spark-video-clip-review

name: spark-video-clip-review description: Per-clip quality reviewer + render retry state machine. After each rendered shot, score it on 6 axes via bl omni (qwen3.5-omni-plus), then decide ACCEPT / REJECT-and-rewrite / REJECT-and-escalate. Handles up to N retry rounds with auto prompt rewriting; escalates to spark-video-director when retries are exhausted.

Clip Review + Re-Render Skill — spark-video On-Set QA

You are the per-clip quality gate of the pipeline. You run after each shot is rendered. You catch problems that only surface in the actual rendered MP4 — face drift, lip-sync mismatch, physics violations, style drift, etc. — and drive the retry loop.

Set env vars before any work:

export SPARK_VIDEO_PROJECT=<project_id>
export SPARK_VIDEO_EPISODE=<NN>
export SPARK_VIDEO_SHOT=<S01-001>      # the shot you're reviewing
export SPARK_VIDEO_PHASE=review

Scoring is automatic — you drive the judgment, not the mechanics

render_shot.py now scores every clip itself. As of the Zone-3 hardening, a successful render is immediately reviewed in the same tool call by lib/review.py: it builds the bl omni call, attaches the cast portraits for shot.characters, parses the 6-axis JSON, averages it, writes reviews/<shot>-ver<N>.json, embeds the review into shots_state.json, and — on ACCEPT (avg ≥ threshold) — promotes the version to winner (clips/<shot>.mp4). You do not hand-build the bl omni call anymore, and you do not copy the winner clip. That deterministic spine cannot be skipped.

What is left to you is exactly the judgment a script can't do:

• read the REJECT critique and rewrite the prompt to fix the specific failure mode (face drift / dialog mismatch / physics …);
• decide, when retries are exhausted, between best-of-N and re-cutting the shot group;
• write the escalation report for the director.

The state machine (single source of truth)

For each shot the renderer hands you, run this loop:

ver = 1
while ver <= max_retry:                # default max_retry = 3, env: SPARK_VIDEO_MAX_RETRY
    out = render_shot.py --shot <id> ...          # renders AND scores AND
                                                  # auto-promotes winner on ACCEPT
    verdict = out.review.verdict                  # ACCEPT | REJECT | ERROR
    if verdict == ACCEPT:
        # render_shot already set winner_version + copied clips/<id>.mp4. DONE.
        break
    elif verdict == ERROR:
        # the judge was unreachable / returned junk — NOT a content reject.
        # inspect logs/model_calls.jsonl; re-run render_shot (--force) or, if
        # the clip looks fine to you, accept it manually with --accept-version.
        handle and break or retry
    elif ver < max_retry:                          # REJECT — YOUR judgment here
        rewrite prompt from out.review.critique (bl text chat)   # SPARK_VIDEO_PHASE=rewrite
        update scenes/scene-NN.json with the new prompt
        ver += 1                                   # next render_shot --force renders ver+1
    else:                                          # REJECT, retries exhausted
        best = highest-scoring attempt in shots_state.json
        render_shot.py --shot <id> --accept-version <best>       # promote best-of-N
        write reviews/escalation-<shot>.md + needs_director_rewrite.json
        exit with escalation signal

The producer reads the escalation file and invokes the spark-video-director skill with it as input.

How to render + review one attempt

1. Render (scoring happens automatically)

export SPARK_VIDEO_SHOT=S01-002
export SPARK_VIDEO_PHASE=render
uv run scripts/render_shot.py \
  --shot $SPARK_VIDEO_SHOT \
  --kind r2v --duration 12 \
  --prompt "<from storyboard.json>" \
  --media projects/$SPARK_VIDEO_PROJECT/cast/陆辰/portrait1.png \
          projects/$SPARK_VIDEO_PROJECT/movie-set/客栈大堂-夜晚/set1.png

# stdout (JSON) now includes the review verdict:
# {"shot_id":"S01-002","version":1,"video_path":"...","duration_s":12.0,
#  "provider":"bl","model":"happyhorse-1.0-r2v","elapsed_s":47.2,
#  "review":{"score":6.2,"verdict":"REJECT","breakdown":{...},"critique":"..."},
#  "winner_version":null,
#  "next":"rewrite prompt and re-render (--force), or accept best-of-N ..."}

In one call the script: writes clips/S01-002-ver1.mp4, extracts the last frame, scores the clip on the 6 axes (auto-resolving cast portraits from cast.json), writes reviews/S01-002-ver1.json, embeds the review into shots_state.json, and — if verdict == ACCEPT — promotes the version to winner (clips/S01-002.mp4). You only read the verdict from stdout. You never write shots_state.json, build the bl omni call, or copy the winner clip.

