@oh-story/dsh 0.1.0

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

@@ -0,0 +1,153 @@
+---
+name: short-drama
+description: 基于文件系统初始化、继续和交付短剧或漫剧项目，并提供面向创作者的本地 Dashboard；也负责项目级制作形态、视觉方向和 Look Development 路由。用户提出“创建/继续短剧项目”“看进度/下一步”“做 Look Development”“打开 dashboard/短剧创作台”“导出制作资料”，或任务跨多个环节而需要判断负责技能时使用；明确的写作、资产、提示词、分镜或审查请求由对应子 skill 直接处理。
+license: MIT
+---
+
+# 短剧创作路由
+
+本技能负责找到项目、显示当前状态并把工作交给正确的创作技能，不代写各阶段内容。
+
+## Quick Start
+
+先离线确认本技能被单独复制后仍能初始化、读取和保护项目路径：
+
+```bash
+python3 {技能目录}/scripts/selftest.py
+python3 {技能目录}/scripts/project_tool.py init ./my-drama --title "示例短剧"
+```
+
+创作者可读内容使用 `short-drama.json#/language`；交给图片或视频生成器的提示词正文使用
+`#/format/prompt_language`。两者互不推断，详见
+[contract-and-ownership.md](references/contract-and-ownership.md#output-language-contract)。
+
+## 每次请求的起点
+
+1. 使用用户给出的路径，或从当前目录向上寻找最近的 `short-drama.json`。
+2. 运行 `status`，只读完成当前任务所需的直接输入，不批量加载整个项目。
+3. 按用户眼下要完成的工作路由；未安装的阶段技能只影响对应路由，不阻断其他工作。
+4. 明确的单阶段任务直接交给独立安装的对应技能，按创作者要的那一步做。
+
+统一预检见 [runtime-preflight.md](references/runtime-preflight.md)，创作入口见
+[creator-workflow.md](references/creator-workflow.md)，所有权见
+[contract-and-ownership.md](references/contract-and-ownership.md)。
+
+## 意图路由
+
+| 创作者意图 | 路由 |
+|---|---|
+| 开发点子、系列承诺、改编和分集地图 | `$short-drama-develop` |
+| 导入已有多集完整剧本/散稿并生成或补分集地图 | `$short-drama-develop` 按文件实际结构建立索引，每次只读当前集并续跑 |
+| 判断长篇是否值得改、建立原文索引 | `$short-drama-novel-analyze` |
+| 写或改单集契约、节拍、剧本 | `$short-drama-write` |
+| 拆人物、造型、地点、道具和状态 | `$short-drama-assets` |
+| 写人物、地点、道具和 Look Development 图片提示词 | `$short-drama-image-prompts` |
+| 做覆盖设计、镜头和冻结关键帧 | `$short-drama-storyboard` |
+| 写动作、表演、运镜、声音视频提示词 | `$short-drama-video-prompts` |
+| 写主题曲/配乐意图或校验时间线音乐规格 | `$short-drama-video-prompts`，歌词必须由创作者提供并接受 |
+| 按已确认规格实际生成图片、视频、TTS 或时间线音乐 | `$short-drama-produce` 先展示本次 job，得到明确确认后才执行 |
+| 定制作形态、视觉方向或 Look Development 路径 | `$short-drama` |
+| 校验、审稿或发修订请求 | `$short-drama-review` |
+| 打开创作台、看内容和进度 | `$short-drama dashboard` |
+
+用户的明确意图优先。Look Development 是可选分支；资产图片提示词和分镜在资产确认后可以
+并行，不互相等待。需要把已接受的视觉方向变成可比较的代表帧时，读
+[look-development.md](references/look-development.md)。
+
+**产物可以直接入场**。 上游的可选环节缺失不构成阻塞：现成剧本可直接进入资产；已接受资产
+可直接进入资产图片提示词；已接受剧本加资产可直接进入分镜；已接受镜头与关键帧可直接进入
+视频提示词。只补建对应技能拥有的、确实缺少的必要产物，绝不倒补虚构的创作简报或系列大纲。
+
+**请求含糊时不要追问技术细节**。 遇到"帮我把这集做完"这类请求，先展示当前的创作者可读
+状态，再给出不超过三个有意义的动作，优先能解开当前交付阻塞的那个；不要拿 schema、事务或
