ai-agent-book 精选快照(<2MB 代码与文档,来自 github.com/bojieli/ai-agent-book)
Build latest book artifacts / build (push) Canceled after 0s
dependency resolution / resolve (3.11) (push) Canceled after 0s
dependency resolution / resolve (3.13) (push) Canceled after 0s
deploy-pages / build (push) Canceled after 0s
deploy-pages / deploy (push) Canceled after 0s
i18n consistency check / check (push) Canceled after 0s
provider adoption tests / test (chapter2/context-compression) (push) Canceled after 0s
provider adoption tests / test (chapter2/prompt-injection) (push) Canceled after 0s
provider adoption tests / test (chapter2/system-hint) (push) Canceled after 0s
provider adoption tests / test (chapter3/log-sanitization) (push) Canceled after 0s
web-search-agent tests / test (push) Canceled after 0s
web-search-agent tests / agentbook (push) Canceled after 0s

This commit is contained in:
2026-08-20 13:12:50 +00:00
commit b119135836
10275 changed files with 3284984 additions and 0 deletions
+457
View File
@@ -0,0 +1,457 @@
# Experiment 5-6: API-Driven Smart Video Editing / 实验 5-6:基于 API 的智能视频剪辑
> Companion lab for *AI Agents in Depth*, Chapter 5 — NL request + multi-scene video → two-step Vision locate → Blender bpy script / ffmpeg cut → ProposerReviewer.
> 《深入理解 AI Agent》配套:自然语言需求 + 多场景视频 → 两步 Vision 定位 → 生成 Blender bpy / ffmpeg 剪辑 → 提议者-审核者。
← [Chapter 5 index / 返回第 5 章目录](../README.md)
---
## English
### Purpose
Given a multi-scene video and one NL request (e.g. “cut out the surfing part”), the Agent locates the target scene, **generates a Blender Python API script**, cuts the clip, and self-reviews.
Three mechanisms in multimedia processing:
1. **Two-step Vision locate**: Proposer cannot “watch” video directly; a **video-analysis sub-agent** uses ffmpeg frame extraction + Vision LLM to find time bounds.
2. **Code generation (Blender Python API)**: Proposer turns the edit plan into a **bpy** script—import / trim / subtitle / speed / render as API calls, run with `blender --background --python edit.py`. Without Blender, scripts still generate; render falls back to ffmpeg (see “edit backends”).
3. **Proposer / Reviewer**: after cut, Reviewer samples key frames with Vision; fail → feedback → iterate.
### Two-step locate
Scanning every frame is slow/expensive → coarse then fine:
- **Coarse**: one frame every **10s**; Vision gets a rough interval (e.g. surfing 2030s).
- **Fine**: expand the coarse window by one coarse step; one frame every **1s**; Vision returns precise bounds (e.g. 1529s).
Encapsulated as a **sub-agent**: tens of screenshots live only in the sub-agents one-shot context and do not pollute Proposer/Reviewer history. Demo prints token comparison at the end.
### ProposerReviewer flow
```
NL request ──► Proposer parses intent (scene + effects)
Video-analysis sub-agent two-step locate ──► [start, end]
Proposer emits Blender bpy script (edit.py) ──► render cut (+ subtitle/slow-mo)
│ Blender if installed else ffmpeg
Reviewer samples start/mid/end frames ──► Vision pass/fail + feedback
│pass? no → Proposer adjusts bounds, re-cut (max 3 rounds)
▼yes
final.mp4
```
### Run
```bash
# From the repository root: use the shared Chapter 5 environment
uv sync --locked --python 3.12 --extra ch5
# Activate it before changing directories:
# macOS/Linux:
source .venv/bin/activate
# Windows PowerShell: .\.venv\Scripts\Activate.ps1
# Windows cmd: .venv\Scripts\activate.bat
# pip fallback when uv is not installed:
# python -m pip install -e ".[ch5]"
cd chapter5/video-edit
# Single-project compatibility path, still supported during migration:
# python -m pip install -r requirements.txt
cp env.example .env # OPENAI_API_KEY (or OPENROUTER_API_KEY fallback)
python demo.py # default: "把冲浪的部分剪出来" (full pipeline)
python demo.py "把滑雪部分剪出来,并加上字幕 Winter" # custom request
python demo.py -i my.mp4 -o out.mp4 "把演讲开场剪出来" # own video + output
python demo.py --backend blender # force Blender headless (Blender required)
python demo.py --vision-model gpt-5.6-luna # also --text-model
python demo.py --quick # coarse sampling + single review round (cheapest Vision)
python demo.py --smoke # smoke: edit path + bpy script only; no API
python demo.py --help
```
Common flags (see `--help`): `--input/-i`, `--output/-o`, `--backend {auto,blender,ffmpeg}`, `--text-model` / `--vision-model`.
