agentdiff 0.1.0

Audit and trace autonomous AI code contributions in git repositories
agentdiff-0.1.0 is not a library.

agentdiff

Audit and trace autonomous AI code contributions in git repositories.

What is agentdiff?

agentdiff tracks who (which AI agent) wrote what code in your repository. It captures the agent name, model, prompt, and line-level attribution for every AI-assisted edit — storing this metadata for audit trails, compliance, and accountability.

Installation

# Install latest release (recommended)
curl -fsSL https://raw.githubusercontent.com/codeprakhar25/agentdiff/master/install.sh | bash

# Install a specific version
curl -fsSL https://raw.githubusercontent.com/codeprakhar25/agentdiff/master/install.sh | bash -s -- --version v0.1.0

# Install from crate source
cargo install --path ~/agentdiff

# Build release and copy manually
cargo build --release
cp target/release/agentdiff ~/.cargo/bin/
cp target/release/agentdiff-mcp ~/.cargo/bin/

Quick Start

# Initialize tracking in a repository
cd ~/your-project
agentdiff init

# Start MCP server (stdio) for agent integrations
agentdiff-mcp

# Make some AI-assisted edits, then commit

# List all captured entries
agentdiff list

# Show one commit's full ledger details
agentdiff show <sha>

# See line-by-line blame for a file
agentdiff blame src/main.rs

# View statistics
agentdiff stats

# Generate CI report (stdout, or --out-md / --out-annotations for files)
agentdiff report --format markdown
agentdiff report --format annotations --out-annotations annotations.json
# optional filters: --since RFC3339 --agent substring --model substring

Commands

Command Description
agentdiff init Initialize tracking in current repository
agentdiff list List all captured attribution entries
agentdiff blame <file> Show line-level attribution (like git-blame)
agentdiff stats Show aggregate statistics by agent/file/model
agentdiff report CI report: markdown and/or GitHub-style annotations (--out-md, --out-annotations, --agent, --model, --since)
agentdiff diff [<commit>] Show attribution changes in a commit
agentdiff log Show chronological history
agentdiff show <sha> Show one ledger entry by commit SHA
agentdiff ledger repair Normalize and de-duplicate .agentdiff/ledger.jsonl
agentdiff ledger import-notes Import legacy refs/notes/agentdiff into ledger
agentdiff sync-notes Legacy: fetch refs/notes/agentdiff for migration
agentdiff config Manage configuration

How It Works

  1. Hook Installation — agentdiff init installs:

    • Git pre-commit hook to snapshot staged AI attribution context
    • Git post-commit hook to append one commit-scoped ledger line
    • Claude Code PostToolUse hook
    • Cursor afterFileEdit hook
    • Codex notify hook
    • Gemini/Antigravity BeforeTool/AfterTool hooks
    • Windsurf repo-level hooks
    • OpenCode repo-level plugin
    • VS Code Copilot extension (~/.vscode/extensions/agentdiff-copilot-0.1.0/)

  2. Capture — When you use AI tools:

    • Claude Code → PostToolUse hook fires on Edit/Write/MultiEdit
    • Cursor → afterFileEdit/afterTabFileEdit hooks fire
    • Codex → notify hook fires on turn completion
    • Gemini/Antigravity → BeforeTool/AfterTool hooks fire on write_file/replace
    • Windsurf → post_write_code hook fires on code write
    • OpenCode → plugin fires on tool.execute.after
    • GitHub Copilot → VS Code extension captures inline completions and chat edits on save
    • Each capture writes to <repo>/.git/agentdiff/session.jsonl
    • Optional MCP writes context to <repo>/.git/agentdiff/pending.json

  3. Commit — On git commit:

    • Pre-commit hook writes <repo>/.git/agentdiff/pending-ledger.json
    • Post-commit hook finalizes one line into <repo>/.agentdiff/ledger.jsonl
    • By default, post-commit auto-amends the commit to include ledger changes
    • Pending files are cleared after finalize

