fix(ollama): honor native model capabilities

extensions/ollama/index.test.ts

@@ -528,6 +528,32 @@ describe("ollama plugin", () => {
     expect((payloadSeen?.options as Record<string, unknown> | undefined)?.think).toBeUndefined();
   });
 
+  it("keeps native Ollama thinking off by default while exposing an opt-in toggle", () => {
+    const provider = registerProvider();
+
+    expect(
+      provider.resolveThinkingProfile?.({
+        provider: "ollama",
+        modelId: "llama3.2:latest",
+        reasoning: false,
+      }),
+    ).toEqual({
+      levels: [{ id: "off" }],
+      defaultLevel: "off",
+    });
+
+    expect(
+      provider.resolveThinkingProfile?.({
+        provider: "ollama",
+        modelId: "gemma4:31b",
+        reasoning: true,
+      }),
+    ).toEqual({
+      levels: [{ id: "off" }, { id: "low", label: "on" }],
+      defaultLevel: "off",
+    });
+  });
+
   it("wraps native Ollama payloads with top-level think=true when thinking is enabled", () => {
     const { baseStreamFn, payloadSeen } = captureWrappedOllamaPayload("low");
     expect(baseStreamFn).toHaveBeenCalledTimes(1);

extensions/ollama/index.ts

@@ -166,6 +166,10 @@ export default definePluginEntry({
       contributeResolvedModelCompat: ({ model }) =>
         usesOllamaOpenAICompatTransport(model) ? { supportsUsageInStreaming: true } : undefined,
       resolveReasoningOutputMode: () => "native",
+      resolveThinkingProfile: ({ reasoning }) => ({
+        levels: reasoning === true ? [{ id: "off" }, { id: "low", label: "on" }] : [{ id: "off" }],
+        defaultLevel: "off",
+      }),
       wrapStreamFn: createConfiguredOllamaCompatStreamWrapper,
       createEmbeddingProvider: async ({ config, model, remote }) => {
         const { provider, client } = await createOllamaEmbeddingProvider({

extensions/ollama/src/provider-models.test.ts

@@ -203,13 +203,26 @@ describe("ollama provider models", () => {
       "vision",
       "completion",
       "tools",
+      "thinking",
     ]);
     expect(visionModel.input).toEqual(["text", "image"]);
+    expect(visionModel.reasoning).toBe(true);
+    expect(visionModel.compat?.supportsTools).toBe(true);
 
     const textModel = buildOllamaModelDefinition("glm-5.1:cloud", 202752, ["completion", "tools"]);
     expect(textModel.input).toEqual(["text"]);
+    expect(textModel.reasoning).toBe(false);
+    expect(textModel.compat?.supportsTools).toBe(true);
 
     const noCapabilities = buildOllamaModelDefinition("unknown-model", 65536);
     expect(noCapabilities.input).toEqual(["text"]);
+    expect(noCapabilities.compat).toBeUndefined();
+  });
+
+  it("disables tool support when Ollama capabilities omit tools", () => {
+    const model = buildOllamaModelDefinition("embeddinggemma:latest", 2048, ["embedding"]);
+
+    expect(model.reasoning).toBe(false);
+    expect(model.compat?.supportsTools).toBe(false);
   });
 });

extensions/ollama/src/provider-models.ts

@@ -218,14 +218,25 @@ export function buildOllamaModelDefinition(
 ): ModelDefinitionConfig {
   const hasVision = capabilities?.includes("vision") ?? false;
   const input: ("text" | "image")[] = hasVision ? ["text", "image"] : ["text"];
+  const reasoning =
+    capabilities === undefined
+      ? isReasoningModelHeuristic(modelId)
+      : capabilities.includes("thinking");
+  const compat =
+    capabilities === undefined
+      ? undefined
+      : {
+          supportsTools: capabilities.includes("tools"),
+        };
   return {
     id: modelId,
     name: modelId,
-    reasoning: isReasoningModelHeuristic(modelId),
+    reasoning,
     input,
     cost: OLLAMA_DEFAULT_COST,
     contextWindow: contextWindow ?? OLLAMA_DEFAULT_CONTEXT_WINDOW,
     maxTokens: OLLAMA_DEFAULT_MAX_TOKENS,
+    ...(compat ? { compat } : {}),
   };
 }