One command runs: generate/read video → two-step locate → bpy cut → review → final. Each run clears `output/` (idempotent). Full path calls Vision many times; use `--smoke` (zero API) or `--quick` first.
### Sample outputs
#### `--smoke` (zero API, reproducible)
Real output of `python demo.py --smoke` (no OpenAI key; needs ffmpeg):
```text
==========================================================================
冒烟自检 | 剪辑链路 + bpy 脚本生成,不调用任何 API
==========================================================================
[1/3] 生成测试视频 OKoutput/source.mp4(场景真值={'hiking': (0, 15), 'surfing': (15, 30), 'skiing': (30, 42), 'cycling': (42, 54)}
[2/3] 抽帧 OKoutput/frames/smoke.png
[3/3] 剪辑+字幕 OK(后端=ffmpeg(未装 Blender,回退)):
文件: smoke_cut.mp4
时长: 5.03s
容器: mov,mp4,m4a,3gp,3g2,mj2
大小: 121.4 KB
视频流: h264 1280x720 @ 30/1 fps
音频流: aac 44100Hz 1ch
已生成 Proposer 的 Blender 脚本:output/edit.py
(这正是书中'生成 Blender Python API 代码'的产物;装好 Blender 后可直接
`blender --background --python output/edit.py` 无头渲染。)
✓ 冒烟自检通过:剪辑链路正常 + bpy 脚本已生成(未调用 OpenAI)。
```
`output/edit.py` is an executable Blender bpy script (`new_movie`, `frame_offset_start` / `frame_final_duration`, `new_effect(type='TEXT')`, `bpy.ops.render.render`); syntax-checked with `py_compile`.
#### `--quick` (full path, needs API)
Excerpt from `python demo.py --quick` (default surfing request); locate/error/token parts independent of bpy/ffmpeg:
```text
步骤 1 | Proposer 解析自然语言需求
解析结果:目标场景='surfing scene' 特效=[]
步骤 2 | 视频分析子 Agent:两步 Vision 定位(--quick 快速采样)
[粗粒度] 每 15s 采样 5 帧 → Vision 得区间 [15, 30]s(依据:The word 'SURFING' appears at t=15s and changes at t=30s.
[细粒度] 窗口 [0.0, 45.0] 内每 2s 采样 23 帧 → 精确边界 [16.0, 28.0]s
>>> 最终定位:起 16.0s 止 28.0s
真值 [15, 30]s → 起点误差 1.0s,终点误差 2.0s(验收要求 ≤ 3s)
步骤 3-4 | Proposer 剪辑 + Reviewer 审查(迭代)
Proposer 剪出片段 [16.0, 28.0]s,成片时长 12.0s
Reviewerpass=... score=... 检查帧=['0.5', '6.0', '11.5']
Token 统计(子 Agent 隔离截图,主上下文不被污染)
主 AgentProposer+Reviewer):573 tokens
子 Agent(两步定位截图) 2934 tokens
```
Artifacts under `output/`:
| File | Duration | Note |
| --- | --- | --- |
| `source.mp4` | 54.0s | Procedural 4-scene test source |
| `edit_round1.py` | — | Proposer bpy script (portable) |
| `cut_round1.mp4` | 12.0s | Round-1 candidate |
| `final.mp4` | 12.0s | Chosen final (H.264 + AAC, 1280x720@30fps) |
Token stats: tens of screenshots (2934 tok) only in **sub-agent**; main history (~573 tok) almost unpolluted. (Synthetic test shows “SURFING” text, not real surfing—Reviewer may fail on content; real video avoids that.)
### Dependencies
- **ffmpeg / ffprobe** for cuts and frames. `brew install ffmpeg` (macOS) / `apt install ffmpeg` (Ubuntu). Validated on ffmpeg 8.0.
- **OPENAI_API_KEY**: `gpt-5.6-luna` for vision locate/review and text planning (vision model must accept images); or `OPENROUTER_API_KEY` fallback via OpenRouter.