+存储格式去问创作者。
+
+## 有界续跑
+
+全链预览或“继续”每轮只授权一个 owner 的一个有界工作单元。写作、资产、图片提示词、分镜和
+视频提示词按各自「每轮的工作单元」分批工作；完成当前批次后报告已覆盖范围、剩余范围与
+下一步，并把控制权交还创作者。下一阶段、审查和生产各自需要明确的后续请求；生产还要先展示
+准确 job 并取得确认。
+
+审查尽量交给没有参与该版本创作的 reviewer，以减少自证偏差；运行环境不支持隔离上下文时，
+如实标注为自检即可。CLI 只记录 verdict、reviewer 标签和备注，不收集或伪造上下文证明。
+Reviewer 给出证据和修改请求，内容仍由原 owner 修改。
+
+## 初始化
+
+没有项目且用户要初始化时：
+
+1. 确定标题、项目语言、提示词语言、画幅和路径；集数或时长未知就留空。
+2. 运行 `init` 建立最小目录和空状态，不覆盖现有文件。
+3. 将制作形态与视觉方向保持为 `unset`，告诉创作者下一个最有用的选择。
+4. 创作者定下形态、方向、播放面或集长目标后，把这条决定发布并接受为一条创作者决策记录，
+   再用 `set-authority` 写回 `short-drama.json`，让下游读到已 `accepted` 的值。
+
+初始化不自动生成故事、剧本或资产设定。
+
+## 本地 Dashboard
+
+Dashboard 是项目内容的轻量展示与有限文本编辑层，核心能力仍在各 skill 中。当用户调用
+`$short-drama dashboard` 时：
+
+1. 项目存在时读取 `status`；浏览普通目录时直接使用该目录。
+2. 项目内启动时用项目根作为 workspace；否则使用用户给出的容器目录或当前目录。
+3. 从本技能安装目录运行：
+
+   ```text
+   python3 <short-drama-skill-dir>/scripts/dashboard_server.py --workspace <workspace> --port 0 --open
+   ```
+
+4. 回报脚本打印的完整回环地址与停止方式，并保持进程运行。
+
+Dashboard 只读取 workspace 内的文件；密钥、生产 adapter 与工作流编排都在它之外。它按项目
+和剧集展示正文、结构化卡片以及已有图片/音频/视频；保存只表示保存文件，不代表创作者确认。
+每次启动使用独立本机会话，项目 API 只接受该会话。参数与安全边界见
+[lifecycle-commands.md](references/lifecycle-commands.md#dashboard-启动)。
+
+## 项目命令
+
+从本技能安装目录调用 `scripts/project_tool.py`：
+
+| 命令 | 用途 |
+|---|---|
+| `init` | 初始化最小项目 |
+| `status` | 显示项目与产物状态 |
+| `publish` | 原子发布一项产物并记录它的直接输入 |
+| `accept` | 记录创作者接受或拒绝当前版本 |
+| `review` | 记录当前版本的复核结论 |
+| `set-authority` | 把已接受的创作者决策写回制作形态、视觉方向、播放面或集长目标 |
+| `package` | 打包当前已确认且复核通过的文本/JSON |
+| `verify` | 用交付校验和检查打包结果是否被改动 |
+
+完整示例见 [lifecycle-commands.md](references/lifecycle-commands.md)。单文件发布用临时文件与
+原子替换完成。
+
+## 状态与修订
+
+对创作者只使用六种状态：
+
+- `draft`：尚未发布；
+- `needs_confirmation`：等待创作者决定；
+- `accepted`：创作者已接受；
+- `revise`：创作者或 reviewer 要求修改；
+- `approved`：当前版本已确认且复核通过；
+- `update_needed`：输出或某个直接输入已经变化，需要重新发布。
+
+状态在读取时核对当前文件，不把过期传播写回整个项目。修订时只说明负责技能、语义变化、
+直接受影响的产物和下一步；重新发布会清除该产物旧的接受与复核结果，不触碰无关产物。
+除非诊断需要，不向创作者显示内部 hash 或原始状态文件。
+
+## 交付
+
+先由 `$short-drama-review` 复核，再用 `package` 打包。包内只收录明确选择的、当前状态为
+`approved` 的文本和 JSON，并记录有意省略项；二进制媒体、非公开输入、机器状态、绝对路径、
+凭据和未批准草稿不进入文本交付包。`verify` 检查清单、校验和及未登记新增文件。
+
+## 边界
+
+- 图片、视频和 TTS 生产只由 `$short-drama-produce` 在展示准确 job 并取得本次明确确认后执行；
+  其他技能不直接调用生成服务。
+- 运行时不检索外部或非公开生产来源。
+- 不把别处案例提升为项目定律。
+- 语义冲突不静默修复；不明外部改动由创作者选择保留或重做。

