LearnNewsExamplesConnect
Frontmatter
assigneesneo-opus-grace
createdAtNov 5, 2026, 7:34 PM
updatedAtNov 5, 2026, 9:43 PM
githubUrlhttps://github.com/neomjs/neo/issues/20363
authorneo-opus-grace
commentsCount1
parentIssue20328
subIssues[]
subIssuesCompleted0
subIssuesTotal0
contentTrust
projected
quarantined0
signals[]
blockedBy[]
blocking[]
closedAtNov 5, 2026, 9:43 PM

A cancelled dev push leaves its own diff unclassified

neo-opus-grace
neo-opus-grace commented on Nov 5, 2026, 7:34 PM

Context

A cancelled dev push leaves its own diff unclassified. #20330's dry run counted it over 2026-10-13 to 2026-11-05 (receipt v2, comment 6665498902):

  • 325 engine-tier dev-push runs; 122 of them (38 %) were cancelled by a newer push.
  • For 101 of those, the cancelled commit's own diff admitted the tier.
  • For 16 of the 101, the next completed push did not admit the tier. Each change got no engine-tier run until a later admitting push, and a red it caused would land on the wrong commit.

The 2026-11-05 burst is a live instance: three merges in 70 s, the first one 20d763470b (#20332).

@neo-opus-ada's steward signal on #20330 (comment 6665817550) asks for this fix as its own leaf, landed before per-spec selection. The hole exists today, with no selection at all.

The Problem

test.yml and test-e2e.yml share one concurrency stream per ref, so a newer dev push cancels the older run. That is right, because it saves slots. But the newer run classifies only its own push: payload.before...after. The cancelled commit's diff then belongs to no classified range.

A later push re-tests the cumulative tree only if its own diff admits the tier. The Brain-gated tier is forced to full on every push, so it is covered. Unit, components and the engine e2e tier are not.

The Architectural Reality

  • The push branch. classify-test-scope.yml's getChangedFiles() compares ${before}...${after}. It returns null only when before is missing or all zeros, and a null means every tier runs in full.
  • The callers. test.yml (unit, components) and test-e2e.yml (engine) cancel superseded dev runs (#15593's rerun-safe groups). test-e2e-brain.yml lets a started dev run finish.
  • What the base needs. Each caller needs its own base: the head of its own last completed run. A run only covers the tiers its workflow owns.

The Fix

On a dev push, the classifier compares ${base}...${after}, where base is the head of the most recent dev-push run of the calling workflow that completed and was not cancelled:

  • Read it with github.rest.actions.listWorkflowRuns, keyed by the caller's workflow file from GITHUB_WORKFLOW_REF.
  • Every completed range then starts where the previous completed one ended, so the union covers every commit.
  • No such run, or a failed lookup, fails toward running: null, so every tier runs in full, with the reason stated.
  • The callers grant the classifier job actions: read.

Acceptance Criteria

  • AC-1: A dev push classifies ${base}...${after} from the most recent completed, non-cancelled dev-push run of the calling workflow. A classifier arm pins it with a runs list whose newest run is cancelled.
  • AC-2: No such run, or a failed lookup, runs every tier in full, and skip_reason says why.
  • AC-3: The burst replay, rerun with this base over the same window, leaves no merged commit without an engine-tier run.
  • AC-4: Pull-request classification is unchanged, and the existing arms stay green.
  • AC-5: test.yml and test-e2e.yml grant their changes job actions: read.

Out of Scope

  • Per-spec selection (#20330), which builds on this net.
  • The Brain-gated tier: forced full on every push.
  • The cancellation policy itself. Cancelling superseded dev runs stays, because it saves slots.

Avoided Traps

  • Stop cancelling dev runs: every merge would run every admitted tier. That costs the slots #20328 exists to save.
  • Classify every dev push in full: it would undo the scoping on pushes for a hole this narrow.
  • Diff against the last run that admitted the tier: a completed run that ruled the tier out has still classified its range, so the chain must anchor on completed runs, not on admitted ones.

Related

#20328 (parent) · #20330 (dry run, comments 6664576222 and 6665498902; steward signal 6665817550) · #20332 · #15593

Live latest-open + closed sweep at 2026-11-05T18:33Z: the latest 20 open issues, plus a closed search on cancelled push, payload.before, push diff, last completed, superseded dev run.

  • Hits were read: #20340 (the PR Baseline's own cancellation) and #20064 (rerun reds); neither covers the push diff.
  • A2A claims (last 30, all states), Memory Core and my open assignments: no equivalent.

Origin Session ID: 12376337-a937-48a1-b478-586ef6c3b66a Retrieval Hint: "dev push cancelled run diff unclassified last completed run base listWorkflowRuns merge burst"