AGENT SKILL

rr-review

One adversarial review of the task's change, at the tier its mode sets — BALANCED in guarded, STRONG in full, security-focused when the domain axis is 2, and in light only when close reports a red fact. A fresh ai-reviewer reads the diff with bounded inputs, returns FINDINGS and a VERDICT, and state.py records them; in-scope BLOCKER and HIGH findings are fixed as one batch, closed by a green verify or one scoped re-review. The pass past the budget is the human's. Use on "/rr-review", "review this", after /rr-build in guarded or full mode, or when close says a review is required.

License 许可
MIT
最近更新
2026-11-14
Stars
★ 3
Forks
0

01 前言 Front Matter

name
rr-review
description
One adversarial review of the task's change, at the tier its mode sets — BALANCED in guarded, STRONG in full, security-focused when the domain axis is 2, and in light only when close reports a red fact. A fresh ai-reviewer reads the diff with bounded inputs, returns FINDINGS and a VERDICT, and state.py records them; in-scope BLOCKER and HIGH findings are fixed as one batch, closed by a green verify or one scoped re-review. The pass past the budget is the human's. Use on "/rr-review", "review this", after /rr-build in guarded or full mode, or when close says a review is required.
argument-hint
[ "B1", "H1 … — a scoped re-review of these findings" ]

03 SKILL.md 正文

/rr-review $ARGUMENTS

One reviewer, one pass, on a context that did not write the code. Resolve the install root once:

for AI_HOME in "${CLAUDE_CONFIG_DIR:-$HOME/.claude}" "${CODEX_HOME:-$HOME/.codex}"; do
  [ -d "$AI_HOME/skills/rr-review" ] && break
done
STATE="python3 $AI_HOME/skills/ai-task/state.py"

1. Which review

$STATE status gives the mode, the axes and the review so far.

Mode Review Tier
light only when close exits 2 with review required (a red fact) BALANCED
guarded one BALANCED
full, full_irreversible one; briefed with the security focus when the domain axis is 2 STRONG

review refuses a tier below the mode's. Two passes per task are the budget (review_rounds_max): one full review and one scoped re-review. The pass past it exits 6 REVIEW_BUDGET — the next verdict is the human's, from their own terminal, with the command it prints, or the pick in the dialog (DIALOG_PENDING, then the picked command verbatim, once). Never a third pass. A batch task (/rr-build-all) has its own budget, batch.review_rounds_max full passes (shipped three): after a blockers_open fix batch the next pass is a full one, and the pass past the budget is the human's the same way.

2. Brief one ai-reviewer

Spawn exactly one ai-reviewer at the tier above — under Claude Code with model: $($STATE profile --tier <TIER>); under Codex its agent file's model is the tier, so name the tier in the brief. Give it paths, not pasted text:

  • the goal, the mode and the workflow ($STATE status);
  • the change: git diff against the task's base plus new files (git status --short), and the task's files from .ai/reports/<task-id>/task.md;
  • the tests the task named, and the last verify result and log path;
  • in full mode with domain 2: the security focus — authentication, authorisation, secrets, personal and payment data, input validation, injection — and record the pass with --focus security.

Bounded inputs in light and guarded: the goal, the diff, the named tests, and the project's own threat model (.ai/policies/safety.md, the guards' defaults, the test fixtures). A finding that needs anything outside them comes back [INFO] with in_scope: no; record it with $STATE memory risk add as one entry — it never opens a fix. You may move a finding to in_scope: yes, never the other way. In full mode the reviewer's own "look for" list applies unbounded. Tell it the cap: about 40 tool calls in light and guarded, about 80 in full, then FINDINGS and VERDICT.

Before you delegate, probe the change yourself for a minute: the few inputs its own threat model makes interesting, through the real code, and hand the reviewer what you saw — it starts where you stopped. When the change touches a guard regex's running time, run the timed probe yourself (/usr/bin/grep -E, or Python timeit on inputs of at most 4096 chars) and hand over the measurement: the bounded reviewer may not run timing loops.

