0) 结论先行：Medeo 的“现象级”来自工程化，而不只是模型

1) 产品形态与交互范式：Chat + Timeline + Generator Recipes

2) 三层架构：DSL / Context / Environment

┌─────────────────────────────── Client ───────────────────────────────┐
│  Chat UI  │  Timeline UI  │  Asset Browser  │  Export / Preview       │
└───────────┬───────────────┬────────────────┬──────────────────────────┘
            │               │                │
            ▼               ▼                ▼
┌────────────────────────── API Gateway / BFF ─────────────────────────┐
│ Auth / RateLimit / Project CRUD / WS Updates / Upload / Billing       │
└───────────────────────────┬──────────────────────────────────────────┘
                            ▼
┌──────────────────────── Planning Layer ──────────────────────────────┐
│ LLM Router → Context Builder → DSL Planner → Validator → Cost Estimator│
└───────────────────────────┬──────────────────────────────────────────┘
                            ▼
┌──────────────────────── Orchestration Layer ─────────────────────────┐
│ DAG Builder → Priority Queue → Worker Dispatch → Retries/Cancel       │
└───────────────────────────┬──────────────────────────────────────────┘
                            ▼
┌──────────────────────── Execution Layer (Workers) ────────────────────┐
│ Stock Search │ TTS │ Music Mix │ ASR/Caption │ Gen Image/Video │ Render│
└───────────────────────────┬──────────────────────────────────────────┘
                            ▼
┌──────────────────────── Data Layer ───────────────────────────────────┐
│ Project Store │ Asset Store │ Job Store │ Cache │ Metrics/Logs/Traces  │
└───────────────────────────────────────────────────────────────────────┘

3) DSL：视频工程中间表示（IR）

目标：让“对话式修改”变成可验证、可回滚、可计费的工程操作。

3.1 DSL 的设计原则（强烈建议）

3.2 最小 DSL 结构（示例）

{
  "project_id": "p_123",
  "base_revision": 42,
  "ops": [
    {
      "op_id": "op_9f31",
      "type": "trim_clip",
      "target": { "clip_id": "clip_7" },
      "params": { "in": 1.20, "out": 4.10 },
      "cost_hint": { "unit": "cpu", "weight": 0.2 }
    },
    {
      "op_id": "op_aa02",
      "type": "replace_voiceover",
      "target": { "audio_id": "aud_vo_2" },
      "params": {
        "text": "更新后的旁白文案…",
        "voice_id": "voice_zh_male_01",
        "pace": 1.05
      },
      "cost_hint": { "unit": "tts_chars", "weight": 530 }
    }
  ],
  "intent": {
    "user_utterance": "把第 12 秒那句旁白更简洁，并把那段节奏加快一点"
  }
}

3.3 原子操作（建议清单）

4) Context Engineering：把“对话 + 工程状态 + 约束”变成可控上下文

4.1 Context 的分层（推荐）

4.2 意图到 patch 的标准流水线

1) Locate: 识别用户说的“那段/第二个观点/12秒附近”对应的 Scene/Clip IDs
2) Plan: 生成候选 edits（可能多方案：A/B），但最终输出一个可执行 patch
3) Validate: schema/type/range/dependency 校验 + 约束（品牌、预算、分辨率）
4) Estimate: 估算 credits + 预计耗时 + 是否触发重渲染
5) Execute: 进入 DAG + Queue
6) Summarize back: 用人话解释“我做了什么”，同时把 patch 写入版本历史

5) 编排器与任务系统：DAG + 优先级队列 + 可取消

5.1 Job 状态机（建议）

QUEUED → PLANNED → RUNNING → (SUCCEEDED | FAILED | CANCELED)
                    ↑
                 RETRYING

5.2 DAG 示例：替换旁白 + 重新对齐字幕 + 局部重渲染

[TTS Generate] ─┬─> [Audio Mix & Loudness] ─┬─> [Caption Align] ─┬─> [Preview Render]
                │                          │                   │
                └─> [Phoneme/Timing]       └─> [Ducking BGM]    └─> [Final Render (optional)]

