菜单
Review tasks.json for correct, complete, executable translation of its reviewed plan before execute.md is finalized. Use for comprehensive task-graph review or explicit re-review. Do not review plan quality, finished code, index formatting, or change scope.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/joenandez/spectre --skill spectre-task-review命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | spectre-task_review |
| description | Review tasks.json for correct, complete, executable translation of its reviewed plan before execute.md is finalized. Use for comprehensive task-graph review or explicit re-review. Do not review plan quality, finished code, index formatting, or change scope. |
| user-invocable | true |
Run one independent semantic review of the complete task graph, apply authorized task corrections, and prove the reviewed graph can be implemented correctly without avoidable rework. State WHAT must hold; trust the reviewer to determine findings and edits.
$ARGUMENTS: explicit feature root/name or descendant task artifact; --mode adversarial|full; optional --auto-apply scope-safe; --review-again only when the user's latest instruction explicitly requests another completed review.specs/plan.md and parseable specs/tasks.json (or explicit scoped .tasks.json). execute.md is not an input.FEATURE_ROOT for this work from explicit/current-thread evidence only (physical directory wins; never branch/recency/lifecycle/scans). If none is confirmed, including when the candidate path is occupied, standalone MUST first load and follow @skill-spectre:spectre-feature-root through DONE; orchestrated calls escalate. Keep writes beneath it and pass it unchanged.TASKS_JSON: selected task graph. : , or a timestamped sibling only for authorized .REVIEW_REPORTreviews/task_review.md--review-againREVIEW_ATTEMPT: reviews/task_review_attempt.json, the durable one-round recovery ledger with round_status: in_progress|report_ready|complete|incomplete, report/authorization/route state, allowed writes, and pre/post task hashes.PREFLIGHT_JSON: reviews/task_review_safety.json, produced by the adjacent scripts/task-review-safety.mjs helper.The reviewer owns TASKS_JSON and REVIEW_REPORT; no other canonical artifact may change. The report contains:
# | Severity | Lens | Location | Finding | Suggested Edit;unresolved|applied|skipped|scope-change, with resulting task edit when applied;DONE when the attempt is complete; findings preceded edits; every finding has evidence, a concrete consequence/edit, and a disposition; authorized edits touch only the task graph/report; task JSON parses; deterministic report/post-write checks pass; protected plan/scope hashes match; route and pre/post hashes are recorded; and Scope Change Required findings remain unapplied.
Establish safety. Run the helper's preflight before reviewer tokens. Hard failures stop; advisories inform the semantic reviewer. The helper validates consumer safety, not semantic quality, and never authorizes or restarts review work; do not use its retired impact operation.
Run one semantic review per authorized round. A completed report ends the round; only an explicit user --review-again authorizes another. Otherwise use the available opposite runtime at pinned medium effort (Codex → Claude opus; Claude → Codex gpt-5.6-sol). Allow up to 20 minutes; quiet output alone is not failure. Adversarial mode reviews the whole graph in one pass; Full mode may use its permitted lens workers within that same review. If the opposing route fails to produce a usable report, dispatch one clean-context @spectre:reviewer as the fallback under the same ledger/report. Resume incomplete or report_ready state instead of starting another review.
Judge the translation as a whole. The goal is not checklist completion: determine whether tasks.json correctly and completely translates the reviewed plan into an executable graph that can be implemented once, the right way, without avoidable rework. Use these lenses as guidance, not an exhaustive taxonomy or a limit on evidence-backed reviewer judgment:
Reject terminal verification/E2E tasks owned by Execute or Prove unless they produce an explicit prerequisite or product-consumed artifact at a product-owned path.
The reviewer may raise any evidence-backed translation risk that threatens correctness, completeness, integration, executability, or creates likely rework. Severity is Blocker|High|Medium|Low|Scope Change Required. Canonical scope and plan.md remain immutable; findings that require changing either are Scope Change Required.
Apply only authorized findings. Write all findings before edits. With --auto-apply scope-safe, apply scope-safe Blocker/High and unambiguous translation-only Medium/Low findings. Otherwise obtain the user's selection and continue the same reviewer route for writeback only. Preserve task IDs when possible. The primary may repair mechanical report/schema metadata from existing reviewer evidence, but may not invent findings, reinterpret them, or perform semantic task edits.
Close deterministically. Run helper validate-report, reparse and postflight the task graph, verify protected inputs, and record hashes/state. Preserve report_ready or incomplete recovery state when closure fails. The planning caller finalizes execute.md only after completion, then runs validate-pair.
Return runtime/fallback, findings and dispositions, report/task paths, parse/validation status, and any unapplied scope-change recommendation. Under --orchestrated, omit user-facing next steps. Standalone unresolved Blocker/High stays in remediation; resolved review recommends /spectre:execute.
/spectre:create_plan or /spectre:create_tasks.--review-again.