Plan

Respond in the user's preferred language (detect from their recent messages, or fall back to the language setting in CLAUDE.md).

Generate a technical implementation plan for Unity. This is NOT a game design document — it translates an existing design into an engineering plan that /qq:execute can consume.

When the plan can use live Unity editing instead of code: some implementation steps are simpler as direct tykit calls (scene tweaks, prefab overrides, UI adjustments, one-off data fixes) rather than writing C# utilities. If a step is a one-shot editor-state change, see shared/tykit-reference.md and mark the step as "execute via tykit command X" in the plan. Reserve code-writing for changes that need version control, compile-time validation, or repeatable behavior.

Arguments: $ARGUMENTS

• A file path to a game design document
• A brief description (1-2 sentences) of what to build
• No arguments: check conversation context for a recent design discussion

1. Understand the Input

Best case: user provides a game design document (from Docs/qq/, Docs/design/, Notion export, or inline). Read it fully, extract the technical requirements.

Minimal case: user gives a one-liner like "add a health system" or "weapons need ammo reloading". Ask 3-5 targeted technical questions before proceeding:

• What existing systems does this interact with? (or: let me explore the codebase to find out)
• Any data format preferences? (ScriptableObject, CSV config, etc.)
• Any hard constraints? (no singletons, must work with existing event bus, etc.)

Do NOT ask more than 5 questions. If something is unclear, explore the codebase to find the answer yourself. Prefer reading code over asking the user.

2. Explore the Codebase

Before writing the plan, understand what already exists:

• Read CLAUDE.md for coding standards
• Read AGENTS.md for architecture layers and module boundaries (if it exists)
• Explore the relevant directories (Assets/Scripts/, service modules, existing interfaces)
• Check .asmdef structure to understand module boundaries
• Identify existing patterns the new code should follow (event bus, service locator, dependency injection, etc.)

This step is critical — do not design in a vacuum.

2.5. Cross-cutting Seams (跨切面接缝)

If .claude/seams.yml exists, this project maintains a registry of known fan-out points — places where adding one thing (a new enum value / interface impl / WorkType / MonsterType / event / registration / config row) requires synchronized edits in several other places. Miss one and it compiles clean but breaks or silently no-ops at runtime (classic: "added a WorkType but forgot the second switch").

For every such addition the plan introduces, match it against the keys in .claude/seams.yml and copy the matched seam's sites[].grep commands into the plan's 跨切面接缝清单 section (template below). Read seams.yml as the deterministic seed — do not enumerate seam points from memory (the model forgets/invents; the file does not). Newly discovered seams should be appended to .claude/seams.yml, not buried in one plan.

If .claude/seams.yml is absent, skip this section (no regression).

3. Write the Plan

Output a single markdown document following this format. Keep it concise — 1-3 pages max. No filler.

# [Feature Name] — Implementation Plan

## Goal
One sentence. What technical capability is added.

## Architecture
```mermaid
graph LR
    A[ComponentA] --> B[ComponentB]
    B --> C[ComponentC]

Key Types

Interfaces

public interface IFoo {
    void DoSomething(SomeEvent e);
    float Value { get; }
    event Action<float> OnValueChanged;
}

Data Schema

Any new config fields, serialized data, or save structures. Use actual field names and types.

Steps

Ordered, each step is a shippable increment. Include:

• Exact file paths (create or modify)
• What to implement
• Dependencies on previous steps
• Done criteria (how to verify this step works)

1. Create IFoo interface — Assets/Scripts/Systems/IFoo.cs

   • Define the contract shown above
   • No deps
   • Done: compiles

2. Implement FooSystem — Assets/Scripts/Systems/FooSystem.cs

   • MonoBehaviour implementing IFoo
   • [SerializeField] private fields for config
   • Depends on: step 1
   • Done: compiles + can attach to GameObject

3. Wire into existing BarSystem — Assets/Scripts/Systems/BarSystem.cs

   • Add IFoo dependency, call on trigger
   • Depends on: step 1, 2
   • Done: compiles + integration test passes

4. Tests — Assets/Tests/EditMode/FooSystemTests.cs

   • Test damage calculation, edge cases (zero, negative, overflow)
   • Depends on: step 2
   • Done: /qq:add-tests can implement this coverage without ambiguity, then all tests green

跨切面接缝清单 (Cross-cutting Seams)

Only when .claude/seams.yml exists. Seeded from it — one row per fan-out point this change touches. Omit the whole section if the change introduces no new enum/registration/event/config-row.

Constraints

• What NOT to do (anti-patterns to avoid)
• Assembly definition placement
• Execution order dependencies
• Existing systems that must not break

Testing Strategy

• EditMode: [what pure logic to test]
• PlayMode: [what integration to test]

Open Questions

• Anything unresolved that might change the plan

## 4. Save the Plan

Get branch name: `git branch --show-current | tr '/' '_'`

Save to `Docs/qq/<branch-name>/<feature-name>_implementation.md`.

## 5. Record Decisions

After saving the plan, record key technical decisions:
```bash
qq-decisions.py add --project . --phase plan --key "<decision>" --value "<choice>" --reason "<why>"

Record architecture choices, pattern decisions, key interface designs.

6. Handoff

Plan review is mandatory before execution. Do NOT offer /qq:execute directly.

First, check if Codex CLI is available by running which codex 2>/dev/null || where codex 2>/dev/null.

• Codex available → recommend /qq:codex-plan-review (cross-model review catches blind spots that same-model review misses)
• Codex not available → recommend /qq:claude-plan-review

--auto mode: run qq-execute-checkpoint.py pipeline-advance --project . --completed-skill "/qq:plan" --next-skill "/qq:codex-plan-review" --plan-doc "<saved-plan-path>", then run the check and invoke the appropriate review skill with --auto.

Self-Review (REQUIRED before saving)

Before saving the plan, verify:

1. File paths: Every step has exact file paths (create or modify), not descriptions
2. Step size: Each step touches 1-3 files, not more. If a step is too big, split it.
3. Dependencies: The depends-on chain is correct — no step uses something not yet created
4. Compile independence: Each step compiles on its own after implementation
5. Interface signatures: Actual code signatures are written, not prose descriptions
6. No placeholders: No "TBD", "TODO", "implement later", or "similar to step N"

If any check fails, fix the plan before saving.

Notes

• The plan must be consumable by /qq:execute — ordered steps with file paths and done criteria
• Test steps must be concrete enough that /qq:add-tests can implement them without re-planning
• Write actual interface signatures in the plan, not prose descriptions
• Use Mermaid for architecture diagrams (GitHub renders them)
• If the design doc is ambiguous, call it out in Open Questions — don't guess silently
• Follow existing project patterns. If the project uses a service container, use it. If it uses events, use events. Don't introduce new patterns unless the design requires it.
• When facing a non-trivial technical decision (e.g., choosing a pathfinding algorithm, structuring a state machine), invoke /qq:tech-research to search for proven approaches before committing to one in the plan.
• Concise over comprehensive. A 1-page plan that an engineer can follow beats a 10-page plan nobody reads.