package/lib/drama/skills/short-drama/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "短剧创作"
+  short_description: "基于文件系统初始化、恢复、路由、可视化管理并交付短剧项目"
+  default_prompt: "使用 $short-drama 初始化、查看、继续、恢复或打开 Dashboard 管理我的文件系统短剧项目。"

package/lib/drama/skills/short-drama/assets/creator-decision.example.jsonl

@@ -0,0 +1,4 @@
+{"record_type":"sources","schema_version":"1.0.0","sources":{"short-drama":{"owner":"short-drama","artifact":"short-drama.json"}}}
+{"example_only":true,"decision_id":"CD-VISUAL-001","decision_kind":"visual_direction","status":"accepted","scope":"series","accepted_value":{"composition":"9:16 纵向构图优先人物关系与关键手部动作","camera_taste":"克制、动机明确，静止镜头可以是主动选择","light_taste":"光源方向服从 Location/View，不为气氛随镜重置"},"target_locators":[{"src":"short-drama","field":"/creator_authority/visual_direction"}],"decided_by":"creator","decided_at":"<ISO-8601>","supersedes_decision_id":null}
+{"example_only":true,"decision_id":"CD-EP001-SCRIPT-001","decision_kind":"artifact_acceptance","artifact_id":"EP001:script","status":"accepted","decided_by":"creator","decided_at":"<ISO-8601>","supersedes_decision_id":null}
+{"example_only":true,"decision_id":"CD-EP001-COVERAGE-AUDITION-SC001","decision_kind":"artifact_acceptance","artifact_id":"EP001:coverage-audition:SC001","status":"accepted","selected_audition_record_id":"AUD-EP001-SC001","selected_approach_id":"APPROACH-A","decided_by":"creator","decided_at":"<ISO-8601>","supersedes_decision_id":null}

package/lib/drama/skills/short-drama/assets/production-observation.example.jsonl

@@ -0,0 +1,2 @@
+{"record_type":"sources","schema_version":"1.0.0","sources":{"prompt-or-spec":{"owner":"short-drama-image-prompts | short-drama-video-prompts | short-drama-storyboard","artifact":"<project-relative-text-artifact>"},"reference-record":{"owner":"<reference-owner>","artifact":"<project-relative-reference-record>"},"short-drama":{"owner":"creator","artifact":"short-drama.json"}}}
+{"example_only":true,"observation_id":"PROD-OBS-001","evidence_state":"active","observation_kind":"generated_result","method":"creator_supplied_description | authorized_visual_inspection","observer":{"owner":"<creator-or-authorized-observer>","independent_from_prompt_owner":false},"prompt_or_spec_refs":[{"src":"prompt-or-spec","record_id":"<stable-record-id>","field":"/<exact-field>"}],"reference_conditions":[{"slot_id":"REF-01","order":1,"artifact_ref":{"src":"reference-record","record_id":"<stable-record-id>"},"role":"identity | style | geography | composition | scale | effect | start_frame | end_frame"}],"production_configuration":{"profile_ref":{"src":"short-drama","field":"/creator_authority/production_profile"},"external_configuration_note":"<private-safe description or unknown; no provider task fields>"},"observations":[{"region_or_interval":"<bounded region or interval>","direct_observation":"<what the authorized observer actually saw/heard>","limitations":["<occlusion/resolution/unseen interval/unknown cause>"]}],"valid_only_for":{"project_id":"<project-id>"},"limitations":["This observation does not prove general model behavior or cross-project quality"],"observed_at":"<ISO-8601>","provenance":"creator_project"}

package/lib/drama/skills/short-drama/assets/project-template/short-drama.json

@@ -0,0 +1,36 @@
+{
+  "schema_version": "1.0.0-draft",
+  "project_id": "LOCAL-PROJECT-ID",
+  "title": "未命名短剧",
+  "language": "zh-CN",
+  "format": {
+    "aspect_ratio": "9:16",
+    "prompt_language": "en",
+    "episode_count": null,
+    "target_seconds_per_episode": null,
+    "pacing": {
+      "spoken_characters_per_second": null,
+      "seconds_per_action_paragraph": null
+    }
+  },
+  "creator_authority": {
+    "decisions_artifact": "创作者决策/",
+    "constraints": [],
+    "visual_direction": {
+      "status": "unset",
+      "choices": {
+        "look_development": null
+      }
+    },
+    "production_profile": {
+      "status": "unset",
+      "choices": {}
+    },
+    "delivery_surface": {
+      "status": "unset",
+      "aspect": null,
+      "overlay_regions": [],
+      "source_ref": null
+    }
+  }
+}

