Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    PR Review

    1

    Guided GitHub pull request review: finds issues by the criteria you pick, walks you through each finding with solutions and draft inline comments, and publishes one review via gh only after you confirm.

    Free

    Security scanned
    PR Review

    PR Review

    Example session with this skill installed

    /pr-review 42

    • Read your context and instructions
    • Compiled the pr review

    Connected to GitHub as

    @reviewer. Review language:

    English. Criteria:

    Bugs/logic,

    Security,

    Performance.

    PR #42 in acme/billing-api · Add invoice export endpoint · 7 files changed, +312 −48

    Summary

    PR #42 · 4 findings

    SeverityCount
    🔴 Critical1
    🟠 High1
    🟡 Medium1
    🔵 Low1
    • #1 🔴 critical: API token written to logs (src/export/client.ts:15-19)
    • #2 🟠 high: Missing null check on customer profile (src/export/invoice.ts:42)
    • #3 🟡 medium: N+1 query when loading invoice lines (src/export/repository.ts:88-97)
    • #4 🔵 low: Magic number for page size (src/export/controller.ts:23)

    Session saved, so you can stop and resume this review at any time.


    Finding 1/4 · 🔴 Critical · Security

    API token written to logs

    This is the problem:
    The export client logs the full request config, which includes the Authorization header. Every export call writes the customer's API token to the application logs, where anyone with log access can read it.

    // src/export/client.ts:15-19
    const config = { headers: { Authorization: `Bearer ${token}` }, timeout: 5000 };
    logger.info('Calling export service', config);
    return http.post(EXPORT_URL, payload, config);
    

    These are the possible solutions:

    S1: Remove the log of the request config (recommended)
    Log only non-sensitive fields such as the URL and invoice id.

    💬 src/export/client.ts:15-19 (inline)
    This logs the whole config, including the Authorization header, so the token ends up in our logs. Could we log only the URL and the invoice id here?

    S2: Mask sensitive headers in the logger
    Add a redaction step in the shared logger so headers like Authorization are always masked.

    💬 src/export/client.ts:15-19 (inline)
    The token is being logged here. Besides fixing this call, it might be worth masking Authorization in the logger so it can't happen again.
    💬 src/logger.ts:8 ⚠️ outside the diff → will go in the review body
    We could add a redaction list (Authorization, Cookie, x-api-key) to the formatter.

    → Which solutions do you want to include in the review? S1 (Recommended) · S2 · Discard finding


    (Findings 2-4 follow the same flow.)

    Final review

    #SeverityCommentLocation
    1🔴Don't log the Authorization headersrc/export/client.ts:15-19
    2🟠Guard against a missing customer profilesrc/export/invoice.ts:42
    3🟡Load invoice lines in one querysrc/export/repository.ts:88-97

    Recommended event: REQUEST_CHANGES (there are critical/high findings).
    Signature check passed: no bot signatures or attribution footers in any comment.

    → Publish review · Go back and review · Cancel

    Next steps

    • Confirm Publish review to post 3 inline comments as one REQUEST_CHANGES review on PR #42.
    • The skill prints the review URL and removes the local session file.
    • Prefer to finish later? Just stop; run /pr-review again and choose Continue review of #42.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Reviewing someone else's pull request well takes time: reading the diff, spotting real problems, explaining them clearly and writing comments that land on the right lines. PR Review turns that into a guided, step-by-step session where the agent does the heavy lifting and you decide everything. Nothing is posted to GitHub until you explicitly confirm at the very end.

