Triage all open GitHub issues — correct the issue type (Bug/Feature/Task), set the Priority field and the lane/* label, link related issues into attackable clusters, and add codebase-backed detail comments to thin tickets. Use when the user says "triage", "triage issues", or "clean up the backlog".
Instrucciones de origen · Vista previa de solo lectura
name
triage
description
Triage all open GitHub issues — correct the issue type (Bug/Feature/Task), set the Priority field and the lane/* label, link related issues into attackable clusters, and add codebase-backed detail comments to thin tickets. Use when the user says "triage", "triage issues", or "clean up the backlog".
argument-hint
[optional issue numbers to limit scope]
allowed-tools
Bash, Read, Grep, Glob
Issue Triage
Triage every open issue (or only $ARGUMENTS if given). Read-only on code; writes go to GitHub only. Never close an issue, never remove a label you cannot justify in the comment trail, never edit issue bodies except through the log-issue backfill, which preserves all information.
PRs NEVER close issues. Beyond never closing issues yourself, triage must prevent GitHub from doing it: any open PR whose body, title, or commits carry a closing keyword (Fixes/Closes/Resolves/Fixed/Closed #n) referencing an issue will auto-close it on merge. During every triage run, scan open PR bodies (gh pr list --state open --json number,body,title) for those keywords; on a hit, remove the keyword from the PR body (gh pr edit <pr> --body) and flag it in the final report. Issues close only through release verification — see the log-issue skill for the full rule.
The published atlas (scripts/issues/generate_issue_report.py) reads the Priority field and the lane/* label straight off each issue. Getting those two right (Steps 3 and 4) is what makes the site correct — there is no derivation to second-guess.
Comment voice — applies to every comment posted. Comments are mechanical triage records attached to the ticket, not replies to the reporter. Never address anyone: no "you", no "thanks for reporting", no "could you confirm", no greetings or sign-offs. State facts about the defect and the code in impersonal declaratives. Where information is missing, record what is missing as a fact ("Repro command not specified"), never as a request.
Step 1: Fetch
gh issue list --state open --limit 200 --json number,title,body,labels,issueType,comments
Task — chore, refactor, docs, CI, tech-debt with no user-visible defect.
Set with gh issue edit <n> --type <Bug|Feature|Task> (fallback if the flag is unavailable: gh api graphql with the updateIssueIssueType mutation; list type IDs via the org's issueTypes).
Step 3: Priority — the GitHub Priority field, not a label
Priority is the org-level Priority issue field. The showstopper and critical labels are deleted; never re-create them. gh issue edit cannot set a field — use scripts/issues/set_priority.sh <issue> <option>.
showstopper — a regression on main that has not been released yet, or a recent regression in a release.
critical — seriously impacting accuracy or the usefulness of the tool. A confirmed false positive/negative already sitting in a release is critical, not showstopper: there is nothing left to hold back.
normal — a problem that impacts the usefulness of the tool.
low — don't worry about it for now.
Every open issue carries exactly one value. An issue with none shows on the site as No priority set — a triage defect.
Domain labels stay labels: false-positive, false-negative, spec-violation, ignored-test, tech-debt. An accuracy issue missing its false-* label is a triage defect — gh issue edit <n> --add-label ....
fixed-on-main means fixed but not yet verified in a release — leave the issue open, keep its Priority, and note in the cluster comment that it awaits release verification.
Apply fixed-on-main yourself when the evidence is strong. If the reported defect is gone on mainand a test covers it — the named failing test now passes, or a test asserting the corrected behaviour exists and is not #[ignore]d — add the label and comment the evidence: commit/PR that landed it, and the test name and path that pins it. The ticket then closes after release verification. Never apply it on code reading alone; no covering test means no label.
Priority flows up the dependency tree. A root cause is never lower-priority than anything it blocks. After clustering (Step 5), re-walk each tree from the leaves and raise ancestors as needed.
Step 4: Lane — exactly one lane/* label
The published lane is the issue's lane/<id> label. Nothing is inferred from title or body.
lane/accuracy — wrong clusters reported or missed.
Judge the lane by where the fix lands, not by the issue's vocabulary. Apply with gh issue edit <n> --add-label lane/<id> --remove-label lane/<wrong>; two lane labels or none is a triage defect, and none publishes the issue as Unassigned.
Step 5: Cluster related issues
Group issues sharing a root cause, pipeline stage (parse → normalize → fingerprint → cluster → LSH → embeddings → fuse → rank → render), or file/module. For each cluster:
Pick the most upstream issue — the root cause — as the anchor.
Mark every dependency. Any issue that depends on another MUST be recorded as blocked by it (sub-issue via gh api graphqladdSubIssue, or the blocked-by relationship where available) — this directionality is how the backlog gets attacked in order. Loose thematic relations with no blocking edge get comment cross-links only.
Post one identical comment on every member: **Cluster: <name>** — #a #b #c. Root cause: #a — fix it first; the rest may collapse or shrink once it lands. Fix order: #a → #b → #c. Always steer work toward the root, not the symptoms.
Retire edges the root outgrew. Before recording anything, walk the existing blocked-by edges: any edge whose blocking issue is fixed-on-main (or closed) is stale — the work it gated has landed. Remove it (gh api graphqlremoveBlockedBy) and comment on the unblocked issue saying which edge went, that the blocker is fixed on main, and what remains. Where the dependent still needs the fix to ship, say so and keep a cross-link in place of the edge — a release-verification wait is not a blocked backlog item.
Step 6: Enrich thin tickets
A ticket is thin if a dev could not start work from it alone. For each, investigate the codebase (Grep/Read the named stage, check spec IDs, run the referenced test if cheap) and comment with facts only — no speculation:
Affected files with paths and line numbers.
Suspected root cause and the spec ID ([GROUP-TOPIC]) it violates.
Repro: exact command or failing-test name, expected vs actual.
Structure enrichment comments with the canonical section names defined in the log-issue skill (.agents/skills/log-issue/SKILL.md — "Canonical issue body") so humans and agents read issues and comments the same way; do not restate the section list here. That skill also owns the plain-language rule for human sections — apply it to enrichment comments too. The issue atlas excerpt (plain_excerpt in scripts/issues/generate_issue_report.py) is drawn from the ## TL;DR section — a backfill changes the site card text, so a missing or empty TL;DR on an open issue is a triage finding.
If investigation uncovers a new accuracy defect, the CLAUDE.md quarantine rule applies — stop triage and follow it.
Step 7: Report
End with a table: issue #, type set, Priority set, lane label, other labels added/removed, cluster, blocked-by, comment posted (y/n). List the root-cause issues separately as the recommended attack order. Flag anything you could not resolve (e.g. type API rejected, ambiguous severity, dependency API unavailable) for the user to decide.
