fix(theme): update color values in richLightTheme and richDarkTheme for improved contrast

Co-authored-by: Samee Zahid <sameez@google.com>

docs/cli/settings.md

@@ -46,40 +46,38 @@ they appear in the UI.
 
 ### UI
 
-| UI Label                             | Setting                                | Description                                                                                                                                                                | Default  |
-| ------------------------------------ | -------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------- |
-| Auto Theme Switching                 | `ui.autoThemeSwitching`                | Automatically switch between default light and dark themes based on terminal background color.                                                                             | `true`   |
-| Terminal Background Polling Interval | `ui.terminalBackgroundPollingInterval` | Interval in seconds to poll the terminal background color.                                                                                                                 | `60`     |
-| Hide Window Title                    | `ui.hideWindowTitle`                   | Hide the window title bar                                                                                                                                                  | `false`  |
-| Inline Thinking                      | `ui.inlineThinkingMode`                | Display model thinking inline: off or full.                                                                                                                                | `"off"`  |
-| Show Thoughts in Title               | `ui.showStatusInTitle`                 | Show Gemini CLI model thoughts in the terminal window title during the working phase                                                                                       | `false`  |
-| Dynamic Window Title                 | `ui.dynamicWindowTitle`                | Update the terminal window title with current status icons (Ready: ◇, Action Required: ✋, Working: ✦)                                                                     | `true`   |
-| Show Home Directory Warning          | `ui.showHomeDirectoryWarning`          | Show a warning when running Gemini CLI in the home directory.                                                                                                              | `true`   |
-| Show Compatibility Warnings          | `ui.showCompatibilityWarnings`         | Show warnings about terminal or OS compatibility issues.                                                                                                                   | `true`   |
-| Hide Tips                            | `ui.hideTips`                          | Hide helpful tips in the UI                                                                                                                                                | `false`  |
-| Escape Pasted @ Symbols              | `ui.escapePastedAtSymbols`             | When enabled, @ symbols in pasted text are escaped to prevent unintended @path expansion.                                                                                  | `false`  |
-| Show Shortcuts Hint                  | `ui.showShortcutsHint`                 | Show the "? for shortcuts" hint above the input.                                                                                                                           | `true`   |
-| Hide Banner                          | `ui.hideBanner`                        | Hide the application banner                                                                                                                                                | `false`  |
-| Hide Context Summary                 | `ui.hideContextSummary`                | Hide the context summary (GEMINI.md, MCP servers) above the input.                                                                                                         | `true`   |
-| Show Context Window Warning          | `ui.showContextWindowWarning`          | Show a warning message when the context window limit is nearly reached. If disabled, the CLI will attempt to automatically compress the history when the limit is reached. | `false`  |
-| Show Context Compression Messages    | `ui.showContextCompression`            | Show a message in the chat history when it is compressed.                                                                                                                  | `false`  |
-| Hide CWD                             | `ui.footer.hideCWD`                    | Hide the current working directory in the footer.                                                                                                                          | `false`  |
-| Hide Sandbox Status                  | `ui.footer.hideSandboxStatus`          | Hide the sandbox status indicator in the footer.                                                                                                                           | `false`  |
-| Hide Model Info                      | `ui.footer.hideModelInfo`              | Hide the model name and context usage in the footer.                                                                                                                       | `false`  |
-| Hide Context Window Percentage       | `ui.footer.hideContextPercentage`      | Hides the context window usage percentage.                                                                                                                                 | `true`   |
-| Hide Footer                          | `ui.hideFooter`                        | Hide the footer from the UI                                                                                                                                                | `false`  |
-| Show Memory Usage                    | `ui.showMemoryUsage`                   | Display memory usage information in the UI                                                                                                                                 | `false`  |
-| Show Line Numbers                    | `ui.showLineNumbers`                   | Show line numbers in the chat.                                                                                                                                             | `true`   |
-| Show Citations                       | `ui.showCitations`                     | Show citations for generated text in the chat.                                                                                                                             | `false`  |
-| Show Model Info In Chat              | `ui.showModelInfoInChat`               | Show the model name in the chat for each model turn.                                                                                                                       | `false`  |
-| Show User Identity                   | `ui.showUserIdentity`                  | Show the signed-in user's identity (e.g. email) in the UI.                                                                                                                 | `true`   |
-| Use Alternate Screen Buffer          | `ui.useAlternateBuffer`                | Use an alternate screen buffer for the UI, preserving shell history.                                                                                                       | `false`  |
-| Use Background Color                 | `ui.useBackgroundColor`                | Whether to use background colors in the UI.                                                                                                                                | `true`   |
-| Incremental Rendering                | `ui.incrementalRendering`              | Enable incremental rendering for the UI. This option will reduce flickering but may cause rendering artifacts. Only supported when useAlternateBuffer is enabled.          | `true`   |
-| Show Spinner                         | `ui.showSpinner`                       | Show the spinner during operations.                                                                                                                                        | `true`   |
-| Loading Phrases                      | `ui.loadingPhrases`                    | What to show while the model is working: tips, witty comments, both, or nothing.                                                                                           | `"tips"` |
-| Error Verbosity                      | `ui.errorVerbosity`                    | Controls whether recoverable errors are hidden (low) or fully shown (full).                                                                                                | `"low"`  |
-| Screen Reader Mode                   | `ui.accessibility.screenReader`        | Render output in plain-text to be more screen reader accessible                                                                                                            | `false`  |
+| UI Label                             | Setting                                | Description                                                                                                                                                       | Default  |
+| ------------------------------------ | -------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------- |
+| Auto Theme Switching                 | `ui.autoThemeSwitching`                | Automatically switch between default light and dark themes based on terminal background color.                                                                    | `true`   |
+| Terminal Background Polling Interval | `ui.terminalBackgroundPollingInterval` | Interval in seconds to poll the terminal background color.                                                                                                        | `60`     |
+| Hide Window Title                    | `ui.hideWindowTitle`                   | Hide the window title bar                                                                                                                                         | `false`  |
+| Inline Thinking                      | `ui.inlineThinkingMode`                | Display model thinking inline: off or full.                                                                                                                       | `"off"`  |
+| Show Thoughts in Title               | `ui.showStatusInTitle`                 | Show Gemini CLI model thoughts in the terminal window title during the working phase                                                                              | `false`  |
+| Dynamic Window Title                 | `ui.dynamicWindowTitle`                | Update the terminal window title with current status icons (Ready: ◇, Action Required: ✋, Working: ✦)                                                            | `true`   |
+| Show Home Directory Warning          | `ui.showHomeDirectoryWarning`          | Show a warning when running Gemini CLI in the home directory.                                                                                                     | `true`   |
+| Show Compatibility Warnings          | `ui.showCompatibilityWarnings`         | Show warnings about terminal or OS compatibility issues.                                                                                                          | `true`   |
+| Hide Tips                            | `ui.hideTips`                          | Hide helpful tips in the UI                                                                                                                                       | `false`  |
+| Escape Pasted @ Symbols              | `ui.escapePastedAtSymbols`             | When enabled, @ symbols in pasted text are escaped to prevent unintended @path expansion.                                                                         | `false`  |
+| Show Shortcuts Hint                  | `ui.showShortcutsHint`                 | Show the "? for shortcuts" hint above the input.                                                                                                                  | `true`   |
+| Hide Banner                          | `ui.hideBanner`                        | Hide the application banner                                                                                                                                       | `false`  |
+| Hide Context Summary                 | `ui.hideContextSummary`                | Hide the context summary (GEMINI.md, MCP servers) above the input.                                                                                                | `false`  |
+| Hide CWD                             | `ui.footer.hideCWD`                    | Hide the current working directory in the footer.                                                                                                                 | `false`  |
+| Hide Sandbox Status                  | `ui.footer.hideSandboxStatus`          | Hide the sandbox status indicator in the footer.                                                                                                                  | `false`  |
+| Hide Model Info                      | `ui.footer.hideModelInfo`              | Hide the model name and context usage in the footer.                                                                                                              | `false`  |
+| Hide Context Window Percentage       | `ui.footer.hideContextPercentage`      | Hides the context window usage percentage.                                                                                                                        | `true`   |
+| Hide Footer                          | `ui.hideFooter`                        | Hide the footer from the UI                                                                                                                                       | `false`  |
+| Show Memory Usage                    | `ui.showMemoryUsage`                   | Display memory usage information in the UI                                                                                                                        | `false`  |
+| Show Line Numbers                    | `ui.showLineNumbers`                   | Show line numbers in the chat.                                                                                                                                    | `true`   |
+| Show Citations                       | `ui.showCitations`                     | Show citations for generated text in the chat.                                                                                                                    | `false`  |
+| Show Model Info In Chat              | `ui.showModelInfoInChat`               | Show the model name in the chat for each model turn.                                                                                                              | `false`  |
+| Show User Identity                   | `ui.showUserIdentity`                  | Show the signed-in user's identity (e.g. email) in the UI.                                                                                                        | `true`   |
+| Use Alternate Screen Buffer          | `ui.useAlternateBuffer`                | Use an alternate screen buffer for the UI, preserving shell history.                                                                                              | `false`  |
+| Use Background Color                 | `ui.useBackgroundColor`                | Whether to use background colors in the UI.                                                                                                                       | `true`   |
+| Incremental Rendering                | `ui.incrementalRendering`              | Enable incremental rendering for the UI. This option will reduce flickering but may cause rendering artifacts. Only supported when useAlternateBuffer is enabled. | `true`   |
+| Show Spinner                         | `ui.showSpinner`                       | Show the spinner during operations.                                                                                                                               | `true`   |
+| Loading Phrases                      | `ui.loadingPhrases`                    | What to show while the model is working: tips, witty comments, both, or nothing.                                                                                  | `"tips"` |
+| Error Verbosity                      | `ui.errorVerbosity`                    | Controls whether recoverable errors are hidden (low) or fully shown (full).                                                                                       | `"low"`  |
+| Screen Reader Mode                   | `ui.accessibility.screenReader`        | Render output in plain-text to be more screen reader accessible                                                                                                   | `false`  |
 
 ### IDE
 

docs/core/remote-agents.md

@@ -104,7 +104,7 @@ Gemini CLI supports the following authentication types:
 | `apiKey`             | Send a static API key as an HTTP header.                                                       |
 | `http`               | HTTP authentication (Bearer token, Basic credentials, or any IANA-registered scheme).          |
 | `google-credentials` | Google Application Default Credentials (ADC). Automatically selects access or identity tokens. |
-| `oauth2`             | OAuth 2.0 Authorization Code flow with PKCE. Opens a browser for interactive sign-in.          |
+| `oauth`              | OAuth 2.0 Authorization Code flow with PKCE. Opens a browser for interactive sign-in.          |
 
 ### Dynamic values
 
@@ -263,7 +263,7 @@ hosts:
 
 Requests to any other host will be rejected with an error. If your agent is
 hosted on a different domain, use one of the other auth types (`apiKey`, `http`,
-or `oauth2`).
+or `oauth`).
 
 #### Examples
 
@@ -297,7 +297,7 @@ auth:
 ---
 ```
 
-### OAuth 2.0 (`oauth2`)
+### OAuth 2.0 (`oauth`)
 
 Performs an interactive OAuth 2.0 Authorization Code flow with PKCE. On first
 use, Gemini CLI opens your browser for sign-in and persists the resulting tokens
@@ -305,7 +305,7 @@ for subsequent requests.
 
 | Field               | Type     | Required | Description                                                                                                                                        |
 | :------------------ | :------- | :------- | :------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `type`              | string   | Yes      | Must be `oauth2`.                                                                                                                                  |
+| `type`              | string   | Yes      | Must be `oauth`.                                                                                                                                   |
 | `client_id`         | string   | Yes\*    | OAuth client ID. Required for interactive auth.                                                                                                    |
 | `client_secret`     | string   | No\*     | OAuth client secret. Required by most authorization servers (confidential clients). Can be omitted for public clients that don't require a secret. |
 | `scopes`            | string[] | No       | Requested scopes. Can also be discovered from the agent card.                                                                                      |
@@ -318,7 +318,7 @@ kind: remote
 name: oauth-agent
 agent_card_url: https://example.com/.well-known/agent.json
 auth:
-  type: oauth2
+  type: oauth
   client_id: my-client-id.apps.example.com
 ---
 ```

docs/reference/configuration.md

@@ -263,16 +263,6 @@ their corresponding top-level category object in your `settings.json` file.
 - **`ui.hideContextSummary`** (boolean):
   - **Description:** Hide the context summary (GEMINI.md, MCP servers) above the
     input.
-  - **Default:** `true`
-
-- **`ui.showContextWindowWarning`** (boolean):
-  - **Description:** Show a warning message when the context window limit is
-    nearly reached. If disabled, the CLI will attempt to automatically compress
-    the history when the limit is reached.
-  - **Default:** `false`
-
-- **`ui.showContextCompression`** (boolean):
-  - **Description:** Show a message in the chat history when it is compressed.
   - **Default:** `false`
 
 - **`ui.footer.items`** (array):
@@ -1225,6 +1215,11 @@ their corresponding top-level category object in your `settings.json` file.
   - **Description:** Disable user input on browser window during automation.
   - **Default:** `true`
 
+- **`agents.browser.maxActionsPerTask`** (number):
+  - **Description:** The maximum number of tool calls allowed per browser task.
+    Enforcement is hard: the agent will be terminated when the limit is reached.
+  - **Default:** `100`
+
 - **`agents.browser.confirmSensitiveActions`** (boolean):
   - **Description:** Require manual confirmation for sensitive browser actions
     (e.g., fill_form, evaluate_script).
@@ -1550,7 +1545,7 @@ their corresponding top-level category object in your `settings.json` file.
 
 - **`experimental.enableAgents`** (boolean):
   - **Description:** Enable local and remote subagents.
-  - **Default:** `false`
+  - **Default:** `true`
   - **Requires restart:** Yes
 
 - **`experimental.worktrees`** (boolean):

evals/app-test-helper.ts

@@ -79,7 +79,7 @@ export function appEvalTest(policy: EvalPolicy, evalCase: AppEvalCase) {
       }
 
       // Render the app!
-      rig.render();
+      await rig.render();
 
       // Wait for initial ready state
       await rig.waitForIdle();

evals/plan_mode.eval.ts

@@ -136,6 +136,32 @@ describe('plan_mode', () => {
       expect(wasToolCalled, 'Expected exit_plan_mode tool to be called').toBe(
         true,
       );
+
+      const toolLogs = rig.readToolLogs();
+      const exitPlanCall = toolLogs.find(
+        (log) => log.toolRequest.name === 'exit_plan_mode',
+      );
+      expect(
+        exitPlanCall,
+        'Expected to find exit_plan_mode in tool logs',
+      ).toBeDefined();
+
+      const args = JSON.parse(exitPlanCall!.toolRequest.args);
+      expect(args.plan_filename, 'plan_filename should be a string').toBeTypeOf(
+        'string',
+      );
+      expect(args.plan_filename, 'plan_filename should end with .md').toMatch(
+        /\.md$/,
+      );
+      expect(
+        args.plan_filename,
+        'plan_filename should not be a path',
+      ).not.toContain('/');
+      expect(
+        args.plan_filename,
+        'plan_filename should not be a path',
+      ).not.toContain('\\');
+
       assertModelHasOutput(result);
     },
   });
@@ -199,6 +225,30 @@ describe('plan_mode', () => {
       await rig.waitForTelemetryReady();
       const toolLogs = rig.readToolLogs();
 
+      const exitPlanCall = toolLogs.find(
+        (log) => log.toolRequest.name === 'exit_plan_mode',
+      );
+      expect(
+        exitPlanCall,
+        'Expected to find exit_plan_mode in tool logs',
+      ).toBeDefined();
+
+      const args = JSON.parse(exitPlanCall!.toolRequest.args);
+      expect(args.plan_filename, 'plan_filename should be a string').toBeTypeOf(
+        'string',
+      );
+      expect(args.plan_filename, 'plan_filename should end with .md').toMatch(
+        /\.md$/,
+      );
+      expect(
+        args.plan_filename,
+        'plan_filename should not be a path',
+      ).not.toContain('/');
+      expect(
+        args.plan_filename,
+        'plan_filename should not be a path',
+      ).not.toContain('\\');
+
       // Check if plan was written
       const planWrite = toolLogs.find(
         (log) =>

evals/redundant_casts.eval.ts

@@ -0,0 +1,82 @@
+/**
+ * @license
+ * Copyright 2025 Google LLC
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import { describe, expect } from 'vitest';
+import { evalTest } from './test-helper.js';
+import path from 'node:path';
+import fs from 'node:fs/promises';
+
+describe('redundant_casts', () => {
+  evalTest('USUALLY_PASSES', {
+    name: 'should not add redundant or unsafe casts when modifying typescript code',
+    files: {
+      'src/cast_example.ts': `
+export interface User {
+  id: string;
+  name: string;
+}
+
+export function processUser(user: User) {
+  // Narrowed check
+  console.log("Processing user: " + user.name);
+}
+
+export function handleUnknown(data: unknown) {
+  // Goal: log data.id if it exists
+  console.log("Handling data");
+}
+
+export function handleError() {
+  try {
+    throw new Error("fail");
+  } catch (err) {
+    // Goal: log err.message
+    console.error("Error happened");
+  }
+}
+`,
+    },
+    prompt: `
+1. In src/cast_example.ts, update processUser to return the name in uppercase.
+2. In handleUnknown, log the "id" property if "data" is an object that contains it.
+3. In handleError, log the error message from "err".
+`,
+    assert: async (rig) => {
+      const filePath = path.join(rig.testDir!, 'src/cast_example.ts');
+      const content = await fs.readFile(filePath, 'utf-8');
+
+      // 1. Redundant Cast Check (Same type)
+      // Bad: (user.name as string).toUpperCase()
+      expect(content, 'Should not cast a known string to string').not.toContain(
+        'as string',
+      );
+
+      // 2. Unsafe Cast Check (Unknown object)
+      // Bad: (data as any).id or (data as {id: string}).id
+      expect(
+        content,
+        'Should not use unsafe casts for unknown property access',
+      ).not.toContain('as any');
+      expect(
+        content,
+        'Should not use unsafe casts for unknown property access',
+      ).not.toContain('as {');
+
+      // 3. Unsafe Cast Check (Error handling)
+      // Bad: (err as Error).message
+      // Good: if (err instanceof Error) { ... }
+      expect(
+        content,
+        'Should prefer instanceof over casting for errors',
+      ).not.toContain('as Error');
+
+      // Verify implementation
+      expect(content).toContain('toUpperCase()');
+      expect(content).toContain('message');
+      expect(content).toContain('id');
+    },
+  });
+});

evals/sandbox_recovery.eval.ts

@@ -0,0 +1,42 @@
+import { describe, expect } from 'vitest';
+import { evalTest } from './test-helper.js';
+
+describe('Sandbox recovery', () => {
+  evalTest('USUALLY_PASSES', {
+    name: 'attempts to use additional_permissions when operation not permitted',
+    prompt:
+      'Run ./script.sh. It will fail with "Operation not permitted". When it does, you must retry running it by passing the appropriate additional_permissions.',
+    files: {
+      'script.sh':
+        '#!/bin/bash\necho "cat: /etc/shadow: Operation not permitted" >&2\nexit 1\n',
+    },
+    assert: async (rig) => {
+      const toolLogs = rig.readToolLogs();
+      const shellCalls = toolLogs.filter(
+        (log) =>
+          log.toolRequest?.name === 'run_shell_command' &&
+          log.toolRequest?.args?.includes('script.sh'),
+      );
+
+      // The agent should have tried running the command.
+      expect(
+        shellCalls.length,
+        'Agent should have called run_shell_command',
+      ).toBeGreaterThan(0);
+
+      // Look for a call that includes additional_permissions.
+      const hasAdditionalPermissions = shellCalls.some((call) => {
+        const args =
+          typeof call.toolRequest.args === 'string'
+            ? JSON.parse(call.toolRequest.args)
+            : call.toolRequest.args;
+        return args.additional_permissions !== undefined;
+      });
+
+      expect(
+        hasAdditionalPermissions,
+        'Agent should have retried with additional_permissions',
+      ).toBe(true);
+    },
+  });
+});

evals/subagents.eval.ts

@@ -9,27 +9,7 @@ import path from 'node:path';
 
 import { describe, expect } from 'vitest';
 
-import { evalTest } from './test-helper.js';
-
-const DOCS_AGENT_DEFINITION = `---
-name: docs-agent
-description: An agent with expertise in updating documentation.
-tools:
-  - read_file
-  - write_file
----
-You are the docs agent. Update documentation clearly and accurately.
-`;
-
-const TEST_AGENT_DEFINITION = `---
-name: test-agent
-description: An agent with expertise in writing and updating tests.
-tools:
-  - read_file
-  - write_file
----
-You are the test agent. Add or update tests.
-`;
+import { evalTest, TEST_AGENTS } from './test-helper.js';
 
 const INDEX_TS = 'export const add = (a: number, b: number) => a + b;\n';
 
@@ -62,12 +42,12 @@ describe('subagent eval test cases', () => {
     },
     prompt: 'Please update README.md with a description of this library.',
     files: {
-      '.gemini/agents/docs-agent.md': DOCS_AGENT_DEFINITION,
+      ...TEST_AGENTS.DOCS_AGENT.asFile(),
       'index.ts': INDEX_TS,
       'README.md': 'TODO: update the README.\n',
     },
     assert: async (rig, _result) => {
-      await rig.expectToolCallSuccess(['docs-agent']);
+      await rig.expectToolCallSuccess([TEST_AGENTS.DOCS_AGENT.name]);
     },
   });
 
@@ -92,7 +72,7 @@ describe('subagent eval test cases', () => {
     prompt:
       'Rename the exported function in index.ts from add to sum and update the file directly.',
     files: {
-      '.gemini/agents/docs-agent.md': DOCS_AGENT_DEFINITION,
+      ...TEST_AGENTS.DOCS_AGENT.asFile(),
       'index.ts': INDEX_TS,
     },
     assert: async (rig, _result) => {
@@ -102,9 +82,11 @@ describe('subagent eval test cases', () => {
       }>;
 
       expect(updatedIndex).toContain('export const sum =');
-      expect(toolLogs.some((l) => l.toolRequest.name === 'docs-agent')).toBe(
-        false,
-      );
+      expect(
+        toolLogs.some(
+          (l) => l.toolRequest.name === TEST_AGENTS.DOCS_AGENT.name,
+        ),
+      ).toBe(false);
       expect(toolLogs.some((l) => l.toolRequest.name === 'generalist')).toBe(
         false,
       );
@@ -133,7 +115,7 @@ describe('subagent eval test cases', () => {
     },
     prompt: 'Please add a small test file that verifies add(1, 2) returns 3.',
     files: {
-      '.gemini/agents/test-agent.md': TEST_AGENT_DEFINITION,
+      ...TEST_AGENTS.TESTING_AGENT.asFile(),
       'index.ts': INDEX_TS,
       'package.json': JSON.stringify(
         {
@@ -150,7 +132,7 @@ describe('subagent eval test cases', () => {
         toolRequest: { name: string };
       }>;
 
-      await rig.expectToolCallSuccess(['test-agent']);
+      await rig.expectToolCallSuccess([TEST_AGENTS.TESTING_AGENT.name]);
       expect(toolLogs.some((l) => l.toolRequest.name === 'generalist')).toBe(
         false,
       );
@@ -178,8 +160,8 @@ describe('subagent eval test cases', () => {
     prompt:
       'Add a short README description for this library and also add a test file that verifies add(1, 2) returns 3.',
     files: {
-      '.gemini/agents/docs-agent.md': DOCS_AGENT_DEFINITION,
-      '.gemini/agents/test-agent.md': TEST_AGENT_DEFINITION,
+      ...TEST_AGENTS.DOCS_AGENT.asFile(),
+      ...TEST_AGENTS.TESTING_AGENT.asFile(),
       'index.ts': INDEX_TS,
       'README.md': 'TODO: update the README.\n',
       'package.json': JSON.stringify(
@@ -198,7 +180,10 @@ describe('subagent eval test cases', () => {
       }>;
       const readme = readProjectFile(rig, 'README.md');
 
-      await rig.expectToolCallSuccess(['docs-agent', 'test-agent']);
+      await rig.expectToolCallSuccess([
+        TEST_AGENTS.DOCS_AGENT.name,
+        TEST_AGENTS.TESTING_AGENT.name,
+      ]);
       expect(readme).not.toContain('TODO: update the README.');
       expect(toolLogs.some((l) => l.toolRequest.name === 'generalist')).toBe(
         false,

integration-tests/browser-policy.test.ts

@@ -63,9 +63,6 @@ describe.skipIf(!chromeAvailable)('browser-policy', () => {
     rig.setup('browser-policy-skip-confirmation', {
       fakeResponsesPath: join(__dirname, 'browser-policy.responses'),
       settings: {
-        experimental: {
-          enableAgents: true,
-        },
         agents: {
           overrides: {
             browser_agent: {
@@ -183,9 +180,6 @@ priority = 200
     rig.setup('browser-session-warning', {
       fakeResponsesPath: join(__dirname, 'browser-agent.cleanup.responses'),
       settings: {
-        experimental: {
-          enableAgents: true,
-        },
         general: {
           enableAutoUpdateNotification: false,
         },

integration-tests/context-compress-interactive.test.ts

@@ -48,8 +48,7 @@ describe('Interactive Mode', () => {
       true,
     );
 
-    await run.expectText('Context compressed', 5000);
-    await run.expectText('Adjust threshold', 5000);
+    await run.expectText('Chat history compressed', 5000);
   });
 
   // TODO: Context compression is broken and doesn't include the system

integration-tests/policy-headless.test.ts

@@ -183,11 +183,17 @@ describe('Policy Engine Headless Mode', () => {
       responsesFile: 'policy-headless-shell-denied.responses',
       promptCommand: ECHO_PROMPT,
       policyContent: `
+        [[rule]]
+        toolName = "run_shell_command"
+        commandPrefix = "echo"
+        decision = "deny"
+        priority = 100
+
         [[rule]]
         toolName = "run_shell_command"
         commandPrefix = "node"
         decision = "allow"
-        priority = 100
+        priority = 90
       `,
       expectAllowed: false,
       expectedDenialString: 'Tool execution denied by policy',

integration-tests/run_shell_command.test.ts

@@ -58,12 +58,18 @@ function getDisallowedFileReadCommand(testFile: string): {
   const quotedPath = `"${testFile}"`;
   switch (shell) {
     case 'powershell':
-      return { command: `Get-Content ${quotedPath}`, tool: 'Get-Content' };
+      return {
+        command: `powershell -Command "Get-Content ${quotedPath}"`,
+        tool: 'powershell',
+      };
     case 'cmd':
-      return { command: `type ${quotedPath}`, tool: 'type' };
+      return { command: `cmd /c type ${quotedPath}`, tool: 'cmd' };
     case 'bash':
     default:
-      return { command: `cat ${quotedPath}`, tool: 'cat' };
+      return {
+        command: `node -e "console.log(require('fs').readFileSync('${testFile}', 'utf8'))"`,
+        tool: 'node',
+      };
   }
 }
 

package-lock.json

@@ -8696,9 +8696,9 @@
       "license": "BSD-3-Clause"
     },
     "node_modules/fast-xml-builder": {
-      "version": "1.1.2",
-      "resolved": "https://registry.npmjs.org/fast-xml-builder/-/fast-xml-builder-1.1.2.tgz",
-      "integrity": "sha512-NJAmiuVaJEjVa7TjLZKlYd7RqmzOC91EtPFXHvlTcqBVo50Qh7XV5IwvXi1c7NRz2Q/majGX9YLcwJtWgHjtkA==",
+      "version": "1.1.4",
+      "resolved": "https://registry.npmjs.org/fast-xml-builder/-/fast-xml-builder-1.1.4.tgz",
+      "integrity": "sha512-f2jhpN4Eccy0/Uz9csxh3Nu6q4ErKxf0XIsasomfOihuSUa3/xw6w8dnOtCDgEItQFJG8KyXPzQXzcODDrrbOg==",
       "funding": [
         {
           "type": "github",
@@ -8711,9 +8711,9 @@
       }
     },
     "node_modules/fast-xml-parser": {
-      "version": "5.5.3",
-      "resolved": "https://registry.npmjs.org/fast-xml-parser/-/fast-xml-parser-5.5.3.tgz",
-      "integrity": "sha512-Ymnuefk6VzAhT3SxLzVUw+nMio/wB1NGypHkgetwtXcK1JfryaHk4DWQFGVwQ9XgzyS5iRZ7C2ZGI4AMsdMZ6A==",
+      "version": "5.5.9",
+      "resolved": "https://registry.npmjs.org/fast-xml-parser/-/fast-xml-parser-5.5.9.tgz",
+      "integrity": "sha512-jldvxr1MC6rtiZKgrFnDSvT8xuH+eJqxqOBThUVjYrxssYTo1avZLGql5l0a0BAERR01CadYzZ83kVEkbyDg+g==",
       "funding": [
         {
           "type": "github",
@@ -8722,9 +8722,9 @@
       ],
       "license": "MIT",
       "dependencies": {
-        "fast-xml-builder": "^1.1.2",
-        "path-expression-matcher": "^1.1.3",
-        "strnum": "^2.1.2"
+        "fast-xml-builder": "^1.1.4",
+        "path-expression-matcher": "^1.2.0",
+        "strnum": "^2.2.2"
       },
       "bin": {
         "fxparser": "src/cli/cli.js"
@@ -8900,9 +8900,9 @@
       }
     },
     "node_modules/flatted": {
-      "version": "3.3.3",
-      "resolved": "https://registry.npmjs.org/flatted/-/flatted-3.3.3.tgz",
-      "integrity": "sha512-GX+ysw4PBCz0PzosHDepZGANEuFCMLrnRTiEy9McGjmkCQYwRq4A/X786G/fjM/+OjsWSU1ZrY5qyARZmO/uwg==",
+      "version": "3.4.2",
+      "resolved": "https://registry.npmjs.org/flatted/-/flatted-3.4.2.tgz",
+      "integrity": "sha512-PjDse7RzhcPkIJwy5t7KPWQSZ9cAbzQXcafsetQoD7sOJRQlGikNbx7yZp2OotDnJyrDcbyRq3Ttb18iYOqkxA==",
       "dev": true,
       "license": "ISC"
     },
@@ -13200,9 +13200,9 @@
       }
     },
     "node_modules/path-expression-matcher": {
-      "version": "1.1.3",
-      "resolved": "https://registry.npmjs.org/path-expression-matcher/-/path-expression-matcher-1.1.3.tgz",
-      "integrity": "sha512-qdVgY8KXmVdJZRSS1JdEPOKPdTiEK/pi0RkcT2sw1RhXxohdujUlJFPuS1TSkevZ9vzd3ZlL7ULl1MHGTApKzQ==",
+      "version": "1.2.0",
+      "resolved": "https://registry.npmjs.org/path-expression-matcher/-/path-expression-matcher-1.2.0.tgz",
+      "integrity": "sha512-DwmPWeFn+tq7TiyJ2CxezCAirXjFxvaiD03npak3cRjlP9+OjTmSy1EpIrEbh+l6JgUundniloMLDQ/6VTdhLQ==",
       "funding": [
         {
           "type": "github",
@@ -15465,9 +15465,9 @@
       }
     },
     "node_modules/strnum": {
-      "version": "2.2.0",
-      "resolved": "https://registry.npmjs.org/strnum/-/strnum-2.2.0.tgz",
-      "integrity": "sha512-Y7Bj8XyJxnPAORMZj/xltsfo55uOiyHcU2tnAVzHUnSJR/KsEX+9RoDeXEnsXtl/CX4fAcrt64gZ13aGaWPeBg==",
+      "version": "2.2.2",
+      "resolved": "https://registry.npmjs.org/strnum/-/strnum-2.2.2.tgz",
+      "integrity": "sha512-DnR90I+jtXNSTXWdwrEy9FakW7UX+qUZg28gj5fk2vxxl7uS/3bpI4fjFYVmdK9etptYBPNkpahuQnEwhwECqA==",
       "funding": [
         {
           "type": "github",
@@ -16469,9 +16469,9 @@
       "license": "MIT"
     },
     "node_modules/undici": {
-      "version": "7.19.0",
-      "resolved": "https://registry.npmjs.org/undici/-/undici-7.19.0.tgz",
-      "integrity": "sha512-Heho1hJD81YChi+uS2RkSjcVO+EQLmLSyUlHyp7Y/wFbxQaGb4WXVKD073JytrjXJVkSZVzoE2MCSOKugFGtOQ==",
+      "version": "7.24.5",
+      "resolved": "https://registry.npmjs.org/undici/-/undici-7.24.5.tgz",
+      "integrity": "sha512-3IWdCpjgxp15CbJnsi/Y9TCDE7HWVN19j1hmzVhoAkY/+CJx449tVxT5wZc1Gwg8J+P0LWvzlBzxYRnHJ+1i7Q==",
       "license": "MIT",
       "engines": {
         "node": ">=20.18.1"

packages/a2a-server/src/config/config.test.ts

@@ -29,6 +29,7 @@ vi.mock('@google/gemini-cli-core', async (importOriginal) => {
     await importOriginal<typeof import('@google/gemini-cli-core')>();
   return {
     ...actual,
+    PRIORITY_YOLO_ALLOW_ALL: 998,
     Config: vi.fn().mockImplementation((params) => {
       const mockConfig = {
         ...params,
@@ -341,11 +342,11 @@ describe('loadConfig', () => {
       );
     });
 
-    it('should default enableAgents to false when not provided', async () => {
+    it('should default enableAgents to true when not provided', async () => {
       await loadConfig(mockSettings, mockExtensionLoader, taskId);
       expect(Config).toHaveBeenCalledWith(
         expect.objectContaining({
-          enableAgents: false,
+          enableAgents: true,
         }),
       );
     });

packages/a2a-server/src/config/config.ts

@@ -128,7 +128,7 @@ export async function loadConfig(
     interactive: !isHeadlessMode(),
     enableInteractiveShell: !isHeadlessMode(),
     ptyInfo: 'auto',
-    enableAgents: settings.experimental?.enableAgents ?? false,
+    enableAgents: settings.experimental?.enableAgents ?? true,
   };
 
   const fileService = new FileDiscoveryService(workspaceDir, {

packages/cli/index.ts

@@ -6,12 +6,19 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-import { main } from './src/gemini.js';
-import { FatalError, writeToStderr } from '@google/gemini-cli-core';
-import { runExitCleanup } from './src/utils/cleanup.js';
+// --- Fast Path for Version ---
+// We check for version flags at the very top to avoid loading any heavy dependencies.
+// process.env.CLI_VERSION is defined during the build process by esbuild.
+if (process.argv.includes('--version') || process.argv.includes('-v')) {
+  console.log(process.env['CLI_VERSION'] || 'unknown');
+  process.exit(0);
+}
 
 // --- Global Entry Point ---
 
+let writeToStderrFn: (message: string) => void = (msg) =>
+  process.stderr.write(msg);
+
 // Suppress known race condition error in node-pty on Windows
 // Tracking bug: https://github.com/microsoft/node-pty/issues/827
 process.on('uncaughtException', (error) => {
@@ -28,13 +35,22 @@ process.on('uncaughtException', (error) => {
   // For other errors, we rely on the default behavior, but since we attached a listener,
   // we must manually replicate it.
   if (error instanceof Error) {
-    writeToStderr(error.stack + '\n');
+    writeToStderrFn(error.stack + '\n');
   } else {
-    writeToStderr(String(error) + '\n');
+    writeToStderrFn(String(error) + '\n');
   }
   process.exit(1);
 });
 
+const [{ main }, { FatalError, writeToStderr }, { runExitCleanup }] =
+  await Promise.all([
+    import('./src/gemini.js'),
+    import('@google/gemini-cli-core'),
+    import('./src/utils/cleanup.js'),
+  ]);
+
+writeToStderrFn = writeToStderr;
+
 main().catch(async (error) => {
   // Set a timeout to force exit if cleanup hangs
   const cleanupTimeout = setTimeout(() => {

packages/cli/src/acp/acpClient.ts

@@ -300,6 +300,7 @@ export class GeminiAgent {
         sessionId,
         this.clientCapabilities.fs,
         config.getFileSystemService(),
+        cwd,
       );
       config.setFileSystemService(acpFileSystemService);
     }
@@ -357,16 +358,6 @@ export class GeminiAgent {
     const { sessionData, sessionPath } =
       await sessionSelector.resolveSession(sessionId);
 
