@oh-story/dsh 0.1.9 → 0.1.11

package/lib/drama/skills/short-drama-produce/scripts/provider_adapters.py

@@ -41,13 +41,21 @@ MAX_ERROR_BODY_BYTES = 64 * 1024
 TERMINAL_FAILURES = {"failed", "cancelled", "canceled", "timeout", "expired"}
 SEEDANCE_RATIOS = {"adaptive", "1:1", "3:4", "4:3", "9:16", "16:9", "21:9"}
 SEEDANCE_REFERENCE_ROLES = {
+    "first_frame": "image_url",
+    "last_frame": "image_url",
     "reference_image": "image_url",
     "reference_video": "video_url",
     "reference_audio": "audio_url",
 }
+# One opening frame and one closing frame, because each names a single position
+# in the take. Note what is deliberately absent: the sibling adapter forbids
+# mixing frame conditioning with reference conditioning, and that rule belongs
+# to that provider's reference, not to this one. Do not carry it across.
+SEEDANCE_SINGULAR_ROLES = {"first_frame", "last_frame"}
 MINIMAX_VIDEO_RATIOS = {"adaptive", "1:1", "3:4", "4:3", "9:16", "16:9", "21:9"}
 MINIMAX_VIDEO_RESOLUTIONS = {"480P", "768P", "2K"}
 MINIMAX_VIDEO_PROMPT_LIMIT = 7000
+MINIMAX_SPEECH_TEXT_LIMIT = 5000
 MINIMAX_VIDEO_ROLES = {
     "first_frame": "image_url",
     "last_frame": "image_url",
@@ -369,10 +377,18 @@ def compile_seedance_payload(
     counters = {"image_url": 0, "video_url": 0, "audio_url": 0}
     labels = {"image_url": "图片", "video_url": "视频", "audio_url": "音频"}
     reference_tokens: list[str] = []
+    singular_seen: set[str] = set()
     for role in reference_roles:
         field = SEEDANCE_REFERENCE_ROLES.get(role)
         if field is None:
             raise ValueError(f"unsupported Seedance reference role: {role}")
+        if role in SEEDANCE_SINGULAR_ROLES:
+            if role in singular_seen:
+                raise ValueError(f"Seedance accepts one {role} reference")
+            singular_seen.add(role)
+        # A frame role still occupies a picture slot, so it keeps counting: the
+        # tokens are positional, and skipping one would point every later
+        # `@图片N` in the prose at the wrong picture.
         counters[field] += 1
         reference_tokens.append(f"@{labels[field]}{counters[field]}")
     if task_type in {"edit", "extend"} and "reference_video" not in reference_roles:
@@ -738,6 +754,82 @@ def compile_gpt_image_2_payload(job: Mapping[str, Any]) -> dict[str, Any]:
     }
 
 
+MINIMAX_SPEECH_EMOTIONS = {
+    "happy", "sad", "angry", "fearful", "disgusted", "surprised", "neutral",
+}
+
+
+def compile_minimax_speech_payload(job: Mapping[str, Any]) -> dict[str, Any]:
+    """Compile a tts job into the official MiniMax speech JSON body.
+
+    The voice catalogue is deliberately not enumerated here. Which preset voices
+    an account can reach depends on the model and the account, no published list
+    is authoritative for both, and a list frozen into this file would either
+    refuse a voice that works or vouch for one that does not. The document owns
+    the value and a reviewer reads it; the adapter only checks its shape.
+    """
+
+    text, parameters = _require_job(job, "tts")
+    parameters = _take(
+        parameters,
+        {"model", "voice_id", "speed", "vol", "pitch", "emotion",
+         "sample_rate", "bitrate", "format"},
+    )
+    model = parameters.pop("model", "")
+    if not isinstance(model, str) or not model.strip():
+        raise ValueError("MiniMax speech model must be explicitly configured")
+    voice_id = parameters.pop("voice_id", "")
+    if not isinstance(voice_id, str) or not voice_id.strip() or voice_id != voice_id.strip():
+        raise ValueError("MiniMax speech requires a confirmed voice_id")
+    if len(text) > MINIMAX_SPEECH_TEXT_LIMIT:
+        raise ValueError("MiniMax speech text exceeds the provider limit")
+
+    target_format = Path(job["outputs"][0]).suffix.casefold().lstrip(".")
+    requested_format = parameters.pop("format", target_format)
+    if requested_format != target_format or requested_format not in {"mp3", "wav"}:
+        raise ValueError("MiniMax output format must match a supported target extension")
+    sample_rate = parameters.pop("sample_rate", 32000)
+    bitrate = parameters.pop("bitrate", 128000)
+    if sample_rate not in {16000, 24000, 32000, 44100}:
+        raise ValueError("unsupported MiniMax sample rate")
+    if bitrate not in {32000, 64000, 128000, 256000}:
+        raise ValueError("unsupported MiniMax bitrate")
+
+    voice_setting: dict[str, Any] = {"voice_id": voice_id}
+    emotion = parameters.pop("emotion", None)
+    if emotion is not None:
+        if emotion not in MINIMAX_SPEECH_EMOTIONS:
+            raise ValueError("unsupported MiniMax speech emotion")
+        voice_setting["emotion"] = emotion
+    for name, low, high in (("speed", 0.5, 2.0), ("vol", 0.1, 10.0)):
+        value = parameters.pop(name, None)
+        if value is None:
+            continue
+        if isinstance(value, bool) or not isinstance(value, (int, float)):
+            raise ValueError(f"MiniMax speech {name} must be a number")
+        if not low <= float(value) <= high:
+            raise ValueError(f"MiniMax speech {name} is outside the supported range")
+        voice_setting[name] = float(value)
+    pitch = parameters.pop("pitch", None)
+    if pitch is not None:
+        if isinstance(pitch, bool) or not isinstance(pitch, int) or not -12 <= pitch <= 12:
+            raise ValueError("MiniMax speech pitch is outside the supported range")
+        voice_setting["pitch"] = pitch
+
+    return {
+        "model": model.strip(),
+        "text": text,
+        "stream": False,
+        "output_format": "hex",
+        "voice_setting": voice_setting,
+        "audio_setting": {
+            "sample_rate": sample_rate,
+            "bitrate": bitrate,
+            "format": requested_format,
+        },
+    }
+
+
 def compile_minimax_music_payload(job: Mapping[str, Any]) -> dict[str, Any]:
     """Compile an audio job into the official MiniMax Music 3.0 JSON body."""
     prompt, parameters = _require_job(job, "music")
@@ -1372,6 +1464,44 @@ def _run_minimax(job: Mapping[str, Any]) -> tuple[Path, str | None]:
     )
 
