claude-blame

git blame says WHO. claude-blame says WHICH CLAUDE SESSION.

Link every git commit to the Claude Code conversation that wrote it — then jump back into it with one command.

$ claude-blame src/auth.ts:42
commit a1b2c3d4e5f6  fix: handle expired refresh tokens
session 9e297d8a-ef72-483e-8599-d2d6c2874a6a
first prompt: refresh tokens are getting rejected after 24h, debug this

→ opening: claude --resume 9e297d8a-ef72-483e-8599-d2d6c2874a6a

A tiny CLI that records which Claude Code session was active when each commit was made. Later, you can run claude-blame <file>:<line> and jump straight back into the exact conversation that produced that code — full context, intact.

No cloud. No telemetry. Everything stays local in .git/ai-sessions.json.

Why this exists

You're 3 weeks into a project. A bug surfaces. You open the file, find the suspect line, and think:

"I wrote this with Claude... last Tuesday? Wednesday? Which session was it?"

Claude Code already saves every session as a .jsonl file. The conversation is on your disk. The problem is finding the specific one that wrote the specific line you're staring at.

claude-blame solves that with a post-commit hook + git blame.

Install

npm install -g claude-blame

Then, in any git repository where you use Claude Code:

claude-blame install

This adds a post-commit hook. From now on, every commit gets linked to your currently-active Claude Code session.

For commits made before you installed, try:

claude-blame backfill

It matches past commits to sessions by timestamp (best-effort — not perfect, but usually close).

Usage

# By file + line — jumps to the session that authored that line
claude-blame src/auth.ts:42

# By commit SHA — direct lookup
claude-blame a1b2c3d

# Print the transcript instead of opening Claude Code
claude-blame src/auth.ts:42 --print

# List recent commits and their linked sessions
claude-blame list
claude-blame list -n 25

How it works

1. claude-blame install writes a post-commit hook into .git/hooks/post-commit.
2. After every commit, the hook calls claude-blame _record, which:
   • reads ~/.claude/projects/<encoded-cwd>/*.jsonl,
   • picks the most recently modified session file (= the one you were just using),
   • saves commit_sha → session_id into .git/ai-sessions.json.
3. When you run claude-blame <file>:<line>, it:
   • runs git blame on that line to find the commit,
   • looks up the session in .git/ai-sessions.json,
   • runs claude --resume <session_id> to drop you back into the conversation.

All local. No daemon. No network call.

How matches are scored

Claude Code stores each session under ~/.claude/projects/<encoded-cwd>/, where <encoded-cwd> is the directory Claude Code was launched from. So which session wrote your commit depends on where you launched Claude Code, not where the commit lives.

claude-blame reports three match qualities:

| Tag | Meaning | |---|---| | (no tag) | Exact match. Claude was launched from the repo root. Most accurate. | | ~ / (parent-dir match) | Claude was launched from a parent directory (e.g., your home folder) and you edited files inside this repo. Almost always correct, but not guaranteed. | | ? / (guess) | Set by claude-blame backfill. Matched by timestamp only — if multiple sessions overlap, this can mislink. Use sparingly. |

Best practice: launch Claude Code from the repo root (cd ~/my-project && claude) so every commit gets an exact match.

What gets stored

.git/ai-sessions.json:

{
  "version": 1,
  "commits": {
    "a1b2c3d4...": {
      "sessionId": "9e297d8a-ef72-483e-8599-d2d6c2874a6a",
      "sessionPath": "/Users/you/.claude/projects/-Users-you-proj/9e297d8a-....jsonl",
      "recordedAt": "2026-06-09T20:35:00.000Z"
    }
  }
}

The actual session transcript stays where Claude Code put it. claude-blame only stores the pointer.

You can .gitignore ai-sessions.json if you don't want to share session IDs across the team, or commit it if you do.

Requirements

• Node.js ≥ 18
• Git
• Claude Code installed (claude on PATH) — only needed for the --resume jump, not for recording

Roadmap

• [ ] MCP server version (so Claude Code can answer "why was this written?" without leaving the editor)
• [ ] Cursor / Aider session detection
• [ ] claude-blame --since=2.weeks (range lookups)
• [ ] HTML transcript viewer fallback when Claude Code isn't installed
• [ ] Team mode: optional shared session metadata via a separate sync repo

PRs welcome. See CONTRIBUTING.md.

FAQ

Does this send my conversations anywhere? No. Everything is local. The tool reads ~/.claude/projects/ (already on your disk) and writes .git/ai-sessions.json (in your repo).

What if I use multiple Claude Code sessions per commit? claude-blame records the most recently active one at the moment of commit. If you switch sessions a lot, the post-commit hook captures whichever was touched last.

Does it work for Cursor / Aider / Continue? Not yet — only Claude Code for now. PRs welcome.

Will it leak my session IDs if I git push ai-sessions.json? The session ID itself is harmless — but the path to the session file may reveal your home directory. If you push, consider .gitignore-ing the file.

License

MIT — see LICENSE.

Built by Cristian Tanase because I kept losing track of which Claude session wrote which line.