### Adapt / extend
#### Model / provider
All via env (`env.example`); no code change:
- `TEXT_MODEL` (default `gpt-5.6-luna`).
- `VISION_MODEL` must support images (default `gpt-5.6-luna`).
- `OPENAI_BASE_URL` + matching `OPENAI_API_KEY` for compatible endpoints.
```bash
export OPENAI_BASE_URL=https://your-gateway.example.com/v1
export VISION_MODEL=gpt-5.6-luna
export TEXT_MODEL=gpt-5.6-luna
```
`agents.py` `OpenAI()` client reads these lazily via `client()`.
#### Own input video
`make_test_video.py` builds a 54s video with 4 distinct scenes (HIKING green / SURFING blue / SKIING white / CYCLING orange) plus large scene-name and timecode watermarks for Vision reproducibility.
Own video: `python demo.py -i your.mp4 -o out.mp4 "edit request"` (skips test generation; no ground-truth error print).
#### Blender vs ffmpeg (edit backends)
Book uses **Blender Python API (bpy)** on the VSE. First-class here: `blender_editor.generate_bpy_script()` emits real bpy (`new_movie`, frame offsets, `TEXT`/`SPEED` effects, render); `render_with_blender()` runs `blender --background --python edit.py`.
- `--backend blender`: force Blender (`blender --version` required);
- `--backend ffmpeg`: force ffmpeg;
- `--backend auto` (default): bpy if Blender installed, else ffmpeg.
**Always**: Proposers bpy script is written to `output/edit_round*.py` (`output/edit.py` under `--smoke`)—core “generate Blender Python API code” artifact. This environment may not have Blender, so render is validated with ffmpeg; bpy is `py_compile`-checked but **not rendered in real Blender** unless you install it and use `--backend blender`.
| | ffmpeg | Blender (bpy) |
| --- | --- | --- |
| Fit | 2D cut/join/subtitle/speed pipelines | 3D scenes, compositing, keyframe anim, particles/camera |
| Ops | single binary, no GUI, CI-friendly | full Blender install; larger/slower |
| When | most “cut a clip + simple effects” | only when 3D/compositing/complex transitions needed |
Two-step Vision + ProposerReviewer is decoupled from the execution layer; same edit plan for both backends.
### Files
| File | Role |
| --- | --- |
| `demo.py` | CLI orchestration, self-check, iteration, token stats |
| `agents.py` | `VideoAnalyzerAgent` / `ProposerAgent` / `ReviewerAgent` |
| `blender_editor.py` | **bpy script gen + headless render** (book path) |
| `video_editor.py` | `apply_edit()`; Blender/ffmpeg backends |
| `make_test_video.py` | Procedural 4-scene test video |
| `ffmpeg_utils.py` | Thin ffmpeg/ffprobe wrappers |
`output/` is gitignored.
### Limitations
- Locate accuracy depends on visual distinctness; gradual scene transitions increase boundary error vs solid-color test film.
- Fine step fixed at 1s → boundary precision about ±1s (book acceptance ±3s).
- Slow-mo audio uses `atempo`; extreme ratios hurt quality; complex transitions/multi-track not covered.
- Reviewer samples only start/mid/end; mid-clip glitches can be missed (raise sample density if needed).