 
+def _run_minimax_speech(job: Mapping[str, Any]) -> tuple[Path, str | None]:
+    token = _credential("MINIMAX_API_KEY")
+    body = compile_minimax_speech_payload(job)
+    base = _base_url("MINIMAX_BASE_URL", MINIMAX_BASE_URL)
+    result, _ = _request_json(
+        f"{base}/t2a_v2",
+        provider="minimax-speech",
+        body=body,
+        token=token,
+    )
+    base_resp = result.get("base_resp")
+    if not isinstance(base_resp, Mapping) or base_resp.get("status_code") != 0:
+        raise AdapterFailure(
+            "MiniMax speech synthesis failed",
+            code=_provider_code(result) or "generation_failed",
+            request_id=_safe_token(result.get("trace_id")),
+        )
+    data = result.get("data")
+    audio = data.get("audio") if isinstance(data, Mapping) else None
+    if not isinstance(audio, str) or not audio:
+        raise AdapterFailure(
+            "MiniMax did not return audio data",
+            code="missing_audio",
+            request_id=_safe_token(result.get("trace_id")),
+        )
+    try:
+        content = bytes.fromhex(audio)
+    except ValueError as exc:
+        raise AdapterFailure("MiniMax returned invalid audio data") from exc
+    if not content:
+        raise AdapterFailure("MiniMax returned empty audio data")
+    trace_id = result.get("trace_id")
+    return (
+        _temporary_output(job, job["outputs"][0], content),
+        trace_id if isinstance(trace_id, str) else None,
+    )
+
+
 def _run_minimax_video(job: Mapping[str, Any]) -> tuple[Path, str]:
     token = _credential("MINIMAX_API_KEY")
     base = _base_url("MINIMAX_VIDEO_BASE_URL", MINIMAX_VIDEO_BASE_URL)
@@ -1484,8 +1614,15 @@ def _selftest() -> None:
         "modality": "music", "prompt": "cinematic", "references": [],
         "outputs": ["制作成果/a.mp3"], "parameters": {"lyrics": "[Verse]\nHello"},
     }
+    speech = {
+        "modality": "tts", "prompt": "你们做了多久？", "references": [],
+        "outputs": ["制作成果/a.mp3"],
+        "parameters": {"model": "configured-speech-model", "voice_id": "a-preset"},
+    }
     if compile_gpt_image_2_payload(image)["model"] != OPENAI_MODEL:
         raise RuntimeError("GPT Image 2 self-test failed")
+    if compile_minimax_speech_payload(speech)["voice_setting"]["voice_id"] != "a-preset":
+        raise RuntimeError("MiniMax speech self-test failed")
     seedance = compile_seedance_payload(
         video,
         model="configured-model",
@@ -1553,7 +1690,10 @@ def main() -> int:
     parser.add_argument(
         "provider",
         nargs="?",
-        choices=("seedance", "gpt-image-2", "minimax-music", "minimax-h3"),
+        choices=(
+            "seedance", "gpt-image-2", "minimax-music", "minimax-h3",
+            "minimax-speech",
+        ),
     )
     parser.add_argument("--selftest", action="store_true")
     args = parser.parse_args()
@@ -1571,6 +1711,7 @@ def main() -> int:
             "gpt-image-2": _run_openai,
             "minimax-music": _run_minimax,
             "minimax-h3": _run_minimax_video,
+            "minimax-speech": _run_minimax_speech,
         }
         path, provider_job_id = runners[args.provider](job)
         response: dict[str, Any] = {

