pack/.claude/commands/retry-intermittent-gha.md

Retry intermittent GitHub Actions failures from the last week. Re-run CI only — do not change code.

Steps

1. List failures

SINCE=$(date -u -v-7d +%Y-%m-%d 2>/dev/null || date -u -d '7 days ago' +%Y-%m-%d)
gh run list --status failure --created "${SINCE}.." --limit 200 \
  --json databaseId,displayTitle,workflowName,headBranch,event,url,createdAt,attempt,headSha
gh pr list --state open --json headRefName --jq '.[].headRefName'

2. Filter candidates

Keep	Drop
main (push / schedule)	Closed/merged PR branches
Open PR head branches	Older than latest failed run for same (workflowName, headBranch)
	attempt >= 3 (already flaky enough; report only)

3. Classify (per run)

Fast path — failed step names only:

gh run view "$RUN_ID" --json jobs \
  --jq '[.jobs[]|select(.conclusion=="failure")|{name,steps:[.steps[]|select(.conclusion=="failure")|.name]}]'

Failed steps	Likely	Next
only Start BuildKit / Stop BuildKit	intermittent (Docker pull flake)	confirm with logs
unit/lint job (Check and Test, etc.)	real	confirm with logs
mix of setup + test	unknown	logs required

Confirm with logs (always before retry; stop early):

gh run view "$RUN_ID" --log-failed 2>&1 \
  | rg -i -m 40 'i/o timeout|dial tcp|ECONNRESET|TLS handshake|toomanyrequests|429|502|503|504|rate.?limit|failed to resolve reference|temporary failure|--- FAIL:|snapshot failed|FAIL\s+github\.com|expected .+ got |error=.+timeout'

Batch ~3–5 log fetches at a time. Ignore cascade noise: Stop BuildKit / No such container: buildkit after a failed start is not a separate root cause.

Decision rule (check real signals first):

1. Any real signal → do not retry
   --- FAIL:, snapshot failed, compile/lint errors, deterministic expected … got …, version/API break (not transport)
2. Else any intermittent signal → retry
   registry/network timeouts, dial tcp, 429/5xx from registries, runner lost connection, Docker Hub pull flakes (moby/buildkit, ghcr, npm, PyPI)
3. Else → needs review (do not retry)

All failed jobs in a run must be intermittent (or pure cascades) to retry the run.

4. Rerun

gh run rerun "$RUN_ID" --failed

Continue on individual errors. Do not watch runs unless asked.

5. Report

Category	run	workflow	branch	jobs	signal/reason	url
retried					e.g. docker.io i/o timeout
real (skipped)					e.g. snapshot failed
filtered (skipped)					closed PR / not latest / attempt≥3
needs review					why unclear

Counts at the top. One line per run.