From 684f41692d4b78faf2d621aaae4678f77dafedcb Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E6=9D=A8=E8=89=BA=E9=9F=AC=28yangyitao=29?=
 <yangyitao100@sina.com>
Date: Sun, 15 Mar 2026 02:31:00 +0000
Subject: [PATCH] fix(ollama): send think param for thinking models

Ollama thinking models (deepseek-r1, qwq, etc.) require an explicit
`think` boolean in the /api/chat request body to control reasoning.
Previously this parameter was never sent, so thinking models always used
their default behavior regardless of the user's thinking-level config.

Now the Ollama stream function forwards the reasoning level as
`think: true` when reasoning is enabled and `think: false` when
reasoning is disabled.

Closes #46680
---
 src/agents/ollama-stream.test.ts | 54 ++++++++++++++++++++++++++++++++
 src/agents/ollama-stream.ts      | 13 ++++++++
 2 files changed, 67 insertions(+)

diff --git a/src/agents/ollama-stream.test.ts b/src/agents/ollama-stream.test.ts
index ded8064ea19..5382aaddb45 100644
--- a/src/agents/ollama-stream.test.ts
+++ b/src/agents/ollama-stream.test.ts
@@ -544,6 +544,60 @@ describe("createOllamaStreamFn", () => {
       [{ type: "text", text: "final answer" }],
     );
   });
+
+  it("sends think:true when reasoning level is set", async () => {
+    await withMockNdjsonFetch(
+      [
+        '{"model":"m","created_at":"t","message":{"role":"assistant","content":"ok"},"done":false}',
+        '{"model":"m","created_at":"t","message":{"role":"assistant","content":""},"done":true,"prompt_eval_count":1,"eval_count":1}',
+      ],
+      async (fetchMock) => {
+        const streamFn = createOllamaStreamFn("http://ollama-host:11434");
+        const stream = streamFn(
+          {
+            id: "deepseek-r1:32b",
+            api: "ollama",
+            provider: "ollama",
+            contextWindow: 131072,
+          } as never,
+          { messages: [{ role: "user", content: "hello" }] } as never,
+          { reasoning: "medium" } as never,
+        );
+        await collectStreamEvents(stream);
+
+        const [, reqInit] = fetchMock.mock.calls[0] as unknown as [string, RequestInit];
+        const body = JSON.parse(reqInit.body as string) as { think?: boolean };
+        expect(body.think).toBe(true);
+      },
+    );
+  });
+
+  it("sends think:false when reasoning is not set but options are present", async () => {
+    await withMockNdjsonFetch(
+      [
+        '{"model":"m","created_at":"t","message":{"role":"assistant","content":"ok"},"done":false}',
+        '{"model":"m","created_at":"t","message":{"role":"assistant","content":""},"done":true,"prompt_eval_count":1,"eval_count":1}',
+      ],
+      async (fetchMock) => {
+        const streamFn = createOllamaStreamFn("http://ollama-host:11434");
+        const stream = streamFn(
+          {
+            id: "deepseek-r1:32b",
+            api: "ollama",
+            provider: "ollama",
+            contextWindow: 131072,
+          } as never,
+          { messages: [{ role: "user", content: "hello" }] } as never,
+          {} as never,
+        );
+        await collectStreamEvents(stream);
+
+        const [, reqInit] = fetchMock.mock.calls[0] as unknown as [string, RequestInit];
+        const body = JSON.parse(reqInit.body as string) as { think?: boolean };
+        expect(body.think).toBe(false);
+      },
+    );
+  });
 });
 
 describe("resolveOllamaBaseUrlForRun", () => {
diff --git a/src/agents/ollama-stream.ts b/src/agents/ollama-stream.ts
index f332ad1fd83..e92b0897144 100644
--- a/src/agents/ollama-stream.ts
+++ b/src/agents/ollama-stream.ts
@@ -42,6 +42,7 @@ interface OllamaChatRequest {
   model: string;
   messages: OllamaChatMessage[];
   stream: boolean;
+  think?: boolean;
   tools?: OllamaTool[];
   options?: Record<string, unknown>;
 }
@@ -459,10 +460,22 @@ export function createOllamaStreamFn(
           ollamaOptions.num_predict = options.maxTokens;
         }
 
+        // Ollama thinking models (e.g. deepseek-r1, qwq) respect a top-level
+        // `think` boolean. Forward the reasoning level so `think: false` is
+        // sent explicitly when thinking is disabled (#46680).
+        const thinkParam: { think?: boolean } = {};
+        if (options?.reasoning) {
+          thinkParam.think = true;
+        } else if (options && !options.reasoning) {
+          // Thinking explicitly disabled – tell Ollama not to think.
+          thinkParam.think = false;
+        }
+
         const body: OllamaChatRequest = {
           model: model.id,
           messages: ollamaMessages,
           stream: true,
+          ...thinkParam,
           ...(ollamaTools.length > 0 ? { tools: ollamaTools } : {}),
           options: ollamaOptions,
         };