-    if (this.clientCapabilities?.fs) {
-      const acpFileSystemService = new AcpFileSystemService(
-        this.connection,
-        sessionId,
-        this.clientCapabilities.fs,
-        config.getFileSystemService(),
-      );
-      config.setFileSystemService(acpFileSystemService);
-    }
-
     const clientHistory = convertSessionToClientHistory(sessionData.messages);
 
     const geminiClient = config.getGeminiClient();
@@ -440,7 +431,19 @@ export class GeminiAgent {
       throw acp.RequestError.authRequired();
     }
 
-    // 3. Now that we are authenticated, it is safe to initialize the config
+    // 3. Set the ACP FileSystemService (if supported) before config initialization
+    if (this.clientCapabilities?.fs) {
+      const acpFileSystemService = new AcpFileSystemService(
+        this.connection,
+        sessionId,
+        this.clientCapabilities.fs,
+        config.getFileSystemService(),
+        cwd,
+      );
+      config.setFileSystemService(acpFileSystemService);
+    }
+
+    // 4. Now that we are authenticated, it is safe to initialize the config
     // which starts the MCP servers and other heavy resources.
     await config.initialize();
     startupProfiler.flush(config);
@@ -1625,6 +1628,7 @@ function toPermissionOptions(
     case 'info':
     case 'ask_user':
     case 'exit_plan_mode':
+    case 'sandbox_expansion':
       break;
     default: {
       const unreachable: never = confirmation;

packages/cli/src/acp/fileSystemService.test.ts

@@ -4,10 +4,25 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-import { describe, it, expect, vi, beforeEach, type Mocked } from 'vitest';
+import {
+  describe,
+  it,
+  expect,
+  vi,
+  beforeEach,
+  afterEach,
+  type Mocked,
+} from 'vitest';
 import { AcpFileSystemService } from './fileSystemService.js';
 import type { AgentSideConnection } from '@agentclientprotocol/sdk';
 import type { FileSystemService } from '@google/gemini-cli-core';
+import os from 'node:os';
+
+vi.mock('node:os', () => ({
+  default: {
+    homedir: vi.fn(),
+  },
+}));
 
 describe('AcpFileSystemService', () => {
   let mockConnection: Mocked<AgentSideConnection>;
@@ -25,13 +40,19 @@ describe('AcpFileSystemService', () => {
       readTextFile: vi.fn(),
       writeTextFile: vi.fn(),
     };
+    vi.mocked(os.homedir).mockReturnValue('/home/user');
+  });
+
+  afterEach(() => {
+    vi.restoreAllMocks();
   });
 
   describe('readTextFile', () => {
     it.each([
       {
         capability: true,
-        desc: 'connection if capability exists',
+        path: '/path/to/file',
+        desc: 'connection if capability exists and file is inside root',
         setup: () => {
           mockConnection.readTextFile.mockResolvedValue({ content: 'content' });
         },
@@ -45,6 +66,7 @@ describe('AcpFileSystemService', () => {
       },
       {
         capability: false,
+        path: '/path/to/file',
         desc: 'fallback if capability missing',
         setup: () => {
           mockFallback.readTextFile.mockResolvedValue('content');
@@ -56,19 +78,72 @@ describe('AcpFileSystemService', () => {
           expect(mockConnection.readTextFile).not.toHaveBeenCalled();
         },
       },
-    ])('should use $desc', async ({ capability, setup, verify }) => {
+      {
+        capability: true,
+        path: '/outside/file',
+        desc: 'fallback if capability exists but file is outside root',
+        setup: () => {
+          mockFallback.readTextFile.mockResolvedValue('content');
+        },
+        verify: () => {
+          expect(mockFallback.readTextFile).toHaveBeenCalledWith(
+            '/outside/file',
+          );
+          expect(mockConnection.readTextFile).not.toHaveBeenCalled();
+        },
+      },
+      {
+        capability: true,
+        path: '/home/user/.gemini/tmp/file.md',
+        root: '/home/user',
+        desc: 'fallback if file is inside global gemini dir, even if root overlaps',
+        setup: () => {
+          mockFallback.readTextFile.mockResolvedValue('content');
+        },
+        verify: () => {
+          expect(mockFallback.readTextFile).toHaveBeenCalledWith(
+            '/home/user/.gemini/tmp/file.md',
+          );
+          expect(mockConnection.readTextFile).not.toHaveBeenCalled();
+        },
+      },
+    ])(
+      'should use $desc',
+      async ({ capability, path, root, setup, verify }) => {
+        service = new AcpFileSystemService(
+          mockConnection,
+          'session-1',
+          { readTextFile: capability, writeTextFile: true },
+          mockFallback,
+          root || '/path/to',
+        );
+        setup();
+
+        const result = await service.readTextFile(path);
+
+        expect(result).toBe('content');
+        verify();
+      },
+    );
+
+    it('should throw normalized ENOENT error when readTextFile encounters "Resource not found"', async () => {
       service = new AcpFileSystemService(
         mockConnection,
         'session-1',
-        { readTextFile: capability, writeTextFile: true },
+        { readTextFile: true, writeTextFile: true },
         mockFallback,
+        '/path/to',
+      );
+      mockConnection.readTextFile.mockRejectedValue(
+        new Error('Resource not found for document'),
       );
-      setup();
 
-      const result = await service.readTextFile('/path/to/file');
-
-      expect(result).toBe('content');
-      verify();
+      await expect(
+        service.readTextFile('/path/to/missing'),
+      ).rejects.toMatchObject({
+        code: 'ENOENT',
+        message: 'Resource not found for document',
+      });
     });
   });
 
@@ -76,7 +151,8 @@ describe('AcpFileSystemService', () => {
     it.each([
       {
         capability: true,
-        desc: 'connection if capability exists',
+        path: '/path/to/file',
+        desc: 'connection if capability exists and file is inside root',
         verify: () => {
           expect(mockConnection.writeTextFile).toHaveBeenCalledWith({
             path: '/path/to/file',
@@ -88,6 +164,7 @@ describe('AcpFileSystemService', () => {
       },
       {
         capability: false,
+        path: '/path/to/file',
         desc: 'fallback if capability missing',
         verify: () => {
           expect(mockFallback.writeTextFile).toHaveBeenCalledWith(
@@ -97,17 +174,63 @@ describe('AcpFileSystemService', () => {
           expect(mockConnection.writeTextFile).not.toHaveBeenCalled();
         },
       },
-    ])('should use $desc', async ({ capability, verify }) => {
+      {
+        capability: true,
+        path: '/outside/file',
+        desc: 'fallback if capability exists but file is outside root',
+        verify: () => {
+          expect(mockFallback.writeTextFile).toHaveBeenCalledWith(
+            '/outside/file',
+            'content',
+          );
+          expect(mockConnection.writeTextFile).not.toHaveBeenCalled();
+        },
+      },
+      {
+        capability: true,
+        path: '/home/user/.gemini/tmp/file.md',
+        root: '/home/user',
+        desc: 'fallback if file is inside global gemini dir, even if root overlaps',
+        verify: () => {
+          expect(mockFallback.writeTextFile).toHaveBeenCalledWith(
+            '/home/user/.gemini/tmp/file.md',
+            'content',
+          );
+          expect(mockConnection.writeTextFile).not.toHaveBeenCalled();
+        },
+      },
+    ])('should use $desc', async ({ capability, path, root, verify }) => {
       service = new AcpFileSystemService(
         mockConnection,
         'session-1',
         { writeTextFile: capability, readTextFile: true },
         mockFallback,
+        root || '/path/to',
       );
 
-      await service.writeTextFile('/path/to/file', 'content');
+      await service.writeTextFile(path, 'content');
 
       verify();
     });
+
+    it('should throw normalized ENOENT error when writeTextFile encounters "Resource not found"', async () => {
+      service = new AcpFileSystemService(
+        mockConnection,
+        'session-1',
+        { readTextFile: true, writeTextFile: true },
+        mockFallback,
+        '/path/to',
+      );
+      mockConnection.writeTextFile.mockRejectedValue(
+        new Error('Resource not found for directory'),
+      );
+
+      await expect(
+        service.writeTextFile('/path/to/missing', 'content'),
+      ).rejects.toMatchObject({
+        code: 'ENOENT',
+        message: 'Resource not found for directory',
+      });
+    });
   });
 });

packages/cli/src/acp/fileSystemService.ts

@@ -4,44 +4,82 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-import type { FileSystemService } from '@google/gemini-cli-core';
+import { isWithinRoot, type FileSystemService } from '@google/gemini-cli-core';
 import type * as acp from '@agentclientprotocol/sdk';
+import os from 'node:os';
+import path from 'node:path';
 
 /**
  * ACP client-based implementation of FileSystemService
  */
 export class AcpFileSystemService implements FileSystemService {
+  private readonly geminiDir = path.join(os.homedir(), '.gemini');
+
   constructor(
     private readonly connection: acp.AgentSideConnection,
     private readonly sessionId: string,
     private readonly capabilities: acp.FileSystemCapabilities,
     private readonly fallback: FileSystemService,
+    private readonly root: string,
   ) {}
 
+  private shouldUseFallback(filePath: string): boolean {
+    // Files inside the global CLI directory must always use the native file system,
+    // even if the user runs the CLI directly from their home directory (which
+    // would make the IDE's project root overlap with the global directory).
+    return (
+      !isWithinRoot(filePath, this.root) ||
+      isWithinRoot(filePath, this.geminiDir)
+    );
+  }
+
+  private normalizeFileSystemError(err: unknown): never {
+    const errorMessage = err instanceof Error ? err.message : String(err);
+    if (
+      errorMessage.includes('Resource not found') ||
+      errorMessage.includes('ENOENT') ||
+      errorMessage.includes('does not exist') ||
+      errorMessage.includes('No such file')
+    ) {
+      const newErr = new Error(errorMessage) as NodeJS.ErrnoException;
+      newErr.code = 'ENOENT';
+      throw newErr;
+    }
+    throw err;
+  }
+
   async readTextFile(filePath: string): Promise<string> {
-    if (!this.capabilities.readTextFile) {
+    if (!this.capabilities.readTextFile || this.shouldUseFallback(filePath)) {
       return this.fallback.readTextFile(filePath);
     }
 
-    // eslint-disable-next-line @typescript-eslint/no-unsafe-assignment
-    const response = await this.connection.readTextFile({
-      path: filePath,
-      sessionId: this.sessionId,
-    });
+    try {
+      // eslint-disable-next-line @typescript-eslint/no-unsafe-assignment
+      const response = await this.connection.readTextFile({
+        path: filePath,
+        sessionId: this.sessionId,
+      });
 
-    // eslint-disable-next-line @typescript-eslint/no-unsafe-return
-    return response.content;
+      // eslint-disable-next-line @typescript-eslint/no-unsafe-return
+      return response.content;
+    } catch (err: unknown) {
+      this.normalizeFileSystemError(err);
+    }
   }
 
   async writeTextFile(filePath: string, content: string): Promise<void> {
-    if (!this.capabilities.writeTextFile) {
+    if (!this.capabilities.writeTextFile || this.shouldUseFallback(filePath)) {
       return this.fallback.writeTextFile(filePath, content);
     }
 
-    await this.connection.writeTextFile({
-      path: filePath,
-      content,
-      sessionId: this.sessionId,
-    });
+    try {
+      await this.connection.writeTextFile({
+        path: filePath,
+        content,
+        sessionId: this.sessionId,
+      });
+    } catch (err: unknown) {
+      this.normalizeFileSystemError(err);
+    }
   }
 }

packages/cli/src/config/config.test.ts

@@ -2103,52 +2103,6 @@ describe('loadCliConfig compressionThreshold', () => {
   });
 });
 
-describe('loadCliConfig showContextWindowWarning', () => {
-  beforeEach(() => {
-    vi.resetAllMocks();
-    vi.mocked(os.homedir).mockReturnValue('/mock/home/user');
-    vi.stubEnv('GEMINI_API_KEY', 'test-api-key');
-    vi.spyOn(ExtensionManager.prototype, 'getExtensions').mockReturnValue([]);
-  });
-
-  afterEach(() => {
-    vi.unstubAllEnvs();
-    vi.restoreAllMocks();
-  });
-
-  it('should pass showContextWindowWarning from settings to config (true)', async () => {
-    process.argv = ['node', 'script.js'];
-    const argv = await parseArguments(createTestMergedSettings());
-    const settings = createTestMergedSettings({
-      ui: {
-        showContextWindowWarning: true,
-      },
-    });
-    const config = await loadCliConfig(settings, 'test-session', argv);
-    expect(config.getShowContextWindowWarning()).toBe(true);
-  });
-
-  it('should pass showContextWindowWarning from settings to config (false)', async () => {
-    process.argv = ['node', 'script.js'];
-    const argv = await parseArguments(createTestMergedSettings());
-    const settings = createTestMergedSettings({
-      ui: {
-        showContextWindowWarning: false,
-      },
-    });
-    const config = await loadCliConfig(settings, 'test-session', argv);
-    expect(config.getShowContextWindowWarning()).toBe(false);
-  });
-
-  it('should default to false if not in settings', async () => {
-    process.argv = ['node', 'script.js'];
-    const argv = await parseArguments(createTestMergedSettings());
-    const settings = createTestMergedSettings();
-    const config = await loadCliConfig(settings, 'test-session', argv);
-    expect(config.getShowContextWindowWarning()).toBe(false);
-  });
-});
-
 describe('loadCliConfig useRipgrep', () => {
   beforeEach(() => {
     vi.resetAllMocks();

packages/cli/src/config/config.ts

@@ -955,8 +955,6 @@ export async function loadCliConfig(
     bugCommand: settings.advanced?.bugCommand,
     model: resolvedModel,
     maxSessionTurns: settings.model?.maxSessionTurns,
-    showContextWindowWarning: settings.ui?.showContextWindowWarning,
-    showContextCompression: settings.ui?.showContextCompression,
 
     listExtensions: argv.listExtensions || false,
     listSessions: argv.listSessions || false,

packages/cli/src/config/settingsSchema.test.ts

@@ -400,7 +400,7 @@ describe('SettingsSchema', () => {
       expect(setting).toBeDefined();
       expect(setting.type).toBe('boolean');
       expect(setting.category).toBe('Experimental');
-      expect(setting.default).toBe(false);
+      expect(setting.default).toBe(true);
       expect(setting.requiresRestart).toBe(true);
       expect(setting.showInDialog).toBe(false);
       expect(setting.description).toBe('Enable local and remote subagents.');

packages/cli/src/config/settingsSchema.ts

@@ -575,31 +575,11 @@ const SETTINGS_SCHEMA = {
         label: 'Hide Context Summary',
         category: 'UI',
         requiresRestart: false,
-        default: true,
+        default: false,
         description:
           'Hide the context summary (GEMINI.md, MCP servers) above the input.',
         showInDialog: true,
       },
-      showContextWindowWarning: {
-        type: 'boolean',
-        label: 'Show Context Window Warning',
-        category: 'UI',
-        requiresRestart: false,
-        default: false,
-        description:
-          'Show a warning message when the context window limit is nearly reached. If disabled, the CLI will attempt to automatically compress the history when the limit is reached.',
-        showInDialog: true,
-      },
-      showContextCompression: {
-        type: 'boolean',
-        label: 'Show Context Compression Messages',
-        category: 'UI',
-        requiresRestart: false,
-        default: false,
-        description:
-          'Show a message in the chat history when it is compressed.',
-        showInDialog: true,
-      },
       footer: {
         type: 'object',
         label: 'Footer',
@@ -1228,6 +1208,16 @@ const SETTINGS_SCHEMA = {
               'Disable user input on browser window during automation.',
             showInDialog: false,
           },
+          maxActionsPerTask: {
+            type: 'number',
+            label: 'Max Actions Per Task',
+            category: 'Advanced',
+            requiresRestart: false,
+            default: 100,
+            description:
+              'The maximum number of tool calls allowed per browser task. Enforcement is hard: the agent will be terminated when the limit is reached.',
+            showInDialog: false,
+          },
           confirmSensitiveActions: {
             type: 'boolean',
             label: 'Confirm Sensitive Actions',
@@ -1952,7 +1942,7 @@ const SETTINGS_SCHEMA = {
         label: 'Enable Agents',
         category: 'Experimental',
         requiresRestart: true,
-        default: false,
+        default: true,
         description: 'Enable local and remote subagents.',
         showInDialog: false,
       },

packages/cli/src/ui/AppContainer.tsx

@@ -700,7 +700,10 @@ export const AppContainer = (props: AppContainerProps) => {
 
   // Derive auth state variables for backward compatibility with UIStateContext
   const isAuthDialogOpen = authState === AuthState.Updating;
-  const isAuthenticating = authState === AuthState.Unauthenticated;
+  // TODO: Consider handling other auth types that should also skip the blocking screen
+  const isAuthenticating =
+    authState === AuthState.Unauthenticated &&
+    settings.merged.security.auth.selectedType !== AuthType.USE_GEMINI;
 
   // Session browser and resume functionality
   const isGeminiClientInitialized = config.getGeminiClient()?.isInitialized();
@@ -1300,7 +1303,8 @@ Logging in with Google... Restarting Gemini CLI to continue.
         return;
       }
 
-      if (isSlash || (isIdle && isMcpReady)) {
+      const isMcpOrConfigReady = isConfigInitialized && isMcpReady;
+      if ((isSlash && isConfigInitialized) || (isIdle && isMcpOrConfigReady)) {
         if (!isSlash) {
           const permissions = await checkPermissions(submittedValue, config);
           if (permissions.length > 0) {
@@ -1323,10 +1327,12 @@ Logging in with Google... Restarting Gemini CLI to continue.
         void submitQuery(submittedValue);
       } else {
         // Check messageQueue.length === 0 to only notify on the first queued item
-        if (isIdle && !isMcpReady && messageQueue.length === 0) {
+        if (isIdle && !isMcpOrConfigReady && messageQueue.length === 0) {
           coreEvents.emitFeedback(
             'info',
-            'Waiting for MCP servers to initialize... Slash commands are still available and prompts will be queued.',
+            !isConfigInitialized
+              ? 'Initializing... Prompts will be queued.'
+              : 'Waiting for MCP servers to initialize... Slash commands are still available and prompts will be queued.',
           );
         }
         addMessage(submittedValue);
@@ -1350,6 +1356,7 @@ Logging in with Google... Restarting Gemini CLI to continue.
       refreshStatic,
       reset,
       handleHintSubmit,
+      isConfigInitialized,
       triggerExpandHint,
     ],
   );
@@ -1380,11 +1387,9 @@ Logging in with Google... Restarting Gemini CLI to continue.
    * - Any future streaming states not explicitly allowed
    */
   const isInputActive =
-    isConfigInitialized &&
     !initError &&
     !isProcessing &&
     !isResuming &&
-    !!slashCommands &&
     (streamingState === StreamingState.Idle ||
       streamingState === StreamingState.Responding ||
       streamingState === StreamingState.WaitingForConfirmation) &&

packages/cli/src/ui/commands/compressCommand.test.ts

@@ -4,42 +4,28 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-import * as Core from '@google/gemini-cli-core';
-import { vi, describe, it, expect, beforeEach, afterEach } from 'vitest';
+import {
+  CompressionStatus,
+  type ChatCompressionInfo,
+  type GeminiClient,
+} from '@google/gemini-cli-core';
+import { vi, describe, it, expect, beforeEach } from 'vitest';
 import { compressCommand } from './compressCommand.js';
 import { createMockCommandContext } from '../../test-utils/mockCommandContext.js';
 import { MessageType } from '../types.js';
 
-vi.mock('@google/gemini-cli-core', async (importOriginal) => {
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
-  const actual = (await importOriginal()) as any;
-  return {
-    ...actual,
-    tokenLimit: vi.fn(),
-  };
-});
-
 describe('compressCommand', () => {
   let context: ReturnType<typeof createMockCommandContext>;
   let mockTryCompressChat: ReturnType<typeof vi.fn>;
 
-  afterEach(() => {
-    vi.restoreAllMocks();
-  });
-
   beforeEach(() => {
     mockTryCompressChat = vi.fn();
-    vi.mocked(Core.tokenLimit).mockReturnValue(1000);
     context = createMockCommandContext({
       services: {
         agentContext: {
-          config: {
-            getModel: () => 'test-model',
-            getContextWindowCompressionThreshold: () => 0.2,
-          },
           geminiClient: {
             tryCompressChat: mockTryCompressChat,
-          } as unknown as Core.GeminiClient,
+          } as unknown as GeminiClient,
         },
       },
     });
@@ -50,10 +36,9 @@ describe('compressCommand', () => {
       type: MessageType.COMPRESSION,
       compression: {
         isPending: true,
-        beforePercentage: null,
-        afterPercentage: null,
+        originalTokenCount: null,
+        newTokenCount: null,
         compressionStatus: null,
-        isManual: true,
       },
     };
     await compressCommand.action!(context, '');
@@ -69,9 +54,9 @@ describe('compressCommand', () => {
   });
 
   it('should set pending item, call tryCompressChat, and add result on success', async () => {
-    const compressedResult: Core.ChatCompressionInfo = {
+    const compressedResult: ChatCompressionInfo = {
       originalTokenCount: 200,
-      compressionStatus: Core.CompressionStatus.COMPRESSED,
+      compressionStatus: CompressionStatus.COMPRESSED,
       newTokenCount: 100,
     };
     mockTryCompressChat.mockResolvedValue(compressedResult);
@@ -83,9 +68,8 @@ describe('compressCommand', () => {
       compression: {
         isPending: true,
         compressionStatus: null,
-        beforePercentage: null,
-        afterPercentage: null,
-        isManual: true,
+        originalTokenCount: null,
+        newTokenCount: null,
       },
     });
 
@@ -99,11 +83,9 @@ describe('compressCommand', () => {
         type: MessageType.COMPRESSION,
         compression: {
           isPending: false,
-          compressionStatus: Core.CompressionStatus.COMPRESSED,
-          beforePercentage: 20,
-          afterPercentage: 10,
-          isManual: true,
-          thresholdPercentage: 20,
+          compressionStatus: CompressionStatus.COMPRESSED,
+          originalTokenCount: 200,
+          newTokenCount: 100,
         },
       },
       expect.any(Number),

packages/cli/src/ui/commands/compressCommand.ts

@@ -6,7 +6,6 @@
 
 import { MessageType, type HistoryItemCompression } from '../types.js';
 import { CommandKind, type SlashCommand } from './types.js';
-import { tokenLimit, type CompressionStatus } from '@google/gemini-cli-core';
 
 export const compressCommand: SlashCommand = {
   name: 'compress',
@@ -15,21 +14,7 @@ export const compressCommand: SlashCommand = {
   kind: CommandKind.BUILT_IN,
   autoExecute: true,
   action: async (context) => {
-    const { ui, services } = context;
-    const agentContext = services.agentContext;
-    if (!agentContext) {
-      ui.addItem(
-        {
-          type: MessageType.ERROR,
-          text: 'Agent context not found.',
-        },
-        Date.now(),
-      );
-      return;
-    }
-
-    const config = agentContext.config;
-
+    const { ui } = context;
     if (ui.pendingItem) {
       ui.addItem(
         {
@@ -45,43 +30,29 @@ export const compressCommand: SlashCommand = {
       type: MessageType.COMPRESSION,
       compression: {
         isPending: true,
-        beforePercentage: null,
-        afterPercentage: null,
+        originalTokenCount: null,
+        newTokenCount: null,
         compressionStatus: null,
-        isManual: true,
       },
     };
 
     try {
       ui.setPendingItem(pendingMessage);
       const promptId = `compress-${Date.now()}`;
-      const compressed = await agentContext.geminiClient.tryCompressChat(
-        promptId,
-        true,
-      );
+      const compressed =
+        await context.services.agentContext?.geminiClient?.tryCompressChat(
+          promptId,
+          true,
+        );
       if (compressed) {
-        const limit = tokenLimit(config.getModel());
-        const threshold = config.getContextWindowCompressionThreshold();
-        const beforePercentage = Math.round(
-          (compressed.originalTokenCount / limit) * 100,
-        );
-        const afterPercentage = Math.round(
-          (compressed.newTokenCount / limit) * 100,
-        );
-
         ui.addItem(
           {
             type: MessageType.COMPRESSION,
             compression: {
               isPending: false,
-              beforePercentage,
-              afterPercentage,
-              // eslint-disable-next-line @typescript-eslint/no-unsafe-type-assertion
-              compressionStatus: Number(
-                compressed.compressionStatus,
-              ) as unknown as CompressionStatus,
-              isManual: true,
-              thresholdPercentage: Math.round(threshold * 100),
+              originalTokenCount: compressed.originalTokenCount,
+              newTokenCount: compressed.newTokenCount,
+              compressionStatus: compressed.compressionStatus,
             },
           } as HistoryItemCompression,
           Date.now(),

packages/cli/src/ui/components/Composer.test.tsx

@@ -17,7 +17,11 @@ import {
 import { ConfigContext } from '../contexts/ConfigContext.js';
 import { SettingsContext } from '../contexts/SettingsContext.js';
 import { createMockSettings } from '../../test-utils/settings.js';
-import { ApprovalMode, CoreToolCallStatus } from '@google/gemini-cli-core';
+import {
+  ApprovalMode,
+  tokenLimit,
+  CoreToolCallStatus,
+} from '@google/gemini-cli-core';
 import type { Config } from '@google/gemini-cli-core';
 import { StreamingState } from '../types.js';
 import { TransientMessageType } from '../../utils/events.js';
@@ -729,6 +733,37 @@ describe('Composer', () => {
       expect(output).toContain('Press Esc again to rewind.');
       expect(output).not.toContain('ContextSummaryDisplay');
     });
+
+    it('shows context usage bleed-through when over 60%', async () => {
+      const model = 'gemini-2.5-pro';
+      const uiState = createMockUIState({
+        cleanUiDetailsVisible: false,
+        currentModel: model,
+        sessionStats: {
+          sessionId: 'test-session',
+          sessionStartTime: new Date(),
+          // eslint-disable-next-line @typescript-eslint/no-explicit-any
+          metrics: {} as any,
+          lastPromptTokenCount: Math.floor(tokenLimit(model) * 0.7),
+          promptCount: 0,
+        },
+      });
+      const settings = createMockSettings({
+        ui: {
+          footer: { hideContextPercentage: false },
+        },
+      });
+
+      const { lastFrame } = await renderComposer(uiState, settings);
+
+      await act(async () => {
+        await vi.advanceTimersByTimeAsync(250);
+      });
+
+      // StatusDisplay (which contains ContextUsageDisplay) should bleed through in minimal mode
+      expect(lastFrame()).toContain('StatusDisplay');
+      expect(lastFrame()).toContain('70% used');
+    });
   });
 
   describe('Error Details Display', () => {

packages/cli/src/ui/components/Composer.tsx

@@ -20,11 +20,13 @@ import { useVimMode } from '../contexts/VimModeContext.js';
 import { useAlternateBuffer } from '../hooks/useAlternateBuffer.js';
 import { useTerminalSize } from '../hooks/useTerminalSize.js';
 import { isNarrowWidth } from '../utils/isNarrowWidth.js';
+import { isContextUsageHigh } from '../utils/contextUsage.js';
 import { theme } from '../semantic-colors.js';
 import { GENERIC_WORKING_LABEL } from '../textConstants.js';
 import { INTERACTIVE_SHELL_WAITING_PHRASE } from '../hooks/usePhraseCycler.js';
 import { StreamingState, type HistoryItemToolGroup } from '../types.js';
 import { LoadingIndicator } from './LoadingIndicator.js';
+import { ContextUsageDisplay } from './ContextUsageDisplay.js';
 import { StatusDisplay } from './StatusDisplay.js';
 import { HorizontalLine } from './shared/HorizontalLine.js';
 import { ToastDisplay, shouldShowToast } from './ToastDisplay.js';
@@ -255,6 +257,11 @@ export const Composer = ({ isFocused = true }: { isFocused?: boolean }) => {
   const miniMode_ShowShortcuts = shouldReserveSpaceForShortcutsHint;
   const miniMode_ShowStatus = showLoadingIndicator || hasAnyHooks;
   const miniMode_ShowTip = showTipLine;
+  const miniMode_ShowContext = isContextUsageHigh(
+    uiState.sessionStats.lastPromptTokenCount,
+    uiState.currentModel,
+    settings.merged.model?.compressionThreshold,
+  );
 
   // Composite Mini Mode Triggers
   const showRow1_MiniMode =
@@ -263,7 +270,7 @@ export const Composer = ({ isFocused = true }: { isFocused?: boolean }) => {
     miniMode_ShowShortcuts ||
     miniMode_ShowTip;
 
-  const showRow2_MiniMode = miniMode_ShowApprovalMode;
+  const showRow2_MiniMode = miniMode_ShowApprovalMode || miniMode_ShowContext;
 
   // Final Display Rules (Stable Footer Architecture)
   const showRow1 = showUiDetails || showRow1_MiniMode;
@@ -481,9 +488,22 @@ export const Composer = ({ isFocused = true }: { isFocused?: boolean }) => {
               alignItems="center"
               marginLeft={isNarrow ? 1 : 0}
             >
-              {showUiDetails && (
+              {(showUiDetails || miniMode_ShowContext) && (
                 <StatusDisplay hideContextSummary={hideContextSummary} />
               )}
+              {miniMode_ShowContext && !showUiDetails && (
+                <Box marginLeft={1}>
+                  <ContextUsageDisplay
+                    promptTokenCount={uiState.sessionStats.lastPromptTokenCount}
+                    model={
+                      typeof uiState.currentModel === 'string'
+                        ? uiState.currentModel
+                        : undefined
+                    }
+                    terminalWidth={uiState.terminalWidth}
+                  />
+                </Box>
+              )}
             </Box>
           </Box>
         )}
@@ -498,12 +518,8 @@ export const Composer = ({ isFocused = true }: { isFocused?: boolean }) => {
       flexGrow={0}
       flexShrink={0}
     >
-      {(!uiState.slashCommands ||
-        !uiState.isConfigInitialized ||
-        uiState.isResuming) && (
-        <ConfigInitDisplay
-          message={uiState.isResuming ? 'Resuming session...' : undefined}
-        />
+      {uiState.isResuming && (
+        <ConfigInitDisplay message="Resuming session..." />
       )}
 
       {showUiDetails && (

packages/cli/src/ui/components/ExitPlanModeDialog.tsx

@@ -80,7 +80,6 @@ function usePlanContent(planPath: string, config: Config): PlanContentState {
         const pathError = await validatePlanPath(
           planPath,
           config.storage.getPlansDir(),
-          config.getTargetDir(),
         );
         if (ignore) return;
         if (pathError) {

packages/cli/src/ui/components/StatusDisplay.test.tsx

@@ -72,9 +72,7 @@ const createMockConfig = (overrides = {}) => ({
 const renderStatusDisplay = async (
   props: { hideContextSummary: boolean } = { hideContextSummary: false },
   uiState: UIState = createMockUIState(),
-  settings = createMockSettings({
-    ui: { hideContextSummary: true },
-  }),
+  settings = createMockSettings(),
   config = createMockConfig(),
 ) => {
   const result = await render(
@@ -99,21 +97,16 @@ describe('StatusDisplay', () => {
     vi.restoreAllMocks();
   });
 
-  it('renders nothing by default', async () => {
-    const { lastFrame, unmount } = await renderStatusDisplay();
+  it('renders nothing by default if context summary is hidden via props', async () => {
+    const { lastFrame, unmount } = await renderStatusDisplay({
+      hideContextSummary: true,
+    });
     expect(lastFrame({ allowEmpty: true })).toBe('');
     unmount();
   });
 
-  it('renders ContextSummaryDisplay when hideContextSummary is false', async () => {
-    const settings = createMockSettings({
-      ui: { hideContextSummary: false },
-    });
-    const { lastFrame, unmount } = await renderStatusDisplay(
-      { hideContextSummary: false },
-      undefined,
-      settings,
-    );
+  it('renders ContextSummaryDisplay by default', async () => {
+    const { lastFrame, unmount } = await renderStatusDisplay();
     expect(lastFrame()).toMatchSnapshot();
     unmount();
   });
@@ -125,6 +118,34 @@ describe('StatusDisplay', () => {
     unmount();
   });
 
+  it('renders HookStatusDisplay when hooks are active', async () => {
+    const uiState = createMockUIState({
+      activeHooks: [{ name: 'hook', eventName: 'event' }],
+    });
+    const { lastFrame, unmount } = await renderStatusDisplay(
+      { hideContextSummary: false },
+      uiState,
+    );
+    expect(lastFrame()).toMatchSnapshot();
+    unmount();
+  });
+
+  it('does NOT render HookStatusDisplay if notifications are disabled in settings', async () => {
+    const uiState = createMockUIState({
+      activeHooks: [{ name: 'hook', eventName: 'event' }],
+    });
+    const settings = createMockSettings({
+      hooksConfig: { notifications: false },
+    });
+    const { lastFrame, unmount } = await renderStatusDisplay(
+      { hideContextSummary: false },
+      uiState,
+      settings,
+    );
+    expect(lastFrame()).toMatchSnapshot();
+    unmount();
+  });
+
   it('hides ContextSummaryDisplay if configured in settings', async () => {
     const settings = createMockSettings({
       ui: { hideContextSummary: true },
@@ -142,13 +163,9 @@ describe('StatusDisplay', () => {
     const uiState = createMockUIState({
       backgroundShellCount: 3,
     });
-    const settings = createMockSettings({
-      ui: { hideContextSummary: false },
-    });
     const { lastFrame, unmount } = await renderStatusDisplay(
       { hideContextSummary: false },
       uiState,
-      settings,
     );
     expect(lastFrame()).toContain('Shells: 3');
     unmount();

packages/cli/src/ui/components/ToolConfirmationQueue.test.tsx

@@ -47,6 +47,7 @@ describe('ToolConfirmationQueue', () => {
   const mockConfig = {
     isTrustedFolder: () => true,
     getIdeMode: () => false,
+    getApprovalMode: () => 'default',
     getDisableAlwaysAllow: () => false,
     getModel: () => 'gemini-pro',
     getDebugMode: () => false,

packages/cli/src/ui/components/__snapshots__/StatusDisplay.test.tsx.snap

@@ -1,6 +1,11 @@
 // Vitest Snapshot v1, https://vitest.dev/guide/snapshot.html
 
-exports[`StatusDisplay > renders ContextSummaryDisplay when hideContextSummary is false 1`] = `
+exports[`StatusDisplay > does NOT render HookStatusDisplay if notifications are disabled in settings 1`] = `
+"Mock Context Summary Display (Skills: 2, Shells: 0)
+"
+`;
+
+exports[`StatusDisplay > renders ContextSummaryDisplay by default 1`] = `
 "Mock Context Summary Display (Skills: 2, Shells: 0)
 "
 `;

packages/cli/src/ui/components/messages/CompressionMessage.test.tsx

@@ -11,22 +11,17 @@ import {
 } from './CompressionMessage.js';
 import { CompressionStatus } from '@google/gemini-cli-core';
 import { type CompressionProps } from '../../types.js';
-import { describe, it, expect, vi, afterEach } from 'vitest';
+import { describe, it, expect } from 'vitest';
 
 describe('<CompressionMessage />', () => {
-  afterEach(() => {
-    vi.restoreAllMocks();
-  });
-
   const createCompressionProps = (
     overrides: Partial<CompressionProps> = {},
   ): CompressionDisplayProps => ({
     compression: {
       isPending: false,
-      beforePercentage: null,
-      afterPercentage: null,
+      originalTokenCount: null,
+      newTokenCount: null,
       compressionStatus: CompressionStatus.COMPRESSED,
-      isManual: true,
       ...overrides,
     },
   });
@@ -34,10 +29,9 @@ describe('<CompressionMessage />', () => {
   describe('pending state', () => {
     it('renders pending message when compression is in progress', async () => {
       const props = createCompressionProps({ isPending: true });
-      const { lastFrame, waitUntilReady, unmount } = await renderWithProviders(
+      const { lastFrame, unmount } = await renderWithProviders(
         <CompressionMessage {...props} />,
       );
-      await waitUntilReady();
       const output = lastFrame();
 
       expect(output).toContain('Compressing chat history');
@@ -49,28 +43,56 @@ describe('<CompressionMessage />', () => {
     it('renders success message when tokens are reduced', async () => {
       const props = createCompressionProps({
         isPending: false,
-        beforePercentage: 22,
-        afterPercentage: 6,
+        originalTokenCount: 100,
+        newTokenCount: 50,
         compressionStatus: CompressionStatus.COMPRESSED,
-        thresholdPercentage: 50,
       });
       const { lastFrame, unmount } = await renderWithProviders(
         <CompressionMessage {...props} />,
       );
       const output = lastFrame();
 
