ai-attr

Line-level AI code attribution for your git history.

ai-attr automatically tracks which lines of code were written by AI agents — Claude Code, GitHub Copilot, Codex, Cline — and stores that attribution as git notes. No manual comment markers. No PATH shimming. No binary interposition.

The problem

Many organizations require developers to mark AI-generated code with comment blocks like /*START GENAI*/ ... /*END GENAI*/. This is error-prone, noisy, and only works when developers remember to do it. Other tools solve this automatically, but they work by shimming the git binary on PATH — something enterprise security teams often flag as a supply chain risk.

How ai-attr works

ai-attr uses each AI agent's native hook system to detect edits, and a standard git post-commit hook to record the results. No binary interposition required.

Developer uses Claude Code to edit src/main.rs
  ↓
Claude's PostToolUse hook fires
  ↓
ai-attr checkpoint claude --hook-input stdin
  → diffs the file against its last known state
  → writes a checkpoint to .git/ai-attr/
  ↓
Developer runs: git commit
  ↓
Post-commit hook fires
  ↓
ai-attr commit
  → consolidates checkpoints into an AuthorshipLog
  → stores it as a git note on refs/notes/ai-attribution

The result is a git note on every commit that records exactly which lines were AI-generated, by which agent, in which session.

Quick start

# Build
make build

# Install hooks for Claude Code in the current repo
./ai-attr install --agents claude

# That's it. Use Claude Code normally.
# Attribution is recorded automatically on every commit.

# View attribution for the latest commit
./ai-attr show

# View attribution for a specific commit
./ai-attr show abc123

Commands

Command	Description
ai-attr install --agents <list>	Set up hooks for specified agents
ai-attr checkpoint <agent>	Record a checkpoint (called by agent hooks, not manually)
ai-attr commit	Consolidate checkpoints into a git note (called by post-commit hook)
ai-attr show [commit]	Display the attribution note for a commit
ai-attr blame <file>	Line-level AI attribution, like git blame
ai-attr stats [range]	AI composition statistics across a commit range
ai-attr uninstall	Remove hooks and agent configurations

Supported agents

Agent	Hook mechanism	Granularity	Status
Claude Code	PreToolUse / PostToolUse	Per-edit	Implemented
GitHub Copilot	preToolUse / postToolUse (agent mode)	Per-edit	Implemented
Cline	PostToolUse	Per-edit	Implemented
Codex CLI	notify (agent-turn-complete)	Per-turn	Implemented

Note: Copilot's inline tab completions cannot be tracked — there is no hook for accepting a completion. Only agent mode edits are captured.

What the output looks like

ai-attr show HEAD displays the attribution note:

src/main.rs
  a1b2c3d4e5f6a7b8 1-10,15-20
src/lib.rs
  a1b2c3d4e5f6a7b8 1-50
  9f8e7d6c5b4a3210 55-60
---
{
  "schema_version": "authorship/3.0.0",
  "tool_version": "ai-attr/0.2.0",
  "prompts": {
    "a1b2c3d4e5f6a7b8": {
      "agent_id": {
        "tool": "claude",
        "model": "sonnet-4",
        "id": "session-abc123"
      },
      "human_author": "alice",
      "total_additions": 60,
      "total_deletions": 5,
      "accepted_lines": 55
    }
  }
}

Lines 1-10 and 15-20 of src/main.rs were written by Claude (sonnet-4) in session session-abc123, invoked by alice.

Design principles

Principle	Details
No PATH shimming	Uses native agent hooks, not binary interposition
Git-native storage	Attribution stored as git notes — travels with the repo
Agent-agnostic core	One binary, agent differences isolated behind AgentAdapter interface
Automatic	No manual markers — hooks capture attribution transparently

Building from source

Requires Go 1.21+.

git clone <repo-url>
cd ai-attr
make build
# Binary at ./ai-attr

Make targets

Target	Description
make build	Build the ai-attr binary
make test	Run all tests
make test-v	Run all tests with verbose output
make vet	Run go vet
make clean	Remove the built binary
make install	Build and copy binary to $GOPATH/bin

Running tests

make test

Tests create temporary git repos and don't require network access.

Project status

ai-attr is under active development. All core functionality is implemented with 96 passing tests: checkpoint recording, commit consolidation, install/uninstall, blame, stats, and adapters for Claude Code, GitHub Copilot, Cline, Codex, and a generic adapter.

See docs/PLANS.md for the full roadmap.

License

MIT