package/lib/drama/skills/short-drama-review/SKILL.md

@@ -49,8 +49,8 @@ license: MIT
 只读对应资料，不预加载整个项目：
 
 - 原著分析：[原著分析审查表](references/rubric-source-analysis.md)
-- 故事、场景、行动与对白：[故事剧本审查表](references/rubric-story-script.md)
-- 身份、变体、连续性与图片提示词：[资产提示词审查表](references/rubric-assets-prompts.md)
+- 故事、场景、行动与对白，改编是否守住原著人物与主线，已接受节奏档案的核对，开场、结尾与中段的局部检查：[故事剧本审查表](references/rubric-story-script.md)
+- 身份、变体、连续性、时代与图片提示词：[资产提示词审查表](references/rubric-assets-prompts.md)
 - 原文落实、镜头、关键帧与视频运动：[视觉运动审查表](references/rubric-visual-motion.md)
 - 完整审查方法：[审查方法](references/review-method.md)
 - 制作端常见缺陷：[生产质量门](references/production-quality-gates.md)
@@ -71,9 +71,8 @@ license: MIT
 - 剧本对白、动作、声音、画面文字是否都有画面/声音承载；
 - 人物、地点、道具状态和镜头起止是否连续；
 - 提示词是否从准确起点到准确终点；
-- 目标模型画音同轨时，可复制正文的声音段是否写成正面闭合事件集：无对白镜不靠 `No dialogue`、
-  `无对白` 关闭语音通道，完整性声明也不枚举本镜没有的层（`VID-22`、`VID-23`，见
-  [视觉运动审查表](references/rubric-visual-motion.md)）；
+- 目标模型画音同轨时，声音内容、时序和发声边界是否与来源一致；无对白镜是否交代底声或静默
+  （`VID-22`、`VID-23`，见 [视觉运动审查表](references/rubric-visual-motion.md)）；
 - 私有输入、凭据、绝对路径或内部流程文字是否泄漏。
 
 缺少必要输入时只停止依赖它的判断，其他问题可以继续汇总。

package/lib/drama/skills/short-drama-review/references/production-quality-gates.md

@@ -71,25 +71,12 @@
 - 角色名/资产引用进入台词内文；
 - 音效、环境底噪与对白方向没有来源或相互遮蔽。
 
-目标模型画音同轨时（`VID-23`），这一节多四条只在同轨条件下成立的核对。四条里只有配乐那条
-有实测后果，其余三条是「正文要如实、完整地交代本镜声音」，按这个理由报，不要替它们编一套
-生成机制：
-
-- **配乐层在正文里整个缺席**（实测有后果）：写了空形式（H3 为 `non_diegetic_music: N/A`）的
-  14 次生成没有一次自带配乐，完全不提配乐的 3 次里 2 次返回了自造配乐。这一条按 major 起报。
-  **空形式本身就是合规写法**：`non_diegetic_music: N/A` 是「本镜没有非画内配乐」的正面陈述，
-  不是待填占位符。只有整行不存在时才报，永远不要要求 owner 删掉已经写好的空形式；
-- 用排除句代替声轨描述：正文出现 `No dialogue`、`无对白` 而没有正面写出这条声轨由什么组成、
-  覆盖多长——问题在于正文没交代本镜该有什么声音，不在于"否定会招来它"；
-- 完整性声明枚举了本镜没有的层：无对白镜写出
-  `The audio timeline is complete after the listed ambience, effects and timed dialogue.`
-  是一句与本镜事实不符的话；
-- 无对白镜从没说清声轨**是什么**：既没有覆盖全时长的底声构成，也没给在场人物一个可见的
-  不说话状态（嘴唇闭合）。
-
-中间两条同时出现时按一个 finding 报，证据引这两句话本身：它们互相矛盾，本身就是声音契约
-没写清。修订要求写成正面声轨，不能是"再加一句不要对白"。本套件实测未观察到这组写法导致
-模型补出台词，报 finding 时不要这样断言。
+目标模型画音同轨时（`VID-23`），检查声音内容、时序、发声边界是否符合来源，
+总结是否错误地引入不存在的声音层。无对白镜也应交代需要的底声或静默。
+H3 无非画内配乐时，`non_diegetic_music: N/A` 是有效的空形式。
+
+按实际遗漏、矛盾与影响报告问题，不因出现 `No dialogue` 或第二个否定句判错，
+不把 H3 的小样本表现套到所有同轨模型上。相关证据指向同一个问题时合并报告。
 
 ### 5. 负面补丁淹没动作
 

package/lib/drama/skills/short-drama-review/references/review-method.md

@@ -71,7 +71,12 @@ Resolve upstream meaning first; do not polish prompt wording when the shot or as
 ## Revision and re-review
 
 Group duplicate findings and route them to the owning skill. A revision request states the target outcome, preserve
-set and review scope to rerun. On a later explicit re-review request, read the current version, confirm the finding is addressed, check the
+set and review scope to rerun.
+
+`REV-14`: a revision request is a patch, not a rewrite. It names the smallest change at a cited location; local
+findings never add up to a request to rewrite the episode. When the review compares a revision with the version before
+it, any strong hook, payoff line, or comic beat that polishing removed is listed under `## 被删的强钩子` with its prior
+location and a short quote, so the owner can restore it. The reviewer lists it; the owner decides. On a later explicit re-review request, read the current version, confirm the finding is addressed, check the
 preserve set and close or retain each affected finding. A different reviewer is useful but not mechanically required.
 
