issue-triage

name	issue-triage
description	3-phase issue backlog management with audit, deep analysis, and validated triage actions. Use when triaging GitHub issues, sorting bug reports, cleaning up stale tickets, or detecting duplicate issues. Args: 'all' to analyze all, issue numbers to focus (e.g. '42 57'), 'en'/'fr' for language, no arg = audit only.
tags	["github","issue","triage","maintainer","multi-agent"]
effort	medium

Issue Triage

3-phase workflow for maintainers: automated audit of all open issues, opt-in deep analysis via parallel agents, and validated triage actions (comments, labels, closures).

When to Use This Skill

Skill	Usage	Output
/issue-triage	Sort, analyze, and act on an issue backlog	Triage tables + analysis + executed actions
/pr-triage	Sort, review, and comment on a PR backlog	Triage table + reviews + posted comments

Triggers:

• Manually: /issue-triage or /issue-triage all or /issue-triage 42 57
• Proactively: when >10 open issues without label, or stale issues >30 days detected

---

Language

• Check the argument passed to the skill
• If en or english → tables and summary in English
• If fr, french, or no argument → French (default)
• Note: GitHub comments and labels (Phase 3) are ALWAYS in English (international audience)

---

Configuration

Thresholds used throughout the workflow. Edit to match your project:

Parameter	Default	Description
staleness_days	30	Days without activity before flagging as stale
very_stale_days	90	Days without activity before flagging as very stale
jaccard_threshold	60%	Minimum Jaccard similarity to flag two issues as duplicates
closed_compare_count	20	Number of recent closed issues to compare for duplicate detection
open_limit	100	Maximum open issues to fetch and analyze

---

Preconditions

git rev-parse --is-inside-work-tree
gh auth status

If either fails, stop and explain what is missing.

---

Phase 1 — Audit (always executed)

Data Gathering (parallel commands)

# Repo identity
gh repo view --json nameWithOwner -q .nameWithOwner

# Open issues (exclude PRs, limit 100)
gh issue list --state open --limit 100 \
  --json number,title,author,createdAt,updatedAt,labels,body,comments,assignees,milestone

# Recent closed issues (for duplicate detection)
gh issue list --state closed --limit 20 \
  --json number,title,body,labels,stateReason

# Open PRs (bodies for cross-reference detection)
gh pr list --state open --limit 50 --json number,title,body

# Collaborators (to distinguish reporter types)
gh api "repos/{owner}/{repo}/collaborators" --jq '.[].login'

Collaborators fallback: if gh api .../collaborators returns 403/404:

# Extract authors from last 10 merged PRs
gh pr list --state merged --limit 10 --json author --jq '.[].author.login' | sort -u

If still ambiguous, ask via AskUserQuestion.

Note: comments field in gh issue list --json comments returns the count, not content. For Phase 2, fetch full content separately: gh issue view {num} --json comments.

Analysis Dimensions

Run all 6 dimensions for each open issue:

1. Categorization

Classify each issue by reading title + first 200 chars of body:

Category	Label	Criteria
Bug	bug	Describes broken behavior, unexpected output, crash
Feature Request	enhancement	Asks for new functionality
Question / Support	question	User asking how something works
Documentation	documentation	Missing or incorrect docs
Out of Scope	wontfix	Clearly outside project boundaries
Unclear	needs-info	Body empty, too vague to categorize

If body is empty → category is always Unclear (never assume).

2. Cross-reference to PRs

Scan each open PR body for references to the issue number:

• Patterns: fixes #N, closes #N, resolves #N, fix #N, close #N (case-insensitive, N = issue number)
• Use regex locally on the body fields already fetched — do NOT make N additional API calls
• If found: flag issue as "PR-linked" with PR number

3. Duplicate Detection via Jaccard Similarity

Algorithm (self-contained — no external library):

For each open issue, compute Jaccard similarity against all other open issues AND the 20 most recent closed issues.

Step 1 — Normalize title + first 300 chars of body:
  - Lowercase the full text
  - Strip category prefixes: "feat:", "fix:", "bug:", "chore:", "docs:", "test:", "refactor:"
  - Remove punctuation: .,!?;:'"()[]{}-_/\@#

Step 2 — Tokenize:
  - Split on whitespace
  - Remove stop words: the a an is in on to for of and or with this that it can not no be
  - Remove tokens shorter than 3 characters

Step 3 — Compute Jaccard:
  tokens_A = set of tokens from issue A
  tokens_B = set of tokens from issue B
  jaccard = |tokens_A ∩ tokens_B| / |tokens_A ∪ tokens_B|

Step 4 — Flag:
  - If jaccard >= 0.60: mark as potential duplicate
  - Report: "Similar to #N (Jaccard: 0.72)"
  - Keep the OLDER issue as canonical; newer = duplicate candidate

Jaccard is computed at runtime using the fetched data — no API calls beyond Phase 1 gather.

