ai code review

Findings, not opinions

Every review Buddy posts is anchored to a line you actually changed and names the failure it causes. No nitpicks about brace style, no summary of what the diff already says, and never the same finding twice.

Get Started Review locally first vs CodeRabbit

$ buddy review 128

             __
    (\,------'()'--o    reading pull request #128
     (_    ___    /~"   14 files, 412 changed lines
      (_)_)  (_)_)

  src/auth/session.ts:82  major  correctness
    Token refresh drops the error, so an
    expired session reads as a good login.

  src/queue/worker.ts:31  minor  correctness
    The retry counter resets inside the loop,
    so a failing job retries forever.

  .github/workflows/ci.yml:19  minor  security
    Interpolates a PR title into `run:`.

  3 findings. 2 inline, 1 already reported.

buddy  commented on src/auth/session.ts

  major · correctness

  `refreshToken()` returns `null` on a network
  error and on an expired token alike, and the
  caller treats `null` as "no refresh needed".
  An expired session therefore reads as a valid
  login until the next hard reload.

  Distinguish the two, or throw on the error
  path so the caller cannot silently continue.

  ✔ resolved by a later push

📍

Anchored to your diff

Buddy reviews the lines the pull request changed, with enough surrounding context to judge them. A finding that cannot be attached to a changed line is reported in the summary rather than pinned to an innocent line, so nothing lands where it did not belong.

🎚️

Severity that means something

major, minor and nit. A major finding is a bug you would revert for. --fail-on turns that scale into an exit code, so severity is a gate rather than a decoration.

♻️

Incremental by design

Buddy records what it already said. Push again and you get findings about the new changes — not a fresh copy of the review you read this morning.

🧰

Analyzers included

Secret scanning, actionlint, shellcheck, hadolint, markdownlint and syntax checks run alongside the model, so one command reports everything instead of you merging two reports by hand.

✅

Suggestions you can commit

Where a fix is unambiguous, the finding carries a committable suggestion. Locally, buddy review --fix applies them bottom-up so one edit cannot shift the line numbers of the next.

🔒

Prompt injection is assumed

Pull request bodies, comments and contributor branch content are written by third parties. They never reach the system prompt — they arrive as tool output inside an explicit untrusted-content marker, with the escape sequence neutralised so a payload cannot close the block early and appear to be trusted context.

What a review run does

buddy review 128            # review a PR and post the result
buddy review 128 --dry-run  # print it instead of posting

In CI it is one step in a workflow you own:

- uses: actions/checkout@v4
  with: { fetch-depth: 0 }
- uses: oven-sh/setup-bun@v2
- run: bunx @buddysh/buddy review ${{ github.event.pull_request.number }}
  env:
    GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
    ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}

The diff is read by your runner, sent to the provider you configured, and the findings are posted by your token. There is no Buddy server in the path, because there is no Buddy server.

Pick the model

// buddy.config.ts
export default {
  ai: {
    provider: 'anthropic',
    model: 'claude-opus-5',
    maxTokensPerRun: 200_000,
  },
} satisfies BuddyConfig

Anthropic, OpenAI, Google, OpenRouter and any OpenAI-compatible endpoint are supported. Model aliases (opus, sonnet, haiku) resolve to current Anthropic models; every other provider takes a concrete ID. See AI providers.

Teach it your house rules

Review guidelines are read from your repository, so the standards a reviewer enforces are the ones your team wrote down rather than a vendor's defaults. @buddy remember ... appends a note that later runs read back, which is how a one-off correction in a thread becomes a standing rule.

What it deliberately does not do

It does not rewrite your code. Review mode draws only from the read and comment tool tiers; a write tool is never advertised to the model, so it cannot be requested at all. Changing code is fix-ci and finishing touches, and both are opt-in.
It does not block on the model. A provider outage degrades the review to the analyzers rather than failing the run.
It does not act on a drive-by comment. @buddy commands are permission-checked against the actor's access to the repository.

In-thread conversations · Local review · Merge gates · The agent runtime · Buddy vs CodeRabbit