Handle ci failure

Handle CI failure

Project CI context

cat .github/HANDLE_CI_FAILURE.md 2>/dev/null || echo "No .github/HANDLE_CI_FAILURE.md found. Derive the branch from \`git branch --show-current\`; find the check command from the Makefile, CI workflow, or package scripts (or ask the user); push without force unless you confirm the branch is rebased on a cycle. There are no project-specific failure patterns to consult."

The context file is the project's contract with this skill. It declares the working branch, the check command, the push policy, the commands behind any auto-fixable failures (formatter, regenerate, dependency tidy), and any project-specific failure patterns. Read it before matching.

Inputs

The user may provide:

• A bare run ID (24375242526).
• A View results: URL from a GitHub notification email — extract the numeric run ID with grep -oE 'actions/runs/[0-9]+' <<<"$input" | cut -d/ -f3.
• Nothing — list recent failures on the working branch and ask.

Workflow

1. Resolve the run ID

If the user gave one, use it. Otherwise:

gh run list --branch <branch> --status=failure --limit 5

Show the list with run IDs and ages, and ask which one to handle. Don't silently pick "most recent" — it drifts between sessions.

2. Extract failure context

"$CLAUDE_SKILL_DIR"/scripts/ci-failure-context.sh <run-id>

Output JSON shape:

{
  "run_id": "...", "run_url": "...", "workflow": "...",
  "branch": "...", "head_sha": "...", "conclusion": "failure",
  "failed_jobs": [
    {
      "id": 123, "name": "lint", "url": "...",
      "failed_step": "Run go vet",
      "annotations": [{"path": "...", "line": 15, "level": "failure", "message": "..."}],
      "log_tail": "... last 40 lines ..."
    }
  ]
}

Read the JSON once. Don't re-query gh for data the script already pulled.

3. Identify the real signal

annotations always contains runner-wrapper noise like {path: ".github", message: "Process completed with exit code 1."}. Filter path != ".github" when looking for root cause. Real signals point at real file paths with real messages. The log_tail is a fallback when annotations don't tell the whole story.

4. Match a known failure pattern

Generic patterns the skill handles directly:

Pattern	Reference
A checked-in file no longer matches a deterministic command's output — a formatting, code-generation/snapshot, or dependency-lockfile check fails	references/checked-in-file-drift.md
A tool the workflow needs isn't installed on the runner (command not found, No such file or directory)	references/workflow-tool-missing.md
A single source file flagged with one clear, unambiguous lint/vet/type error	Fix it directly — see the note below

Project-specific patterns live in the context file under a "Project failure patterns" heading (merge-conflict resolution, framework-specific failures, and the like). Consult those too — they point at project-local reference docs.

The direct-fix row is for cases where an annotation names a file and line and the fix is obvious (an unused import after a rename, a symbol that moved). If the fix needs judgment, spans several files, or you're guessing, treat it as open-ended debugging and escalate per the rules below rather than committing a speculative change.

Quick triage heuristic:

• Annotation points at a real source file with a compile/lint/vet message and one obvious fix → fix directly.
• A *-check / diff / snapshot step shows a checked-in file differs from generated output, or log_tail lists files that "would be reformatted" / "is not tidy" → regenerate and commit.
• log_tail shows command not found or No such file or directory for a tool → workflow tool missing.
• The failure matches a pattern described in the context file → follow that reference.

If none fits, stop and report the diagnosis. This skill is for mechanical, well-understood failure classes — not open-ended debugging. Adding a new generic pattern is cheap (drop a file in references/, add a row here); a project-specific one goes in the context file instead. Either should be done deliberately, after seeing the failure more than once.

5. Verify

Run the project's check command (the context file names it — e.g. make check, rake, npm test). Must be green before committing. If it isn't, the fix is incomplete — don't paper over with an inline suppression or by deleting tests.

6. Commit

Confirm the working branch matches what the context file specifies (git branch --show-current). If it doesn't, stop.

Follow the project's commit style: terse but informative, prose for a single complex fix, bullets for multiple distinct changes. Explain why, not what.

Stage explicitly — git add -u for tracked files, git add <path> for new ones. Never git add -A. Before committing, confirm any project instruction files the repo keeps out of version control (e.g. CLAUDE.md, CLAUDE.local.md) are not staged.

7. Push

A normal branch takes a plain push:

git push origin <branch>

Use --force-with-lease only when the context file says the working branch is rebased or force-pushed on a cycle (a syncing fork, for example), where a fast-forward push would be rejected:

git push --force-with-lease origin <branch>

Never plain --force. If --force-with-lease is rejected because an automated workflow updated the remote between your fetch and push, re-fetch and re-evaluate rather than overriding the lease.

8. Confirm green

gh run list --branch <branch> --limit 3

Watch the new run with gh run watch <run-id> --exit-status. If it fails again with a different error, start over at step 1.

Guardrails

• Never commit project instruction files the repo keeps out of version control (e.g. CLAUDE.md, CLAUDE.local.md). Check .gitignore if unsure.
• Never pass --no-verify or skip commit signing.
• Never use git add -A or git add .. Always stage explicitly.
• Never amend a pushed commit. New fix → new commit.
• Never force-push a protected branch (main, master, a release branch).

When to stop and escalate

• Failure doesn't match a generic pattern above or a project-specific one in the context file.
• Local reproduction gives a different error than CI reported — local state diverged; investigate before fixing.
• Check command won't go green after what looks like the right fix — don't commit a partial fix. Report what's left.
• Multiple jobs failed for unrelated reasons — handle them one at a time.