菜单
Canonical code/infra closeout workflow — freezes scope, resolves the exact review target (local/commit/branch/PR), runs a non-mutating review, resolves a cross-family reviewer, and requires separate behavior proof for user-visible code or infrastructure changes. Never use as a learner-content semantic gate.
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
Run or resume the manifest-derived curriculum lifecycle for one active track. Use when an operator asks to process a track in manifest order, select built, unbuilt, stale, or one-module scope, or resume an exact coordinator run. Route every acquired module through curriculum-preparation or track-completion; do not use for an uncoordinated standalone module.
Prepare, audit, or resume missing or stale prerequisite evidence for one manifest target, one track missing-only scan, or one bounded homogeneous packet by using the canonical curriculum readiness evaluator. Use for preparation-scoped work before a learner-module build or when curriculum-lifecycle explicitly delegates preparation. Do not use to build learner modules, coordinate a track, run post-build review, certify, publish, or deploy.
Complete one active Ukrainian curriculum module through a canonical CORE-or-seminar workflow. Use for built modules that need the versioned post-build gate, unbuilt modules that need plan review and V7 construction before that gate, interrupted module completion runs that must resume from a durable ledger, or non-PASS modules that need root-cause repair, fresh review, independent cross-family review, and publication.
Model-agnostic orchestrator playbook for driving ONE epic/track lane end-to-end over the fleet-comms message plane. Invoke this when you are launched as an epic/track driver (Grok, Sonnet-5, Gemini/AGY, Kimi, or Claude) via a start-<model>.sh --epic launch and told to "drive this epic". Teaches the METHOD — topology → route → dispatch → settle → cross-family review → merge → handoff — and defers all live roster/routing DATA to /api/rules and model_catalog.yaml. Not for main-orchestrator cold-start (that has its own hook) and not for writing a single module (use $track-completion).
How to monitor V7 curriculum builds — the Monitor tool command template, mandatory --worktree isolation, the JSONL event fields v7_build.py/linear_pipeline.py emit, and the Monitor API for state queries without running a build. Use whenever launching or watching a `scripts/build/v7_build.py` run.
Audit prompt sizes across build orchestration dirs. Surfaces sections that dominate the prompt (>40%), tracks per-phase budget vs the model's context window, and checks whether directives in the prompt were actually covered in the response.
| name | local-code-review |
| description | Canonical code/infra closeout workflow — freezes scope, resolves the exact review target (local/commit/branch/PR), runs a non-mutating review, resolves a cross-family reviewer, and requires separate behavior proof for user-visible code or infrastructure changes. Never use as a learner-content semantic gate. |
| argument-hint | [local | commit <sha> | branch <branch> <base> | pr <number>] |
This skill supports only code and infra review profiles. A normal BIO
learner-content request routes to $curriculum-lifecycle (or directly to
$track-completion for one module), which owns the single bounded
$post-build-review gate. resolve-reviewer fails before route selection for
any learner-content semantic profile or domain. Do not duplicate post-build
semantic review here.
Scope: this is the closeout gate for a change you (or another agent) already made — not a PR-comment bot. For posting inline PR comments, use the official
/code-review:code-reviewplugin instead; this skill can still drive the target-selection and scope-freeze steps underneath it.
This skill defines orchestration and policy for closeout review. It does not implement the strict finding verifier or reviewer process isolation — those are separate, later slices (issue #5281 children 2 and 3). What it does guarantee:
$ARGUMENTS selects the review target mode, explicitly — never inferred
from git status:
local — staged + unstaged + untracked changes in the working tree.commit <sha> — one already-made commit, diffed against its parent.branch <branch> <base> — a branch diffed against an explicit base ref.pr <number> — a PR diffed against its actual base (not an assumed
default branch).If the user gives you a PR number, a commit SHA, or says "review my branch,"
that is the mode — do not fall back to local because the working tree
looks clean. A clean working tree is not evidence that a commit or PR was
reviewed (see Step 1 below).
Read and follow local-code-review-checklist.md
in full. It drives .venv/bin/python -m scripts.review.closeout_cli for every
step that has a deterministic answer (target resolution, scope-baseline
freeze and breakers, reviewer resolution, findings adjudication) and tells
you exactly what to reason about yourself.
Print a structured report to the conversation (the checklist's Step 8 gives the exact shape). Do NOT write a file unless specifically asked.