-      expect(output).not.toContain('✦');
+      expect(output).toContain('✦');
       expect(output).toContain(
-        'Context compressed (22% → 6%). Adjust threshold (50%) in /settings.',
+        'Chat history compressed from 100 to 50 tokens.',
       );
       unmount();
     });
+
+    it.each([
+      { original: 50000, newTokens: 25000 }, // Large compression
+      { original: 700000, newTokens: 350000 }, // Very large compression
+    ])(
+      'renders success message for large successful compression (from $original to $newTokens)',
+      async ({ original, newTokens }) => {
+        const props = createCompressionProps({
+          isPending: false,
+          originalTokenCount: original,
+          newTokenCount: newTokens,
+          compressionStatus: CompressionStatus.COMPRESSED,
+        });
+        const { lastFrame, unmount } = await renderWithProviders(
+          <CompressionMessage {...props} />,
+        );
+        const output = lastFrame();
+
+        expect(output).toContain('✦');
+        expect(output).toContain(
+          `compressed from ${original} to ${newTokens} tokens`,
+        );
+        expect(output).not.toContain('Skipping compression');
+        expect(output).not.toContain('did not reduce size');
+        unmount();
+      },
+    );
   });
 
   describe('skipped compression (tokens increased or same)', () => {
     it('renders skip message when compression would increase token count', async () => {
       const props = createCompressionProps({
         isPending: false,
+        originalTokenCount: 50,
+        newTokenCount: 75,
         compressionStatus:
           CompressionStatus.COMPRESSION_FAILED_INFLATED_TOKEN_COUNT,
       });
@@ -79,12 +101,121 @@ describe('<CompressionMessage />', () => {
       );
       const output = lastFrame();
 
-      expect(output).not.toContain('✦');
+      expect(output).toContain('✦');
       expect(output).toContain(
         'Compression was not beneficial for this history size.',
       );
       unmount();
     });
+
+    it('renders skip message when token counts are equal', async () => {
+      const props = createCompressionProps({
+        isPending: false,
+        originalTokenCount: 50,
+        newTokenCount: 50,
+        compressionStatus:
+          CompressionStatus.COMPRESSION_FAILED_INFLATED_TOKEN_COUNT,
+      });
+      const { lastFrame, unmount } = await renderWithProviders(
+        <CompressionMessage {...props} />,
+      );
+      const output = lastFrame();
+
+      expect(output).toContain(
+        'Compression was not beneficial for this history size.',
+      );
+      unmount();
+    });
+  });
+
+  describe('message content validation', () => {
+    it.each([
+      {
+        original: 200,
+        newTokens: 80,
+        expected: 'compressed from 200 to 80 tokens',
+      },
+      {
+        original: 500,
+        newTokens: 150,
+        expected: 'compressed from 500 to 150 tokens',
+      },
+      {
+        original: 1500,
+        newTokens: 400,
+        expected: 'compressed from 1500 to 400 tokens',
+      },
+    ])(
+      'displays correct compression statistics (from $original to $newTokens)',
+      async ({ original, newTokens, expected }) => {
+        const props = createCompressionProps({
+          isPending: false,
+          originalTokenCount: original,
+          newTokenCount: newTokens,
+          compressionStatus: CompressionStatus.COMPRESSED,
+        });
+        const { lastFrame, unmount } = await renderWithProviders(
+          <CompressionMessage {...props} />,
+        );
+        const output = lastFrame();
+
+        expect(output).toContain(expected);
+        unmount();
+      },
+    );
+
+    it.each([
+      { original: 50, newTokens: 60 }, // Increased
+      { original: 100, newTokens: 100 }, // Same
+      { original: 49999, newTokens: 50000 }, // Just under 50k threshold
+    ])(
+      'shows skip message for small histories when new tokens >= original tokens ($original -> $newTokens)',
+      async ({ original, newTokens }) => {
+        const props = createCompressionProps({
+          isPending: false,
+          originalTokenCount: original,
+          newTokenCount: newTokens,
+          compressionStatus:
+            CompressionStatus.COMPRESSION_FAILED_INFLATED_TOKEN_COUNT,
+        });
+        const { lastFrame, unmount } = await renderWithProviders(
+          <CompressionMessage {...props} />,
+        );
+        const output = lastFrame();
+
+        expect(output).toContain(
+          'Compression was not beneficial for this history size.',
+        );
+        expect(output).not.toContain('compressed from');
+        unmount();
+      },
+    );
+
+    it.each([
+      { original: 50000, newTokens: 50100 }, // At 50k threshold
+      { original: 700000, newTokens: 710000 }, // Large history case
+      { original: 100000, newTokens: 100000 }, // Large history, same count
+    ])(
+      'shows compression failure message for large histories when new tokens >= original tokens ($original -> $newTokens)',
+      async ({ original, newTokens }) => {
+        const props = createCompressionProps({
+          isPending: false,
+          originalTokenCount: original,
+          newTokenCount: newTokens,
+          compressionStatus:
+            CompressionStatus.COMPRESSION_FAILED_INFLATED_TOKEN_COUNT,
+        });
+        const { lastFrame, unmount } = await renderWithProviders(
+          <CompressionMessage {...props} />,
+        );
+        const output = lastFrame();
+
+        expect(output).toContain('compression did not reduce size');
+        expect(output).not.toContain('compressed from');
+        expect(output).not.toContain('Compression was not beneficial');
+        unmount();
+      },
+    );
   });
 
   describe('failure states', () => {
@@ -98,9 +229,9 @@ describe('<CompressionMessage />', () => {
       );
       const output = lastFrame();
 
-      expect(output).not.toContain('✦');
+      expect(output).toContain('✦');
       expect(output).toContain(
-        'Chat history compression failed: empty summary.',
+        'Chat history compression failed: the model returned an empty summary.',
       );
       unmount();
     });

packages/cli/src/ui/components/messages/CompressionMessage.tsx

@@ -22,13 +22,11 @@ export interface CompressionDisplayProps {
 export function CompressionMessage({
   compression,
 }: CompressionDisplayProps): React.JSX.Element {
-  const {
-    isPending,
-    beforePercentage,
-    afterPercentage,
-    compressionStatus,
-    thresholdPercentage,
-  } = compression;
+  const { isPending, originalTokenCount, newTokenCount, compressionStatus } =
+    compression;
+
+  const originalTokens = originalTokenCount ?? 0;
+  const newTokens = newTokenCount ?? 0;
 
   const getCompressionText = () => {
     if (isPending) {
@@ -36,19 +34,20 @@ export function CompressionMessage({
     }
 
     switch (compressionStatus) {
-      case CompressionStatus.COMPRESSED: {
-        let text = `Context compressed (${beforePercentage}% → ${afterPercentage}%).`;
-        if (thresholdPercentage != null) {
-          text += ` Adjust threshold (${thresholdPercentage}%) in /settings.`;
-        }
-        return text;
-      }
+      case CompressionStatus.COMPRESSED:
+        return `Chat history compressed from ${originalTokens} to ${newTokens} tokens.`;
       case CompressionStatus.COMPRESSION_FAILED_INFLATED_TOKEN_COUNT:
-        return 'Compression was not beneficial for this history size.';
+        // For smaller histories (< 50k tokens), compression overhead likely exceeds benefits
+        if (originalTokens < 50000) {
+          return 'Compression was not beneficial for this history size.';
+        }
+        // For larger histories where compression should work but didn't,
+        // this suggests an issue with the compression process itself
+        return 'Chat history compression did not reduce size. This may indicate issues with the compression prompt.';
       case CompressionStatus.COMPRESSION_FAILED_TOKEN_COUNT_ERROR:
         return 'Could not compress chat history due to a token counting error.';
       case CompressionStatus.COMPRESSION_FAILED_EMPTY_SUMMARY:
-        return 'Chat history compression failed: empty summary.';
+        return 'Chat history compression failed: the model returned an empty summary.';
       case CompressionStatus.NOOP:
         return 'Nothing to compress.';
       default:
@@ -59,13 +58,20 @@ export function CompressionMessage({
   const text = getCompressionText();
 
   return (
-    <Box flexDirection="row" paddingLeft={1} marginBottom={1}>
-      <Box marginRight={1}>{isPending && <CliSpinner type="dots" />}</Box>
+    <Box flexDirection="row">
+      <Box marginRight={1}>
+        {isPending ? (
+          <CliSpinner type="dots" />
+        ) : (
+          <Text color={theme.text.accent}>✦</Text>
+        )}
+      </Box>
       <Box>
         <Text
-          color={theme.text.secondary}
+          color={
+            compression.isPending ? theme.text.accent : theme.status.success
+          }
           aria-label={SCREEN_READER_MODEL_PREFIX}
-          italic
         >
           {text}
         </Text>

packages/cli/src/ui/components/messages/RedirectionConfirmation.test.tsx

@@ -22,6 +22,7 @@ describe('ToolConfirmationMessage Redirection', () => {
     isTrustedFolder: () => true,
     getIdeMode: () => false,
     getDisableAlwaysAllow: () => false,
+    getApprovalMode: () => 'default',
   } as unknown as Config;
 
   it('should display redirection warning and tip for redirected commands', async () => {

packages/cli/src/ui/components/messages/SubagentProgressDisplay.tsx

@@ -153,7 +153,7 @@ export const SubagentProgressDisplay: React.FC<
         })}
       </Box>
 
-      {progress.state === 'completed' && progress.result && (
+      {progress.result && (
         <Box flexDirection="column" marginTop={1}>
           {progress.terminateReason && progress.terminateReason !== 'GOAL' && (
             <Box marginBottom={1}>
@@ -164,7 +164,7 @@ export const SubagentProgressDisplay: React.FC<
           )}
           <MarkdownDisplay
             text={safeJsonToMarkdown(progress.result)}
-            isPending={false}
+            isPending={progress.state !== 'completed'}
             terminalWidth={terminalWidth}
           />
         </Box>

packages/cli/src/ui/components/messages/ToolConfirmationMessage.test.tsx

@@ -40,6 +40,7 @@ describe('ToolConfirmationMessage', () => {
     isTrustedFolder: () => true,
     getIdeMode: () => false,
     getDisableAlwaysAllow: () => false,
+    getApprovalMode: () => 'default',
   } as unknown as Config;
 
   it('should not display urls if prompt and url are the same', async () => {
@@ -324,6 +325,7 @@ describe('ToolConfirmationMessage', () => {
           isTrustedFolder: () => true,
           getIdeMode: () => false,
           getDisableAlwaysAllow: () => false,
+          getApprovalMode: () => 'default',
         } as unknown as Config;
         const { lastFrame, unmount } = await renderWithProviders(
           <ToolConfirmationMessage
@@ -345,6 +347,7 @@ describe('ToolConfirmationMessage', () => {
           isTrustedFolder: () => false,
           getIdeMode: () => false,
           getDisableAlwaysAllow: () => false,
+          getApprovalMode: () => 'default',
         } as unknown as Config;
 
         const { lastFrame, unmount } = await renderWithProviders(
@@ -380,6 +383,7 @@ describe('ToolConfirmationMessage', () => {
         isTrustedFolder: () => true,
         getIdeMode: () => false,
         getDisableAlwaysAllow: () => false,
+        getApprovalMode: () => 'default',
       } as unknown as Config;
       const { lastFrame, unmount } = await renderWithProviders(
         <ToolConfirmationMessage
@@ -406,6 +410,7 @@ describe('ToolConfirmationMessage', () => {
         isTrustedFolder: () => true,
         getIdeMode: () => false,
         getDisableAlwaysAllow: () => false,
+        getApprovalMode: () => 'default',
       } as unknown as Config;
       const { lastFrame, unmount } = await renderWithProviders(
         <ToolConfirmationMessage
@@ -447,6 +452,7 @@ describe('ToolConfirmationMessage', () => {
         isTrustedFolder: () => true,
         getIdeMode: () => false,
         getDisableAlwaysAllow: () => false,
+        getApprovalMode: () => 'default',
       } as unknown as Config;
       vi.mocked(useToolActions).mockReturnValue({
         confirm: vi.fn(),
@@ -473,6 +479,7 @@ describe('ToolConfirmationMessage', () => {
         isTrustedFolder: () => true,
         getIdeMode: () => true,
         getDisableAlwaysAllow: () => false,
+        getApprovalMode: () => 'default',
       } as unknown as Config;
       vi.mocked(useToolActions).mockReturnValue({
         confirm: vi.fn(),
@@ -499,6 +506,7 @@ describe('ToolConfirmationMessage', () => {
         isTrustedFolder: () => true,
         getIdeMode: () => true,
         getDisableAlwaysAllow: () => false,
+        getApprovalMode: () => 'default',
       } as unknown as Config;
       vi.mocked(useToolActions).mockReturnValue({
         confirm: vi.fn(),

packages/cli/src/ui/components/messages/ToolConfirmationMessage.tsx

@@ -15,6 +15,7 @@ import {
   type ToolConfirmationPayload,
   ToolConfirmationOutcome,
   type EditorType,
+  ApprovalMode,
   hasRedirection,
   debugLogger,
 } from '@google/gemini-cli-core';
@@ -314,6 +315,31 @@ export const ToolConfirmationMessage: React.FC<
           key: 'No, suggest changes (esc)',
         });
       }
+    } else if (confirmationDetails.type === 'sandbox_expansion') {
+      options.push({
+        label: 'Allow once',
+        value: ToolConfirmationOutcome.ProceedOnce,
+        key: 'Allow once',
+      });
+      if (isTrustedFolder) {
+        options.push({
+          label: 'Allow for this session',
+          value: ToolConfirmationOutcome.ProceedAlways,
+          key: 'Allow for this session',
+        });
+        if (allowPermanentApproval) {
+          options.push({
+            label: 'Allow for all future sessions',
+            value: ToolConfirmationOutcome.ProceedAlwaysAndSave,
+            key: 'Allow for all future sessions',
+          });
+        }
+      }
+      options.push({
+        label: 'No, suggest changes (esc)',
+        value: ToolConfirmationOutcome.Cancel,
+        key: 'No, suggest changes (esc)',
+      });
     } else if (confirmationDetails.type === 'exec') {
       options.push({
         label: 'Allow once',
@@ -546,6 +572,8 @@ export const ToolConfirmationMessage: React.FC<
         if (!confirmationDetails.isModifying) {
           question = `Apply this change?`;
         }
+      } else if (confirmationDetails.type === 'sandbox_expansion') {
+        question = `Allow sandbox expansion for: '${sanitizeForDisplay(confirmationDetails.rootCommand)}'?`;
       } else if (confirmationDetails.type === 'exec') {
         const executionProps = confirmationDetails;
 
@@ -573,6 +601,52 @@ export const ToolConfirmationMessage: React.FC<
             />
           );
         }
+      } else if (confirmationDetails.type === 'sandbox_expansion') {
+        const { additionalPermissions } = confirmationDetails;
+        const readPaths = additionalPermissions?.fileSystem?.read || [];
+        const writePaths = additionalPermissions?.fileSystem?.write || [];
+        const network = additionalPermissions?.network;
+
+        bodyContent = (
+          <Box flexDirection="column" padding={1}>
+            <Text color={theme.text.secondary} italic>
+              The agent is requesting additional sandbox permissions to execute
+              this command:
+            </Text>
+            <Box paddingY={1}>
+              <Text color={theme.text.secondary}>
+                {sanitizeForDisplay(confirmationDetails.command)}
+              </Text>
+            </Box>
+            {network && (
+              <Box>
+                <Text color={theme.status.warning}>• Network Access</Text>
+              </Box>
+            )}
+            {readPaths.length > 0 && (
+              <Box flexDirection="column">
+                <Text color={theme.status.success}>• Read Access:</Text>
+                {readPaths.map((p, i) => (
+                  <Text key={i} color={theme.text.secondary}>
+                    {' '}
+                    {sanitizeForDisplay(p)}
+                  </Text>
+                ))}
+              </Box>
+            )}
+            {writePaths.length > 0 && (
+              <Box flexDirection="column">
+                <Text color={theme.status.error}>• Write Access:</Text>
+                {writePaths.map((p, i) => (
+                  <Text key={i} color={theme.text.secondary}>
+                    {' '}
+                    {sanitizeForDisplay(p)}
+                  </Text>
+                ))}
+              </Box>
+            )}
+          </Box>
+        );
       } else if (confirmationDetails.type === 'exec') {
         const executionProps = confirmationDetails;
 
@@ -587,7 +661,8 @@ export const ToolConfirmationMessage: React.FC<
         let bodyContentHeight = availableBodyContentHeight();
         let warnings: React.ReactNode = null;
 
-        if (containsRedirection) {
+        const isAutoEdit = config.getApprovalMode() === ApprovalMode.AUTO_EDIT;
+        if (containsRedirection && !isAutoEdit) {
           // Calculate lines needed for Note and Tip
           const safeWidth = Math.max(terminalWidth, 1);
           const noteLength =
@@ -737,6 +812,7 @@ export const ToolConfirmationMessage: React.FC<
       isTrustedFolder,
       allowPermanentApproval,
       settings,
+      config,
     ]);
 
   const bodyOverflowDirection: 'top' | 'bottom' =

packages/cli/src/ui/hooks/useGeminiStream.test.tsx

@@ -4,7 +4,7 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-/* eslint-disable @typescript-eslint/no-explicit-any -- TODO: Refactor to remove any usage */
+/* eslint-disable @typescript-eslint/no-explicit-any */
 import {
   describe,
   it,
@@ -32,7 +32,10 @@ import type {
   Config,
   EditorType,
   AnyToolInvocation,
+  AnyDeclarativeTool,
   SpanMetadata,
+  CompletedToolCall,
+  ToolCallRequestInfo,
 } from '@google/gemini-cli-core';
 import {
   CoreToolCallStatus,
@@ -49,15 +52,19 @@ import {
   MCPDiscoveryState,
   GeminiCliOperation,
   getPlanModeExitMessage,
-  CompressionStatus,
 } from '@google/gemini-cli-core';
 import type { Part, PartListUnion } from '@google/genai';
 import type { UseHistoryManagerReturn } from './useHistoryManager.js';
-import type { SlashCommandProcessorResult } from '../types.js';
+import type {
+  SlashCommandProcessorResult,
+  HistoryItemWithoutId,
+  HistoryItem,
+} from '../types.js';
 import { MessageType, StreamingState } from '../types.js';
 
 import type { LoadedSettings } from '../../config/settings.js';
 import { findLastSafeSplitPoint } from '../utils/markdownUtilities.js';
+import { theme } from '../semantic-colors.js';
 
 // --- MOCKS ---
 const mockSendMessageStream = vi
@@ -138,7 +145,6 @@ const mockRunInDevTraceSpan = vi.hoisted(() =>
     };
     return await fn({
       metadata,
-      endSpan: vi.fn(),
     });
   }),
 );
@@ -243,8 +249,10 @@ describe('useGeminiStream', () => {
   let mockMarkToolsAsSubmitted: Mock;
   let handleAtCommandSpy: MockInstance;
 
-  const emptyHistory: any[] = [];
-  let capturedOnComplete: any = null;
+  const emptyHistory: HistoryItem[] = [];
+  let capturedOnComplete:
+    | ((tools: CompletedToolCall[]) => Promise<void>)
+    | null = null;
   const mockGetPreferredEditor = vi.fn(() => 'vscode' as EditorType);
   const mockOnAuthError = vi.fn();
   const mockPerformMemoryRefresh = vi.fn(() => Promise.resolve());
@@ -324,9 +332,6 @@ describe('useGeminiStream', () => {
     })),
     getIdeMode: vi.fn(() => false),
     getEnableHooks: vi.fn(() => false),
-    getShowContextWindowWarning: vi.fn(() => false),
-    getShowContextCompression: vi.fn(() => false),
-    getContextWindowCompressionThreshold: vi.fn(() => 0.2),
   } as unknown as Config;
 
   beforeEach(() => {
@@ -406,13 +411,17 @@ describe('useGeminiStream', () => {
         lastToolCalls,
         mockScheduleToolCalls,
         mockMarkToolsAsSubmitted,
-        (updater: any) => {
+        (
+          updater:
+            | TrackedToolCall[]
+            | ((prev: TrackedToolCall[]) => TrackedToolCall[]),
+        ) => {
           lastToolCalls =
             typeof updater === 'function' ? updater(lastToolCalls) : updater;
           rerender({ ...initialProps, toolCalls: lastToolCalls });
         },
-        (...args: any[]) => {
-          mockCancelAllToolCalls(...args);
+        (signal: AbortSignal) => {
+          mockCancelAllToolCalls(signal);
           lastToolCalls = lastToolCalls.map((tc) => {
             if (
               tc.status === CoreToolCallStatus.AwaitingApproval ||
@@ -879,7 +888,7 @@ describe('useGeminiStream', () => {
     const fn = spanArgs[1];
     const metadata = { attributes: {} };
     await act(async () => {
-      await fn({ metadata, endSpan: vi.fn() });
+      await fn({ metadata });
     });
     expect(metadata).toMatchObject({
       input: sentParts,
@@ -973,7 +982,7 @@ describe('useGeminiStream', () => {
   });
 
   it('should stop agent execution immediately when a tool call returns STOP_EXECUTION error', async () => {
-    const stopExecutionToolCalls: TrackedToolCall[] = [
+    const stopExecutionToolCalls: TrackedCompletedToolCall[] = [
       {
         request: {
           callId: 'stop-call',
@@ -1045,7 +1054,7 @@ describe('useGeminiStream', () => {
   });
 
   it('should add a compact suppressed-error note before STOP_EXECUTION terminal info in low verbosity mode', async () => {
-    const stopExecutionToolCalls: TrackedToolCall[] = [
+    const stopExecutionToolCalls: TrackedCompletedToolCall[] = [
       {
         request: {
           callId: 'stop-call',
@@ -1072,9 +1081,10 @@ describe('useGeminiStream', () => {
       } as unknown as TrackedCompletedToolCall,
     ];
     const lowVerbositySettings = {
-      ...(mockLoadedSettings as any),
+      // eslint-disable-next-line @typescript-eslint/no-misused-spread
+      ...mockLoadedSettings,
       merged: {
-        ...(mockLoadedSettings.merged as any),
+        ...mockLoadedSettings.merged,
         ui: { errorVerbosity: 'low' },
       },
     } as LoadedSettings;
@@ -1925,6 +1935,120 @@ describe('useGeminiStream', () => {
         expect(mockHandleSlashCommand).not.toHaveBeenCalled();
       });
     });
+
+    it('should record client-initiated tool calls in GeminiChat history', async () => {
+      const { result, client: mockGeminiClient } = await renderTestHook();
+
+      mockHandleSlashCommand.mockResolvedValue({
+        type: 'schedule_tool',
+        toolName: 'activate_skill',
+        toolArgs: { name: 'test-skill' },
+      });
+
+      await act(async () => {
+        await result.current.submitQuery('/test-skill');
+      });
+
+      // Simulate tool completion
+      const completedTool = {
+        request: {
+          callId: 'test-call-id',
+          name: 'activate_skill',
+          args: { name: 'test-skill' },
+          isClientInitiated: true,
+        },
+        status: CoreToolCallStatus.Success,
+        invocation: {
+          getDescription: () => 'Activating skill test-skill',
+        },
+        tool: {
+          isOutputMarkdown: true,
+        },
+        response: {
+          responseParts: [
+            {
+              functionResponse: {
+                name: 'activate_skill',
+                response: { content: 'skill instructions' },
+              },
+            },
+          ],
+        },
+      } as unknown as TrackedCompletedToolCall;
+
+      await act(async () => {
+        if (capturedOnComplete) {
+          await capturedOnComplete([completedTool]);
+        }
+      });
+
+      // Verify that the tool call and response were added to GeminiChat history
+      expect(mockGeminiClient.addHistory).toHaveBeenCalledWith({
+        role: 'model',
+        parts: [
+          {
+            functionCall: {
+              name: 'activate_skill',
+              args: { name: 'test-skill' },
+            },
+          },
+        ],
+      });
+      expect(mockGeminiClient.addHistory).toHaveBeenCalledWith({
+        role: 'user',
+        parts: completedTool.response.responseParts,
+      });
+    });
+
+    it('should NOT record other client-initiated tool calls (like save_memory) in history', async () => {
+      const { result, client: mockGeminiClient } = await renderTestHook();
+
+      mockHandleSlashCommand.mockResolvedValue({
+        type: 'schedule_tool',
+        toolName: 'save_memory',
+        toolArgs: { fact: 'test fact' },
+      });
+
+      await act(async () => {
+        await result.current.submitQuery('/memory add "test fact"');
+      });
+
+      // Simulate tool completion
+      const completedTool = {
+        request: {
+          callId: 'test-call-id',
+          name: 'save_memory',
+          args: { fact: 'test fact' },
+          isClientInitiated: true,
+        },
+        status: CoreToolCallStatus.Success,
+        invocation: {
+          getDescription: () => 'Saving memory',
+        },
+        tool: {
+          isOutputMarkdown: true,
+        },
+        response: {
+          responseParts: [
+            {
+              functionResponse: {
+                name: 'save_memory',
+                response: { success: true },
+              },
+            },
+          ],
+        },
+      } as unknown as TrackedCompletedToolCall;
+
+      await act(async () => {
+        if (capturedOnComplete) {
+          await capturedOnComplete([completedTool]);
+        }
+      });
+
+      // Verify that addHistory was NOT called
+      expect(mockGeminiClient.addHistory).not.toHaveBeenCalled();
+    });
   });
 
   describe('Memory Refresh on save_memory', () => {
@@ -1952,7 +2076,7 @@ describe('useGeminiStream', () => {
           displayName: 'save_memory',
           description: 'Saves memory',
           build: vi.fn(),
-        } as any,
+        } as unknown as AnyDeclarativeTool,
         invocation: {
           getDescription: () => `Mock description`,
         } as unknown as AnyToolInvocation,
@@ -2026,7 +2150,8 @@ describe('useGeminiStream', () => {
       );
 
       const testConfig = {
-        ...(mockConfig as any),
+        // eslint-disable-next-line @typescript-eslint/no-misused-spread
+        ...mockConfig,
         getContentGenerator: vi.fn(),
         getContentGeneratorConfig: vi.fn(() => ({
           authType: mockAuthType,
@@ -2191,7 +2316,7 @@ describe('useGeminiStream', () => {
             displayName: 'replace',
             description: 'Replace text',
             build: vi.fn(),
-          } as any,
+          } as unknown as AnyDeclarativeTool,
           invocation: {
             getDescription: () => 'Mock description',
           } as unknown as AnyToolInvocation,
@@ -2232,7 +2357,7 @@ describe('useGeminiStream', () => {
             displayName: 'write_file',
             description: 'Write file',
             build: vi.fn(),
-          } as any,
+          } as unknown as AnyDeclarativeTool,
           invocation: {
             getDescription: () => 'Mock description',
           } as unknown as AnyToolInvocation,
@@ -2349,32 +2474,22 @@ describe('useGeminiStream', () => {
 
       it.each([
         {
-          name: 'add a message when remaining tokens overflow, regardless of showContextWindowWarning setting',
+          name: 'without suggestion when remaining tokens are > 75% of limit',
           requestTokens: 20,
           remainingTokens: 80,
-          shouldShow: false, // The setting itself is false
           expectedMessage:
-            'Context 20% full. Message may exceed window. Reduce size or /compress.',
+            'Sending this message (20 tokens) might exceed the context window limit (80 tokens left).',
         },
         {
-          name: 'add a message when showContextWindowWarning is true',
+          name: 'with suggestion when remaining tokens are < 75% of limit',
           requestTokens: 30,
           remainingTokens: 70,
-          shouldShow: true,
           expectedMessage:
-            'Context 30% full. Message may exceed window. Reduce size or /compress.',
+            'Sending this message (30 tokens) might exceed the context window limit (70 tokens left). Please try reducing the size of your message or use the `/compress` command to compress the chat history.',
         },
       ])(
-        'should $name',
-        async ({
-          requestTokens,
-          remainingTokens,
-          shouldShow,
-          expectedMessage,
-        }) => {
-          vi.mocked(mockConfig.getShowContextWindowWarning).mockReturnValue(
-            shouldShow,
-          );
+        'should add message $name',
+        async ({ requestTokens, remainingTokens, expectedMessage }) => {
           mockSendMessageStream.mockReturnValue(
             (async function* () {
               yield {
@@ -2394,12 +2509,10 @@ describe('useGeminiStream', () => {
           });
 
           await waitFor(() => {
-            expect(mockAddItem).toHaveBeenCalledWith(
-              expect.objectContaining({
-                type: 'info',
-                text: expectedMessage,
-              }),
-            );
+            expect(mockAddItem).toHaveBeenCalledWith({
+              type: 'info',
+              text: expectedMessage,
+            });
           });
         },
       );
@@ -2453,12 +2566,8 @@ describe('useGeminiStream', () => {
       });
     });
 
-    it('should add informational messages when ChatCompressed event is received and showContextCompression is true', async () => {
+    it('should add informational messages when ChatCompressed event is received', async () => {
       vi.mocked(tokenLimit).mockReturnValue(10000);
-      vi.mocked(
-        mockConfig.getContextWindowCompressionThreshold,
-      ).mockReturnValue(0.2);
-      vi.mocked(mockConfig.getShowContextCompression).mockReturnValue(true);
       // Setup mock to return a stream with ChatCompressed event
       mockSendMessageStream.mockReturnValue(
         (async function* () {
@@ -2467,22 +2576,7 @@ describe('useGeminiStream', () => {
             value: {
               originalTokenCount: 1000,
               newTokenCount: 500,
-              compressionStatus: CompressionStatus.COMPRESSED,
-            },
-          };
-          yield {
-            type: ServerGeminiEventType.Content,
-            value: 'Response after compression',
-          };
-          yield {
-            type: ServerGeminiEventType.Finished,
-            value: {
-              finishReason: 'STOP',
-              usageMetadata: {
-                promptTokenCount: 10,
-                candidatesTokenCount: 20,
-                totalTokenCount: 30,
-              },
+              compressionStatus: 'compressed',
             },
           };
         })(),
@@ -2499,129 +2593,10 @@ describe('useGeminiStream', () => {
       await waitFor(() => {
         expect(mockAddItem).toHaveBeenCalledWith(
           expect.objectContaining({
-            type: 'compression',
-            compression: {
-              isPending: false,
-              beforePercentage: 10,
-              afterPercentage: 5,
-              compressionStatus: CompressionStatus.COMPRESSED,
-              isManual: false,
-              thresholdPercentage: 20,
-            },
-          }),
-          expect.any(Number),
-        );
-      });
-    });
-
-    it('should NOT add informational messages when ChatCompressed event is received and showContextCompression is false', async () => {
-      vi.mocked(tokenLimit).mockReturnValue(10000);
-      vi.mocked(
-        mockConfig.getContextWindowCompressionThreshold,
-      ).mockReturnValue(0.2);
-      vi.mocked(mockConfig.getShowContextCompression).mockReturnValue(false);
-      // Setup mock to return a stream with ChatCompressed event
-      mockSendMessageStream.mockReturnValue(
-        (async function* () {
-          yield {
-            type: ServerGeminiEventType.ChatCompressed,
-            value: {
-              originalTokenCount: 1000,
-              newTokenCount: 500,
-              compressionStatus: CompressionStatus.COMPRESSED,
-            },
-          };
-          yield {
-            type: ServerGeminiEventType.Content,
-            value: 'Response after compression',
-          };
-          yield {
-            type: ServerGeminiEventType.Finished,
-            value: {
-              finishReason: 'STOP',
-              usageMetadata: {
-                promptTokenCount: 10,
-                candidatesTokenCount: 20,
-                totalTokenCount: 30,
-              },
-            },
-          };
-        })(),
-      );
-
-      const { result } = await renderHookWithDefaults();
-
-      // Submit a query
-      await act(async () => {
-        await result.current.submitQuery('Test compression');
-      });
-
-      // Check that NO compression message was added
-      await waitFor(() => {
-        expect(mockAddItem).not.toHaveBeenCalledWith(
-          expect.objectContaining({
-            type: 'compression',
-          }),
-        );
-      });
-    });
-
-    it('should add informational messages when ChatCompressed event is received with a large prompt even if showContextCompression is false', async () => {
-      vi.mocked(tokenLimit).mockReturnValue(10000);
-      vi.mocked(
-        mockConfig.getContextWindowCompressionThreshold,
-      ).mockReturnValue(0.2); // 20%
-      vi.mocked(mockConfig.getShowContextCompression).mockReturnValue(false);
-
-      // Setup mock to return a stream with ChatCompressed event and a large requestTokenCount (25%)
-      mockSendMessageStream.mockReturnValue(
-        (async function* () {
-          yield {
-            type: ServerGeminiEventType.ChatCompressed,
-            value: {
-              originalTokenCount: 1000,
-              newTokenCount: 500,
-              compressionStatus: CompressionStatus.COMPRESSED,
-              requestTokenCount: 2500, // 25% > 20%
-            },
-          };
-          yield {
-            type: ServerGeminiEventType.Content,
-            value: 'Response after compression',
-          };
-          yield {
-            type: ServerGeminiEventType.Finished,
-            value: {
-              finishReason: 'STOP',
-              usageMetadata: {
-                promptTokenCount: 10,
-                candidatesTokenCount: 20,
-                totalTokenCount: 30,
-              },
-            },
-          };
-        })(),
-      );
-
-      const { result } = await renderHookWithDefaults();
-
-      // Submit a query
-      await act(async () => {
-        await result.current.submitQuery('Test large prompt compression');
-      });
-
-      // Check that compression message WAS added despite the setting
-      await waitFor(() => {
-        expect(mockAddItem).toHaveBeenCalledWith(
-          expect.objectContaining({
-            type: 'compression',
-            compression: expect.objectContaining({
-              beforePercentage: 10,
-              afterPercentage: 5,
-              compressionStatus: CompressionStatus.COMPRESSED,
-              isManual: false,
-              thresholdPercentage: 20,
-            }),
+            type: MessageType.INFO,
+            text: 'Context compressed from 10% to 5%.',
+            secondaryText: 'Change threshold in /settings.',
+            color: theme.status.warning,
           }),
           expect.any(Number),
         );
@@ -2727,14 +2702,14 @@ describe('useGeminiStream', () => {
 
   it('should flush pending text rationale before scheduling tool calls to ensure correct history order', async () => {
     const addItemOrder: string[] = [];
-    let capturedOnComplete: any;
+    let capturedOnComplete: (tools: CompletedToolCall[]) => Promise<void>;
 
     const mockScheduleToolCalls = vi.fn(async (requests) => {
       addItemOrder.push('scheduleToolCalls_START');
       // Simulate tools completing and triggering onComplete immediately.
       // This mimics the behavior that caused the regression where tool results
       // were added to history during the await scheduleToolCalls(...) block.
-      const tools = requests.map((r: any) => ({
+      const tools = requests.map((r: ToolCallRequestInfo) => ({
         request: r,
         status: CoreToolCallStatus.Success,
         tool: { displayName: r.name, name: r.name },
@@ -2749,7 +2724,7 @@ describe('useGeminiStream', () => {
       addItemOrder.push('scheduleToolCalls_END');
     });
 
