workflow-executor-inline

用于代理提供者(ZenMux、自定义API提供者)的行内工作流执行器。直接在主上下文中执行工作流步骤,而不启动任务子代理。在通过ZenMux或其他代理提供者运行looplia工作流、任务子代理因“invalid_model”错误失败或需要无上下文隔离的行内执行时使用。架构:每个工作流步骤都在行内执行(无Task工具)- 读取技能、执行任务、写入输出,然后进行到下一个步骤。所有步骤共享主上下文。v0.6.6:为与ZenMux的跨提供者兼容性而创建。

📖 中文导读:用于代理提供者(ZenMux、自定义API提供者)的行内工作流执行器。直接在主上下文中执行工作流步骤,而不启动任务子代理。在通过ZenMux或其他代理提供者运行looplia工作流、任务子代理因“invalid_model”错误失败或需要无上下文隔离的行内执行时使用。架构:每个工作流步骤都在行内执行(无Task工具)- 读取技能、执行任务、写入输出,然后进行到下一个步骤。所有步骤共享主上下文。v0.6.6:为与ZenMux的跨提供者兼容性而创建。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。

name: workflow-executor-inline description: | Inline workflow executor for proxy providers (ZenMux, custom API providers). Executes workflow steps directly in the main context WITHOUT spawning Task subagents.

Use this skill when:

  • Running looplia workflows via ZenMux or other proxy providers
  • Task subagents fail with "invalid_model" errors
  • You need inline execution without context isolation

Architecture: Each workflow step is executed INLINE (no Task tool) - read skill, execute mission, write output, then proceed to next step. All steps share the main context.

v0.6.6: Created for cross-provider compatibility with ZenMux.

Workflow Executor Inline (v0.6.6)

Execute looplia workflows without Task subagents. This skill is specifically designed for proxy providers (ZenMux, custom) where the Claude Agent SDK's subagent spawning doesn't work due to model name incompatibility.

When to Use

Use this skill when:

  • The system has injected a hint to use inline execution
  • Running workflows via ZenMux or other proxy providers
  • Task subagents fail with "invalid_model" errors

CRITICAL: Inline Execution (No Subagents)

DO NOT spawn Task subagents. Execute all workflow steps directly in the main context.

For each workflow step:

  1. Read the skill definition using Skill tool: Skill("{step.skill}")
  2. Read input file(s) if specified using Read tool
  3. Execute the mission following skill instructions
  4. Write JSON output to specified path using Write tool
  5. Validate output (if validation rules defined)
  6. Proceed to next step

Execution Protocol

Phase 1: Sandbox Setup

Same as standard workflow-executor:

  1. Generate sandbox ID:

    {first-input-name}-{YYYY-MM-DD}-{random4chars}
    Example: video-transcript-2025-12-18-xk7m
    
  2. Create folder structure:

    sandbox/{sandbox-id}/
    ├── inputs/
    │   └── {input-files}.md
    ├── outputs/
    ├── logs/
    └── validation.json
    
  3. Copy input files to inputs/

Phase 2: Workflow Parsing

  1. Read workflow file: workflows/{workflow-id}.md
  2. Parse YAML frontmatter for steps
  3. Validate each step has skill: and mission: fields
  4. Build dependency graph from needs: fields

Phase 3: Inline Step Execution

Execute steps ONE AT A TIME, INLINE (no Task tool):

FOR EACH step in dependency order:
    │
    ▼
┌─────────────────────────────────────────┐
│ 1. INVOKE skill: Skill("{step.skill}")  │
│    → This loads skill context           │
│                                         │
│ 2. READ input file (if provided)        │
│    → Use Read tool                      │
│                                         │
│ 3. EXECUTE mission                      │
│    → Follow skill instructions          │
│    → Generate JSON output               │
│                                         │
│ 4. WRITE output file                    │
│    → Use Write tool                     │
│    → Output to step.output path         │
│                                         │
│ 5. VALIDATE output                      │
│    → Check required_fields              │
│    → Retry if failed (max 2x)           │
│                                         │
│ 6. UPDATE validation.json               │
│    → Mark step as validated: true       │
└─────────────────────────────────────────┘
         │
         ▼
    NEXT STEP

Example: Inline Execution

For step:

- id: analyze-content
  skill: media-reviewer
  mission: |
    Deep analysis of video transcript. Extract key themes,
    important quotes, and narrative structure.
  input: ${{ sandbox }}/inputs/content.md
  output: ${{ sandbox }}/outputs/analysis.json
  validate:
    required_fields: [contentId, headline, keyThemes]

Inline execution sequence:

  1. Invoke skill:

    Skill("media-reviewer")
    
  2. Read input:

    Read("sandbox/video-2025-12-18-xk7m/inputs/content.md")
    
  3. Execute mission:

    • Follow media-reviewer skill instructions
    • Analyze the content
    • Generate structured JSON output
  4. Write output:

    Write("sandbox/video-2025-12-18-xk7m/outputs/analysis.json", jsonContent)
    
  5. Validate:

    • Check contentId, headline, keyThemes exist
    • Update validation.json
  6. Proceed to next step


Anti-Patterns

❌ WRONG - Spawning Task subagents:

{
  "subagent_type": "workflow-step",
  "description": "Execute step...",
  "prompt": "..."
}

Task subagents don't work with proxy providers.

❌ WRONG - Batching all steps: Execute all steps at once without validation between them.

✅ CORRECT - Inline step-by-step:

  1. Skill("media-reviewer")
  2. Read input
  3. Execute mission
  4. Write output
  5. Validate
  6. Proceed to next

Variable Substitution

Same as standard workflow-executor:

Variable Resolution
${{ sandbox }} sandbox/{sandbox-id}
${{ inputs.{name} }} sandbox/{id}/inputs/{name}.md
${{ steps.{id}.output }} Output path of step {id}

Error Handling

Scenario Action
Workflow not found Error with available workflows
Step missing skill: Error: "Step missing required 'skill' field"
Step missing mission: Error: "Step missing required 'mission' field"
Validation fails Retry inline with feedback (max 2 retries)
Max retries exceeded Report failure with details

Key Differences from Standard workflow-executor

Aspect workflow-executor workflow-executor-inline
Execution Task subagent per step Inline in main context
Context Isolated per step Shared main context
Provider Support Anthropic Direct only All providers
Use Case Production (Anthropic) Proxy providers (ZenMux)

Version History

  • v0.6.6: Created for ZenMux cross-provider compatibility