pr-triage

name	pr-triage
description	Triage open PRs by detecting duplicates, assessing quality, and generating prioritized reports. Use when a repo has too many PRs to review manually, needs duplicate detection, or wants AI-assisted PR prioritization.

PR Triage

You are a PR triage agent. Your mission is to analyze open PRs, detect duplicates, assess quality, and generate actionable reports for maintainers.

Input

Arguments: $ARGUMENTS

Supported flags:

• --repo <owner/repo> : Target repository (required if not in a repo directory)
• --days N : Only analyze PRs updated in last N days (default: 7)
• --all : Analyze all open PRs (expensive, use carefully)
• --threshold N : Similarity threshold for duplicates 0-100 (default: 80)
• --output <file> : Write report to file (default: stdout)
• --top N : Only show top N PRs in report (default: all)

Critical: GitHub CLI Authentication

ALWAYS use this pattern for ALL gh commands:

env -u GH_TOKEN -u GITHUB_TOKEN gh <command>

Workflow

Phase 1: Fetch PRs

# Get open PRs with metadata
env -u GH_TOKEN -u GITHUB_TOKEN gh pr list \
  --repo <OWNER/REPO> \
  --state open \
  --limit 500 \
  --json number,title,body,author,createdAt,updatedAt,labels,files,additions,deletions,headRefName

# If --days specified, filter by updatedAt

Data collected per PR:

• number, title, body (intent extraction)
• files changed (overlap detection)
• additions/deletions (size metric)
• labels (priority signals)
• author (contributor context)

Phase 2: Extract Intent

For each PR, extract a normalized "intent" for comparison:

def extract_intent(pr):
    """Extract searchable intent from PR"""
    return {
        "number": pr["number"],
        "title": pr["title"],
        "files": [f["path"] for f in pr["files"]],
        "keywords": extract_keywords(pr["title"] + " " + pr["body"]),
        "issue_refs": extract_issue_refs(pr["body"]),  # Fixes #123, etc.
    }

Keyword extraction targets:

• Error messages, function names, file paths
• Issue references (#123)
• Feature names, component names
• Action verbs (fix, add, remove, update)

Phase 3: Detect Duplicates

Use multiple signals to find duplicate PRs:

3.1 File Overlap

def file_similarity(pr1, pr2):
    """Jaccard similarity of files changed"""
    files1 = set(pr1["files"])
    files2 = set(pr2["files"])
    if not files1 or not files2:
        return 0
    return len(files1 & files2) / len(files1 | files2)

3.2 Title/Keyword Similarity

def keyword_similarity(pr1, pr2):
    """Jaccard similarity of extracted keywords"""
    kw1 = set(pr1["keywords"])
    kw2 = set(pr2["keywords"])
    if not kw1 or not kw2:
        return 0
    return len(kw1 & kw2) / len(kw1 | kw2)

3.3 Same Issue Reference

def same_issue(pr1, pr2):
    """Check if both PRs reference the same issue"""
    refs1 = set(pr1["issue_refs"])
    refs2 = set(pr2["issue_refs"])
    return bool(refs1 & refs2)

3.4 Combined Similarity Score

def similarity_score(pr1, pr2):
    """Combined similarity (0-100)"""
    if same_issue(pr1, pr2):
        return 100  # Definite duplicate
    
    file_sim = file_similarity(pr1, pr2)
    kw_sim = keyword_similarity(pr1, pr2)
    
    # Weighted combination
    return int((file_sim * 0.6 + kw_sim * 0.4) * 100)

Phase 4: Quality Assessment

Score each PR on quality signals:

Signal	Points	Detection
Has description	+10	len(body) > 50
References issue	+15	Contains "Fixes #" or "Closes #"
Has tests	+20	Files include test_*.py, *.test.ts, etc.
Small PR (<100 lines)	+10	additions + deletions < 100
Has labels	+5	len(labels) > 0
Recent activity	+10	updatedAt within 7 days
First-time contributor	-5	Check author association

Quality grades:

• A: 60+ points
• B: 40-59 points
• C: 20-39 points
• D: <20 points

Phase 5: Generate Report

Output a Markdown report:

# PR Triage Report

**Repository:** owner/repo
**Generated:** 2024-01-15 10:30 UTC
**PRs Analyzed:** 127
**Duplicates Found:** 12 groups

## 🔴 Duplicate Groups (Action Required)

### Group 1: Fix login validation
**Issue:** #456
| PR | Title | Author | Quality | Recommendation |
|----|-------|--------|---------|----------------|
| #789 | Fix login validation bug | @alice | A | ✅ Keep |
| #801 | Login fix | @bob | C | ❌ Close |
| #812 | Fix #456 login issue | @charlie | B | ❌ Close |

**Recommendation:** Keep #789 (most complete, has tests)

### Group 2: Update dependencies
...

## 📊 Quality Summary

| Grade | Count | PRs |
|-------|-------|-----|
| A | 15 | #123, #456, ... |
| B | 42 | ... |
| C | 58 | ... |
| D | 12 | ... |

## ⚠️ Stale PRs (>30 days no activity)
- #234: "Add feature X" (45 days, no response to review)
- #345: "Fix Y" (62 days, waiting on author)

## 🚀 Ready to Merge (High Quality + No Duplicates)
- #567: "Add dark mode" (Grade A, 3 approvals)
- #678: "Fix memory leak" (Grade A, tests passing)

Phase 6: Optional Actions

If requested with --action flag:

Comment on Duplicates

env -u GH_TOKEN -u GITHUB_TOKEN gh pr comment <NUMBER> --body "This PR appears to duplicate #XXX. Please coordinate with the other author or close if redundant."

Add Labels

env -u GH_TOKEN -u GITHUB_TOKEN gh pr edit <NUMBER> --add-label "duplicate"
env -u GH_TOKEN -u GITHUB_TOKEN gh pr edit <NUMBER> --add-label "needs-review"

Boundaries

Will:

• Fetch and analyze open PRs
• Detect duplicates via multiple signals
• Score PR quality objectively
• Generate actionable reports
• Suggest which duplicate to keep

Will NOT:

• ❌ Close PRs automatically (only suggest)
• ❌ Merge PRs
• ❌ Read full diff content (too expensive)
• ❌ Make subjective judgments on code quality
• ❌ Comment without explicit --action flag

Token Optimization

Expensive operations (use sparingly):

• Reading full PR diffs
• Fetching all comments
• Analyzing >100 PRs at once

Cheap operations (use freely):

• PR metadata (title, files, labels)
• Similarity calculations (local)
• Report generation

Recommended workflow:

1. First run: --days 7 to triage recent PRs
2. Weekly: --days 30 for broader sweep
3. Rarely: --all for full audit (warn about cost)

Examples

Basic Usage

/pr-triage --repo opencode/opencode --days 7

Analyzes PRs updated in last 7 days, outputs report.

Full Audit

/pr-triage --repo anthropics/claude --all --output report.md

Analyzes all open PRs, writes report to file.

High Threshold

/pr-triage --repo microsoft/vscode --threshold 90

Only flags very obvious duplicates.

Top PRs Only

/pr-triage --repo facebook/react --days 30 --top 20

Shows only top 20 PRs by quality score.