package/lib/drama/skills/short-drama/assets/reference-observation.example.jsonl

@@ -0,0 +1,2 @@
+{"record_type":"sources","schema_version":"1.0.0","sources":{"characters":{"owner":"short-drama-assets","artifact":"设定集/characters.jsonl"}}}
+{"example_only":true,"observation_id":"REF-OBS-001","evidence_state":"active","observation_kind":"input_reference","admission_status":"creator_described | visually_inspected","reference_artifact_ref":{"src":"characters","record_id":"CHAR-EXAMPLE","authority":"candidate"},"method":"creator_supplied_description | authorized_visual_inspection","observer":{"owner":"<creator-or-reference-owner-or-authorized-inspector>","independent_from_prompt_owner":false},"observed_regions":[{"region":"<whole frame or bounded region>","visible_text_or_glyphs":"<none observed or private-safe description>","identity_or_geography_evidence":"<only the visible facts this reference may support>","crop_mask_or_cleanup":"<none or declared treatment>","limitations":["<occlusion/resolution/unseen area>"]}],"text_policy_conclusion":"compatible | conflict | unresolved","unresolved_risks":["<risk kept unresolved when evidence is absent or incomplete>"],"observed_at":"<ISO-8601>","provenance":"creator_project"}

package/lib/drama/skills/short-drama/references/audience-reveal.md

@@ -0,0 +1,74 @@
+# 观众此刻可以知道什么
+
+背影、腿部入画、门缝、画外声、浅景深和遮挡不只是画面风格。它们同时决定观众**何时**
+知道一个人是谁、看见那张字条、或者发现门后有人。
+
+这件事容易被当成构图偏好处理，于是出两种错，而且两种都不会被"这镜好看吗"抓到：
+
+- 本镜本来要交代证据，却被一个漂亮的遮挡挡住了——**该给的没给**；
+- 剧本只允许观众先看到鞋和停步位置，结果正脸提前露出来——**不该给的给了**。
+
+所以每个镜头要按**事实**分开记录可见性，而不是给整镜写一句"稍后揭示"。
+
+## 一条记录只管一个事实
+
+这些记录写在镜头的 `audience_visibility[]` 里（字段属于 storyboard 的 shot 记录）。
+
+同一个镜头里，"这人是谁"和"他手里有什么"是两件独立的事，可以有完全不同的揭示条件。
+揉成一句，下游就只能一起放或一起收——所以 `audience_visibility` **按事实分条**，
+不是每镜一条。
+
+每条记录写清五项：
+
+| 字段 | 回答什么 |
+|---|---|
+| `fact` | 这条管的是哪一个具体事实 |
+| `source_ref` | 这个揭示义务来自剧本的哪一段 |
+| `permission` | 现在就给（`show_now`）还是先扣住（`withhold_now`） |
+| `carrier` | 观众此刻能看见或听见的是什么 |
+| `reveal_trigger` / `protection_method` / `rationale` | 什么时候放、靠什么扣住、为什么 |
+
+### 合成示例
+
+一个镜头里，观众要听见有人叫出主角的旧称呼（说明来者认识他），但**不能**看到来者的脸。
+
+`audience_visibility` 所在的 `shots.jsonl` 在第一行声明本文件引用到的上游快照：
+
+```json
+{"record_type":"sources","schema_version":"1.0.0","sources":{"screenplay":{"owner":"short-drama-write","artifact":"剧集/EP001/screenplay.md"}}}
+```
+
+数组里每条只写快照键和记录 ID：
+
+```json
+[
+  {
+    "fact": "来者认识主角的旧身份",
+    "source_ref": {"src": "screenplay", "record_id": "BLK-EP001-SC003-A02"},
+    "permission": "show_now",
+    "carrier": "画外一声旧称呼，主角闻声停手",
+    "reveal_trigger": "对方开口",
+    "rationale": "本镜必须让观众意识到主角被认出来了"
+  },
+  {
+    "fact": "来者是谁",
+    "source_ref": {"src": "screenplay", "record_id": "BLK-EP001-SC003-A05"},
+    "permission": "withhold_now",
+    "carrier": "只有门框边的一只手和影子",
+    "protection_method": "来者留在景深之外，不进正面光",
+    "reveal_trigger": "下一场推门进屋时",
+    "rationale": "身份揭示是场末的转折，提前露脸会让那一下失效"
+  }
+]
+```
+
+两条各有各的条件：**声音放，脸扣住**。写成一条就做不到这件事。
+
+## 不是"神秘感越多越好"
+
+扣住信息本身没有价值，它只在为后面某个具体时刻服务时才有价值。所以
+`withhold_now` 必须写出 `reveal_trigger`——说不出什么时候放，就说明这不是设计，
+只是没想清楚。
+
+审查时同时核对两侧：原文此刻**允许**透露什么，以及本镜**必须**传达什么。
+只评价构图是否漂亮的审查，抓不到这类缺陷。

