Read-only by design
The reviewer can read files, search, and see the diff. It cannot write, execute, push, approve, or merge.
An extra reviewer that never takes the wheel.
Hammer reads the diff and the pull-request checkout with a small set of read-only tools, checks its own findings, and leaves one review with inline comments plus a summary. It never edits files, runs your code, pushes, approves, or merges.
Few, checked comments
A verification pass scores every candidate against the code and drops weak or out-of-diff findings before they post.
Incremental by default
Each push reviews only new changes, does not repeat earlier findings, and resolves threads whose code was fixed.
Your key, your runner
Anthropic, OpenAI, Gemini, Azure OpenAI, or any compatible gateway. Your code goes only to the model provider you configure.
Read-only is not a setting you can loosen.
The reviewer can read files, list directories, search, see the diff, and read the PR discussion. That is the whole tool set. There is no shell, no package install, no network tool, and no path outside the checkout.
Untrusted input
PR title, body, comments, filenames, and file contents are treated as data, never as instructions.
Read-only reviewer
A scoped PR-head checkout with hooks disabled and config read from the base branch.
review comments only
One review
Inline comments anchored to changed lines, plus one summary for the run.
No side effects
No file edits, command execution, package installs, pushes, approvals, or merges.
Narrow token
Contents read and pull requests write; optionally statuses write for comment-triggered reviews.
Pull requests from forks are not reviewed automatically by the default workflow, because they do not receive repository secrets. A maintainer can run /hammer review instead.
It checks its own findings before you do.
The first pass produces candidate findings. A second pass reopens the code, confirms the trigger, relocates the lines, and scores each finding from 0 to 10. Anything below the threshold never becomes a comment.
- 01
Review
Read the diff and the surrounding code with read-only tools.
- 02
Verify
Re-open the code, confirm the trigger, and score each finding from 0 to 10.
- 03
Post
Keep verified, in-diff findings as one review with inline comments and a summary.
- 04
Next push
Review only what changed, suppress repeats, and resolve threads whose code was fixed.
Scores, not confidence
A finding below verify_threshold is dropped, even if the reviewer was confident about it.
No duplicate noise
Fingerprints ignore line numbers, so code moving by ten lines does not repost the same finding.
Threads close when fixed
Threads whose flagged lines changed and whose finding was not reported again are resolved on the next run.
A review that reads like a review.
One high-severity finding, anchored to the line that caused it, with the fix in context. Everything else stays in the summary, so the diff remains readable.
acme/apiPublic
WatchFork 84Star 2.1k
Pull requests / #128
Retry webhook delivery with backoff Open
hammer[bot] reviewed just now
2 findingsHammer Review · 2 findings (1 high)
- High
src/webhook/client.ts:88Retries a non-idempotent POST without an idempotency key - Medium
src/webhook/client.ts:120Backoff has no upper bound
@@ -84,7 +84,13 @@ export async function deliver(event: WebhookEvent) {
- 8484
export async function deliver(event: WebhookEvent) { - 8585
const headers = { 'content-type': 'application/json' } - 86
for (let attempt = 1; attempt <= 3; attempt++) { - 87
await post(url, event, { headers, attempt }) - 88
} - 8689
} - hammer[bot] just nowHigh · correctness
Retries a non-idempotent POST without an idempotency key. A timeout after the server commits will deliver the webhook twice.
headers: { ...headers, 'Idempotency-Key': event.id },
@@ -118,4 +124,7 @@ async function backoff(attempt: number) {
- 118124
return 1000 * 2 ** attempt - 125
// retry immediately after a provider timeout
1 earlier thread resolved after its flagged lines changed.
Illustrative review with sample data. Posts as github-actions[bot] by default, or as hammer[bot] through the hosted platform.
Your rules, in your repository.
.github/hammer.yml holds rules, path instructions, ignores, severity thresholds, and review limits. Hammer reads it from the base branch, so a pull request cannot change the rules that review it. It also reads AGENTS.md, CLAUDE.md, and CONTRIBUTING.md as guidance.
1min_severity: medium2max_comments: 103fail_on: none4ignore:5 - "**/*.snap"6 - "docs/generated/**"7rules:8 - id: no-raw-sql9 text: Queries must go through the db helper, never string-built SQL.10path_instructions:11 - path: "src/api/**"12 text: Check that every handler validates input and returns typed errors.13verify: true14verify_threshold: 6Base branch wins
Config and repository guidance come from the base branch, not from the pull request under review.
Rule ids show up
A finding that cites a house rule names its id in the comment, so the rule stays traceable.
Gate on purpose
Set fail_on: high to fail the job and let branch protection block the merge.
Your key. Your runner. Your code path.
Hammer runs inside your GitHub Actions workflow with the model key you already have. Your code is sent only to the model provider you choose, and Hammer collects no telemetry.
