AGENT SKILL

grok-review

Use when the user requests a Grok review, or when another skill such as review-loop runs one.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 0
Forks
0

01 前言 Front Matter

name
grok-review
description
Use when the user requests a Grok review, or when another skill such as review-loop runs one.

02 资源文件 1

03 SKILL.md 正文

Grok review

Run Grok Build's built-in /review locally with an installed, authenticated grok CLI. Use the user's selected model and effort; otherwise model {{defaults.model}} and effort {{defaults.effort}}. Do not post findings or edit the reviewed files.

Prepare

Identify the repository and exact review target from the request. Record the current commit and git status --short. Create a unique directory under .plans/scratch/ for prompt.md, the report, and the log; exclude it from review.

Write prompt.md as one /review line for the targets it covers:

  • Uncommitted changes: /review --local.
  • The current branch with a clean tree: /review --main. It compares against origin/main or origin/master.
  • Another branch: /review --branch <name>.

Never pass --pr, --stack, --submit, or --approve; they post to GitHub. /review --local reviews every uncommitted change. If the tree holds user changes outside the requested target, say so when presenting the findings.

For a commit, a plan, or selected files, write a self-contained prompt instead, with the exact target and the command or paths that show it. Ask for concrete bugs, regressions, and requirements mismatches with severity, file and line, failure scenario, and fix direction; for plans, feasibility and missing decisions. Request an explicit statement if no substantive issues are found, and instruct it to review without edits or tests.

Review a branch against a base

When the base is origin/main or origin/master, prompt.md is /review --branch <branch>. It ignores uncommitted files in the tree. For any other base, prompt.md holds only the target and git diff <base>...<branch>. That prompt, like every custom prompt, runs through the CLI command under "Run and monitor", never as a T3 delegate.

Run as a T3 delegate

In T3 Code, when prompt.md is a /review line, run it as a delegate instead of the command below. Call orchestrator_capabilities and check that provider grok offers the selected model and effort. Then call delegate_task with mode: "async", target: {providerInstanceId: "grok", model: <model>, options: {reasoningEffort: <effort>}}, and the /review line as the whole task. Set no role: role: "review" puts a sentence in front of the task, and Grok then no longer reads it as a slash command.

Once every other task you need is started, end your turn and wait for the completion notice. The result is the review's closing summary, which names its findings file under /tmp/grok-<user>/. The delegate has no read-only sandbox, so compare git status --short and git rev-parse HEAD with the ones you recorded. Report any change it made, and never treat its files or commits as part of the reviewed work. If the delegate is unavailable or fails, run the command below.

Run and monitor

Run grok --version first. If it is 1.0.66 or older, stop and ask the user to update Grok; do not update it yourself. Those versions can keep running after the review has finished.

Start the command below using the calling tool's background or resumable execution mode, and retain its task or session ID so you can check its status and exit code. Do not set a review runtime limit or a tool timeout that kills the process.

Set REVIEW_REPO, REVIEW_MODEL, REVIEW_EFFORT, and REVIEW_DIR to the absolute repository path, selected model, selected effort, and artifact directory:

grok --cwd "$REVIEW_REPO" --model "$REVIEW_MODEL" --effort "$REVIEW_EFFORT" \
  --sandbox read-only --always-approve --disable-web-search \
  --output-format plain -p "$(cat "$REVIEW_DIR/prompt.md")" \
  </dev/null >"$REVIEW_DIR/report.md" 2>"$REVIEW_DIR/run.log"

/review gives its reviewer a shell and a write tool, so --always-approve is safe only inside the read-only sandbox: the kernel blocks writes outside ~/.grok and temp directories, and blocks network access for its commands. /review writes its full findings to a file under the temp directory and names that file in the report. This is a static review: the caller remains responsible for tests.

Wait until the review finishes, then report. If the calling tool notifies you when a background command exits, end your turn and wait for that notice. Never start a polling loop or a monitor to watch the review. Without such a notice, check its status with short waits that return control and leave it running.

A running process alone does not prove progress, and silence or elapsed time alone does not prove a hang and must not trigger cancellation. If the review appears stuck, inspect its process state and logs. Report what you can establish and keep waiting when the evidence is inconclusive.

When the task exits, check its exit code and report. Report failures promptly and identify any partial output; never present an incomplete review as clean. Do not silently switch models or effort, or drop the sandbox to make the invocation work.

Assess the result

Read the report and the findings file it names, and verify substantive findings against the source before presenting them. Distinguish confirmed issues from unresolved concerns, name the reviewed target, model, and effort, and state verification gaps. A clean review is not evidence that tests passed. Check the final Git state; do not revert user work.

04 同仓库更多 Skills