package/lib/drama/skills/short-drama/references/contract-and-ownership.md

@@ -0,0 +1,94 @@
+# Contract And Ownership
+
+## Contents
+
+1. Project flow
+2. Runtime path ownership
+3. Output language contract
+4. Stable identities and references
+5. Rule classes
+6. Trust and file safety
+
+## Project flow
+
+```text
+development? -> screenplay -> assets -> image prompts
+                         \-> storyboard -> video prompts
+                         \-> confirmed production -> review -> delivery
+```
+
+This is a routing map, not a mandatory waterfall. Direct entry is valid, and Look Development is optional.
+Each artifact records only the project files it directly read. There is no recursive stale-propagation graph.
+
+## Runtime path ownership
+
+`project_tool.py` does not contain a registry mapping business paths to skill names. The first published artifact
+claims its output paths in project state, and another artifact cannot silently overwrite those paths. Input,
+delivery and machine-state roots remain protected regardless of owner name.
+
+Each skill's own stage contract describes its semantic responsibility. An artifact may project another artifact's
+facts but may not silently redefine them. A reviewer requests changes; the responsible skill edits its source.
+
+## Output language contract
+
+| Field | Governs | Default |
+|---|---|---|
+| `short-drama.json#/language` | creator-facing artifacts, findings, status and Dashboard | `zh-CN` |
+| `short-drama.json#/format/prompt_language` | copyable image/video prompt bodies | `en` |
+
+Creator-facing prose follows `language`. Prompt bodies follow `prompt_language`. Depicted speech and on-screen text
+come from accepted story/asset facts, not from either description-language setting.
+
+## Stable identities and references
+
+Use stable IDs such as `EP001`, `SC001`, `BLK-...`, `CHAR-...`, `LOOK-...`, `LOC-...`, `PROP-...`,
+`SHOT-...`, `KEY-...` and `MOTION-...`. Display names may change without changing identity.
+
+For cross-artifact references, a file declares each upstream snapshot once, and every reference names that
+declaration plus the record it points at.
+
+A `.jsonl` file declares them on its first line:
+
+```json
+{"record_type":"sources","schema_version":"1.0.0","sources":{"screenplay-index":{"owner":"short-drama-write","artifact":"剧集/EP001/screenplay-index.jsonl"}}}
+```
+
+A `.json` file uses a top-level `"sources"` object with the same entries. A reference is then:
+
+```json
+{"src": "screenplay-index", "record_id": "BLK-EP001-SC001-A01", "field": "/text"}
+```
+
+`sources` keys are short, lowercase and derived from the artifact filename; each is unique and stable inside its
+file, and one artifact is declared under one key. `owner` and the project-relative `artifact` identify authority.
+`record_id` and `field` are optional selectors, and a reference pointing at a whole artifact keeps `src` alone.
+`authority` (`accepted` or `candidate`) belongs to the individual reference, because one snapshot can be read at
+different authority levels by different references. Do not copy whole authoritative values into consumers, and do
+not create self-references. A `sources` entry names the artifact the file was written against by `owner` and
+`artifact` only — no digest. Bytes are compared by the tool, in lifecycle state and delivery checksums, and that is not
+user-supplied lifecycle evidence and does not drive transitive invalidation.
+
+## Rule classes
+
+- `structural_invariant`: locally provable; a script may block.
+- `reviewed_invariant`: semantic obligation; a reviewer may request revision.
+- `craft_default`: usually useful, but creator intent may override it.
+- `taste_option`: present alternatives; never block by itself.
+
+Word counts, shot counts, emotional curves and patterns seen elsewhere are not structural invariants unless the
+creator explicitly chose them as format constraints.
+
+## Trust and file safety
+
+Deterministic suite scripts do not make outbound connections. Confirmed production may launch a project-external,
+locally configured adapter without a shell; that adapter owns any provider SDK/API access and reads credentials from
+its environment or system store. Released skills do not retrieve external or private production sources at runtime.
+Do not ship private identifiers, plot passages, prompt sentences, URLs, user data or credentials.
+
+Creator input under `输入/` is immutable to publication commands. Machine state lives under `.short-drama/` and
+delivery under `交付/`; neither is an owner publication target. Text/JSON publication validates all sources before
+writing and replaces each target atomically under one project lock. A crash cannot expose a partially written file;
+rerun the command to continue. Unknown external edits are never silently overwritten.
+
+Legacy v0.3 state remains readable. The next state-changing command rewrites only the compact artifact record; old
+transaction and propagation fields are not carried forward.

