CI Debug

/ci-debug — classify a failing CI run

Direct response to the recurring CI-debug pattern surfaced by /insights: ~12 sessions in 3 weeks doing the same classification dance. This skill encodes the 11 patterns so the dance becomes a lookup.

Input

User invokes with one of:

• PR number: /ci-debug 822 (default repo from context; ask if ambiguous)
• Run URL: /ci-debug https://github.com/owner/repo/actions/runs/12345
• Job URL: /ci-debug https://github.com/owner/repo/actions/runs/X/job/Y

Execution

1. Resolve the failing job

# From PR number:
gh pr checks <n> --repo <owner>/<repo> --json bucket,link,name \
  --jq '.[] | select(.bucket=="fail") | "\(.name)|\(.link)"'

# From run URL:
gh api repos/<owner>/<repo>/actions/runs/<run-id>/jobs \
  --jq '.jobs[] | select(.conclusion=="failure")
                | {id, name, runner_name, started_at, completed_at,
                   steps: [.steps[] | select(.conclusion=="failure") | {name, number}]}'

If multiple jobs failed, pick the one with the shortest duration — root cause is usually the first failure; later jobs cascade.

No job in the fail bucket but a check won't settle? If gh pr checks shows zero fail-bucket entries yet a status sits in pending that never resolves (and gh pr view --json mergeStateStatus returns UNSTABLE while mergeable=MERGEABLE), this is a stuck external status, not a failure — jump straight to Pattern #11. There is no failing log to fetch; classify on the commit-status metadata (gh api repos/<o>/<r>/commits/<sha>/status).

2. Fetch the failing log

gh api repos/<owner>/<repo>/actions/jobs/<job_id>/logs 2>&1 \
  | grep -iE '(error|fail|ERR_|CONFLICT|Process completed with exit code)' \
  | head -30

Capture the FIRST distinct error message (later lines often echo).

3. Classify against the playbook

Walk the patterns in order. First match wins.

The memory references point at user-curated memory files (~/.claude/projects/<project>/memory/*.md). If your memory doesn't have them yet, the signature column is enough to classify — the memory citation is a nice-to-have, not required.

4. Report

For a matched pattern:

## CI Debug: <repo> · <pr-or-run-ref>

**Failing job:** `<job name>` (<duration>s) on runner `<runner_name>`
**Failing step:** <step name> (#<step number>)
**Error excerpt:**
\`\`\`
<first 3 lines of grep'd error>
\`\`\`

**Classification:** Pattern #<n> — <pattern name>
**Reference:** memory `<memory-file.md>`

**Proposed fix:**
<exact commands, one per line>

**Will I apply this?** No — awaiting your approval. Reply "go" to ship.

For an UNMATCHED failure:

## CI Debug: <repo> · <pr-or-run-ref> · NOVEL

**Failing step:** <step name>
**Unique log lines:**
\`\`\`
<top 10 distinct error lines>
\`\`\`

This doesn't match any of the 11 playbook patterns. Surfacing the raw
evidence for your read. Once you identify the root cause, consider
adding it to the playbook (in this SKILL.md) so the next run catches
it automatically.

Headless invocation (permission mode)

/ci-debug REQUIRES Bash tool access — gh pr checks, gh run view, and gh api .../jobs/.../logs are how it fetches evidence. In headless claude -p runs (ci-sentinel, cron):

• Use --permission-mode acceptEdits — the headless "use tools without prompting" mode. The skill is read-only by design, so this grants nothing risky.
• NEVER dontAsk — it silently REFUSES permission-requiring tools (including Bash), so every analysis returns empty output with no error (#1862 Bug C).
• claude -p reports auth/permission failures as JSON on stdout, not stderr — capture and log both streams on failure (see shared/rules/cc-bare-auth-gotcha.md for the auth side).

CRITICAL guardrails

• NEVER auto-apply a fix. This skill proposes; the user approves. Wasted cycles from misclassification are exactly the failure mode we're hardening against.
• NEVER skip the fetch-log step. Without the actual error text, classification is guessing. If logs are gone (>90 days old, run deleted), say so and stop.
• Cite the exact memory entry (filename + section heading where relevant) so the user can verify the analogy.
• Each classification must have falsifying evidence — the signature must match. Don't pattern-match on hope.

When to invoke

• A specific CI run / job / PR check has failed and the user wants the diagnosis.
• The user pastes a run URL or "PR #N is failing — why?".

When NOT to invoke

• For broad "what's failing across the org?" — use /status instead.
• For application-level test failures (the playbook is CI-infra-specific, not e.g. pytest failures).
• Before reading the actual log — don't speculate.

Adding patterns

When a novel CI failure surfaces (the UNMATCHED report fires), add a row to the table above:

1. Signature: the unique-enough log line / metadata combination. Must be falsifiable.
2. Memory ref: name the memory file you wrote with the full incident write-up.
3. Proposed fix: the EXACT command line. No prose, no "you might want to". The skill produces commands the user runs; ambiguity defeats the purpose.

Commit the SKILL.md change as docs(ci-debug): add pattern #N (<short-name>). The /ork:ci-sentinel workflow picks up new patterns automatically on its next sweep — no separate plumbing needed.

Related Skills

• Composes with — /ork:ci-sentinel (the autonomous hourly trigger for this skill against open red PRs).
• Anti-pattern — manually re-running gh pr checks and eyeballing the logs across N repos. That's exactly the toil this skill kills.
• Upstream — /status for org-wide sweeps that surface WHICH PRs are red; this skill answers WHY.