task-decomposer（任务分解器）

Task Decomposer Protocol

Break a complex task into independently executable, verifiable subtasks. Produce a structured plan first, get approval, then execute step by step with progress tracking.

Workflow

1. Analyze the task
2. Decompose into subtasks (with dependency map)
3. Present the plan and get approval
4. Execute sequentially — complete each step before moving to next
5. Produce summary report

Step 1: Analyze the Task

Before decomposing, answer these questions:

• What is the goal? What does success look like?
• What is the scope? What is included, what is excluded?
• Are there constraints? Technology, time, dependencies?
• Is anything unclear? If so, ask one clarifying question — never multiple.
• Domain Awareness Check: If the user's request clearly aligns with a known domain (e.g., Mobile apps, DB Design, API Architecture), you MUST read the skills/.curated/domains/<domain>/ECOSYSTEM.md file (if it exists) to understand the pre-defined workflows and available skills *before* creating your Dependency Map. Do not hallucinate the workflow.

Do not generate subtasks until the task is clearly understood.

Step 2: Decompose into Subtasks

Each subtask must meet these criteria:

• Atomic: Does one thing, contains no internal decision branches
• Verifiable: You can definitively say "done" when completed
• Independent (when possible): Can run without waiting for other subtasks
• Has a concrete output: A file, a function, a document, a decision

Dependency Map

Define relationships between subtasks:

• Sequential (->): B cannot start until A completes
• Parallel (||): Can run simultaneously — no shared dependencies
• Optional (?): Not core, nice to have

How to handle parallel steps: Group || steps together. Complete all steps in the group, then move to the next sequential step.

Group A (parallel): #1 || #2 || #3  -> complete all, then continue
Group B (sequential): #4 -> #5 -> #6

Dependency Visualization

Use mermaid for complex dependency graphs:

graph LR
    T1["#1 Data Model"] --> T3["#3 CRUD Operations"]
    T2["#2 DB Connection"] --> T3
    T3 --> T4["#4 API Endpoints"]
    T4 --> T5["#5 Tests"]

Step 3: Present the Plan

Show the plan to the user using the format in templates/task-plan.md.tmpl and ask for approval.

Do not start execution without explicit user approval.

Step 4: Execute Step by Step

After approval, complete each subtask:

1. Before starting, print a header: ### [#N] Task Name
2. Execute the task — write code, create files, research, document
3. On completion, print: Completed: [what was produced]
4. Move to the next subtask

On failure:

• Explain the error
• Suggest an alternative path
• Ask the user: "Continue or fix this first?"

Escalation rule: If 2 consecutive subtasks fail, stop automatically and ask:

"2 consecutive steps failed. Should we review the plan together before continuing?"

Cross-Skill Integration

• checkpoint-guardian: During execution, critical actions (file deletion, DB writes, deploys) should trigger risk assessment checkpoints.
• multi-brain: For complex subtasks where the approach is unclear, use 3-perspective evaluation before committing to an implementation.

Step 5: Summary Report

When all subtasks are done, present a summary using templates/task-summary.md.tmpl.

When to Skip

Do not decompose — respond directly when:

• Single-step, clear task ("translate this file", "fix this bug")
• User already knows the steps and only wants execution
• Task can be completed in under 3 minutes

See references/SKIP_CONDITIONS.md for the full decision matrix.

References

• See references/PLAN_TEMPLATE.md for the plan output structure.
• See references/EXAMPLES.md for worked examples including a real code output.
• See references/SKIP_CONDITIONS.md for when to skip decomposition.

Templates

• Use templates/task-plan.md.tmpl for presenting the plan to the user.
• Use templates/task-summary.md.tmpl for the completion report.