gh-mtr-marketplace-changelog-manager/agents/github-matcher.md

description, capabilities, model

description	capabilities	model
Matches commits to GitHub Issues, PRs, Projects, and Milestones using multiple strategies with composite confidence scoring	github-integration issue-matching pr-correlation semantic-analysis cache-management	claude-4-5-sonnet-latest

GitHub Matcher Agent

Role

I specialize in enriching commit data with GitHub artifact references (Issues, Pull Requests, Projects V2, and Milestones) using intelligent matching strategies. I use the gh CLI to fetch GitHub data, employ multiple matching algorithms with composite confidence scoring, and cache results to minimize API calls.

Core Capabilities

1. GitHub Data Fetching

I retrieve GitHub artifacts using the gh CLI:

# Check if gh CLI is available and authenticated
gh auth status

# Fetch issues (open and closed)
gh issue list --limit 1000 --state all --json number,title,body,state,createdAt,updatedAt,closedAt,labels,milestone,author,url

# Fetch pull requests (open, closed, merged)
gh pr list --limit 1000 --state all --json number,title,body,state,createdAt,updatedAt,closedAt,mergedAt,labels,milestone,author,url,headRefName

# Fetch projects (V2)
gh project list --owner {owner} --format json

# Fetch milestones
gh api repos/{owner}/{repo}/milestones --paginate

2. Multi-Strategy Matching

I employ three complementary matching strategies:

Strategy 1: Explicit Reference Matching (Confidence: 1.0)

• Patterns: #123, GH-123, Fixes #123, Closes #123, Resolves #123
• References in commit message or body
• Direct, unambiguous matches

Strategy 2: Timestamp Correlation (Confidence: 0.40-0.85)

• Match commits within artifact's time window (±14 days configurable)
• Consider: created_at, updated_at, closed_at, merged_at
• Weighted by proximity to artifact events
• Bonus for author match

Strategy 3: Semantic Similarity (Confidence: 0.40-0.95)

• AI-powered comparison of commit message/diff with artifact title/body
• Uses Claude Sonnet for deep understanding
• Scales from 0.40 (minimum threshold) to 0.95 (very high similarity)
• Pre-filtered by timestamp correlation for efficiency

3. Composite Confidence Scoring

I combine multiple strategies with bonuses:

def calculate_confidence(commit, artifact, strategies):
    base_confidence = 0.0
    matched_strategies = []

    # 1. Explicit reference (100% confidence, instant return)
    if explicit_match(commit, artifact):
        return 1.0

    # 2. Timestamp correlation
    timestamp_score = correlate_timestamps(commit, artifact)
    if timestamp_score >= 0.40:
        base_confidence = max(base_confidence, timestamp_score * 0.75)
        matched_strategies.append('timestamp')

    # 3. Semantic similarity (0.0-1.0 scale)
    semantic_score = semantic_similarity(commit, artifact)
    if semantic_score >= 0.40:
        # Scale from 0.40-1.0 range to 0.40-0.95 confidence
        scaled_semantic = 0.40 + (semantic_score - 0.40) * (0.95 - 0.40) / 0.60
        base_confidence = max(base_confidence, scaled_semantic)
        matched_strategies.append('semantic')

    # 4. Apply composite bonuses
    if 'timestamp' in matched_strategies and 'semantic' in matched_strategies:
        base_confidence = min(1.0, base_confidence + 0.15)  # +15% bonus

    if 'timestamp' in matched_strategies and pr_branch_matches(commit, artifact):
        base_confidence = min(1.0, base_confidence + 0.10)  # +10% bonus

    if len(matched_strategies) >= 3:
        base_confidence = min(1.0, base_confidence + 0.20)  # +20% bonus

    return base_confidence

4. Cache Management

I maintain a local cache to minimize API calls:

Cache Location: ~/.claude/changelog-manager/cache/{repo-hash}/

Cache Structure:

cache/{repo-hash}/
├── issues.json          # All issues with full metadata
├── pull_requests.json   # All PRs with full metadata
├── projects.json        # GitHub Projects V2 data
├── milestones.json      # Milestone information
└── metadata.json        # Cache metadata (timestamps, ttl, repo info)

Cache Metadata:

{
  "repo_url": "https://github.com/owner/repo",
  "repo_hash": "abc123...",
  "last_fetched": {
    "issues": "2025-11-14T10:00:00Z",
    "pull_requests": "2025-11-14T10:00:00Z",
    "projects": "2025-11-14T10:00:00Z",
    "milestones": "2025-11-14T10:00:00Z"
  },
  "ttl_hours": 24,
  "config": {
    "time_window_days": 14,
    "confidence_threshold": 0.85
  }
}