 ## Anti-template review

package/lib/drama/skills/short-drama-review/references/rubric-assets-prompts.md

@@ -72,6 +72,17 @@ validity.
 
 Fail when a prop teleports, changes scale/material, or readable evidence is erased.
 
+## Era
+
+The era comes from source facts: devices, platforms, slang, stated years. Fail
+when the era shown in `视觉设定.md`, `图片提示词.md` or `视频提示词.md` contradicts
+them, or when a retro or aged descriptor (旧/老式/复古/褪色/斑驳, old, worn, dusty,
+period, CRT) has neither a source basis nor an explicit creator choice (`AST-14`).
+This is the visual layer of the source-fidelity check (`STY-07`): cite the source
+fact and the conflicting line. A visual direction with
+no era anchors, or entries that leave the era-specific form unstated, is a craft
+note (`AST-15`).
+
 ## Prompt recipe review
 
 All prompt types need purpose, exact binding, identifying facts, current variant,

package/lib/drama/skills/short-drama-review/references/rubric-story-script.md

@@ -5,9 +5,12 @@
 - [Promise](#story-promise-and-engine)
 - [Episode](#episode-shape)
 - [Entry and serial memory](#entry-character-and-serial-memory)
+- [Adaptation fidelity](#adaptation-fidelity)
 - [Scene](#scene-test)
 - [Action](#action-and-production-meaning)
 - [Dialogue](#dialogue)
+- [Rhythm profile](#rhythm-profile-rev-12)
+- [Local passes](#local-passes-rev-13)
 - [Findings](#common-findings)
 
 ## Story promise and engine
@@ -55,8 +58,9 @@
 A capacity estimate (`STY-16`) is informational, never a blocking finding. If a
 planned episode's shot/duration magnitude is far from the project's own accepted
 ratios, note it as an advisory observation with the sampled basis, and leave the
-resolution to the creator. Never derive a word, line, or shot quota from it, and
-never compare against another project's numbers.
+resolution to the creator. Do not derive a word, line, or shot quota from it. The only
+numbers a review checks are the ones in an accepted rhythm profile
+([below](#rhythm-profile-rev-12)), and never another project's numbers.
 
 When the premise grants foreknowledge or externally authorized ability (`STY-17`), keep
 two layers apart and cite each from its own place:
@@ -117,7 +121,28 @@ What is reportable:
   A breathing scene should process a consequence rather than suspend the story.
 
 Do not require a fixed hook type, beat count, reversal, climax timestamp, or
-dialogue quota.
+dialogue quota beyond what an accepted rhythm profile declares.
+
+## Adaptation fidelity
+
+Active only when the project adapts a source. Read the source span each mapping in
+`项目开发/adaptation-map.jsonl` cites, then the episode record and screenplay.
+
+- For each core source character: is their stance toward the protagonist, their way of
+  acting, and their relationships the same on screen as in the source (`STY-07`)?
+  Pressure made louder or spoken aloud is fine; a supportive figure turned oppressor,
+  or a relationship reversed, is a finding unless the creator accepted the change.
+  Cite the source fact and the conflicting screenplay line (`REV-03`).
+- Does the main line still run the source's way, and does any invented detail
+  contradict a source fact?
+- Is every character, scene, or opponent the source lacks registered as its own `add`
+  mapping, with the function it serves, why existing source material cannot carry it,
+  the source facts checked against it, and a line in the brief for the creator (`STY-30`)?
+  An addition hidden inside `change_carrier`, `merge`, or `move_earlier` notes is a
+  `craft_default` finding; an unregistered addition that also contradicts the source
+  is a `STY-07` finding.
+- When an episode's opponent was manufactured to meet `opposed_reversals_per_episode_min`,
+  say so: the profile deviation is the honest outcome, not a new antagonist.
 
 ## Scene test
 
@@ -197,6 +222,54 @@ them fit this checklist. A loud scene can fail when nothing changes.
 Silence, interruption, slang, narration, and sentence length are creator options,
 not universal scoring ratios.
 
