Give Claude context for PR reviews

claude-pr-review is a GitHub Action that wraps anthropics/claude-code-action with everything it needs to act as a recurring PR reviewer, not just a one-shot summarizer. Add it to a workflow and Claude will:

• See the whole conversation, not just the diff. The full PR discussion — every comment and thread, each shown with its nearest diff fragment — is rendered with gh-pr-render and added to the prompt. Claude can see what’s already been said and responds rather than repeating a nearly identical review.
• Re-review only what changed. On every push, the new diff is compared against what was last reviewed. A rebase with no real code changes triggers nothing; a real change gets Claude a diff-of-diffs showing exactly what’s new, so a re-review addresses the update, not the entire PR again.
• Post like a real review, as a real reviewer. A bundled CLI lets Claude queue inline comments and submit them as a single grouped review — not one review per comment, which is what you get from anthropics/claude-code-action’s built-in inline-comment tool.
• Fail safe. If Claude’s turn errors out partway through, a final sweep step still posts whatever was queued, and any pending approval is automatically downgraded to a plain comment rather than going out unearned.
• Stay inside the bounds you set. The default configuration restricts what Claude can do while allowing you to pick and choose additional specialized tooling appropriate to your repo.

- uses: danielparks/claude-pr-review@fdca20601bf0d709ee55c3b799c2cdcc9adcccf7 # v1.0.2
  with:
    # You can set either of these secrets. If both are set, the API key
    # seems to win.
    anthropic-api-key: ${{ secrets.ANTHROPIC_API_KEY }}
    claude-code-oauth-token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}

See example workflow-review.yaml for a working PR-triggered workflow, and example workflow-response.yaml for a working comment (@claude) triggered workflow. This repo uses both.

Anthropic API key for token-based billing. Either this or claude-code-oauth-token is required.

Claude Code OAuth token for use with a subscription account. Generate with claude setup-token. Either this or anthropic-api-key is required.

Prompt sent to Claude for both initial reviews and re-reviews. Claude determines the current state of the PR from context. Defaults to:

Please review this pull request. /tmp/pr.patch contains the overall diff
for the PR, the PR branch is checked out in the current working directory,
and the PR discussion is included below.

If there are prior review comments from you in the PR discussion, treat
this as a re-review. Respond to discussion as appropriate, and review new
changes. If nothing has changed there is no need to note it unless the
user seems to think that changes were made.

Only post GitHub comments — don't submit review text as messages.

For the commands below: write the comment body to a file under
/tmp/pr-review-scratch first with the Write tool, then pass that file's
path with `--body-file`. This avoids shell special characters.

1. Use `pr-review reply-inline-comment --comment-id ID --body-file PATH`
   to reply to existing inline comments if they have been resolved, if
   changes have been made that haven't resolved them, or if you have new
   information. This posts immediately.
2. Use `pr-review queue-inline-comment --path PATH --line N --body-file
   PATH` for every specific code issue you find in new code. This only
   queues the comment; nothing is posted yet.
3. Use `pr-review comment-review --body-file PATH` once at the end if
   you want to leave top-level feedback. This will submit all queued
   comments; if you don't call it queued comments will be submitted
   after you finish. Do not repeat comments made inline, though you may
   briefly list issues that remain unresolved from a previous review.

If it's useful you may link to specific comments, but don't refer to them
by ID — the user does not have easy access to that.

- `#discussion_rID` — an inline comment.
- `#pullrequestreview-ID` — a top-level review comment.
- `#issuecomment-ID` — a top-level non-review comment.

Appended after prompt. Empty default.

This is useful if you just want to add text to the default prompt.

List of tools to allow Claude to use, one per line. Passed to --allowedTools. Defaults to:

Read
Bash(find:*)
Bash(grep:*)
Bash(git log:*)
Bash(git diff:*)
Bash(git blame:*)
Bash(gh pr diff:*)
Bash(gh pr view:*)
Bash(gh run list:*)
Bash(gh run view:*)
Bash(pr-review --help:*)
Bash(pr-review queue-inline-comment:*)
Bash(pr-review reply-inline-comment:*)
Bash(pr-review comment-review:*)
Bash(pr-review list-queue:*)
Bash(pr-review discard-queue:*)
Bash(pr-review post-comment:*)
Bash(pr-review list-available-labels:*)
Bash(pr-review add-label:*)
Bash(pr-review remove-label:*)

# From https://code.claude.com/docs/en/agent-sdk/permissions.md:
#
# > Use `//path` for an absolute filesystem path: a deny rule of
# > `Edit(//secrets/**)` blocks writes anywhere under `/secrets` on disk.
# > With a single leading slash, `Edit(/secrets/**)` anchors at the rule's
# > source instead.
Edit(//tmp/pr-review-scratch/**)

# Needed to view long inline comment threads:
Bash(gh-pr-render:*)

List of additional tools to allow Claude to use, one per line. Added to allowed-tools and then passed to --allowedTools. Empty default.

