ai-code-stats

Measure how much code from CodingAgents such as Claude Code and Codex is accepted into git commits, including AI-authored line counts, adoption rate, AI share of each commit, and token usage. Metrics are reported by git repository x committer on every commit.

Report delivery is pluggable: HTTP webhook, local JSONL files, or a custom command. All payloads are defined by versioned JSON Schemas and work across macOS, Windows, and Linux.

Full installation, configuration, reporting, and troubleshooting details are in docs/USAGE.md. Chinese documentation is available in README_zh.md.

What It Answers

• How many AI-generated lines were present in a commit, and how many were ultimately accepted.
• Total added lines, AI-added lines, and AI share of each commit, using both raw and effective-code counting modes.
• Token usage associated with the commit, including input, output, and cache read tokens.

How It Works

 AI edit (PostToolUse hook)             git commit (post-commit / post-merge hook)
 +----------------------------+         +------------------------------------------+
 | Parse Edit/Write/apply_patch|        | Read commit diff, including renames       |
 | Added lines -> normalize+hash| ----> | Match AI fingerprints within the window   |
 | Mark effective-code lines   | pending| Compute adoption, AI share, and tokens    |
 | Store under .git/ai-code-stats/      | Build JSON envelope -> dispatch reporters |
 +----------------------------+         +------------------------------------------+

• Adoption rate = AI lines accepted into the commit / AI lines generated in the attribution window.
• AI share = commit added lines matched to AI fingerprints / total commit added lines.
• Matching is based on normalized content hashes, so moved or renamed AI code can still be attributed.

Installation

Requires Python 3.9 or newer and git.

pip install ai-code-stats          # or: pip install -e . for development

# Run from the target repository root. Installs git + Claude + Codex hooks.
ai-code-stats install

# Install only selected hooks, or preview without writing files.
ai-code-stats install --git
ai-code-stats install --claude --scope user      # writes ~/.claude/settings.json
ai-code-stats install --codex --dry-run

# Uninstall. This is idempotent and preserves your own hook content.
ai-code-stats uninstall

Codex hooks are written to $CODEX_HOME/config.toml, defaulting to ~/.codex/config.toml. Because the Codex hook schema is still evolving, run ai-code-stats install --codex --dry-run after installation and confirm that your Codex version supports inline [[hooks.PostToolUse]] entries.

Configuration

Configuration is merged in this order, with later sources overriding earlier ones: built-in defaults, user-level config.json, repository .ai-code-stats.json, then the file pointed to by AI_CODE_STATS_CONFIG. String values support ${ENV:VAR} placeholders for secret injection.

{
  "enabled": true,
  "privacy": {
    "store_plaintext": true,     // Keep AI line plaintext locally under .git/.
    "redact_in_reports": true    // Reports include metrics only, not source code.
  },
  "files": {
    "include": [],               // Empty = known code extensions; non-empty = only matching files.
    "exclude": ["**/node_modules/**", "**/*.min.js", "package-lock.json"]
  },
  "attribution": {
    "count_modes": ["raw", "effective"],
    "primary": "effective",      // Primary metric uses the effective-code mode.
    "merge_strategy": "skip",    // Merge commits: skip or first_parent.
    "detect_renames": true
  },
  "reporters": [
    { "type": "json_file", "path": "{repo_data}/reports.jsonl" },
    { "type": "http_webhook",
      "url": "https://metrics.example.com/ingest",
      "headers": { "Authorization": "Bearer ${ENV:AI_CODE_STATS_TOKEN}" },
      "mapping": {                // Map the envelope into any backend schema.
        "repo": "data.repo_id",
        "rate": "data.ai.effective.adoption_rate",
        "tokens": "data.tokens.total"
      }
    },
    { "type": "command", "argv": ["my-forwarder"] }  // Envelope JSON is passed via stdin.
  ]
}

Counting Modes

• raw: all added and removed lines.
• effective: blank lines and pure comment lines are excluded based on the file language's comment syntax.

File Filtering

By default, ai-code-stats counts files with known code-language extensions and excludes lock files, generated artifacts, vendored directories, and binaries. Use files.include and files.exclude glob patterns, including **, to customize this behavior.

Data Contract

The schemas/ directory contains three versioned JSON Schemas:

Example envelope:

{
  "schema_version": "1.0",
  "kind": "commit_stat",
  "produced_at": "2026-06-15T08:00:00Z",
  "producer": { "plugin": "ai-code-stats", "version": "0.1.1", "os": "darwin" },
  "data": {
    "repo_id": "github.com/org/repo",
    "commit": { "sha": "...", "branch": "main", "is_merge": false },
    "committer": { "name": "Dev", "email": "dev@x.com" },
    "totals": { "files_changed": 2, "raw": { "lines_added": 5 }, "effective": { "lines_added": 3 } },
    "ai": {
      "raw":       { "ai_lines_added": 4, "adoption_rate": 1.0, "ai_share_of_commit": 0.8 },
      "effective": { "ai_lines_added": 3, "adoption_rate": 1.0, "ai_share_of_commit": 1.0 }
    },
    "tokens": { "input": 120, "output": 30, "total": 150 }
  }
}

Common Commands

ai-code-stats status              # Show pending attribution events and token snapshots.
ai-code-stats report              # Print the current HEAD envelope without sending or consuming it.
ai-code-stats flush               # Retry failed report deliveries.

Privacy

• AI line plaintext is stored only under the repository's .git/ai-code-stats/ directory and is never committed.
• Reports default to redact_in_reports=true, so they contain metrics only and no source code.
• For stronger privacy, set privacy.store_plaintext=false; local storage will keep only hashes.

Known Limitations

• Merge commits skip attribution by default to avoid noisy diffs. Set merge_strategy=first_parent if you want first-parent merge attribution.
• Adoption rate is approximate under rebase, cherry-pick, and commit --amend.
• Token attribution is estimated from the session-level cumulative delta since the previous commit, so concurrent work across multiple repositories can be approximate.

Development

PYTHONPATH=src python3 -m pytest
PYTHONPATH=src python3 -m ai_code_stats.cli --help

Architecture overview: agents/ adapts CodingAgent payloads, classify filters and classifies lines, attribution computes matches, tokens aggregates token usage, reporters/ provides pluggable delivery, githook/ computes commit statistics, and install/ manages hook installation.

To add a reporter, implement reporters/base.Reporter and register it in reporters/registry.REPORTER_TYPES. To add an Agent, implement agents/base.AgentAdapter and register it in agents/registry.