manastalukdar Issue Triage
Keep the issue backlog clean and actionable. Scans open issues, identifies duplicates, extracts priority signals, and produces a structured queue the team (or other loops) can consume.
Usage
/issue-triage # triage all open issues
/issue-triage --repo <owner/repo> # triage a specific repo
/issue-triage --since <date> # only issues opened since date (ISO 8601)
/issue-triage --label <label> # filter to a specific label
/issue-triage --report-only # produce report without applying any labels
Run on a loop for continuous triage:
/loop 2h /issue-triage
/loop 1d /issue-triage --since yesterday
Behavior
Step 1 — Load open issues
# Fetch open issues with metadata
gh issue list --state open --limit 100 \
--json number,title,body,labels,createdAt,updatedAt,comments,reactions,assignees \
| jq 'sort_by(.reactions.totalCount) | reverse'
If 0 issues → report "No open issues" and exit.
Step 2 — Deduplicate
For each issue, check for existing duplicates:
# Search for similar titles (keyword extraction)
KEYWORDS=$(echo "$ISSUE_TITLE" | tr '[:upper:]' '[:lower:]' | grep -oE '\b\w{4,}\b' | head -5 | tr '\n' ' ')
gh issue list --state open --search "$KEYWORDS" --json number,title | head -5
If a likely duplicate is found:
- Add
duplicatelabel to the newer issue - Comment with the original issue number
- Do not close — human confirms duplicates
Step 3 — Extract priority signals
For each issue, score priority (1–5) from:
| Signal | Weight | How to detect |
|---|---|---|
| Reactions (👍, ❤️) | High | reactions.totalCount |
| Comments count | Medium | comments length |
| Age (older = more urgent) | Medium | createdAt |
| Bug label present | High | labels[] |
| "crash", "data loss", "security" in title | Critical | keyword grep |
| Assignee present | Low (already tracked) | assignees |
| Linked PR or "fix:" commit | Reduces priority | search commits |
# Check for critical keywords
echo "$ISSUE_TITLE $ISSUE_BODY" | grep -iE "crash|data.loss|security|vulnerability|CVE|regression|production|outage" && PRIORITY=1
Step 4 — Suggest labels
Based on issue content, suggest labels:
| Content signal | Suggested label |
|---|---|
| Stack trace / error message | bug |
| "would be nice", "could you add" | enhancement |
| "docs", "documentation", "typo" | documentation |
| First comment from new contributor | good first issue |
| No steps to reproduce | needs-repro |
| Ambiguous or unclear ask | needs-clarification |
| Priority score 1–2 | priority: high |
# Apply suggested labels (unless --report-only)
gh issue edit <number> --add-label "<label>"
Only add labels — never remove existing human-assigned labels.
Step 5 — Produce structured output
ISSUE TRIAGE — <repo> (2026-06-30 10:45 UTC)
Open: 23 issues | New since last run: 4 | Needs human: 2
TOP 5 (by priority score)
#487 [BUG] Crash on export with files > 100MB P1 7👍 3 comments
Labels: bug, needs-repro | Suggested: priority:high
Action: Request reproduction steps (no steps provided)
#412 [FEAT] Dark mode support P2 24👍 12 comments
Labels: enhancement | Suggested: add to roadmap discussion
Action: None — actively discussed
#501 Security: SSRF via webhook URL parameter P1 1👍 0 comments
Labels: (none) | Suggested: bug, security, priority:high
Action: Label applied; escalate to security review
DUPLICATES FOUND (1)
#499 is likely a duplicate of #412 (dark mode) — duplicate label added
NEEDS HUMAN (2)
#489 Ambiguous feature request — unclear acceptance criteria
#493 Potential duplicate of #234 but different enough to confirm manually
NOISE / IGNORED (5)
#480–#484 Dependabot alerts (handled by dep-sweep)
State written to: .claude/cache/issue-triage/state.json
Step 6 — Update state
mkdir -p .claude/cache/issue-triage
jq -n \
--arg ts "$(date -u +%Y-%m-%dT%H:%M:%SZ)" \
--arg open "$OPEN_COUNT" \
--arg new "$NEW_COUNT" \
'{last_run: $ts, open_count: $open, new_since_last_run: $new}' \
> .claude/cache/issue-triage/state.json
Edge Cases
- No GitHub auth: Report the error; suggest
gh auth login. - Large backlog (100+ issues): Process in batches of 50; note total count in report.
- Private repo: Works the same;
ghhandles auth. - No labels configured in repo: Skip label suggestions; only produce priority report.
- Issue already has
blockedorwontfix: Skip — do not re-triage human decisions.
Token Optimization
Expected range: 400–1,200 tokens; 50–100 tokens (no issues, early exit)
Patterns used: Bash (gh CLI), early exit (0 issues), progressive disclosure (summary then top-5 then full list on request), caching (state avoids re-reading unchanged issues)
Caching: Stores last-run timestamp and issue count in .claude/cache/issue-triage/state.json. On resume, reads only issues updated since last_run.
Early exit: 0 open issues → single-line report and exit.