feat(core): always allow topic tool and refine reasoning mandate

- Added explicit policy rules to always allow 'create_new_topic' in all modes.
- Updated topic tool output to use 'Current topic' phrasing for clarity.
- Replaced the 'Thinking Protocol' with a mandate for internal reasoning to prevent literal thought block output in the UI.
- Added and updated unit tests to verify policy and tool behavior.

packages/core/src/policy/policies/plan.toml

@@ -101,6 +101,13 @@ decision = "allow"
 priority = 70
 modes = ["plan"]
 
+# Topic grouping tool is innocuous and used for UI organization.
+[[rule]]
+toolName = "create_new_topic"
+decision = "allow"
+priority = 70
+modes = ["plan"]
+
 [[rule]]
 toolName = ["ask_user", "save_memory"]
 decision = "ask_user"

packages/core/src/policy/policies/read-only.toml

@@ -55,4 +55,10 @@ priority = 50
 [[rule]]
 toolName = ["codebase_investigator", "cli_help", "get_internal_docs"]
 decision = "allow"
+priority = 50
+
+# Topic grouping tool is innocuous and used for UI organization.
+[[rule]]
+toolName = "create_new_topic"
+decision = "allow"
 priority = 50

packages/core/src/policy/topic-policy.test.ts

@@ -0,0 +1,67 @@
+/**
+ * @license
+ * Copyright 2026 Google LLC
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import { describe, it, expect } from 'vitest';
+import * as path from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { loadPoliciesFromToml } from './toml-loader.js';
+import { PolicyEngine } from './policy-engine.js';
+import { ApprovalMode, PolicyDecision } from './types.js';
+import { CREATE_NEW_TOPIC_TOOL_NAME } from '../tools/tool-names.js';
+
+const __filename = fileURLToPath(import.meta.url);
+const __dirname = path.dirname(__filename);
+
+describe('Topic Tool Policy', () => {
+  async function loadDefaultPolicies() {
+    const policiesDir = path.resolve(__dirname, 'policies');
+    const getPolicyTier = () => 1; // Default tier
+    const result = await loadPoliciesFromToml([policiesDir], getPolicyTier);
+    return result.rules;
+  }
+
+  it('should allow create_new_topic in DEFAULT mode', async () => {
+    const rules = await loadDefaultPolicies();
+    const engine = new PolicyEngine({
+      rules,
+      approvalMode: ApprovalMode.DEFAULT,
+    });
+
+    const result = await engine.check(
+      { name: CREATE_NEW_TOPIC_TOOL_NAME },
+      undefined,
+    );
+    expect(result.decision).toBe(PolicyDecision.ALLOW);
+  });
+
+  it('should allow create_new_topic in PLAN mode', async () => {
+    const rules = await loadDefaultPolicies();
+    const engine = new PolicyEngine({
+      rules,
+      approvalMode: ApprovalMode.PLAN,
+    });
+
+    const result = await engine.check(
+      { name: CREATE_NEW_TOPIC_TOOL_NAME },
+      undefined,
+    );
+    expect(result.decision).toBe(PolicyDecision.ALLOW);
+  });
+
+  it('should allow create_new_topic in YOLO mode', async () => {
+    const rules = await loadDefaultPolicies();
+    const engine = new PolicyEngine({
+      rules,
+      approvalMode: ApprovalMode.YOLO,
+    });
+
+    const result = await engine.check(
+      { name: CREATE_NEW_TOPIC_TOOL_NAME },
+      undefined,
+    );
+    expect(result.decision).toBe(PolicyDecision.ALLOW);
+  });
+});

packages/core/src/prompts/snippets.ts

@@ -591,10 +591,10 @@ function mandateTopicUpdateModel(): string {
   - **Silent Mode:** No conversational filler, no "I will now...", and no summaries between tools. 
   - Only the tool execution is permitted to define the state. Everything in between must be silent.
 
-- **3. Thinking Protocol:**
-  - Use internal thought blocks to keep track of what tools you have called, plan your next steps, and reason about the task.
-  - Without reasoning and tracking in thought blocks, you may lose context.
-  - Always use the required syntax for thought blocks to ensure they remain hidden from the user interface.
+- **3. Internal Reasoning:**
+  - You MUST reason about your plan, track tool calls, and strategize internally before executing tools.
+  - This reasoning process must remain internal. You are strictly FORBIDDEN from including your reasoning, "thoughts," or explanations in your text response.
+  - Between tool calls, your text output MUST remain completely empty (Zero-Noise).
 
 - **4. Completion:**
   - Only when the entire task is finalized do you provide a **Final Summary**.

packages/core/src/tools/topicTool.test.ts

@@ -59,7 +59,7 @@ describe('CreateNewTopicTool', () => {
     const invocation = tool.build({ [TOPIC_PARAM_TITLE]: 'New Chapter' });
     const result = await invocation.execute(new AbortController().signal);
 
-    expect(result.llmContent).toContain('New Chapter');
+    expect(result.llmContent).toBe('Current topic: "New Chapter"');
     expect(TopicManager.getInstance().getTopic()).toBe('New Chapter');
   });
 

packages/core/src/tools/topicTool.ts

@@ -78,8 +78,8 @@ class CreateNewTopicInvocation extends BaseToolInvocation<
     TopicManager.getInstance().setTopic(title.trim());
 
     return {
-      llmContent: `Topic changed to: "${title.trim()}"`,
-      returnDisplay: `Topic changed to: **${title.trim()}**`,
+      llmContent: `Current topic: "${title.trim()}"`,
+      returnDisplay: `Current topic: **${title.trim()}**`,
     };
   }
 }