GitHub Actions hardening P2900 (20/10/2026, Paris time)
Audit (last 200 runs per branch, up to 19/10/2026)
| Branch | Runs | Failed | Cancelled | Publish workflow | Publish failed | Publish median time |
|---|---|---|---|---|---|---|
| master | 200 (since 19/10) | 17 (8.5%) | 9 | Auto-Publish on Push | 6/27 (22%) | 802 s |
| stable-v5 | 200 (since 14/10) | 61 (30.5%) | 13 | Publish Stable to Test | 9/35 (26%) | 636 s |
| bastien-home | 200 (since 10/10) | 40 (20%) | 16 | Publish Zigbee Bastien | 10/62 (16%) | 189 s |
Failure causes: on stable-v5 and bastien-home, most red runs are test regressions (Node 22 runtime tests, Syntax Check, all fixed since). On the publish workflows, the causes are Athom processing_failed (socket hang up, AggregateError, invalid_state) and gh EOF/TLS errors. All workflows already had concurrency, timeouts and permissions. Publish groups use cancel-in-progress: false, so queued publishes are not cancelled. The cancelled runs come from code-quality and Node 22 superseding themselves, which is intended.
Changes
.github/scripts/athom-retry-policy.jsis byte-identical on all three branches, and test p2900 checks its hash. It sets the budget to 2 transient retries (socket hang up, ECONNRESET, S3 key, timeouts), 2 AggregateError retries and then a bisect, and 1 invalid_state retry, only after an idle wait. Unknown failures get no retry. Backoff is 60 s × 2^(n-1), capped at 240 s, plus 0 to 29 s of jitter. Every retry uses a strictly higher version.- master:
transient-publish-retry.jsreads its budgets and backoff from this policy. The values are unchanged. - stable-v5: the in-run re-upload loop uses the policy budget and backoff. It used 30 s × 2^n before. The failure kind is logged.
- bastien-home: there was no retry before. A new
reuploadstep runs when processing_failed comes with a retryable kind. Each attempt applies the version floor, bumps the patch, checks the floor gate, rebuilds, prunes, runs direct-api--force, and waits again. Tip-lag and in-flight builds are never re-uploaded. .github/scripts/publish-size-telemetry.jsmeasures each upload: gz archive, payload bytes and files, app.json bytes, drivers and flow cards. It records the outcome and kind in a CSV, uploaded as thepublish-size-telemetry-<run>artifact (kept 90 days). The CSV lives under$RUNNER_TEMP, so no bot commit picks it up.scripts/direct-api-publish.js(S3 PUT): addsAbortSignal.timeout(HOMEY_UPLOAD_TIMEOUT_MS, 300 s), so a stalled upload gets retried instead of hanging the step. Retry backoff now has jitter, andduplex: 'half'is set for the stream body. The default API timeout is now 180 s (was 120 s)..github/scripts/gh-retry.shretries gh only on transport errors (TLS, EOF, 5xx, secondary rate limit): 4 tries, 5/15/45 s plus jitter. master uses it for the retry dispatch.- Every
actions/*@vNreference in workflows and composite actions is pinned to a full SHA (master 48 refs, stable 20, bastien 29), and test p2900 enforces it.
Correlating failures with size
Download the artifacts (gh run download -n publish-size-telemetry-<id>), concatenate the CSVs, then group outcome/kind by archive_gz_bytes.