scribe 🪶

scribe maintains a per-repo pool of tool usage events. Each time the harness invokes a tool, a hook adds an entry to the pool.

Run scribe amend to drain the pool and annotate the current commit with:

Assisted-By: anthropic:claude-sonnet-4-6, github:gpt-4o

The pool is stored at .git/scribe/<branch>/pool.jsonl — scoped to the current branch, local to the repo, inside .git/ so it is never committed and requires no .gitignore entry.

Installation

go install github.com/gnugomez/scribe@latest

Usage

In order to capture tool usage, you need to set up hooks for your AI tools (see below). Once that's done, the typical workflow is:

Just run scribe amend after your commit to annotate it with the tools you used. This is a manual step so you can choose which commits to annotate and when.

scribe amend

This could be paired with pre-commit hooks.

Some other useful commands:

scribe amend --dry-run   # Preview the annotation without clearing the pool
scribe pool # View the list of captured tool events in the pool
scribe clear # Discard pool without amending

---

Claude Code CLI

Claude Code fires PostToolUse hooks after every file write. Add the following to ~/.claude/settings.json:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Write|Edit|MultiEdit",
        "hooks": [
          {
            "type": "command",
            "command": "scribe hook --vendor anthropic"
          }
        ]

      }
    ],
    "SessionStart": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "scribe hook --vendor anthropic"
          }
        ]
      }
    ]
  }
}

scribe reads the model from the payload. If a PostToolUse payload doesn't include a model, scribe will reuse the latest known model from that same session_id.

---

GitHub Copilot Chat

VS Code Copilot Chat supports Agent Hooks (preview) that fire when the agent uses tools.

Create .copilot/hooks/scribe.json in your repo or globally at ~/.copilot/hooks/scribe.json with the following content:

{
  "hooks": {
    "PostToolUse": [
      {
        "command": "scribe hook --vendor github --format copilot",
        "type": "command"
      }
    ],
    "SessionStart": [
      {
        "command": "scribe hook --vendor github --format copilot",
        "type": "command"
      }
    ]
  }
}

scribe reads the model from the payload. If a PostToolUse payload doesn't include a model, scribe will reuse the latest known model from that same session_id.

Warning

Copilot hook payloads may not always include model metadata. The SessionStart hook lets scribe capture the model at session start so it can be reused when later tool events omit it.

Note

The VS Code Agent Hooks API is in preview. See the VS Code Copilot hooks documentation for the latest config format and payload schema.

---

Pool design

The pool is a single file at .git/scribe/pool.jsonl shared across all branches — just like git's staging area. When you switch branches carrying uncommitted work, the pool entries follow you. The pool is cleared explicitly by scribe amend (after annotating) or scribe clear (to discard).

Tip

If you discard your work (e.g. git reset --hard) without committing, run scribe clear to remove the now-irrelevant pool entries.

---

Workflow

# Work session with Claude Code and Copilot Chat:
# → hooks fire automatically as each tool writes files

git add -A
git commit -m "feat: implement user auth"

scribe amend --dry-run   # preview: Assisted-By: anthropic:claude-sonnet-4-6, github:gpt-4o
scribe amend             # apply and clear pool

---

Adding support for a new AI tool

Adding a new hook format requires only a new vendor hook package:

1. Create vendors/<toolname>/hook/parser.go
2. Implement the hook.Parser interface (Name() + Parse())
3. Self-register via func init() { hook.Register(&Parser{}) }
4. Add a blank import in cmd/root.go for github.com/gnugomez/scribe/vendors/<toolname>/hook

The core (cmd/amend.go, cmd/hook.go) never needs to change.

---

Pool file

.git/scribe/<branch>/pool.jsonl — one JSON object per line:

{"timestamp":"2026-04-13T10:00:00Z","vendor":"anthropic","model":"claude-sonnet-4-6"}
{"timestamp":"2026-04-13T10:01:00Z","vendor":"github","model":"gpt-4o"}

With scribe pool --debug, each entry also shows model source so you can see where attribution came from: payload, session, flag, or default.