トークン効率戦略

┌──────────────────────────────────────────────────────────┐
│                    Token Consumption                      │
├────────────────┬────────────────┬────────────────────────┤
│   Context      │   Tool Calls   │   Model Output         │
│   (Input)      │   (Overhead)   │   (Response)           │
├────────────────┼────────────────┼────────────────────────┤
│ • AGENTS.md    │ • Read files   │ • Explanations         │
│ • Memories     │ • Grep/search  │ • Generated code       │
│ • Conversation │ • Execute cmds │ • Analysis             │
│ • File content │ • Each retry   │ • Thinking tokens      │
└────────────────┴────────────────┴────────────────────────┘

## In your AGENTS.md

## Testing
- Run single file: `npm test -- path/to/file.test.ts`
- Run fast smoke tests: `npm test -- --testPathPattern=smoke`
- Full suite takes ~3 minutes, use `--bail` for early exit on failure

## In your AGENTS.md

## Validation Commands
- Lint (auto-fix): `npm run lint:fix`
- Type check: `npm run typecheck`
- Full validation: `npm run validate` (lint + typecheck + test)

Always run `npm run lint:fix` after making changes.

## In your AGENTS.md

## Project Structure
- `src/components/` - React components (one per file)
- `src/hooks/` - Custom React hooks
- `src/services/` - API and business logic
- `src/types/` - TypeScript type definitions
- `tests/` - Test files mirror src/ structure

When adding a new component:
1. Create component in `src/components/ComponentName/`
2. Add index.ts for exports
3. Add ComponentName.test.tsx in same directory

droid
> /readiness-report

Simple edit, formatting      → Haiku 4.5 (0.4×)
Implement feature from spec  → GPT-5.1-Codex (0.5×)
Debug complex issue          → Sonnet 4.5 (1.2×)
Architecture planning        → Opus 4.7 (2×, 4月30日まで1×) or Opus 4.6 (2×)
Bulk file processing         → Droid Core (GLM-4.7 at 0.25× or GLM-5 at 0.4×)

Turn 1: Start implementing → wrong approach → wasted tokens
Turn 2: Undo and try different approach → more tokens
Turn 3: Finally get it right
Total: 3 turns of implementation tokens

Turn 1: Plan with exploration → correct approach identified
Turn 2: Implement correctly
Total: 1 turn of planning + 1 turn of implementation

Read file A → Read file B → Read file C → Understand context → Work
(4 tool calls before actual work)

Work (IDE provides open files, errors, selection)
(0 extra tool calls for context)

"Fix the bug in the auth module"

"Fix the timeout bug in src/auth/session.ts line 45 where the session expires after 5 minutes instead of 24 hours"

Turn 1: "Add logging to userService"
Turn 2: "Add logging to orderService"
Turn 3: "Add logging to paymentService"
(3 turns, context rebuilt each time)

Turn 1: "Add structured logging to all services in src/services/. Use the pattern from src/lib/logger.ts. Services: user, order, payment."
(1 turn, pattern established once)

"Add the feature. Return only the changed code, no explanations unless something is unclear."

"Review this code. Format: bullet list of issues only, no preamble."

"Debug this test failure. Show me the fix, then explain in 2-3 sentences."

droid
> /cost

1. Set up your project
   └─ AGENTS.md with commands
   └─ Fast tests
   └─ Linting configured
   └─ IDE plugin installed

2. Start each task right
   └─ Use Spec Mode for complex work
   └─ Be specific about the goal
   └─ Reference existing patterns

3. Choose the right model
   └─ Simple → Haiku/Droid Core
   └─ Standard → Codex/Sonnet
   └─ Complex → Opus (with reasoning)

4. Monitor and adjust
   └─ Check /cost periodically
   └─ Identify expensive patterns
   └─ Refine your approach