/git:triage

Unified issue and PR triage: scan, categorize, cross-link, and optionally act.

When to Use This Skill

Context

• Repo remote: !git remote -v
• Repo toplevel: !git rev-parse --show-toplevel
• Current branch: !git branch --show-current
• Recent merged PRs: !git log --merges --format='%h %s' --max-count=15

Parameters

Parse these from $ARGUMENTS (all optional):

Writes are disabled by default. --auto-close and --auto-merge still require a per-batch AskUserQuestion confirmation before any gh issue close or gh pr merge.

Execution

Execute this triage workflow:

Step 1: Resolve target repo and gather batches

Run the data-gathering script. It fetches issue/PR batches, computes age in days per item, categorizes each PR via the pure first-match table (over isDraft, mergeable, mergeStateStatus, reviewDecision, and statusCheckRollup[].conclusion), flags stale-candidate issues, carries each issue's title through so the output is readable without a second gh issue list pass, and extracts each PR's closing keywords:

bash "${CLAUDE_SKILL_DIR}/scripts/git-triage.sh" --home-dir "$HOME" --project-dir "$(pwd)" --type "$TYPE" --batch "$BATCH" --days-stale-issue "$STALE_ISSUE" --days-stale-pr "$STALE_PR"

Read the coverage keys first — they say what the batch actually covers:

--batch caps each fetch, and gh returns the N most recently created open items, so the default batch is the newest slice of the backlog. It therefore leaves out the oldest items, which are exactly the stale candidates. When TRUNCATED=true, say "triaged N of M" wherever the report states a count, and re-run with --batch set to the _TOTAL value for a full sweep. When TRUNCATED=unknown, report the coverage as unverified rather than complete. --oldest-first reorders only the fetched batch; it never changes which items were fetched.

STATUS=, ISSUE_COUNT=, and the ISSUES: block are the collector's own diagnostics (for example, a fetch that returned non-JSON), not GitHub issues: ISSUE_COUNT=0 beside ten issue blocks means the collector hit no problems. Read STATUS= to decide whether the data can be trusted, and the ISSUES: rows (present only when ISSUE_COUNT is above 0) for why.

Per item it emits ISSUE_<n>_TITLE / ISSUE_<n>_AGE_DAYS / ISSUE_<n>_REFS / ISSUE_<n>_COMMENTS / ISSUE_<n>_STALE_CANDIDATE and PR_<n>_CATEGORY / PR_<n>_AGE_DAYS / PR_<n>_CLOSES (plus the underlying enum fields). It also rolls up SYSTEMATIC_FAILURE_* groups (see Step 4). If --repo was provided, pass it through; the script reads the current repo from origin otherwise. Sort each set by age (oldest first if --oldest-first) and track one entry per item — via TodoWrite when the session has the task tools (see .claude/rules/agentic-permissions.md § Task-tool availability), otherwise as a checklist you keep in the response.

Step 2: Investigate each issue (skip if --type prs)

For each open issue in parallel (batch reads), gather evidence:

1. Extract referenced PR numbers from title, body, and comments (regex #(\d+)).
2. Check status of each referenced PR:

   gh pr view <n> --repo $REPO --json number,state,mergedAt,title
   

3. Search the codebase for concrete nouns in the issue (file paths, function names, resource names, command names) using Grep. Evidence that described artefacts exist (or no longer exist) feeds the categorization.
4. Compute age in days from updatedAt.

Step 3: Categorize each issue

Record the winning PR number (if any) with each implemented entry.

Step 4: Read each PR's category (skip if --type issues)

The script already categorized every PR in Step 1 — read PR_<n>_CATEGORY straight from its output. The category is a pure first-match over the enum fields (the script owns this deterministic table, top to bottom):