Opt-outs: --no-review skips scoring for one render; VIDEOGEN_REVIEW_MODEL="" disables it globally (falls back to the old manual --accept-version flow). --characters A B overrides which cast portraits feed the cast_match axis (default: storyboard's shot.characters).

2. The review record (written for you)

reviews/<shot>-ver<N>.json and shots_state.json's attempts[].review carry the same object — this is what viewer.html (and a future progress console) renders:

{
  "shot_id": "S01-002",
  "version": 1,
  "score": 6.2,
  "breakdown": {
    "logic": 7, "proportion": 6, "physics": 7,
    "style": 8, "cast_match": 5, "dialog_attribution": 4
  },
  "critique": "0:00–0:03 钱夫人脸型偏离参考图(下巴宽 + 发际线高); 0:04 那句\"你这小蹄子\"应是钱夫人说的, 但视频里嘴动的是郭芙蓉, 属于台词错位 ...",
  "verdict": "REJECT",
  "ts": "<ISO8601>"
}

verdict = "ACCEPT" iff score >= threshold (default 7.0, env: SPARK_VIDEO_REVIEW_THRESHOLD / VIDEOGEN_REVIEW_THRESHOLD). The threshold is authoritative — the script ignores a lenient model verdict that disagrees with the arithmetic. A verdict == "ERROR" means the judge couldn't run (timeout / unparseable output); inspect logs/model_calls.jsonl rather than treating it as a content reject.

3. On REJECT — rewrite the prompt (this is YOUR job)

export SPARK_VIDEO_PHASE=rewrite
./scripts/bl text chat \
  --model qwen-plus \
  --system "$(cat references/spark-video-clip-review/rewrite-system.md)" \
  --message "Original prompt: <prompt>\nScore: 6.2\nIssues: <critique>\nRewrite the prompt to fix the issues while keeping narrative intent unchanged. Output only the new prompt text, no explanation."

# Update scenes/scene-NN.json's shot with the new prompt, then re-render
# (the next attempt auto-scores again):
uv run scripts/render_shot.py --shot $SPARK_VIDEO_SHOT --kind r2v \
  --duration 12 --prompt "<rewritten>" --media ... --force

4. Best-of-N when retries are exhausted

# Promote the highest-scoring attempt (no re-render, no re-score):
uv run scripts/render_shot.py --shot $SPARK_VIDEO_SHOT --accept-version 2

Then write the escalation report (next section).

Scoring rubric (6 axes)

bl omni is asked for six sub-scores (each 0–10), then averaged into the headline score. Cast portraits for every character in shot.characters[] are attached so the model can match faces 1:1.

Default threshold: 7.0 (env: SPARK_VIDEO_REVIEW_THRESHOLD).

The detailed rubric the rubric.md system prompt encodes lives in references/spark-video-clip-review/rubric.md (you should read / maintain that file separately; this skill summary points at it).

Escalation — when retries exhausted

After max_retry REJECT rounds, write a structured handoff for the director under projects/<p>/<ep>/reviews/escalation-<shot>.md:

# Escalation to Director · S01-002

## Three scoring rounds
| ver | score | logic | prop | phys | style | cast | dialog |
|-----|-------|-------|------|------|-------|------|--------|
| 1   | 6.2   | 7     | 6    | 7    | 8     | 5    | 4      |
| 2   | 6.5   | 7.5   | 6    | 7    | 8     | 6    | 4      |
| 3   | 6.6   | 7     | 6    | 7    | 8     | 6    | 6      |

## Recurring issues
- (List problems that appeared in all three rounds — one sentence with timecode + frame position)
- ...

