defog CLAUDE.md

General Behaviour

• Do not be obsequious. Do not say "You are absolutely right" or "You are right". You do not have to please the user – you just need to make the code work.
• Do NOT use emotional words or exclaimations like "Perfect!", or "Great!". Be more stoic

Updating documentation

• After you implement or change code, always remember to check if you need to update documentation. Documentation is in README.md and in the docs/ folder

Linting and Checking

• Use ruff check to make sure your implementations are not broken
• Use ruff format for code formatting

Reviewing PRs

• Be honest, technically focused, and blunt. Pay special attention to security issues.

Using LLMs

• Recall that this is 2025 and that new LLM models have been released.
  • Claude 4 Sonnet and Claude 4 Opus are now valid models
  • OpenAI's gpt-4.1, o3, and o4-mini are now valid models
  • Gemini's 2.5 flash and 2.5 pro are now valid models
• Recall that you should consider using the chat_async function and its associated parameters (including response_format, reasoning_effort, and tools) - instead of using LLM clients directly or rolling your JSON parsing functions for structure outputs

Testing

• Remember to run all tests with python -m pytest .... This is to ensure you are using the correct version of python and pytest
• Add PYTHONPATH=. when running tests, so that the tests/examples use the version of defog in this repo - instead of the machine installed version
• Use the --envfile .env argument to load environment variables (API keys etc.) when running tests
• Try to run scoped tests that directly affect changed code instead of running all tests at once

Model Pricing

• Prices live in defog/llm/cost/models.py as per-1k-token rates.
• When adding or updating an entry, use ONLY the provider's official pricing page:
  • OpenAI: https://developers.openai.com/api/docs/pricing
  • Anthropic (Claude): https://platform.claude.com/docs/en/about-claude/models/overview
• Convert the published per-1M-token numbers to per-1k (divide by 1000). Keys are input_cost_per1k, cached_input_cost_per1k, output_cost_per1k, and (for Anthropic) cache_creation_input_cost_per1k.
• Do not infer prices from neighbouring entries or prior versions. A newer tier name (e.g. gpt-5.4-mini) does not necessarily share pricing with the matching gpt-5-mini entry — look each model up on the source page.