跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
搜索与研究

comic-asset-review-loop

Phase-1 (S5) UPSTREAM ref-asset gate — the bounded cross-model adversarial loop that LOCKS one reusable identity-locked asset (character sheet / location plate / prop cutout / text-panel / logo-free symbol) BEFORE it can be composited into any panel. This is NOT the panel_gate (composed-output) or the assembly_gate (cross-panel) — it gates the refs that FEED both, so identity drift is caught at the source, not downstream. Two layers: (1) a MANDATORY static single-source collision check (check_asset_collisions.py — 'one visual dialect, never two', enforced by tool) + zero-text/literal build-asserts; (2) a per-round 3-reviewer panel (Claude narrative ‖ Gemini visual ‖ Codex synth) that blind-scores 5 dims and LOCKS only at 准×3 unanimity (prior_lock_count>=3 approvals in the SAME round) — else regenerate (route to comic-asset-ref-generator) or, at MAX_REVIEW_ROUNDS, escalate the asset REQUIREMENT back to outline/storyboard. Hard IP veto. Use when you say 'lock this character ref', 'asset

wanshuiyin · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add wanshuiyin/ARIS-Movie-Director --skill comic-asset-review-loop

技能说明


name: comic-asset-review-loop description: "Phase-1 (S5) UPSTREAM ref-asset gate — the bounded cross-model adversarial loop that LOCKS one reusable identity-locked asset (character sheet / location plate / prop cutout / text-panel / logo-free symbol) BEFORE it can be composited into any panel. This is NOT the panel_gate (composed-output) or the assembly_gate (cross-panel) — it gates the refs that FEED both, so identity drift is caught at the source, not downstream. Two layers: (1) a MANDATORY static single-source collision check (check_asset_collisions.py — 'one visual dialect, never two', enforced by tool) + zero-text/literal build-asserts; (2) a per-round 3-reviewer panel (Claude narrative ‖ Gemini visual ‖ Codex synth) that blind-scores 5 dims and LOCKS only at 准×3 unanimity (prior_lock_count>=3 approvals in the SAME round) — else regenerate (route to comic-asset-ref-generator) or, at MAX_REVIEW_ROUNDS, escalate the asset REQUIREMENT back to outline/storyboard. Hard IP veto. Use when you say 'lock this character ref', 'asset gate', 'ref review', 'is this sheet reusable', '锁角色 ref', before any metered panel bake." argument-hint: [asset_id | --batch-from-outline | --lite] allowed-tools: Bash(*), Read, Write, Edit, Grep, Glob, mcp__codex__codex, mcp__codex__codex-reply, mcp__gemini__analyzeFile, mcp__gemini-cli__ask-gemini, mcp__oracle__consult

comic-asset-review-loop — the UPSTREAM ref gate (Phase 1 · S5)

The ref gate the comic side was missing. comic-director's panel_gate reviews composed output; the assembly_gate reviews cross-panel consistency. Neither gates the identity refs that feed both — so without this skill, identity drift is only caught downstream, after credits are spent on a panel bake. This skill is the Layer-2 gate on a SINGLE reusable reference asset (the character sheets / location plates / prop cutouts the ART_BIBLE.md points every bake at), run before any panel composition.

A reusable asset is a one-way door: once a hundred panels condition on duo_canonical_ref_v001.png, you cannot quietly swap it. So the lock must be earned cross-model, never self-certified by the generator. This skill locks an asset only at 准×3 — Claude and Gemini and Codex each independently pass the same review round — or it sends the asset back to be regenerated, or (at the round ceiling) escalates the asset requirement itself back to the outline/storyboard layer. The executor (and the generation model family) never acquit their own ref.

Pipeline position (the Phase-1 asset DAG — the documented contract): this lock runs AFTER the human-approved outline (OUTLINE_DRAFT_VALID only requires every referenced asset to be DECLARED with a complete, generatable request — never pre-locked assets) and the provisional storyboard's consolidated_asset_requests, and BEFORE OUTLINE_FINAL_LOCK (cheap re-check that the locked assets still match the approved outline) → storyboard FINAL asset-resolution validation → blueprint authoring, where the hard locked-asset barrier sits. (The old single-stage contract — outline gate demanding locked assets — deadlocked and is retired.)

  asset (review_status:"pending")  ◀── comic-asset-ref-generator (S4: generates, NEVER self-locks)
        │
        ▼
   LAYER 1 — STATIC CI GATE (deterministic, zero credits)
        check_asset_collisions.py  (one canonical owner per filename — "one visual dialect, never two")
        + build asserts: zero-text contract · literal contract · raster-ref 6-field completeness
        │  pass
        ▼
   LAYER 2 — VISUAL LOOP   (round R = 1 … MAX_REVIEW_ROUNDS=3)
        ① Claude narrative  ‖  ② Gemini visual (primary: identity/iso)  ‖  ③ Codex synth (policy/semantic)
        each reads the REAL asset file · scores 5 dims (0–5 ints) · flags failure_modes from the fixed vocab
        │
        ▼ Codex synthesises: MIN-fuse per dim (max for inverted) · threshold · per-reviewer unanimity
   准×3 ? ── no, a family HARD-FAILS a floor ─▶ comic-asset-ref-generator --regenerate (bump version) ─▶ ① R+1
        │ no, third family just ABSENT (e.g. Gemini down) ─▶ re-VOTE (re-collect the missing reviewer; NO re-bake)
        │ no (safety_ip<3)  ─▶ HARD abandon_shot (IP veto, no retries)
        │ no (R == MAX_REVIEW_ROUNDS) ─▶ ESCALATE the asset REQUIREMENT to outline/storyboard (abandon_shot)
        │ yes
        ▼
   LOCK  — asset.status=locked + payload.review_status=locked  (ONE-WAY DOOR, immutable for the rest of Phase 1)
        + score_progression audit table (ASSET_REVIEW.md) + decision/review/failure_mode wiki nodes + edges

Constants

  • REVIEWERS = Claude (this agent, narrative) ‖ mcp__gemini__analyzeFile (model: "auto-gemini-3", NEVER raw gemini-3.1-pro — 429 silent-downgrade to 2.5) ‖ mcp__codex__codex r1 / mcp__codex__codex-reply r2+ (model_reasoning_effort: "xhigh", save the threadId; pin no model — Codex reviewer calls follow the local codex config, currently gpt-5.6-sol, exactly like the shipped Codex CLI reviewers in run_comic.py, which pass no -m; only the upstream BAKE payload pins gpt-5.5 + xhigh, the run_comic.get_bake_plan() compat default — a config-driven override there is planned, not yet implemented).
  • 准×3 UNANIMITY — the lock predicate. A Codex-synth approve is NOT enough: cc_approves AND gemini_approves AND codex_approves must ALL be true in the same round (each reviewer's own dim scores must independently clear threshold). Codex-approve-without-unanimity is DOWNGRADED to revise (the "准×3 bite"). prior_lock_count = the count of approvals accumulated this round; lock iff prior_lock_count >= 3.
  • MAX_REVIEW_ROUNDS = 3 (hard ceiling). Even if Codex keeps saying regenerate, the loop STOPS at 3 and routes back to outline/storyboard — the requirement changes; the gate does not grind an image forever.
  • CROSS-MODEL ACQUITTAL — Codex is the **synth/pol
阅读完整来源 ↗