plan-decompose

/hypo-workflow:plan-decompose

Output Language Rules

📌 输出语言规则： 读取 config.yaml → output.language

• zh-CN / zh：所有用户可见的输出使用中文（PROGRESS、报告、状态提示、错误消息、交互提问）
• en：使用英文
• auto：跟随用户对话语言 内部日志（log.yaml、state.yaml）始终英文。

Use this skill for P2 Decompose only.

For /hw:plan --batch, this phase decomposes Feature Queue entries according to batch.decompose_mode.

Preconditions

• P1 Discover has already clarified the project enough to define milestones

Execution Flow

1. Read the current design summary and repo context.
2. Split work into serial milestones.
3. Each milestone must include:
   • objective
   • implementation scope
   • test spec
   • expected artifacts
4. For implementation work, prefer a runnable vertical slice: one narrow behavior that crosses only the layers needed to run and validate it.
5. Flag horizontal-only splits when database/API/UI/schema-only milestones do not produce runnable behavior.
6. Prefer narrow milestones when architecture may shift later prompts.
7. Preserve append-mode safety:
   • do not silently renumber executed prompts
   • append new prompts after the highest safe sequence number

Interactive Behavior

• in interactive mode, show the proposed milestone split and ask follow-up questions if dependencies or scope boundaries are still ambiguous
• after P2 produces the split, stop at a checkpoint before P3
• the checkpoint must show:
  • milestone number and name
  • objective
  • implementation scope
  • test spec
  • expected artifacts
  • runnable vertical slice quality, including touched layers and real validation evidence
  • unresolved assumptions
• wait for explicit user confirmation before entering P3 Generate
• do not generate .pipeline/ files, prompt files, or architecture files from P2 directly
• if the user asks for changes, revise the split and present the checkpoint again
• in auto mode, finalize the milestone split directly unless blocked

P2 Checkpoint Gate

Interactive P2 completion is not permission to write files. The only valid next step is to display the proposed decomposition and ask the user to confirm it. P3 may start only after the user explicitly approves the milestone split.

Batch Decompose

When --batch is present:

1. Read .plan-state/batch-discover.yaml or the current Feature candidate table.
2. Resolve batch.decompose_mode as project config > global config > default upfront.
3. If mode is upfront, decompose every Feature into initial Milestones before P3.
4. If mode is just_in_time, create Feature-level queue entries and leave Milestone arrays empty with JIT decomposition pending.
5. For upfront mode, produce:
   • a Feature Queue Markdown table
   • a Mermaid dependency graph
   • a Feature-level architecture impact section
6. Preserve single-feature /hw:plan behavior when --batch is absent.

Reference Files

• plan/PLAN-SKILL.md — Decompose phase rules
• references/commands-spec.md — command routing
• SKILL.md — broader planning context

Analysis Decomposition Notes

For workflow_kind: analysis, decompose by question rather than by implementation slice. Each analysis Milestone should be able to define a question, gather context, hypothesize, experiment, interpret, and conclude with a ledger-backed report.