메뉴
Produce a read-only code review report from a diff. Runs parallel reviewers, folds their findings through adjudication and humanization, and saves the report under docs/superpowers/reviews/. Shadows the built-in /review command.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Clean up code in current branch - simplify, remove AI artifacts, challenge decisions
Save session context to ferrex before context clear
Delete ferrex memories by ID
Search ferrex memory
Audit ferrex memory health
Store in ferrex memory (no args = session summary, with args = store that info)
| name | review |
| description | Produce a read-only code review report from a diff. Runs parallel reviewers, folds their findings through adjudication and humanization, and saves the report under docs/superpowers/reviews/. Shadows the built-in /review command. |
You are the review controller. Produce a single humanized review report and print it inline. Every phase is read-only — you never commit, never touch branches, never edit files outside the staging directory and the final report path.
Announce at start: "Using the review skill." Then run the phases below in order.
The invocation grammar, evaluated left-to-right, first match wins:
| Invocation | Meaning | Asks base? |
|---|---|---|
/review | current branch vs main | no |
/review onto <target> | current branch vs <target> | no |
/review <ref> onto <target> | <ref> vs <target> | no |
/review <int> | GitHub PR number for the current repo | no |
/review <sha> | commit vs current-branch-or-main | yes |
/review <branch> | branch vs current-branch-or-main | yes |
Ambiguity resolution:
^[0-9]+$) → PR mode.git rev-parse --verify <arg> and git rev-parse --verify <arg>^{commit}:
<arg> matches ^[0-9a-f]{7,40}$ and resolves → treat as SHA.<arg> as a SHA, branch, or PR number." and stop.main or HEAD is detached and /review was given no args (or onto <target> with no ref), respond: "No branch to review. Pass a commit, branch, or PR number." and stop.Base ambiguity. In SHA and bare-branch modes (no onto clause), ask a single AskUserQuestion:
current branch (<name>@<short-sha>) — only offered when the current branch is not main and is not detached.mainOther — free-form ref. If the entered ref doesn't resolve, re-ask once, then stop.PR mode. Derive <owner>/<repo> from git remote get-url origin (require an origin remote pointing at github.com — otherwise respond: "PR mode requires an origin remote on github.com." and stop). Call pull_request_read from the GitHub MCP with that owner/repo. If the call errors (nonexistent PR, permission denied, network), print the error and stop — do not attempt a worktree.
On success, record the PR head SHA, base branch, title, body, URL, and linked issues. Fetch the head locally so reviewers can browse files:
git fetch origin pull/<n>/head:refs/pr/<n>
Set head = refs/pr/<n>, base = <pr-base-branch>, name = pr-<n>.
Fill in ref names for other modes:
base = main, head = <current-branch>, name = main-vs-<current-branch>onto <target> alone: base = <target>, head = <current-branch>, name = <target>-vs-<current-branch><ref> onto <target>: base = <target>, head = <ref>. If <ref> is a SHA use its 8-char short form for name; else use the branch name. name = <target>-vs-<ref-or-short-sha><sha>: base from the ambiguity prompt, head = <sha>, name = <base>-vs-<short-sha> (8-char short)<branch>: base from the ambiguity prompt, head = <branch>, name = <base>-vs-<branch>Compute short_run_id as a random 8-hex string (e.g., first 8 chars of a date +%s%N | sha256sum hash — any 8 hex chars are fine as long as they don't collide with an existing worktree).
Set repo_root from git rev-parse --show-toplevel. Set worktree_path = <repo_root>/.git-worktree-review-<short_run_id> and staging_dir = <repo_root>/docs/superpowers/reviews/_<name>.
Run: git worktree add --detach <worktree_path> <head> from the repo root. <head> is whatever Phase 0 resolved — a branch tip, a SHA, or a refs/pr/<n> ref for PR mode. --detach is required so a branch checked out at the repo root doesn't collide.
If the command fails, print the error and stop.
Track two flags for cleanup:
created_worktree = truepr_ref_created — true only if Phase 0 fetched a refs/pr/<n> ref (PR mode)Create <staging_dir> (absolute path). Any existing contents are removed first (rm -rf <staging_dir>). This is the drop zone for subagent reports and lives in the repo root, not the worktree — every subagent prompt uses the absolute path so writes don't land in the worktree's own gitignored copy.
From the worktree, compute a diff summary. The commands below assume you have already cd'd into <worktree_path> — do that first.