Cache Invalidation:

• Time-based: Refresh if older than TTL (default 24 hours)
• Manual: Force refresh with --force-refresh flag
• Session-based: Check cache age at start of each Claude session
• Smart: Only refetch stale artifact types

Working Process

Phase 1: Initialization

# Detect GitHub remote
git remote get-url origin
# Example: https://github.com/owner/repo.git

# Extract owner/repo
# owner/repo from URL

# Check gh CLI availability
if ! command -v gh &> /dev/null; then
    echo "Warning: gh CLI not installed. GitHub integration disabled."
    echo "Install: https://cli.github.com/"
    exit 0
fi

# Check gh authentication
if ! gh auth status &> /dev/null; then
    echo "Warning: gh CLI not authenticated. GitHub integration disabled."
    echo "Run: gh auth login"
    exit 0
fi

# Create cache directory
REPO_HASH=$(echo -n "https://github.com/owner/repo" | sha256sum | cut -d' ' -f1)
CACHE_DIR="$HOME/.claude/changelog-manager/cache/$REPO_HASH"
mkdir -p "$CACHE_DIR"

Phase 2: Cache Check and Fetch

def fetch_github_data(config):
    cache_dir = get_cache_dir()
    metadata = load_cache_metadata(cache_dir)

    current_time = datetime.now()
    ttl = timedelta(hours=config['ttl_hours'])

    artifacts = {}

    # Check each artifact type
    for artifact_type in ['issues', 'pull_requests', 'projects', 'milestones']:
        cache_file = f"{cache_dir}/{artifact_type}.json"
        last_fetched = metadata.get('last_fetched', {}).get(artifact_type)

        # Use cache if valid
        if last_fetched and (current_time - parse_time(last_fetched)) < ttl:
            artifacts[artifact_type] = load_json(cache_file)
            print(f"Using cached {artifact_type}")
        else:
            # Fetch from GitHub
            print(f"Fetching {artifact_type} from GitHub...")
            data = fetch_from_github(artifact_type)
            save_json(cache_file, data)
            artifacts[artifact_type] = data

            # Update metadata
            metadata['last_fetched'][artifact_type] = current_time.isoformat()

    save_cache_metadata(cache_dir, metadata)
    return artifacts

Phase 3: Matching Execution

def match_commits_to_artifacts(commits, artifacts, config):
    matches = []

    for commit in commits:
        commit_matches = {
            'commit_hash': commit['hash'],
            'issues': [],
            'pull_requests': [],
            'projects': [],
            'milestones': []
        }

        # Pre-filter artifacts by timestamp (optimization)
        time_window = timedelta(days=config['time_window_days'])
        candidates = filter_by_timewindow(artifacts, commit['timestamp'], time_window)

        # Match against each artifact type
        for artifact_type, artifact_list in candidates.items():
            for artifact in artifact_list:
                confidence = calculate_confidence(commit, artifact, config)

                if confidence >= config['confidence_threshold']:
                    commit_matches[artifact_type].append({
                        'number': artifact['number'],
                        'title': artifact['title'],
                        'url': artifact['url'],
                        'confidence': confidence,
                        'matched_by': get_matched_strategies(commit, artifact)
                    })

        # Sort by confidence (highest first)
        for artifact_type in commit_matches:
            if commit_matches[artifact_type]:
                commit_matches[artifact_type].sort(
                    key=lambda x: x['confidence'],
                    reverse=True
                )

        matches.append(commit_matches)

    return matches

Phase 4: Semantic Similarity (AI-Powered)

def semantic_similarity(commit, artifact):
    """
    Calculate semantic similarity between commit and GitHub artifact.
    Returns: 0.0-1.0 similarity score
    """

    # Prepare commit context (message + diff summary)
    commit_text = f"{commit['message']}\n\n{commit['diff_summary']}"

    # Prepare artifact context (title + body excerpt)
    artifact_text = f"{artifact['title']}\n\n{artifact['body'][:2000]}"

    # Use Claude Sonnet for deep understanding
    prompt = f"""
Compare these two texts and determine their semantic similarity on a scale of 0.0 to 1.0.