+## Rhythm profile (`REV-12`)
+
+Active only when `short-drama.json#/creator_authority/rhythm_profile` has `status`
+`accepted`. A missing, `unset`, or `proposed` profile means nothing here is checked.
+The review does arithmetic on the accepted values against the current screenplay and
+nothing else:
+
+| Field | What the reviewer measures |
+|---|---|
+| `first_hook_seconds` | Seconds from the episode start to the first hook: crisis, conflict, anomaly, or promised spectacle already happening on screen. |
+| `beat_interval_seconds_max` | The longest gap between consecutive emotional touchpoints: a conflict line, a situation-changing action, or new information. |
+| `opposed_reversals_per_episode_min` | Reversals driven or suffered by a human opponent with an agenda; changes completed by a system, a number, or the environment alone do not count. |
+| `end_on_peak` | Whether the last beat sits on or just before the peak, with no afterglow, summary, or quiet coda after it. |
+| `reprise_previous_last_beat` | Whether the opening replays the previous episode's last beat. Not applicable to the first episode. |
+| `vo_share_max` | `[VO]` spoken characters as a share of all spoken characters; `$short-drama-write`'s duration estimate reports it when run with the project file. |
+| `first_major_payoff_by_episode` | Series review only: the episode in which the first major payoff lands. |
+
+Seconds come from the project's declared pacing, or from a real-time read-through
+labelled as a text estimate; state which. A deviation is a `craft_default` finding that
+cites the field, the accepted value, the measured value, and the location. Severity
+follows audience impact and is never `blocker`; the creator may keep the episode with a
+stated reason. Storyboard fields (`target_avg_shot_seconds`, `close_shot_share_min`)
+are measured in the [visual rubric](rubric-visual-motion.md#rhythm-profile-shots).
+
+## Local passes (`REV-13`)
+
+After the whole-episode read, run three passes that each read one part only. A whole
+read lets a strong middle excuse a flat opening; a part read cannot.
+
+1. **Opening hook**: read only the first segment, up to `first_hook_seconds` or, with
+   no accepted profile, the first scene. Would a viewer who saw only this stay? Name
+   the hook and say whether it is happening or only being set up. Then read the first
+   beat alone, the opening frame: does it raise a question by itself,
+   or does it need on-screen text read, a character recognized, or the next line heard
+   first (`SCR-22`)? A still image that poses its own question passes. For a first
+   episode, also check that the brief compared the entry candidates on that same image
+   (`STY-25`).
+2. **Ending suspense**: read only the last segment, the final scene's closing beats.
+   What question is left open? Does it grow out of this episode's result? Is it a line
+   hook (an unfinished question, an identity hint) or an image hook (a door half open,
+   an envelope half torn)? Does it stop before the answer?
+3. **Mid-episode reversal density**: read only the middle. List each reversal with its
+   location and the person driving or suffering it; flag long stretches with no
+   touchpoint and reversals bunched at the end.
+
+Each finding cites its own segment. The passes apply with or without a profile; only
+the numeric comparisons need one.
+
 ## Replaceable realization (`SCR-10`)
 
 Only for beats the creator marked. **An unmarked beat leaves this section
@@ -230,6 +303,7 @@ the benefit is not, and the mark exists precisely so the strong version can stay
 - cliffhanger withholds all payoff;
 - episode pauses an unfinished action without producing a local result (`STY-13`);
 - first episode dumps the backstory reservoir instead of selecting an active entry;
+- an adaptation hardens a source character or invents an opponent without an `add` mapping or creator acceptance;
 - a declared character arc has no pressure test, choice, cost, or changed strategy;
 - serial summary omits an information permission or unpaid setup obligation;
 - generic emotion adjectives replace visible performance;

package/lib/drama/skills/short-drama-review/references/rubric-visual-motion.md

@@ -7,6 +7,7 @@
 - [Keyframe](#frozen-keyframe)
 - [Motion](#motion)
 - [Continuity](#cross-shot-continuity)
+- [Rhythm profile](#rhythm-profile-shots)
 - [Findings](#common-findings)
 - [Production risks and authorized observations](#production-risk-and-authorized-observation-checklist-rev-08)
 
@@ -31,6 +32,11 @@
 - Why is this a new shot?
 - Do framing and camera behavior serve attention, alignment, pressure, reveal, or
   rhythm rather than decorate the prose?
+- Does each shot's `运镜` line state a behavior and a reason chosen from the shot's
+  purpose (`SHT-04`)? A stated lock-off with its reason passes. An episode locked on
+  nearly every shot while its peak gets no movement is a finding; so is a missing line,
+  which leaves the choice to the video-prompt stage; a bare 「运镜：固定」 with no reason
+  is also a finding.
 - Are Location/View, axis, screen direction, eyelines, entrances, positions,
   hands, and props coherent?
 - Are exact asset variants bound?
@@ -39,6 +45,15 @@
 - Do crop, occlusion, focus, back view, and offscreen space serve that information
   permission rather than accidentally reveal or conceal it?
 