3. Record the verdict

The reviewer returns ## FINDINGS (- [BLOCKER] B1 …, [HIGH] H1, [MEDIUM], [LOW], [INFO], each with in_scope: yes|no) and ## VERDICT (verdict: pass | pass_with_findings | blockers_open). Pass both through verbatim:

$STATE review --verdict <verdict> --tier <BALANCED|STRONG> [--focus security] --report - <<'EOF'
<the reviewer's FINDINGS and VERDICT, unchanged>
EOF

review checks the verdict against the findings: an in-scope BLOCKER is blockers_open, an in-scope HIGH is not a pass. It keeps the text in the task's record (task.md), so there is no separate report to write.

4. After the verdict

Continue from the verdict §3 recorded; do not record it again. Whatever the verdict, each MEDIUM and LOW is one $STATE memory risk add (.ai/project/known-risks.md amended into the task if missing), not a fix.

  • pass: /rr-ship, i.e. $STATE close directly: the verdict is already recorded. (To record and close in one call, §3's review --verdict pass … <<'EOF' && $STATE close; review exits 0 for every verdict, so that chain is for pass only.)
  • pass_with_findings: fix every in-scope HIGH in one batch, then $STATE verify --scope all --after-fix "H1,H2: <one line>" and $STATE review --closes "H1,H2". The findings close on that green run, on a tree the fix changed — no second review. MEDIUM and LOW: memory risk add as above; only a BLOCKER, a HIGH or a regression buys an --after-fix run.
  • blockers_open: fix every BLOCKER and HIGH together, verify --scope all --after-fix "…", then one scoped re-review (/rr-review B1,H1): brief a fresh ai-reviewer with the named findings and the fix's diff only — do they close, and what did the fix introduce — and record it as above. Never one finding, one fix, one re-review at a time.

The verify after a fix is often the task's third whole-suite run: past tests.max_suite_runs it exits 6 SUITE_BUDGET, and that run is the human's, from their own terminal, with the command it prints, or the pick in the dialog.

A finding whose cause is the plan goes back to /rr-plan for that part only. A real problem outside the task is recorded with $STATE memory risk add; a fix in this task closes its risk with $STATE memory risk close <rkey> --status fixed --why "…". A mistake this repository has seen before — the second time a review flags it — gets one line in the project instruction file, in this task.

04 同仓库更多 Skills

vanssata
vanssata/claude-agentic-sdlc-module

sylius-fixtures

Load the Sylius fixture suite into the local database

★ 3
vanssata
vanssata/claude-agentic-sdlc-module

deploy-staging

Deploy the current branch to staging and smoke-test the checkout. Use on "deploy to staging".

★ 3
vanssata
vanssata/claude-agentic-sdlc-module

rr-status

暂无简介

★ 3
vanssata
vanssata/claude-agentic-sdlc-module

rr-spec

Turn a rough idea into one file, docs/sdlc/specs/<slug>.md — an `## Intent` section (problem, outcome, affected users and systems, constraints, decisions taken) from a short one-question-at-a-time brainstorm recorded through state.py, then requirements and design from the architect agent on the STRONG tier, checked against the project's constitution, instruction files, skills and ADRs, ending with explicitly flagged concerns (especially contradicting policies). Accepts an existing docs/sdlc/intent/<slug>.md instead of the brainstorm. Use when starting a feature, "write the spec", "let's define what we want", "design this", before any plan or code.

★ 3
vanssata
vanssata/claude-agentic-sdlc-module

rr-ship

Close the task — state.py close measures the two facts (a green whole-suite verify on the tree as it is now; every changed file inside the task's files), applies the mode table and either archives the task or, in full mode, requests the human's approval and prints the exact approve command for their own terminal. Writes the commit message body and suggests the commit; never merges or deploys, and commits, pushes or approves only as the command the human picked in the dialog, verbatim, once. Use on "/rr-ship", "close the task", "ship it", "ready for approval", after /rr-build or /rr-review.

★ 3
vanssata
vanssata/claude-agentic-sdlc-module

rr-setup

暂无简介

★ 3