---
## 中文
### 目的
用户给一段含多个场景的视频 + 一句自然语言需求(如"把冲浪部分剪出来"),Agent 自动定位目标场景、**生成 Blender Python API 脚本**剪出片段并自我审查。
验证三个核心机制在多媒体处理中的作用:
1. **两步 Vision 定位**Proposer 无法直接"看懂"视频,于是委托一个**视频分析子 Agent**,
用 ffmpeg 抽帧 + Vision LLM 读图来定位目标场景的时间边界。
2. **代码生成(Blender Python API**:Proposer 把剪辑计划翻译成一段调用 **Blender
Python APIbpy** 的脚本——导入 / 裁剪 / 字幕 / 变速 / 渲染各对应一个 API 调用,
`blender --background --python edit.py` 无头执行。这正是书中"把视频编辑重构为
API 调用和代码生成问题"的落地。未装 Blender 时脚本照常生成(代码生成产物),
实际渲染回退到 ffmpeg(见下文"剪辑后端")。
3. **提议者-审核者(Proposer / Reviewer**:Proposer 剪辑后无法自证效果,
由 Reviewer 抽取成片关键帧、用 Vision LLM 检查是否剪对,不合格则反馈、迭代。
### 两步定位原理
Vision LLM 逐帧扫全片既慢又贵,因此采用"先粗后细":
- **第一步(粗粒度)**:每 **10 秒**抽一帧,把全片的稀疏截图连同"要找哪个场景"一起
交给 Vision,得到大致区间(如"冲浪在 2030s")。
- **第二步(细粒度)**:在粗区间上下各外扩一个粗间隔,每 **1 秒**抽一帧,
再问 Vision 精确边界(如"1529s")。
把这套抽帧-读图封装成**独立子 Agent**:几十张截图只进入子 Agent 的一次性上下文,
不会污染主 AgentProposer/Reviewer)的对话历史。demo 末尾会打印两者的 token 对比。
### 提议者-审核者
```
NL 需求 ──► Proposer 解析意图(目标场景 + 特效)
视频分析子 Agent 两步定位 ──► [start, end]
Proposer 生成 Blender bpy 脚本(edit.py)──► 渲染剪辑(可加字幕/慢动作)
│ 装了 Blender 用 bpy,否则回退 ffmpeg
Reviewer 抽首/中/尾关键帧 ──► Vision 检查 pass/fail + 反馈
│pass? 否 → Proposer 据反馈修正边界,重剪(最多 3 轮)
▼是
输出成片 final.mp4
```
### 运行
```bash
# 在仓库根目录使用统一的第 5 章环境
uv sync --locked --python 3.12 --extra ch5
# 切换目录前先激活环境:
# macOS/Linux
source .venv/bin/activate
# Windows PowerShell.\.venv\Scripts\Activate.ps1
# Windows cmd.venv\Scripts\activate.bat
# 未安装 uv 时可用 pip 兜底:
# python -m pip install -e ".[ch5]"
cd chapter5/video-edit
# 迁移期间仍支持单项目兼容路径:
# python -m pip install -r requirements.txt
cp env.example .env # 填入 OPENAI_API_KEY(未配置时设 OPENROUTER_API_KEY 自动改走 OpenRouter
python demo.py # 默认需求"把冲浪的部分剪出来"(完整流程)
python demo.py "把滑雪部分剪出来,并加上字幕 Winter" # 自定义需求
python demo.py -i my.mp4 -o out.mp4 "把演讲开场剪出来" # 用自己的视频 + 自定义输出
python demo.py --backend blender # 强制用 Blender Python API 无头渲染(需装 Blender
python demo.py --vision-model gpt-5.6-luna # 覆盖模型(也可用 --text-model
python demo.py --quick # 快速模式:粗采样 + 单轮审查,Vision 调用最少(省时省钱)
python demo.py --smoke # 冒烟自检:仅剪辑链路 + 生成 bpy 脚本,不调用任何 API
python demo.py --help # 查看全部参数
```
常用参数(完整见 `--help`):`--input/-i` 输入视频、`--output/-o` 成片路径、
`--backend {auto,blender,ffmpeg}` 剪辑后端、`--text-model`/`--vision-model` 覆盖模型。
一条命令即可跑通:生成/读取视频 → 两步定位 → 生成 bpy 脚本剪辑 → 审查 → 输出成片。
每次运行都会清空 `output/`,从干净状态开始(幂等可重复)。完整流程会多次调用
Vision 模型(较慢/耗费额度);只想验证链路时先跑 `--smoke`(零 API),或用 `--quick`
### 预期输出示例
#### `--smoke`(零 API,可复现)
以下为 `python demo.py --smoke` 的**真实输出**(无需 OpenAI Key,仅需 ffmpeg):
```text
==========================================================================
冒烟自检 | 剪辑链路 + bpy 脚本生成,不调用任何 API
==========================================================================
[1/3] 生成测试视频 OKoutput/source.mp4(场景真值={'hiking': (0, 15), 'surfing': (15, 30), 'skiing': (30, 42), 'cycling': (42, 54)}
[2/3] 抽帧 OKoutput/frames/smoke.png
[3/3] 剪辑+字幕 OK(后端=ffmpeg(未装 Blender,回退)):
文件: smoke_cut.mp4
时长: 5.03s
容器: mov,mp4,m4a,3gp,3g2,mj2
大小: 121.4 KB
视频流: h264 1280x720 @ 30/1 fps
音频流: aac 44100Hz 1ch
已生成 Proposer 的 Blender 脚本:output/edit.py
(这正是书中'生成 Blender Python API 代码'的产物;装好 Blender 后可直接
`blender --background --python output/edit.py` 无头渲染。)
✓ 冒烟自检通过:剪辑链路正常 + bpy 脚本已生成(未调用 OpenAI)。
```
生成的 `output/edit.py` 是一段**可执行的 Blender bpy 脚本**`new_movie` 导入、
`frame_offset_start`/`frame_final_duration` 裁剪、`new_effect(type='TEXT')` 字幕、
`bpy.ops.render.render` 渲染),本机对其做过 `py_compile` 语法校验。
#### `--quick`(完整链路,需 API
以下为 `python demo.py --quick`(默认需求"把冲浪的部分剪出来")的真实节选(定位/误差/
token 部分与剪辑后端无关,故不受 bpy/ffmpeg 后端切换影响):
```text
步骤 1 | Proposer 解析自然语言需求
解析结果:目标场景='surfing scene' 特效=[]
步骤 2 | 视频分析子 Agent:两步 Vision 定位(--quick 快速采样)
[粗粒度] 每 15s 采样 5 帧 → Vision 得区间 [15, 30]s(依据:The word 'SURFING' appears at t=15s and changes at t=30s.