package/lib/drama/skills/short-drama/references/creator-workflow.md

@@ -0,0 +1,82 @@
+# Creator Workflow
+
+## Creator authority
+
+`short-drama.json#/creator_authority` stores accepted creator constraints, visual direction and production
+profile choices. A skill may draft choices, but only the creator or an explicitly authorized delegate can accept
+them. If nobody has decided, keep the artifact pending; do not infer acceptance from silence or from a request to
+preview the whole pipeline.
+
+Use project-relative paths and stable record IDs when one artifact points to another. Human-facing records do not
+need lifecycle digests. Deterministic source indexes may still carry byte spans because their job
+is exact source extraction, not workflow state.
+
+## Entry points
+
+| Creator asks for | Owner | Minimum input |
+|---|---|---|
+| Develop an idea or episode map | `short-drama-develop` | brief or conversation |
+| Analyze a long source | `short-drama-novel-analyze` | preserved source file |
+| Write or revise an episode | `short-drama-write` | idea, episode card, outline or script |
+| Extract assets | `short-drama-assets` | current script |
+| Create image prompt specs | `short-drama-image-prompts` | current asset facts |
+| Design shots/keyframes | `short-drama-storyboard` | current script and asset facts |
+| Create video prompt specs | `short-drama-video-prompts` | current shots/keyframes |
+| Produce images, video or TTS | `short-drama-produce` | current bounded job plus explicit confirmation after preview |
+| Review | `short-drama-review` | bounded artifact set |
+| Initialize, continue, show or deliver | `short-drama` | project path or target path |
+
+Direct entry is valid. Do not fabricate upstream files merely to make a nominal pipeline complete.
+
+## Branches
+
+```text
+direction / story development (when needed)
+              |
+           screenplay
+              |
+            assets
+          /        \
+ image prompts   storyboard -> video prompts
+          \        /
+       confirmed production
+                |
+             review -> delivery
+```
+
+Look Development is optional. Image prompts and storyboard are sibling branches after asset facts exist.
+
+## Preview, confirmation and revision
+
+A request for an end-to-end preview authorizes drafting, not acceptance. Its runtime contract is **one bounded work unit per turn**:
+
+1. choose one owner stage; for a high-fanout stage, choose one explicit contiguous batch;
+2. read only that unit's direct inputs, draft or revise it, and run its local structural checks;
+3. persist a candidate without inventing creator acceptance;
+4. report the included scope, remaining scope and next useful action, then **return control to the creator**.
+
+The turn ends at step 4. A creator message such as “continue” authorizes the next bounded unit. The next owner
+stage, review and production each begin on their own explicit request; review also runs when an explicitly
+requested delivery needs a verdict. Undecided work stays in `needs_confirmation`, and delivery waits until the
+creator accepts the relevant outputs.
+
+For a revision:
+
+1. identify the artifact and owner;
+2. read only its direct current inputs;
+3. show the semantic change and what should remain unchanged;
+4. let the owner publish the revision;
+5. request creator acceptance when canonical meaning changed;
+6. report whether a separate bounded review is still needed; do not start it automatically.
+
+Changing an upstream file does not recursively rewrite project state. A consumer that directly recorded that file
+shows `update_needed`; republish the consumer when the creator actually wants it refreshed.
+
+Rendered prompt Markdown is a view of its structured spec. If someone edits the rendered view, either regenerate
+it, adopt the change into the spec, or reconcile both explicitly.
+
+## Delivery boundary
+
+Deliver only explicitly selected, current, approved text and JSON in the text package. Produced media remains in
+its project production directory and may be exported by a separate creator-approved media handoff. Exclude private
+sources, credentials, absolute paths and machine state. `verify` checks checksums; it does not judge creative quality.

package/lib/drama/skills/short-drama/references/form-cards/Q/347/211/210/350/241/250/350/276/276.md

