Buddy

Review

AI Code Review Inline findings anchored to the lines you changed Conversations @buddy answers, re-reviews, resolves and pauses Local Review Read the working tree before the PR exists Finishing Touches Tests, docstrings and autofix, as a stacked PR

Automate

Merge Gates A real check run, not a comment nobody reads CI Repair Classify a failing run and fix what is unambiguous Dependency Updates Eleven ecosystems, real changelogs, auto-merge

Platform

Workflow Security Supply-chain footguns in .github/workflows Your CI, Your Keys No hosted app, no diff leaving your pipeline
Nine jobs, one binary. All features →

By team

Open Source A reviewer a maintainer cannot hire Startups No spare reviewer, no time for updates Platform & Enterprise Standards as check runs, on your infrastructure Agencies Twenty client repositories, one config

By job

Monorepos Every manifest at every depth, no setting Security & Compliance Licences, advisories and supply-chain audits Migrating Bring the Renovate config you already wrote AI Coding Agents Close the loop before the PR exists
Find the shape of your problem. All use cases →
GuideCLIAIConfigAPIGitHub
stacked Never a surprise commit on your branch

finishing touches

The work everyone means to do

Docstrings, tests, the simplification you noticed and skipped. Tick a box on the pull request and Buddy does one of them — then hands it back as a stacked pull request targeting your branch, so accepting it is a merge and refusing it is a close.

Get Started How the agent is sandboxed
buddy  commented on #128

  Finishing touches — tick one and I will
  open it as a pull request against this
  branch.

  - [ ] unit-tests   cover the changed behaviour
  - [ ] docstrings   document the changed APIs
  - [ ] autofix      apply the review's fixes
  - [ ] simplify     tidy without changing behaviour
  - [ ] plan         write an implementation plan

─────────────────────────────────────────────

buddy  commented 3 minutes ago

  **Write tests covering the changed behaviour**
  — opened #129 targeting this branch.

  3 files changed · `bun test` passed (214)
🧪

unit-tests

Writes tests for the behaviour the branch changed, following the suite's existing structure and naming — then runs them. A generated test that does not pass is worse than none, so a failure is reported rather than committed.

📚

docstrings

Documents the public APIs the branch changed, matching the density and conventions of the surrounding code rather than importing a different house style. Untouched code is left alone.

🩹

autofix

Applies exactly what the review's findings describe — and nothing else. An autofix commit that also refactors is one a reviewer has to disentangle.

✂️

simplify

Removes duplicated logic and unnecessary indirection in the changed code, and stops where a simplification would alter an edge case. Delivered as a suggestion by default, because taste is not a bot's call.

🗺️

plan

Names the files that would change and why, the order of the work, the risks, and how the result gets verified — written so another agent could execute it. Runs in review mode, so it cannot change anything.

✅

It verifies its own work

Every touch that changes code runs the project's tests or build before reporting done, and reports the failure instead of claiming success when verification fails. Buddy also reconciles what the agent says it changed against what actually changed on disk — a report that does not match the working tree is not trusted.

Delivered as a stacked pull request

A touch never appends a quiet commit to the branch you are reviewing. It opens a branch of its own — <your-branch>/buddy-<touch> — and a pull request targeting your branch, labelled buddy and finishing-touch.

That shape is the whole point:

  • Accepting is a merge. The diff is reviewed like any other.
  • Refusing is a close. Nothing to revert, nothing to force-push.
  • Your branch is unchanged until you say otherwise, so a touch cannot invalidate the review already in flight.

Touches whose output is a suggestion rather than a commit — simplify and plan — never open a branch at all. They post what they would do and stop.

Running one

From the thread, tick the checkbox Buddy posts. From the CLI:

buddy touch 128                      # read the ticked boxes and run them
buddy touch 128 --name unit-tests    # run one directly
buddy touch 128 --test-command "bun test"
buddy touch 128 --dry-run            # report what would run

--test-command names the command that verifies the change. Without it Buddy uses what it can discover in the project, and a repository with no test command gets a touch that says so rather than one that claims a green run it never had.

The sandbox is the same one

Touches that write run in implement mode; plan runs in review mode and has no write tool available to request. Both are bounded by the agent runtime's limits: an allowlisted environment with no credentials in it, a workspace no path can escape, and independent caps on tool calls, wall clock and tokens.

AI code review · Conversations · CI repair · The agent runtime