[细粒度] 窗口 [0.0, 45.0] 内每 2s 采样 23 帧 → 精确边界 [16.0, 28.0]s
>>> 最终定位:起 16.0s 止 28.0s
真值 [15, 30]s → 起点误差 1.0s,终点误差 2.0s(验收要求 ≤ 3s)
步骤 3-4 | Proposer 剪辑 + Reviewer 审查(迭代)
Proposer 剪出片段 [16.0, 28.0]s,成片时长 12.0s
Reviewerpass=... score=... 检查帧=['0.5', '6.0', '11.5']
Token 统计(子 Agent 隔离截图,主上下文不被污染)
主 AgentProposer+Reviewer):573 tokens
子 Agent(两步定位截图) 2934 tokens
```
产物(`output/` 目录,真实文件):
| 文件 | 时长 | 说明 |
| --- | --- | --- |
| `source.mp4` | 54.0s | 程序化生成的 4 场景测试原片 |
| `edit_round1.py` | — | Proposer 生成的 Blender bpy 脚本(代码生成产物,可换机执行) |
| `cut_round1.mp4` | 12.0s | 第 1 轮剪出的候选片段 |
| `final.mp4` | 12.0s | 采用的成片(H.264 + AAC1280x720@30fps |
Token 统计印证了核心结论:几十张截图(2934 tokens)只进入**子 Agent**的一次性
上下文,主 Agent 的对话历史(573 tokens)几乎不受截图污染。
(注:合成测试片仅显示"SURFING"字样而非真实冲浪画面,Reviewer 有时会据此判为
不通过——这正是审核者按画面内容如实反馈的体现;换真实视频即无此现象。)
### 依赖
- **ffmpeg / ffprobe**:本机实际剪辑与抽帧。`brew install ffmpeg`macOS/
`apt install ffmpeg`Ubuntu)。本项目在 ffmpeg 8.0 上验证通过。
- **OPENAI_API_KEY**:用 `gpt-5.6-luna` 做视觉定位/审查与文本规划(视觉模型须支持图像输入);未配置时用 `OPENROUTER_API_KEY` 兜底,自动改走 OpenRouter。
### 如何适配 / 扩展
#### 换模型 / 供应商
模型与端点全部通过**环境变量**注入(见 `env.example`),无需改代码:
- `TEXT_MODEL`:规划/边界修正的文本模型(默认 `gpt-5.6-luna`)。
- `VISION_MODEL`:定位/审查的视觉模型,**必须支持图像输入**(默认 `gpt-5.6-luna`)。
- `OPENAI_BASE_URL`:换成任何兼容 OpenAI 协议的端点(自建代理、Azure OpenAI、
或其他厂商网关),配合对应的 `OPENAI_API_KEY` 即可。
```bash
export OPENAI_BASE_URL=https://your-gateway.example.com/v1
export VISION_MODEL=gpt-5.6-luna # 例:用当前廉价旗舰视觉模型
export TEXT_MODEL=gpt-5.6-luna
```
`agents.py` 里的 `OpenAI()` 客户端会自动读取上述变量(`client()` 惰性初始化)。
#### 换输入视频
`make_test_video.py` 用 ffmpeg **程序化生成**一段 54s 的视频,含 4 个明显不同的场景
HIKING 绿 / SURFING 蓝 / SKIING 白 / CYCLING 橙),每段都叠加大号场景名与时间码水印,
让 Vision 仅凭画面就能准确定位——便于复现验收。
换成**你自己的真实视频**:直接 `python demo.py -i 你的.mp4 -o 输出.mp4 "剪辑需求"`
即可(无需改代码)。此时跳过测试片生成,也不再打印定位误差(外部视频无真值)。
#### Blender vs. ffmpeg(剪辑后端)
书中原方案用 **Blender Python APIbpy** 驱动视频序列编辑器(VSE)完成剪辑。
