Conversations hang after first response when provider returns a reasoning field (using Ollama/Qwen3)

[fenrir-labs76]

Description

When using local thinking models on Ollama (e.g. Qwen3.6-35b), the model appears
to think but then stops and returns no response.

The @ai-sdk/openai-compatible SDK extracts the reasoning field from the response and re-sends it as reasoning_content in message history on the next turn. Ollama does not accept reasoning_content in
message history, causing the conversation to hang with no output returned to the user.

Plugins

None

OpenCode version

1.17.13

Steps to reproduce

1. Configure an Ollama provider with a thinking model (e.g. Qwen3.6-35b)
2. Start a session and send a message, like: "Explain in detail how distributed consensus algorithms work. Compare PBFT, Tendermint, and Raft protocols, their tradeoffs, fault tolerance models, and real-world applications. Walk through a concrete example of message ordering in each protocol."
3. Note that opencode just ends - as if it responded to you.

Screenshot and/or share link

No response

Operating System

Ubuntu 24.04 LTS

Terminal

Ghostty

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Cerebras zai-glm-4.7 fails on follow-up turn with reasoning_content #26762: Same root cause — reasoning_content in assistant message history rejected on follow-up turns by providers that don't support it (Cerebras)
• DeepSeek V4: reasoning_content 400 error persists on v1.14.31 — complete fix exists in unmerged PRs #25311: Same root cause — reasoning_content must/must-not be replayed in history depending on the provider; multi-turn failures with OpenAI-compatible providers
• Bug: 400 Error (Extra inputs not permitted) when sending 'reasoning' field in chat history for kimi-k2.6 #27852: Same root cause — 400 errors when reasoning/reasoning_content fields from prior turns are sent back in history to providers that reject them

The underlying issue (the AI SDK's @ai-sdk/openai-compatible re-injecting reasoning_content into subsequent message history for providers that don't accept it) appears to be the same problem described in those issues, just observed with Ollama instead of Cerebras/DeepSeek/Kimi.

[twhittock]

This breaks all usage of reasoning models in my ollama-based environment - I am running a patched build (#36068) locally to work around this.