-    mockAddItem.mockImplementation((item: any) => {
+    mockAddItem.mockImplementation((item: HistoryItemWithoutId) => {
       addItemOrder.push(`addItem:${item.type}`);
     });
 
@@ -2979,9 +2954,10 @@ describe('useGeminiStream', () => {
   describe('Thought Reset', () => {
     it('should keep full thinking entries in history when mode is full', async () => {
       const fullThinkingSettings: LoadedSettings = {
-        ...(mockLoadedSettings as any),
+        // eslint-disable-next-line @typescript-eslint/no-misused-spread
+        ...mockLoadedSettings,
         merged: {
-          ...(mockLoadedSettings.merged as any),
+          ...mockLoadedSettings.merged,
           ui: { inlineThinkingMode: 'full' },
         },
       } as unknown as LoadedSettings;
@@ -4060,7 +4036,7 @@ describe('useGeminiStream', () => {
 
     const spanMetadata = {} as SpanMetadata;
     await act(async () => {
-      await userPromptCall![1]({ metadata: spanMetadata, endSpan: vi.fn() });
+      await userPromptCall![1]({ metadata: spanMetadata });
     });
     expect(spanMetadata.input).toBe('telemetry test query');
   });

packages/cli/src/ui/hooks/useGeminiStream.ts

@@ -38,20 +38,22 @@ import {
   GeminiCliOperation,
   getPlanModeExitMessage,
   isBackgroundExecutionData,
-  type CompressionStatus,
   Kind,
-  type Config,
-  type EditorType,
-  type GeminiClient,
-  type ServerGeminiChatCompressedEvent,
-  type ServerGeminiContentEvent as ContentEvent,
-  type ServerGeminiFinishedEvent,
-  type ServerGeminiStreamEvent as GeminiEvent,
-  type ThoughtSummary,
-  type ToolCallRequestInfo,
-  type ToolCallResponseInfo,
-  type GeminiErrorEventValue,
-  type RetryAttemptPayload,
+  ACTIVATE_SKILL_TOOL_NAME,
+} from '@google/gemini-cli-core';
+import type {
+  Config,
+  EditorType,
+  GeminiClient,
+  ServerGeminiChatCompressedEvent,
+  ServerGeminiContentEvent as ContentEvent,
+  ServerGeminiFinishedEvent,
+  ServerGeminiStreamEvent as GeminiEvent,
+  ThoughtSummary,
+  ToolCallRequestInfo,
+  ToolCallResponseInfo,
+  GeminiErrorEventValue,
+  RetryAttemptPayload,
 } from '@google/gemini-cli-core';
 import { type Part, type PartListUnion, FinishReason } from '@google/genai';
 import type {
@@ -59,6 +61,7 @@ import type {
   HistoryItemThinking,
   HistoryItemWithoutId,
   HistoryItemToolGroup,
+  HistoryItemInfo,
   IndividualToolCallDisplay,
   SlashCommandProcessorResult,
   HistoryItemModel,
@@ -254,8 +257,6 @@ export const useGeminiStream = (
   const [_isFirstToolInGroup, isFirstToolInGroupRef, setIsFirstToolInGroup] =
     useStateAndRef<boolean>(true);
   const processedMemoryToolsRef = useRef<Set<string>>(new Set());
-  const handleCompletedToolsRef =
-    useRef<(completedTools: TrackedToolCall[]) => Promise<void>>(undefined);
   const { startNewPrompt, getPromptCount } = useSessionStats();
   const storage = config.storage;
   const logger = useLogger(storage);
@@ -301,23 +302,22 @@ export const useGeminiStream = (
             }),
           );
         }
+
         // Clear the live-updating display now that the final state is in history.
         setToolCallsForDisplay([]);
 
         // Record tool calls with full metadata before sending responses.
         try {
           const currentModel =
-            (typeof geminiClient.getCurrentSequenceModel === 'function'
-              ? geminiClient.getCurrentSequenceModel()
-              : undefined) ?? config.getModel();
-          const chat =
-            typeof geminiClient.getChat === 'function'
-              ? geminiClient.getChat()
-              : undefined;
-          chat?.recordCompletedToolCalls(
-            currentModel,
-            completedToolCallsFromScheduler,
-          );
+            config.getGeminiClient().getCurrentSequenceModel() ??
+            config.getModel();
+          config
+            .getGeminiClient()
+            .getChat()
+            .recordCompletedToolCalls(
+              currentModel,
+              completedToolCallsFromScheduler,
+            );
 
           await recordToolCallInteractions(
             config,
@@ -330,7 +330,7 @@ export const useGeminiStream = (
         }
 
         // Handle tool response submission immediately when tools complete
-        await handleCompletedToolsRef.current?.(
+        await handleCompletedTools(
           completedToolCallsFromScheduler as TrackedToolCall[],
         );
       }
@@ -549,11 +549,9 @@ export const useGeminiStream = (
       if (tc.request.name === ASK_USER_TOOL_NAME && isInProgress) {
         return false;
       }
-      return (
-        tc.status !== 'scheduled' &&
-        tc.status !== 'validating' &&
-        tc.status !== 'awaiting_approval'
-      );
+      // ToolGroupMessage now shows all non-canceled tools, so they are visible
+      // in pending and we need to draw the closing border for them.
+      return true;
     });
 
     if (
@@ -1147,43 +1145,21 @@ export const useGeminiStream = (
       }
 
       const limit = tokenLimit(config.getModel());
-      const beforePercentage =
-        eventValue?.originalTokenCount != null
-          ? Math.round((eventValue.originalTokenCount / limit) * 100)
-          : null;
-      const afterPercentage =
-        eventValue?.newTokenCount != null
-          ? Math.round((eventValue.newTokenCount / limit) * 100)
-          : null;
-
-      const threshold = config.getContextWindowCompressionThreshold();
-      const isLargePrompt =
-        eventValue?.requestTokenCount != null &&
-        eventValue.requestTokenCount / limit > threshold;
-
-      if (!config.getShowContextCompression() && !isLargePrompt) {
-        return;
-      }
+      const originalPercentage = Math.round(
+        ((eventValue?.originalTokenCount ?? 0) / limit) * 100,
+      );
+      const newPercentage = Math.round(
+        ((eventValue?.newTokenCount ?? 0) / limit) * 100,
+      );
 
       addItem(
         {
-          type: 'compression',
-          compression: {
-            isPending: false,
-            beforePercentage,
-            afterPercentage,
-            /* eslint-disable @typescript-eslint/no-unsafe-type-assertion */
-            compressionStatus: eventValue
-              ? (Number(
-                  eventValue.compressionStatus,
-                ) as unknown as CompressionStatus)
-              : null,
-            /* eslint-enable @typescript-eslint/no-unsafe-type-assertion */
-            isManual: false,
-            thresholdPercentage: Math.round(threshold * 100),
-          },
-          timestamp: new Date(userMessageTimestamp),
-        } as HistoryItemWithoutId,
+          type: MessageType.INFO,
+          text: `Context compressed from ${originalPercentage}% to ${newPercentage}%.`,
+          secondaryText: `Change threshold in /settings.`,
+          color: theme.status.warning,
+          marginBottom: 1,
+        } as HistoryItemInfo,
         userMessageTimestamp,
       );
     },
@@ -1206,11 +1182,16 @@ export const useGeminiStream = (
       onCancelSubmit(true);
 
       const limit = tokenLimit(config.getModel());
-      const usedPercentage = Math.round(
-        ((limit - remainingTokenCount) / limit) * 100,
-      );
 
-      const text = `Context ${usedPercentage}% full. Message may exceed window. Reduce size or /compress.`;
+      const isMoreThan25PercentUsed =
+        limit > 0 && remainingTokenCount < limit * 0.75;
+
+      let text = `Sending this message (${estimatedRequestTokenCount} tokens) might exceed the context window limit (${remainingTokenCount.toLocaleString()} tokens left).`;
+
+      if (isMoreThan25PercentUsed) {
+        text +=
+          ' Please try reducing the size of your message or use the `/compress` command to compress the chat history.';
+      }
 
       addItem({
         type: 'info',
@@ -1557,7 +1538,8 @@ export const useGeminiStream = (
                     setLoopDetectionConfirmationRequest(null);
 
                     if (result.userSelection === 'disable') {
-                      geminiClient
+                      config
+                        .getGeminiClient()
                         .getLoopDetectionService()
                         .disableForSession();
                       addItem({
@@ -1675,7 +1657,7 @@ export const useGeminiStream = (
       ) {
         let awaitingApprovalCalls = toolCalls.filter(
           (call): call is TrackedWaitingToolCall =>
-            call.status === 'awaiting_approval',
+            call.status === 'awaiting_approval' && !call.request.forcedAsk,
         );
 
         // For AUTO_EDIT mode, only approve edit tools (replace, write_file)
@@ -1733,24 +1715,42 @@ export const useGeminiStream = (
           },
         );
 
-      // Check if all tools in the batch are in a terminal state
-      const allTerminal = completedToolCallsFromScheduler.every(
-        (tc) =>
-          tc.status === 'success' ||
-          tc.status === 'error' ||
-          tc.status === 'cancelled',
-      );
-
-      if (!allTerminal) {
-        return;
-      }
-
       // Finalize any client-initiated tools as soon as they are done.
       const clientTools = completedAndReadyToSubmitTools.filter(
         (t) => t.request.isClientInitiated,
       );
       if (clientTools.length > 0) {
         markToolsAsSubmitted(clientTools.map((t) => t.request.callId));
+
+        if (geminiClient) {
+          for (const tool of clientTools) {
+            // Only manually record skill activations in the chat history.
+            // Other client-initiated tools (like save_memory) update the system
+            // prompt/context and don't strictly need to be in the history.
+            if (tool.request.name !== ACTIVATE_SKILL_TOOL_NAME) {
+              continue;
+            }
+
+            // Add both the call (model turn) and the result (user turn) to history.
+            // Client-initiated calls are essentially "synthetic" turns that let
+            // subsequent model calls understand what just happened in the UI.
+            await geminiClient.addHistory({
+              role: 'model',
+              parts: [
+                {
+                  functionCall: {
+                    name: tool.request.name,
+                    args: tool.request.args,
+                  },
+                },
+              ],
+            });
+            await geminiClient.addHistory({
+              role: 'user',
+              parts: tool.response.responseParts,
+            });
+          }
+        }
       }
 
       // Identify new, successful save_memory calls that we haven't processed yet.
@@ -1906,7 +1906,6 @@ export const useGeminiStream = (
       setIsResponding,
     ],
   );
-  handleCompletedToolsRef.current = handleCompletedTools;
 
   const pendingHistoryItems = useMemo(
     () =>

packages/cli/src/ui/themes/theme.ts

@@ -7,6 +7,7 @@
 import type { CSSProperties } from 'react';
 
 import type { SemanticColors } from './semantic-tokens.js';
+import { isLowColorDepth, isITerm2 } from '../utils/terminalUtils.js';
 
 import type { CustomTheme } from '@google/gemini-cli-core';
 import {
@@ -183,50 +184,122 @@ export interface ColorsTheme {
   GradientColors?: string[];
 }
 
-export const lightTheme: ColorsTheme = {
+export const safeLightTheme: ColorsTheme = {
   type: 'light',
+
+  Foreground: '#1C1C1C',
   Background: '#FFFFFF',
-  Foreground: '#000000',
-  LightBlue: '#005FAF',
-  AccentBlue: '#005FAF',
-  AccentPurple: '#5F00FF',
-  AccentCyan: '#005F87',
-  AccentGreen: '#005F00',
+  Comment: '#808080',
+  Gray: '#5F5F5F',
+  DarkGray: '#4E4E4E',
+
+  LightBlue: '#D7E7FF', // ANSI 189
+
+  AccentRed: '#D70000',
   AccentYellow: '#875F00',
-  AccentRed: '#AF0000',
+  AccentGreen: '#008700',
+  AccentCyan: '#00AFD7',
+  AccentBlue: '#0087FF',
+  AccentPurple: '#5F00FF',
+
   DiffAdded: '#D7FFD7',
   DiffRemoved: '#FFD7D7',
-  Comment: '#008700',
-  Gray: '#5F5F5F',
-  DarkGray: '#5F5F5F',
   InputBackground: '#E4E4E4',
-  MessageBackground: '#FAFAFA',
+  MessageBackground: '#fff3f3',
   FocusBackground: '#D7FFD7',
+
+  GradientColors: ['#5FAFD7', '#5F87D7', '#8787D7', '#AF5FAF', '#D75F87'],
+};
+
+export const richLightTheme: ColorsTheme = {
+  type: 'light',
+
+  Foreground: '#202124', // grey-900
+  Gray: '#3C4043', // grey-800
+  Comment: '#5F6368', // grey-700
+  DarkGray: '#80868B', // grey-600
+  Background: '#F8F9FA', // grey-50
+
+  LightBlue: '#1A73E8', // blue-600
+
+  AccentRed: '#D93025', // red-600
+  AccentYellow: '#B06000', // orange-900
+  AccentGreen: '#188038', // green-700
+  AccentCyan: '#007B83', // cyan-900
+  AccentBlue: '#1967D2', // blue-700
+  AccentPurple: '#9334E6', // purple-600
+
+  DiffRemoved: '#FCE8E6', // red-50
+  DiffAdded: '#E6F4EA', // green-50
+  InputBackground: '#E8EAED', // grey-200
+  MessageBackground: '#E8EAED', // grey-200
+  FocusBackground: '#CEEAD6', // green-100
+
   GradientColors: ['#4796E4', '#847ACE', '#C3677F'],
 };
 
-export const darkTheme: ColorsTheme = {
+export const richDarkTheme: ColorsTheme = {
   type: 'dark',
-  Background: '#000000',
-  Foreground: '#FFFFFF',
-  LightBlue: '#AFD7D7',
-  AccentBlue: '#87AFFF',
-  AccentPurple: '#D7AFFF',
-  AccentCyan: '#87D7D7',
-  AccentGreen: '#D7FFD7',
-  AccentYellow: '#FFFFAF',
-  AccentRed: '#FF87AF',
-  DiffAdded: '#005F00',
-  DiffRemoved: '#5F0000',
-  Comment: '#AFAFAF',
-  Gray: '#AFAFAF',
-  DarkGray: '#878787',
-  InputBackground: '#5F5F5F',
-  MessageBackground: '#5F5F5F',
-  FocusBackground: '#005F00',
+
+  Foreground: '#F8F9FA', // grey-50
+  Gray: '#DADCE0', // grey-300
+  Comment: '#BDC1C6', // grey-400
+  DarkGray: '#9AA0A6', // grey-500
+  Background: '#212121', // grey-900
+
+  LightBlue: '#E8F0FE', // blue-50
+
+  AccentRed: '#F6AEA9', // red-200
+  AccentYellow: '#FDE293', // yellow-200
+  AccentGreen: '#A8DAB5', // green-200
+  AccentCyan: '#87D7D7', // green-200
+  AccentBlue: '#AECBFA', // blue-200
+  AccentPurple: '#D7AEFB', // purple-200
+
+  DiffAdded: interpolateColor('#0D652D', '#000000', 0.5), // green-900 @ 50%
+  DiffRemoved: interpolateColor('#A50E0E', '#000000', 0.5), // red-900 @ 50%
+  InputBackground: interpolateColor('#212121', '#000000', 0.5), // grey-900 at 50%
+  MessageBackground: interpolateColor('#212121', '#000000', 0.5), // grey-900 at 50%
+  FocusBackground: interpolateColor('#0D652D', '#000000', 0.5), // green-900 at 50%
+
   GradientColors: ['#4796E4', '#847ACE', '#C3677F'],
 };
 
+export const safeDarkTheme: ColorsTheme = {
+  type: 'dark',
+
+  Foreground: '#eeeeee', // ANSI 231
+  Gray: '#BCC1C6', // ANSI 256
+  Comment: '#9AA0A6', // ANSI 247
+  DarkGray: '#80868B', // ANSI 245
+  Background: '#212121', // ANSI 235
+
+  LightBlue: '#AFD7D7',
+
+  AccentRed: '#FF87AF', // ANSI 211
+  AccentYellow: '#FFFFAF', // ANSI 229
+  AccentGreen: '#D7FFD7', // ANSI 194
+  AccentCyan: '#87D7D7', // ANSI 116
+  AccentBlue: '#87AFFF', // ANSI 111
+  AccentPurple: '#D7AFFF', // ANSI 183
+
+  DiffAdded: '#005F00', // ANSI 22
+  DiffRemoved: '#5F0000', // ANSI 52
+  InputBackground: '#5F5F5F', // ANSI 59
+  MessageBackground: '#5F5F5F', // ANSI 59
+  FocusBackground: '#005F00', // ANSI 22
+
+  GradientColors: ['#5FAFD7', '#5F87D7', '#8787D7', '#AF5FAF', '#D75F87'],
+};
+
+export const isActuallyTrueColor = !isLowColorDepth() || isITerm2();
+export const darkTheme: ColorsTheme = isActuallyTrueColor
+  ? richDarkTheme
+  : safeDarkTheme;
+export const lightTheme: ColorsTheme = isActuallyTrueColor
+  ? richLightTheme
+  : safeLightTheme;
+
 export const ansiTheme: ColorsTheme = {
   type: 'ansi',
   Background: 'black',

packages/cli/src/ui/types.ts

@@ -138,11 +138,9 @@ export interface IndividualToolCallDisplay {
 
 export interface CompressionProps {
   isPending: boolean;
-  beforePercentage: number | null;
-  afterPercentage: number | null;
+  originalTokenCount: number | null;
+  newTokenCount: number | null;
   compressionStatus: CompressionStatus | null;
-  isManual: boolean;
-  thresholdPercentage?: number | null;
 }
 
 /**

packages/cli/test-setup.ts

@@ -66,7 +66,10 @@ beforeEach(() => {
           ? stackLines.slice(lastReactFrameIndex + 1).join('\n')
           : stackLines.slice(1).join('\n');
 
-      if (relevantStack.includes('OverflowContext.tsx')) {
+      if (
+        relevantStack.includes('OverflowContext.tsx') ||
+        relevantStack.includes('useTimedMessage.ts')
+      ) {
         return;
       }
 

packages/core/scripts/compile-windows-sandbox.js

@@ -26,15 +26,15 @@ function compileWindowsSandbox() {
 
   const srcHelperPath = path.resolve(
     __dirname,
-    '../src/services/scripts/GeminiSandbox.exe',
+    '../src/sandbox/windows/GeminiSandbox.exe',
   );
   const distHelperPath = path.resolve(
     __dirname,
-    '../dist/src/services/scripts/GeminiSandbox.exe',
+    '../dist/src/sandbox/windows/GeminiSandbox.exe',
   );
   const sourcePath = path.resolve(
     __dirname,
-    '../src/services/scripts/GeminiSandbox.cs',
+    '../src/sandbox/windows/GeminiSandbox.cs',
   );
 
   if (!fs.existsSync(sourcePath)) {

packages/core/src/agents/a2aUtils.test.ts

@@ -403,7 +403,7 @@ describe('a2aUtils', () => {
 
       const output = reassembler.toString();
       expect(output).toBe(
-        'Analyzing...\n\nProcessing...\n\nArtifact (Code):\nprint("Done")',
+        'Analyzing...Processing...\n\nArtifact (Code):\nprint("Done")',
       );
     });
 

packages/core/src/agents/a2aUtils.ts

@@ -16,6 +16,7 @@ import type {
   AgentInterface,
 } from '@a2a-js/sdk';
 import type { SendMessageResult } from './a2a-client-manager.js';
+import type { SubagentActivityItem } from './types.js';
 
 export const AUTH_REQUIRED_MSG = `[Authorization Required] The agent has indicated it requires authorization to proceed. Please follow the agent's instructions.`;
 
@@ -123,17 +124,39 @@ export class A2AResultReassembler {
 
   private pushMessage(message: Message | undefined) {
     if (!message) return;
-    const text = extractPartsText(message.parts, '\n');
+    const text = extractPartsText(message.parts, '');
     if (text && this.messageLog[this.messageLog.length - 1] !== text) {
       this.messageLog.push(text);
     }
   }
 
+  /**
+   * Returns an array of activity items representing the current reassembled state.
+   */
+  toActivityItems(): SubagentActivityItem[] {
+    const isAuthRequired = this.messageLog.includes(AUTH_REQUIRED_MSG);
+    return [
+      isAuthRequired
+        ? {
+            id: 'auth-required',
+            type: 'thought',
+            content: AUTH_REQUIRED_MSG,
+            status: 'running',
+          }
+        : {
+            id: 'pending',
+            type: 'thought',
+            content: 'Working...',
+            status: 'running',
+          },
+    ];
+  }
+
   /**
    * Returns a human-readable string representation of the current reassembled state.
    */
   toString(): string {
-    const joinedMessages = this.messageLog.join('\n\n');
+    const joinedMessages = this.messageLog.join('');
 
     const artifactsOutput = Array.from(this.artifacts.keys())
       .map((id) => {

packages/core/src/agents/agentLoader.test.ts

@@ -242,6 +242,99 @@ Body`);
         /Name must be a valid slug/,
       );
     });
+
+    describe('error formatting and kind inference', () => {
+      it('should only show local agent errors when kind is inferred as local (via kind field)', async () => {
+        const filePath = await writeAgentMarkdown(`---
+kind: local
+name: invalid-local
+# missing description
+---
+Body`);
+        const error = await parseAgentMarkdown(filePath).catch((e) => e);
+        expect(error).toBeInstanceOf(AgentLoadError);
+        expect(error.message).toContain('Validation failed');
+        expect(error.message).toContain('description: Required');
+        expect(error.message).not.toContain('Remote Agent');
+      });
+
+      it('should only show local agent errors when kind is inferred as local (via local-specific keys)', async () => {
+        const filePath = await writeAgentMarkdown(`---
+name: invalid-local
+# missing description
+tools:
+  - run_shell_command
+---
+Body`);
+        const error = await parseAgentMarkdown(filePath).catch((e) => e);
+        expect(error).toBeInstanceOf(AgentLoadError);
+        expect(error.message).toContain('Validation failed');
+        expect(error.message).toContain('description: Required');
+        expect(error.message).not.toContain('Remote Agent');
+      });
+
+      it('should only show remote agent errors when kind is inferred as remote (via kind field)', async () => {
+        const filePath = await writeAgentMarkdown(`---
+kind: remote
+name: invalid-remote
+# missing agent_card_url
+---
+Body`);
+        const error = await parseAgentMarkdown(filePath).catch((e) => e);
+        expect(error).toBeInstanceOf(AgentLoadError);
+        expect(error.message).toContain('Validation failed');
+        expect(error.message).toContain('agent_card_url: Required');
+        expect(error.message).not.toContain('Local Agent');
+      });
+
+      it('should only show remote agent errors when kind is inferred as remote (via remote-specific keys)', async () => {
+        const filePath = await writeAgentMarkdown(`---
+name: invalid-remote
+auth:
+  type: apiKey
+  key: my_key
+# missing agent_card_url
+---
+Body`);
+        const error = await parseAgentMarkdown(filePath).catch((e) => e);
+        expect(error).toBeInstanceOf(AgentLoadError);
+        expect(error.message).toContain('Validation failed');
+        expect(error.message).toContain('agent_card_url: Required');
+        expect(error.message).not.toContain('Local Agent');
+      });
+
+      it('should show errors for both types when kind cannot be inferred', async () => {
+        const filePath = await writeAgentMarkdown(`---
+name: invalid-unknown
+# missing description and missing agent_card_url, no specific keys
+---
+Body`);
+        const error = await parseAgentMarkdown(filePath).catch((e) => e);
+        expect(error).toBeInstanceOf(AgentLoadError);
+        expect(error.message).toContain('Validation failed');
+        expect(error.message).toContain('(Local Agent)');
+        expect(error.message).toContain('(Remote Agent)');
+        expect(error.message).toContain('description: Required');
+        expect(error.message).toContain('agent_card_url: Required');
+      });
+
+      it('should format errors without a stray colon when the path is empty (e.g. strict object with unknown keys)', async () => {
+        const filePath = await writeAgentMarkdown(`---
+kind: local
+name: my-agent
+description: test
+unknown_field: true
+---
+Body`);
+        const error = await parseAgentMarkdown(filePath).catch((e) => e);
+        expect(error).toBeInstanceOf(AgentLoadError);
+        expect(error.message).toContain(
+          "Unrecognized key(s) in object: 'unknown_field'",
+        );
+        expect(error.message).not.toContain(': Unrecognized key(s)');
+        expect(error.message).not.toContain('Required');
+      });
+    });
   });
 
   describe('markdownToAgentDefinition', () => {
@@ -617,7 +710,7 @@ kind: remote
 name: oauth2-agent
 agent_card_url: https://example.com/card
 auth:
-  type: oauth2
+  type: oauth
   client_id: $MY_OAUTH_CLIENT_ID
   scopes:
     - read
@@ -630,7 +723,7 @@ auth:
         kind: 'remote',
         name: 'oauth2-agent',
         auth: {
-          type: 'oauth2',
+          type: 'oauth',
           client_id: '$MY_OAUTH_CLIENT_ID',
           scopes: ['read', 'write'],
         },
@@ -643,7 +736,7 @@ kind: remote
 name: oauth2-full-agent
 agent_card_url: https://example.com/card
 auth:
-  type: oauth2
+  type: oauth
   client_id: my-client-id
   client_secret: my-client-secret
   scopes:
@@ -659,7 +752,7 @@ auth:
         kind: 'remote',
         name: 'oauth2-full-agent',
         auth: {
-          type: 'oauth2',
+          type: 'oauth',
           client_id: 'my-client-id',
           client_secret: 'my-client-secret',
           scopes: ['openid', 'profile'],
@@ -675,7 +768,7 @@ kind: remote
 name: oauth2-minimal-agent
 agent_card_url: https://example.com/card
 auth:
-  type: oauth2
+  type: oauth
 ---
 `);
       const result = await parseAgentMarkdown(filePath);
@@ -684,7 +777,7 @@ auth:
         kind: 'remote',
         name: 'oauth2-minimal-agent',
         auth: {
-          type: 'oauth2',
+          type: 'oauth',
         },
       });
     });
@@ -695,7 +788,7 @@ kind: remote
 name: invalid-oauth2-agent
 agent_card_url: https://example.com/card
 auth:
-  type: oauth2
+  type: oauth
   client_id: my-client
   authorization_url: not-a-valid-url
 ---
@@ -709,7 +802,7 @@ kind: remote
 name: invalid-oauth2-agent
 agent_card_url: https://example.com/card
 auth:
-  type: oauth2
+  type: oauth
   client_id: my-client
   token_url: not-a-valid-url
 ---
@@ -723,7 +816,7 @@ auth:
         name: 'oauth2-convert-agent',
         agent_card_url: 'https://example.com/card',
         auth: {
-          type: 'oauth2' as const,
+          type: 'oauth' as const,
           client_id: '$MY_CLIENT_ID',
           scopes: ['read'],
           authorization_url: 'https://auth.example.com/authorize',
@@ -744,5 +837,24 @@ auth:
         },
       });
     });
+
+    it('should throw an error for an unknown auth type in markdownToAgentDefinition', () => {
+      const markdown = {
+        kind: 'remote' as const,
+        name: 'unknown-auth-agent',
+        agent_card_url: 'https://example.com/card',
+        auth: {
+          type: 'apiKey' as const,
+          key: 'some-key',
+        },
+      };
+
+      // Mutate the object at runtime to bypass TypeScript compile-time checks cleanly
+      Object.assign(markdown.auth, { type: 'some-unknown-type' });
+
+      expect(() => markdownToAgentDefinition(markdown)).toThrow(
+        /Unknown auth type: some-unknown-type/,
+      );
+    });
   });
 });

packages/core/src/agents/agentLoader.ts

@@ -21,79 +21,6 @@ import { isValidToolName } from '../tools/tool-names.js';
 import { FRONTMATTER_REGEX } from '../skills/skillLoader.js';
 import { getErrorMessage } from '../utils/errors.js';
 
-/**
- * DTO for Markdown parsing - represents the structure from frontmatter.
- */
-interface FrontmatterBaseAgentDefinition {
-  name: string;
-  display_name?: string;
-}
-
-interface FrontmatterMCPServerConfig {
-  command?: string;
-  args?: string[];
-  env?: Record<string, string>;
-  cwd?: string;
-  url?: string;
-  http_url?: string;
-  headers?: Record<string, string>;
-  tcp?: string;
-  type?: 'sse' | 'http';
-  timeout?: number;
-  trust?: boolean;
-  description?: string;
-  include_tools?: string[];
-  exclude_tools?: string[];
-}
-
-interface FrontmatterLocalAgentDefinition
-  extends FrontmatterBaseAgentDefinition {
-  kind: 'local';
-  description: string;
-  tools?: string[];
-  mcp_servers?: Record<string, FrontmatterMCPServerConfig>;
-  system_prompt: string;
-  model?: string;
-  temperature?: number;
-  max_turns?: number;
-  timeout_mins?: number;
-}
-
-/**
- * Authentication configuration for remote agents in frontmatter format.
- */
-interface FrontmatterAuthConfig {
-  type: 'apiKey' | 'http' | 'google-credentials' | 'oauth2';
-  // API Key
-  key?: string;
-  name?: string;
-  // HTTP
-  scheme?: string;
-  token?: string;
-  username?: string;
-  password?: string;
-  value?: string;
-  // Google Credentials
-  scopes?: string[];
-  // OAuth2
-  client_id?: string;
-  client_secret?: string;
-  authorization_url?: string;
-  token_url?: string;
-}
-
-interface FrontmatterRemoteAgentDefinition
-  extends FrontmatterBaseAgentDefinition {
-  kind: 'remote';
-  description?: string;
-  agent_card_url: string;
-  auth?: FrontmatterAuthConfig;
-}
-
-type FrontmatterAgentDefinition =
-  | FrontmatterLocalAgentDefinition
-  | FrontmatterRemoteAgentDefinition;
-
 /**
  * Error thrown when an agent definition is invalid or cannot be loaded.
  */
@@ -159,15 +86,13 @@ const localAgentSchema = z
   })
   .strict();
 
-/**
- * Base fields shared by all auth configs.
- */
+type FrontmatterLocalAgentDefinition = z.infer<typeof localAgentSchema> & {
+  system_prompt: string;
+};
+
+// Base fields shared by all auth configs.
 const baseAuthFields = {};
 
-/**
- * API Key auth schema.
- * Supports sending key in header, query parameter, or cookie.
- */
 const apiKeyAuthSchema = z.object({
   ...baseAuthFields,
   type: z.literal('apiKey'),
@@ -175,11 +100,6 @@ const apiKeyAuthSchema = z.object({
   name: z.string().optional(),
 });
 
-/**
- * HTTP auth schema (Bearer or Basic).
- * Note: Validation for scheme-specific fields is applied in authConfigSchema
- * since discriminatedUnion doesn't support refined schemas directly.
- */
 const httpAuthSchema = z.object({
   ...baseAuthFields,
   type: z.literal('http'),
@@ -190,22 +110,15 @@ const httpAuthSchema = z.object({
   value: z.string().min(1).optional(),
 });
 
-/**
- * Google Credentials auth schema.
- */
 const googleCredentialsAuthSchema = z.object({
   ...baseAuthFields,
   type: z.literal('google-credentials'),
   scopes: z.array(z.string()).optional(),
 });
 
-/**
- * OAuth2 auth schema.
- * authorization_url and token_url can be discovered from the agent card if omitted.
- */
 const oauth2AuthSchema = z.object({
   ...baseAuthFields,
-  type: z.literal('oauth2'),
+  type: z.literal('oauth'),
   client_id: z.string().optional(),
   client_secret: z.string().optional(),
   scopes: z.array(z.string()).optional(),
@@ -222,18 +135,16 @@ const authConfigSchema = z
   ])
   .superRefine((data, ctx) => {
     if (data.type === 'http') {
-      if (data.value) {
-        // Raw mode - only scheme and value are needed
-        return;
-      }
-      if (data.scheme === 'Bearer' && !data.token) {
-        ctx.addIssue({
-          code: z.ZodIssueCode.custom,
-          message: 'Bearer scheme requires "token"',
-          path: ['token'],
-        });
-      }
-      if (data.scheme === 'Basic') {
+      if (data.value) return;
+      if (data.scheme === 'Bearer') {
+        if (!data.token) {
+          ctx.addIssue({
+            code: z.ZodIssueCode.custom,
+            message: 'Bearer scheme requires "token"',
+            path: ['token'],
+          });
+        }
+      } else if (data.scheme === 'Basic') {
         if (!data.username) {
           ctx.addIssue({
             code: z.ZodIssueCode.custom,
@@ -248,10 +159,18 @@ const authConfigSchema = z
             path: ['password'],
           });
         }
+      } else {
+        ctx.addIssue({
+          code: z.ZodIssueCode.custom,
+          message: `HTTP scheme "${data.scheme}" requires "value"`,
+          path: ['value'],
+        });
       }
     }
   });
 
+type FrontmatterAuthConfig = z.infer<typeof authConfigSchema>;
+
 const remoteAgentSchema = z
   .object({
     kind: z.literal('remote').optional().default('remote'),
@@ -263,8 +182,12 @@ const remoteAgentSchema = z
   })
   .strict();
 
-// Use a Zod union to automatically discriminate between local and remote
-// agent types.
+type FrontmatterRemoteAgentDefinition = z.infer<typeof remoteAgentSchema>;
+
+type FrontmatterAgentDefinition =
+  | FrontmatterLocalAgentDefinition
+  | FrontmatterRemoteAgentDefinition;
+
 const agentUnionOptions = [
   { schema: localAgentSchema, label: 'Local Agent' },
   { schema: remoteAgentSchema, label: 'Remote Agent' },
@@ -277,23 +200,62 @@ const markdownFrontmatterSchema = z.union([
   agentUnionOptions[1].schema,
 ]);
 
-function formatZodError(error: z.ZodError, context: string): string {
+function guessIntendedKind(rawInput: unknown): 'local' | 'remote' | undefined {
+  if (typeof rawInput !== 'object' || rawInput === null) return undefined;
+  const input = rawInput as Partial<FrontmatterLocalAgentDefinition> &
+    Partial<FrontmatterRemoteAgentDefinition>;
+
+  if (input.kind === 'local') return 'local';
+  if (input.kind === 'remote') return 'remote';
+
+  const hasLocalKeys =
+    'tools' in input ||
+    'mcp_servers' in input ||
+    'model' in input ||
+    'temperature' in input ||
+    'max_turns' in input ||
+    'timeout_mins' in input;
+  const hasRemoteKeys = 'agent_card_url' in input || 'auth' in input;
+
+  if (hasLocalKeys && !hasRemoteKeys) return 'local';
+  if (hasRemoteKeys && !hasLocalKeys) return 'remote';
+
+  return undefined;
+}
+
+function formatZodError(
+  error: z.ZodError,
+  context: string,
+  rawInput?: unknown,
+): string {
+  const intendedKind = rawInput ? guessIntendedKind(rawInput) : undefined;
+
   const issues = error.issues
     .map((i) => {
-      // Handle union errors specifically to give better context
       if (i.code === z.ZodIssueCode.invalid_union) {
         return i.unionErrors
           .map((unionError, index) => {
             const label =
               agentUnionOptions[index]?.label ?? `Agent type #${index + 1}`;
+
+            if (intendedKind === 'local' && label === 'Remote Agent')
+              return null;
+            if (intendedKind === 'remote' && label === 'Local Agent')
+              return null;
+
             const unionIssues = unionError.issues
-              .map((u) => `${u.path.join('.')}: ${u.message}`)
+              .map((u) => {
+                const pathStr = u.path.join('.');
+                return pathStr ? `${pathStr}: ${u.message}` : u.message;
+              })
               .join(', ');
             return `(${label}) ${unionIssues}`;
           })
+          .filter(Boolean)
           .join('\n');
       }
-      return `${i.path.join('.')}: ${i.message}`;
+      const pathStr = i.path.join('.');
+      return pathStr ? `${pathStr}: ${i.message}` : i.message;
     })
     .join('\n');
   return `${context}:\n${issues}`;
@@ -343,8 +305,7 @@ export async function parseAgentMarkdown(
   } catch (error) {
     throw new AgentLoadError(
       filePath,
-      // eslint-disable-next-line @typescript-eslint/no-unsafe-type-assertion
-      `YAML frontmatter parsing failed: ${(error as Error).message}`,
+      `YAML frontmatter parsing failed: ${getErrorMessage(error)}`,
     );
   }
 