本项目把它实现为**一等后端**`blender_editor.generate_bpy_script()` 把剪辑计划翻译成
一段真实可执行的 bpy 脚本(`new_movie` 导入、`frame_offset_start`/`frame_final_duration`
裁剪、`new_effect(type='TEXT'/'SPEED')` 字幕/变速、`bpy.ops.render.render` 渲染),
`render_with_blender()` 再用 `blender --background --python edit.py` 无头执行。
- `--backend blender`:强制走 Blender(需 `blender --version` 可用);
- `--backend ffmpeg`:强制走 ffmpeg
- `--backend auto`(默认):装了 Blender 用 bpy,否则回退 ffmpeg。
**关键点:无论哪个后端,Proposer 生成的 bpy 脚本都会落盘到 `output/edit_round*.py`**
`--smoke` 下为 `output/edit.py`)——即"生成 Blender Python API 代码"这一核心产物,
可人工核对、也可拷到装了 Blender 的机器上执行。本机未安装 Blender,故本仓库的实际
渲染由 ffmpeg 完成并验证;bpy 脚本已通过 `py_compile` 语法校验,但**未在真实 Blender
上跑过渲染**(装好 Blender 后即可用 `--backend blender` 端到端执行)。两种后端的取舍:
| | ffmpeg | Blenderbpy |
| --- | --- | --- |
| 定位 | 裁剪/拼接/字幕/变速等 2D 流水线 | 3D 场景、合成、关键帧动画、粒子/摄像机 |
| 上手 | 单二进制、无 GUI、CI 友好 | 需装完整 Blender,体积大、渲染慢 |
| 适用 | 绝大多数"剪一段 + 简单特效"需求 | 需要 3D 合成/复杂转场/图层混合时才值得 |
核心的"两步 Vision 定位 + 提议者-审核者"与执行层解耦,两个后端共用同一份剪辑计划,
`agents.py`/`demo.py` 无需为切换后端改动逻辑。
### 文件
| 文件 | 作用 |
| --- | --- |
| `demo.py` | 一条命令跑通的编排入口(CLI、启动自检、迭代循环、token 统计) |
| `agents.py` | `VideoAnalyzerAgent`(两步定位)/ `ProposerAgent` / `ReviewerAgent` |
| `blender_editor.py` | **Blender bpy 脚本生成 + 无头渲染**(书中原方案,核心实验点) |
| `video_editor.py` | 剪辑执行层:`apply_edit()` 统一入口,调度 Blender/ffmpeg 双后端 |
| `make_test_video.py` | 程序化生成含 4 个场景的测试视频 |
| `ffmpeg_utils.py` | ffmpeg/ffprobe 薄封装(统一错误检查、抽帧、探测时长/流) |
`output/`(生成的视频、截图、成片)已被 `.gitignore` 忽略,避免仓库膨胀。
### 局限
- 定位精度取决于场景在画面上的可辨识度;真实视频若场景过渡渐变,边界误差会大于纯色测试片。
- 细粒度步长固定 1s,边界精度上限即 ±1s 量级(满足书中 ±3s 验收)。
- 慢动作音频用 `atempo` 变速,倍率过大时音质下降;转场/多轨混音等复杂特效未覆盖。
- Reviewer 仅抽首/中/尾三帧,长片段中段的偶发错误可能漏检(可调高抽帧密度)。
---
## Notes / 说明
- Prefer `--smoke` then `--quick` before a full Vision run. / 完整 Vision 前先 `--smoke`,再 `--quick`
- Commands/code/paths/env vars are identical in both language sections. / 命令、代码、路径与环境变量在中英文两侧保持一致。