading wrong after 2 jobs and the live reading wrong after 0.

Requirements

• FR-001: The shape job MUST judge the issue body as it is when the job runs, not as the event payload carried it.
• FR-002: The shape job MUST decide close or reopen from the live labels and state.
• FR-003: specs/claw/shape_race.t27 MUST replay a timeline under both readings, and MUST read unknown when any row is unreadable.

Success Criteria

• t27c test-report specs/claw/shape_race.t27 --specs-dir specs reports 16 of 16 passing with 0 vacuous passes, and t27b test specs/claw/shape_race.t27 exits 0.
• t27c seal specs/claw/shape_race.t27 --verify exits 0.
• After the workflow change: an issue opened with a failing body and edited to pass within one minute ends open, without shape-refused, once both gate jobs have run.

Boundary

• .github/workflows/needs-boundary.yml
• specs/claw/shape_race.t27
• .trinity/seals/claw_ClawShapeRace.json
• docs/claw/README.md
• docs/reports/t27b_expectations.json

🤖 Generated with Claude Code