Decompose
Decompose tests into atomic task DAG with wave-based parallel execution and module stewardshipFrom its SKILL.md
npx -y skills add zxdxjtu/harness --skill decomposeAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 8 stars8 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
5.4 KB, as published. Nobody here has run it
Decompose — Atomic Task DAG + Module Stewardship
Parameter: $ARGUMENTS — feature ID (e.g. F001)
Precondition
- Spec must be
approved(orreviewedif spec-review was in the flow) - Test files must exist and ALL be RED
Step 1: Analyze Test Cases
Read all test files for this feature. List every test ID.
Step 2: Build Atomic Tasks
Map each test ID (or group of closely related tests) to one atomic task.
Each task must satisfy:
- ≤ 2 hours of agent work
- One clear completion criterion: corresponding tests go from RED → GREEN
- Explicit inputs: which task outputs (files, interfaces) are required
- Explicit outputs: which files are produced + which tests turn green
- Independently verifiable: can run the test in isolation
Step 3: Analyze Dependencies
Determine task ordering:
- Type/interface definitions → implementations using those types
- Data layer → business logic → UI layer
- Unit test tasks → integration test tasks → E2E test tasks
Step 4: Assign Waves
- No dependencies → Wave 1 (parallel)
- Depends on Wave 1 → Wave 2 (parallel)
- ... and so on
Step 5: Module Stewardship Analysis
Analyze which code modules each task will touch:
- Identify module boundaries: Based on directory structure, package boundaries, or
.harness/full-design.md - Map tasks to modules: For each task, determine which module(s) it affects
- Generate AGENT.md files: For each affected module that doesn't already have one:
Read the template from the harness plugin's templates/agent-md-template.md and generate a module-specific AGENT.md:
# Module: {module-name}
## Path: {relative/path}
## Last Updated: {ISO timestamp}
## Responsibilities
{derived from code analysis and design doc}
## Boundaries
- **Owns**: {files in this directory}
- **Depends On**: {modules imported by this module}
- **Depended By**: {modules that import this module}
- **Forbidden Dependencies**: {inferred from architecture — e.g. UI modules should not import from CLI}
## Quality Standards
{derived from project config and existing linting rules}
## Interface Contract
- **Exports**: {detected public API}
- **Data Flow**: {input/output shapes}
## Change History
| Date | Change | Impact | Reviewer |
|------|--------|--------|----------|
Place AGENT.md in the actual module directory (e.g. src/auth/AGENT.md).
- Generate module dependency graph: Write
.harness/module-graph.json:
{
"modules": {
"src/auth": {
"agent_md": "src/auth/AGENT.md",
"depends_on": ["src/db", "src/config"],
"depended_by": ["src/api", "src/middleware"],
"forbidden": ["src/ui"]
}
},
"last_updated": "{ISO timestamp}"
}
- Add stewardship metadata to tasks: Each task entry includes which module(s) it touches.
Step 6: Write tasks.md
Write to .harness/tasks.md:
# Tasks — $ARGUMENTS
## DAG Visualization
Wave 1: [T001] [T002] [T005] ← parallel, no dependencies
Wave 2: [T003] [T006] ← depends on Wave 1
Wave 3: [T004] ← depends on Wave 2
Wave 4: [T007] [T008] ← E2E verification
## Task List
| ID | Name | Wave | Status | Tests | Retries | BlockedBy | Files | Modules |
|----|------|------|--------|-------|---------|-----------|-------|---------|
| T001 | desc | 1 | pending | T001-V1 | 0 | - | src/... | src/auth |
| T002 | desc | 1 | pending | T002-V1 | 0 | - | src/... | src/db |
| T003 | desc | 2 | pending | T003-V1,V2 | 0 | T001 | src/... | src/auth, src/api |
## Module Stewardship
| Module | AGENT.md | Tasks Affecting | Guardian Active |
|--------|----------|----------------|-----------------|
| src/auth | src/auth/AGENT.md | T001, T003 | Yes |
| src/db | src/db/AGENT.md | T002 | Yes |
## Summary
- Total tasks: N
- Waves: M
- Modules affected: K
- AGENT.md files generated: J
- Estimated parallel speedup: Nx → Mx
Step 7: Update Progress
Append to .harness/progress.md:
## [timestamp] Task Decomposition — $ARGUMENTS
- Total tasks: N
- Waves: M
- Modules with stewardship: K
- Status: awaiting human approval
Step 8: Request Approval
Present the DAG and module stewardship plan to the user. Ask:
"任务拆解完成:{N} 个原子任务,分 {M} 个 Wave,涉及 {K} 个模块(已生成 AGENT.md)。任务依赖关系是否正确?有需要拆分或合并的任务吗?"
Do NOT proceed until approved.
Next Step — Auto-Navigate
- Read the
flowfield from the spec - Display progress visualization:
📋 SDD 进度 — {FXXX}: {feature name}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
✅ proposal → ✅ tdd-align → ✅ decompose → ⬜ sprint → ...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
当前阶段: decompose (任务拆解 + 责任田) ✅ 已完成
下一阶段: sprint (自动执行)
- Ask: "任务拆解完成,共 {N} 个原子任务分 {M} 个 Wave。下一步是 Sprint 自动执行(按 Wave 并行开发,每 Wave 后触发模块守护和评估检查)。是否继续?"
- If user confirms → execute sprint logic directly