菜单
Use to review a git ref, branch, or PR with the local kagura-engineer reviewer — shells out to `kagura-engineer review [target]`, returning a structured verdict and findings. With `--fix` it is a HARNESS that edits files and commits; without `--fix` it is read-only.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when kagura-engineer:doctor reports a blocked or failing environment — scaffolds missing config, installs and authenticates harness prerequisites, and generates Memory Cloud MCP config by shelling out to `kagura-engineer setup`, then re-checks with doctor.
Use before any kagura-engineer run/review, or when the harness reports a blocked environment — diagnoses the local setup (git, selected brain CLI, gh, ollama, memory backend, and gh-issue-driven plugin) by shelling out to `kagura-engineer doctor` and reports what must be fixed.
Use to measure whether memory grounding actually improves PR quality — shells out to `kagura-engineer eval` with one or more issue numbers, running the SAME fixed issue set in two arms (recall ON vs OFF) and printing an A/B table. HARNESS — high cost; it runs the full loop twice per issue and mutates the repo. Confirm with the user before launching.
Use to drive a single GitHub issue to a pull request with the kagura-engineer harness — shells out to `kagura-engineer run` with an issue number (guard → recall → worktree → start → implement → ship → persist). HARNESS — this mutates the repo, creates a PR, and spends model budget; confirm with the user before launching.
Use to bootstrap a fresh checkout for the kagura-engineer harness — scaffolds a commented repo.yaml template and adds it to .gitignore by shelling out to `kagura-engineer init`. Run this before setup/run/review when no repo.yaml exists yet. Idempotent and never overwrites.
| name | review |
| description | Use to review a git ref, branch, or PR with the local kagura-engineer reviewer — shells out to `kagura-engineer review [target]`, returning a structured verdict and findings. With `--fix` it is a HARNESS that edits files and commits; without `--fix` it is read-only. |
Thin wrapper around the kagura-engineer review CLI verb. It runs the multi-angle
reviewer over a target (working tree, branch, or PR number) and returns a structured
verdict. It reimplements nothing — it discovers config, optionally gates on doctor,
shells out, and surfaces the report.
⚠️ Two modes — know which you are in:
- Default (read-only) — reviews the diff and reports. No repo mutation.
--fix(Harness) — on a red verdict it invokes the selected brain backend to edit blocking findings and commit them, then re-reviews (bounded byconfig.review.max_loops). This mutates the repo and spends model budget — confirm with the user first.
Announce: "Using the kagura-engineer:review skill" (add "— with --fix, this is a Harness that commits" when fixing).
Usage: kagura-engineer:review [target] — target defaults to HEAD; may be a git ref, branch, or PR number. Add --fix to auto-fix blocking findings (Harness).
Validate the target. The optional positional TARGET defaults to HEAD; it may
be a git ref, branch name, or PR number. Validate before shell interpolation against a
first-char-anchored allow-list — a positive integer ^[1-9][0-9]{0,8}$ for a PR
number, or ^[A-Za-z0-9][A-Za-z0-9._/-]{0,199}$ for a ref/branch (leading - rejected
so it cannot be parsed as an option) — and pass it after a -- separator (step 4). Do
not pass free-form text.
Discover config. repo.yaml in CWD by default; pass --config <path> otherwise.
Precondition — run doctor:
kagura-engineer doctor --json
review itself has no phase-0 guard, but running doctor first surfaces a missing
reviewer (ollama / model) before you wait on a review. If overall == "fail", hand
off to kagura-engineer:setup; otherwise proceed.
Review (read-only). Options go before the -- separator, the validated target
after it, so a target beginning with - can never be parsed as a flag:
kagura-engineer review --base <ref> --json -- <target>
Or, to auto-fix blocking findings (confirm the Harness warning first):
kagura-engineer review --fix --json -- <target>
Interpret the status. ok (exit 0) → green/yellow or nothing to review.
blocked (exit 2) → red verdict with blocking findings; show them.
fail (exit 1) → could not review (reviewer missing, timeout, fix failed).
Surface & hand off. Print the verdict and findings (dimension / severity /
file:line / title); the full report is at .kagura/review.json. On a blocked verdict,
offer --fix or surface the findings for manual resolution.