Highlight Manifest 示例（Electron v0.1）

版本：0.1
状态：示例
说明：本示例展示 Highlight 这类本地 API 服务如何以 zip + manifest.yaml 交付给 FlowyAIPC Electron v0.1 安装器。配置模板位于 Skill 的 assets/ 目录，安装后 Skill 可通过相对路径 assets/config.json 直接读取；服务启动由 Agent 根据 Skill 文档自行执行。

目录结构

text

highlight/
  manifest.yaml
  payload/
    AGENTS.md
    SOUL.md
    TOOLS.md
    IDENTITY.md
    USER.md
    MEMORY.md
    HEARTBEAT.md
    service.exe
    llama-server/
      llama-server.exe
    models/
      model-Q4_K_M.gguf
      mmproj-F16.gguf
    skills/
      highlight/
        SKILL.md
        assets/
          config.example.json
        references/
          workflow-start-recording.md
          workflow-fetch-highlights.md
  docs/
    openapi.yaml
    integration.md
    error-codes.md
    privacy.md

manifest.yaml

yaml

spec_version: 0.1
package_name: highlight
package_version: 1.0.0
type: api-skill

agent:
  id: highlight-agent
  name: "足球精彩时刻检测"

summary: "本地足球精彩时刻检测服务"

display_name:
  zh: "精彩时刻检测"
  en: "Highlight Detection"
  jp: "ハイライト検出"

description:
  zh: "本地足球精彩时刻检测服务，实时录制屏幕并 AI 检测进球"
  en: "Local soccer highlight detection service with real-time AI goal detection"
  jp: "ローカルサッカーハイライト検出サービス"

runtime:
  mode: managed
  entrypoint: payload/service.exe
  working_dir: payload
  config_template: payload/skills/highlight/assets/config.example.json
  generated_config: payload/skills/highlight/assets/config.json
  health_check:
    method: GET
    path: /health
  ready_check:
    method: GET
    path: /health
    ready_field: status
    ready_value: ready
  stop:
    kind: api
    method: DELETE
    path: /api/v1/sessions/{session_id}

artifacts:
  - id: highlight-service
    kind: binary
    name: service.exe
    required: true
    source:
      kind: embedded
      path: payload/service.exe
    target: "{install_dir}/payload/service.exe"

  - id: ffmpeg-win32
    kind: archive
    name: ffmpeg
    required: true
    version: "6.1"
    source:
      kind: download
      url: https://example.com/ffmpeg-release-full.zip
      archive: zip
      extract: true
      strip_components: 0
      checksum:
        algorithm: sha256
        value: "<sha256>"
    target: "{package_shared_dir}/bin/ffmpeg"
    bins:
      - ffmpeg.exe

  - id: llama-server
    kind: binary
    name: llama-server.exe
    required: true
    source:
      kind: embedded
      path: payload/llama-server/llama-server.exe
    target: "{install_dir}/payload/llama-server/llama-server.exe"

  - id: model-q4
    kind: file
    name: model-Q4_K_M.gguf
    required: true
    source:
      kind: embedded
      path: payload/models/model-Q4_K_M.gguf
    target: "{package_shared_dir}/models/model-Q4_K_M.gguf"

  - id: mmproj
    kind: file
    name: mmproj-F16.gguf
    required: true
    source:
      kind: embedded
      path: payload/models/mmproj-F16.gguf
    target: "{package_shared_dir}/models/mmproj-F16.gguf"

config:
  template: payload/skills/highlight/assets/config.example.json
  output: payload/skills/highlight/assets/config.json
  placeholders:
    http_host: "127.0.0.1"
    http_port: 8088
    auth_token: "{generated_secret}"
    service_path: "{artifact:highlight-service:path}"
    ffmpeg_path: "{artifact:ffmpeg-win32:bin:ffmpeg.exe}"
    llama_server_path: "{artifact:llama-server:path}"
    llama_model: "{artifact:model-q4:path}"
    llama_mmproj: "{artifact:mmproj:path}"
    output_dir: "{package_shared_dir}/output"
    log_dir: "{package_shared_dir}/logs"

resources:
  scheduling_class: gpu-exclusive
  gpu:
    required: true
    vendor: intel
    min_vram_gb: 8
  memory:
    min_gb: 16
  disk:
    min_free_gb: 20
    growth_hint: "3.6 GB/hour"
  ports:
    - port: 8088
      protocol: tcp
      required: true
      purpose: api
    - port: 18080
      protocol: tcp
      required: true
      purpose: internal_llama_server

