app: model picker tooltip always shows "No reasoning" for V2 catalog models #55906

Summary

In the V2 app model picker, the hover tooltip's Reasoning row always reads "No reasoning" / 「不支持推理」, even for models that clearly support reasoning. For example opencode/deepseek-v4.1-flash-free shows "No reasoning" in the tooltip, yet still exposes low / high / max thinking-effort variants and streams a thinking block. The tooltip reads a capability field that V2 no longer populates.

Environment

• opencode version: desktop app 2.0.30 (C:\Users\liche\AppData\Local\Programs\@opencode-aidesktop\OpenCode.exe); the npm CLI on PATH is a separate opencode-ai@1.18.31
• OS: Windows_NT 10.0.26200 (win32 x64)
• Terminal: xterm-256color (bug is in the app UI, not the terminal)
• Shell: C:\WINDOWS\system32\cmd.exe
• Install/channel: latest (desktop app installer)
• Active plugins: none (global config only sets $schema)

Reproduction

1. Open the app model picker (/models).
2. Hover a reasoning model such as opencode/deepseek-v4.1-flash-free (or opencode-go/deepseek-v4.1-flash).
3. Observe the tooltip's Reasoning row reads "No reasoning" / 「不支持推理」.
4. Select the model: it still offers the low / high / max variants (thinking effort) and streams a thinking block.

Expected Behavior

When a model advertises reasoning (has reasoning-effort variants and/or reasoning metadata), the tooltip should report that it supports reasoning.

Actual Behavior

The tooltip's Reasoning row reports "No reasoning" regardless of the model.

Additional Context

Source analysis:

• The label is decided in packages/app/src/providers/models/tooltip.tsx:110-118 from model.capabilities.reasoning (falling back to model.reasoning).
• V2 catalog models do not carry a reasoning capability: packages/core/src/model.ts:32 (mergeCapabilities) and packages/core/src/models-dev.ts:232 build capabilities with only tools / input / output.
• The client sync hardcodes the value: packages/app/src/runtime/server/global-sync/utils.ts:84 sets reasoning: false, so the tooltip reads false for every catalog model.

The real reasoning behavior is driven by separate data that is already correct:

• Thinking effort = variants, derived from models.dev reasoning_options (packages/core/src/models-dev.ts:115 via Variant.resolve). For deepseek-v4.1-flash this is { "type": "effort", "values": ["low", "high", "max"] }.
• Thinking text = streamed reasoning parts, using the field from interleaved / compatibility.reasoningField ({ "field": "reasoning_content" } for DeepSeek).

models.dev reports reasoning: true and interleaved: { field: "reasoning_content" } for deepseek-v4.1-flash, so the catalog data is correct and only the tooltip's field is disconnected/stale.

• Frequency: always (deterministic).
• Impact: cosmetic/misleading only; reasoning features still work. No workaround needed.