File-level stats:
git diff --name-status <base>...<head> → files changed / added / removed / renamedgit diff --numstat <base>...<head> → lines added / removed per file (raw)Effective LOC: raw lines minus every line in files matching any of these globs (matched with git check-attr or, more simply, plain shell glob against each file path):
*.lockCargo.lock, flake.lock, pnpm-lock.yaml, package-lock.json*.min.*vendor/****/*.snapPath buckets. Assign each changed file to exactly one bucket, first match wins:
**/auth/**, **/crypto/**, **/security/**, **/migrations/**, **/*.sql, **/policy/**, **/rbac/**, **/keychain*, **/secret*# generated / // generated marker in the first 5 lines, or paths containing generated/, .pb.go, .pb.rs, .pb.tstest/, tests/, __tests__/, or names ending in _test.{ext} / .test.{ext} / .spec.{ext}**/*.toml, **/*.yaml, **/*.yml, **/*.ini, **/*.json (unless already generated), **/*.nix**/*.md, **/*.rst, **/*.txt, docs/**Log the bucket assignment as a short table so the user can see the reasoning.
Complexity tier:
mechanical — no logic edits: pure renames (per --name-status R), formatting-only diffs (no non-whitespace changes in a hunk), comment-only edits, config bumpsminor-logical — logic edits confined to non-crucial pathscrucial-logical — any hit in the crucial bucket, regardless of sizeReviewer count formula:
| Tier | ≤300 | 301–800 | >800 |
|---|---|---|---|
| mechanical | 1 | 1 | 2 |
| minor-logical | 1 | 2 | 2 |
| crucial-logical | 2 | 2 | 3 |
For effective LOC > 800: reviewers = max(table_value, ceil(effective_LOC / 800)).
Per-tier caps: mechanical 2, minor-logical 4, crucial-logical 5.
Codex reviewer is proposed when tier is crucial-logical OR effective LOC > 600 OR ≥4 crucial-bucket files.
Gap-finder pass is proposed under the same triggers.
Record the results as plan_proposal:
tier, raw_loc, effective_loc, bucket_countsreviewers (int)codex (bool)gap_finder (bool)Use the AskUserQuestion tool. The question body carries the plan proposal (tier, effective LOC, reviewers, codex, gap-finder, bucket breakdown). Options:
crucial-logical.Store the final choice as plan (same fields as plan_proposal).
Unsatisfiable constraint. If tier is crucial-logical and the resolved plan.reviewers is 1 (only reachable via Other), the ≥2-reviewers-per-crucial-slice rule in Phase 2.7 can't be met. Re-ask once with a note: "This diff touches crucial paths and needs at least 2 reviewers. Pick a value ≥2 or type override to proceed with 1 anyway." Accept override verbatim — the user knows.
If effective_loc ≤ 800 × plan.reviewers, every reviewer gets the whole diff. Skip the rest of this phase — set slices = [{ reviewer: i, paths: <all> } for i in 1..plan.reviewers].
Otherwise partition:
src/auth/, modules/services/, docs/).Now assign slices to reviewers under these constraints:
crucial-bucket file gets ≥2 reviewers.Each reviewer also receives the full diff as reference context. Instruct them: "Deep-review your assigned slices only. Skim the rest and flag cross-slice smells in a separate section."
For each changed file, look up its extension in this table and collect the union of rule names:
| Extension | Rule name |
|---|---|
.rs | languages/rust |
.py | languages/python |
.ts, .tsx, .js | languages/typescript |
.go | languages/go |
.java | languages/java |
.kt | languages/kotlin |
.swift | languages/swift |
.css | languages/css |
.html | languages/html |
Extensions with no mapping (e.g. .nix) are skipped without error. Markdown and other prose files are excluded — only the prose reviewer touches them, via the humanizer rubric.
Add base/core and base/git unconditionally. The final list is code_rules.
Every code, idiomatic, and cleanup reviewer's prompt includes a line: Load these rule files from ~/.config/claude/rules/: <comma-separated code_rules>.
The prose reviewer does not get language rules — it uses the humanizer skill's rubric instead.
In a single message, issue N parallel Agent tool calls — one per reviewer below. Every prompt starts with:
cd <worktree_path> before any git command.<base> and <head> refs.git diff <base>...<head>.Critical / Important / Minor, each with file:line + rationale, and a separate "Cross-slice smells" section.Every output path below is joined against <staging_dir> (absolute), not the worktree.