permissions:
  screen_capture: required
  audio_capture: required
  file_read:
    - user_selected_source
  file_write:
    - "{package_shared_dir}/output"
    - "{package_shared_dir}/logs"
  network:
    local_only: true
  process_spawn: required
  gpu_compute: required
  model_download: none

privacy:
  local_processing: true
  reads:
    - screen_content
    - system_audio
  stores:
    - video_clips
    - session_metadata
  llm_allowed_fields:
    - event_count
    - event_timestamp
    - short_summary
  llm_blocked_fields:
    - raw_frame
    - screenshot
    - full_video
    - absolute_file_path
    - auth_token

model:
  mode: local
  provider: llama.cpp
  default_model: model-q4
  timeout_seconds: 60

skills:
  - path: payload/skills/highlight/SKILL.md

compatibility:
  min_flowyaipc_version: 0.8.0
  protocol_version: api-v1
  platform:
    os: [win32]
    arch: [x64]

install:
  mode: managed
  steps:
    - id: resolve-artifacts
      kind: resolve-artifacts
      artifacts:
        - highlight-service
        - ffmpeg-win32
        - llama-server
        - model-q4
        - mmproj
    - id: render-config
      kind: render-config
      template: payload/skills/highlight/assets/config.example.json
      output: payload/skills/highlight/assets/config.json
    - id: scan-skills
      kind: scan-skills
      path: payload/skills

uninstall:
  preserve_user_data: true
  remove_paths: []
  keep_paths:
    - shared

payload/skills/highlight/assets/config.example.json

config.placeholders 中的值会替换 JSON 字符串值。数组和对象应直接写在 JSON 模板中，不要作为 placeholder 值声明。

json

{
  "http": {
    "host": "{http_host}",
    "port": "{http_port}",
    "allowed_origins": [
      "http://127.0.0.1:8088",
      "http://localhost:8088"
    ]
  },
  "auth": {
    "token": "{auth_token}"
  },
  "paths": {
    "service": "{service_path}",
    "ffmpeg": "{ffmpeg_path}",
    "llama_server": "{llama_server_path}",
    "llama_model": "{llama_model}",
    "llama_mmproj": "{llama_mmproj}",
    "output_dir": "{output_dir}",
    "log_dir": "{log_dir}"
  }
}

payload/skills/highlight/SKILL.md

Skill 通过 skills.load.extraDirs 引用包内 skills 目录被发现。安装器不写入 marker 文件。

markdown

---
name: highlight
description: "通过 Highlight 服务管理录屏、检测和视频片段生成。"
metadata:
  openclaw:
    os: [win32]
    requires:
      bins:
        - ffmpeg.exe
      config:
        - configPath
---

# Highlight

使用 manifest 中 `agent` 字段的配置和 `assets/config.json` 中的生成配置调用 Highlight 服务或启动相关流程。Skill 目录通过 `skills.load.extraDirs` 被 OpenClaw 加载。

说明

顶层 package zip 的 sha256 来自后端分发元数据，不写在包内 manifest 中。
ffmpeg-win32 使用 archive，平台负责下载、校验、解压和定位 ffmpeg.exe。
model-q4 和 mmproj 使用 file，示例中表示模型随包提供。
如果模型希望单独下载，可把 source.kind 改成 download 并提供 HTTPS url 与 sha256 checksum。
Skill 注册由顶层 skills 字段驱动，不通过 directory artifact 的 post_install 驱动。
artifact 落地路径通过 {artifact:<id>:path} / {artifact:<id>:bin:<name>} 在 config.placeholders 中引用。
Electron v0.1 安装器不会执行 launch、health-check、ready-check install step。
Electron v0.1 安装器会写 openclaw.json（agents.list + skills.load.extraDirs）。

Highlight Manifest 示例（Electron v0.1） ​

目录结构 ​

manifest.yaml ​

payload/skills/highlight/assets/config.example.json ​

payload/skills/highlight/SKILL.md ​

说明 ​

Highlight Manifest 示例（Electron v0.1）

目录结构

manifest.yaml

payload/skills/highlight/assets/config.example.json

payload/skills/highlight/SKILL.md

说明