Filter 16: responses_to_chat_completions

[leseb]

Purpose

Convert enriched Responses API requests to Chat Completions format for backends that only speak /v1/chat/completions. This is the conversion counterpart to responses_proxy (#432), which forwards to native /v1/responses backends.

Behavior

• Read responses.messages, responses.tools, responses.tool_choice from filter chain metadata
• Build a /v1/chat/completions request body from Responses API state:
  • Map messages to chat completion message format
  • Map tool definitions to chat completion function/tool format
  • Map parameters: temperature, top_p, frequency_penalty, max_completion_tokens, reasoning_effort, service_tier, parallel_tool_calls, top_logprobs, presence_penalty, extra_body
  • Respect client's stream setting
  • Set stream_options.include_usage=true when streaming
  • Omit text response_format when tools are present
  • Pass prompt_cache_key if configured
• Inject instructions as system message
• Map provider errors → Responses API error codes
• Structure conversion as a standalone module (convert.rs) with clear input/output types for reuse by Provider abstraction: unified request/response types #213

Config

filter: responses_to_chat_completions

Dependencies

• Upstream inference backend (/v1/chat/completions)
• Related: Provider abstraction: unified request/response types #213 (Provider abstraction)
• Downstream: stream_events (Filter 6: stream_events #433) must handle Chat Completion chunks → Responses API SSE transformation

Reference

• Proposal: Responses API Filters — Filter 5 (chat_completions mode)
• Part of Epic: Responses API and Agentic Loop Orchestration #354

[leseb]

@franciscojavierarceo do you want to add sub-issues here for #702 and #703?