LearnNewsExamplesConnect
Frontmatter
assigneesneo-gpt-emmy
createdAtNov 5, 2026, 5:29 PM
updatedAtNov 5, 2026, 7:17 PM
githubUrlhttps://github.com/neomjs/neo/issues/20347
authorneo-gpt-emmy
commentsCount0
parentIssue20328
subIssues[]
subIssuesCompleted0
subIssuesTotal0
contentTrust
projected
quarantined0
signals[]
blockedBy[]
blocking[]
closedAtNov 5, 2026, 7:17 PM

Run units and timezone probes in one prepared CI job

Closed v13.4.0/archive-v13-4-0-chunk-3 enhancementaitestingperformancebuild
neo-gpt-emmy
neo-gpt-emmy commented on Nov 5, 2026, 5:29 PM

Context

Under #20328, the operator asked whether several tasks can share one runner instead of consuming one org slot per shard. The measured exploration is here. This first slice consolidates the browserless unit family; it does not change component or E2E execution.

At Engine dev 83006c9747c552ed0e4e0a496049f46211f9247a, test.yml creates one main-unit job and two timezone jobs. In run 42048502578, main units took 146s; the +14 and −11 probes took 3s and 4s, each after its own 9s dependency install and 11s bundle step. This is a one-run baseline, not a promised speedup.

The Problem

Three separate runner allocations repeat preparation for seven seconds of additional timezone work. The org queue makes those short jobs wait behind larger jobs. Dropping either timezone is not acceptable: #19005 establishes why UTC alone cannot catch the relevant regressions.

The Architectural Reality

  • .github/workflows/test.yml owns the matrix and dispatch probe interface.
  • playwright.config.unit.mjs already uses four workers in CI; the timezone config launches a separate Node process and sets TZ from NEO_UNIT_TZ.
  • The current jobs share /tmp/test-run.log, a unit JSON reporter path, and (for both timezone probes) an output directory. Even sequential consolidation must preserve each result before the next invocation writes.
  • The live dev ruleset requires unit, not the two unit-tz (...) contexts individually.

Design authority: #20328's terminal predicate says “the fixed checks run as a handful of jobs, not dozens”; the operator requested this aggregation exploration. Preserve #19005's non-UTC falsifiers and the current classifier as scope authority.

The Fix

Run main units and both timezone probes in one prepared unit job. Keep them as separate, sequential processes: main units retain their four-worker policy, followed by +14 and −11 probes. Seven seconds of sampled probe work does not justify another layer of concurrent processes.

Share checkout, Node setup, installation and browser-dependency bundling. Keep one aggregate conclusion that fails for any failed invocation, while reporting all three outcomes. Isolate logs, JSON and failure artifacts. Preserve job-level scope skipping, stale-head protection and the existing dispatch semantics.

Contract Ledger

Surface Authority Behavior Fallback / failure Evidence
Required unit context Current dev ruleset and #20328 One job covers main, +14 and −11 on ordinary selected runs Any invocation failure fails the job; canceled or unexecuted work is not success Forced failure in each position
Timezone probes #19005 and current timezone config Both signs execute in separate processes without changing the main unit timezone One failure does not hide the other result Existing mutation controls and distinct reports
Dispatch suite/specs/repeatEach Existing test.yml interface unit selects main only, unit-tz selects both probes, empty preserves ordinary selection Existing argument validation and measured-denominator reporting remain Focused dispatch-selection controls
Classifier and current-head gates classify-test-scope.yml Preserve run_unit, full dev selection and stale-head refusal Skipped work is explicitly reported, not silently omitted Selected, skipped and stale controls

Acceptance Criteria

  • Ordinary selected CI runs allocate one unit-family runner job instead of three, preserving the unit check name and all three populations.
  • Main, +14 and −11 results, logs and artifacts remain distinct; each timezone is set only for its own process.
  • A controlled failure in any of the three invocations fails the aggregate, while the other selected invocations still produce their results; cancellation/unexecuted cases cannot produce a false green.
  • Classifier skips, current-head checks, full dev coverage and dispatch suite/specs/repeatEach behavior retain focused positive/negative controls.
  • Record the candidate's hosted runner-job count, test-step duration, total job duration and coverage against the cited baseline. No reduced test population, relaxed timeout or added retry substitutes for consolidation.

Out of Scope

Component/E2E worker counts or packing; native parallel-step rollout; the shared baseline (#281 in Skills); Julia's #20344/#20345 surfaces; ruleset changes; timezone production-code changes.

Avoided Traps

  • A new matrix still allocates multiple runners.
  • Parallelizing the main four-worker pool with other CPU-heavy suites can add contention; this slice does not do that.
  • Reusing report paths erases an earlier failure.
  • A combined job must retain all outcomes rather than stopping after the first red.

Sequencing and Ownership

Owner: @neo-gpt-emmy. The existing #3782 author response is finished before this implementation starts. Grace confirmed this lands before #20338; that leaf then rebases for component-only job-level skipping. Not good first issue: already owned and coordinated with an active shared-workflow change.

Decision Record impact: none. Structure map: N/A, Engine workflow/config work with no new module-placement prescription.

Related

#20328 · #20338 · #19005 · #18916 · neomjs/neo-agent-skills#281

Live latest-open + closed sweep: latest 20 open issues and closed searches for timezone CI, unit runner and unit-tz checked on 2026-11-05; #19005 and #18916 decisions read. No equivalent unit-family packing leaf found. Recent 30-message all-state A2A sweep confirms Grace's sequencing and Julia's separate #20344/#20345 claim. Own assignments were #3782 and #16151, neither a duplicate. MC/KB searches on repeated setup and timezone runner waits supplied no contrary decision; live source and the cited issues govern.

Origin Session ID: 3270b5c6-cc24-4a5b-984c-b3a1a4967dcd

Retrieval Hint: "unit unit-tz runner allocation repeated setup three jobs one prepared unit context"

Authored by Emmy (GPT-6 Astra Ultra, Codex).