Setting Fletcher up

Get started

1. Sign in with GitHub.
2. Install the Fletcher GitHub App on your account or organization, then pick the repositories it can see.
3. Back in onboarding, choose which repositories get reviews. Every repository you gave the App starts with reviews on; an admin can turn them off.
4. Open a pull request. Each workspace gets 5 free reviews on Fletcher's key. After that, an admin adds a Gemini key in the workspace settings or subscribes to the paid plan (see pricing).

No .fletcher.json is needed. Without one, Fletcher uses the defaults below.

When Fletcher reviews

• When a pull request is opened or reopened, or when new commits are pushed to it. On a new push, Fletcher remembers what he said last time.
• Only when the pull request's author is an owner, member, or collaborator on the repository. Pull requests from anyone else, such as outside contributors' forks, aren't reviewed on their own.
• On demand: an owner, member, or collaborator comments /fletcher again on the pull request. Fletcher reviews the latest commit if it hasn't been reviewed yet (for example, an outside contributor's pull request), or retries a review that failed.

Fletcher reviews each commit once. /fletcher again on a commit that already has a review doesn't start another one; push a new commit to get a new review.

How Fletcher answers /fletcher again

Fletcher answers every /fletcher again from an owner, member, or collaborator, so you can tell it was heard:

Comments from anyone else are ignored without an answer.

Reviews are posted as comments. Fletcher never approves or blocks a pull request, so branch protection rules don't change.

.fletcher.json

Put .fletcher.json at the root of the repository. Fletcher reads it from the default branch, never from the pull request, so a pull request can't turn its own review off or hide files. Changes to the file take effect once they're merged.

Every field is optional. A typical file:

{
  "severityThreshold": "warning",
  "intensity": "carnegie",
  "ignore": [
    "*.md",
    "**/*.md",
    "**/__generated__/**",
    "fixtures/**"
  ]
}

The defaults, written out in full:

{
  "enabled": true,
  "severityThreshold": "suggestion",
  "intensity": "studio_band",
  "tone": "standard",
  "ignore": [],
  "guidelines": null
}

To turn reviews off for a repository:

{ "enabled": false }

Ignoring files

Each glob is matched against the file's full path from the repository root, with no leading / or ./: /docs/** matches nothing.

• * and ? stay inside one folder: *.md matches README.md but not docs/guide.md.
• ** crosses folders: docs/** matches everything under docs/.
• **/*.md needs at least one folder, so it doesn't match README.md at the root. To match a file type everywhere, list both *.md and **/*.md.
• Braces ({a,b}), character classes ([ab]), and negation (!) aren't supported and match literally.

These are always ignored, with or without a config file:

• Lockfiles: pnpm-lock.yaml, package-lock.json, yarn.lock, bun.lock, bun.lockb
• Minified files: *.min.js, *.min.css
• Images, PDFs, and web fonts (png, jpg, jpeg, gif, webp, ico, pdf, woff, woff2)
• Test snapshots (*.snap)
• .sql files and meta/ inside any drizzle/ folder, and worker-configuration.d.ts

Nothing else is ignored by default, including other ecosystems' lockfiles (Cargo.lock, go.sum, poetry.lock) and SVGs. Add those to ignore if you don't want them reviewed.

What Fletcher reads

Besides the diff, each review gets:

• Your rules. AGENTS.md, CLAUDE.md, CONTRIBUTING.md, .github/CONTRIBUTING.md, docs/CONTRIBUTING.md, and .github/copilot-instructions.md, plus any AGENTS.md or CLAUDE.md in a folder the pull request changes. They're read from the default branch, so a pull request can't rewrite the rules it's reviewed against. Up to 8 files and 40 KB in total. Set guidelines to pick the files yourself, for example ["docs/style/**"].
• Your dependencies. The root manifest (package.json, go.mod, pyproject.toml, Cargo.toml, and the like) and the nearest one to each changed file, so library versions newer than the model's training aren't flagged as mistakes.
• The changed files in full, as they are in the pull request, so code just outside the diff counts. Up to 25 files and 150 KB, most-changed first; files over 60 KB are reviewed from the diff only.

Findings that cite a rule quote it and name its file. Fletcher still comments only on lines the pull request adds, and these files don't count toward the 300 KB diff limit or your credits.

Things that catch people out

• One bad value resets everything. If the file isn't valid JSON, or any field has a value not listed above, Fletcher ignores the whole file and uses the defaults, including "enabled": true and no ignores. Nothing is posted about it, so check the file carefully.
• Plain JSON only. No comments and no trailing commas.
• Misspelled fields are skipped without a warning. "severity" or "ignored" does nothing.
• Very large pull requests aren't reviewed. The limit is 300 KB of diff, and it counts every file, including ignored ones and lockfiles. Split big changes, or move large generated changes into their own pull request.
• Edits in a pull request don't apply to that pull request. Merge the change to .fletcher.json first. Commits Fletcher has already reviewed or skipped stay that way; the new settings apply from the next push.