@@ -0,0 +1,73 @@
+# 形态卡：Q 版表达（含 Q 版知识与教育短片）
+
+上层选择、构成差异与跨阶段责任见 [production-form-profiles.md](../production-form-profiles.md)。
+本卡回答：选了这个形态之后，可执行文本本身变成什么样。示例均为合成改写。
+
+## 叙事职责
+
+用**降低威胁感和降低认知负担**换取信息通过率：夸张比例让情绪一眼可读，简化形体让因果链
+只剩必要元素。它擅长把规则、流程、风险和抽象概念变成可看的因果单位，也擅长让沉重题材
+被接受。它不擅长细节写实、精确尺度和严肃的暴力后果；需要这些时改由旁白或文字承担，
+不要靠画面硬撑。
+
+## 身份锚点载体
+
+**比例体系 + 大色块 + 一个被放大的记号**。头身比一旦选定就是硬约束：不同角色可以有不同
+头身比，但同一角色跨镜不能变。识别靠发型剪影、主色块和一处刻意放大的特征
+（眼镜、呆毛、围裙、工牌）。五官细节不做锚点——它在简化体系里承载不了识别。
+
+## 连续性载体
+
+- **必须传递**：头身比与角色之间的相对尺度、主色块、放大记号、当前手中物、
+  可读文字与后期文字的预留位置。
+- **可以省**：写实材质、精确光向、背景细节。
+- **最常断裂处**：角色间尺度关系逐镜变化、同一道具时大时小、放大记号在某些角度消失。
+
+## 层级拆分
+
+- 身份层：比例体系、色块、记号；
+- 环境层：简化背景，只保留承担因果的功能物；
+- 可动层：本镜承担因果的那一个动作或那一个物体；
+- 效果层：符号层（问号、汗滴、箭头、高亮圈、放大镜）与后期文字预留区。
+
+符号层是这个形态特有的可执行载体：它可以直接承担注意力指向与因果标注，但必须先声明
+是画内元素还是后期层，而且**不能替代事实本身**——箭头指向不等于展示了后果。
+
+## 光与材质词汇
+
+可执行词汇是**平涂 + 两段明暗 + 高饱和主色**，加少量指示性高光。「危险物用画面里唯一的
+高饱和红，其余降饱和」可执行；「可爱风格」「明快色调」不可执行。材质通常只需分三类——
+软（布、皮肤）、硬（金属、塑料）、透（玻璃、液体）——够用即可，再细分不会被画出来。
+
+## 运动预算
+
+一个画面或一个运动段**只承载一个因果单位**。默认 hold 加局部动：动的是执行因果的那只手
+或那个物体，其余保持。可爱不等于持续弹跳——给每个元素都加弹性循环会让因果读不出来。
+转场可以用符号层或尺度变化（推到局部）承接。危险动作与错误示范是否允许被完整演示，
+属于形态卡的安全边界决定，不由提示词阶段临时判断。
+
+## 声音与文字职责
+
+旁白通常是主要信息载体，画面负责证据与后果，两者分工要写死：**旁白说规则，画面说后果**，
+画面不复述旁白已经说完的内容。可读文字（标签、数字、步骤号）在这个形态里高频出现，
+必须走资产 text policy 并明确是画内还是后期层；后期文字要在构图里预留位置，
+不能事后压在脸或记号上。事实准确性不因为风格可爱而下调。
+
+## 跨阶段传递
+
+| 阶段 | 这个形态额外要写的 |
+|---|---|
+| develop | 头身比体系、旁白与画面的分工、符号层策略、安全边界 |
+| assets | 比例体系、色块、记号、符号层清单、文字政策 |
+| image-prompts | 单帧的比例、色块、符号位置与文字预留区 |
+| storyboard | 一个画面一个因果单位、注意力指向、后期文字预留 |
+| video-prompts | 会动的那一层与其因果、符号层出现时机、旁白相对动作的位置 |
+| review | 因果单位是否被稀释、尺度关系是否稳定、文字预留是否被占用 |
+
+## 常见失效
+
+- 全员持续弹跳，因果读不出来；
+- 一镜塞进多个因果单位，观众只记住最后一个；
+- 符号层替代了事实，箭头指向被当成已经展示；
+- 后期文字压住脸或放大记号；
+- 为了可爱牺牲事实准确性或安全边界。

package/lib/drama/skills/short-drama/references/form-cards//344/272/214/347/273/264/345/212/250/346/200/201/346/274/253.md