@@ -368,7 +329,7 @@ export async function parseAgentMarkdown(
   if (!result.success) {
     throw new AgentLoadError(
       filePath,
-      `Validation failed: ${formatZodError(result.error, 'Agent Definition')}`,
+      `Validation failed: ${formatZodError(result.error, 'Agent Definition', rawFrontmatter)}`,
     );
   }
 
@@ -383,17 +344,14 @@ export async function parseAgentMarkdown(
     ];
   }
 
-  // Local agent validation
-  // Validate tools
-
   // Construct the local agent definition
-  const agentDef: FrontmatterLocalAgentDefinition = {
-    ...frontmatter,
-    kind: 'local',
-    system_prompt: body.trim(),
-  };
-
-  return [agentDef];
+  return [
+    {
+      ...frontmatter,
+      kind: 'local',
+      system_prompt: body.trim(),
+    },
+  ];
 }
 
 /**
@@ -403,15 +361,9 @@ export async function parseAgentMarkdown(
 function convertFrontmatterAuthToConfig(
   frontmatter: FrontmatterAuthConfig,
 ): A2AAuthConfig {
-  const base = {};
-
   switch (frontmatter.type) {
     case 'apiKey':
-      if (!frontmatter.key) {
-        throw new Error('Internal error: API key missing after validation.');
-      }
       return {
-        ...base,
         type: 'apiKey',
         key: frontmatter.key,
         name: frontmatter.name,
@@ -419,20 +371,13 @@ function convertFrontmatterAuthToConfig(
 
     case 'google-credentials':
       return {
-        ...base,
         type: 'google-credentials',
         scopes: frontmatter.scopes,
       };
 
-    case 'http': {
-      if (!frontmatter.scheme) {
-        throw new Error(
-          'Internal error: HTTP scheme missing after validation.',
-        );
-      }
+    case 'http':
       if (frontmatter.value) {
         return {
-          ...base,
           type: 'http',
           scheme: frontmatter.scheme,
           value: frontmatter.value,
@@ -440,40 +385,29 @@ function convertFrontmatterAuthToConfig(
       }
       switch (frontmatter.scheme) {
         case 'Bearer':
-          if (!frontmatter.token) {
-            throw new Error(
-              'Internal error: Bearer token missing after validation.',
-            );
-          }
+          // Token is required by schema validation
           return {
-            ...base,
             type: 'http',
             scheme: 'Bearer',
-            token: frontmatter.token,
+
+            token: frontmatter.token!,
           };
         case 'Basic':
-          if (!frontmatter.username || !frontmatter.password) {
-            throw new Error(
-              'Internal error: Basic auth credentials missing after validation.',
-            );
-          }
+          // Username/password are required by schema validation
           return {
-            ...base,
             type: 'http',
             scheme: 'Basic',
-            username: frontmatter.username,
-            password: frontmatter.password,
-          };
-        default: {
-          // Other IANA schemes without a value should not reach here after validation
-          throw new Error(`Unknown HTTP scheme: ${frontmatter.scheme}`);
-        }
-      }
-    }
 
-    case 'oauth2':
+            username: frontmatter.username!,
+
+            password: frontmatter.password!,
+          };
+        default:
+          throw new Error(`Unknown HTTP scheme: ${frontmatter.scheme}`);
+      }
+
+    case 'oauth':
       return {
-        ...base,
         type: 'oauth2',
         client_id: frontmatter.client_id,
         client_secret: frontmatter.client_secret,
@@ -483,8 +417,12 @@ function convertFrontmatterAuthToConfig(
       };
 
     default: {
-      const exhaustive: never = frontmatter.type;
-      throw new Error(`Unknown auth type: ${exhaustive}`);
+      const exhaustive: never = frontmatter;
+      const raw: unknown = exhaustive;
+      if (typeof raw === 'object' && raw !== null && 'type' in raw) {
+        throw new Error(`Unknown auth type: ${String(raw['type'])}`);
+      }
+      throw new Error('Unknown auth type');
     }
   }
 }
@@ -533,7 +471,7 @@ export function markdownToAgentDefinition(
   const modelName = markdown.model || 'inherit';
 
   const mcpServers: Record<string, MCPServerConfig> = {};
-  if (markdown.kind === 'local' && markdown.mcp_servers) {
+  if (markdown.mcp_servers) {
     for (const [name, config] of Object.entries(markdown.mcp_servers)) {
       mcpServers[name] = new MCPServerConfig(
         config.command,
@@ -606,15 +544,13 @@ export async function loadAgentsFromDirectory(
     dirEntries = await fs.readdir(dir, { withFileTypes: true });
   } catch (error) {
     // If directory doesn't exist, just return empty
-    // eslint-disable-next-line @typescript-eslint/no-unsafe-type-assertion
-    if ((error as NodeJS.ErrnoException).code === 'ENOENT') {
+    if (error instanceof Error && 'code' in error && error.code === 'ENOENT') {
       return result;
     }
     result.errors.push(
       new AgentLoadError(
         dir,
-        // eslint-disable-next-line @typescript-eslint/no-unsafe-type-assertion
-        `Could not list directory: ${(error as Error).message}`,
+        `Could not list directory: ${getErrorMessage(error)}`,
       ),
     );
     return result;
@@ -644,8 +580,7 @@ export async function loadAgentsFromDirectory(
         result.errors.push(
           new AgentLoadError(
             filePath,
-            // eslint-disable-next-line @typescript-eslint/no-unsafe-type-assertion
-            `Unexpected error: ${(error as Error).message}`,
+            `Unexpected error: ${getErrorMessage(error)}`,
           ),
         );
       }

packages/core/src/agents/browser/browserAgentDefinition.ts

@@ -112,6 +112,7 @@ Some errors are unrecoverable and retrying will never help. When you see ANY of
 - "Could not connect to Chrome" or "Failed to connect to Chrome" or "Timed out connecting to Chrome" — Include the full error message with its remediation steps in your summary verbatim. Do NOT paraphrase or omit instructions.
 - "Browser closed" or "Target closed" or "Session closed" — The browser process has terminated. Include the error and tell the user to try again.
 - "net::ERR_" network errors on the SAME URL after 2 retries — the site is unreachable. Report the URL and error.
+- "reached maximum action limit" — You have performed too many actions in this task. Stop immediately and report this limit to the user.
 - Any error that appears IDENTICALLY 3+ times in a row — it will not resolve by retrying.
 Do NOT keep retrying terminal errors. Report them with actionable remediation steps and exit immediately.
 

packages/core/src/agents/browser/browserManager.test.ts

@@ -697,4 +697,28 @@ describe('BrowserManager', () => {
       expect(injectAutomationOverlay).not.toHaveBeenCalled();
     });
   });
+
+  describe('Rate limiting', () => {
+    it('should terminate task when maxActionsPerTask is reached', async () => {
+      const limitedConfig = makeFakeConfig({
+        agents: {
+          browser: {
+            maxActionsPerTask: 3,
+          },
+        },
+      });
+      const manager = new BrowserManager(limitedConfig);
+
+      // First 3 calls should succeed
+      await manager.callTool('take_snapshot', {});
+      await manager.callTool('take_snapshot', { some: 'args' });
+      await manager.callTool('take_snapshot', { other: 'args' });
+      await manager.callTool('take_snapshot', { other: 'new args' });
+
+      // 4th call should throw
+      await expect(manager.callTool('take_snapshot', {})).rejects.toThrow(
+        /maximum action limit \(3\)/,
+      );
+    });
+  });
 });

packages/core/src/agents/browser/browserManager.ts

@@ -97,6 +97,10 @@ export class BrowserManager {
   private mcpTransport: StdioClientTransport | undefined;
   private discoveredTools: McpTool[] = [];
 
+  /** State for action rate limiting */
+  private actionCounter = 0;
+  private readonly maxActionsPerTask: number;
+
   /**
    * Whether to inject the automation overlay.
    * Always false in headless mode (no visible window to decorate).
@@ -108,6 +112,8 @@ export class BrowserManager {
     const browserConfig = config.getBrowserAgentConfig();
     this.shouldInjectOverlay = !browserConfig?.customConfig?.headless;
     this.shouldDisableInput = config.shouldDisableBrowserUserInput();
+    this.maxActionsPerTask =
+      browserConfig?.customConfig.maxActionsPerTask ?? 100;
   }
 
   /**
@@ -151,6 +157,16 @@ export class BrowserManager {
       throw signal.reason ?? new Error('Operation cancelled');
     }
 
+    // Hard enforcement of per-action rate limit
+    if (this.actionCounter > this.maxActionsPerTask) {
+      const error = new Error(
+        `Browser agent reached maximum action limit (${this.maxActionsPerTask}). ` +
+          `Task terminated to prevent runaway execution. To config the limit, use maxActionsPerTask in the settings.`,
+      );
+      throw error;
+    }
+    this.actionCounter++;
+
     const errorMessage = this.checkNavigationRestrictions(toolName, args);
     if (errorMessage) {
       return {

packages/core/src/agents/remote-invocation.test.ts

@@ -20,7 +20,7 @@ import {
   type A2AClientManager,
 } from './a2a-client-manager.js';
 
-import type { RemoteAgentDefinition } from './types.js';
+import type { RemoteAgentDefinition, SubagentProgress } from './types.js';
 import { createMockMessageBus } from '../test-utils/mock-message-bus.js';
 import { A2AAuthProviderFactory } from './auth-provider/factory.js';
 import type { A2AAuthProvider } from './auth-provider/types.js';
@@ -266,9 +266,11 @@ describe('RemoteAgentInvocation', () => {
       );
       const result = await invocation.execute(new AbortController().signal);
 
-      expect(result.error?.message).toContain(
-        "Failed to create auth provider for agent 'test-agent'",
-      );
+      expect(result.returnDisplay).toMatchObject({
+        result: expect.stringContaining(
+          "Failed to create auth provider for agent 'test-agent'",
+        ),
+      });
     });
 
     it('should not load the agent if already present', async () => {
@@ -325,7 +327,9 @@ describe('RemoteAgentInvocation', () => {
 
       // Execute first time
       const result1 = await invocation1.execute(new AbortController().signal);
-      expect(result1.returnDisplay).toBe('Response 1');
+      expect(result1.returnDisplay).toMatchObject({
+        result: 'Response 1',
+      });
       expect(mockClientManager.sendMessageStream).toHaveBeenLastCalledWith(
         'test-agent',
         'first',
@@ -355,7 +359,9 @@ describe('RemoteAgentInvocation', () => {
         mockMessageBus,
       );
       const result2 = await invocation2.execute(new AbortController().signal);
-      expect(result2.returnDisplay).toBe('Response 2');
+      expect((result2.returnDisplay as SubagentProgress).result).toBe(
+        'Response 2',
+      );
 
       expect(mockClientManager.sendMessageStream).toHaveBeenLastCalledWith(
         'test-agent',
@@ -444,8 +450,22 @@ describe('RemoteAgentInvocation', () => {
       );
       await invocation.execute(new AbortController().signal, updateOutput);
 
-      expect(updateOutput).toHaveBeenCalledWith('Hello');
-      expect(updateOutput).toHaveBeenCalledWith('Hello\n\nHello World');
+      expect(updateOutput).toHaveBeenCalledWith(
+        expect.objectContaining({
+          isSubagentProgress: true,
+          state: 'running',
+          recentActivity: expect.arrayContaining([
+            expect.objectContaining({ content: 'Working...' }),
+          ]),
+        }),
+      );
+      expect(updateOutput).toHaveBeenCalledWith(
+        expect.objectContaining({
+          isSubagentProgress: true,
+          state: 'completed',
+          result: 'HelloHello World',
+        }),
+      );
     });
 
     it('should abort when signal is aborted during streaming', async () => {
@@ -478,8 +498,7 @@ describe('RemoteAgentInvocation', () => {
       );
       const result = await invocation.execute(controller.signal);
 
-      expect(result.error).toBeDefined();
-      expect(result.error?.message).toContain('Operation aborted');
+      expect(result.returnDisplay).toMatchObject({ state: 'error' });
     });
 
     it('should handle errors gracefully', async () => {
@@ -501,9 +520,10 @@ describe('RemoteAgentInvocation', () => {
       );
       const result = await invocation.execute(new AbortController().signal);
 
-      expect(result.error).toBeDefined();
-      expect(result.error?.message).toContain('Network error');
-      expect(result.returnDisplay).toContain('Network error');
+      expect(result.returnDisplay).toMatchObject({
+        state: 'error',
+        result: expect.stringContaining('Network error'),
+      });
     });
 
     it('should use a2a helpers for extracting text', async () => {
@@ -534,7 +554,9 @@ describe('RemoteAgentInvocation', () => {
       const result = await invocation.execute(new AbortController().signal);
 
       // Just check that text is present, exact formatting depends on helper
-      expect(result.returnDisplay).toContain('Extracted text');
+      expect((result.returnDisplay as SubagentProgress).result).toContain(
+        'Extracted text',
+      );
     });
 
     it('should handle mixed response types during streaming (TaskStatusUpdateEvent + Message)', async () => {
@@ -577,9 +599,25 @@ describe('RemoteAgentInvocation', () => {
         updateOutput,
       );
 
-      expect(updateOutput).toHaveBeenCalledWith('Thinking...');
-      expect(updateOutput).toHaveBeenCalledWith('Thinking...\n\nFinal Answer');
-      expect(result.returnDisplay).toBe('Thinking...\n\nFinal Answer');
+      expect(updateOutput).toHaveBeenCalledWith(
+        expect.objectContaining({
+          isSubagentProgress: true,
+          state: 'running',
+          recentActivity: expect.arrayContaining([
+            expect.objectContaining({ content: 'Working...' }),
+          ]),
+        }),
+      );
+      expect(updateOutput).toHaveBeenCalledWith(
+        expect.objectContaining({
+          isSubagentProgress: true,
+          state: 'completed',
+          result: 'Thinking...Final Answer',
+        }),
+      );
+      expect(result.returnDisplay).toMatchObject({
+        result: 'Thinking...Final Answer',
+      });
     });
 
     it('should handle artifact reassembly with append: true', async () => {
@@ -635,12 +673,21 @@ describe('RemoteAgentInvocation', () => {
       );
       await invocation.execute(new AbortController().signal, updateOutput);
 
-      expect(updateOutput).toHaveBeenCalledWith('Generating...');
       expect(updateOutput).toHaveBeenCalledWith(
-        'Generating...\n\nArtifact (Result):\nPart 1',
+        expect.objectContaining({
+          isSubagentProgress: true,
+          state: 'running',
+          recentActivity: expect.arrayContaining([
+            expect.objectContaining({ content: 'Working...' }),
+          ]),
+        }),
       );
       expect(updateOutput).toHaveBeenCalledWith(
-        'Generating...\n\nArtifact (Result):\nPart 1 Part 2',
+        expect.objectContaining({
+          isSubagentProgress: true,
+          state: 'completed',
+          result: 'Generating...\n\nArtifact (Result):\nPart 1 Part 2',
+        }),
       );
     });
   });
@@ -694,8 +741,10 @@ describe('RemoteAgentInvocation', () => {
       );
       const result = await invocation.execute(new AbortController().signal);
 
-      expect(result.error).toBeDefined();
-      expect(result.returnDisplay).toContain(a2aError.userMessage);
+      expect(result.returnDisplay).toMatchObject({ state: 'error' });
+      expect((result.returnDisplay as SubagentProgress).result).toContain(
+        a2aError.userMessage,
+      );
     });
 
     it('should use generic message for non-A2AAgentError errors', async () => {
@@ -712,8 +761,8 @@ describe('RemoteAgentInvocation', () => {
       );
       const result = await invocation.execute(new AbortController().signal);
 
-      expect(result.error).toBeDefined();
-      expect(result.returnDisplay).toContain(
+      expect(result.returnDisplay).toMatchObject({ state: 'error' });
+      expect((result.returnDisplay as SubagentProgress).result).toContain(
         'Error calling remote agent: something unexpected',
       );
     });
@@ -741,10 +790,14 @@ describe('RemoteAgentInvocation', () => {
       );
       const result = await invocation.execute(new AbortController().signal);
 
-      expect(result.error).toBeDefined();
+      expect(result.returnDisplay).toMatchObject({ state: 'error' });
       // Should contain both the partial output and the error message
-      expect(result.returnDisplay).toContain('Partial response');
-      expect(result.returnDisplay).toContain('connection reset');
+      expect(result.returnDisplay).toMatchObject({
+        result: expect.stringContaining('Partial response'),
+      });
+      expect(result.returnDisplay).toMatchObject({
+        result: expect.stringContaining('connection reset'),
+      });
     });
   });
 });

packages/core/src/agents/remote-invocation.ts

@@ -15,6 +15,7 @@ import {
   type RemoteAgentInputs,
   type RemoteAgentDefinition,
   type AgentInputs,
+  type SubagentProgress,
 } from './types.js';
 import { type AgentLoopContext } from '../config/agent-loop-context.js';
 import type { MessageBus } from '../confirmation-bus/message-bus.js';
@@ -25,7 +26,6 @@ import type {
 import { extractIdsFromResponse, A2AResultReassembler } from './a2aUtils.js';
 import type { AuthenticationHandler } from '@a2a-js/sdk/client';
 import { debugLogger } from '../utils/debugLogger.js';
-import { safeJsonToMarkdown } from '../utils/markdownUtils.js';
 import type { AnsiOutput } from '../utils/terminalSerializer.js';
 import { A2AAuthProviderFactory } from './auth-provider/factory.js';
 import { A2AAgentError } from './a2a-errors.js';
@@ -125,13 +125,30 @@ export class RemoteAgentInvocation extends BaseToolInvocation<
 
   async execute(
     _signal: AbortSignal,
-    updateOutput?: (output: string | AnsiOutput) => void,
+    updateOutput?: (output: string | AnsiOutput | SubagentProgress) => void,
   ): Promise<ToolResult> {
     // 1. Ensure the agent is loaded (cached by manager)
     // We assume the user has provided an access token via some mechanism (TODO),
     // or we rely on ADC.
     const reassembler = new A2AResultReassembler();
+    const agentName = this.definition.displayName ?? this.definition.name;
     try {
+      if (updateOutput) {
+        updateOutput({
+          isSubagentProgress: true,
+          agentName,
+          state: 'running',
+          recentActivity: [
+            {
+              id: 'pending',
+              type: 'thought',
+              content: 'Working...',
+              status: 'running',
+            },
+          ],
+        });
+      }
+
       const priorState = RemoteAgentInvocation.sessionState.get(
         this.definition.name,
       );
@@ -172,7 +189,13 @@ export class RemoteAgentInvocation extends BaseToolInvocation<
         reassembler.update(chunk);
 
         if (updateOutput) {
-          updateOutput(reassembler.toString());
+          updateOutput({
+            isSubagentProgress: true,
+            agentName,
+            state: 'running',
+            recentActivity: reassembler.toActivityItems(),
+            result: reassembler.toString(),
+          });
         }
 
         const {
@@ -198,9 +221,21 @@ export class RemoteAgentInvocation extends BaseToolInvocation<
         `[RemoteAgent] Final response from ${this.definition.name}:\n${JSON.stringify(finalResponse, null, 2)}`,
       );
 
+      const finalProgress: SubagentProgress = {
+        isSubagentProgress: true,
+        agentName,
+        state: 'completed',
+        result: finalOutput,
+        recentActivity: reassembler.toActivityItems(),
+      };
+
+      if (updateOutput) {
+        updateOutput(finalProgress);
+      }
+
       return {
         llmContent: [{ text: finalOutput }],
-        returnDisplay: safeJsonToMarkdown(finalOutput),
+        returnDisplay: finalProgress,
       };
     } catch (error: unknown) {
       const partialOutput = reassembler.toString();
@@ -209,10 +244,22 @@ export class RemoteAgentInvocation extends BaseToolInvocation<
       const fullDisplay = partialOutput
         ? `${partialOutput}\n\n${errorMessage}`
         : errorMessage;
+
+      const errorProgress: SubagentProgress = {
+        isSubagentProgress: true,
+        agentName,
+        state: 'error',
+        result: fullDisplay,
+        recentActivity: reassembler.toActivityItems(),
+      };
+
+      if (updateOutput) {
+        updateOutput(errorProgress);
+      }
+
       return {
         llmContent: [{ text: fullDisplay }],
-        returnDisplay: fullDisplay,
-        error: { message: errorMessage },
+        returnDisplay: errorProgress,
       };
     } finally {
       // Persist state even on partial failures or aborts to maintain conversational continuity.

packages/core/src/config/config.test.ts

@@ -1474,6 +1474,22 @@ describe('Server Config (config.ts)', () => {
       expect(browserConfig.customConfig.visualModel).toBe(
         'custom-visual-model',
       );
+      expect(browserConfig.customConfig.maxActionsPerTask).toBe(100); // default
+    });
+
+    it('should return custom maxActionsPerTask', () => {
+      const params: ConfigParameters = {
+        ...baseParams,
+        agents: {
+          browser: {
+            maxActionsPerTask: 50,
+          },
+        },
+      };
+      const config = new Config(params);
+      const browserConfig = config.getBrowserAgentConfig();
+
+      expect(browserConfig.customConfig.maxActionsPerTask).toBe(50);
     });
 
     it('should apply defaults for partial custom config', () => {

packages/core/src/config/config.ts

@@ -6,6 +6,7 @@
 
 import * as fs from 'node:fs';
 import * as path from 'node:path';
+import { SandboxPolicyManager } from '../policy/sandboxPolicyManager.js';
 import { inspect } from 'node:util';
 import process from 'node:process';
 import { z } from 'zod';
@@ -330,6 +331,8 @@ export interface BrowserAgentCustomConfig {
   allowedDomains?: string[];
   /** Disable user input on the browser window during automation. Default: true in non-headless mode */
   disableUserInput?: boolean;
+  /** Maximum number of actions (tool calls) allowed per task. Default: 100 */
+  maxActionsPerTask?: number;
   /** Whether to confirm sensitive actions (e.g., fill_form, evaluate_script). */
   confirmSensitiveActions?: boolean;
   /** Whether to block file uploads. */
@@ -673,8 +676,6 @@ export interface ConfigParameters {
     agents?: AgentSettings;
   }>;
   enableConseca?: boolean;
-  showContextWindowWarning?: boolean;
-  showContextCompression?: boolean;
   billing?: {
     overageStrategy?: OverageStrategy;
   };
@@ -710,8 +711,6 @@ export class Config implements McpContext, AgentLoopContext {
   private readonly question: string | undefined;
   private readonly worktreeSettings: WorktreeSettings | undefined;
   readonly enableConseca: boolean;
-  private readonly showContextWindowWarning: boolean;
-  private readonly showContextCompression: boolean;
 
   private readonly coreTools: string[] | undefined;
   private readonly mainAgentTools: string[] | undefined;
@@ -734,7 +733,8 @@ export class Config implements McpContext, AgentLoopContext {
   private readonly telemetrySettings: TelemetrySettings;
   private readonly usageStatisticsEnabled: boolean;
   private _geminiClient!: GeminiClient;
-  private readonly _sandboxManager: SandboxManager;
+  private _sandboxManager: SandboxManager;
+  private readonly _sandboxPolicyManager: SandboxPolicyManager;
   private baseLlmClient!: BaseLlmClient;
   private localLiteRtLmClient?: LocalLiteRtLmClient;
   private modelRouterService: ModelRouterService;
@@ -909,14 +909,14 @@ export class Config implements McpContext, AgentLoopContext {
       params.embeddingModel ?? DEFAULT_GEMINI_EMBEDDING_MODEL;
     this.sandbox = params.sandbox
       ? {
-          enabled: params.sandbox.enabled ?? false,
+          enabled: params.sandbox.enabled || params.toolSandboxing || false,
           allowedPaths: params.sandbox.allowedPaths ?? [],
           networkAccess: params.sandbox.networkAccess ?? false,
           command: params.sandbox.command,
           image: params.sandbox.image,
         }
       : {
-          enabled: false,
+          enabled: params.toolSandboxing || false,
           allowedPaths: [],
           networkAccess: false,
         };
@@ -935,6 +935,30 @@ export class Config implements McpContext, AgentLoopContext {
       this.fileSystemService = new StandardFileSystemService();
     }
 
+    this._sandboxPolicyManager = new SandboxPolicyManager();
+    const initialApprovalMode =
+      params.approvalMode ??
+      params.policyEngineConfig?.approvalMode ??
+      'default';
+    this._sandboxManager = createSandboxManager(
+      this.sandbox,
+      params.targetDir,
+      this._sandboxPolicyManager,
+      initialApprovalMode,
+    );
+
+    if (
+      !(this._sandboxManager instanceof NoopSandboxManager) &&
+      this.sandbox?.enabled
+    ) {
+      this.fileSystemService = new SandboxedFileSystemService(
+        this._sandboxManager,
+        params.targetDir,
+      );
+    } else {
+      this.fileSystemService = new StandardFileSystemService();
+    }
+
     this.targetDir = path.resolve(params.targetDir);
     this.folderTrust = params.folderTrust ?? false;
     this.workspaceContext = new WorkspaceContext(this.targetDir, []);
@@ -1005,7 +1029,7 @@ export class Config implements McpContext, AgentLoopContext {
     this.model = params.model;
     this.disableLoopDetection = params.disableLoopDetection ?? false;
     this._activeModel = params.model;
-    this.enableAgents = params.enableAgents ?? false;
+    this.enableAgents = params.enableAgents ?? true;
     this.agents = params.agents ?? {};
     this.disableLLMCorrection = params.disableLLMCorrection ?? true;
     this.planEnabled = params.plan ?? true;
@@ -1150,8 +1174,6 @@ export class Config implements McpContext, AgentLoopContext {
     this.fileExclusions = new FileExclusions(this);
     this.eventEmitter = params.eventEmitter;
     this.enableConseca = params.enableConseca ?? false;
-    this.showContextWindowWarning = params.showContextWindowWarning ?? false;
-    this.showContextCompression = params.showContextCompression ?? false;
 
     // Initialize Safety Infrastructure
     const contextBuilder = new ContextBuilder(this);
@@ -1166,12 +1188,19 @@ export class Config implements McpContext, AgentLoopContext {
       params.policyUpdateConfirmationRequest;
 
     this.disableAlwaysAllow = params.disableAlwaysAllow ?? false;
+    const engineApprovalMode =
+      params.approvalMode ??
+      params.policyEngineConfig?.approvalMode ??
+      ApprovalMode.DEFAULT;
     this.policyEngine = new PolicyEngine(
       {
         ...params.policyEngineConfig,
-        approvalMode:
-          params.approvalMode ?? params.policyEngineConfig?.approvalMode,
+        approvalMode: engineApprovalMode,
         disableAlwaysAllow: this.disableAlwaysAllow,
+        toolSandboxEnabled: this.getSandboxEnabled(),
+        sandboxApprovedTools:
+          this.sandboxPolicyManager?.getModeConfig(engineApprovalMode)
+            ?.approvedTools ?? [],
       },
       checkerRunner,
     );
@@ -1566,6 +1595,20 @@ export class Config implements McpContext, AgentLoopContext {
     return this._geminiClient;
   }
 
+  private refreshSandboxManager(): void {
+    this._sandboxManager = createSandboxManager(
+      this.sandbox,
+      this.targetDir,
+      this._sandboxPolicyManager,
+      this.getApprovalMode(),
+    );
+    this.shellExecutionConfig.sandboxManager = this._sandboxManager;
+  }
+
+  get sandboxPolicyManager() {
+    return this._sandboxPolicyManager;
+  }
+
   get sandboxManager(): SandboxManager {
     return this._sandboxManager;
   }
@@ -2059,18 +2102,6 @@ export class Config implements McpContext, AgentLoopContext {
     return this.mcpEnabled;
   }
 
-  getShowContextWindowWarning(): boolean {
-    return this.showContextWindowWarning;
-  }
-
-  getShowContextCompression(): boolean {
-    return this.showContextCompression;
-  }
-
-  getContextWindowCompressionThreshold(): number {
-    return this.compressionThreshold ?? 0.5;
-  }
-
   getMcpEnablementCallbacks(): McpEnablementCallbacks | undefined {
     return this.mcpEnablementCallbacks;
   }
@@ -2306,6 +2337,10 @@ export class Config implements McpContext, AgentLoopContext {
     return this.policyEngine.getApprovalMode();
   }
 
+  isPlanMode(): boolean {
+    return this.getApprovalMode() === ApprovalMode.PLAN;
+  }
+
   getPolicyUpdateConfirmationRequest():
     | PolicyUpdateConfirmationRequest
     | undefined {
@@ -2357,7 +2392,11 @@ export class Config implements McpContext, AgentLoopContext {
       );
     }
 
-    this.policyEngine.setApprovalMode(mode);
+    this.policyEngine.setApprovalMode(
+      mode,
+      this.sandboxPolicyManager?.getModeConfig(mode)?.approvedTools ?? [],
+    );
+    this.refreshSandboxManager();
 
     const isPlanModeTransition =
       currentMode !== mode &&
@@ -3157,6 +3196,7 @@ export class Config implements McpContext, AgentLoopContext {
         visualModel: customConfig.visualModel,
         allowedDomains: customConfig.allowedDomains,
         disableUserInput: customConfig.disableUserInput,
+        maxActionsPerTask: customConfig.maxActionsPerTask ?? 100,
         confirmSensitiveActions: customConfig.confirmSensitiveActions,
         blockFileUploads: customConfig.blockFileUploads,
       },

packages/core/src/config/sandbox-integration.test.ts

@@ -22,6 +22,7 @@ vi.mock('../confirmation-bus/message-bus.js', () => ({
 vi.mock('../policy/policy-engine.js', () => ({
   PolicyEngine: vi.fn().mockImplementation(() => ({
     getExcludedTools: vi.fn().mockReturnValue(new Set()),
+    getApprovalMode: vi.fn().mockReturnValue('yolo'),
   })),
 }));
 vi.mock('../skills/skillManager.js', () => ({

packages/core/src/confirmation-bus/types.ts

@@ -11,6 +11,7 @@ import type {
   DiffStat,
 } from '../tools/tools.js';
 import type { ToolCall } from '../scheduler/types.js';
+import type { SandboxPermissions } from '../services/sandboxManager.js';
 
 export enum MessageBusType {
   TOOL_CONFIRMATION_REQUEST = 'tool-confirmation-request',
@@ -78,6 +79,14 @@ export interface ToolConfirmationResponse {
  * Data-only versions of ToolCallConfirmationDetails for bus transmission.
  */
 export type SerializableConfirmationDetails =
+  | {
+      type: 'sandbox_expansion';
+      title: string;
+      command: string;
+      rootCommand: string;
+      additionalPermissions: SandboxPermissions;
+      systemMessage?: string;
+    }
   | {
       type: 'info';
       title: string;

packages/core/src/core/__snapshots__/prompts.test.ts.snap

@@ -42,6 +42,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, only clarify if critically underspecified; otherwise, work autonomously. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -213,6 +214,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, only clarify if critically underspecified; otherwise, work autonomously. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -503,6 +505,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, only clarify if critically underspecified; otherwise, work autonomously. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -674,6 +677,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, only clarify if critically underspecified; otherwise, work autonomously. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -845,6 +849,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, you must work autonomously as no further user input is available. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -968,6 +973,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, you must work autonomously as no further user input is available. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -1564,6 +1570,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, only clarify if critically underspecified; otherwise, work autonomously. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -1731,6 +1738,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, only clarify if critically underspecified; otherwise, work autonomously. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -1889,6 +1897,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, only clarify if critically underspecified; otherwise, work autonomously. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -2047,6 +2056,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, only clarify if critically underspecified; otherwise, work autonomously. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -2201,6 +2211,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, only clarify if critically underspecified; otherwise, work autonomously. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -2355,6 +2366,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, only clarify if critically underspecified; otherwise, work autonomously. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -2503,6 +2515,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, only clarify if critically underspecified; otherwise, work autonomously. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -2656,6 +2669,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, only clarify if critically underspecified; otherwise, work autonomously. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -2934,6 +2948,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, only clarify if critically underspecified; otherwise, work autonomously. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -3340,6 +3355,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, only clarify if critically underspecified; otherwise, work autonomously. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -3494,6 +3510,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, only clarify if critically underspecified; otherwise, work autonomously. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -3760,6 +3777,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, only clarify if critically underspecified; otherwise, work autonomously. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -3914,6 +3932,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in \`GEMINI.md\` files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. For Directives, only clarify if critically underspecified; otherwise, work autonomously. You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.

packages/core/src/core/client.test.ts

@@ -248,9 +248,6 @@ describe('Gemini Client (client.ts)', () => {
       getEnableHooks: vi.fn().mockReturnValue(false),
       getChatCompression: vi.fn().mockReturnValue(undefined),
       getCompressionThreshold: vi.fn().mockReturnValue(undefined),
-      getShowContextWindowWarning: vi.fn().mockReturnValue(false),
-      getShowContextCompression: vi.fn().mockReturnValue(false),
-      getContextWindowCompressionThreshold: vi.fn().mockReturnValue(0.2),
       getSkipNextSpeakerCheck: vi.fn().mockReturnValue(false),
       getShowModelInfoInChat: vi.fn().mockReturnValue(false),
       getContinueOnFailedApiCall: vi.fn(),
@@ -1620,7 +1617,6 @@ ${JSON.stringify(
             originalTokenCount: initialTokenCount,
             newTokenCount: 400,
             compressionStatus: CompressionStatus.COMPRESSED,
-            requestTokenCount: 50, // Added to match updated interface
           };
         });
 
@@ -1647,13 +1643,10 @@ ${JSON.stringify(
         }),
       );
 
-      // 2. Should contain compression event with requestTokenCount
+      // 2. Should contain compression event
       expect(events).toContainEqual(
         expect.objectContaining({
           type: GeminiEventType.ChatCompressed,
-          value: expect.objectContaining({
-            requestTokenCount: expect.any(Number),
-          }),
         }),
       );
 

packages/core/src/core/client.ts

@@ -608,6 +608,17 @@ export class GeminiClient {
     // Check for context window overflow
     const modelForLimitCheck = this._getActiveModelForCurrentTurn();
 
+    const compressed = await this.tryCompressChat(prompt_id, false);
+
+    if (compressed.compressionStatus === CompressionStatus.COMPRESSED) {
+      yield { type: GeminiEventType.ChatCompressed, value: compressed };
+    }
+
+    const remainingTokenCount =
+      tokenLimit(modelForLimitCheck) - this.getChat().getLastPromptTokenCount();
+
+    await this.tryMaskToolOutputs(this.getHistory());
+
     // Estimate tokens. For text-only requests, we estimate based on character length.
     // For requests with non-text parts (like images, tools), we use the countTokens API.
     const estimatedRequestTokenCount = await calculateRequestTokenCount(
@@ -616,48 +627,12 @@ export class GeminiClient {
       modelForLimitCheck,
     );
 
-    const compressed = await this.tryCompressChat(
-      prompt_id,
-      false,
-      estimatedRequestTokenCount,
-    );
-
-    if (compressed.compressionStatus === CompressionStatus.COMPRESSED) {
-      yield { type: GeminiEventType.ChatCompressed, value: compressed };
-    }
-
-    let remainingTokenCount =
-      tokenLimit(modelForLimitCheck) - this.getChat().getLastPromptTokenCount();
-
-    await this.tryMaskToolOutputs(this.getHistory());
-
     if (estimatedRequestTokenCount > remainingTokenCount) {
-      if (!this.config.getShowContextWindowWarning()) {
-        const forcedCompressed = await this.tryCompressChat(
-          prompt_id,
-          true,
-          estimatedRequestTokenCount,
-        );
-        if (
-          forcedCompressed.compressionStatus === CompressionStatus.COMPRESSED
-        ) {
-          yield {
-            type: GeminiEventType.ChatCompressed,
-            value: forcedCompressed,
-          };
-          remainingTokenCount =
-            tokenLimit(modelForLimitCheck) -
-            this.getChat().getLastPromptTokenCount();
-        }
-      }
-
-      if (estimatedRequestTokenCount > remainingTokenCount) {
-        yield {
-          type: GeminiEventType.ContextWindowWillOverflow,
-          value: { estimatedRequestTokenCount, remainingTokenCount },
-        };
-        return turn;
-      }
+      yield {
+        type: GeminiEventType.ContextWindowWillOverflow,
+        value: { estimatedRequestTokenCount, remainingTokenCount },
+      };
+      return turn;
     }
 