  4. View — Use CLI commands to inspect captured data

Configuration

Config stored at ~/.agentdiff/config.toml

schema_version = "1.0"
data_dir = "~/.agentdiff/spillover" # optional spillover for no-repo captures
scripts_dir = "~/.agentdiff/scripts"
auto_amend_ledger = true

[[repos]]
path = "/home/user/project"
slug = "-home-user-project"

Disable same-commit amend behavior:

agentdiff config set auto_amend_ledger false
agentdiff init

Supported Agents

  • Claude Code — via PostToolUse hook
  • Cursor — via afterFileEdit/afterTabFileEdit hooks
  • Codex CLI — via notify hook
  • Gemini/Antigravity — via ~/.gemini/settings.json hooks (write_file|replace)
  • Windsurf — via post_write_code hook
  • OpenCode — via .opencode/plugins/agentdiff.ts
  • GitHub Copilot (VS Code) — via VS Code extension installed to ~/.vscode/extensions/
  • MCP context writer — via record-context.py (writes .git/agentdiff/pending.json)
  • Manual batch fallback — capture-antigravity.py --prompt "..." --model "..."

MCP Server

agentdiff-mcp is a real MCP stdio server.
Supported methods:

  • initialize
  • tools/list
  • tools/call for tool record_context

record_context writes context to <repo>/.git/agentdiff/pending.json, which is consumed by agentdiff pre/post commit hooks and attached to the next ledger line.

Example record_context arguments:

{
  "cwd": "/path/to/repo",
  "agent": "codex",
  "model_id": "gpt-5.4",
  "session_id": "sess_abc123",
  "prompt": "add auth middleware",
  "files_read": ["src/auth.rs", "src/config.rs"],
  "intent": "auth hardening",
  "trust": 92,
  "flags": ["security"]
}

MCP Client Config

Minimal MCP stdio config is provided at:

  • examples/mcp/agentdiff-mcp.json

Generic server block:

{
  "mcpServers": {
    "agentdiff": {
      "command": "agentdiff-mcp",
      "args": [],
      "env": {}
    }
  }
}

Use this same command in Claude/Cursor/Codex/OpenCode MCP settings.

End-to-End Testing

1) Reinstall latest binaries and scripts

cargo install --path /home/prakh/agentdiff --force
cd /path/to/your/repo
agentdiff init

2) MCP flow test (commit should include prompt/model/session)

# In repo root
agentdiff-mcp
# In another terminal, send tools/call record_context via your MCP client
# then make a small edit + git add + git commit

agentdiff list
agentdiff show HEAD

3) Hook-only fallback test (no MCP)

# Make an AI edit from Cursor/Codex/Windsurf/Claude, then commit
agentdiff list
agentdiff show HEAD

4) Cursor diagnostics (if entry is missing)

export AGENTDIFF_DEBUG=1
# make one Cursor edit + commit
tail -n 100 ~/.agentdiff/logs/capture-cursor.log
tail -n 50 .git/agentdiff/session.jsonl
tail -n 20 .agentdiff/ledger.jsonl

4b) Codex / Gemini diagnostics (if entry is missing)

export AGENTDIFF_DEBUG=1
# make one edit in Codex or Gemini, then commit
tail -n 100 ~/.agentdiff/logs/codex-notify-fired.log
tail -n 100 ~/.agentdiff/logs/capture-codex.log
tail -n 50 .git/agentdiff/session.jsonl
tail -n 20 .agentdiff/ledger.jsonl

5) Edge cases to validate

  • Empty or malformed hook payloads should not crash capture scripts.
  • afterTabFileEdit events should still attribute file/line/model without prompt.
  • Events with Windows-style backslash paths should resolve to repo-local paths.
  • If hook cwd is not repo (for example ~/.cursor), file-path-based repo resolution should still write to <repo>/.git/agentdiff/session.jsonl.
  • Commits with no staged overlap against captured lines should fall back to human.
  • Duplicate finalize on same SHA should not append duplicate ledger rows.