Model clients
Anthropic
OpenAI'%20fill-rule='nonzero'%3e%3c/path%3e%3c/svg%3e)
Gemini'%3e%3c/path%3e%3cpath%20d='M18.397%2015.296H7.4a.51.51%200%2000-.347.882l7.066%206.595c.206.192.477.298.758.298h6.226l-2.706-7.775z'%20fill='%230078D4'%3e%3c/path%3e%3cpath%20d='M15.272.857H7.497L0%2023.071h7.775l1.596-4.73%205.068%204.73h6.665l-2.707-7.775h-7.998L15.272.857z'%20fill='url(%23lobe-icons-azure-fill-1)'%3e%3c/path%3e%3cpath%20d='M17.193%201.613a1.11%201.11%200%2000-1.052-.756h-7.81.035c.477%200%20.9.304%201.052.756l6.748%2019.992a1.11%201.11%200%2001-1.052%201.466h-.12%207.895a1.11%201.11%200%20001.052-1.466L17.193%201.613z'%20fill='url(%23lobe-icons-azure-fill-2)'%3e%3c/path%3e%3cdefs%3e%3clinearGradient%20gradientUnits='userSpaceOnUse'%20id='lobe-icons-azure-fill-0'%20x1='8.247'%20x2='1.002'%20y1='1.626'%20y2='23.03'%3e%3cstop%20stop-color='%23114A8B'%3e%3c/stop%3e%3cstop%20offset='1'%20stop-color='%230669BC'%3e%3c/stop%3e%3c/linearGradient%3e%3clinearGradient%20gradientUnits='userSpaceOnUse'%20id='lobe-icons-azure-fill-1'%20x1='14.042'%20x2='12.324'%20y1='15.302'%20y2='15.888'%3e%3cstop%20stop-opacity='.3'%3e%3c/stop%3e%3cstop%20offset='.071'%20stop-opacity='.2'%3e%3c/stop%3e%3cstop%20offset='.321'%20stop-opacity='.1'%3e%3c/stop%3e%3cstop%20offset='.623'%20stop-opacity='.05'%3e%3c/stop%3e%3cstop%20offset='1'%20stop-opacity='0'%3e%3c/stop%3e%3c/linearGradient%3e%3clinearGradient%20gradientUnits='userSpaceOnUse'%20id='lobe-icons-azure-fill-2'%20x1='12.841'%20x2='20.793'%20y1='1.626'%20y2='22.814'%3e%3cstop%20stop-color='%233CCBF4'%3e%3c/stop%3e%3cstop%20offset='1'%20stop-color='%232892DF'%3e%3c/stop%3e%3c/linearGradient%3e%3c/defs%3e%3c/svg%3e)
Azure OpenAI- OpenAI-compatible
Any Linux runner
A bundled Node 24 action on GitHub-hosted, Blacksmith, or self-hosted Linux, x64 and arm64.
Visible spend
Every run reports tokens and, when pricing is known, cost. Add token or cost limits when you want a hard stop.
One outbound path
Your code goes only to the configured model provider. There is no Hammer telemetry.
Post as hammer[bot] without building an App.
The optional hosted platform issues a short-lived GitHub App token, scoped to one repository, so Hammer can post as the bot. A repository admin signs in with GitHub, installs the Hammer App, and registers the repository. The platform never sees your code or your model key.
Your workflow
Runs the Action with your model key and an optional repository key.
Token broker
Verifies the repository registration, then mints one short-lived installation token.
installation token
hammer[bot]
The review is attributed to the Hammer App instead of github-actions[bot].
Rotate a key
Regenerate from the console; the old key stops working immediately.
No App to build
You do not create, install, or maintain a GitHub App yourself.
Optional. Without a key, Hammer reviews with the workflow token and posts as github-actions[bot].
One workflow file. Then open a pull request.
Add a model key as a repository secret, add the workflow, and let Hammer review the next PR. The bot identity and repository rules are optional.
- 01
Add your model key
Store ANTHROPIC_API_KEY, OPENAI_API_KEY, or the key for the provider you use as a repository secret.
- 02
Add the workflow
Use anvia-hq/hammer@v1 with client, model, and api-key. No checkout step is needed.
- 03
Optional: post as Hammer
Install the Hammer App, register the repository, and pass its key as hammer-api-key.
- 04
Open a pull request
Hammer reviews it. Add the command trigger to re-run a review with /hammer review.
1name: Hammer Review23on:4 pull_request:5 types: [opened, synchronize, reopened, ready_for_review]67permissions:8 contents: read9 pull-requests: write1011jobs:12 review:13 concurrency:14 group: hammer-review-${{ github.event.pull_request.number }}15 cancel-in-progress: true16 runs-on: ubuntu-latest17 timeout-minutes: 3018 steps:19 - uses: anvia-hq/hammer@v120 with:21 client: anthropic22 model: claude-sonnet-5-523 api-key: ${{ secrets.ANTHROPIC_API_KEY }}24 hammer-api-key: ${{ secrets.HAMMER_API_KEY }}The workflow guide adds the issue_comment trigger for /hammer review, /hammer review <focus>, and /hammer help. Commands need write access to the repository.
v1, in the open.
Hammer is dogfooded on its own repository and has run live reviews on GitHub-hosted Linux x64. A first 31-case synthetic evaluation passed its automated metric gate. Human review of review quality, the full evaluation corpus, and wider runner acceptance are still open.
- Source and releasesAGPL-3.0 source, issue tracker, and tagged releases.Open GitHub
- Security coverageHow each control is tested, from untrusted input to comment sanitizing.Read the coverage
- Sandbox evidenceThe live reviews, measured baseline, and what remains open.Read the evidence
Hammer is AGPL-3.0. The hosted token broker is a separate, private service.