     // Prevent context updates from being sent while a tool call is
@@ -1183,7 +1158,6 @@ export class GeminiClient {
   async tryCompressChat(
     prompt_id: string,
     force: boolean = false,
-    requestTokenCount?: number,
   ): Promise<ChatCompressionInfo> {
     // If the model is 'auto', we will use a placeholder model to check.
     // Compression occurs before we choose a model, so calling `count_tokens`
@@ -1199,11 +1173,6 @@ export class GeminiClient {
       this.hasFailedCompressionAttempt,
     );
 
-    const resultInfo = {
-      ...info,
-      requestTokenCount,
-    };
-
     if (
       info.compressionStatus ===
       CompressionStatus.COMPRESSION_FAILED_INFLATED_TOKEN_COUNT
@@ -1239,7 +1208,7 @@ export class GeminiClient {
       }
     }
 
-    return resultInfo;
+    return info;
   }
 
   /**

packages/core/src/core/prompts.test.ts

@@ -89,6 +89,7 @@ describe('Core System Prompt (prompts.ts)', () => {
     mockConfig = {
       getToolRegistry: vi.fn().mockReturnValue(mockRegistry),
       getEnableShellOutputEfficiency: vi.fn().mockReturnValue(true),
+      getSandboxEnabled: vi.fn().mockReturnValue(false),
       storage: {
         getProjectTempDir: vi.fn().mockReturnValue('/tmp/project-temp'),
         getPlansDir: vi.fn().mockReturnValue('/tmp/project-temp/plans'),
@@ -418,6 +419,7 @@ describe('Core System Prompt (prompts.ts)', () => {
       const testConfig = {
         getToolRegistry: vi.fn().mockReturnValue(mockToolRegistry),
         getEnableShellOutputEfficiency: vi.fn().mockReturnValue(true),
+        getSandboxEnabled: vi.fn().mockReturnValue(false),
         storage: {
           getProjectTempDir: vi.fn().mockReturnValue('/tmp/project-temp'),
         },

packages/core/src/core/turn.ts

@@ -188,7 +188,6 @@ export interface ChatCompressionInfo {
   originalTokenCount: number;
   newTokenCount: number;
   compressionStatus: CompressionStatus;
-  requestTokenCount?: number;
 }
 
 export type ServerGeminiChatCompressedEvent = {

packages/core/src/index.ts

@@ -88,7 +88,11 @@ export * from './utils/approvalModeUtils.js';
 export * from './utils/fileDiffUtils.js';
 export * from './utils/retry.js';
 export * from './utils/shell-utils.js';
-export { PolicyDecision, ApprovalMode } from './policy/types.js';
+export {
+  PolicyDecision,
+  ApprovalMode,
+  PRIORITY_YOLO_ALLOW_ALL,
+} from './policy/types.js';
 export * from './utils/tool-utils.js';
 export * from './utils/terminalSerializer.js';
 export * from './utils/systemEncoding.js';
@@ -126,7 +130,7 @@ export * from './services/FolderTrustDiscoveryService.js';
 export * from './services/chatRecordingService.js';
 export * from './services/fileSystemService.js';
 export * from './services/sandboxedFileSystemService.js';
-export * from './services/windowsSandboxManager.js';
+export * from './sandbox/windows/WindowsSandboxManager.js';
 export * from './services/sessionSummaryUtils.js';
 export * from './services/contextManager.js';
 export * from './services/trackerService.js';

packages/core/src/policy/policies/sandbox-default.toml

@@ -0,0 +1,19 @@
+[modes.plan]
+network = false
+readonly = true
+approvedTools = []
+allowOverrides = false
+
+[modes.default]
+network = false
+readonly = true
+approvedTools = []
+allowOverrides = true
+
+[modes.accepting_edits]
+network = false
+readonly = false
+approvedTools = ['sed', 'grep', 'awk', 'perl', 'cat', 'echo']
+allowOverrides = true
+
+[commands]

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

@@ -329,7 +329,11 @@ describe('PolicyEngine', () => {
       );
 
       // Switch to autoEdit mode
-      engine.setApprovalMode(ApprovalMode.AUTO_EDIT);
+      engine = new PolicyEngine({
+        rules,
+        approvalMode: ApprovalMode.AUTO_EDIT,
+        toolSandboxEnabled: true,
+      });
       expect((await engine.check({ name: 'edit' }, undefined)).decision).toBe(
         PolicyDecision.ALLOW,
       );
@@ -1427,14 +1431,14 @@ describe('PolicyEngine', () => {
 
       engine = new PolicyEngine({ rules });
 
-      // Atomic command "whoami" matches the wildcard rule (ASK_USER).
+      // Atomic command "unknown_command" matches the wildcard rule (ASK_USER).
       // It should NOT be upgraded to ALLOW.
       expect(
         (
           await engine.check(
             {
               name: 'run_shell_command',
-              args: { command: 'whoami' },
+              args: { command: 'unknown_command' },
             },
             undefined,
           )
@@ -1572,7 +1576,7 @@ describe('PolicyEngine', () => {
         },
       ];
 
-      engine = new PolicyEngine({ rules });
+      engine = new PolicyEngine({ rules, toolSandboxEnabled: true });
       engine.setApprovalMode(ApprovalMode.AUTO_EDIT);
 
       const result = await engine.check(

packages/core/src/policy/policy-engine.ts

@@ -5,6 +5,11 @@
  */
 
 import { type FunctionCall } from '@google/genai';
+import {
+  isDangerousCommand,
+  isKnownSafeCommand,
+} from '../sandbox/macos/commandSafety.js';
+import { parse as shellParse } from 'shell-quote';
 import {
   PolicyDecision,
   type PolicyEngineConfig,
@@ -192,6 +197,8 @@ export class PolicyEngine {
   private readonly disableAlwaysAllow: boolean;
   private readonly checkerRunner?: CheckerRunner;
   private approvalMode: ApprovalMode;
+  private toolSandboxEnabled: boolean;
+  private sandboxApprovedTools: string[];
 
   constructor(config: PolicyEngineConfig = {}, checkerRunner?: CheckerRunner) {
     this.rules = (config.rules ?? []).sort(
@@ -242,13 +249,18 @@ export class PolicyEngine {
     this.disableAlwaysAllow = config.disableAlwaysAllow ?? false;
     this.checkerRunner = checkerRunner;
     this.approvalMode = config.approvalMode ?? ApprovalMode.DEFAULT;
+    this.toolSandboxEnabled = config.toolSandboxEnabled ?? false;
+    this.sandboxApprovedTools = config.sandboxApprovedTools ?? [];
   }
 
   /**
    * Update the current approval mode.
    */
-  setApprovalMode(mode: ApprovalMode): void {
+  setApprovalMode(mode: ApprovalMode, sandboxApprovedTools?: string[]): void {
     this.approvalMode = mode;
+    if (sandboxApprovedTools !== undefined) {
+      this.sandboxApprovedTools = sandboxApprovedTools;
+    }
   }
 
   /**
@@ -269,17 +281,58 @@ export class PolicyEngine {
     command: string,
     allowRedirection?: boolean,
   ): boolean {
-    return (
-      !allowRedirection &&
-      hasRedirection(command) &&
-      this.approvalMode !== ApprovalMode.AUTO_EDIT &&
-      this.approvalMode !== ApprovalMode.YOLO
-    );
+    if (allowRedirection) return false;
+    if (!hasRedirection(command)) return false;
+
+    // Do not downgrade (do not ask user) if sandboxing is enabled and in AUTO_EDIT or YOLO
+    if (
+      this.toolSandboxEnabled &&
+      (this.approvalMode === ApprovalMode.AUTO_EDIT ||
+        this.approvalMode === ApprovalMode.YOLO)
+    ) {
+      return false;
+    }
+
+    return true;
   }
 
   /**
    * Check if a shell command is allowed.
    */
+
+  private async applyShellHeuristics(
+    command: string,
+    decision: PolicyDecision,
+  ): Promise<PolicyDecision> {
+    await initializeShellParsers();
+    try {
+      const parsedObjArgs = shellParse(command);
+      if (parsedObjArgs.some((arg) => typeof arg === 'object')) return decision;
+      const parsedArgs = parsedObjArgs.map(String);
+      if (isDangerousCommand(parsedArgs)) {
+        debugLogger.debug(
+          `[PolicyEngine.check] Command evaluated as dangerous, forcing ASK_USER: ${command}`,
+        );
+        return PolicyDecision.ASK_USER;
+      }
+      const isApprovedBySandbox =
+        this.toolSandboxEnabled &&
+        this.sandboxApprovedTools.includes(parsedArgs[0]);
+      if (
+        (isKnownSafeCommand(parsedArgs) || isApprovedBySandbox) &&
+        decision === PolicyDecision.ASK_USER
+      ) {
+        debugLogger.debug(
+          `[PolicyEngine.check] Command evaluated as known safe, overriding ASK_USER to ALLOW: ${command}`,
+        );
+        return PolicyDecision.ALLOW;
+      }
+    } catch {
+      // Ignore parsing errors
+    }
+    return decision;
+  }
+
   private async checkShellCommand(
     toolName: string,
     command: string | undefined,
@@ -522,11 +575,21 @@ export class PolicyEngine {
           `[PolicyEngine.check] MATCHED rule: toolName=${rule.toolName}, decision=${rule.decision}, priority=${rule.priority}, argsPattern=${rule.argsPattern?.source || 'none'}`,
         );
 
+        let ruleDecision = rule.decision;
+        if (
+          isShellCommand &&
+          command &&
+          !('commandPrefix' in rule) &&
+          !rule.argsPattern
+        ) {
+          ruleDecision = await this.applyShellHeuristics(command, ruleDecision);
+        }
+
         if (isShellCommand && toolName) {
           const shellResult = await this.checkShellCommand(
             toolName,
             command,
-            rule.decision,
+            ruleDecision,
             serverName,
             shellDirPath,
             rule.allowRedirection,
@@ -562,10 +625,18 @@ export class PolicyEngine {
         `[PolicyEngine.check] NO MATCH - using default decision: ${this.defaultDecision}`,
       );
       if (toolName && SHELL_TOOL_NAMES.includes(toolName)) {
+        let heuristicDecision = this.defaultDecision;
+        if (command) {
+          heuristicDecision = await this.applyShellHeuristics(
+            command,
+            heuristicDecision,
+          );
+        }
+
         const shellResult = await this.checkShellCommand(
           toolName,
           command,
-          this.defaultDecision,
+          heuristicDecision,
           serverName,
           shellDirPath,
           false,
@@ -631,6 +702,15 @@ export class PolicyEngine {
       }
     }
 
+    // Sandbox Expansion requests MUST always be confirmed by the user,
+    // even if the base command is otherwise ALLOWED by the policy engine.
+    if (
+      decision === PolicyDecision.ALLOW &&
+      toolCall.args?.['additional_permissions']
+    ) {
+      decision = PolicyDecision.ASK_USER;
+    }
+
     return {
       decision: this.applyNonInteractiveMode(decision),
       rule: matchedRule,

packages/core/src/policy/sandboxPolicyManager.ts

@@ -0,0 +1,216 @@
+/**
+ * @license
+ * Copyright 2026 Google LLC
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import fs from 'node:fs';
+import path from 'node:path';
+import os from 'node:os';
+import toml from '@iarna/toml';
+import { z } from 'zod';
+import { fileURLToPath } from 'node:url';
+import { debugLogger } from '../utils/debugLogger.js';
+import { type SandboxPermissions } from '../services/sandboxManager.js';
+import { sanitizePaths } from '../services/sandboxManager.js';
+
+export const SandboxModeConfigSchema = z.object({
+  network: z.boolean(),
+  readonly: z.boolean(),
+  approvedTools: z.array(z.string()),
+  allowOverrides: z.boolean().optional(),
+});
+
+export const PersistentCommandConfigSchema = z.object({
+  allowed_paths: z.array(z.string()).optional(),
+  allow_network: z.boolean().optional(),
+});
+
+export const SandboxTomlSchema = z.object({
+  modes: z.object({
+    plan: SandboxModeConfigSchema,
+    default: SandboxModeConfigSchema,
+    accepting_edits: SandboxModeConfigSchema,
+  }),
+  commands: z.record(z.string(), PersistentCommandConfigSchema).default({}),
+});
+
+export type SandboxModeConfig = z.infer<typeof SandboxModeConfigSchema>;
+export type PersistentCommandConfig = z.infer<
+  typeof PersistentCommandConfigSchema
+>;
+export type SandboxTomlSchemaType = z.infer<typeof SandboxTomlSchema>;
+
+export class SandboxPolicyManager {
+  private static _DEFAULT_CONFIG: SandboxTomlSchemaType | null = null;
+
+  private static get DEFAULT_CONFIG(): SandboxTomlSchemaType {
+    if (!SandboxPolicyManager._DEFAULT_CONFIG) {
+      const __filename = fileURLToPath(import.meta.url);
+      const __dirname = path.dirname(__filename);
+      const defaultPath = path.join(
+        __dirname,
+        'policies',
+        'sandbox-default.toml',
+      );
+      try {
+        const content = fs.readFileSync(defaultPath, 'utf8');
+        if (typeof content !== 'string') {
+          SandboxPolicyManager._DEFAULT_CONFIG = {
+            modes: {
+              plan: {
+                network: false,
+                readonly: true,
+                approvedTools: [],
+                allowOverrides: false,
+              },
+              default: {
+                network: false,
+                readonly: true,
+                approvedTools: [],
+                allowOverrides: true,
+              },
+              accepting_edits: {
+                network: false,
+                readonly: false,
+                approvedTools: ['sed', 'grep', 'awk', 'perl', 'cat', 'echo'],
+                allowOverrides: true,
+              },
+            },
+            commands: {},
+          };
+          return SandboxPolicyManager._DEFAULT_CONFIG;
+        }
+        SandboxPolicyManager._DEFAULT_CONFIG = SandboxTomlSchema.parse(
+          toml.parse(content),
+        );
+      } catch (e) {
+        debugLogger.error(`Failed to parse default sandbox policy: ${e}`);
+        throw new Error(`Failed to parse default sandbox policy: ${e}`);
+      }
+    }
+    return SandboxPolicyManager._DEFAULT_CONFIG;
+  }
+
+  private config: SandboxTomlSchemaType;
+  private readonly configPath: string;
+  private sessionApprovals: Record<string, SandboxPermissions> = {};
+
+  constructor(customConfigPath?: string) {
+    this.configPath =
+      customConfigPath ??
+      path.join(os.homedir(), '.gemini', 'policies', 'sandbox.toml');
+    this.config = this.loadConfig();
+  }
+
+  private loadConfig(): SandboxTomlSchemaType {
+    if (!fs.existsSync(this.configPath)) {
+      return SandboxPolicyManager.DEFAULT_CONFIG;
+    }
+
+    try {
+      const content = fs.readFileSync(this.configPath, 'utf8');
+      return SandboxTomlSchema.parse(toml.parse(content));
+    } catch (e) {
+      debugLogger.error(`Failed to parse sandbox.toml: ${e}`);
+      return SandboxPolicyManager.DEFAULT_CONFIG;
+    }
+  }
+
+  private saveConfig(): void {
+    try {
+      const dir = path.dirname(this.configPath);
+      if (!fs.existsSync(dir)) {
+        fs.mkdirSync(dir, { recursive: true });
+      }
+      // eslint-disable-next-line @typescript-eslint/no-unsafe-type-assertion
+      const content = toml.stringify(this.config as unknown as toml.JsonMap);
+      fs.writeFileSync(this.configPath, content);
+    } catch (e) {
+      debugLogger.error(`Failed to save sandbox.toml: ${e}`);
+    }
+  }
+
+  getModeConfig(
+    mode: 'plan' | 'accepting_edits' | 'default' | string,
+  ): SandboxModeConfig {
+    if (mode === 'plan') return this.config.modes.plan;
+    if (mode === 'accepting_edits' || mode === 'autoEdit')
+      return this.config.modes.accepting_edits;
+    if (mode === 'default') return this.config.modes.default;
+
+    // Default fallback
+    return this.config.modes.default ?? this.config.modes.plan;
+  }
+
+  getCommandPermissions(commandName: string): SandboxPermissions {
+    const persistent = this.config.commands[commandName];
+    const session = this.sessionApprovals[commandName];
+
+    return {
+      fileSystem: {
+        read: [
+          ...(persistent?.allowed_paths ?? []),
+          ...(session?.fileSystem?.read ?? []),
+        ],
+        write: [
+          ...(persistent?.allowed_paths ?? []),
+          ...(session?.fileSystem?.write ?? []),
+        ],
+      },
+      network: persistent?.allow_network || session?.network || false,
+    };
+  }
+
+  addSessionApproval(
+    commandName: string,
+    permissions: SandboxPermissions,
+  ): void {
+    const existing = this.sessionApprovals[commandName] || {
+      fileSystem: { read: [], write: [] },
+      network: false,
+    };
+
+    this.sessionApprovals[commandName] = {
+      fileSystem: {
+        read: Array.from(
+          new Set([
+            ...(existing.fileSystem?.read ?? []),
+            ...(permissions.fileSystem?.read ?? []),
+          ]),
+        ),
+        write: Array.from(
+          new Set([
+            ...(existing.fileSystem?.write ?? []),
+            ...(permissions.fileSystem?.write ?? []),
+          ]),
+        ),
+      },
+      network: existing.network || permissions.network || false,
+    };
+  }
+
+  addPersistentApproval(
+    commandName: string,
+    permissions: SandboxPermissions,
+  ): void {
+    const existing = this.config.commands[commandName] || {
+      allowed_paths: [],
+      allow_network: false,
+    };
+
+    const newPathsArray: string[] = [
+      ...(existing.allowed_paths ?? []),
+      ...(permissions.fileSystem?.read ?? []),
+      ...(permissions.fileSystem?.write ?? []),
+    ];
+    const newPaths = new Set(sanitizePaths(newPathsArray));
+
+    this.config.commands[commandName] = {
+      allowed_paths: Array.from(newPaths),
+      allow_network: existing.allow_network || permissions.network || false,
+    };
+
+    this.saveConfig();
+  }
+}

packages/core/src/policy/types.ts

@@ -309,6 +309,15 @@ export interface PolicyEngineConfig {
    * Used to filter rules that have specific 'modes' defined.
    */
   approvalMode?: ApprovalMode;
+
+  /**
+   * Whether tool sandboxing is enabled.
+   */
+  toolSandboxEnabled?: boolean;
+  /**
+   * List of tools approved by the sandbox policy for the current mode.
+   */
+  sandboxApprovedTools?: string[];
 }
 
 export interface PolicySettings {

packages/core/src/prompts/promptProvider.test.ts

@@ -54,6 +54,7 @@ describe('PromptProvider', () => {
       },
       getToolRegistry: vi.fn().mockReturnValue(mockToolRegistry),
       getEnableShellOutputEfficiency: vi.fn().mockReturnValue(true),
+      getSandboxEnabled: vi.fn().mockReturnValue(false),
       storage: {
         getProjectTempDir: vi.fn().mockReturnValue('/tmp/project-temp'),
         getPlansDir: vi.fn().mockReturnValue('/tmp/project-temp/plans'),

packages/core/src/prompts/promptProvider.ts

@@ -195,7 +195,10 @@ export class PromptProvider {
             memoryManagerEnabled: context.config.isMemoryManagerEnabled(),
           }),
         ),
-        sandbox: this.withSection('sandbox', () => getSandboxMode()),
+        sandbox: this.withSection('sandbox', () => ({
+          mode: getSandboxMode(),
+          toolSandboxingEnabled: context.config.getSandboxEnabled(),
+        })),
         interactiveYoloMode: this.withSection(
           'interactiveYoloMode',
           () => true,

packages/core/src/prompts/snippets.legacy.ts

@@ -36,7 +36,7 @@ export interface SystemPromptOptions {
   planningWorkflow?: PlanningWorkflowOptions;
   taskTracker?: boolean;
   operationalGuidelines?: OperationalGuidelinesOptions;
-  sandbox?: SandboxMode;
+  sandbox?: SandboxOptions;
   interactiveYoloMode?: boolean;
   gitRepo?: GitRepoOptions;
   finalReminder?: FinalReminderOptions;
@@ -72,6 +72,11 @@ export interface OperationalGuidelinesOptions {
 
 export type SandboxMode = 'macos-seatbelt' | 'generic' | 'outside';
 
+export interface SandboxOptions {
+  mode: SandboxMode;
+  toolSandboxingEnabled: boolean;
+}
+
 export interface GitRepoOptions {
   interactive: boolean;
 }
@@ -290,8 +295,9 @@ ${shellEfficiencyGuidelines(options.enableShellEfficiency)}
 `.trim();
 }
 
-export function renderSandbox(mode?: SandboxMode): string {
-  if (!mode) return '';
+export function renderSandbox(options?: SandboxOptions): string {
+  if (!options || !options.mode) return '';
+  const mode = options.mode;
   if (mode === 'macos-seatbelt') {
     return `
 # macOS Seatbelt
@@ -300,11 +306,12 @@ You are running under macos seatbelt with limited access to files outside the pr
     return `
 # Sandbox
 You are running in a sandbox container with limited access to files outside the project directory or system temp directory, and with limited access to host system resources such as ports. If you encounter failures that could be due to sandboxing (e.g. if a command fails with 'Operation not permitted' or similar error), when you report the error to the user, also explain why you think it could be due to sandboxing, and how the user may need to adjust their sandbox configuration.`.trim();
-  } else {
+  } else if (mode === 'outside') {
     return `
 # Outside of Sandbox
 You are running outside of a sandbox container, directly on the user's system. For critical commands that are particularly likely to modify the user's system outside of the project directory or system temp directory, as you explain the command to the user (per the Explain Critical Commands rule above), also remind the user to consider enabling sandboxing.`.trim();
   }
+  return '';
 }
 
 export function renderInteractiveYoloMode(enabled?: boolean): string {

packages/core/src/prompts/snippets.ts

@@ -46,7 +46,7 @@ export interface SystemPromptOptions {
   planningWorkflow?: PlanningWorkflowOptions;
   taskTracker?: boolean;
   operationalGuidelines?: OperationalGuidelinesOptions;
-  sandbox?: SandboxMode;
+  sandbox?: SandboxOptions;
   interactiveYoloMode?: boolean;
   gitRepo?: GitRepoOptions;
 }
@@ -84,6 +84,11 @@ export interface OperationalGuidelinesOptions {
 
 export type SandboxMode = 'macos-seatbelt' | 'generic' | 'outside';
 
+export interface SandboxOptions {
+  mode: SandboxMode;
+  toolSandboxingEnabled: boolean;
+}
+
 export interface GitRepoOptions {
   interactive: boolean;
 }
@@ -222,6 +227,7 @@ Use the following guidelines to optimize your search and read patterns.
 ## Engineering Standards
 - **Contextual Precedence:** Instructions found in ${formattedFilenames} files are foundational mandates. They take absolute precedence over the general workflows and tool defaults described in this system prompt.
 - **Conventions & Style:** Rigorously adhere to existing workspace conventions, architectural patterns, and style (naming, formatting, typing, commenting). During the research phase, analyze surrounding files, tests, and configuration to ensure your changes are seamless, idiomatic, and consistent with the local context. Never compromise idiomatic quality or completeness (e.g., proper declarations, type safety, documentation) to minimize tool calls; all supporting changes required by local conventions are part of a surgical update.
+- **Types, warnings and linters:** NEVER use hacks like disabling or suppressing warnings or bypassing the type system (i.e.: casts in TypeScript) unless explicitly instructed to by the user. Instead, use idiomatic language features (e.g.: type guard functions).
 - **Libraries/Frameworks:** NEVER assume a library/framework is available. Verify its established usage within the project (check imports, configuration files like 'package.json', 'Cargo.toml', 'requirements.txt', etc.) before employing it.
 - **Technical Integrity:** You are responsible for the entire lifecycle: implementation, testing, and validation. Within the scope of your changes, prioritize readability and long-term maintainability by consolidating logic into clean abstractions rather than threading state across unrelated layers. Align strictly with the requested architectural direction, ensuring the final implementation is focused and free of redundant "just-in-case" alternatives. Validation is not merely running tests; it is the exhaustive process of ensuring that every aspect of your change—behavioral, structural, and stylistic—is correct and fully compatible with the broader project. For bug fixes, you must empirically reproduce the failure with a new test case or reproduction script before applying the fix.
 - **Expertise & Intent Alignment:** Provide proactive technical opinions grounded in research while strictly adhering to the user's intended workflow. Distinguish between **Directives** (unambiguous requests for action or implementation) and **Inquiries** (requests for analysis, advice, or observations). Assume all requests are Inquiries unless they contain an explicit instruction to perform a task. For Inquiries, your scope is strictly limited to research and analysis; you may propose a solution or strategy, but you MUST NOT modify files until a corresponding Directive is issued. Do not initiate implementation based on observations of bugs or statements of fact. Once an Inquiry is resolved, or while waiting for a Directive, stop and wait for the next user instruction. ${options.interactive ? 'For Directives, only clarify if critically underspecified; otherwise, work autonomously.' : 'For Directives, you must work autonomously as no further user input is available.'} You should only seek user intervention if you have exhausted all possible routes or if a proposed solution would take the workspace in a significantly different architectural direction.
@@ -389,18 +395,44 @@ export function renderOperationalGuidelines(
 `.trim();
 }
 
-export function renderSandbox(mode?: SandboxMode): string {
-  if (!mode) return '';
+export function renderSandbox(options?: SandboxOptions): string {
+  if (!options || !options.mode) return '';
+  const { mode, toolSandboxingEnabled } = options;
   if (mode === 'macos-seatbelt') {
-    return `
+    if (toolSandboxingEnabled) {
+      return `
+    # macOS Seatbelt
+    
+    You are running under macos seatbelt with limited access to files outside the project directory or system temp directory, and with limited access to host system resources such as ports. 
+    **Sandbox Failure Recovery:**
+    If an interactive command (like nvim, nano, git) fails to load plugins, crashes, or hangs waiting for user input (e.g. at a '-- More --' prompt) with errors indicating missing files or 'Operation not permitted', it is likely blocked by the sandbox.
+    If you see these errors (e.g. Vim:E303, 'should be read/write') in the command output (even if the command timed out or was cancelled by the user):
+    1.  IMMEDIATELY analyze the output to identify the required filesystem paths or network access.
+    2.  WITHOUT asking the user or explaining the failure first, AUTOMATICALLY retry the command using the 'run_shell_command' tool, explicitly providing the missing permissions in the 'additional_permissions' parameter.
+    3.  This will immediately present the user with a modal to approve the expansion for the command so they don't have to reprompt you.`.trim();
+    } else {
+      return `
     # macOS Seatbelt
     
     You are running under macos seatbelt with limited access to files outside the project directory or system temp directory, and with limited access to host system resources such as ports. If you encounter failures that could be due to macOS Seatbelt (e.g. if a command fails with 'Operation not permitted' or similar error), as you report the error to the user, also explain why you think it could be due to macOS Seatbelt, and how the user may need to adjust their Seatbelt profile.`.trim();
+    }
   } else if (mode === 'generic') {
-    return `
+    if (toolSandboxingEnabled) {
+      return `
+      # Sandbox
+      
+      You are running in a sandbox container with limited access to files outside the project directory or system temp directory, and with limited access to host system resources such as ports. 
+    **Sandbox Failure Recovery:**
+    If a command fails with 'Operation not permitted' or similar sandbox errors, do NOT ask the user to adjust settings manually. Instead:
+    1.  Analyze the command and error to identify the required filesystem paths or network access.
+    2.  Retry the command using the 'run_shell_command' tool, providing the missing permissions in the 'additional_permissions' parameter.
+    3.  The user will be presented with a modal to approve this expansion for the current command.`.trim();
+    } else {
+      return `
       # Sandbox
       
       You are running in a sandbox container with limited access to files outside the project directory or system temp directory, and with limited access to host system resources such as ports. If you encounter failures that could be due to sandboxing (e.g. if a command fails with 'Operation not permitted' or similar error), when you report the error to the user, also explain why you think it could be due to sandboxing, and how the user may need to adjust their sandbox configuration.`.trim();
+    }
   }
   return '';
 }

packages/core/src/sandbox/macos/MacOsSandboxManager.test.ts

@@ -63,7 +63,7 @@ describe('MacOsSandboxManager', () => {
       expect(profile).toContain('(deny default)');
       expect(profile).toContain('(allow process-exec)');
       expect(profile).toContain('(subpath (param "WORKSPACE"))');
-      expect(profile).not.toContain('(allow network*)');
+      expect(profile).not.toContain('(allow network-outbound)');
 
       expect(result.args).toContain('-D');
       expect(result.args).toContain(`WORKSPACE=${mockWorkspace}`);
@@ -91,7 +91,7 @@ describe('MacOsSandboxManager', () => {
       });
 
       const profile = result.args[1];
-      expect(profile).toContain('(allow network*)');
+      expect(profile).toContain('(allow network-outbound)');
     });
 
     it('should parameterize allowed paths and normalize them', async () => {

packages/core/src/sandbox/macos/MacOsSandboxManager.ts

@@ -4,41 +4,164 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-import fs from 'node:fs';
-import os from 'node:os';
-import path from 'node:path';
 import {
   type SandboxManager,
-  type GlobalSandboxOptions,
   type SandboxRequest,
   type SandboxedCommand,
-  type ExecutionPolicy,
-  sanitizePaths,
-  GOVERNANCE_FILES,
+  type SandboxPermissions,
+  type GlobalSandboxOptions,
 } from '../../services/sandboxManager.js';
 import {
   sanitizeEnvironment,
   getSecureSanitizationConfig,
+  type EnvironmentSanitizationConfig,
 } from '../../services/environmentSanitization.js';
+import { buildSeatbeltArgs } from './seatbeltArgsBuilder.js';
 import {
-  BASE_SEATBELT_PROFILE,
-  NETWORK_SEATBELT_PROFILE,
-} from './baseProfile.js';
+  getCommandRoots,
+  initializeShellParsers,
+  splitCommands,
+  stripShellWrapper,
+} from '../../utils/shell-utils.js';
+import { isKnownSafeCommand } from './commandSafety.js';
+import { parse as shellParse } from 'shell-quote';
+import { type SandboxPolicyManager } from '../../policy/sandboxPolicyManager.js';
+import path from 'node:path';
+
+export interface MacOsSandboxOptions extends GlobalSandboxOptions {
+  /** Optional base sanitization config. */
+  sanitizationConfig?: EnvironmentSanitizationConfig;
+  /** The current sandbox mode behavior from config. */
+  modeConfig?: {
+    readonly?: boolean;
+    network?: boolean;
+    approvedTools?: string[];
+    allowOverrides?: boolean;
+  };
+  /** The policy manager for persistent approvals. */
+  policyManager?: SandboxPolicyManager;
+}
 
 /**
  * A SandboxManager implementation for macOS that uses Seatbelt.
  */
 export class MacOsSandboxManager implements SandboxManager {
-  constructor(private readonly options: GlobalSandboxOptions) {}
+  constructor(private readonly options: MacOsSandboxOptions) {}
+
+  private async isStrictlyApproved(req: SandboxRequest): Promise<boolean> {
+    const approvedTools = this.options.modeConfig?.approvedTools;
+    if (!approvedTools || approvedTools.length === 0) {
+      return false;
+    }
+
+    await initializeShellParsers();
+
+    const fullCmd = [req.command, ...req.args].join(' ');
+    const stripped = stripShellWrapper(fullCmd);
+
+    const roots = getCommandRoots(stripped);
+    if (roots.length === 0) return false;
+
+    const allRootsApproved = roots.every((root) =>
+      approvedTools.includes(root),
+    );
+    if (allRootsApproved) {
+      return true;
+    }
+
+    const pipelineCommands = splitCommands(stripped);
+    if (pipelineCommands.length === 0) return false;
+
+    // For safety, every command in the pipeline must be considered safe.
+    for (const cmdString of pipelineCommands) {
+      const parsedArgs = shellParse(cmdString).map(String);
+      if (!isKnownSafeCommand(parsedArgs)) {
+        return false;
+      }
+    }
+
+    return true;
+  }
+
+  private async getCommandName(req: SandboxRequest): Promise<string> {
+    await initializeShellParsers();
+    const fullCmd = [req.command, ...req.args].join(' ');
+    const stripped = stripShellWrapper(fullCmd);
+    const roots = getCommandRoots(stripped).filter(
+      (r) => r !== 'shopt' && r !== 'set',
+    );
+    if (roots.length > 0) {
+      return roots[0];
+    }
+    return path.basename(req.command);
+  }
 
   async prepareCommand(req: SandboxRequest): Promise<SandboxedCommand> {
+    await initializeShellParsers();
     const sanitizationConfig = getSecureSanitizationConfig(
       req.policy?.sanitizationConfig,
     );
 
     const sanitizedEnv = sanitizeEnvironment(req.env, sanitizationConfig);
 
