anthropic_messages_to_chat_completions

Transforms Anthropic Messages API requests to Chat Completions-compatible request bodies and transforms compatible responses back. The name refers to the Chat Completions wire shape, not the OpenAI Responses API; any Chat Completions-compatible backend is a valid target, not only OpenAI.

Configuration Notes

Request fields the translation does not map are forwarded untouched for the backend to validate. Fields whose effect the translated response could not report truthfully (service_tier, container, inference_geo, mcp_servers, and Chat Completions fields such as n or logprobs whose output the translated response would discard) are rejected with a 400. Unsupported semantic content is rejected because Chat Completions cannot represent it faithfully.

allow_lossy_features opts specific Anthropic-only features into operator-approved degradation instead of that 400. A listed feature’s wire markers are validated and then stripped so an unmodified client (for example Claude Code, which always sends cache_control and thinking) can still drive a Chat Completions backend. Each degradation is reported to the operator — one WARN log per request, a praxis_anthropic_messages_to_chat_completions_degraded_total counter per feature, and an x-degraded-features response header — because the request succeeds but the feature was silently dropped. prompt_caching removes cache_control markers (prompt and tool content are preserved, but explicit cache breakpoints are not honored, so cost and latency may differ); extended_thinking removes thinking and thinking-only context_management edits (the translated response carries no thinking blocks). Features absent from the allowlist, malformed markers, and other context_management edits are still rejected with a 400.

Configuration

Examples

Example 1

filter: anthropic_messages_to_chat_completions

Example 2

filter: anthropic_messages_to_chat_completions
max_body_bytes: 1048576
allow_lossy_features:
  - prompt_caching
  - extended_thinking