@@ -0,0 +1,74 @@
+# 形态卡：二维动态漫（限动画 / 绘制动画）
+
+上层选择、构成差异与跨阶段责任见 [production-form-profiles.md](../production-form-profiles.md)。
+本卡回答：选了这个形态之后，可执行文本本身变成什么样。示例均为合成改写。
+
+## 叙事职责
+
+用**明确的形状与线条**把复杂信息压缩成一眼可读的画面：谁是谁、谁在看谁、哪件东西重要，
+由轮廓和留白直接分配注意力。它擅长用很少的全动作推进大量剧情——省下的动作预算换成
+信息密度；也擅长表现被夸张过的力度与情绪。它不擅长连续的复杂三维调度和真实材质的
+细微变化；需要这些时先问能不能改成一次姿势切换或一次效果层承接。
+
+## 身份锚点载体
+
+**轮廓剪影 + 线条特征 + 固定色块**。跨镜可辨认靠发型剪影、脸型比例、服装的色块划分，
+以及一两个不可省的图形记号（领口形状、发饰、袖口条纹）。不要拿「眼神」或写实五官
+细节做锚点——绘制形态每个角度都会重画，只有形状级特征稳定。剪影测试是这个形态特有的
+自检：把角色填成纯黑，是否还认得出是谁。
+
+## 连续性载体
+
+- **必须传递**：轮廓与色块、阴影分区（阴影落在哪一块）、线条粗细与是否描边、
+  本造型相对基准的色卡差异、本镜可动层清单。
+- **可以省**：写实材质细节、精确布光角度、背景纵深。
+- **最常断裂处**：阴影分区换了逻辑、描边时有时无、同一角色跨镜配色饱和度漂移。
+
+## 层级拆分
+
+这是本形态最核心的可执行结构，每镜都要显式写：
+
+- 身份层：角色线稿与固定色块，跨镜不重画；
+- 环境层：背景板，通常整段复用，只在机位改变时更换；
+- 可动层：本镜真正会动的部分（一只手、嘴、发梢、衣摆、一件道具）；
+- 效果层：光斑、粒子、速度线、冲击波、光晕，与角色层分开声明谁负责。
+
+每镜点名：哪层 hold、哪层局部循环、哪层做视差、哪层用姿势切换。
+
+## 光与材质词汇
+
+可执行词汇是**色块关系**而不是物理光：主色 / 阴影色 / 高光色三段怎么分、阴影边界是硬边
+还是渐变、有没有反射光块、描边颜色是否随环境变化。材质写绘制方式：平涂、笔刷纹理、
+分层上色、渗色。不写次表面散射、物理材质这类三维词，也不写「高质量插画」。
+
+## 运动预算
+
+默认 **limited motion**，静止是主动的注意力选择而不是缺陷。分配原则：一镜只让承担因果的
+那一层全动；其余交给 hold pose、局部循环（呼吸、发梢、衣摆、闪烁）、背景视差或效果层。
+转折瞬间可以用姿势切换代替连续动作，冲击可以用效果层加顿帧承接，位移可以由剪辑承接。
+不要给所有角色、头发、衣摆和背景同时加无因运动——那是这个形态最典型的廉价感来源。
+
+## 声音与文字职责
+
+口型策略（不做口型 / 有限口型开合 / 完整口型）要在形态卡里先决定并跨集一致；决定之后，
+台词的写法随之改变——不做口型时长台词优先改成画外或背身。画面文字（拟声字、心声框、
+标题字）是本形态的合法叙事载体，但必须走资产 text policy，并声明是画内元素还是后期层。
+
+## 跨阶段传递
+
+| 阶段 | 这个形态额外要写的 |
+|---|---|
+| develop | 限动比例、口型策略、是否使用画面文字与效果层承接 |
+| assets | 轮廓、色块、阴影分区、描边规则、可动层清单、表情与姿势库 |
+| image-prompts | 单帧的线条与色块投影、层次关系、背景密度 |
+| storyboard | 注意力分配、遮挡、本镜哪一层冻结、可动层边界 |
+| video-prompts | 只写会动的层与循环方式、姿势切换点、效果层时机、口型策略 |
+| review | 每镜是否点名可动层，是否出现无因全局运动 |
+
+## 常见失效
+
+- 全部图层同时轻微飘动，观众读不出谁在动；
+- 阴影分区逐镜换逻辑，或描边时有时无；
+- 把三维光影词写进二维说明，导致质感撕裂；
+- 效果层盖掉身份锚点，剪影认不出人；
+- 背景密度过高抢走注意力，人物退成贴纸。