plan（计划）

PLAN MODE: Five-Phase Planning Process

WARNING: DO NOT use the built-in ExitPlanMode or EnterPlanMode tools. This project has its own planning workflow using /plan, /implement, and /verify skill commands that are automatically invoked via the /spec slash-command. The built-in Claude Code plan mode tools write to different paths and are incompatible. When planning is complete, simply inform the user and wait for confirmation - no special tool needed.

Using AskUserQuestion - Core Planning Tool

The AskUserQuestion tool is essential for effective planning. Questions are grouped into batches so users answer related questions together without interruption.

Question Batching Strategy

Questions are consolidated into two batches for a smooth user experience:

Don't scatter questions across phases - gather them and present together.

When to Use AskUserQuestion

How to Structure Questions

Use the AskUserQuestion tool with clear options:

Question: "Which authentication approach should we use?"
Options:
- Option A: JWT tokens (stateless, scalable, requires token refresh logic)
- Option B: Session-based (simpler, requires session storage)
- Option C: OAuth integration (most secure, more complex setup)

Key principles:

• Present options, not open-ended questions when possible
• Include trade-offs for each option
• Batch related questions together - don't interrupt user flow
• Don't proceed with assumptions - ASK

Extending Existing Plans

When adding tasks to existing plan:

1. Load existing plan: Read(file_path="docs/plans/...")
2. Parse structure (architecture, completed tasks, pending tasks)
3. Check git status for partially completed work
4. Verify new tasks are compatible with existing architecture
5. Check total: If original + new > 12 tasks, suggest splitting
6. Mark new tasks with [NEW] prefix
7. Update total count: Total Tasks: X (Originally: Y)
8. Add extension history: > Extended [Date]: Tasks X-Y for [feature]

⚠️ CRITICAL: Migration/Refactoring Tasks

When the task involves migrating, refactoring, or replacing existing code, you MUST complete these additional steps to prevent missing features.

Mandatory Feature Inventory (Phase 1.5)

After exploration but BEFORE creating tasks:

1. List ALL files being replaced: ` ## Feature Inventory - Files Being Replaced | Old File | Functions/Classes | Status | |----------|-------------------|--------| | old/module1.py | func_a(), func_b(), ClassX | ⬜ Not mapped | | old/module2.py | func_c(), func_d() | ⬜ Not mapped | `
2. Map EVERY function/feature to a new task: ` ## Feature Mapping - Old → New | Old Feature | New Location | Task # | |-------------|--------------|--------| | module1.func_a() | new/step1.py | Task 3 | | module1.func_b() | new/step1.py | Task 3 | | module2.func_c() | new/step2.py | Task 5 | | module2.func_d() | ❌ MISSING | ⚠️ NEEDS TASK | `
3. Verify 100% coverage before proceeding:

- Every row must have a Task # or explicit "Out of Scope" justification - "Out of Scope" means the feature is INTENTIONALLY REMOVED (with user confirmation) - "Out of Scope" does NOT mean "migrate as-is" - that still needs a task!

"Out of Scope" Clarification

CRITICAL: "Out of Scope" has a precise meaning:

When in doubt, ASK THE USER:

"The old code has [feature]. Should we:
A) Migrate it as-is (needs implementation task)
B) Intentionally remove it (truly out of scope)
C) Improve it (new feature, needs implementation task)"

Pre-Task Verification Gate

Before finalizing Phase 3 (Implementation Planning), verify:

• All old files listed in Feature Inventory
• All functions/classes from old files identified
• Every feature mapped to a task OR explicitly marked "REMOVED" with user confirmation
• No row in Feature Mapping has "⬜ Not mapped" status
• User has confirmed any features marked for removal

If any checkbox is unchecked, DO NOT proceed to Phase 4.

Creating New Plans

Phase 0: Task Understanding & Clarification

First, clearly state your understanding of the task.

Before any exploration:

1. Restate what the user is asking for in your own words
2. Identify the core problem being solved
3. List any assumptions you're making

Then gather all clarifications needed (Question Batch 1):

Use AskUserQuestion to ask everything upfront in a single interaction:

I want to confirm my understanding and clarify a few things:

Questions:
1. "Is my understanding correct?"
   - Yes, that's correct
   - Partially correct (please clarify)
   - No, let me explain

2. "What's the priority for this feature?"
   - Performance (fast, optimized)
   - Simplicity (easy to maintain)
   - Flexibility (extensible)

3. "Should error handling be comprehensive or minimal?"
   - Comprehensive (recovery, retries, detailed logging)
   - Basic (fail fast, simple error messages)

Don't proceed to exploration until clarifications are complete.