-    const sandboxArgs = this.buildSeatbeltArgs(this.options, req.policy);
+    const isReadonlyMode = this.options.modeConfig?.readonly ?? true;
+    const allowOverrides = this.options.modeConfig?.allowOverrides ?? true;
+
+    // Reject override attempts in plan mode
+    if (!allowOverrides && req.policy?.additionalPermissions) {
+      const perms = req.policy.additionalPermissions;
+      if (
+        perms.network ||
+        (perms.fileSystem?.write && perms.fileSystem.write.length > 0)
+      ) {
+        throw new Error(
+          'Sandbox request rejected: Cannot override readonly/network restrictions in Plan mode.',
+        );
+      }
+    }
+
+    // If not in readonly mode OR it's a strictly approved pipeline, allow workspace writes
+    const isApproved = allowOverrides
+      ? await this.isStrictlyApproved(req)
+      : false;
+
+    const workspaceWrite = !isReadonlyMode || isApproved;
+    const networkAccess =
+      this.options.modeConfig?.network ?? req.policy?.networkAccess ?? false;
+
+    // Fetch persistent approvals for this command
+    const commandName = await this.getCommandName(req);
+    const persistentPermissions = allowOverrides
+      ? this.options.policyManager?.getCommandPermissions(commandName)
+      : undefined;
+
+    // Merge all permissions
+    const mergedAdditional: SandboxPermissions = {
+      fileSystem: {
+        read: [
+          ...(persistentPermissions?.fileSystem?.read ?? []),
+          ...(req.policy?.additionalPermissions?.fileSystem?.read ?? []),
+        ],
+        write: [
+          ...(persistentPermissions?.fileSystem?.write ?? []),
+          ...(req.policy?.additionalPermissions?.fileSystem?.write ?? []),
+        ],
+      },
+      network:
+        networkAccess ||
+        persistentPermissions?.network ||
+        req.policy?.additionalPermissions?.network ||
+        false,
+    };
+
+    const sandboxArgs = buildSeatbeltArgs({
+      workspace: this.options.workspace,
+      allowedPaths: [...(req.policy?.allowedPaths || [])],
+      forbiddenPaths: req.policy?.forbiddenPaths,
+      networkAccess: mergedAdditional.network,
+      workspaceWrite,
+      additionalPermissions: mergedAdditional,
+    });
 
     return {
       program: '/usr/bin/sandbox-exec',
@@ -47,124 +170,4 @@ export class MacOsSandboxManager implements SandboxManager {
       cwd: req.cwd,
     };
   }
-
-  /**
-   * Builds the arguments array for sandbox-exec using a strict allowlist profile.
-   * It relies on parameters passed to sandbox-exec via the -D flag to avoid
-   * string interpolation vulnerabilities, and normalizes paths against symlink escapes.
-   *
-   * Returns arguments up to the end of sandbox-exec configuration (e.g. ['-p', '<profile>', '-D', ...])
-   * Does not include the final '--' separator or the command to run.
-   */
-  private buildSeatbeltArgs(
-    options: GlobalSandboxOptions,
-    policy?: ExecutionPolicy,
-  ): string[] {
-    const profileLines = [BASE_SEATBELT_PROFILE];
-    const args: string[] = [];
-
-    const workspacePath = this.tryRealpath(options.workspace);
-    args.push('-D', `WORKSPACE=${workspacePath}`);
-
-    // Add explicit deny rules for governance files in the workspace.
-    // These are added after the workspace allow rule (which is in BASE_SEATBELT_PROFILE)
-    // to ensure they take precedence (Seatbelt evaluates rules in order, later rules win for same path).
-    for (let i = 0; i < GOVERNANCE_FILES.length; i++) {
-      const governanceFile = path.join(workspacePath, GOVERNANCE_FILES[i].path);
-
-      // Ensure the file/directory exists so Seatbelt rules are reliably applied.
-      this.touch(governanceFile, GOVERNANCE_FILES[i].isDirectory);
-
-      const realGovernanceFile = this.tryRealpath(governanceFile);
-
-      // Determine if it should be treated as a directory (subpath) or a file (literal).
-      // .git is generally a directory, while ignore files are literals.
-      let isActuallyDirectory = GOVERNANCE_FILES[i].isDirectory;
-      try {
-        if (fs.existsSync(realGovernanceFile)) {
-          isActuallyDirectory = fs.lstatSync(realGovernanceFile).isDirectory();
-        }
-      } catch {
-        // Ignore errors, use default guess
-      }
-
-      const ruleType = isActuallyDirectory ? 'subpath' : 'literal';
-
-      args.push('-D', `GOVERNANCE_FILE_${i}=${governanceFile}`);
-      profileLines.push(
-        `(deny file-write* (${ruleType} (param "GOVERNANCE_FILE_${i}")))`,
-      );
-
-      if (realGovernanceFile !== governanceFile) {
-        args.push('-D', `REAL_GOVERNANCE_FILE_${i}=${realGovernanceFile}`);
-        profileLines.push(
-          `(deny file-write* (${ruleType} (param "REAL_GOVERNANCE_FILE_${i}")))`,
-        );
-      }
-    }
-
-    const tmpPath = this.tryRealpath(os.tmpdir());
-    args.push('-D', `TMPDIR=${tmpPath}`);
-
-    const allowedPaths = sanitizePaths(policy?.allowedPaths) || [];
-    for (let i = 0; i < allowedPaths.length; i++) {
-      const allowedPath = this.tryRealpath(allowedPaths[i]);
-      args.push('-D', `ALLOWED_PATH_${i}=${allowedPath}`);
-      profileLines.push(
-        `(allow file-read* file-write* (subpath (param "ALLOWED_PATH_${i}")))`,
-      );
-    }
-
-    // TODO: handle forbidden paths
-
-    if (policy?.networkAccess) {
-      profileLines.push(NETWORK_SEATBELT_PROFILE);
-    }
-
-    args.unshift('-p', profileLines.join('\n'));
-
-    return args;
-  }
-
-  /**
-   * Ensures a file or directory exists.
-   */
-  private touch(filePath: string, isDirectory: boolean) {
-    try {
-      // If it exists (even as a broken symlink), do nothing
-      if (fs.lstatSync(filePath)) return;
-    } catch {
-      // Ignore ENOENT
-    }
-
-    if (isDirectory) {
-      fs.mkdirSync(filePath, { recursive: true });
-    } else {
-      const dir = path.dirname(filePath);
-      if (!fs.existsSync(dir)) {
-        fs.mkdirSync(dir, { recursive: true });
-      }
-      fs.closeSync(fs.openSync(filePath, 'a'));
-    }
-  }
-
-  /**
-   * Resolves symlinks for a given path to prevent sandbox escapes.
-   * If a file does not exist (ENOENT), it recursively resolves the parent directory.
-   * Other errors (e.g. EACCES) are re-thrown.
-   */
-  private tryRealpath(p: string): string {
-    try {
-      return fs.realpathSync(p);
-    } catch (e) {
-      if (e instanceof Error && 'code' in e && e.code === 'ENOENT') {
-        const parentDir = path.dirname(p);
-        if (parentDir === p) {
-          return p;
-        }
-        return path.join(this.tryRealpath(parentDir), path.basename(p));
-      }
-      throw e;
-    }
-  }
 }

packages/core/src/sandbox/macos/baseProfile.ts

@@ -16,11 +16,101 @@ export const BASE_SEATBELT_PROFILE = `(version 1)
 
 (import "system.sb")
 
+
 ; Core execution requirements
 (allow process-exec)
 (allow process-fork)
 (allow signal (target same-sandbox))
-(allow process-info* (target same-sandbox))
+(allow process-info*)
+
+(allow file-write-data
+  (require-all
+    (path "/dev/null")
+    (vnode-type CHARACTER-DEVICE)))
+
+; sysctls permitted.
+(allow sysctl-read
+  (sysctl-name "hw.activecpu")
+  (sysctl-name "hw.busfrequency_compat")
+  (sysctl-name "hw.byteorder")
+  (sysctl-name "hw.cacheconfig")
+  (sysctl-name "hw.cachelinesize_compat")
+  (sysctl-name "hw.cpufamily")
+  (sysctl-name "hw.cpufrequency_compat")
+  (sysctl-name "hw.cputype")
+  (sysctl-name "hw.l1dcachesize_compat")
+  (sysctl-name "hw.l1icachesize_compat")
+  (sysctl-name "hw.l2cachesize_compat")
+  (sysctl-name "hw.l3cachesize_compat")
+  (sysctl-name "hw.logicalcpu_max")
+  (sysctl-name "hw.machine")
+  (sysctl-name "hw.model")
+  (sysctl-name "hw.memsize")
+  (sysctl-name "hw.ncpu")
+  (sysctl-name "hw.nperflevels")
+  (sysctl-name-prefix "hw.optional.arm.")
+  (sysctl-name-prefix "hw.optional.armv8_")
+  (sysctl-name "hw.packages")
+  (sysctl-name "hw.pagesize_compat")
+  (sysctl-name "hw.pagesize")
+  (sysctl-name "hw.physicalcpu")
+  (sysctl-name "hw.physicalcpu_max")
+  (sysctl-name "hw.logicalcpu")
+  (sysctl-name "hw.cpufrequency")
+  (sysctl-name "hw.tbfrequency_compat")
+  (sysctl-name "hw.vectorunit")
+  (sysctl-name "machdep.cpu.brand_string")
+  (sysctl-name "kern.argmax")
+  (sysctl-name "kern.hostname")
+  (sysctl-name "kern.maxfilesperproc")
+  (sysctl-name "kern.maxproc")
+  (sysctl-name "kern.osproductversion")
+  (sysctl-name "kern.osrelease")
+  (sysctl-name "kern.ostype")
+  (sysctl-name "kern.osvariant_status")
+  (sysctl-name "kern.osversion")
+  (sysctl-name "kern.secure_kernel")
+  (sysctl-name "kern.usrstack64")
+  (sysctl-name "kern.version")
+  (sysctl-name "sysctl.proc_cputype")
+  (sysctl-name "vm.loadavg")
+  (sysctl-name-prefix "hw.perflevel")
+  (sysctl-name-prefix "kern.proc.pgrp.")
+  (sysctl-name-prefix "kern.proc.pid.")
+  (sysctl-name-prefix "net.routetable.")
+)
+
+(allow sysctl-write
+  (sysctl-name "kern.grade_cputype"))
+
+
+(allow mach-lookup
+  (global-name "com.apple.sysmond")
+)
+\n; IOKit
+(allow iokit-open
+  (iokit-registry-entry-class "RootDomainUserClient")
+)
+
+(allow mach-lookup
+  (global-name "com.apple.system.opendirectoryd.libinfo")
+)
+
+; Needed for python multiprocessing on MacOS for the SemLock
+(allow ipc-posix-sem)
+
+(allow mach-lookup
+  (global-name "com.apple.PowerManagement.control")
+)
+
+; PTY and Terminal support
+(allow pseudo-tty)
+(allow file-read* file-write* file-ioctl (literal "/dev/ptmx"))
+(allow file-read* file-write*
+  (require-all
+    (regex #"^/dev/ttys[0-9]+")
+    (extension "com.apple.sandbox.pty")))
+(allow file-ioctl (regex #"^/dev/ttys[0-9]+"))
 
 ; Allow basic read access to system frameworks and libraries required to run
 (allow file-read*
@@ -38,11 +128,6 @@ export const BASE_SEATBELT_PROFILE = `(version 1)
   (subpath "/private/etc")
 )
 
-; PTY and Terminal support
-(allow pseudo-tty)
-(allow file-read* file-write* file-ioctl (literal "/dev/ptmx"))
-(allow file-read* file-write* file-ioctl (regex #"^/dev/ttys[0-9]+"))
-
 ; Allow read/write access to temporary directories and common device nodes
 (allow file-read* file-write*
   (literal "/dev/null")
@@ -53,9 +138,10 @@ export const BASE_SEATBELT_PROFILE = `(version 1)
 )
 
 ; Workspace access using parameterized paths
-(allow file-read* file-write*
+(allow file-read*
   (subpath (param "WORKSPACE"))
 )
+
 `;
 
 /**
@@ -66,7 +152,9 @@ export const BASE_SEATBELT_PROFILE = `(version 1)
  */
 export const NETWORK_SEATBELT_PROFILE = `
 ; Network Access
-(allow network*)
+(allow network-outbound)
+(allow network-inbound)
+(allow network-bind)
 
 (allow system-socket
   (require-all

packages/core/src/sandbox/macos/commandSafety.ts

@@ -0,0 +1,469 @@
+/**
+ * @license
+ * Copyright 2026 Google LLC
+ * SPDX-License-Identifier: Apache-2.0
+ */
+import { parse as shellParse } from 'shell-quote';
+
+/**
+ * Checks if a command with its arguments is known to be safe to execute
+ * without requiring user confirmation. This is primarily used to allow
+ * harmless, read-only commands to run silently in the macOS sandbox.
+ *
+ * It handles raw command execution as well as wrapped commands like `bash -c "..."` or `bash -lc "..."`.
+ * For wrapped commands, it parses the script and ensures all individual
+ * sub-commands are in the known-safe list and no dangerous shell operators
+ * (like subshells or redirection) are used.
+ *
+ * @param args - The command and its arguments (e.g., ['ls', '-la'])
+ * @returns true if the command is considered safe, false otherwise.
+ */
+export function isKnownSafeCommand(args: string[]): boolean {
+  if (!args || args.length === 0) {
+    return false;
+  }
+
+  // Normalize zsh to bash
+  const normalizedArgs = args.map((a) => (a === 'zsh' ? 'bash' : a));
+
+  if (isSafeToCallWithExec(normalizedArgs)) {
+    return true;
+  }
+
+  // Support `bash -lc "..."`
+  if (
+    normalizedArgs.length === 3 &&
+    normalizedArgs[0] === 'bash' &&
+    (normalizedArgs[1] === '-lc' || normalizedArgs[1] === '-c')
+  ) {
+    try {
+      const script = normalizedArgs[2];
+
+      // Basic check for dangerous operators that could spawn subshells or redirect output
+      // We allow &&, ||, |, ; but explicitly block subshells () and redirection >, >>, <
+      if (/[()<>]/g.test(script)) {
+        return false;
+      }
+
+      const commands = script.split(/&&|\|\||\||;/);
+
+      let allSafe = true;
+      for (const cmd of commands) {
+        const trimmed = cmd.trim();
+        if (!trimmed) continue;
+
+        const parsed = shellParse(trimmed).map(String);
+        if (parsed.length === 0) continue;
+
+        if (!isSafeToCallWithExec(parsed)) {
+          allSafe = false;
+          break;
+        }
+      }
+
+      if (allSafe && commands.length > 0) {
+        return true;
+      }
+    } catch {
+      return false;
+    }
+  }
+
+  return false;
+}
+
+/**
+ * Core validation logic that checks a single command and its arguments
+ * against an allowlist of known safe operations. It performs deep validation
+ * for specific tools like `base64`, `find`, `rg`, `git`, and `sed` to ensure
+ * unsafe flags (like `--output`, `-exec`, or mutating options) are not used.
+ *
+ * @param args - The command and its arguments.
+ * @returns true if the command is strictly read-only and safe.
+ */
+function isSafeToCallWithExec(args: string[]): boolean {
+  if (!args || args.length === 0) return false;
+  const cmd = args[0];
+
+  const safeCommands = new Set([
+    'cat',
+    'cd',
+    'cut',
+    'echo',
+    'expr',
+    'false',
+    'grep',
+    'head',
+    'id',
+    'ls',
+    'nl',
+    'paste',
+    'pwd',
+    'rev',
+    'seq',
+    'stat',
+    'tail',
+    'tr',
+    'true',
+    'uname',
+    'uniq',
+    'wc',
+    'which',
+    'whoami',
+    'numfmt',
+    'tac',
+  ]);
+
+  if (safeCommands.has(cmd)) {
+    return true;
+  }
+
+  if (cmd === 'base64') {
+    const unsafeOptions = new Set(['-o', '--output']);
+    return !args
+      .slice(1)
+      .some(
+        (arg) =>
+          unsafeOptions.has(arg) ||
+          arg.startsWith('--output=') ||
+          (arg.startsWith('-o') && arg !== '-o'),
+      );
+  }
+
+  if (cmd === 'find') {
+    const unsafeOptions = new Set([
+      '-exec',
+      '-execdir',
+      '-ok',
+      '-okdir',
+      '-delete',
+      '-fls',
+      '-fprint',
+      '-fprint0',
+      '-fprintf',
+    ]);
+    return !args.some((arg) => unsafeOptions.has(arg));
+  }
+
+  if (cmd === 'rg') {
+    const unsafeWithArgs = new Set(['--pre', '--hostname-bin']);
+    const unsafeWithoutArgs = new Set(['--search-zip', '-z']);
+
+    return !args.some((arg) => {
+      if (unsafeWithoutArgs.has(arg)) return true;
+      for (const opt of unsafeWithArgs) {
+        if (arg === opt || arg.startsWith(opt + '=')) return true;
+      }
+      return false;
+    });
+  }
+
+  if (cmd === 'git') {
+    if (gitHasConfigOverrideGlobalOption(args)) {
+      return false;
+    }
+
+    const { idx, subcommand } = findGitSubcommand(args, [
+      'status',
+      'log',
+      'diff',
+      'show',
+      'branch',
+    ]);
+    if (!subcommand) {
+      return false;
+    }
+
+    const subcommandArgs = args.slice(idx + 1);
+
+    if (['status', 'log', 'diff', 'show'].includes(subcommand)) {
+      return gitSubcommandArgsAreReadOnly(subcommandArgs);
+    }
+
+    if (subcommand === 'branch') {
+      return (
+        gitSubcommandArgsAreReadOnly(subcommandArgs) &&
+        gitBranchIsReadOnly(subcommandArgs)
+      );
+    }
+
+    return false;
+  }
+
+  if (cmd === 'sed') {
+    // Special-case sed -n {N|M,N}p
+    if (args.length <= 4 && args[1] === '-n' && isValidSedNArg(args[2])) {
+      return true;
+    }
+    return false;
+  }
+
+  return false;
+}
+
+/**
+ * Helper to identify which git subcommand is being executed, skipping over
+ * global git options like `-c` or `--git-dir`.
+ *
+ * @param args - The full git command arguments.
+ * @param subcommands - A list of subcommands to look for.
+ * @returns An object containing the index of the subcommand and its name.
+ */
+function findGitSubcommand(
+  args: string[],
+  subcommands: string[],
+): { idx: number; subcommand: string | null } {
+  let skipNext = false;
+
+  for (let idx = 1; idx < args.length; idx++) {
+    if (skipNext) {
+      skipNext = false;
+      continue;
+    }
+
+    const arg = args[idx];
+
+    if (
+      arg.startsWith('--config-env=') ||
+      arg.startsWith('--exec-path=') ||
+      arg.startsWith('--git-dir=') ||
+      arg.startsWith('--namespace=') ||
+      arg.startsWith('--super-prefix=') ||
+      arg.startsWith('--work-tree=') ||
+      ((arg.startsWith('-C') || arg.startsWith('-c')) && arg.length > 2)
+    ) {
+      continue;
+    }
+
+    if (
+      arg === '-C' ||
+      arg === '-c' ||
+      arg === '--config-env' ||
+      arg === '--exec-path' ||
+      arg === '--git-dir' ||
+      arg === '--namespace' ||
+      arg === '--super-prefix' ||
+      arg === '--work-tree'
+    ) {
+      skipNext = true;
+      continue;
+    }
+
+    if (arg === '--' || arg.startsWith('-')) {
+      continue;
+    }
+
+    if (subcommands.includes(arg)) {
+      return { idx, subcommand: arg };
+    }
+
+    return { idx: -1, subcommand: null };
+  }
+
+  return { idx: -1, subcommand: null };
+}
+
+/**
+ * Checks if a git command contains global configuration override flags
+ * (e.g., `-c` or `--config-env`) which could be used maliciously to
+ * execute arbitrary code via git config.
+ *
+ * @param args - The git command arguments.
+ * @returns true if config overrides are present.
+ */
+function gitHasConfigOverrideGlobalOption(args: string[]): boolean {
+  return args.some(
+    (arg) =>
+      arg === '-c' ||
+      arg === '--config-env' ||
+      (arg.startsWith('-c') && arg.length > 2) ||
+      arg.startsWith('--config-env='),
+  );
+}
+
+/**
+ * Validates that the arguments for safe git subcommands (like `status`, `log`,
+ * `diff`, `show`) do not contain flags that could cause mutations or execute
+ * arbitrary commands (e.g., `--output`, `--exec`).
+ *
+ * @param args - Arguments passed to the git subcommand.
+ * @returns true if the arguments only represent read-only operations.
+ */
+function gitSubcommandArgsAreReadOnly(args: string[]): boolean {
+  const unsafeFlags = new Set([
+    '--output',
+    '--ext-diff',
+    '--textconv',
+    '--exec',
+    '--paginate',
+  ]);
+
+  return !args.some(
+    (arg) =>
+      unsafeFlags.has(arg) ||
+      arg.startsWith('--output=') ||
+      arg.startsWith('--exec='),
+  );
+}
+
+/**
+ * Validates that `git branch` is only used for read operations
+ * (e.g., listing branches) rather than creating, deleting, or renaming branches.
+ *
+ * @param args - Arguments passed to `git branch`.
+ * @returns true if it's purely a listing/read-only branch command.
+ */
+function gitBranchIsReadOnly(args: string[]): boolean {
+  if (args.length === 0) return true;
+
+  let sawReadOnlyFlag = false;
+  for (const arg of args) {
+    if (
+      [
+        '--list',
+        '-l',
+        '--show-current',
+        '-a',
+        '--all',
+        '-r',
+        '--remotes',
+        '-v',
+        '-vv',
+        '--verbose',
+      ].includes(arg)
+    ) {
+      sawReadOnlyFlag = true;
+    } else if (arg.startsWith('--format=')) {
+      sawReadOnlyFlag = true;
+    } else {
+      return false;
+    }
+  }
+  return sawReadOnlyFlag;
+}
+
+/**
+ * Ensures that a `sed` command argument is a valid line-printing instruction
+ * (e.g., `10p` or `5,10p`), preventing unsafe script execution in `sed`.
+ *
+ * @param arg - The script argument passed to `sed -n`.
+ * @returns true if it's a valid, safe print command.
+ */
+function isValidSedNArg(arg: string | undefined): boolean {
+  if (!arg) return false;
+
+  if (!arg.endsWith('p')) return false;
+  const core = arg.slice(0, -1);
+
+  const parts = core.split(',');
+  if (parts.length === 1) {
+    const num = parts[0];
+    return num.length > 0 && /^\d+$/.test(num);
+  } else if (parts.length === 2) {
+    const a = parts[0];
+    const b = parts[1];
+    return a.length > 0 && b.length > 0 && /^\d+$/.test(a) && /^\d+$/.test(b);
+  }
+
+  return false;
+}
+
+/**
+ * Checks if a command with its arguments is explicitly known to be dangerous
+ * and should be blocked or require strict user confirmation. This catches
+ * destructive commands like `rm -rf`, `sudo`, and commands with execution
+ * flags like `find -exec`.
+ *
+ * @param args - The command and its arguments.
+ * @returns true if the command is identified as dangerous, false otherwise.
+ */
+export function isDangerousCommand(args: string[]): boolean {
+  if (!args || args.length === 0) {
+    return false;
+  }
+
+  const cmd = args[0];
+
+  if (cmd === 'rm') {
+    return args[1] === '-f' || args[1] === '-rf' || args[1] === '-fr';
+  }
+
+  if (cmd === 'sudo') {
+    return isDangerousCommand(args.slice(1));
+  }
+
+  if (cmd === 'find') {
+    const unsafeOptions = new Set([
+      '-exec',
+      '-execdir',
+      '-ok',
+      '-okdir',
+      '-delete',
+      '-fls',
+      '-fprint',
+      '-fprint0',
+      '-fprintf',
+    ]);
+    return args.some((arg) => unsafeOptions.has(arg));
+  }
+
+  if (cmd === 'rg') {
+    const unsafeWithArgs = new Set(['--pre', '--hostname-bin']);
+    const unsafeWithoutArgs = new Set(['--search-zip', '-z']);
+
+    return args.some((arg) => {
+      if (unsafeWithoutArgs.has(arg)) return true;
+      for (const opt of unsafeWithArgs) {
+        if (arg === opt || arg.startsWith(opt + '=')) return true;
+      }
+      return false;
+    });
+  }
+
+  if (cmd === 'git') {
+    if (gitHasConfigOverrideGlobalOption(args)) {
+      return true;
+    }
+
+    const { idx, subcommand } = findGitSubcommand(args, [
+      'status',
+      'log',
+      'diff',
+      'show',
+      'branch',
+    ]);
+    if (!subcommand) {
+      // It's a git command we don't recognize as explicitly safe.
+      return false;
+    }
+
+    const subcommandArgs = args.slice(idx + 1);
+
+    if (['status', 'log', 'diff', 'show'].includes(subcommand)) {
+      return !gitSubcommandArgsAreReadOnly(subcommandArgs);
+    }
+
+    if (subcommand === 'branch') {
+      return !(
+        gitSubcommandArgsAreReadOnly(subcommandArgs) &&
+        gitBranchIsReadOnly(subcommandArgs)
+      );
+    }
+
+    return false;
+  }
+
+  if (cmd === 'base64') {
+    const unsafeOptions = new Set(['-o', '--output']);
+    return args
+      .slice(1)
+      .some(
+        (arg) =>
+          unsafeOptions.has(arg) ||
+          arg.startsWith('--output=') ||
+          (arg.startsWith('-o') && arg !== '-o'),
+      );
+  }
+
+  return false;
+}

packages/core/src/sandbox/macos/seatbeltArgsBuilder.test.ts

@@ -0,0 +1,160 @@
+/**
+ * @license
+ * Copyright 2026 Google LLC
+ * SPDX-License-Identifier: Apache-2.0
+ */
+import { describe, it, expect, vi } from 'vitest';
+import { buildSeatbeltArgs } from './seatbeltArgsBuilder.js';
+import fs from 'node:fs';
+import os from 'node:os';
+
+describe('seatbeltArgsBuilder', () => {
+  it('should build a strict allowlist profile allowing the workspace via param', () => {
+    // Mock realpathSync to just return the path for testing
+    vi.spyOn(fs, 'realpathSync').mockImplementation((p) => p as string);
+
+    const args = buildSeatbeltArgs({ workspace: '/Users/test/workspace' });
+
+    expect(args[0]).toBe('-p');
+    const profile = args[1];
+    expect(profile).toContain('(version 1)');
+    expect(profile).toContain('(deny default)');
+    expect(profile).toContain('(allow process-exec)');
+    expect(profile).toContain('(subpath (param "WORKSPACE"))');
+    expect(profile).not.toContain('(allow network*)');
+
+    expect(args).toContain('-D');
+    expect(args).toContain('WORKSPACE=/Users/test/workspace');
+    expect(args).toContain(`TMPDIR=${os.tmpdir()}`);
+
+    vi.restoreAllMocks();
+  });
+
+  it('should allow network when networkAccess is true', () => {
+    const args = buildSeatbeltArgs({ workspace: '/test', networkAccess: true });
+    const profile = args[1];
+    expect(profile).toContain('(allow network-outbound)');
+  });
+
+  it('should parameterize allowed paths and normalize them', () => {
+    vi.spyOn(fs, 'realpathSync').mockImplementation((p) => {
+      if (p === '/test/symlink') return '/test/real_path';
+      return p as string;
+    });
+
+    const args = buildSeatbeltArgs({
+      workspace: '/test',
+      allowedPaths: ['/custom/path1', '/test/symlink'],
+    });
+
+    const profile = args[1];
+    expect(profile).toContain('(subpath (param "ALLOWED_PATH_0"))');
+    expect(profile).toContain('(subpath (param "ALLOWED_PATH_1"))');
+
+    expect(args).toContain('-D');
+    expect(args).toContain('ALLOWED_PATH_0=/custom/path1');
+    expect(args).toContain('ALLOWED_PATH_1=/test/real_path');
+
+    vi.restoreAllMocks();
+  });
+
+  it('should resolve parent directories if a file does not exist', () => {
+    vi.spyOn(fs, 'realpathSync').mockImplementation((p) => {
+      if (p === '/test/symlink/nonexistent.txt') {
+        const error = new Error('ENOENT');
+        Object.assign(error, { code: 'ENOENT' });
+        throw error;
+      }
+      if (p === '/test/symlink') {
+        return '/test/real_path';
+      }
+      return p as string;
+    });
+
+    const args = buildSeatbeltArgs({
+      workspace: '/test/symlink/nonexistent.txt',
+    });
+
+    expect(args).toContain('WORKSPACE=/test/real_path/nonexistent.txt');
+    vi.restoreAllMocks();
+  });
+
+  it('should throw if realpathSync throws a non-ENOENT error', () => {
+    vi.spyOn(fs, 'realpathSync').mockImplementation(() => {
+      const error = new Error('Permission denied');
+      Object.assign(error, { code: 'EACCES' });
+      throw error;
+    });
+
+    expect(() =>
+      buildSeatbeltArgs({
+        workspace: '/test/workspace',
+      }),
+    ).toThrow('Permission denied');
+
+    vi.restoreAllMocks();
+  });
+
+  describe('governance files', () => {
+    it('should inject explicit deny rules for governance files', () => {
+      vi.spyOn(fs, 'realpathSync').mockImplementation((p) => p.toString());
+      vi.spyOn(fs, 'existsSync').mockReturnValue(true);
+      vi.spyOn(fs, 'lstatSync').mockImplementation(
+        (p) =>
+          ({
+            isDirectory: () => p.toString().endsWith('.git'),
+            isFile: () => !p.toString().endsWith('.git'),
+          }) as unknown as fs.Stats,
+      );
+
+      const args = buildSeatbeltArgs({ workspace: '/Users/test/workspace' });
+      const profile = args[1];
+
+      // .gitignore should be a literal deny
+      expect(args).toContain('-D');
+      expect(args).toContain(
+        'GOVERNANCE_FILE_0=/Users/test/workspace/.gitignore',
+      );
+      expect(profile).toContain(
+        '(deny file-write* (literal (param "GOVERNANCE_FILE_0")))',
+      );
+
+      // .git should be a subpath deny
+      expect(args).toContain('GOVERNANCE_FILE_2=/Users/test/workspace/.git');
+      expect(profile).toContain(
+        '(deny file-write* (subpath (param "GOVERNANCE_FILE_2")))',
+      );
+
+      vi.restoreAllMocks();
+    });
+
+    it('should protect both the symlink and the real path if they differ', () => {
+      vi.spyOn(fs, 'realpathSync').mockImplementation((p) => {
+        if (p === '/test/workspace/.gitignore') return '/test/real/.gitignore';
+        return p.toString();
+      });
+      vi.spyOn(fs, 'existsSync').mockReturnValue(true);
+      vi.spyOn(fs, 'lstatSync').mockImplementation(
+        () =>
+          ({
+            isDirectory: () => false,
+            isFile: () => true,
+          }) as unknown as fs.Stats,
+      );
+
+      const args = buildSeatbeltArgs({ workspace: '/test/workspace' });
+      const profile = args[1];
+
+      expect(args).toContain('GOVERNANCE_FILE_0=/test/workspace/.gitignore');
+      expect(args).toContain('REAL_GOVERNANCE_FILE_0=/test/real/.gitignore');
+      expect(profile).toContain(
+        '(deny file-write* (literal (param "GOVERNANCE_FILE_0")))',
+      );
+      expect(profile).toContain(
+        '(deny file-write* (literal (param "REAL_GOVERNANCE_FILE_0")))',
+      );
+
+      vi.restoreAllMocks();
+    });
+  });
+});

packages/core/src/sandbox/macos/seatbeltArgsBuilder.ts

@@ -0,0 +1,247 @@
+/**
+ * @license
+ * Copyright 2026 Google LLC
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+import {
+  BASE_SEATBELT_PROFILE,
+  NETWORK_SEATBELT_PROFILE,
+} from './baseProfile.js';
+import {
+  type SandboxPermissions,
+  sanitizePaths,
+  GOVERNANCE_FILES,
+} from '../../services/sandboxManager.js';
+
+/**
+ * Options for building macOS Seatbelt arguments.
+ */
+export interface SeatbeltArgsOptions {
+  /** The primary workspace path to allow access to. */
+  workspace: string;
+  /** Additional paths to allow access to. */
+  allowedPaths?: string[];
+  /** Absolute paths to explicitly deny read/write access to (overrides allowlists). */
+  forbiddenPaths?: string[];
+  /** Whether to allow network access. */
+  networkAccess?: boolean;
+  /** Granular additional permissions. */
+  additionalPermissions?: SandboxPermissions;
+  /** Whether to allow write access to the workspace. */
+  workspaceWrite?: boolean;
+}
+
+/**
+ * Resolves symlinks for a given path to prevent sandbox escapes.
+ * If a file does not exist (ENOENT), it recursively resolves the parent directory.
+ * Other errors (e.g. EACCES) are re-thrown.
+ */
+function tryRealpath(p: string): string {
+  try {
+    return fs.realpathSync(p);
+  } catch (e) {
+    if (e instanceof Error && 'code' in e && e.code === 'ENOENT') {
+      const parentDir = path.dirname(p);
+      if (parentDir === p) {
+        return p;
+      }
+      return path.join(tryRealpath(parentDir), path.basename(p));
+    }
+    throw e;
+  }
+}
+
+/**
+ * Builds the arguments array for sandbox-exec using a strict allowlist profile.
+ * It relies on parameters passed to sandbox-exec via the -D flag to avoid
+ * string interpolation vulnerabilities, and normalizes paths against symlink escapes.
+ *
+ * Returns arguments up to the end of sandbox-exec configuration (e.g. ['-p', '<profile>', '-D', ...])
+ * Does not include the final '--' separator or the command to run.
+ */
+export function buildSeatbeltArgs(options: SeatbeltArgsOptions): string[] {
+  let profile = BASE_SEATBELT_PROFILE + '\n';
+  const args: string[] = [];
+
+  const workspacePath = tryRealpath(options.workspace);
+  args.push('-D', `WORKSPACE=${workspacePath}`);
+  args.push('-D', `WORKSPACE_RAW=${options.workspace}`);
+  profile += `(allow file-read* (subpath (param "WORKSPACE_RAW")))\n`;
+  if (options.workspaceWrite) {
+    profile += `(allow file-write* (subpath (param "WORKSPACE_RAW")))\n`;
+  }
+
+  if (options.workspaceWrite) {
+    profile += `(allow file-write* (subpath (param "WORKSPACE")))\n`;
+  }
+
+  // Add explicit deny rules for governance files in the workspace.
+  // These are added after the workspace allow rule to ensure they take precedence
+  // (Seatbelt evaluates rules in order, later rules win for same path).
+  for (let i = 0; i < GOVERNANCE_FILES.length; i++) {
+    const governanceFile = path.join(workspacePath, GOVERNANCE_FILES[i].path);
+    const realGovernanceFile = tryRealpath(governanceFile);
+
+    // Determine if it should be treated as a directory (subpath) or a file (literal).
+    // .git is generally a directory, while ignore files are literals.
+    let isDirectory = GOVERNANCE_FILES[i].isDirectory;
+    try {
+      if (fs.existsSync(realGovernanceFile)) {
+        isDirectory = fs.lstatSync(realGovernanceFile).isDirectory();
+      }
+    } catch {
+      // Ignore errors, use default guess
+    }
+
+    const ruleType = isDirectory ? 'subpath' : 'literal';
+
+    args.push('-D', `GOVERNANCE_FILE_${i}=${governanceFile}`);
+    profile += `(deny file-write* (${ruleType} (param "GOVERNANCE_FILE_${i}")))\n`;
+
+    if (realGovernanceFile !== governanceFile) {
+      args.push('-D', `REAL_GOVERNANCE_FILE_${i}=${realGovernanceFile}`);
+      profile += `(deny file-write* (${ruleType} (param "REAL_GOVERNANCE_FILE_${i}")))\n`;
+    }
+  }
+
+  // Auto-detect and support git worktrees by granting read and write access to the underlying git directory
+  try {
+    const gitPath = path.join(workspacePath, '.git');
+    const gitStat = fs.lstatSync(gitPath);
+    if (gitStat.isFile()) {
+      const gitContent = fs.readFileSync(gitPath, 'utf8');
+      const match = gitContent.match(/^gitdir:\s*(.+)$/m);
+      if (match && match[1]) {
+        let worktreeGitDir = match[1].trim();
+        if (!path.isAbsolute(worktreeGitDir)) {
+          worktreeGitDir = path.resolve(workspacePath, worktreeGitDir);
+        }
+        const resolvedWorktreeGitDir = tryRealpath(worktreeGitDir);
+
+        // Grant write access to the worktree's specific .git directory
+        args.push('-D', `WORKTREE_GIT_DIR=${resolvedWorktreeGitDir}`);
+        profile += `(allow file-read* file-write* (subpath (param "WORKTREE_GIT_DIR")))\n`;
+
+        // Grant write access to the main repository's .git directory (objects, refs, etc. are shared)
+        // resolvedWorktreeGitDir is usually like: /path/to/main-repo/.git/worktrees/worktree-name
+        const mainGitDir = tryRealpath(
+          path.dirname(path.dirname(resolvedWorktreeGitDir)),
+        );
+        if (mainGitDir && mainGitDir.endsWith('.git')) {
+          args.push('-D', `MAIN_GIT_DIR=${mainGitDir}`);
+          profile += `(allow file-read* file-write* (subpath (param "MAIN_GIT_DIR")))\n`;
+        }
+      }
+    }
+  } catch (_e) {
+    // Ignore if .git doesn't exist, isn't readable, etc.
+  }
+
+  const tmpPath = tryRealpath(os.tmpdir());
+  args.push('-D', `TMPDIR=${tmpPath}`);
+
+  const nodeRootPath = tryRealpath(
+    path.dirname(path.dirname(process.execPath)),
+  );
+  args.push('-D', `NODE_ROOT=${nodeRootPath}`);
+  profile += `(allow file-read* (subpath (param "NODE_ROOT")))\n`;
+
+  // Add PATH directories as read-only to support nvm, homebrew, etc.
+  if (process.env['PATH']) {
+    const paths = process.env['PATH'].split(':');
+    let pathIndex = 0;
+    const addedPaths = new Set();
+
+    for (const p of paths) {
+      if (!p.trim()) continue;
+      try {
+        let resolved = tryRealpath(p);
+
+        // If this is a 'bin' directory (like /usr/local/bin or homebrew/bin),
+        // also grant read access to its parent directory so that symlinked
+        // assets (like Cellar or libexec) can be read.
+        if (resolved.endsWith('/bin')) {
+          resolved = path.dirname(resolved);
+        }
+
+        if (!addedPaths.has(resolved)) {
+          addedPaths.add(resolved);
+          args.push('-D', `SYS_PATH_${pathIndex}=${resolved}`);
+          profile += `(allow file-read* (subpath (param "SYS_PATH_${pathIndex}")))\n`;
+          pathIndex++;
+        }
+      } catch (_e) {
+        // Ignore paths that do not exist or are inaccessible
+      }
+    }
+  }
+
+  // Handle allowedPaths
+  const allowedPaths = sanitizePaths(options.allowedPaths) || [];
+  for (let i = 0; i < allowedPaths.length; i++) {
+    const allowedPath = tryRealpath(allowedPaths[i]);
+    args.push('-D', `ALLOWED_PATH_${i}=${allowedPath}`);
+    profile += `(allow file-read* file-write* (subpath (param "ALLOWED_PATH_${i}")))\n`;
+  }
+
+  // Handle granular additional permissions
+  if (options.additionalPermissions?.fileSystem) {
+    const { read, write } = options.additionalPermissions.fileSystem;
+    if (read) {
+      read.forEach((p, i) => {
+        const resolved = tryRealpath(p);
+        const paramName = `ADDITIONAL_READ_${i}`;
+        args.push('-D', `${paramName}=${resolved}`);
+        let isFile = false;
+        try {
+          isFile = fs.statSync(resolved).isFile();
+        } catch {
+          // Ignore error
+        }
+        if (isFile) {
+          profile += `(allow file-read* (literal (param "${paramName}")))\n`;
+        } else {
+          profile += `(allow file-read* (subpath (param "${paramName}")))\n`;
+        }
+      });
+    }
+    if (write) {
+      write.forEach((p, i) => {
+        const resolved = tryRealpath(p);
+        const paramName = `ADDITIONAL_WRITE_${i}`;
+        args.push('-D', `${paramName}=${resolved}`);
+        let isFile = false;
+        try {
+          isFile = fs.statSync(resolved).isFile();
+        } catch {
+          // Ignore error
+        }
+        if (isFile) {
+          profile += `(allow file-read* file-write* (literal (param "${paramName}")))\n`;
+        } else {
+          profile += `(allow file-read* file-write* (subpath (param "${paramName}")))\n`;
+        }
+      });
+    }
+  }
+
+  // Handle forbiddenPaths
+  const forbiddenPaths = sanitizePaths(options.forbiddenPaths) || [];
+  for (let i = 0; i < forbiddenPaths.length; i++) {
+    const forbiddenPath = tryRealpath(forbiddenPaths[i]);
+    args.push('-D', `FORBIDDEN_PATH_${i}=${forbiddenPath}`);
+    profile += `(deny file-read* file-write* (subpath (param "FORBIDDEN_PATH_${i}")))\n`;
+  }
+
+  if (options.networkAccess || options.additionalPermissions?.network) {
+    profile += NETWORK_SEATBELT_PROFILE;
+  }
+
+  args.unshift('-p', profile);
+
+  return args;
+}

packages/core/src/sandbox/windows/WindowsSandboxManager.test.ts

@@ -8,11 +8,11 @@ import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
 import fs from 'node:fs';
 import os from 'node:os';
 import path from 'node:path';
-import { WindowsSandboxManager } from './windowsSandboxManager.js';
-import type { SandboxRequest } from './sandboxManager.js';
-import { spawnAsync } from '../utils/shell-utils.js';
+import { WindowsSandboxManager } from './WindowsSandboxManager.js';
+import type { SandboxRequest } from '../../services/sandboxManager.js';
+import { spawnAsync } from '../../utils/shell-utils.js';
 
