[Bug] Variant body-level fields (e.g. chat_template_kwargs) are dropped before the request #42876

Description

Body-level fields declared in a model's variants entry never reach the HTTP request. The variant is merged into options correctly inside opencode, but the fields are lost before the wire.

I verified this with a logging reverse proxy between opencode and the server, so this is the actual request body, not an inference.

Config declares a variant carrying chat_template_kwargs (needed to enable thinking on Qwen3.8 via its chat template):

"variants": {
  "low": {
    "temperature": 1.0, "top_p": 0.95, "top_k": 20,
    "chat_template_kwargs": { "enable_thinking": true, "reasoning_effort": "low" }
  }
}

Agent uses "variant": "low". Captured request body:

{
  "model": "Qwen3.8-27B-oQ6e-mtp",
  "max_tokens": 32000,
  "top_p": 1,
  "stream": true
}

chat_template_kwargs, temperature and top_k are all gone; top_p is the hardcoded value from ProviderTransform.topP(), not the variant's.

Sending the same chat_template_kwargs straight to the same endpoint with curl works, so the server accepts it:

POST /v1/messages  {"chat_template_kwargs":{"enable_thinking":true,"reasoning_effort":"banana"}}
-> {"error":{"message":"Chat template error: Unexpected reasoning effort banana. ..."}}

As far as I can tell the loss happens at packages/opencode/src/provider/transform.ts:1405:

const key = sdkKey(model.api.npm) ?? (usesDotSplitOptions ? model.providerID.split(".")[0] : model.providerID)
return { [key]: normalized }

Options are handed to the AI SDK as providerOptions[key], and the adapter validates that object against its own schema, so keys it does not know are dropped. mergeOptions at session/llm/request.ts:91 does include the variant, so this is not a merge bug.

This may affect shipped functionality, not just custom providers. ProviderTransform.variants() (transform.ts:729) returns chat_template_kwargs for MiniMax-M3 on @ai-sdk/anthropic and @ai-sdk/openai-compatible:

return {
  none:     { chat_template_kwargs: { thinking_mode: "disabled" } },
  thinking: { chat_template_kwargs: { thinking_mode: "enabled"  } },
}

If the analysis above is right those built-in variants go the same way. I have not tested MiniMax-M3 myself, so I am flagging it rather than asserting it.

Previously reported in #42300 for @ai-sdk/openai-compatible; that issue was auto-closed for template non-compliance rather than on merit. This report adds a second adapter and a captured request body.

specs/v2/config.md:265 mentions agent-local options carrying body overrides, which would be a natural place to fix this.

Plugins

none (reproduced with a minimal standalone config)

OpenCode version

1.18.18

Steps to reproduce

1. Start any OpenAI/Anthropic-compatible server that accepts extra body fields. I used oMLX serving Qwen3.8-27B on 127.0.0.1:8001.

2. Put a logging proxy in front of it that dumps the request body and forwards to the server (any tiny HTTP proxy will do), listening on 127.0.0.1:8009.

3. Minimal config pointing at the proxy:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "px": {
      "npm": "@ai-sdk/anthropic",
      "options": { "baseURL": "http://127.0.0.1:8009/v1", "apiKey": "x" },
      "models": {
        "Qwen3.8-27B-oQ6e-mtp": {
          "reasoning": true,
          "limit": { "context": 262144, "output": 262144 },
          "variants": {
            "low": { "chat_template_kwargs": { "enable_thinking": true, "reasoning_effort": "low" } }
          }
        }
      }
    }
  },
  "model": "px/Qwen3.8-27B-oQ6e-mtp",
  "agent": { "probe": { "model": "px/Qwen3.8-27B-oQ6e-mtp", "variant": "low" } }
}

4. OPENCODE_CONFIG=<that file> opencode run --agent probe "hi"

5. Look at the proxy log: the body has no chat_template_kwargs.

Screenshot and/or share link

n/a — captured request bodies are inline above.

Operating System

macOS (Darwin 25.5.0), Apple Silicon

Terminal

n/a — reproduced headless with opencode run