fix(ai): make custom output token budget configurable

apps/desktop/src/lib/__tests__/ai/ai.spec.ts

@@ -158,6 +158,44 @@ describe("AI SQL dialect prompt", () => {
     expect(prompt).toContain("Ask mode");
   });
 
+  it("uses the configured output budget for non-thinking requests", () => {
+    const request = buildAgentRequest({
+      config: {
+        provider: "deepseek",
+        apiKey: "test",
+        apiUrl: "https://example.invalid",
+        model: "deepseek-v4-flash",
+        enableThinking: false,
+        maxOutputTokens: 1024,
+      },
+      action: "general",
+      mode: "ask",
+      instruction: "generate a long migration",
+      context: context(),
+    });
+
+    expect(request.maxTokens).toBe(1024);
+  });
+
+  it("uses the configured output budget above the thinking minimum", () => {
+    const request = buildAgentRequest({
+      config: {
+        provider: "deepseek",
+        apiKey: "test",
+        apiUrl: "https://example.invalid",
+        model: "deepseek-v4-flash",
+        enableThinking: true,
+        maxOutputTokens: 32768,
+      },
+      action: "general",
+      mode: "ask",
+      instruction: "generate a long migration",
+      context: context(),
+    });
+
+    expect(request.maxTokens).toBe(32768);
+  });
+
   it("MongoDB agent mode uses shell commands instead of SQL", () => {
     const prompt = buildSystemPrompt("general", context({ databaseType: "mongodb", connectionName: "MongoDB", database: "benchmark" }), "agent");
 

apps/desktop/src/lib/ai/ai.ts

@@ -179,7 +179,8 @@ export function buildAgentRequest(input: AiRequestInput, history?: api.AiMessage
   ];
 
   const params = actionParams(input.action);
-  const maxTokens = input.config.enableThinking ? Math.max(params.maxTokens, 8192) : params.maxTokens;
+  const baseMaxTokens = input.config.maxOutputTokens ?? params.maxTokens;
+  const maxTokens = input.config.enableThinking ? Math.max(baseMaxTokens, 8192) : baseMaxTokens;
   return { messages, systemPrompt, taskContract, maxTokens };
 }