bug: opencode-go gateway returns reasoning_content instead of content in streaming responses #37635

OpenCode Version: 1.18.3
OS: Windows 10
Terminal: PowerShell 7.6

Summary

The opencode-go gateway wraps ALL model streaming responses in a Chain-of-Thought wrapper, putting the assistant's actual output in delta.reasoning_content instead of delta.content. This causes the agent loop to receive empty content and exit immediately ("exiting loop" after step 1).

Steps to reproduce

1. opencode providers login > authenticate with opencode-go API key
2. opencode -m opencode-go/grok-4.5 (or any opencode-go model)
3. Send any message, e.g. "Say hello"
4. Observe: session exits after ~2s with no response

Root cause

Direct API test of the opencode-go gateway streaming endpoint:

POST https://opencode.ai/zen/go/v1/chat/completions
{"model":"grok-4.5","messages":[{"role":"user","content":"What is 1+1? Reply ONLY with the number."}],"max_tokens":200,"stream":true}

Actual output — every chunk uses reasoning_content, the actual answer appears only in the final chunk as a single content delta:

data: {"choices":[{"delta":{"role":"assistant","reasoning_content":"The"}}]}
data: {"choices":[{"delta":{"reasoning_content":" user"}}]}
data: {"choices":[{"delta":{"reasoning_content":" asks"}}]}
... (18 more reasoning_content-only chunks) ...
data: {"choices":[{"delta":{"content":"2"}}]}
data: {"choices":[{"finish_reason":"stop","delta":{}}]}

Token breakdown: reasoning_tokens=33, completion_tokens=1 — 97% consumed by the forced CoT wrapper.

The non-streaming endpoint returns content correctly — only streaming is affected.

Expected behavior

delta.content should contain the streaming response text, matching the non-streaming behavior.

Affected models

ALL opencode-go models: grok-4.5, deepseek-v4-pro, deepseek-v4-flash, kimi-k3, kimi-k2.7-code, kimi-k2.6, glm-5.1, glm-5.2, mimo-v2.5, mimo-v2.5-pro, minimax-m2.7, minimax-m3, qwen3.6-plus, qwen3.7-max, qwen3.7-plus

Additional observation

When requesting "model":"deepseek-v4-pro", the streaming response returns "model":"grok-4.5" in the chunk metadata — possible model routing issue.