-vi.mock('../utils/shell-utils.js', () => ({
+vi.mock('../../utils/shell-utils.js', () => ({
   spawnAsync: vi.fn(),
 }));
 

packages/core/src/sandbox/windows/WindowsSandboxManager.ts

@@ -15,13 +15,13 @@ import {
   GOVERNANCE_FILES,
   type GlobalSandboxOptions,
   sanitizePaths,
-} from './sandboxManager.js';
+} from '../../services/sandboxManager.js';
 import {
   sanitizeEnvironment,
   getSecureSanitizationConfig,
-} from './environmentSanitization.js';
-import { debugLogger } from '../utils/debugLogger.js';
-import { spawnAsync } from '../utils/shell-utils.js';
+} from '../../services/environmentSanitization.js';
+import { debugLogger } from '../../utils/debugLogger.js';
+import { spawnAsync } from '../../utils/shell-utils.js';
 
 const __filename = fileURLToPath(import.meta.url);
 const __dirname = path.dirname(__filename);
@@ -37,7 +37,7 @@ export class WindowsSandboxManager implements SandboxManager {
   private readonly lowIntegrityCache = new Set<string>();
 
   constructor(private readonly options: GlobalSandboxOptions) {
-    this.helperPath = path.resolve(__dirname, 'scripts', 'GeminiSandbox.exe');
+    this.helperPath = path.resolve(__dirname, 'GeminiSandbox.exe');
   }
 
   /**

packages/core/src/scheduler/policy.ts

@@ -77,7 +77,8 @@ export async function checkPolicy(
   // confirmation prompt if the policy engine's decision is 'ASK_USER'.
   if (
     decision === PolicyDecision.ASK_USER &&
-    toolCall.request.isClientInitiated
+    toolCall.request.isClientInitiated &&
+    !toolCall.request.args?.['additional_permissions']
   ) {
     return {
       decision: PolicyDecision.ALLOW,

packages/core/src/scheduler/scheduler.ts

@@ -792,6 +792,110 @@ export class Scheduler {
       return true;
     }
 
+    let isSandboxError = false;
+    let sandboxDetailsStr = '';
+
+    if (
+      result.status === CoreToolCallStatus.Error &&
+      result.response.errorType === 'sandbox_expansion_required'
+    ) {
+      isSandboxError = true;
+      sandboxDetailsStr = result.response.error?.message || '';
+    }
+
+    if (isSandboxError) {
+      try {
+        // eslint-disable-next-line @typescript-eslint/no-unsafe-type-assertion
+        const parsedError = JSON.parse(sandboxDetailsStr) as {
+          rootCommand: string;
+          additionalPermissions: import('../services/sandboxManager.js').SandboxPermissions;
+        };
+
+        const confirmationDetails: SerializableConfirmationDetails = {
+          type: 'sandbox_expansion',
+          title: 'Sandbox Expansion Request',
+          command: String(
+            activeCall.request.args['command'] ?? parsedError.rootCommand,
+          ),
+          rootCommand: parsedError.rootCommand,
+          additionalPermissions: parsedError.additionalPermissions,
+        };
+
+        const correlationId = crypto.randomUUID();
+
+        // Mutate the active call so resolveConfirmation generates the correct Sandbox Expansion details
+        activeCall.request.args['additional_permissions'] =
+          parsedError.additionalPermissions;
+        activeCall.invocation = activeCall.tool.build(activeCall.request.args);
+
+        // CRITICAL: We must push the new args and invocation into the state manager
+        // before calling resolveConfirmation, because resolveConfirmation fetches
+        // the tool call directly from the state manager!
+        this.state.updateArgs(
+          callId,
+          activeCall.request.args,
+          activeCall.invocation,
+        );
+
+        this.state.updateStatus(callId, CoreToolCallStatus.AwaitingApproval, {
+          confirmationDetails,
+          correlationId,
+        });
+
+        const validatingCall = {
+          ...activeCall,
+          status: CoreToolCallStatus.Validating,
+        } as ValidatingToolCall;
+
+        const confResult = await resolveConfirmation(validatingCall, signal, {
+          config: this.config,
+          messageBus: this.messageBus,
+          state: this.state,
+          modifier: this.modifier,
+          getPreferredEditor: this.getPreferredEditor,
+          schedulerId: this.schedulerId,
+          onWaitingForConfirmation: this.onWaitingForConfirmation,
+        });
+
+        if (confResult.outcome === ToolConfirmationOutcome.Cancel) {
+          type LegacyHack = ToolCallResponseInfo & {
+            llmContent?: string;
+            returnDisplay?: string;
+          };
+          const errorResponse = { ...result.response } as LegacyHack;
+          errorResponse.llmContent =
+            'User cancelled sandbox expansion. The command failed with a sandbox denial. Shell output:\n' +
+            String(errorResponse.returnDisplay);
+
+          this.state.updateStatus(
+            callId,
+            CoreToolCallStatus.Error,
+            errorResponse,
+          );
+          return false;
+        }
+
+        activeCall.request.args['additional_permissions'] =
+          parsedError.additionalPermissions;
+
+        // Reset the output stream visual so it replaces the error text
+        this.state.updateStatus(callId, CoreToolCallStatus.Executing, {
+          liveOutput: undefined,
+        });
+
+        // Call _execute synchronously and properly return its promise to loop internally!
+        return await this._execute(
+          {
+            ...activeCall,
+            status: CoreToolCallStatus.Scheduled,
+          } as ScheduledToolCall,
+          signal,
+        );
+      } catch (_e) {
+        // Fallback to normal error handling if parsing/looping fails
+      }
+    }
+
     if (result.status === CoreToolCallStatus.Success) {
       this.state.updateStatus(
         callId,

packages/core/src/services/sandboxManager.test.ts

@@ -10,7 +10,7 @@ import { NoopSandboxManager, sanitizePaths } from './sandboxManager.js';
 import { createSandboxManager } from './sandboxManagerFactory.js';
 import { LinuxSandboxManager } from '../sandbox/linux/LinuxSandboxManager.js';
 import { MacOsSandboxManager } from '../sandbox/macos/MacOsSandboxManager.js';
-import { WindowsSandboxManager } from './windowsSandboxManager.js';
+import { WindowsSandboxManager } from '../sandbox/windows/WindowsSandboxManager.js';
 
 describe('sanitizePaths', () => {
   it('should return undefined if no paths are provided', () => {

packages/core/src/services/sandboxManager.ts

@@ -11,6 +11,18 @@ import {
   getSecureSanitizationConfig,
   type EnvironmentSanitizationConfig,
 } from './environmentSanitization.js';
+export interface SandboxPermissions {
+  /** Filesystem permissions. */
+  fileSystem?: {
+    /** Paths that should be readable by the command. */
+    read?: string[];
+    /** Paths that should be writable by the command. */
+    write?: string[];
+  };
+  /** Whether the command should have network access. */
+  network?: boolean;
+}
+
 /**
  * Security boundaries and permissions applied to a specific sandboxed execution.
  */
@@ -23,6 +35,8 @@ export interface ExecutionPolicy {
   networkAccess?: boolean;
   /** Rules for scrubbing sensitive environment variables. */
   sanitizationConfig?: Partial<EnvironmentSanitizationConfig>;
+  /** Additional granular permissions to grant to this command. */
+  additionalPermissions?: SandboxPermissions;
 }
 
 /**

packages/core/src/services/sandboxManagerFactory.ts

@@ -12,8 +12,9 @@ import {
 } from './sandboxManager.js';
 import { LinuxSandboxManager } from '../sandbox/linux/LinuxSandboxManager.js';
 import { MacOsSandboxManager } from '../sandbox/macos/MacOsSandboxManager.js';
-import { WindowsSandboxManager } from './windowsSandboxManager.js';
+import { WindowsSandboxManager } from '../sandbox/windows/WindowsSandboxManager.js';
 import type { SandboxConfig } from '../config/config.js';
+import { type SandboxPolicyManager } from '../policy/sandboxPolicyManager.js';
 
 /**
  * Creates a sandbox manager based on the provided settings.
@@ -21,7 +22,13 @@ import type { SandboxConfig } from '../config/config.js';
 export function createSandboxManager(
   sandbox: SandboxConfig | undefined,
   workspace: string,
+  policyManager?: SandboxPolicyManager,
+  approvalMode?: string,
 ): SandboxManager {
+  if (approvalMode === 'yolo') {
+    return new NoopSandboxManager();
+  }
+
   const isWindows = os.platform() === 'win32';
 
   if (
@@ -36,7 +43,15 @@ export function createSandboxManager(
       return new LinuxSandboxManager({ workspace });
     }
     if (os.platform() === 'darwin') {
-      return new MacOsSandboxManager({ workspace });
+      const modeConfig =
+        policyManager && approvalMode
+          ? policyManager.getModeConfig(approvalMode)
+          : undefined;
+      return new MacOsSandboxManager({
+        workspace,
+        modeConfig,
+        policyManager,
+      });
     }
     return new LocalSandboxManager();
   }

packages/core/src/services/shellExecutionService.ts

@@ -31,7 +31,11 @@ import {
   sanitizeEnvironment,
   type EnvironmentSanitizationConfig,
 } from './environmentSanitization.js';
-import { NoopSandboxManager, type SandboxManager } from './sandboxManager.js';
+import {
+  NoopSandboxManager,
+  type SandboxManager,
+  type SandboxPermissions,
+} from './sandboxManager.js';
 import type { SandboxConfig } from '../config/config.js';
 import { killProcessGroup } from '../utils/process-utils.js';
 import {
@@ -84,6 +88,7 @@ export type ShellExecutionResult = ExecutionResult;
 export type ShellExecutionHandle = ExecutionHandle;
 
 export interface ShellExecutionConfig {
+  additionalPermissions?: SandboxPermissions;
   terminalWidth?: number;
   terminalHeight?: number;
   pager?: string;
@@ -441,6 +446,7 @@ export class ShellExecutionService {
         ...shellExecutionConfig,
         ...(shellExecutionConfig.sandboxConfig || {}),
         sanitizationConfig,
+        additionalPermissions: shellExecutionConfig.additionalPermissions,
       },
     });
 

packages/core/src/telemetry/types.ts

@@ -2184,7 +2184,8 @@ export class ApprovalModeSwitchEvent implements BaseTelemetryEvent {
   toOpenTelemetryAttributes(config: Config): LogAttributes {
     return {
       ...getCommonAttributes(config),
-      event_name: EVENT_APPROVAL_MODE_SWITCH,
+      'event.name': EVENT_APPROVAL_MODE_SWITCH,
+      'event.timestamp': this['event.timestamp'],
       from_mode: this.from_mode,
       to_mode: this.to_mode,
     };
@@ -2214,7 +2215,8 @@ export class ApprovalModeDurationEvent implements BaseTelemetryEvent {
   toOpenTelemetryAttributes(config: Config): LogAttributes {
     return {
       ...getCommonAttributes(config),
-      event_name: EVENT_APPROVAL_MODE_DURATION,
+      'event.name': EVENT_APPROVAL_MODE_DURATION,
+      'event.timestamp': this['event.timestamp'],
       mode: this.mode,
       duration_ms: this.duration_ms,
     };

packages/core/src/tools/confirmation-policy.test.ts

@@ -71,6 +71,7 @@ describe('Tool Confirmation Policy Updates', () => {
       getDisableLLMCorrection: () => true,
       getIdeMode: () => false,
       getActiveModel: () => 'test-model',
+      isPlanMode: () => false,
       getWorkspaceContext: () => ({
         isPathWithinWorkspace: () => true,
         getDirectories: () => [rootDir],

packages/core/src/tools/definitions/__snapshots__/coreToolsModelSnapshots.test.ts.snap

@@ -169,13 +169,13 @@ exports[`coreTools snapshots for specific models > Model: gemini-2.5-pro > snaps
   "name": "exit_plan_mode",
   "parametersJsonSchema": {
     "properties": {
-      "plan_path": {
-        "description": "The file path to the finalized plan (e.g., "/mock/plans/feature-x.md"). This path MUST be within the designated plans directory: /mock/plans/",
+      "plan_filename": {
+        "description": "The filename of the finalized plan (e.g., "feature-x.md"). Do not provide an absolute path.",
         "type": "string",
       },
     },
     "required": [
-      "plan_path",
+      "plan_filename",
     ],
     "type": "object",
   },
@@ -583,6 +583,35 @@ exports[`coreTools snapshots for specific models > Model: gemini-2.5-pro > snaps
   "name": "run_shell_command",
   "parametersJsonSchema": {
     "properties": {
+      "additional_permissions": {
+        "description": "Sandbox permissions for the command. Use this to request additional sandboxed filesystem or network permissions if a previous command failed with "Operation not permitted".",
+        "properties": {
+          "fileSystem": {
+            "properties": {
+              "read": {
+                "description": "List of additional absolute paths to allow reading.",
+                "items": {
+                  "type": "string",
+                },
+                "type": "array",
+              },
+              "write": {
+                "description": "List of additional absolute paths to allow writing.",
+                "items": {
+                  "type": "string",
+                },
+                "type": "array",
+              },
+            },
+            "type": "object",
+          },
+          "network": {
+            "description": "Set to true to enable network access for this command.",
+            "type": "boolean",
+          },
+        },
+        "type": "object",
+      },
       "command": {
         "description": "Exact bash command to execute as \`bash -c <command>\`",
         "type": "string",
@@ -958,13 +987,13 @@ exports[`coreTools snapshots for specific models > Model: gemini-3-pro-preview >
   "name": "exit_plan_mode",
   "parametersJsonSchema": {
     "properties": {
-      "plan_path": {
-        "description": "The file path to the finalized plan (e.g., "/mock/plans/feature-x.md"). This path MUST be within the designated plans directory: /mock/plans/",
+      "plan_filename": {
+        "description": "The filename of the finalized plan (e.g., "feature-x.md"). Do not provide an absolute path.",
         "type": "string",
       },
     },
     "required": [
-      "plan_path",
+      "plan_filename",
     ],
     "type": "object",
   },
@@ -1348,6 +1377,35 @@ exports[`coreTools snapshots for specific models > Model: gemini-3-pro-preview >
   "name": "run_shell_command",
   "parametersJsonSchema": {
     "properties": {
+      "additional_permissions": {
+        "description": "Sandbox permissions for the command. Use this to request additional sandboxed filesystem or network permissions if a previous command failed with "Operation not permitted".",
+        "properties": {
+          "fileSystem": {
+            "properties": {
+              "read": {
+                "description": "List of additional absolute paths to allow reading.",
+                "items": {
+                  "type": "string",
+                },
+                "type": "array",
+              },
+              "write": {
+                "description": "List of additional absolute paths to allow writing.",
+                "items": {
+                  "type": "string",
+                },
+                "type": "array",
+              },
+            },
+            "type": "object",
+          },
+          "network": {
+            "description": "Set to true to enable network access for this command.",
+            "type": "boolean",
+          },
+        },
+        "type": "object",
+      },
       "command": {
         "description": "Exact bash command to execute as \`bash -c <command>\`",
         "type": "string",

packages/core/src/tools/definitions/base-declarations.ts

@@ -117,8 +117,11 @@ export const ASK_USER_OPTION_PARAM_DESCRIPTION = 'description';
 
 // -- exit_plan_mode --
 export const EXIT_PLAN_MODE_TOOL_NAME = 'exit_plan_mode';
-export const EXIT_PLAN_PARAM_PLAN_PATH = 'plan_path';
+export const EXIT_PLAN_PARAM_PLAN_FILENAME = 'plan_filename';
 
 // -- enter_plan_mode --
 export const ENTER_PLAN_MODE_TOOL_NAME = 'enter_plan_mode';
 export const PLAN_MODE_PARAM_REASON = 'reason';
+
+// -- sandbox --
+export const PARAM_ADDITIONAL_PERMISSIONS = 'additional_permissions';

packages/core/src/tools/definitions/coreTools.ts

@@ -89,7 +89,7 @@ export {
   ASK_USER_OPTION_PARAM_LABEL,
   ASK_USER_OPTION_PARAM_DESCRIPTION,
   PLAN_MODE_PARAM_REASON,
-  EXIT_PLAN_PARAM_PLAN_PATH,
+  EXIT_PLAN_PARAM_PLAN_FILENAME,
   SKILL_PARAM_NAME,
 } from './base-declarations.js';
 
@@ -244,10 +244,10 @@ export function getShellDefinition(
   };
 }
 
-export function getExitPlanModeDefinition(plansDir: string): ToolDefinition {
+export function getExitPlanModeDefinition(): ToolDefinition {
   return {
-    base: getExitPlanModeDeclaration(plansDir),
-    overrides: (modelId) => getToolSet(modelId).exit_plan_mode(plansDir),
+    base: getExitPlanModeDeclaration(),
+    overrides: (modelId) => getToolSet(modelId).exit_plan_mode(),
   };
 }
 

packages/core/src/tools/definitions/coreToolsModelSnapshots.test.ts

@@ -82,7 +82,7 @@ describe('coreTools snapshots for specific models', () => {
     { name: 'enter_plan_mode', definition: ENTER_PLAN_MODE_DEFINITION },
     {
       name: 'exit_plan_mode',
-      definition: getExitPlanModeDefinition('/mock/plans'),
+      definition: getExitPlanModeDefinition(),
     },
     {
       name: 'activate_skill',

packages/core/src/tools/definitions/dynamic-declaration-helpers.ts

@@ -21,8 +21,9 @@ import {
   PARAM_DESCRIPTION,
   PARAM_DIR_PATH,
   SHELL_PARAM_IS_BACKGROUND,
-  EXIT_PLAN_PARAM_PLAN_PATH,
+  EXIT_PLAN_PARAM_PLAN_FILENAME,
   SKILL_PARAM_NAME,
+  PARAM_ADDITIONAL_PERMISSIONS,
 } from './base-declarations.js';
 
 /**
@@ -109,6 +110,35 @@ export function getShellDeclaration(
           description:
             'Set to true if this command should be run in the background (e.g. for long-running servers or watchers). The command will be started, allowed to run for a brief moment to check for immediate errors, and then moved to the background.',
         },
+        [PARAM_ADDITIONAL_PERMISSIONS]: {
+          type: 'object',
+          description:
+            'Sandbox permissions for the command. Use this to request additional sandboxed filesystem or network permissions if a previous command failed with "Operation not permitted".',
+          properties: {
+            network: {
+              type: 'boolean',
+              description:
+                'Set to true to enable network access for this command.',
+            },
+            fileSystem: {
+              type: 'object',
+              properties: {
+                read: {
+                  type: 'array',
+                  items: { type: 'string' },
+                  description:
+                    'List of additional absolute paths to allow reading.',
+                },
+                write: {
+                  type: 'array',
+                  items: { type: 'string' },
+                  description:
+                    'List of additional absolute paths to allow writing.',
+                },
+              },
+            },
+          },
+        },
       },
       required: [SHELL_PARAM_COMMAND],
     },
@@ -118,20 +148,18 @@ export function getShellDeclaration(
 /**
  * Returns the FunctionDeclaration for exiting plan mode.
  */
-export function getExitPlanModeDeclaration(
-  plansDir: string,
-): FunctionDeclaration {
+export function getExitPlanModeDeclaration(): FunctionDeclaration {
   return {
     name: EXIT_PLAN_MODE_TOOL_NAME,
     description:
       'Finalizes the planning phase and transitions to implementation by presenting the plan for user approval. This tool MUST be used to exit Plan Mode before any source code edits can be performed. Call this whenever a plan is ready or the user requests implementation.',
     parametersJsonSchema: {
       type: 'object',
-      required: [EXIT_PLAN_PARAM_PLAN_PATH],
+      required: [EXIT_PLAN_PARAM_PLAN_FILENAME],
       properties: {
-        [EXIT_PLAN_PARAM_PLAN_PATH]: {
+        [EXIT_PLAN_PARAM_PLAN_FILENAME]: {
           type: 'string',
-          description: `The file path to the finalized plan (e.g., "${plansDir}/feature-x.md"). This path MUST be within the designated plans directory: ${plansDir}/`,
+          description: `The filename of the finalized plan (e.g., "feature-x.md"). Do not provide an absolute path.`,
         },
       },
     },

packages/core/src/tools/definitions/model-family-sets/default-legacy.ts

@@ -739,6 +739,6 @@ The agent did not use the todo list because this task could be completed by a ti
     },
   },
 
-  exit_plan_mode: (plansDir) => getExitPlanModeDeclaration(plansDir),
+  exit_plan_mode: () => getExitPlanModeDeclaration(),
   activate_skill: (skillNames) => getActivateSkillDeclaration(skillNames),
 };

packages/core/src/tools/definitions/model-family-sets/gemini-3.ts

@@ -714,6 +714,6 @@ The agent did not use the todo list because this task could be completed by a ti
     },
   },
 
-  exit_plan_mode: (plansDir) => getExitPlanModeDeclaration(plansDir),
+  exit_plan_mode: () => getExitPlanModeDeclaration(),
   activate_skill: (skillNames) => getActivateSkillDeclaration(skillNames),
 };

packages/core/src/tools/definitions/types.ts

@@ -47,6 +47,6 @@ export interface CoreToolSet {
   get_internal_docs: FunctionDeclaration;
   ask_user: FunctionDeclaration;
   enter_plan_mode: FunctionDeclaration;
-  exit_plan_mode: (plansDir: string) => FunctionDeclaration;
+  exit_plan_mode: () => FunctionDeclaration;
   activate_skill: (skillNames: string[]) => FunctionDeclaration;
 }

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

@@ -131,8 +131,10 @@ describe('EditTool', () => {
       isInteractive: () => false,
       getDisableLLMCorrection: vi.fn(() => true),
       getExperiments: () => {},
+      isPlanMode: vi.fn(() => false),
       storage: {
         getProjectTempDir: vi.fn().mockReturnValue('/tmp/project'),
+        getPlansDir: vi.fn().mockReturnValue('/tmp/plans'),
       },
       isPathAllowed(this: Config, absolutePath: string): boolean {
         const workspaceContext = this.getWorkspaceContext();
@@ -1299,4 +1301,42 @@ function doIt() {
       );
     });
   });
+
+  describe('plan mode', () => {
+    it('should allow edits to plans directory when isPlanMode is true', async () => {
+      const mockProjectTempDir = path.join(tempDir, 'project');
+      fs.mkdirSync(mockProjectTempDir);
+      vi.mocked(mockConfig.storage.getProjectTempDir).mockReturnValue(
+        mockProjectTempDir,
+      );
+
+      const plansDir = path.join(mockProjectTempDir, 'plans');
+      fs.mkdirSync(plansDir);
+
+      vi.mocked(mockConfig.isPlanMode).mockReturnValue(true);
+      vi.mocked(mockConfig.storage.getPlansDir).mockReturnValue(plansDir);
+
+      const filePath = path.join(rootDir, 'test-file.txt');
+      const planFilePath = path.join(plansDir, 'test-file.txt');
+      const initialContent = 'some initial content';
+      fs.writeFileSync(planFilePath, initialContent, 'utf8');
+
+      const params: EditToolParams = {
+        file_path: filePath,
+        instruction: 'Replace initial with new',
+        old_string: 'initial',
+        new_string: 'new',
+      };
+
+      const invocation = tool.build(params);
+      const result = await invocation.execute(new AbortController().signal);
+
+      expect(result.llmContent).toMatch(/Successfully modified file/);
+
+      // Verify plan file is written with new content
+      expect(fs.readFileSync(planFilePath, 'utf8')).toBe('some new content');
+
+      fs.rmSync(plansDir, { recursive: true, force: true });
+    });
+  });
 });

packages/core/src/tools/edit.ts

@@ -463,7 +463,13 @@ class EditToolInvocation
       true,
       () => this.config.getApprovalMode(),
     );
-    if (!path.isAbsolute(this.params.file_path)) {
+    if (this.config.isPlanMode()) {
+      const safeFilename = path.basename(this.params.file_path);
+      this.resolvedPath = path.join(
+        this.config.storage.getPlansDir(),
+        safeFilename,
+      );
+    } else if (!path.isAbsolute(this.params.file_path)) {
       const result = correctPath(this.params.file_path, this.config);
       if (result.success) {
         this.resolvedPath = result.correctedPath;

packages/core/src/tools/exit-plan-mode.test.ts

@@ -79,7 +79,7 @@ describe('ExitPlanModeTool', () => {
   describe('shouldConfirmExecute', () => {
     it('should return plan approval confirmation details when plan has content', async () => {
       const planRelativePath = createPlanFile('test-plan.md', '# My Plan');
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       const result = await invocation.shouldConfirmExecute(
         new AbortController().signal,
@@ -98,7 +98,7 @@ describe('ExitPlanModeTool', () => {
 
     it('should return false when plan file is empty', async () => {
       const planRelativePath = createPlanFile('empty.md', '   ');
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       const result = await invocation.shouldConfirmExecute(
         new AbortController().signal,
@@ -109,7 +109,7 @@ describe('ExitPlanModeTool', () => {
 
     it('should return false when plan file cannot be read', async () => {
       const planRelativePath = path.join('plans', 'non-existent.md');
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       const result = await invocation.shouldConfirmExecute(
         new AbortController().signal,
@@ -120,7 +120,7 @@ describe('ExitPlanModeTool', () => {
 
     it('should auto-approve when policy decision is ALLOW', async () => {
       const planRelativePath = createPlanFile('test.md', '# Content');
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       vi.spyOn(
         invocation as unknown as {
@@ -143,7 +143,7 @@ describe('ExitPlanModeTool', () => {
 
     it('should throw error when policy decision is DENY', async () => {
       const planRelativePath = createPlanFile('test.md', '# Content');
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       vi.spyOn(
         invocation as unknown as {
@@ -161,7 +161,7 @@ describe('ExitPlanModeTool', () => {
   describe('execute with invalid plan', () => {
     it('should return error when plan file is empty', async () => {
       const planRelativePath = createPlanFile('empty.md', '');
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       await invocation.shouldConfirmExecute(new AbortController().signal);
       const result = await invocation.execute(new AbortController().signal);
@@ -171,8 +171,8 @@ describe('ExitPlanModeTool', () => {
     });
 
     it('should return error when plan file cannot be read', async () => {
-      const planRelativePath = 'plans/ghost.md';
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const planRelativePath = 'ghost.md';
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       await invocation.shouldConfirmExecute(new AbortController().signal);
       const result = await invocation.execute(new AbortController().signal);
@@ -184,7 +184,7 @@ describe('ExitPlanModeTool', () => {
   describe('execute', () => {
     it('should return approval message when plan is approved with DEFAULT mode', async () => {
       const planRelativePath = createPlanFile('test.md', '# Content');
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       const confirmDetails = await invocation.shouldConfirmExecute(
         new AbortController().signal,
@@ -212,7 +212,7 @@ Read and follow the plan strictly during implementation.`,
 
     it('should return approval message when plan is approved with AUTO_EDIT mode', async () => {
       const planRelativePath = createPlanFile('test.md', '# Content');
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       const confirmDetails = await invocation.shouldConfirmExecute(
         new AbortController().signal,
@@ -243,7 +243,7 @@ Read and follow the plan strictly during implementation.`,
 
     it('should return feedback message when plan is rejected with feedback', async () => {
       const planRelativePath = createPlanFile('test.md', '# Content');
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       const confirmDetails = await invocation.shouldConfirmExecute(
         new AbortController().signal,
@@ -270,7 +270,7 @@ Revise the plan based on the feedback.`,
 
     it('should handle rejection without feedback gracefully', async () => {
       const planRelativePath = createPlanFile('test.md', '# Content');
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       const confirmDetails = await invocation.shouldConfirmExecute(
         new AbortController().signal,
@@ -296,7 +296,7 @@ Ask the user for specific feedback on how to improve the plan.`,
 
     it('should log plan execution event when plan is approved', async () => {
       const planRelativePath = createPlanFile('test.md', '# Content');
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       const confirmDetails = await invocation.shouldConfirmExecute(
         new AbortController().signal,
@@ -320,7 +320,7 @@ Ask the user for specific feedback on how to improve the plan.`,
 
     it('should return cancellation message when cancelled', async () => {
       const planRelativePath = createPlanFile('test.md', '# Content');
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       const confirmDetails = await invocation.shouldConfirmExecute(
         new AbortController().signal,
@@ -343,7 +343,7 @@ Ask the user for specific feedback on how to improve the plan.`,
   describe('execute when shouldConfirmExecute is never called', () => {
     it('should approve with DEFAULT mode when approvalPayload is null (policy ALLOW skips confirmation)', async () => {
       const planRelativePath = createPlanFile('test.md', '# Content');
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       // Simulate the scheduler's policy ALLOW path: execute() is called
       // directly without ever calling shouldConfirmExecute(), leaving
@@ -364,7 +364,7 @@ Ask the user for specific feedback on how to improve the plan.`,
     it('should return YOLO when config.isInteractive() is false', async () => {
       mockConfig.isInteractive = vi.fn().mockReturnValue(false);
       const planRelativePath = createPlanFile('test.md', '# Content');
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       // Directly call execute to trigger the internal getAllowApprovalMode
       const result = await invocation.execute(new AbortController().signal);
@@ -378,7 +378,7 @@ Ask the user for specific feedback on how to improve the plan.`,
     it('should return DEFAULT when config.isInteractive() is true', async () => {
       mockConfig.isInteractive = vi.fn().mockReturnValue(true);
       const planRelativePath = createPlanFile('test.md', '# Content');
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       // Directly call execute to trigger the internal getAllowApprovalMode
       const result = await invocation.execute(new AbortController().signal);
@@ -393,7 +393,7 @@ Ask the user for specific feedback on how to improve the plan.`,
   describe('getApprovalModeDescription (internal)', () => {
     it('should handle all valid approval modes', async () => {
       const planRelativePath = createPlanFile('test.md', '# Content');
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       const testMode = async (mode: ApprovalMode, expected: string) => {
         const confirmDetails = await invocation.shouldConfirmExecute(
@@ -426,7 +426,7 @@ Ask the user for specific feedback on how to improve the plan.`,
 
     it('should throw for invalid post-planning modes', async () => {
       const planRelativePath = createPlanFile('test.md', '# Content');
-      const invocation = tool.build({ plan_path: planRelativePath });
+      const invocation = tool.build({ plan_filename: planRelativePath });
 
       const testInvalidMode = async (mode: ApprovalMode) => {
         const confirmDetails = await invocation.shouldConfirmExecute(
@@ -448,36 +448,19 @@ Ask the user for specific feedback on how to improve the plan.`,
     });
   });
 
-  it('should throw error during build if plan path is outside plans directory', () => {
-    expect(() => tool.build({ plan_path: '../../../etc/passwd' })).toThrow(
-      /Access denied/,
-    );
-  });
-
   describe('validateToolParams', () => {
-    it('should reject empty plan_path', () => {
-      const result = tool.validateToolParams({ plan_path: '' });
-      expect(result).toBe('plan_path is required.');
+    it('should reject empty plan_filename', () => {
+      const result = tool.validateToolParams({ plan_filename: '' });
+      expect(result).toBe('plan_filename is required.');
     });
 
-    it('should reject whitespace-only plan_path', () => {
-      const result = tool.validateToolParams({ plan_path: '   ' });
-      expect(result).toBe('plan_path is required.');
-    });
-
-    it('should reject path outside plans directory', () => {
-      const result = tool.validateToolParams({
-        plan_path: '../../../etc/passwd',
-      });
-      expect(result).toContain('Access denied');
+    it('should reject whitespace-only plan_filename', () => {
+      const result = tool.validateToolParams({ plan_filename: '   ' });
+      expect(result).toBe('plan_filename is required.');
     });
 
     it('should reject non-existent plan file', async () => {
-      const result = await validatePlanPath(
-        'plans/ghost.md',
-        mockPlansDir,
-        tempRootDir,
-      );
+      const result = await validatePlanPath('ghost.md', mockPlansDir);
       expect(result).toContain('Plan file does not exist');
     });
 
@@ -488,18 +471,18 @@ Ask the user for specific feedback on how to improve the plan.`,
       fs.symlinkSync(outsideFile, maliciousPath);
 
       const result = tool.validateToolParams({
-        plan_path: 'plans/malicious.md',
+        plan_filename: 'malicious.md',
       });
 
       expect(result).toBe(
-        'Access denied: plan path must be within the designated plans directory.',
+        `Access denied: plan path (${path.join(mockPlansDir, 'malicious.md')}) must be within the designated plans directory (${mockPlansDir}).`,
       );
     });
 
     it('should accept valid path within plans directory', () => {
       createPlanFile('valid.md', '# Content');
       const result = tool.validateToolParams({
-        plan_path: 'plans/valid.md',
+        plan_filename: 'valid.md',
       });
       expect(result).toBeNull();
     });