This is useful if you just want to add a tool to the default tool list. In particular, see the “Opt-in tools” section of cli/README.md for pr-review subcommands that are deliberately left out of the default list.

Newline-separated glob patterns, e.g. Claude: *, that select which labels Claude is allowed to add or remove from a PR. Only labels whose names match at least one pattern are available; Claude cannot add any others.

Only * (any run of characters) and ? (any single character) are supported instead of full shell glob syntax, so [abc] and {a,b} are matched literally rather than as character classes or alternatives.

When set, fetches matching labels (with descriptions) from the repo and injects them into Claude's prompt so it knows what's available.

When blank (the default), label management is disabled: add-label and remove-label will always fail.

Set to * to allow all repo labels.

Examples:

# Allow all labels
available-labels: "*"

# Allow only labels prefixed with "Claude: "
available-labels: "Claude: *"

# Allow several specific labels
available-labels: |
  bug
  enhancement
  needs-work

# Explicitly disable (same as the default, but self-documenting)
available-labels: ""

Version of the gh-pr-render tool to use. Defaults to the latest version at the time of this action’s release.

How long to keep the Claude thinking log artifact in days. disabled or 0 means not to create it.

The thinking log is an HTML file that contains some of Claude’s internal thoughts and all of its tool uses.

This attempts to redact GitHub tokens, Anthropic API keys, and other Anthropic secrets. There is no guarantee that it will succeed — keys could be broken up into multiple strings.

Defaults to: disabled

This action bundles its own CLI tool, cli/pr-review, that action.yaml puts on PATH for Claude to run via its Bash tool:

• pr-review queue-inline-comment queues an inline comment on disk. Nothing is posted to GitHub until comment-review is called.
• pr-review comment-review posts everything queued by queue-inline-comment, plus an optional top-level body, as a single grouped comment review. Claude doesn’t have to call this itself — action.yaml runs it automatically after Claude’s turn ends if anything is still queued.
• pr-review reply-inline-comment replies to an existing inline comment thread; this posts immediately, since replies attach to an existing thread rather than a new review.
• pr-review post-comment can be used to post a new top-level comment immediately. This differs from gh pr comment in that it doesn’t have the ability to edit or delete existing comments. It is not mentioned in the default prompt.
• pr-review list-queue and pr-review discard-queue --dir PATH let Claude recover if a submission fails for a reason that won’t change on retry (e.g. GitHub rejecting an inline comment’s line number): after fixing the problem and resubmitting successfully, Claude can discard the original failed batch so the automatic post-turn sweep doesn’t keep retrying — and failing on — it.
• pr-review add-label LABEL and pr-review remove-label LABEL add and remove labels. Only labels matching the available-labels input are allowed; label management is disabled when that input is blank. pr-review list-available-labels shows the available labels and is also used by action.yaml to inject label context into Claude’s prompt.
• pr-review --help lists every command; pr-review <command> --help shows that command’s flags.

anthropics/claude-code-action’s built-in inline-comment tool posts each inline comment through GitHub’s single-comment REST endpoint, which creates and submits its own standalone review every time — so a review with five comments shows up as five separate reviews instead of one.

This is a Bash-invoked CLI tool rather than an MCP server so that it can post as claude[bot] instead of github-actions[bot]: the Claude App token only reaches subprocesses that inherit anthropics/claude-code-action’s real environment, which a Bash-tool call does and an MCP server does not. It’s plain, dependency-free JavaScript that action.yaml runs directly with no build step — see cli/README.md for more information, including the opt-in pr-review subcommands that are left out of the default allowed-tools list.

When a PR is updated this compares the version of the PR to the new version. If they are identical, e.g. the PR was rebased on changes to another part of the codebase, it does not trigger Claude to re-review.

If the changes are not identical then this provides Claude with a diff-of-diffs so that it can see what was changed.

The most obvious alternative is anthropics/claude-code-action, on which this action is built. This action was created to close a few specific gaps in it:

• It has no built-in way to skip re-review on a no-op push (e.g. a rebase), or to show Claude only what changed since the last review. This action’s diff caching and diff-of-diffs do that.
• It has no built-in way of providing discussion context to inform re-reviews.
• Its built-in inline-comment tool posts a separate review per comment instead of one grouped review (claude-code-action#1049). The cli/pr-review queue/claim/sweep design exists specifically to fix this.

PR-Agent/Qodo Merge is the closest open-source, self-hosted peer, and supports several LLM providers rather than being Claude-specific.

There are several proprietary SaaS alternatives. The big ones are:

• CodeRabbit
• Greptile
• GitHub’s own Copilot code review

Unless otherwise noted, this project is dual-licensed under the Apache 2 and MIT licenses. You may choose to use either.

• Apache License, Version 2.0
• MIT license

Unless you explicitly state otherwise, any contribution you submit as defined in the Apache 2.0 license shall be dual licensed as above, without any additional terms or conditions.