4. Risk Classification

Assign Red / Yellow / Green based on signals in title + body:

Level	Color	Criteria
Critical	Red	Security vulnerability, data loss, regression blocking users, crash in production
Needs Attention	Yellow	Missing validation, performance degradation, breaking change undocumented, Unclear with no response for >7 days
Normal	Green	Everything else

5. Staleness

Status	Criterion
Active	Updated within 30 days
Stale	No activity 30–90 days
Very Stale	No activity >90 days

Use updatedAt field. Staleness does NOT depend on comments count — a commented-on issue with old updatedAt is still stale.

6. Recommendations

One recommended action per issue:

Situation	Action
Category = Unclear, body empty	Comment requesting details
Jaccard >= 0.60 with known issue	Close as duplicate, link original
Very stale + no assignee	Comment requesting status, suggest close
Risk = Red	Pin to top of triage, escalate immediately
Category = OOS	Close with explanation
PR-linked	No action needed (tracked via PR)
Normal + labeled	No action needed

Output — Triage Tables

## Open Issues ({count})

### Critical — Immediate Attention (Risk: Red)
| # | Title | Category | Reporter | Days Open | Action |
|---|-------|----------|----------|-----------|--------|

### PR-Linked (tracked in open PRs)
| # | Title | Category | PR | Days Open |
|---|-------|----------|----|-----------|

### Active Issues
| # | Title | Category | Labels | Reporter | Days | Action |
|---|-------|----------|--------|----------|------|--------|

### Duplicate Candidates
| # | Title | Similar To | Jaccard | Action |
|---|-------|------------|---------|--------|

### Stale Issues
| # | Title | Category | Last Activity | Reporter | Action |
|---|-------|----------|---------------|----------|--------|

### Summary
- Total open: {N}
- Critical (Red): {count}
- PR-linked: {count}
- Duplicate candidates: {count}
- Stale (30–90d): {count}
- Very stale (>90d): {count}
- Unlabeled: {count}
- Recommended actions: {comment: N, label: N, close: N}

0 issues → display No open issues. and stop.

Protection rules (apply to all phases):

• Never close an issue authored by a collaborator without explicit user confirmation
• Never re-label an issue that already has labels (only add missing labels)
• If body is empty → always request details before any other action
• Never auto-close a Red issue without user confirmation

Automatic Copy

After displaying the triage tables, copy to clipboard using platform-appropriate command:

UNAME=$(uname -s)
if [ "$UNAME" = "Darwin" ]; then
  pbcopy <<'EOF'
{full triage tables}
EOF
elif command -v xclip &>/dev/null; then
  echo "{full triage tables}" | xclip -selection clipboard
elif command -v wl-copy &>/dev/null; then
  echo "{full triage tables}" | wl-copy
elif command -v clip.exe &>/dev/null; then
  echo "{full triage tables}" | clip.exe
fi

Confirm: Triage tables copied to clipboard. (EN) / Tableaux copiés dans le presse-papier. (FR)

---

Phase 2 — Deep Analysis (opt-in)

Issue Selection

If argument passed:

• "all" → all issues with recommended actions
• Numbers ("42 57") → only those issues
• No argument → propose via AskUserQuestion

If no argument, display:

question: "Which issues do you want to analyze in depth?"
header: "Deep Analysis"
multiSelect: true
options:
  - label: "All ({N} issues with recommended actions)"
    description: "Launch parallel analysis agents for each actionable issue"
  - label: "Critical only ({M} Red issues)"
    description: "Focus on high-risk issues requiring immediate action"
  - label: "Duplicate candidates ({K} issues)"
    description: "Verify Jaccard similarity with full body + comments"
  - label: "Stale only ({J} stale issues)"
    description: "Decide which stale issues to close vs. revive"
  - label: "Skip"
    description: "Stop here — audit only"

If "Skip" → end workflow.