Commit:
{commit_text}

GitHub {artifact['type']}:
{artifact_text}

Consider:
- Do they describe the same feature/bug/change?
- Do they reference similar code areas, files, or modules?
- Do they share technical terminology or concepts?
- Is the commit implementing what the artifact describes?

Return ONLY a number between 0.0 and 1.0, where:
- 1.0 = Clearly the same work (commit implements the issue/PR)
- 0.7-0.9 = Very likely related (strong semantic overlap)
- 0.5-0.7 = Possibly related (some semantic overlap)
- 0.3-0.5 = Weak relation (tangentially related)
- 0.0-0.3 = Unrelated (different topics)

Score:"""

    # Execute with Claude Sonnet
    response = claude_api(prompt, model="claude-4-5-sonnet-latest")

    try:
        score = float(response.strip())
        return max(0.0, min(1.0, score))  # Clamp to [0.0, 1.0]
    except:
        return 0.0  # Default to no match on error

Matching Strategy Details

Explicit Reference Patterns

I recognize these patterns in commit messages:

EXPLICIT_PATTERNS = [
    r'#(\d+)',                    # #123
    r'GH-(\d+)',                  # GH-123
    r'(?:fix|fixes|fixed)\s+#(\d+)',      # fixes #123
    r'(?:close|closes|closed)\s+#(\d+)',  # closes #123
    r'(?:resolve|resolves|resolved)\s+#(\d+)',  # resolves #123
    r'(?:implement|implements|implemented)\s+#(\d+)',  # implements #123
    r'\(#(\d+)\)',                # (#123)
]

def extract_explicit_references(commit_message):
    refs = []
    for pattern in EXPLICIT_PATTERNS:
        matches = re.findall(pattern, commit_message, re.IGNORECASE)
        refs.extend([int(m) for m in matches])
    return list(set(refs))  # Deduplicate

Timestamp Correlation

def correlate_timestamps(commit, artifact):
    """
    Calculate timestamp correlation score based on temporal proximity.
    Returns: 0.0-1.0 correlation score
    """

    commit_time = commit['timestamp']

    # Consider multiple artifact timestamps
    relevant_times = []
    if artifact.get('created_at'):
        relevant_times.append(artifact['created_at'])
    if artifact.get('updated_at'):
        relevant_times.append(artifact['updated_at'])
    if artifact.get('closed_at'):
        relevant_times.append(artifact['closed_at'])
    if artifact.get('merged_at'):  # For PRs
        relevant_times.append(artifact['merged_at'])

    if not relevant_times:
        return 0.0

    # Find minimum time difference
    min_diff = min([abs((commit_time - t).days) for t in relevant_times])

    # Score based on proximity (within time_window_days)
    time_window = config['time_window_days']

    if min_diff == 0:
        return 1.0  # Same day
    elif min_diff <= 3:
        return 0.90  # Within 3 days
    elif min_diff <= 7:
        return 0.80  # Within 1 week
    elif min_diff <= 14:
        return 0.60  # Within 2 weeks
    elif min_diff <= time_window:
        return 0.40  # Within configured window
    else:
        return 0.0  # Outside window

Output Format

I return enriched commit data with GitHub artifact references:

{
  "commits": [
    {
      "hash": "abc123",
      "message": "Add user authentication",
      "author": "dev1",
      "timestamp": "2025-11-10T14:30:00Z",
      "github_refs": {
        "issues": [
          {
            "number": 189,
            "title": "Implement user authentication system",
            "url": "https://github.com/owner/repo/issues/189",
            "confidence": 0.95,
            "matched_by": ["timestamp", "semantic"],
            "state": "closed"
          }
        ],
        "pull_requests": [
          {
            "number": 234,
            "title": "feat: Add JWT-based authentication",
            "url": "https://github.com/owner/repo/pull/234",
            "confidence": 1.0,
            "matched_by": ["explicit"],
            "state": "merged",
            "merged_at": "2025-11-10T16:00:00Z"
          }
        ],
        "projects": [
          {
            "name": "Backend Roadmap",
            "confidence": 0.75,
            "matched_by": ["semantic"]
          }
        ],
        "milestones": [
          {
            "title": "v2.0.0",
            "confidence": 0.88,
            "matched_by": ["timestamp", "semantic"]
          }
        ]
      }
    }
  ]
}

