desktop: model picker tooltip shows "No reasoning" for every model (reasoning capability hardcoded to false) #54195

Summary

Desktop 2.0.26: the model picker tooltip shows "No reasoning" for every model, even for models whose models.dev catalog entry has reasoning: true. The cause is an adapter in the desktop renderer that hardcodes reasoning: false (and temperature: false) when converting server model info into the internal catalog format, so the tooltip logic always reads the hardcoded false instead of real capability data.

Environment

• opencode version: 2.0.26
• OS: Windows_NT 10.0.26300 (win32 x64)
• Terminal: Unavailable: desktop app (renderer), terminal env vars not set
• Shell: C:\WINDOWS\system32\cmd.exe
• Install/channel: release (Desktop installer, @opencode-aidesktop)
• Active plugins: ./plugins-src/rate-limit-fallback (unrelated to this bug)

Reproduction

1. Open OpenCode Desktop 2.0.26 with any model provider connected (reproduced with the built-in opencode and nvidia/amd providers).
2. Open the model picker and hover a model to show its capability tooltip.
3. Observe the "Reasoning" row.

Expected Behavior

The tooltip should show "Allows reasoning" for models whose catalog entry has reasoning: true — e.g. opencode/mimo-v2.6-flash-free, nvidia/deepseek-ai/deepseek-v4.1-flash (which also exposes effort variants none/low/high/max), and amd/DeepSeek-V4.1-Flash. Local models.json cache confirms all three have "reasoning": true.

Actual Behavior

Every model shows "No reasoning" in the tooltip.

Two code sites in the bundled renderer (out/renderer/assets/main-*.js) are involved.

1. The adapter that maps server Model.Info to the internal catalog format hardcodes the flags:

t.models[e.id] = {
  id: e.id,
  providerID: e.providerID,
  api: { id: e.modelID, url: ``, npm: e.package ?? t.id },
  name: e.name,
  family: e.family,
  capabilities: {
    temperature: !1,          // hardcoded false
    reasoning: !1,            // hardcoded false — never reads catalog reasoning
    attachment: e.capabilities.input.some((e) => e !== `text`), // derived correctly
    toolcall: e.capabilities.tools,                             // derived correctly
    input: { /* derived from e.capabilities.input */ },
    output: { /* derived from e.capabilities.output */ },
    interleaved: !1,          // hardcoded false
  },
  // ...
};

2. The tooltip logic then trusts capabilities.reasoning whenever capabilities exists (which it always does, since the adapter always sets it):

s = () =>
  e.model.capabilities
    ? e.model.capabilities.reasoning
      ? t.t(`model.tooltip.reasoning.allowed`)  // "Allows reasoning"
      : t.t(`model.tooltip.reasoning.none`)     // "No reasoning"
    : e.model.reasoning
      ? t.t(`model.tooltip.reasoning.allowed`)
      : t.t(`model.tooltip.reasoning.none`)

Because the adapter always attaches capabilities with reasoning: false, the fallback branch (e.model.reasoning, which would be correct for catalog data) is never reached.

Note that the same adapter also hardcodes temperature: !1 and interleaved: !1, so those indicators are wrong for all models as well.

Additional Context

• Runtime verification: reasoning works fine despite the tooltip. A session with opencode/mimo-v2.6-flash-free shows "reasoning": 49 in token stats; an extended coding session on the same model accumulated "reasoning": 256662 tokens. The nvidia deepseek model exposes effort variants (none/low/high/max) that are passed through separately and remain switchable via "Cycle thinking effort", so this is purely a display bug in the tooltip path.
• Server /api/model output for these models includes compatibility.reasoningField: "reasoning_content" and (for nvidia) the effort variants, confirming the server-side data is correct.
• The Model.Info schema (@opencode/schema) no longer carries a top-level reasoning field — capabilities are { tools, input, output } — so the adapter has no runtime flag to read; deriving it (e.g. from variants/compatibility/provider catalog data) or surfacing the catalog reasoning value through the API would fix the tooltip.
• Reproducible every time, 100%.