DeepSeek Python Development Rules

## DeepSeek Python Development Rules

### Project Structure
```
project/
  src/
    deepseek_client.py    # API wrapper with retry logic
    models.py             # Pydantic models for request/response
    prompts/              # Prompt templates as .txt or .j2 files
    utils.py              # Token counting, text processing
  tests/
    test_client.py
    test_prompts.py
    fixtures/             # Sample API responses for testing
  .env                    # API keys (gitignored)
  pyproject.toml
```

### Code Standards
- Python 3.10+ with type hints on all functions
- Use Pydantic v2 for request/response models
- Use httpx for async HTTP calls (not requests)
- Use tenacity for retry logic
- Format with ruff, type-check with mypy

### API Client Pattern
- Create a singleton client with connection pooling
- Use async/await for all API calls
- Implement proper timeout handling (connect: 5s, read: 120s)
- Parse streaming responses with async generators
- Cache identical prompts with functools.lru_cache or Redis

### Prompt Management
- Store prompts as separate template files, not inline strings
- Use Jinja2 templating for dynamic prompt construction
- Version prompts alongside code
- Log prompt templates (not filled values) for debugging
- A/B test prompt variants with feature flags

### Testing
- Mock API calls in unit tests (use respx or pytest-httpx)
- Create fixtures from real API responses
- Test edge cases: empty responses, malformed JSON, timeouts
- Integration tests against the real API in CI (with budget limits)
- Measure and track prompt quality metrics over time