minimal-agent 0.1.9 → 0.3.0

package/workflows/youtube-shorts.yaml

@@ -0,0 +1,171 @@
+name: youtube-shorts
+description: |
+  示例工作流：输入主题 → LLM 写 5 幕脚本 → OpenRouter 生图 ×5 → 中文 TTS ×5
+  → ffmpeg 合成竖屏 mp4。
+  这是把 4 个通用 skill（image-gen-openrouter / tts-zh / subtitle-srt /
+  video-compose）组合的样例 —— 你可以在编辑器里改 inputs / 增删节点 / 换 skill
+  去做任意视频流水线（横屏长视频、广告片、教程片都行）。
+version: "0.1"
+
+inputs:
+  - name: topic
+    type: string
+    required: true
+    description: "视频主题（中文）"
+  - name: openrouter_key
+    type: string
+    required: false
+    description: "OpenRouter API key；不填则用 OPENROUTER_API_KEY 环境变量"
+  - name: aspect
+    type: enum
+    values: ["9:16", "16:9", "1:1", "4:5"]
+    default: "9:16"
+    description: "图片与视频宽高比"
+  - name: width
+    type: number
+    default: 1080
+    description: "视频输出宽度"
+  - name: height
+    type: number
+    default: 1920
+    description: "视频输出高度"
+
+steps:
+  # ---------- 1. 用 LLM 写 5 幕脚本（强制 5 幕以便后续 loop 用静态数组） ----------
+  - id: write_script
+    llm: |
+      你是短视频文案策划。请为主题"${inputs.topic}"写一个 **正好 5 幕** 的竖屏短视频。
+      每一幕包含：
+        - text：一句 8-12 字的中文旁白
+        - image_prompt：一段 30-50 词的英文图片提示词，描述这一幕的画面
+      只输出一个 JSON 对象，结构严格如下，不要包裹在 ``` 里、不要任何额外解释：
+      {"scenes":[
+        {"text":"...","image_prompt":"..."},
+        ...共 5 项...
+      ]}
+    capture:
+      text: script_json_text
+
+  # ---------- 2. 把 LLM 输出落盘为 script.json ----------
+  - id: write_script_json
+    tool: Write
+    args:
+      file_path: "./assets/script.json"
+      content: "${script_json_text}"
+
+  # ---------- 3. 解析 script.json，拆出 prompts.txt / narrations.txt（每行一个） ----------
+  - id: parse_script
+    tool: Bash
+    args:
+      command: |
+        set -euo pipefail
+        mkdir -p ./assets ./out
+        jq -r '.scenes[].image_prompt' ./assets/script.json > ./assets/prompts.txt
+        jq -r '.scenes[].text' ./assets/script.json > ./assets/narrations.txt
+        # 把 narrations 转成 JSON 数组字符串，用于后续 subtitle-srt 调用
+        jq -c '[.scenes[].text]' ./assets/script.json > ./assets/narrations_json.txt
+        wc -l ./assets/prompts.txt ./assets/narrations.txt
+    capture:
+      content: parse_log
+
+  - id: assert_script_parsed
+    type: assert
+    condition: 'fileExists("./assets/prompts.txt") && fileExists("./assets/narrations.txt")'
+    onFail: "脚本解析失败：./assets/prompts.txt 或 ./assets/narrations.txt 不存在"
+
+  # ---------- 4. 生成 5 张配图（loop 静态数组 + 每轮先读对应一行 prompt） ----------
+  - id: gen_images
+    type: loop
+    over: "[0, 1, 2, 3, 4]"
+    as: i
+    do:
+      - id: read_prompt
+        tool: Bash
+        args:
+          command: "sed -n '$((${i}+1))p' ./assets/prompts.txt"
+        capture:
+          content: prompt_text
+      - id: gen_one_image
+        skill: image-gen-openrouter
+        input: 'prompt=${prompt_text} output=./assets/img_${i}.png aspect=${inputs.aspect} key=${inputs.openrouter_key}'
+
+  # ---------- 5. 生成 5 段 TTS 音频 ----------
+  - id: gen_audio
+    type: loop
+    over: "[0, 1, 2, 3, 4]"
+    as: i
+    do:
+      - id: read_narration
+        tool: Bash
+        args:
+          command: "sed -n '$((${i}+1))p' ./assets/narrations.txt"
+        capture:
+          content: narration_text
+      - id: gen_one_tts
+        skill: tts-zh
+        input: 'text=${narration_text} output=./assets/audio_${i}.mp3'
+
+  # ---------- 6. 探测各段音频时长，逗号串联 ----------
+  - id: probe_durations
+    tool: Bash
+    args:
+      command: |
+        set -euo pipefail
+        for f in $(ls ./assets/audio_*.mp3 | sort); do
+          ffprobe -v error -show_entries format=duration -of csv=p=0 "$f"
+        done | paste -sd, -
+    capture:
+      content: durations
+
+  # ---------- 7. 合并所有 TTS 段成一条音轨 ----------
+  - id: merge_audio
+    tool: Bash
+    args:
+      command: |
+        set -euo pipefail
+        LIST=/tmp/audio_concat.txt
+        : > "$LIST"
+        for f in $(ls ./assets/audio_*.mp3 | sort); do
+          echo "file '$(pwd)/${f}'" >> "$LIST"
+        done
+        ffmpeg -y -f concat -safe 0 -i "$LIST" -c copy ./assets/audio.mp3
+        ls -la ./assets/audio.mp3
+
+  # ---------- 8. 读 narrations_json（前面 parse_script 已写好），作 SRT 文本输入 ----------
+  - id: read_narrations_json
+    tool: Read
+    args:
+      file_path: "./assets/narrations_json.txt"
+    capture:
+      content: narrations_json
+
+  # ---------- 9. 生成 SRT 字幕 ----------
+  - id: make_srt
+    skill: subtitle-srt
+    input: 'texts=${narrations_json} durations=${durations} output=./assets/subs.srt'
+
+  # ---------- 10. 合成最终 mp4 ----------
+  - id: compose_video
+    skill: video-compose
+    input: 'images_glob=./assets/img_*.png audio=./assets/audio.mp3 srt=./assets/subs.srt output=./out/shorts.mp4 width=${inputs.width} height=${inputs.height}'
+
+  # ---------- 11. 守门 assert ----------
+  - id: verify_output
+    type: assert
+    condition: 'fileExists("./out/shorts.mp4")'
+    onFail: "最终视频 ./out/shorts.mp4 未生成，请检查 video-compose 日志"
+
+__meta:
+  layout:
+    - { id: write_script,           x: 100, y: 50 }
+    - { id: write_script_json,      x: 100, y: 200 }
+    - { id: parse_script,           x: 100, y: 350 }
+    - { id: assert_script_parsed,   x: 100, y: 500 }
+    - { id: gen_images,             x: 100, y: 650 }
+    - { id: gen_audio,              x: 450, y: 650 }
+    - { id: probe_durations,        x: 450, y: 850 }
+    - { id: merge_audio,            x: 450, y: 1000 }
+    - { id: read_narrations_json,   x: 100, y: 850 }
+    - { id: make_srt,               x: 275, y: 1150 }
+    - { id: compose_video,          x: 275, y: 1300 }
+    - { id: verify_output,          x: 275, y: 1450 }