[bug] Anthropic thinking block is forwarded to openai_responses clients as reasoning.type (OpenAI 400) #1012

Summary

When an Anthropic Messages request that carries a thinking block is routed to an openai_responses client, the translator copies the Anthropic thinking object verbatim into the Responses reasoning field. OpenAI rejects the request:

400 {"error":{"message":"Unknown parameter: 'reasoning.type'.","type":"invalid_request_error","param":"reasoning.type","code":"unknown_parameter"}}

This breaks every Claude Agent SDK / Claude Code session pointed at an OpenAI reasoning model through Switchyard 0.3.0, because the SDK sends thinking ({"type":"adaptive"} on current models, {"type":"enabled","budget_tokens":N} on older ones) on every turn.

Where

• switchyard-translation/src/codecs/anthropic/buffered.rs:80 — ReasoningParams.raw is set to body.get("thinking").cloned() (the Anthropic-shaped object).
• switchyard-translation/src/codecs/responses/buffered.rs:260-270 — the encoder starts from request.reasoning.raw, inserts effort, and emits the result as reasoning. The Anthropic keys (type, budget_tokens) survive.
• The openai_chat encoder only emits reasoning_effort from request.reasoning.effort, so chat clients are unaffected.

Reproduce (0.3.0, switchyard-server, route → openai_responses client, model gpt-5.5)

curl -s http://127.0.0.1:4000/v1/messages -H 'content-type: application/json' -H 'anthropic-version: 2023-06-01' -d '{
  "model": "<route>", "max_tokens": 4000,
  "thinking": {"type": "adaptive"},
  "messages": [{"role": "user", "content": "Call the noop tool with kind=a, then say ok."}],
  "tools": [{"name": "noop", "description": "Probe tool.", "input_schema": {"type": "object", "properties": {"kind": {"type": "string"}}}}]
}'

→ 400 Unknown parameter: 'reasoning.type'. The same request with "output_config": {"effort": "medium"} instead of thinking succeeds (the effort is mapped to reasoning.effort correctly). thinking: {"type":"enabled","budget_tokens":2000} fails the same way.

Expected

raw from a different wire format should not be replayed into the Responses reasoning object. Either carry raw only when the source format was already openai_responses, or build the Responses reasoning from the normalized fields (effort; optionally summary) and drop unknown keys. Mapping Anthropic budget_tokens to a Responses effort tier would be a bonus but is not needed to unbreak clients.

Workaround we use

Set MAX_THINKING_TOKENS=0 on the Claude Agent SDK subprocess for routes that terminate on openai_responses, so the SDK omits the thinking block; the model still reasons at the effort passed through output_config.effort.

Version: switchyard-server 0.3.0 (cargo install --locked switchyard-server --version 0.3.0), static musl build.