6) 渲染与“实时感”：Preview / Final + 增量渲染

6.1 两阶段渲染（强推荐）

6.2 增量渲染：只重算受影响的 segments

Segment = hash(clip_ids + timing + effects + assets + style + resolution + fps)

DirtyRange =
  union(
    affected_clips_time_range,
    neighboring_transitions,
    audio_mix_window
  )

Render strategy:
  - reuse clean segments from cache
  - render dirty segments
  - stitch segments (concat) for preview/final

7) 素材系统与角色一致性：资产是第一等公民

7.1 资产系统（Asset Store）需要存什么

7.2 角色一致性（Character Consistency）常用工程组合

8) Credits：成本模型、估算、扣费、取消与对账

8.1 成本模型（建议从一开始就做）

estimated_credits =
  Σ op_i.cost(
    model, resolution, duration, frames, retries, priority, caching_hit
  )

典型计费维度：
- VideoGen: frames * model_rate
- Render: pixels * duration * fps
- TTS: characters * voice_rate
- ASR/Align: audio_seconds * rate
- Stock search: usually cheap (can be bundled)

8.2 扣费策略（用户体验与风控兼顾）

9) 可观测性与质量控制：让“生成系统”可运营

9.1 必备观测指标（建议最低配）

9.2 “生成质量”如何工程化（可执行）

10) 复刻路线图：从 MVP 到“现象级”护城河

10.1 MVP（4 个模块先跑通）

10.2 护城河（你要追上“现象级”的关键）

附录：数据模型 / 协议示例（可直接改造成你自己的系统）

A) Project（工程文件）最小 schema（示例）

{
  "project_id": "p_123",
  "revision": 42,
  "meta": {
    "title": "科普短视频",
    "aspect": "9:16",
    "fps": 30,
    "style": { "caption": "bold", "theme": "dark" }
  },
  "script": {
    "outline": ["痛点", "原因", "解决方案", "总结"],
    "full_text": "..."
  },
  "scenes": [
    { "scene_id": "s1", "range": [0.0, 6.5], "summary": "开场钩子", "bindings": { "char_id": "c_7" } },
    { "scene_id": "s2", "range": [6.5, 18.0], "summary": "解释原因", "bindings": { "char_id": "c_7" } }
  ],
  "timeline": {
    "tracks": [
      { "track_id": "t_v1", "kind": "video", "clips": ["clip_1","clip_2","clip_7"] },
      { "track_id": "t_a1", "kind": "audio", "clips": ["aud_vo_2","aud_bgm_1"] },
      { "track_id": "t_c1", "kind": "caption", "clips": ["cap_1","cap_2"] }
    ],
    "objects": {
      "clip_7": { "type":"video_clip","asset_id":"as_stock_91","in":0.0,"out":3.1,"start":11.2,"speed":1.0 },
      "aud_vo_2": { "type":"audio_clip","asset_id":"as_tts_33","start":0.0,"gain_db":-2.0 },
      "cap_2": { "type":"caption_seg","start":11.2,"end":14.0,"text":"..." }
    }
  },
  "history": [
    { "rev": 41, "ops": ["op_aa02"], "by":"assistant", "ts":"2026-01-05T10:12:11Z" }
  ]
}

B) Worker 输出规范（artifact + usage_report）

{
  "job_id": "job_88",
  "step_id": "step_tts_1",
  "status": "SUCCEEDED",
  "artifacts": [
    { "artifact_id":"as_tts_33", "type":"audio", "uri":"s3://.../as_tts_33.wav", "hash":"..." }
  ],
  "usage_report": {
    "provider": "tts_vendor_x",
    "model": "zh-male-v1",
    "characters": 530,
    "seconds": 12.4,
    "cache_hit": false,
    "cost_units": { "credits": 14.2 }
  }
}

C) Patch 应用伪代码（幂等 + 版本）

// applyPatch(project, patch):
if patch.base_revision != project.revision:
  return CONFLICT

for op in patch.ops:
  if ledger.has(op.op_id): continue // 幂等：已执行过
  validate(op, project)
  estimateAndReserve(op)
  project = applyOpToAST(project, op)
  ledger.record(op.op_id, op.cost_estimate)

project.revision += 1
store(project)
emitWSUpdate(project_id, revision)