No workflow sets `shell:`, and GitHub runs such a step as `bash -e {0}`,
without pipefail: the exit code of `… | tee` is tee's, and a failing left
side passed silently. Three steps were unprotected:
- _process-resume.yml: an exception of process-resume.mjs (gh, API) left the
step green and the resume event was lost until process-reconcile;
- release-review.yml: a failed `prepare` went on with an incomplete
GITHUB_OUTPUT and proceed=true;
- validate.yml: a failed `classify-changes.mjs --heavy` left `heavy` empty,
heavy jobs were skipped and job `changes` stayed green.
Each gets `set -o pipefail` as the first line of `run` (validate.yml's step
becomes a block), following #727 and #472. test/workflow-pipefail.test.mjs
walks every .github/workflows/*.yml: a `| tee` line in `run` must follow
`set -[a-z]*o pipefail` or the step must have `shell: bash`; on the old tree
it names exactly the three places, and the _process-resume and validate
steps run on real bash under `bash -e` with a failing node.
ci-proof.mjs exports githubApiBase(env) (GITHUB_API_URL or
https://api.github.com, no trailing slash); githubCandidateTree,
loadGithubProofContext and release-gate's workflowRunsUrl take `apiBase`
with that default instead of the hardcoded host. night-red.mjs passes the
base directly and drops the fetch wrapper that rewrote the prefix. On
github.com the runner's GITHUB_API_URL is the same host, so behaviour there
does not change; archive_download_url stays as the API returned it.
The `mode` input for ship-review is out of scope (thin file in main, #716).
Thin files are not touched: _process-resume.yml is a body, validate.yml and
release-review.yml are not thin.
Issue: #751
User-Visible: no
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018qZfe7YS4rqEMKoVeS3GKd
`loadGithubProofContext` спрашивал объявленный `Baseline-Reviewed` run для
любого потребителя, а `evaluateCiProof` судил его при `reviewedRun: null` —
merge и review начинали зависеть от доступности старого run по чужой причине.
Теперь запрос делается только с `withReviewedRun` (release-gate передаёт его
вместе с ожиданиями), а проверка стоит внутри `if (expected)` — рядом со
сверкой evidence, где ей и место. Тест: без ожиданий merge/review green при
reviewedRun undefined/null/пустом; фейковый fetch доказывает, что запроса нет.
Issue: #573
User-Visible: no
Приёмка эталонов на beta.3 (`ad4000f9`) стоила второго полного Validate —
22 минуты, из них 17–22 на шард мутантов. Причина одна: корпус отпечатка
(`source-fingerprint.mjs`) называет `demo/golden` строкой-каталогом, а
замыкание входов раскрывало каталог во все текстовые файлы под ним, включая
`baselines-index.json`. Индекс становился входом smoke, performance_smoke и
каждого гарда через `serve.mjs`: на реальной паре C→B ключи smoke/perf были
DIFFERENT, отпечатки 181 из 183 браузерных свидетелей менялись, журнал их не
пропускал.
- `check-inputs.mjs`: `BASELINE_OVERLAY` — раскрытие каталога не выдаёт
overlay; явный корень golden и явная ссылка на файл — как были. На паре
C→B: ключи smoke/perf/parity/backend same, golden DIFFERENT; отпечатки
743 из 744 равны; план мутантов B с журналом C — 0–1 на шард вместо 38–44
- `ci-proof.mjs`: составное evidence — product tree без overlay, overlay
(tree, sha256 индекса, run из `Baseline-Reviewed`), content-ключи всех
реюзных job (исполненных тоже); `evaluateCiProof({expected, reviewedRun})`
сверяет с локальным расчётом, fail-closed на ключ, tree, индекс, reviewed
run, маркер с чужим ключом; proof без evidence при ожиданиях — stale
- `release-gate.mjs` / `release-prerelease.mjs`: ожидания считаются на
checkout кандидата (`candidateExpectations`), чужой checkout — notice
- мутанты: `baseline-overlay-leaks-into-every-key`,
`proof-trusts-evidence-it-could-verify`,
`reused-marker-key-unchecked-against-candidate`,
`product-tree-identity-counts-baselines`; перенацелен
`ci-proof-ignores-run-attempt`
- docs: TESTING (правило overlay), DEVELOPMENT (evidence в release proof),
STATUS
Issue: #573
User-Visible: no
The gate used to require every Validate run on the tag SHA to be green:
a cancelled duplicate or a red flake that a later re-run had fixed kept
the stable release blocked (v1.73.0, 09.09 — released by hand). Now the
verdict comes from the newest run that was not cancelled: not completed →
wait, success → pass, anything else → fail, no run → wait. The same rule
is documented for the perf workflow and the release runbook.
Mutants: release-gate-counts-cancelled-runs, release-gate-oldest-run-wins.
Issue: #511
User-Visible: no
