More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Review This PR

    1

    Your AI agent reads the author's summary and the project rules before it reviews a pull request.

    $7

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Review This PR

    Review This PR

    Example session with this skill installed

    Use the review-this-pr skill to review this branch.

    • Read your context and instructions
    • Compiled the review this pr

    This is a sample repository. It has a rules file (CLAUDE.md), a branch that adds reset(key) to an alert throttle, and the author's summary for that branch. The review below is the real, unedited output.


    I reviewed feat/throttle-reset against main and found one blocking bug: reset("disk") clears the throttle for every alert, not just "disk". You asked for scrutiny of reset(), and that's where it is. Nothing in your notes covers it.

    What I checked: I read your context notes (.claude/review-context/feat-throttle-reset.md) and the rules in CLAUDE.md, and looked at all three changed files. I ran python3 -m unittest -q and all 4 tests pass, as your notes say. There's no PR description to compare against because the repo has no git remote.

    Blocking

    1. reset(key) ignores key and clears every key (src/alert_throttle.py:20)

    def reset(key):
        """Let the next alert for `key` through, e.g. after an operator acks it."""
        _last_sent.clear()
    

    The docstring says it resets one key, but _last_sent.clear() empties the whole table. I ran this to confirm:

    should_send("disk", now=0); should_send("cpu", now=0)
    reset("disk")
    should_send("cpu", now=10)   # → True; should be False
    

    In production, acking one alert would release every throttled alert at once. That's the alert storm your notes say the 300 s window exists to prevent. The fix is _last_sent.pop(key, None).

    2. The new test can't catch this (tests/test_alerts.py:20-23)

    test_reset_lets_next_alert_through only uses "disk", so a reset that clears everything still passes. CLAUDE.md rule 2 asks for a test of each behaviour change, and the behaviour here is "reset only this key". Please add a test that sends "disk" and "cpu", resets "disk", and asserts that should_send("cpu", now=10) is False. That test fails on the current code.

    Nit (not blocking)

    • tests/test_alerts.py:18 still hard-codes now=301. The code now reads the window from settings, but this test still assumes it's 300. If someone changes ALERT_THROTTLE_WINDOW, the test breaks. settings.ALERT_THROTTLE_WINDOW + 1 would keep it in step. This line was already on main, so it's your call.

    No issues

    • Moving the 300 s window into settings.py follows rule 1, and the setting is defined there.
    • An in-memory dict is a reasonable choice given one process on one VM, so I didn't suggest Redis. Wiring reset() to the ack button is out of scope as your notes say, so I didn't review it.

    Verdict: request changes. The fix is a one-line code change plus one test.

    uses , so a reset that clears everything still passes. CLAUDE.md

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

    About this skill

    The problem

    An AI agent reviews a pull request. It does not know why the author made each choice. So it asks questions that the author already answered, and it flags things that are not problems. The author loses time on replies.

    What this skill does

    Before the agent reads the code changes, it reads:

    • The author's summary for the branch. It says what changed, what the author already tested, which options they did not use and why, and where to look closely.
    • The project rules, for example CLAUDE.md or CONTRIBUTING.md.

    If there is no summary, the agent says so in the review and asks the author for one. It does not guess.

    Every comment needs proof

    Before the agent writes a comment, it must:

    1. Give the file and line.
    2. Show what goes wrong, for example an input and the wrong result.
    3. Check that the code or a test does not already handle it.
    4. Check that the author's summary does not already answer it.

    The agent runs the project's tests when it can. It marks each comment as Blocking, Worth reconsidering, or Nit.

    When the author disagrees

    If the author shows proof that a comment is wrong, the agent checks the proof and takes the comment back. It does not take a comment back only because the author disagrees.

    What you need

    • A local copy of the repository, with the branch checked out.
    • The GitHub command-line tool (gh) is optional. The agent uses it to read the pull request description.
    • No scripts and no packages. This skill is one instruction file.

    Use it with PR Review Defense

    This skill helps the reviewer. PR Review Defense is a separate skill for the author.

    1. The author uses PR Review Defense to write a summary of the branch.
    2. You use this skill. Your agent reads that summary first.
    3. Each of your comments gives a file and line, and shows what goes wrong.
    4. The author's agent answers each comment with proof.

    Each skill works alone. They work best together.

    This skill is one instruction file with no script, so it cannot force the agent to follow every rule. Without the author's summary, the review has less context. If you have other review skills, a different one can load. Then name this skill in your request.

    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

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 29 days ago

    Requires

    GitHub
    git
    gh (GitHub CLI), optional

    Permissions

    Terminal / Shell
    Read Files
    Network Access

    Allowed Hosts

    api.github.com

    Terminal: the agent runs git, the project's tests, and gh. Read files: the agent reads the code, the project rules, and the author's summary. The skill does not write or change files. Network: gh reads the pull request description from GitHub. The skill does not post anything.

    Listed29 days ago

    What's inside

    Frequently Asked Questions

    It helps your AI agent review a pull request. The agent reads the author's summary and the project rules first. Then it gives only comments that point to a file and line and show what goes wrong.

    No. This skill works alone. The agent still reads the project rules and needs proof for each comment. But the author's summary comes from PR Review Defense. With both skills, the review starts with the facts.

    The agent says so in the review and asks the author for one. It does not guess why the author made a choice. It still reads the project rules and needs proof for each comment.

    A local copy of the repository, with the branch checked out. The GitHub command-line tool (gh) is optional. You do not need Python or other packages.

    Any agent that can read SKILL.md files and run commands. I tested it in Claude Code and Codex CLI.

    No. The agent shows the review to you. You decide what to post.

    New in Code Quality & Review