Phase 1: Exploration

Explore the codebase systematically. Run explorations one at a time (sequentially, not in parallel).

🔧 Tools for Exploration

Use these tools to gather context efficiently:

Context7 requires descriptive queries for both tools - see context7-docs.md for full docs.

Before exploring, verify Vexor is available:

vexor --version

First search will trigger automatic indexing. Use vexor search "query" --mode code for semantic code search.

Exploration areas (in order):

1. Architecture - Project structure, entry points, how components connect
2. Similar Features - Existing patterns that relate to the task, what can be reused
3. Dependencies - Imports, modules, what will be impacted
4. Tests - Test infrastructure, existing patterns, available fixtures

⚠️ CRITICAL: NO SUB-AGENTS DURING PLANNING

• DO NOT use the Task tool with any subagent_type during planning
• Perform ALL exploration yourself using direct tool calls (Read, Grep, Glob, Context7)
• Sub-agents lose context and make planning inconsistent
• You must maintain full context throughout the planning process

For each area:

• Use vexor search "query" --mode code for semantic searches like "authentication middleware" or "database connection handling"
• Use Context7 when you need library/framework API details: resolve-library-id(query="your question", libraryName="lib") then query-docs(libraryId, query) - descriptive queries enable server-side reranking
• Use Read, Grep, Glob tools directly for file exploration
• Document hypotheses (not conclusions)
• Note full file paths for relevant code
• Track questions that remain unanswered

After explorations complete:

1. Read each identified file to verify hypotheses
2. Build a complete mental model of current architecture
3. Identify integration points and potential risks
4. Note reusable patterns

Phase 2: Design Decisions

Present findings and gather all design decisions (Question Batch 2).

Summarize what you found, then use AskUserQuestion with all decisions at once:

Based on my exploration, I found [summary of key findings].

I have some design questions:

Questions:
1. "Which architecture approach should we use?"
   - Option A: [description + trade-offs]
   - Option B: [description + trade-offs]
   - Option C: [description + trade-offs]

2. "How should we handle [specific concern found during exploration]?"
   - Option A: [approach]
   - Option B: [approach]

3. "The existing [component] uses [pattern]. Should we follow it or try something different?"
   - Follow existing pattern (consistency)
   - Use new approach (explain why)

After user answers:

• Summarize the chosen design approach
• Confirm: "Does this design work for your needs?"
• Don't proceed until design is validated

Phase 3: Implementation Planning

Task Count Guidance

• Aim for 10-12 tasks as a general guideline for focused plans
• Avoid bloating plans with unnecessary or overly granular tasks
• If the work genuinely requires more tasks, that's fine - the /spec workflow handles multi-session execution, so longer plans will be executed across multiple context windows automatically
• Focus on keeping tasks meaningful and necessary, not on hitting a specific number

Task Structure:

### Task N: [Component Name]

**Objective:** [1-2 sentences describing what to build]

**Files:**
- Create: `exact/path/to/file.py`
- Modify: `exact/path/to/existing.py`
- Test: `tests/exact/path/to/test.py`

**Implementation Steps:**
1. Write failing test - Define expected behavior
2. Implement minimal code - Make test pass
3. Verify execution - Run actual program
4. Integration test - Test with other components

**Definition of Done:**
- [ ] All tests pass (unit, integration if applicable)
- [ ] No diagnostics errors (linting, type checking)
- [ ] Code functions correctly with real data
- [ ] Edge cases handled appropriately
- [ ] Error messages are clear and actionable

Zero-context assumption:

• Assume implementer knows nothing about codebase
• Provide exact file paths
• Explain domain concepts
• List integration points
• Reference similar patterns in codebase

Phase 4: Documentation

Save plan to: docs/plans/YYYY-MM-DD-<feature-name>.md

Required plan template:

# [Feature Name] Implementation Plan

Created: [Date]
Status: PENDING
Approved: No
Iterations: 0

> **Status Lifecycle:** PENDING → COMPLETE → VERIFIED
> **Iterations:** Tracks implement→verify cycles (incremented by /verify)
> - PENDING: Initial state, awaiting implementation
> - COMPLETE: All tasks implemented (set by /implement)
> - VERIFIED: Rules supervisor passed (set automatically)
>
> **Approval Gate:** Implementation CANNOT proceed until `Approved: Yes`
> - Claude will ask for your approval after presenting the plan
> - You can request changes before approving
> - Claude updates this field automatically when you approve