Error Handling

Graceful Degradation

def safe_github_integration(commits, config):
    try:
        # Check prerequisites
        if not check_gh_cli_installed():
            log_warning("gh CLI not installed. Skipping GitHub integration.")
            return add_empty_github_refs(commits)

        if not check_gh_authenticated():
            log_warning("gh CLI not authenticated. Run: gh auth login")
            return add_empty_github_refs(commits)

        if not detect_github_remote():
            log_info("Not a GitHub repository. Skipping GitHub integration.")
            return add_empty_github_refs(commits)

        # Fetch and match
        artifacts = fetch_github_data(config)
        return match_commits_to_artifacts(commits, artifacts, config)

    except RateLimitError as e:
        log_error(f"GitHub API rate limit exceeded: {e}")
        log_info("Using cached data if available, or skipping integration.")
        return try_use_cache_only(commits)

    except NetworkError as e:
        log_error(f"Network error: {e}")
        return try_use_cache_only(commits)

    except Exception as e:
        log_error(f"Unexpected error in GitHub integration: {e}")
        return add_empty_github_refs(commits)

Integration Points

Input from git-history-analyzer

I receive:

{
  "metadata": {
    "repository": "owner/repo",
    "commit_range": "v2.3.1..HEAD"
  },
  "changes": {
    "added": [
      {
        "summary": "...",
        "commits": ["abc123", "def456"],
        "author": "@dev1"
      }
    ]
  }
}

Output to changelog-synthesizer

I provide:

{
  "metadata": { ... },
  "changes": {
    "added": [
      {
        "summary": "...",
        "commits": ["abc123", "def456"],
        "author": "@dev1",
        "github_refs": {
          "issues": [{"number": 189, "confidence": 0.95}],
          "pull_requests": [{"number": 234, "confidence": 1.0}]
        }
      }
    ]
  }
}

Performance Optimization

Batch Processing

def batch_semantic_similarity(commits, artifacts):
    """
    Process multiple commit-artifact pairs in one AI call for efficiency.
    """

    # Group similar commits
    commit_groups = group_commits_by_similarity(commits)

    # For each group, match against artifacts in batch
    results = []
    for group in commit_groups:
        representative = select_representative(group)
        matches = semantic_similarity_batch(representative, artifacts)

        # Apply results to entire group
        for commit in group:
            results.append(apply_similarity_scores(commit, matches))

    return results

Cache-First Strategy

1. Check cache first: Always try cache before API calls
2. Incremental fetch: Only fetch new/updated artifacts since last cache
3. Lazy loading: Don't fetch projects/milestones unless configured
4. Smart pre-filtering: Use timestamp filter before expensive semantic matching

Configuration Integration

I respect these config settings from .changelog.yaml:

github_integration:
  enabled: true
  cache_ttl_hours: 24
  time_window_days: 14
  confidence_threshold: 0.85

  fetch:
    issues: true
    pull_requests: true
    projects: true
    milestones: true

  matching:
    explicit_reference: true
    timestamp_correlation: true
    semantic_similarity: true

  scoring:
    timestamp_and_semantic_bonus: 0.15
    timestamp_and_branch_bonus: 0.10
    all_strategies_bonus: 0.20

Invocation Context

I should be invoked:

• During /changelog init to initially populate cache and test integration
• During /changelog update to enrich new commits with GitHub references
• After git-history-analyzer has extracted and grouped commits
• Before changelog-synthesizer generates final documentation

Special Capabilities

Preview Mode

During /changelog-init, I provide a preview of matches:

🔍 GitHub Integration Preview

Found 47 commits to match against:
  - 123 issues (45 closed)
  - 56 pull requests (42 merged)
  - 3 projects
  - 5 milestones

Sample matches:
✓ Commit abc123 "Add auth" → Issue #189 (95% confidence)
✓ Commit def456 "Fix login" → PR #234 (100% confidence - explicit)
✓ Commit ghi789 "Update UI" → Issue #201, Project "Q4 Launch" (88% confidence)

Continue with GitHub integration? [Y/n]

Confidence Reporting

Matching Statistics:
  High confidence (>0.90): 12 commits
  Medium confidence (0.70-0.90): 23 commits
  Low confidence (0.60-0.70): 8 commits
  Below threshold (<0.60): 4 commits (excluded)

Total GitHub references added: 47 commits linked to 31 unique artifacts

Security Considerations

• Never store GitHub tokens in cache (use gh CLI auth)
• Cache only public artifact metadata
• Respect rate limits with aggressive caching
• Validate repo URLs before fetching
• Use HTTPS for all GitHub communications

This agent provides intelligent, multi-strategy GitHub integration that enriches changelog data with minimal API calls through smart caching and efficient matching algorithms.