skill 引擎参照 claude code 支持运行skill 中定义的 subagent

[zhuzhaoyun]

Pre-checks

• I have searched existing issues and this is not a duplicate.

What problem does this solve?

当前如 deepagents, claude code 都支持在 skill 中定义的 subagent 执行，但是 clawtih 中无法运行 skill 中定义的 subagents , 以 skill-creator 为例，这个 skill 中定义了多个 subagent ,但在 clawith 中是无法运行的。

Proposed solution

No response

Willing to contribute?

• I'd be interested in working on this.

[Clawiee]

Thank you for this feature request! 🙏

Analysis

You're correct that Clawith's skill engine currently doesn't support running subagents defined within skills. This is a valuable feature that many advanced users want.

How OpenClaw/Claude Code Handle Subagents

Based on OpenClaw's implementation:

1. Spawn command/tool: /subagents spawn <agentId> <task> or via sessions_spawn tool
2. Session key pattern: agent:<agentId>:subagent:<uuid>
3. Nested subagents: Supported with maxSpawnDepth config (default: 1, max: 2)
4. Per-subagent config: Model selection, tools, timeout, thinking level

What Would Be Needed

To implement skill-defined subagent support in Clawith:

1. Skill format extension: Define a standard schema for subagents in skill YAML/MD files:

   subagents:
     - name: coder
       model: claude-sonnet
       tools: [execute_code, browser]
       system_prompt: "You are a coding specialist..."
     - name: researcher  
       model: claude-haiku
       tools: [web_search]
       system_prompt: "You are a research specialist..."

2. Skill engine modification: Parse and register subagents from loaded skills

3. Spawn tool: Create a spawn_subagent tool that allows agents to:

   • List available subagents
   • Spawn subagent sessions
   • Communicate with subagents
   • Aggregate results

4. Concurrency control: Like OpenClaw's maxConcurrent: 8 and maxChildrenPerAgent: 5

5. Tool policy by depth: Control which tools subagents can access based on nesting level

Questions for Clarification

1. Use case priority: Would you like subagents primarily for:

   • Parallel task execution (one agent orchestrates multiple workers)
   • Specialized domain expertise (coder vs researcher vs writer)
   • Security isolation (subagents with limited tool access)

2. Subagent type: Should these be:

   • Native Clawith agents (created in the platform)
   • Lightweight task sessions (similar to OpenClaw's approach)
   • OpenClaw-compatible subagent sessions

3. Communication: How should results be aggregated?

   • Direct message back to orchestrator
   • Shared workspace/files
   • Structured result format

---

@yaojin3616 @zanyuzhantong FYI - this is a significant enhancement request that would require architecture work on the skill engine.

---

💡 Tip: For international collaboration, issues are typically written in English so the broader community can understand and contribute. Feel free to continue in any language you're comfortable with!

[zhuzhaoyun]

Parallel task execution
OpenClaw-compatible subagent sessions
Direct message back to orchestrator