DiffButler

A terminal dashboard for human-directed AI code reviews on GitHub.

Keep track of what each pull request needs next: AI review, your review, or author changes. Start an AI review, discuss its findings in the agent's own session, and finish your review in the browser.

Install

Install uv and Git. From a local checkout of this repository, run:

uv tool install --python 3.14 .

uv manages Python and dependencies in an isolated environment. No virtual environment activation is needed. If diffbutler is not on your PATH, run uv tool update-shell and restart your shell.

Install and sign in to Codex CLI or Claude Code before requesting AI reviews. The native presets support Codex CLI or Claude Code on Linux and macOS without version checks. Reviews use your normal agent settings, including model and reasoning effort, with background permission and output overrides. New conversations remain in the agent’s own history after DiffButler cleans up its workspaces; existing private sessions retain their original home for resume. See supported agents and validation for details.

Use

Use authenticated GitHub access, including for public repositories. If you already use gh auth login, export its token before launching DiffButler:

export GITHUB_TOKEN="$(gh auth token)"

Alternatively, create a fine-grained personal access token for the repositories you want to review with Pull requests: Read-only permission, and set it as GITHUB_TOKEN. Anonymous access is a last resort for public repositories; its low API quota makes it unsuitable for regular polling. See the authentication guide for host/account selection, HTTPS versus SSH Git setup, and token permissions.

From the GitHub repository you want to review, launch:

diffbutler

First-run setup detects the local repository, looks up your GitHub username using the selected token, asks for your preferred agent, and saves your choices. If username detection is unavailable, setup asks for it. Choose the recommended token access for public repositories; private repositories require it. Setup asks for the token's environment variable name and validates access. It never saves the token. New configs silently enable both Codex and Claude when found in PATH; setup asks only which detected agent should be the default. Missing agents get commented-out entries you can enable after installation. Later launches reuse your configuration.

The dashboard shows active reviews, including followed PRs whose GitHub review requests have disappeared after you posted feedback. Waiting-for-author reviews stay visible across refreshes and restarts. If an approved PR receives new commits, it reappears as done+ at the end of the queue without starting another AI review. Authenticated polling slows when API quota is scarce and speeds back up when it recovers, accounting for refresh cost and time until the quota resets.

Select a PR with the arrow keys, then work through the review:

1. Optionally press n to change the current configured AI reviewer for all PRs. Press r to start a background AI review. Reviews start only when you request them. The agent is instructed to read the PR description, discussion, and earlier review threads and replies before reviewing the local code, reporting any unavailable context. DiffButler caches full Git history and reviews changes from the PR merge base to its head, excluding unrelated changes made later on the target branch.
2. A successful review shows approve only when the AI explicitly reports no findings; otherwise it shows check. Press Enter to read the results, or c to resume the saved agent conversation. If the agent exits normally but DiffButler cannot parse its output, the task also shows check; use c to inspect the saved conversation or L for captured output. Changing reviewers leaves that conversation available through c; pressing r starts a new run with the current reviewer and replaces the current review details. Ask questions and explicitly tell the agent which comments to post, if any.
3. Exit the agent to return to the dashboard, which refreshes activity from GitHub. Press w if you are waiting for author changes, or o to review in your browser.
4. After approving on GitHub, press m to mark the task complete locally. DiffButler also detects your approval, closure, or merge and cleans up owned session resources while retaining review history.

Space performs the selected PR's next action: start a review, discuss findings, or open the browser. An AI review never counts as your approval. Background analysis uses the agent's native permission controls and read-only review instructions; source sent to the agent may go to its model provider.

• /: Filter the queue.
• a: Toggle monitoring all open PRs in configured repositories.
• b: Browse accessible PRs and follow one.
• R: Refresh from GitHub.
• k: Stop the selected PR's review and reset its local tracking and owned resources. It disappears until a forge refresh rediscovers it; review history is retained.
• h: View completed review history.
• ?: Show all keyboard shortcuts.
• q: Quit.

See the dashboard guide for all actions and configuration guide for settings.

Your own PRs are hidden from the review queue by default, including direct and team requests. Set include_own_prs = true in [ui] to review them, such as PRs an agent created using your account. PRs assigned to you are also hidden by default, even when you are requested directly or through a team. Set include_assigned_prs = true in [ui] to include them.

Reset

Close running DiffButler sessions, then run:

diffbutler reset

DiffButler refuses a second command or dashboard using the same data home, including reset, while another instance is running. Its instance lock releases on process exit; a leftover PID file does not block startup. See the locking details before handling Git lock files or resetting after a crash.

After confirmation, this permanently removes DiffButler's standard configuration, data, and state directories, including review history, saved sessions, and diagnostic logs. Custom files outside those directories are preserved. Use diffbutler reset --force to skip confirmation. Run diffbutler again to repeat first-run setup.

Contributing

See CONTRIBUTING.md for development setup and checks.

License

MIT.