fix(ai): fall back to heuristic for non-builtin custom models in inferReasoning

Custom models auto-saved by ensureCustomProvidersAndModels carry only
capabilities:['chat'], so inferReasoning was returning reasoning=false
even for models like qwen3:8b whose names match the reasoning heuristic.
The UI still showed thinking levels via getSupportedThinkingLevels, but
selections were silently ignored by the backend.

Fix: for non-builtin (builtin=false) model definitions, also apply the
name/provider heuristic as a fallback alongside the capabilities check.
Builtin catalog models remain authoritative.

Adds a regression test covering the exact findModelFn shape produced by
ensureCustomProvidersAndModels.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

packages/node-runtime/src/ai/__tests__/llm-builder.test.ts

@@ -132,6 +132,30 @@ describe('buildPiModel', () => {
     assert.deepEqual(model.compat, { thinkingFormat: 'qwen' })
   })
 
+  it('auto-infers reasoning=true for custom model saved with only capabilities:["chat"]', () => {
+    // Regression: ensureCustomProvidersAndModels saves user-added models with capabilities:['chat']
+    // only. The heuristic must still fire for non-builtin model definitions so that thinking
+    // level selections are not silently ignored in the UI.
+    const model = buildPiModel(
+      { provider: 'openai-compatible', model: 'qwen3:8b', baseUrl: 'http://localhost:11434/v1' },
+      {
+        findModelFn: () => ({
+          id: 'qwen3:8b',
+          providerId: 'openai-compatible',
+          name: 'qwen3:8b',
+          contextWindow: 32768,
+          capabilities: ['chat'] as const,
+          recommendedFor: ['chat'] as const,
+          status: 'stable' as const,
+          builtin: false,
+          editable: true,
+        }),
+      }
+    )
+    assert.equal(model.reasoning, true)
+    assert.deepEqual(model.compat, { thinkingFormat: 'qwen' })
+  })
+
   it('auto-infers reasoning=false and compat=undefined for non-reasoning Anthropic model', () => {
     const model = buildPiModel({
       provider: 'openai-compatible',

packages/node-runtime/src/ai/llm-builder.ts

@@ -86,7 +86,11 @@ interface InferReasoningResult {
 }
 
 function inferReasoning(provider: string, modelId: string, modelDef: ModelDefinition | null): InferReasoningResult {
-  const reasoning = modelDef ? modelDef.capabilities.includes('reasoning') : isReasoningModel(provider, modelId)
+  // Builtin catalog models: trust capabilities array exclusively.
+  // Custom (builtin=false) or unlisted models: also check name heuristic as fallback because
+  // auto-saved custom model entries carry only capabilities:['chat'] even for reasoning models.
+  const reasoning =
+    modelDef?.capabilities.includes('reasoning') || (!modelDef?.builtin && isReasoningModel(provider, modelId))
 
   if (!reasoning) return { reasoning: false, compat: undefined }