If a PR comes back as uncategorized (e.g. mergeStateStatus/mergeable both UNKNOWN), trigger a fresh view and re-run the script, or inspect:

gh pr view <n> --repo $REPO --json mergeable,mergeStateStatus

Systematic failures. When ≥2 bot-authored needs-fix PRs share an identical failing-check signature, the script groups them under SYSTEMATIC_FAILURE_<k>_SIGNATURE (the sorted |-joined failed check names) and SYSTEMATIC_FAILURE_<k>_PRS (the PR list); SYSTEMATIC_FAILURE_COUNT holds the number of groups. These almost always have one shared root cause — e.g. Dependabot can't update bun.lock, so every npm-bump PR fails the bun install --frozen-lockfile step before lint/typecheck/tests run, and "Lint FAILURE / Type Check FAILURE" is misleading (nothing was linted). For each group, read the install step's log once before assuming code defects:

gh pr checks <n> --repo $REPO --json name,state,conclusion,detailsUrl
gh run view <run-id> --repo $REPO --log-failed

Diagnose the shared cause once and present a single grouped row (Step 6) / blocker (Step 8) instead of N independent needs-fix PRs.

For bot PRs whose checks may belong to a pre-rebase SHA, and for pin PRs that smuggle a minor bump, see REFERENCE.md.

Step 5: Cross-link issues and PRs

• For each implemented issue (from Step 3's judgment), attach the merged PR number.
• For each ready-to-merge PR, read the closing keywords the script extracted in PR_<n>_CLOSES and attach those issue numbers.
• For each needs-fix / needs-rebase / changes-requested PR, note the referenced issues so the report can suggest which issues remain blocked.

Step 6: Present the prioritized queue

Ordering: quick wins first. Use AskUserQuestion only when the user will need to pick what to act on next.

Print a status table (one row per item) grouped by category:

## Issues (N of M open, triaged)

| # | Age | Title | Category | Cross-link |
|---|-----|-------|----------|------------|
| 42 | 120d | Remove legacy X | implemented | PR #99 (merged) |
| 17 | 210d | Deprecated docs | stale | — |
| 13 | 14d  | Add retry logic | still-valid | — |

## PRs (N of M open, triaged)

| # | Age | Title | Category | Cross-link |
|---|-----|-------|----------|------------|
| 101 | 2d  | feat(api): X | ready-to-merge | closes #55 |
| 102 | 18d | fix(auth): Y | needs-fix | — |
| 103 | 45d | refactor(ui) | stale | — |

Step 7: Optional writes (guarded)

If --auto-close was set and any issue is implemented or stale, ask before acting:

AskUserQuestion("Close N issues?", options=[
  "Yes — close all implemented + stale",
  "Implemented only",
  "Stale only",
  "No, report only"
])

For each selected issue:

gh issue close <n> --repo $REPO --comment "Closing as <category>.

Evidence: <short summary + cross-link PR>

Triaged by /git:triage on <date>."

If --auto-merge was set and any PR is ready-to-merge, ask similarly. Merge with:

gh pr merge <n> --repo $REPO --squash --auto

(--squash is the repo default for this project; for other repos, read gh repo view --json squashMergeAllowed,mergeCommitAllowed,rebaseMergeAllowed and pick the first allowed strategy.)

Step 8: Synthesize the backlog report

After per-item actions, emit a structured summary:

1. Actions taken — count of issues closed, PRs merged, with numbers.
2. Quick wins — still-valid issues whose body suggests <30 min of work (single file, doc edit, config tweak).
3. Blockers — changes-requested PRs and issues blocked on external factors. For each SYSTEMATIC_FAILURE_* group, emit one "systematic failure — likely shared root cause" line naming the PRs and the shared check signature, rather than N independent needs-fix entries.
4. Decisions needed — still-valid issues whose body ends in a question or "how should we…".
5. Handoff recommendations per category:

Post-actions

• Print a one-line summary: Triaged N of M open issues (X closed), P of Q open PRs (Y merged). See report above.
• If any writes were gated behind confirmation that the user declined, leave the items open and note "no writes — report only".
• Remind the user they can re-run with --type prs or --type issues to focus the sweep.

Agentic Optimizations

See Also

• /git:fix-pr — fix needs-fix PRs
• /git:pr-feedback — address changes-requested PRs
• /git:issue — work on a still-valid issue
• /git:issue-manage — admin ops on issues
• /git:issue-hierarchy — sub-issue relationships
• /git:conflicts — resolve needs-rebase PRs