hierarchical-task-spawn

name	hierarchical-task-spawn
homepage	https://github.com/wd041216-bit/openclaw-hierarchical-task-spawn
description	Plan and execute spawned OpenClaw work as a task tree instead of a flat list. Use when a medium task should be split into fine sub-tasks, or a complex task should be split into medium workstreams that can further split into fine tasks.

Hierarchical Task Spawn

Use this skill when one spawned task is still too big.

Goal

Turn substantial work into a small task tree:

• simple work -> do it directly
• medium work -> split into a few fine tasks
• complex work -> split into medium workstreams, then split those into fine tasks if needed

Complexity Ladder

Simple

Do not spawn.

Use direct execution when the task can be finished in one short flow with little search, no meaningful branching, and no multiple deliverables.

Medium

Split into fine tasks when the request has 2-4 meaningful steps or deliverables.

Examples:

• research + summary + table
• outline + draft + polish
• collect inputs + generate PPT + final QA

Complex

Split into medium workstreams first when the request contains clearly different domains, multiple deliverables, or enough uncertainty that one child task would become bloated.

Examples:

• market scan + PM plan + PPT deck
• source collection + competitor analysis + executive summary + spreadsheet
• requirement clarification + implementation plan + artifact generation

Tree Shape

Prefer a shallow tree with explicit ownership:

1. root request
2. medium workstreams
3. fine execution tasks

Default depth limit: 3 levels including the root.

Workflow

1. Classify the task

Decide whether the request is simple, medium, or complex before spawning.

2. Split only where useful

• medium task -> 2-4 fine tasks
• complex task -> 2-5 medium workstreams
• each medium workstream may split into 2-4 fine tasks if needed

Do not create child tasks for cosmetic micro-steps.

3. Assign the right owner

Use the best specialist for each branch:

• research -> search, source review, verification, comparisons
• office -> team coordination, milestones, meeting outputs, shared execution
• slides -> personal planning, prioritization, private execution
• council -> multi-perspective debate
• current agent -> when the work still belongs here

Keep family and kittypuppy isolated to their own contexts unless the user is already in those agents.

4. Spawn by layer

• root spawns medium branches when the task is complex
• a medium branch may spawn fine tasks if it is still too broad
• leaf tasks should execute real work instead of just rephrasing the plan

Leaf Task Standard

A leaf task should be concrete enough to finish without further decomposition.

Good leaf tasks:

• 搜索 2026 年 AI agent 框架并提取发布时间、价格、定位
• 把研究结果整理成 6 页 PPT 初稿
• 把会议纪要整理成 owner/deadline/action list

Bad leaf tasks:

• 继续想一想
• 处理这个项目
• 完成所有内容

Reporting

Prefer progressive reporting:

• leaf tasks can report their own results when useful
• parent tasks should aggregate only related branches
• do not wait for unrelated branches if one finished result is already valuable

In Feishu, pair this skill with feishu-parallel-dispatch and feishu-task-status.

Guardrails

• Do not recurse forever
• Prefer 3-8 active leaves, not a swarm
• Stop splitting once the next child would be a single-tool or single-deliverable action
• If the user clearly wants one synchronous answer, keep the tree internal and return one combined result
• If one branch stalls, do not block the whole tree from reporting other finished branches

Example

User asks: 帮我做一个行业调研、整理成对比表，再做一份给老板的 PPT

Good split:

1. medium branch: 研究与证据
2. medium branch: 表格整理
3. medium branch: PPT 生成

Then:

• 研究与证据 may split into search / browse / verify
• PPT 生成 may split into outline / slide draft / QA

This keeps each child narrow enough to finish well.