## Summary
**Goal:** [One sentence describing what this builds]

**Architecture:** [2-3 sentences about chosen approach]

**Tech Stack:** [Key technologies/libraries]

## Scope

### In Scope
- [What WILL be changed/built]
- [Specific components affected]

### Out of Scope
- [What will NOT be changed]
- [Explicit boundaries]

## Prerequisites
- [Any requirements before starting]
- [Dependencies that must exist]
- [Environment setup needed]

## Context for Implementer
- [Key codebase convention or pattern]
- [Domain knowledge needed]
- [Integration points or dependencies]

## Feature Inventory (FOR MIGRATION/REFACTORING ONLY)

> **Include this section when replacing existing code. Delete if not applicable.**

### Files Being Replaced

| Old File | Functions/Classes | Mapped to Task |
|----------|-------------------|----------------|
| `old/file1.py` | `func_a()`, `func_b()` | Task 3 |
| `old/file2.py` | `ClassX`, `func_c()` | Task 4, Task 5 |

### Feature Mapping Verification

- [ ] All old files listed above
- [ ] All functions/classes identified
- [ ] Every feature has a task number
- [ ] No features accidentally omitted

**⚠️ If any feature shows "❌ MISSING", add a task before implementation!**

## Progress Tracking

**MANDATORY: Update this checklist as tasks complete. Change `[ ]` to `[x]`.**

- [ ] Task 1: [Brief summary]
- [ ] Task 2: [Brief summary]
- [ ] ...

**Total Tasks:** [Number] | **Completed:** 0 | **Remaining:** [Number]

## Implementation Tasks

### Task 1: [Component Name]
[Full task structure from Phase 3]

### Task 2: [Component Name]
[Full task structure from Phase 3]

## Testing Strategy
- Unit tests: [What to test in isolation]
- Integration tests: [What to test together]
- Manual verification: [Steps to verify manually]

## Risks and Mitigations

| Risk | Likelihood | Impact | Mitigation |
|------|------------|--------|------------|
| [Risk 1] | Low/Med/High | Low/Med/High | [How to mitigate] |

## Open Questions
- [Any remaining questions for the user]
- [Decisions deferred to implementation]

Phase 5: Implementation Handoff

⛔ MANDATORY APPROVAL GATE - This is NON-NEGOTIABLE

After saving plan:

1. Summarize the plan - Provide a brief overview of:

- What will be built (goal) - Key tasks (numbered list) - Tech stack / approach

1. Use AskUserQuestion to request approval: Question: "Do you approve this plan for implementation?" Header: "Plan Review" Options: - "Yes, proceed with implementation" - I've reviewed the plan and it looks good - "No, I need to make changes" - I want to edit the plan first
2. Based on user response: If user approves ("Yes, proceed..."): If user wants changes ("No, I need to make changes"):

- Edit the plan file to change Approved: No to Approved: Yes - Tell user: "Plan approved. Proceeding with implementation..." - EXIT /plan immediately - /spec will invoke /implement - Do NOT write any implementation code - that's /implement's job - Tell user: "Please edit the plan file at <plan-path>, then say 'ready' when done" - Wait for user to confirm they're done editing - Re-read the plan file to see their changes - Ask for approval again using AskUserQuestion

1. DO NOT proceed to implementation until user explicitly approves

⚠️ CRITICAL: Claude handles the Approved: field update - user never edits it manually DO NOT write or edit any implementation files until approved.

Critical Rules

These rules are non-negotiable:

1. NEVER use sub-agents (Task tool) during planning - Perform all exploration yourself with direct tool calls
2. USE AskUserQuestion when uncertain - Don't guess, ask the user
3. Batch questions together - Don't interrupt user with scattered questions
4. Run explorations sequentially - One at a time, never in parallel
5. NEVER write implementation code during planning - Planning and implementing are separate
6. NEVER assume - verify by reading files - Hypotheses must be confirmed
7. ⛔ NEVER proceed without user approval - Use AskUserQuestion to get explicit approval, then update Approved: Yes yourself
8. ALWAYS re-read the plan after user edits - If they chose to make changes, re-read before asking again
9. The plan must be detailed enough that another developer could follow it
10. NEVER use built-in ExitPlanMode or EnterPlanMode tools - This project uses custom /plan, /implement, /verify skill commands. The built-in plan mode tools are incompatible with this workflow. They are orchestrated by the /spec slash command.
11. FOR MIGRATIONS: Create Feature Inventory BEFORE tasks - List every file, function, and class being replaced. Map each to a task. No unmapped features allowed.
12. "Out of Scope" ≠ "Don't implement" - "Out of Scope: Changes to X" means migrate X as-is (still needs a task). Only "Out of Scope: Remove X" means no task needed (requires user confirmation).
13. Ask for approval after creating plan - After Phase 5, use AskUserQuestion to get approval. If approved, update the file and let /spec continue automatically.