Executing Analysis

For each selected issue, launch an analysis agent via Task tool in parallel:

subagent_type: general
model: sonnet
prompt: |
  Analyze GitHub issue #{num}: "{title}"

  **Metadata**: Category={category}, Risk={risk}, Days open={days}, Labels={labels}
  **Reporter**: @{author} ({collaborator? "collaborator" : "external"})
  **Assignees**: {assignees or "none"}

  **Body**:
  {body}

  **Comments** (fetch via: gh issue view {num} --json comments):
  {comments[].body — truncate at 5000 chars total}

  **Duplicate candidates**: {jaccard_results or "none found"}
  **Linked PRs**: {pr_refs or "none"}

  Tasks:
  1. Verify the category assigned in Phase 1 (correct? suggest alternative if not)
  2. If duplicate candidate: confirm or deny similarity with rationale
  3. If Unclear/needs-info: identify exactly what information is missing
  4. Suggest the most appropriate action with exact text if a comment is needed
  5. Estimate effort to fix if it's a Bug or Feature Request (XS/S/M/L/XL)

  Return structured output:
  ### Verification
  ### Duplicate Analysis
  ### Missing Information
  ### Recommended Action
  ### Effort Estimate

Fallback if parallel agents unavailable: run analysis sequentially, one issue at a time. Notify user: Running sequential analysis (parallel agents not available).

Fetch full comments via:

gh issue view {num} --json comments --jq '.comments[].body'

Aggregate all reports. Display a summary after all analyses complete.

---

Phase 3 — Actions (mandatory validation)

Draft Generation

For each analyzed issue, generate the appropriate action using the template templates/issue-comment.md.

3 action types:

Type	Command	When
Comment	gh issue comment {num} --body-file -	Needs info, stale ping, OOS explanation
Label	gh issue edit {num} --add-label "{label}"	Unlabeled issue with clear category
Close	gh issue close {num} --reason "not planned"	Duplicate, OOS, very stale

Rules:

• Language for comments: English (international audience)
• Labels added: use existing repo labels only (fetch with gh label list)
• Close reason: "not planned" for OOS/duplicate, "completed" only if a fix was merged
• Never post a comment AND close in the same action without user seeing both drafts
• Always attach a comment when closing (explain why)

Display and Validation

Display ALL drafted actions in format:

---
### Draft — Issue #{num}: {title}

**Action**: {Comment / Label / Close + Comment}
**Reason**: {1 sentence}

{full comment text if applicable}

---

Then request validation via AskUserQuestion:

question: "These actions are ready. Which ones do you want to execute?"
header: "Execute Triage Actions"
multiSelect: true
options:
  - label: "All ({N} actions)"
    description: "Execute all drafted triage actions"
  - label: "Issue #{x} — {title_truncated} ({action_type})"
    description: "Execute only this action"
  - label: "None"
    description: "Cancel — execute nothing"

(Generate one option per issue + "All" + "None")

Execution

For each validated action:

# Comment
gh issue comment {num} --body-file - <<'TRIAGE_EOF'
{comment}
TRIAGE_EOF

# Label
gh issue edit {num} --add-label "{label}"

# Close with comment
gh issue comment {num} --body-file - <<'TRIAGE_EOF'
{close comment}
TRIAGE_EOF
gh issue close {num} --reason "not planned"

Confirm each action: Action executed on issue #{num}: {title}

If "None" → No actions executed. Workflow complete.

---

Edge Cases

Situation	Behavior
0 open issues	Display No open issues. + stop
Body empty	Category = Unclear, action = request details, never assume
Collaborator as reporter	Protect from auto-close, flag explicitly in table
Jaccard inconclusive (0.55–0.65)	Flag as "possible duplicate — verify manually"
Label not in repo	Skip label action, notify user to create the label first
Issue already closed during workflow	Skip silently, note in summary
gh api .../collaborators 403/404	Fallback to last 10 merged PR authors
Parallel agents unavailable	Run sequential analysis, notify user
Very large body (>5000 chars)	Truncate to 5000 chars with [truncated] note
Milestone assigned	Include in table, never close milestoned issues without confirmation

---

Notes

• Always derive owner/repo via gh repo view, never hardcode
• Use gh CLI (not curl GitHub API) except for collaborators list
• comments in gh issue list --json comments = count only; full content requires gh issue view {num} --json comments
• Never execute any action without explicit user validation in chat
• Drafted actions must be visible BEFORE any gh issue comment or gh issue close
• Jaccard is computed locally — no external API, no library, pure set operations on fetched data
• Signature on all comments: *Triaged via Claude Code /issue-triage*

---

Related: /pr-triage

	/issue-triage	/pr-triage
Scope	Issue backlog	PR backlog
Use when	Catching up on reporter feedback, periodic issue cleanup	Catching up after PR accumulation
Phases	3 (audit + deep analysis + actions)	3 (audit + deep review + comments)
Agents	Parallel sub-agents per issue	Parallel sub-agents per PR
Duplicate detection	Jaccard similarity on title+body	File overlap % between PRs
Actions	Comment / label / close	GitHub review comment
Validation	AskUserQuestion before executing	AskUserQuestion before posting

Decision rule: use /issue-triage for issue backlog management, /pr-triage for code review backlog.