Summary

Follow-up to #62 addressing this Copilot review comment: _strip_nonsemantic_system_lines() was only applied to the Responses API path. convert__anthropic_message_to_openai_chat__request() still forwarded request.system verbatim, so Anthropic->OpenAI Chat requests continued to carry the per-request unique x-anthropic-billing-header: ...; cch=<hash>; prefix and could still defeat caching/prompt stability for the chat-completions endpoint.

Changes

ccproxy/llms/formatters/anthropic_to_openai/requests.py

Apply _strip_nonsemantic_system_lines() to both the str and list[TextBlock] system paths in the chat converter, mirroring the Responses-API path. If the cleaned content is empty, the system message is dropped entirely instead of injecting an empty system role.

# Before: billing header forwarded verbatim into chat system content
sys_content = "".join(block.text for block in request.system)

# After: header stripped per-block
cleaned_blocks = (
    _strip_nonsemantic_system_lines(block.text or "")
    for block in request.system if block.text
)
sys_content = "".join(b for b in cleaned_blocks if b)

Tests

Added 5 chat-converter tests in tests/unit/llms/formatters/test_anthropic_to_openai_helpers.py mirroring the Responses-API coverage:

• test_chat_billing_header_stripped_from_string_system
• test_chat_billing_header_stripped_from_block_system
• test_chat_billing_header_only_system_drops_system_message
• test_chat_billing_header_stripped_case_insensitive
• test_chat_system_stable_across_requests_with_different_cch

Test plan

• uv run pytest tests/unit/llms/formatters/test_anthropic_to_openai_helpers.py -- 21 passed
• uv run mypy ccproxy/llms/formatters/anthropic_to_openai/requests.py -- clean
• uv run ruff check / uv run ruff format --check -- clean
• pre-commit hooks passed on commit