post-review

name	post-review
description	三方会审 — each reviewer identity is run independently by three heterogeneous models (Opus + DeepSeek + Codex) via takeover fan-out + sharp-review's merge engine. Disagreements between models surface genuine issues.
argument-hint	<slug> [platform]
allowed-tools	["Read","Write","Glob","Grep","Agent","Skill","Bash","mcp__plugin_fabric_fabric__call"]

/post-review — 三方会审

Two reviewer identities, each independently run by 3 heterogeneous models by default (Opus + DeepSeek + Codex; --fast drops to the first 2). Reviewers are fanned out directly via the fabric MCP tool (mcp__plugin_fabric_fabric__call); their findings are merged (dedup + confidence tagging) by sharp-review's merge-findings.js engine. post-review handles identity-specific configuration and pipeline integration.

身份A 读者代理人:  [Opus] [DeepSeek V4 Pro] [Codex]  → takeover fan-out → merge-findings → merged findings
身份B 技术核查员:  [Opus] [DeepSeek V4 Pro] [Codex]  → takeover fan-out → merge-findings → merged findings
                              ↓
                    Cross-identity synthesis

Three heterogeneous models independently review with the same prompt. ≥2 agree → confident. Only one flags it → needs human judgment.

参数解析

• <slug> — 文章目录名，对应 ongoing/<slug>/
• [platform] — 可选，指定单一平台；不填则审所有已生成平台

Walk ongoing/<slug>/2-draft/ to find the latest version number (highest N). For each platform, walk back through versions to find the latest copy. Also check for images.md in the version chain.

Phase Table

Phase	File	What Happens
Setup	01-identities.md	审稿身份 A/B prompts + finding JSON schemas + AI味 grades
Setup	02-reviewers.md	Reviewer model config arrays (Opus + DeepSeek + Codex) + --fast + merge-findings.js path resolution
Phase 1-2	03-execution.md	Per-identity takeover fan-out + merge-findings collection
Phase 3-5	04-synthesis.md	分身份合议 (P3) → 综合裁决 (P4) → 全平台总览 (P5)
Phase 6	05-images.md	MANDATORY image plan review (术语一致性, 残留引用, 封面标题匹配, 比例)
Phase 7	06-persist.md	Write fixed articles + corrected images.md + review-verdict.md to v<N+1>/

Phase 6 is not optional — every review round (including single-platform re-reviews) must审 images.md against the latest drafts. If no images.md exists yet, note that in the verdict; otherwise the verdict MUST contain an ## Image plan review section. A title/motivation change in the drafts almost always staleizes the cover hook — Phase 6 is what catches it.

How to Execute

Run phases in order. At the start of each phase, Read the corresponding sub-file and follow its instructions. This SKILL.md is just the map.

Phase 1-2 run per platform. Phase 3-4 synthesize per-platform (per-identity 合议 → cross-identity verdict). Phase 5 runs once for the all-platform overview.

Hard Rules

• Twitter/X skips identity B — no code to verify, text-only platform.
• Default is 3 reviewers per identity (Opus + DeepSeek + Codex). --fast drops to the first 2 (Opus + DeepSeek) to save a Codex call.
• Failed reviewer → ⚠️ 未响应, don't block the other identity.
• Verdict persists: every round writes review-verdict.md to v<N+1>/. Check it to determine review status across sessions.
• Image plan review is mandatory (Phase 6): every round审 images.md; review-verdict.md must include an ## Image plan review section (or note "no images.md yet"). Cover-hook stale after a title/motivation change is the most common miss — always re-check covers against the latest title.
• ❌ → manual rewrite needed, then re-review; ⚠️ → fix then re-check; ✅ → ready for publish.

Invocation

Called from post-new/09-review.md after version diff chain is collected. Also callable standalone via /post-review <slug> [platform].