AgentSkillsCN

execplan-reviewer

严格依据 ExecPlan 阅读与评审评分标准及 JSON 输出规则,对 ExecPlan.md 进行审核;仅在无头审核运行期间被调用时使用。

SKILL.md
--- frontmatter
name: execplan-reviewer
description: Strict reviewer for ExecPlan.md using the ExecPlan review rubric and JSON output rules; use only when invoked during headless review runs.

ExecPlan Reviewer

Invocation keyword: $execplan-reviewer

Use this skill only during review runs to apply the rubric and emit strict JSON output matching the schema. If this skill is unavailable, the embedded rubric in the prompt is the fallback.

Output rules

  • Output JSON only, matching the provided schema exactly.
  • No extra keys, no Markdown, no code fences.
  • Evidence snippets must be short and pinpointed.

Deterministic verdict rules

  • If any issue is severity blocker or high, verdict must be FAIL.
  • If issues exist but all are medium or low, verdict may be PASS_WITH_FIXES.
  • If no issues exist, verdict must be PASS.

Review rubric

Gate 0: File-format compliance (blocker) FAIL if any of these are true:

  • Missing any mandatory living-document sections.
  • Progress is missing, not a checkbox list, or has no timestamped entries.
  • ExecPlan contains any triple-backtick fences (```).
  • Acceptance criteria are not observable/testable.
  • Plan relies on external documentation for key steps instead of embedding needed info.
  • Plan omits reading/following PLANS.md (must be stated early in the plan).

Gate 1: Purpose / outcomes / done (blocker/high)

  • Clear user-visible outcomes and how to see them working.
  • Non-goals and assumptions/constraints stated.
  • Acceptance is testable (commands + expected outputs or fail→pass tests). FAIL (blocker) if acceptance is not testable.

Gate 2: Plan feasibility and concreteness (high/medium)

  • Concrete repo-relative file paths, exact commands with working directory.
  • Expected transcripts where useful, concise and proof-focused.
  • Milestones (if used) are narrative and independently verifiable increments.
  • Idempotence/retry guidance for steps that can fail midway.

Gate 3: Safety / operational readiness (high/medium)

  • Safe execution steps and avoidance of destructive ops without backups/rollback.
  • Security/privacy/secrets/logging considerations when applicable.
  • Validation is comprehensive and not optional (tests + run/exercise system).

Gate 4: Plain language and readability for a novice (medium/low)

  • Terms-of-art defined or avoided.
  • Prose-first style (lists/tables minimized outside Progress).
  • Headings followed by two blank lines; correct Markdown.

Issue format requirements

  • severity: blocker | high | medium | low
  • category: formatting | self_containment | acceptance | scope | plan_quality | validation | safety | idempotence | novice_clarity | dependencies | decision_log
  • evidence: short pinpoint reference
  • fix: 1–3 concrete bullet steps
  • acceptance: how to verify the fix (testable/observable)