models: V2 model metadata has no reasoning flag, so the picker always shows "No reasoning" #50279

Summary

The model picker's info panel always reports "不支持推理" / "No reasoning" for every model, including models that are known to support reasoning. V2 model metadata (GET /api/model) has no reasoning field at all, so the panel's capabilities.reasoning / model.reasoning check can never be true.

Environment

• opencode version: 2.0.11 (v2.0.11)
• OS: Linux 6.18.33.2-microsoft-standard-WSL2 (linux x64)
• Terminal: TERM=xterm-256color, TERM_PROGRAM=(unset), COLORTERM=(unset)
• Shell: /bin/bash
• Install/channel: npm global install of @opencode/cli (release)
• Active plugins: None

Reproduction

1. Configure a custom OpenAI-compatible provider whose model is known to support reasoning (this one is DeepSeek V4.1 Flash through an internal OpenAI-compatible gateway), with compatibility.reasoningField set:

   {
     "providers": {
       "turing": {
         "name": "Turing AI",
         "package": "@opencode/ai/providers/openai-compatible",
         "settings": { "baseURL": "https://<internal-gateway>/api/v1/" },
         "models": {
           "deepseek-v4.1-flash": {
             "name": "Deepseek V4.1 Flash",
             "capabilities": { "tools": true, "input": ["text", "image"], "output": ["text"] },
             "limit": { "context": 1000000, "output": 393216 },
             "compatibility": { "reasoningField": "reasoning_content" }
           }
         }
       }
     }
   }

2. Open the model picker and highlight the model. The info panel shows 推理: 不支持推理 (reasoning: not supported), while 上下文: 1,000,000 renders correctly.

3. Confirm the model actually reasons by calling it directly (streaming) and observing reasoning_content:

   curl https://<internal-gateway>/api/v1/chat/completions \
     -H "Content-Type: application/json" -H "Authorization: Bearer $KEY" \
     -d '{"model":"deepseek-v4.1-flash","messages":[{"role":"user","content":"1+1=?"}],"reasoning_effort":"high","stream":true}'
   # -> 35 chunks contain "reasoning_content":"..."

4. Inspect the model metadata served to the clients:

   opencode api get /api/model

Expected Behavior

The panel should report reasoning as supported for a model that returns reasoning_content and accepts a reasoning effort setting. Either the metadata exposes a reasoning flag, or the configuration provides a way to declare it.

Actual Behavior

GET /api/model returns 45 models and none of them expose any reasoning field. Every entry has capabilities: { tools, input, output } only. This includes catalog reasoning models, e.g. opencode-go/deepseek-v4.1-flash:

{
  "id": "deepseek-v4.1-flash",
  "providerID": "opencode-go",
  "compatibility": { "reasoningField": "reasoning_content" },
  "capabilities": { "tools": true, "input": ["text", "image"], "output": ["text"] },
  "variants": [{ "id": "low" }, { "id": "high" }, { "id": "max" }],
  "limit": { "context": 1000000, "output": 384000 }
}

models.dev marks this model as "reasoning": true, and the same panel therefore mislabels it as "No reasoning" too. There is no config field that changes this.

Additional Context

• In the 2.0.11 binary, the V2 model schema for capabilities is:

  Model.Capabilities = { tools: boolean, input: string[], output: string[] }

  and Model.Info has no reasoning field. Meanwhile the picker tooltip resolves the row as:

  model.capabilities
    ? (model.capabilities.reasoning ? "model.tooltip.reasoning.allowed" : "model.tooltip.reasoning.none")
    : (model.reasoning ? "model.tooltip.reasoning.allowed" : "model.tooltip.reasoning.none")

  Since V2 always provides a capabilities object without reasoning, the first branch wins and the row is always "none".

• Frequency: always reproducible; affects all providers, including official catalog providers such as OpenCode Go / OpenCode Zen.

• Workarounds tried:

  • Adding "reasoning": true on the model: silently dropped by the V2 schema (still absent from /api/model).
  • Adding "canonical": "opencode-go" to inherit catalog metadata: inherits family/variants/cost, but still no reasoning flag; the panel still shows "不支持推理".

• References:

  • OpenCode docs, Models: https://opencode.ai/v2/docs/models/
  • models.dev entry used for comparison: provider opencode-go, model deepseek-v4.1-flash (reasoning: true).