# Runtime deps for the vendored SkillOpt + whetstone env.
# Verified by import smoke test (see README.md / VENDORED.md).
#
# The optimizer/judge backend (claude_chat) and the target backend
# (claude_code_exec) both shell out to the `claude` CLI — no Anthropic SDK
# dependency. `openai` is only pulled because skillopt.model eagerly imports its
# azure backend at package load; the azure path is never exercised by our config.
pyyaml>=6.0
numpy>=1.24.0
openai>=1.30.0
httpx>=0.27.0
pytest>=7.0          # deterministic `hard` check runs the fixture test suite
