Context
#20328 (cause 3: job count against 20 org-wide slots). At 15:43Z on 2026-11-05, 21 PR Baseline runs sat queued in this repository at ten jobs each. Three of them were for heads that had already been superseded; one was #20332's first head, queued for 64 minutes. No other PR workflow leaves superseded runs alive. pr-baseline.yml has no concurrency block, while sixteen sibling PR workflows have one.
The Problem
A push or a body edit starts a new baseline run, and the previous run keeps its place in the queue. When that run finally starts, it spends ten slots judging a head nobody is proposing any more.
The Architectural Reality
.github/workflows/pr-baseline.yml calls neomjs/neo-agent-skills/.github/workflows/reusable-pr-baseline.yml@v0.1.84 on pull_request: [opened, edited, synchronize, ready_for_review]. edited is there so a corrected body goes green without a push.
same-head-check-repair.yml watches PR Baseline and reruns, once, every earlier red run (cancelled included) on the same head after a later run passes. A cancelled run on the same head therefore costs a full rerun later.
- #15593 (
build.yml and the test workflows) set the rerun-safe group: first attempts share one group per ref, and a rerun gets its own (github.run_attempt == '1' && github.ref || github.run_id). Otherwise an obsolete rerun cancels the current head's run.
Design authority: #15593: "Normal first-attempt supersession is valuable and must remain: a new commit should cancel expensive work for the previous commit", plus its rerun-safe group.
The Fix
pr-baseline.yml gains:
group: pr-baseline-${{ github.run_attempt == '1' && github.ref || github.run_id }}
cancel-in-progress: ${{ github.event.action != 'edited' }}
A new head cancels the run for the old one. A body edit waits behind the running run instead of cancelling it, so the same-head repair has nothing to rerun.
A new spec, PrBaselineConcurrency.spec.mjs, pins three things: both expressions, and that the repair still watches PR Baseline, which is why edited must not cancel.
Acceptance Criteria
Out of Scope
- Ten jobs into one, and an
edited run that only needs pr-body: neomjs/neo-agent-skills#281.
- Brain's and Institution's own
shared-pr-baseline.yml callers. Each is its repository's own one-file change.
Avoided Traps
- Plain
cancel-in-progress: true: a body edit would cancel its own head's run, and the same-head repair would rerun all ten jobs.
- A ref-only group: a stale rerun cancels the current head's run (#15593).
Related
#20328 (parent) · #15593 · #20064 (same-head repair) · neomjs/neo-agent-skills#281
Sweeps at 2026-11-05T15:45Z found no equivalent:
- latest open;
- an all-state search on baseline / concurrency / superseded;
- A2A claims;
- Memory Core.
Origin Session ID: 8545cff9-f4d7-4731-b93e-7dfad1347c8a
Retrieval Hint: "PR Baseline concurrency superseded head cancel-in-progress edited same-head repair run_attempt"
Context
#20328 (cause 3: job count against 20 org-wide slots). At 15:43Z on 2026-11-05, 21
PR Baselineruns sat queued in this repository at ten jobs each. Three of them were for heads that had already been superseded; one was #20332's first head, queued for 64 minutes. No other PR workflow leaves superseded runs alive.pr-baseline.ymlhas noconcurrencyblock, while sixteen sibling PR workflows have one.The Problem
A push or a body edit starts a new baseline run, and the previous run keeps its place in the queue. When that run finally starts, it spends ten slots judging a head nobody is proposing any more.
The Architectural Reality
.github/workflows/pr-baseline.ymlcallsneomjs/neo-agent-skills/.github/workflows/reusable-pr-baseline.yml@v0.1.84onpull_request: [opened, edited, synchronize, ready_for_review].editedis there so a corrected body goes green without a push.same-head-check-repair.ymlwatchesPR Baselineand reruns, once, every earlier red run (cancelledincluded) on the same head after a later run passes. A cancelled run on the same head therefore costs a full rerun later.build.ymland the test workflows) set the rerun-safe group: first attempts share one group per ref, and a rerun gets its own (github.run_attempt == '1' && github.ref || github.run_id). Otherwise an obsolete rerun cancels the current head's run.Design authority: #15593: "Normal first-attempt supersession is valuable and must remain: a new commit should cancel expensive work for the previous commit", plus its rerun-safe group.
The Fix
pr-baseline.ymlgains:group: pr-baseline-${{ github.run_attempt == '1' && github.ref || github.run_id }}cancel-in-progress: ${{ github.event.action != 'edited' }}A new head cancels the run for the old one. A body edit waits behind the running run instead of cancelling it, so the same-head repair has nothing to rerun.
A new spec,
PrBaselineConcurrency.spec.mjs, pins three things: both expressions, and that the repair still watchesPR Baseline, which is whyeditedmust not cancel.Acceptance Criteria
actions/runs?status=queued.Out of Scope
editedrun that only needspr-body: neomjs/neo-agent-skills#281.shared-pr-baseline.ymlcallers. Each is its repository's own one-file change.Avoided Traps
cancel-in-progress: true: a body edit would cancel its own head's run, and the same-head repair would rerun all ten jobs.Related
#20328 (parent) · #15593 · #20064 (same-head repair) · neomjs/neo-agent-skills#281
Sweeps at 2026-11-05T15:45Z found no equivalent:
Origin Session ID: 8545cff9-f4d7-4731-b93e-7dfad1347c8a Retrieval Hint: "PR Baseline concurrency superseded head cancel-in-progress edited same-head repair run_attempt"