Standing agents (always dispatched):
| Agent | Subagent type | Model | Output path |
|---|---|---|---|
Code reviewer × plan.reviewers (indices 1..N) | superpowers:code-reviewer | opus | <staging_dir>/code-review-<i>.md |
| Idiomatic-pass reviewer | general-purpose | opus | <staging_dir>/idiomatic.md |
| Cleanup reviewer | general-purpose | opus | <staging_dir>/cleanup.md |
| Prose/comment reviewer | general-purpose | opus | <staging_dir>/prose.md |
Conditional (per the plan):
| Agent | Subagent type | Model | Output path |
|---|---|---|---|
Codex reviewer (when plan.codex is true) | general-purpose | inherit | <staging_dir>/codex.md |
Role rubrics:
superpowers:code-reviewer conventions.cleanup skill.humanizer skill first to load its rubric. Then for every added or changed comment, docstring, and error message in the diff, flag any pattern hit — restating code, AI-obvious phrasing, sycophantic docstrings, promotional language.skill-codex skill to hand the diff to the codex CLI. Ask codex for a grounding pass — verify claims in code (log messages, error strings, boundary asserts) match reality, and surface edge cases the primary reviewers missed. Codex writes back into the staging file at the specified path.If a subagent errors or times out, keep the others' outputs. Write a stub inside <staging_dir> renamed with a .failed.md suffix (e.g. <staging_dir>/codex.failed.md) whose body is one line naming the failure (e.g. "codex CLI timed out after 120s"). Record parallel_failure = true. The pipeline continues — Phase R2's adjudicator sees the gap.
Every subagent output below uses an absolute path under <staging_dir>.
Skip this phase when plan.gap_finder is false.
Use the Agent tool with subagent_type: general-purpose, model: opus. Prompt contents:
cd instruction.<base> and <head> refs.<staging_dir>/ as reference material (list the absolute paths; the subagent reads them).Critical / Important / Minor taxonomy.<staging_dir>/gap-finder.md.Use the Agent tool with subagent_type: general-purpose, model: opus. Prompt contents:
cd instruction.<staging_dir>/ as reference material (including gap-finder.md if it exists, and any .failed.md stubs).accept / modify / remove:
accept — keep verbatim.modify — rewrite when the description is unclear or overreaching, keeping the underlying issue.remove — drop false positives. Each rejection is appended to an "Adjudicator rejections" list with a one-line reason.<staging_dir>/adjudicated.md. Structure: findings grouped by severity, then a "Cross-slice smells" section, then an "Adjudicator rejections" appendix.Use the Agent tool with subagent_type: general-purpose, model: opus. The controller pre-fetched PR title/body/URL/linked-issues in Phase 0 (PR mode only) and passes them into the prompt. The reconciler itself fetches CI status and looks up prior reviews. Prompt contents:
cd instruction.<staging_dir>/adjudicated.md, the diff.<owner>/<repo> from git remote get-url origin (both modes) so the reconciler can build a compare URL and, in PR mode, call the GitHub MCP checks endpoint.<staging_dir>/draft.md:
https://github.com/<owner>/<repo>/compare/<base>...<head>; linked issues from the PR body (PR mode); CI status via the GitHub MCP checks endpoint (PR mode only — skip in branch / SHA modes); related past reviews found by listing <repo_root>/docs/superpowers/reviews/*.md on the working filesystem (these are gitignored, so do not consult any git ref) and matching filenames against the current <name> prefix or, in PR mode, pr-<n>.md.file:line + rationale.adjudicated.md.Use the Agent tool with subagent_type: general-purpose, model: sonnet. Prompt contents:
cd instruction.<staging_dir>/draft.md, <staging_dir>/adjudicated.md, the diff.adjudicated.md but missing from draft.md,file:line references — for each, grep -n <text-fragment> <file> in the worktree to confirm the line exists and matches,<staging_dir>/verified.md.verified.md).Use the Agent tool with subagent_type: general-purpose, model: inherit. Prompt contents:
cd instruction (though this phase does not need git).humanizer skill first to load its rubric, then apply it to every prose section of <staging_dir>/verified.md.file:line refs, and diff quotes untouched.<staging_dir>/final.md.final.md).Copy <staging_dir>/final.md to <repo_root>/docs/superpowers/reviews/<name>.md. If the target already exists (re-run for the same name), overwrite it — no timestamping. The prior report was already picked up as a related past review by the reconciler in R3.
Print the contents of <repo_root>/docs/superpowers/reviews/<name>.md inline to the user.
If final.md is missing (any earlier phase failed catastrophically), tell the user which phase failed and skip the copy.
Silent by default:
git worktree remove <worktree_path> from the repo root.pr_ref_created is true, run git update-ref -d refs/pr/<n> to drop the fetched PR ref.Ask before cleanup only when parallel_failure is true (Phase 3.5 wrote a .failed.md stub). Sequential-phase misses are already surfaced by R6 and don't need a second escalation. Show the user the missing files and let them inspect the worktree before agreeing to remove it.