+## Rhythm profile: shots
+
+Under the `REV-12` conditions in the [story rubric](rubric-story-script.md#rhythm-profile-rev-12),
+measure two storyboard fields: the average shot length per scene (scene duration divided
+by its shot count) against `target_avg_shot_seconds`, and the share of close-range shots
+across the episode, as the profile field defines them, against `close_shot_share_min`. The average is a
+scene-level target; one long or short shot is not a finding. Also report durations folded
+from screenplay word counts instead of the scene's sound timeline (`SHT-28`).
+
 ## Frozen keyframe
 
 - Can all described facts exist in one still instant?
@@ -86,41 +101,15 @@
 - Does each media reference state its exact purpose, what may be copied, and what
   must not be copied, rather than treating all visible pixels as authority?
 
-For a copyable body, ask these separately — they are the two rules a rewrite regresses on
-most often, because a body that reads "complete" can still carry both defects. Findings here
-are about whether the body states this shot's sound and text truthfully and completely; only
-the music-layer item below has a measured generation consequence behind it:
-
-- Does the body carry only what will be filmed (`VID-22`)? Negative action intent must
-  already be rewritten as a visible positive state. The one permitted negation is the
-  no-subtitle/caption/dialogue-overlay declaration; a second exclusion clause is a defect,
-  not extra safety. (The often-repeated reason — that a noun inside a negation gets
-  rendered — did not reproduce in this suite's probes, so cite the body's own clarity,
-  not that mechanism, when you raise this.)
-- When the target model generates audio in the same pass (`VID-23`), is the sound written
-  as a closed positive event set — what the track consists of, and across what span —
-  rather than a list of what must not be heard? Check three failures by name:
-  - an exclusion clause (`No dialogue`, `无对白`, `no music`) standing in for a positive
-    soundscape, so the body never says what this shot sounds like;
-  - a completeness declaration that enumerates a layer this shot does not have. "The audio
-    timeline is complete after the listed ambience, effects and timed dialogue" on a shot
-    with no lines states something untrue of this shot. Read the declaration against the
-    shot's actual layers, not against a template;
-  - a shot that leaves the non-diegetic music layer unstated. This is the one sound defect
-    with measured consequences: bodies that omitted the layer came back with an invented
-    score. **The dialect's empty form satisfies this** — for H3 that is the literal line
-    `non_diegetic_music: N/A`, which is a correct statement of "no non-diegetic music",
-    not a placeholder. Raise this only when the layer is missing from the body altogether;
-    never ask an owner to delete a present empty form;
-  - a shot with no lines whose body never says what the track *is* for the full duration,
-    and never gives the on-screen performers a visible non-speaking state.
-  Report a contradictory pair — an exclusion clause plus a declaration that enumerates the
-  excluded layer — as one finding against the body's sound contract, and quote both
-  sentences. Raise it as an internally inconsistent sound contract, which it is; do not
-  assert that it makes the model invent lines, which this suite has probed and not seen.
-
-These two apply to a body the review itself is about to ask an owner to rewrite. A revision
-request that removes a defect by adding `no ...` has moved the defect, not fixed it.
+For a copyable body (`VID-22`, `VID-23`), check that actions and sound match the
+source and that the timeline accounts for the intended ambience or silence.
+A summary must not introduce a layer absent from the shot. H3's
+`non_diegetic_music: N/A` is a valid empty music layer.
+
+Report a concrete contradiction, omission or ambiguity. A negative clause is
+not a defect by itself; do not impose a clause count or assume that negation
+causes the model to generate the excluded content. Keep model-specific
+observations scoped to their tested model and execution path.
 
 For a pickup or alternate, ask these separately:
 

package/lib/drama/skills/short-drama-review/references/stage-contract.md

@@ -22,6 +22,9 @@
 | REV-09 | reviewed_invariant | After prompt revision or repackaging, recheck source coverage and every applicable accepted directive; correct asset bindings alone do not prove compliance. |
 | REV-10 | reviewed_invariant | A project-calibration finding distinguishes input-reference from generated-result observation, binds the exact project, the prompt/spec records it observed, stable reference slots, production configuration, method and limits, and—when its disposition calls for a change (see REV-11)—proposes the smallest owner-routed one with a preserve set; it does not generalize across projects or infer quality from task state. |
 | REV-11 | reviewed_invariant | A calibration finding carries `disposition` (keep, post_production, targeted_edit, resubmit, rewrite) and `disposition_rationale` before any revision text, justified by whether the defect is text-controllable and whether it has already recurred. Dispositions that call for no change leave `required_change` empty rather than inventing one. Resubmitting identical text and appending quality adjectives are not repairs; a recurring defect routes to a structural change instead. Findings outside project calibration use `not_applicable`. |
+| REV-12 | craft_default | When the project's rhythm profile has status `accepted`, review checks the episode against it by arithmetic on the accepted values only and reports a deviation as a non-blocking finding citing field, accepted value, measured value, and location; the creator may keep the episode with a stated reason. A missing or proposed profile leaves the check inactive, and no finding compares against another project's or a suite-wide number. |
+| REV-13 | craft_default | Beyond the whole-episode read, three local passes each read one part only: an opening-hook pass on the first segment, which also reads the first beat alone as the opening frame, an ending-suspense pass on the last segment, and a reversal-density pass on the middle. |
+| REV-14 | craft_default | A revision request is a patch at a cited location with a preserve set, never a whole-episode rewrite; a strong hook, payoff line, or comic beat removed by polishing is logged with its prior location and quote so the owner can restore it. |
 