Release Process

Tagging v* triggers release workflow:

  • Runs Rust + Python tests and MCP smoke test
  • Builds agentdiff and agentdiff-mcp artifacts for Linux/macOS/Windows
  • Publishes SHA256SUMS
  • Creates GitHub Release with artifacts
  • Optionally publishes crate to crates.io when CARGO_REGISTRY_TOKEN is set

Output Formats

List Output

  # — COMMIT     TIME         AGENT         MODEL                 TOOL      FILE                          LINES    PROMPT
  ──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
  1  a1b2c3d4   Mar 20 17:52 claude-code   sonnet-4-6           Edit      src/auth/tokens.py             17-24   "add auth middleware"
  2  a1b2c3d4   Mar 20 17:55 cursor        gpt-4o               afterFileEdit src/main.rs             1       "add main entry"
  * a1b2c3e5   Mar 20 18:02 claude-code   sonnet-4-6           Write     src/utils.rs                  1-94    "refactor utils"  (uncommitted)

Blame Output

  agentdiff blame — src/main.rs
  ───────────────────────────────────────────────────────────────────────────────────────────────────
     1 human         fn main() {
     2 human         use anyhow::Result;
     3 claude-code   use std::process::Command;  (Edit)
     4 human
     5 cursor        fn run() -> Result<()> {   (afterFileEdit)

Stats Output

  agentdiff — Statistics

  Total lines tracked: 2847

  By Agent:
    claude-code   1847 (65%) ████████████████████
    cursor         647 (23%) ████████
    human         353 (12%) ████

  By File:
    src/main.rs                    340 lines (78% AI) — claude-code
    src/auth/mod.rs                220 lines (65% AI) — cursor

Use Cases

  • Security audits — Know which AI generated risky code
  • Compliance — SOC2, ISO 27001 require knowing what generated production code
  • Code review — Identify AI-authored code requiring extra scrutiny
  • Attribution — Track prompts driving major changes

VS Code Copilot Extension

agentdiff init installs a plain-JavaScript VS Code extension directly into ~/.vscode/extensions/agentdiff-copilot-0.1.0/. No npm install or build step is needed. Restart VS Code after running agentdiff init to activate it.

The extension:

  • Detects GitHub Copilot inline completions (multi-character/multi-line insertions)
  • Flushes captured lines to session.jsonl on file save
  • Exposes command "agentdiff: Capture Copilot edits for current file" for manual capture

To skip Copilot setup: agentdiff init --no-copilot To skip Gemini/Antigravity setup: agentdiff init --no-antigravity

Note: The extension uses a heuristic — insertions of ≥10 characters or multiple lines while Copilot is active are attributed to Copilot. Short single-character insertions are treated as manual typing.

Architecture

~/.agentdiff/
├── config.toml           ← Global configuration
├── scripts/              ← Python capture scripts
│   ├── capture-claude.py
│   ├── capture-cursor.py
│   ├── capture-codex.py
│   ├── capture-windsurf.py
│   ├── capture-opencode.py
│   ├── capture-copilot.py
│   ├── prepare-ledger.py
│   ├── finalize-ledger.py
│   ├── record-context.py
│   ├── capture-antigravity.py
│   └── write-note.py      ← legacy notes writer (migration only)
└── spillover/            ← Optional no-repo event spillover

~/.vscode/extensions/
└── agentdiff-copilot-0.1.0/
    ├── package.json      ← VS Code extension manifest
    └── extension.js      ← Copilot capture logic (plain JS, no build needed)

<repo>/.agentdiff/
└── ledger.jsonl          ← Canonical committed append-only attribution log

<repo>/.git/agentdiff/
├── session.jsonl         ← Uncommitted per-repo buffer from hooks/agents
├── pending.json          ← Ephemeral MCP context handoff
└── pending-ledger.json   ← Pre-commit snapshot finalized in post-commit

<repo>/.windsurf/
└── hooks.json            ← Windsurf repo-level hook config

<repo>/.opencode/plugins/
└── agentdiff.ts          ← OpenCode repo-level plugin

git refs (legacy migration source):
└── refs/notes/agentdiff

Requirements

  • Rust 1.85+
  • Python 3.7+ (for capture scripts)
  • Git repository

License

Dual-licensed under: