fix: forward reasoning_effort on Bedrock Runtime GPT
✨ What changed
- Bedrock Runtime GPT profiles (
us.openai.gpt-6-lunaetc.) now get the caller'sreasoning_effortforwarded instead of dropped. - When the request has function tools and no effort, Manifest sends
reasoning_effort: 'none'. - Non-GPT Runtime models (Kimi K3) are unchanged.
💭 Why
Two Manifest-side gaps made every tool call to these models fail with a 400. Manifest stripped reasoning_effort on this route, and Bedrock Runtime defaults these models to reasoning, which it rejects next to function tools. OpenAI's Chat Completions default for the same models is none, so the same body worked on OpenAI and failed on Bedrock.
Now the same body behaves the same on both. An explicit effort with tools is the caller's choice and fails the same way it does on OpenAI, where Autofix handles it.
👤 For users
Agents using tools on Bedrock GPT CRIS profiles through /v1/chat/completions or /v1/messages work again. A reasoning_effort you set on these models now reaches Bedrock.
📝 Notes
Fixes #3065
Not run against live Bedrock: our AWS account has no OpenAI model entitlement (403). Bedrock's own error asks for reasoning_effort: 'none', and the same fix heals this error on direct OpenAI in production.
#3066 is the alternative: it translates these calls to the Responses API to keep reasoning with tools.