 规则分级由高到低：`structural_invariant`（结构缺陷，阻断）、
 `reviewed_invariant`（需证据判断）、`craft_default`（常用做法，可覆盖）、

package/lib/drama/skills/short-drama-storyboard/SKILL.md

@@ -7,7 +7,8 @@ license: MIT
 # 短剧分镜与冻结关键帧
 
 把剧本和视觉事实转成有镜头职责、空间连续性和可冻结起点的 `剧集/<EP>/分镜.md`。
-每镜使用二级标题 `## SHOT-...`，同镜下用 `### 冻结关键帧提示词` 写起始帧正文。
+每镜使用二级标题 `## SHOT-...`，同镜下用 `### 冻结关键帧提示词` 写起始帧正文；
+需要时另加 `### 收尾关键帧提示词` 写终点正文。
 
 ## Quick Start
 
@@ -25,6 +26,9 @@ license: MIT
 分配镜头时长前读取
 `short-drama.json#/creator_authority/production_profile/choices/native_duration_seconds`。已声明目标模型但
 没有原生时长时先补这一个选择；目标模型也未声明时照常按叙事节奏设计，不猜供应商限制。
+同时读取 `short-drama.json#/creator_authority/rhythm_profile`：`status` 为 `accepted` 时，镜长按
+[镜长从哪来](references/shot-craft.md#镜长从哪来-sht-28) 由声音事件与 `target_avg_shot_seconds` 推出，
+全集近景类镜头占比不低于 `close_shot_share_min`；档案缺失或未接受时不核对这些数值。
 
 冻结关键帧写完之前，谁会出现在这一格画面里还没有定下来。所以三条依据都在关键帧之后回填，
 读的是成稿画面，不是镜头意图。
@@ -47,12 +51,17 @@ license: MIT
 2. 确定观众何时知道什么、对齐谁、空间如何揭示、转折落在哪个可见动作。
    剧本或 Brief 用“前半段 / 后半段 / 前三分之一 / 最后 N 秒”等相对时间限定首次出现、释放或禁止时，
    先用已接受集长换成明确时间边界再分镜；例如 60 秒作品的“后半段首次出现”不得早于 30 秒。
-3. 每镜写清唯一职责、来源、时长、景别/机位、起点、唯一主要动作、终点和声音；用
+3. 每镜写清唯一职责、来源、时长、景别/机位、运镜、起点、唯一主要动作、终点和声音；用
    `起点 → 唯一动作 → 终点` 直接说明人物、双手与持物怎样完成本镜状态转换。三条依据留到第 6 步。
+   「运镜」按本镜目的逐镜选，写行为和理由，固定也写理由，见 [运镜逐镜选](references/shot-craft.md#运镜逐镜选-sht-04)。
    有对白、VO 或 OS 时，先按原文和人物说话方式估算完整发声时间，再安排不能重叠的动作、停顿与
    末尾反应；把估时依据简写在本镜「声音」里。具体做法见 [对白估时](references/shot-craft.md#对白估时)。
+   对白场先按 [对白覆盖默认](references/shot-craft.md#对白场怎么切-sht-29) 放切点；镜长不按剧本字数折算。
 4. 锁定人物朝向、位置、视线、持物、出入口、屏幕方向和必要状态；上游已有会限制动作空间的持续关系时，把关系投影进本镜起点与终点。
 5. 冻结关键帧只投影镜头起点；删除终点才出现的文字、动作、姿态或道具状态。
+   镜头的职责落在收尾那一下时，另写 `### 收尾关键帧提示词` 投影**终点**——同一格画面的
+   另一端，规则完全对称：删除起点才有、终点已经不在的东西。它是可选的，写法与取舍见
+   [首尾成对](references/shot-craft.md#首尾成对)。
 6. 关键帧成稿后，先从成稿正文读出这一格里需要保持身份、造型或地理的人物、地点和道具，
    再用同一份清单一次填满三条依据：
 