    What it does

    • Criteria-driven analysis: pick what the review should focus on: Bugs/logic, Security, Performance, SOLID, DRY, KISS, Scalability, Atomic Design, or your own custom focus (accessibility, testing, naming...).
    • Prioritized findings: every finding gets a severity (🔴 critical, 🟠 high, 🟡 medium, 🔵 low) and an exact path:line location.
    • Problem + solutions, one at a time: for each finding you see the explanation, the relevant snippet and up to 3 alternative solutions, each with its own draft inline comments.
    • You choose and edit: pick one or more solutions, edit any comment, or discard the finding.
    • Smart comment placement: comments outside the diff are moved to the review body automatically, and duplicates on the same lines are merged.
    • Two languages: discuss the review in one language and write the PR comments in another (Spanish / English).
    • Your voice: comments are written as the reviewer, with no bot signatures or attribution footers; a final check strips them before publishing.
    • Resumable sessions: progress is saved locally, so you can pick up an interrupted review later. You get a warning if the PR received new commits in the meantime.
    • One clean review: approved comments are published as a single GitHub review (REQUEST_CHANGES, COMMENT or APPROVE) with a recommended event based on severity.

    How it works

    1. Run /pr-review, /pr-review 42 or /pr-review <PR URL> (with no argument, choose from the open PRs).
    2. The skill checks your gh connection, then asks for the review language and criteria.
    3. It loads the PR files and diff, analyzes them and shows a summary table of findings.
    4. Walk through each finding: choose solutions, edit or approve the comments.
    5. Review the final list, pick the event and an optional general message.
    6. Preview the full review and confirm. Only then is it published.

    Built-in safety

    • Read-only on GitHub until you confirm: only read commands run before the final Publish review.
    • Never touches the code under review: no edits, commits, checkouts or branch switches.
    • Strict input validation: PR numbers, owners, repos, SHAs and file paths are validated before reaching any shell command; unsafe file names are skipped and flagged.
    • Prompt-injection resistant: PR titles, bodies, diffs and repo instructions are treated as data, never as instructions. Text that tries to steer the reviewer is reported as a security finding.
    • No secrets accessed: uses only your existing gh session; no tokens, no extra MCP servers.

    Requirements

    • GitHub CLI (gh) installed and authenticated (gh auth login).
    • An agent that supports SKILL.md skills with interactive multiple-choice questions (built and tested on Claude Code).

    Open source

    Part of abis-skills (MIT): https://github.com/gabrielciprianoo/abis-skills. Install every skill globally with npx abis-skills.

    Requires the GitHub CLI (gh) installed and authenticated. Reviews pull requests only: it never edits code, commits or checks out branches. Only open PRs can be reviewed, and listing mode shows up to 100 open PRs (pass a PR number or URL for others). Files with unsafe characters in their paths are skipped and flagged for manual review. Comment language options are Spanish and English.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 20 days ago

    • Free with an Agensi account

    Requires

    GitHub
    GitHub CLI (gh)

    Permissions

    Terminal / Shell
    Network Access
    Write Files
    Read Files

    Allowed Hosts

    api.github.com

    File Scopes

    ~/.claude/pr-review/sessions/*.json

    Terminal/Shell: runs gh commands to read the PR (view, diff, file contents) and, only after explicit confirmation, a single POST to create the review. Read/Write files: saves resumable review sessions in ~/.claude/pr-review/sessions/ and writes the review payload file before publishing. Network: reaches the GitHub API through gh. The skill never modifies the reviewed code and never reads local secrets.

    Listed20 days ago

    What's inside

    Frequently Asked Questions

    The GitHub CLI (`gh`) installed and authenticated with `gh auth login`, plus an agent that supports SKILL.md skills. It uses your existing gh session: no extra API keys, tokens or MCP servers.

    No. Everything before the final step is read-only. You see a full preview of the review (event, body and every inline comment) and it is published as a single review only after you pick "Publish review". Cancel at any point and nothing is posted.

    Yes. Progress is saved after every step in ~/.claude/pr-review/sessions/. Run /pr-review again and choose "Continue review". If the PR received new commits in the meantime, you are warned and can start over or continue anyway.

    Comments are drafted in the reviewer's voice and you can edit every one of them. Before publishing, a check removes bot signatures, attribution footers and assistant-style self-references, so the review reads like one you wrote.

    Yes. You choose the language for the conversation and, separately, the language for the PR comments (Spanish or English). The skill recommends the language used in the PR title and description.

    New in Code Quality & Review