codegov

Know what AI wrote in your codebase.

CodeGov scans your git history and detects commits authored by AI coding tools — Claude Code, Cursor, GitHub Copilot, Devin, and Aider — from git metadata alone. No config, no API keys, no agents to install.

$ npx @james-wall/codegov scan

82.3% of commits are AI-authored (283/344)
Agents: 280 claude-code, 2 copilot, 1 cursor

Detected commits:
  c7e2a46  claude-code [Claude Opus 4.6]  +0/-22   (95%)
  07d9deb  claude-code [Claude Sonnet 4.6]  +8/-8   (95%)
  5f90db7  copilot  +83/-22   (95%)
  ...

Install

npx @james-wall/codegov scan       # try it now, no install needed
npm install -g @james-wall/codegov # or install globally

Commands

Zero-config (works on any repo)

codegov scan                # scan full history
codegov scan --since 3m     # last 3 months
codegov scan --format html  # generate shareable HTML report
codegov stats               # agent breakdown + monthly trends
codegov query --agent cursor --since 90d
codegov forensics <hash>    # deep dive on a single commit
codegov export --format csv # export for spreadsheets
codegov export --format sbom # SPDX 2.3 compliance artifact
codegov share               # upload HTML report as a GitHub Gist

Ongoing monitoring

codegov init                # install post-commit hook + scan history
codegov server              # start OTEL collector + live dashboard
codegov server --port 8080  # custom port

The server accepts OpenTelemetry traces at POST /v1/traces (Claude Code natively supports OTEL export) and git commit metadata at POST /v1/hooks/commit, then correlates them to build tool-to-commit attribution.

How detection works

CodeGov identifies AI-authored commits by matching patterns in git metadata:

| Tool | Detection signals | |------|------------------| | Claude Code | Co-Authored-By: Claude trailer, Generated by Claude Code marker | | Cursor | cursor[bot] author, [email protected] co-author, Made-with: Cursor trailer | | Copilot | [email protected] co-author, copilot-swe-agent[bot] author, Agent-Logs-Url trailer | | Devin | devin-ai-integration[bot] author email | | Aider | aider (model) <[email protected]> co-author, aider: message prefix |

Each detection carries a confidence score (0-100%). The threshold is 30% — anything below is classified as human-authored.

Only the five named tools above are counted by default — each is matched on a specific, documented signature, so false positives are rare. A low-signal "generic AI" heuristic (e.g. an <ai@…> co-author with no specific tool) is available behind codegov scan --include-generic, but it is off by default to keep the headline number trustworthy. Detection reads commit metadata only, so the percentage is a floor: AI-assisted commits that carry no signature are not counted.

Export formats

• JSON — full audit report with summary + all records
• CSV — tabular, import into any spreadsheet
• SBOM — SPDX 2.3 for compliance workflows
• HTML — dark-themed dashboard, shareable as a standalone file

Use cases

• Engineering leaders: understand what % of your codebase AI wrote and which tools produce the most merged code
• Compliance teams: generate provenance artifacts for SOC2 auditors asking about AI-generated code
• Open source maintainers: add transparency about AI contributions
• Individual devs: track your own AI tool usage across projects

License

MIT