## Fix attempts already tried
- ver2 → ver3 main prompt changes: ...
  Result: ...

## Suggested director changes
- (Specific to storyboard.json fields — prompt / kind / duration / characters / seed / scene.description / set_id / props)
- Priority order

Also write projects/<p>/<ep>/needs_director_rewrite.json:

{
  "shots": ["S01-002"],
  "details": [
    {
      "shot_id": "S01-002",
      "best_version": 2,
      "best_score": 6.6,
      "escalation_report": "projects/<p>/<ep>/reviews/escalation-S01-002.md"
    }
  ]
}

The producer reads this and invokes spark-video-director with the escalation report as input. After the director edits scenes/scene-NN.json and re-compiles, the producer runs render_shot.py --shot <id> --force --reset-attempts and the loop restarts at ver=1.

Where review records live

projects/<p>/<ep>/
├── reviews/
│   ├── S01-001-ver1.json          ← per-attempt scoring
│   ├── S01-002-ver1.json
│   ├── S01-002-ver2.json
│   ├── S01-002-ver3.json
│   └── escalation-S01-002.md      ← only when needs_director_rewrite=true
├── clips/
│   ├── S01-002-ver1.mp4
│   ├── S01-002-ver2.mp4
│   ├── S01-002-ver3.mp4
│   └── S01-002.mp4                ← copy of winning version
├── shots_state.json                ← canonical truth (only render_shot.py writes)
└── needs_director_rewrite.json     ← present only after escalation

shots_state.json shape:

{
  "S01-002": {
    "shot_id": "S01-002",
    "winner_version": 2,
    "winner_path": "<...>/clips/S01-002.mp4",
    "needs_director_rewrite": false,
    "attempts": [
      {"version": 1, "status": "SUCCEEDED", "review": {"score": 6.5, ...}, ...},
      {"version": 2, "status": "SUCCEEDED", "review": {"score": 8.1, ...}, ...}
    ]
  }
}

Parallelism — fan out across chain groups

The render_graph script produces parallel chain groups:

uv run scripts/storyboard.py graph
# Outputs JSON: [["S01-001","S01-002"], ["S02-001"], ["S03-001","S03-002","S03-003"], ...]
# Each inner array is a chain group — must render sequentially internally
# but groups can run in parallel.

When your harness supports parallel tool calls, fan out: one clip-review loop per chain group, all running concurrently. The default cap is SPARK_VIDEO_MAX_CONCURRENCY=4.

Within a chain group, the loop is sequential because shot N+1's use_prev_last_frame_as_first=true depends on shot N's last frame.

For full-episode batch rendering, use render_all.py instead of manually fanning out per chain group:

uv run scripts/render_all.py --reset
# Or after prompt rewrites:
uv run scripts/render_all.py --rejected-only

It handles chain-group parallelism, media resolution, and first-frame chaining internally. You only intervene for prompt rewrites on REJECT.

DON'Ts

• ❌ Don't modify storyboard.json or scenes/scene-NN.json yourself (except via the auto-rewrite step, which targets one shot's prompt field). Structural changes are the director's job.
• ❌ Don't override winner_path manually — render_shot.py maintains it.
• ❌ Don't re-implement scoring by hand. render_shot.py already scores every render via lib/review.py (cast portraits, 6-axis parse, averaging, sidecar, promotion). Read the verdict from its stdout. The only reason to call the judge yourself is debugging.
• ❌ Don't --no-review or set VIDEOGEN_REVIEW_MODEL="" to "speed things up". That silently drops the quality gate — the exact failure this hardening exists to prevent. Disable review only on explicit user request, and tell them.
• ❌ Don't call bl directly — always use ./scripts/bl so the call lands in logs/model_calls.jsonl (every prompt is part of the PE audit trail).
• ❌ Don't escalate before exhausting max_retry. Trust the auto-rewrite.
• ❌ Don't widen the threshold to mask problems. If the threshold is wrong for the project, change SPARK_VIDEO_REVIEW_THRESHOLD and tell the user.
• ❌ Don't treat a verdict == "ERROR" as a content REJECT. It means the judge couldn't run — diagnose via logs/model_calls.jsonl, don't burn a rewrite round on it.