feat: LLM intent router, direct chat execution, and Memory V2 light intervention

Wire chat intent routing with direct_chat on regular sessions, skill-selected
short-circuit to Agent, memory light/heavy intervention tiers, and fix direct
chat UI stuck streaming after completion.

Co-authored-by: Cursor <cursoragent@cursor.com>

tkmind-proxy.test.mjs

@@ -406,7 +406,7 @@ test('startSessionForUser resolves memories through Memory V2 facade', async ()
       userId: 'user-1',
       sessionId: 'session-1',
       query: null,
-      limit: 40,
+      limit: 3,
     });
     assert.ok(
       harnessEntries.some((entry) => String(entry.content ?? '').includes('用户关注 Memory V2 架构边界')),
@@ -577,6 +577,55 @@ test('submitSessionReplyForUser passes current prompt to Memory V2 resolve befor
       },
     );
 
+    assert.deepEqual(resolveInput, {
+      userId: 'user-1',
+      sessionId: 'session-1',
+      query: '帮我设计 memory-chain 下一阶段',
+      limit: 3,
+    });
+  });
+});
+
+test('submitSessionReplyForUser uses heavy memory resolve when deep reasoning is enabled', async () => {
+  let resolveInput = null;
+  await withFakeGoosedSession(async ({ apiTarget, workingDir }) => {
+    const proxy = createTkmindProxy({
+      apiTarget,
+      apiSecret: 'test-secret',
+      userAuth: {
+        ...createMemoryTestUserAuth(workingDir),
+        async ownsSession() {
+          return true;
+        },
+        async canUseChat() {
+          return { ok: true };
+        },
+        async getUserById() {
+          return { id: 'user-1' };
+        },
+        async resolveUserPolicies() {
+          return { unrestricted: true, policies: {} };
+        },
+      },
+      memoryV2: {
+        async resolve(input) {
+          resolveInput = input;
+          return { memories: [] };
+        },
+      },
+    });
+
+    await proxy.submitSessionReplyForUser(
+      'user-1',
+      'session-1',
+      'request-1',
+      {
+        role: 'user',
+        content: [{ type: 'text', text: '帮我设计 memory-chain 下一阶段' }],
+      },
+      { forceDeepReasoning: true },
+    );
+
     assert.deepEqual(resolveInput, {
       userId: 'user-1',
       sessionId: 'session-1',