@@ -71,22 +80,24 @@ license: MIT
 
 - 每镜有唯一职责、明确来源、可见起止状态和合理时长。「来源」以《剧本.md》里真实存在的场景 ID 开头，
   ID 之后可以跟必要短引文，跨场次的镜头用 `、` 连写多条；`creator_markdown_check.py` 会核对每个 ID 都解析得到。
-- **引文是认领，不是背景。** 「来源」里引了哪一句，就等于说这一镜负责拍它。引一个本镜画面里
-  根本没有的人做的动作——把「另一个人低头去翻自己的稿子」写进一个只拍某一个人的特写——
-  那个动作就从所有人的清单上消失了：没有别的镜头认领它，也没有任何一步报它缺失，
-  于是它根本不会被拍出来，而这个缺口只有等人看完成片、问「他的反应呢」才浮出来。
-  确实看不见就在「视觉依据」末尾写 `；画外：<类别>「<名称>」`，机械核对认这个写法。
+- 「来源」短引文应对应本镜承担的动作或声音。正文提到画外人物时，可在视觉依据写
+  `；画外：<类别>「<名称>」`；这只声明其不在画面，不能代替必要动作的落实。
+  需要看见的动作应另有镜头承载，或说明为何省略。
+
 - 剧本里的每个场景都要有镜头承载。确实决定不拍某一场时，在《分镜.md》正文开头写一行
   `- 未拍场次：<场景 ID>（理由：……）`，多项用 `；` 连接。SHT-01 一直允许带理由的省略，
   但省略和漏拍在成稿文档里长得一模一样，所以理由要写下来，而不是留在对话里。
-- 已声明原生时长时，每个一镜一生成的镜头都落在其最短与最长秒数内。短于下限的叙事动作在合法
-  镜长内完成后保持终点；长于上限的动作按可见状态闭合点拆镜，拆出的每镜仍各有唯一职责与可信交界。
+- 已声明原生时长时，每镜照常有自己的唯一职责、起止边界和冻结关键帧，镜长先按叙事需要定。短于最短秒数
+  的镜头，向创作者提议把同场相邻的几镜装进一个多镜容器，在交付档案里声明；目标方言要求时，容器的每一切
+  都要有已接受的关键帧图片，能否装、怎么装由 `$short-drama-video-prompts` 的方言决定。创作者没有声明
+  多镜容器、或方言不支持时，镜头落在最短秒数内，动作完成后保持终点。长于上限的动作按可见状态闭合点拆镜，
+  拆出的每镜仍各有唯一职责与可信交界。
 - 「唯一动作」写从本镜起点到终点的那条可见状态链；涉及交接、换手、放置或取回时，把接触者、
   接触方式和物件落点写在同一链里，不能让合法终点之间靠镜外瞬移衔接。
 - 切镜必须带来信息、权力、情绪、空间或节奏变化；同义重复镜头删掉。
 - 相对时间词约束的是事件首次可见或可听的时刻，不因前置动作已经完成就提前释放；分镜时间轴直接
   写出该边界，不把“稍后”“交接后”等叙事顺序当成数值时间的替代。
-- 同一动作不跨镜重复发生；终点应能成为下一镜可信起点。
+- 同一动作不跨镜重复发生，峰值处声明过的[冲击重复](references/shot-craft.md#动作落实表)除外；终点应能成为下一镜可信起点。
 - 关键信息的可读性优先于装饰性运镜，竖屏构图优先保证主体和反应可见。
 - 复杂群戏先锁空间锚点和屏幕方向，再决定景别变化。
 - 起终边界除了逐人姿态，还要写清本镜依赖的身体—物件—空间关系。关系在镜内不变时两端都保留；确需改变时，过程与终点必须有剧本依据。
@@ -140,11 +151,12 @@ license: MIT
 默认只读本 SKILL、当前剧本和必要视觉事实。遇到对应问题时只打开一份：
 
 - 阶段边界与规则分级：[阶段契约](references/stage-contract.md)
-- 镜头职责、切镜理由和边界：[镜头手艺](references/shot-craft.md)
+- 镜头职责、切镜理由、镜长推导、对白覆盖和边界：[镜头手艺](references/shot-craft.md)
 - 实际制作所需的镜头语法：[生产镜头语法](references/production-shot-grammar.md)
 - 关键场次比较导演方案：[场次视觉计划](references/scene-visual-plan.md)
 - 检查原文是否都有画面承载：[Coverage Audition](references/coverage-audition.md)
 - 多人、群体和复杂空间调度：[调度手册](references/blocking-playbooks.md)
+- 谁被照亮、怎么写才执行得出：[光的设计](references/lighting-craft.md)
 - 冻结瞬间、注意中心和可读性：[关键帧手艺](references/keyframe-craft.md)
 - 漫剧/二维漫画形态的共享画风词汇：[漫剧关键帧词表](references/comic-keyframe-lexicon.md)
 - 剧本事实到关键帧的完整示例：[剧本到关键帧](references/screenplay-to-keyframe-example.md)

package/lib/drama/skills/short-drama-storyboard/references/blocking-playbooks.md

@@ -24,7 +24,7 @@
 - 先保证进出方向和视线连续，再用近景强化选择。
 
 可以按以下顺序设计：空间锚点与入口 → 每人的行动目标和筹码 → 群体注意中心 →
-关键动作前后的站位 → 哪些同时反应必须留在全景中 → 只有丢失信息时才切插入镜头或近景。
+关键动作前后的站位 → 哪些同时反应必须留在全景中 → 再按[对白覆盖默认](shot-craft.md#对白场怎么切-sht-29)切到说话者、听者和插入镜。
 
 常见问题：每句台词都切一个头像近景；站位只为填满画面；背景人物在切镜后无原因换边；
 重要人物一直被前景遮住，但剧情并未利用这种遮挡。