remove topp default value 0.0

* feat: Add Qwen3-VL-Plus token limits (256K input, 32K output)

- Added 256K input context window limit for Qwen3-VL-Plus model
- Updated output token limit from 8K to 32K for Qwen3-VL-Plus
- Added comprehensive tests for both input and output limits

As requested by Qwen maintainers for proper model support.

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <noreply@anthropic.com>

* fix: enable high-res flag for qwen VL models

---------

Co-authored-by: Claude <noreply@anthropic.com>

.vscode/launch.json

@@ -101,6 +101,13 @@
       "env": {
         "GEMINI_SANDBOX": "false"
       }
+    },
+    {
+      "name": "Attach by Process ID",
+      "processId": "${command:PickProcess}",
+      "request": "attach",
+      "skipFiles": ["<node_internals>/**"],
+      "type": "node"
     }
   ],
   "inputs": [
@@ -115,6 +122,12 @@
       "type": "promptString",
       "description": "Enter your prompt for non-interactive mode",
       "default": "Explain this code"
+    },
+    {
+      "id": "debugPort",
+      "type": "promptString",
+      "description": "Enter the debug port number (default: 9229)",
+      "default": "9229"
     }
   ]
 }

CHANGELOG.md

@@ -1,5 +1,44 @@
 # Changelog
 
+## 0.0.14
+
+- Added plan mode support for task planning
+- Fixed unreliable editCorrector that injects extra escape characters
+- Fixed task tool dynamic updates
+- Added Qwen3-VL-Plus token limits (256K input, 32K output) and highres support
+- Enhanced dashScope cache control
+
+## 0.0.13
+
+- Added YOLO mode support for automatic vision model switching with CLI arguments and environment variables.
+- Fixed ripgrep lazy loading to resolve VS Code IDE companion startup issues.
+- Fixed authentication hang when selecting Qwen OAuth.
+- Added OpenAI and Qwen OAuth authentication support to Zed ACP integration.
+- Fixed output token limit for Qwen models.
+- Fixed Markdown list display issues on Windows.
+- Enhanced vision model instructions and documentation.
+- Improved authentication method compatibility across different IDE integrations.
+
+## 0.0.12
+
+- Added vision model support for Qwen-OAuth authentication.
+- Synced upstream `gemini-cli` to v0.3.4 with numerous improvements and bug fixes.
+- Enhanced subagent functionality with system reminders and improved user experience.
+- Added tool call type coercion for better compatibility.
+- Fixed arrow key navigation issues on Windows.
+- Fixed missing tool call chunks for OpenAI logging.
+- Fixed system prompt issues to avoid malformed tool calls.
+- Fixed terminal flicker when subagent is executing.
+- Fixed duplicate subagents configuration when running in home directory.
+- Fixed Esc key unable to cancel subagent dialog.
+- Added confirmation prompt for `/init` command when context file exists.
+- Added `skipLoopDetection` configuration option.
+- Fixed `is_background` parameter reset issues.
+- Enhanced Windows compatibility with multi-line paste handling.
+- Improved subagent documentation and branding consistency.
+- Fixed various linting errors and improved code quality.
+- Miscellaneous improvements and bug fixes.
+
 ## 0.0.11
 
 - Added subagents feature with file-based configuration system for specialized AI assistants.

README.md

@@ -54,6 +54,7 @@ For detailed setup instructions, see [Authorization](#authorization).
 - **Code Understanding & Editing** - Query and edit large codebases beyond traditional context window limits
 - **Workflow Automation** - Automate operational tasks like handling pull requests and complex rebases
 - **Enhanced Parser** - Adapted parser specifically optimized for Qwen-Coder models
+- **Vision Model Support** - Automatically detect images in your input and seamlessly switch to vision-capable models for multimodal analysis
 
 ## Installation
 
@@ -121,6 +122,58 @@ Create or edit `.qwen/settings.json` in your home directory:
 
 > 📝 **Note**: Session token limit applies to a single conversation, not cumulative API calls.
 
+### Vision Model Configuration
+
+Qwen Code includes intelligent vision model auto-switching that detects images in your input and can automatically switch to vision-capable models for multimodal analysis. **This feature is enabled by default** - when you include images in your queries, you'll see a dialog asking how you'd like to handle the vision model switch.
+
+#### Skip the Switch Dialog (Optional)
+
+If you don't want to see the interactive dialog each time, configure the default behavior in your `.qwen/settings.json`:
+
+```json
+{
+  "experimental": {
+    "vlmSwitchMode": "once"
+  }
+}
+```
+
+**Available modes:**
+
+- **`"once"`** - Switch to vision model for this query only, then revert
+- **`"session"`** - Switch to vision model for the entire session
+- **`"persist"`** - Continue with current model (no switching)
+- **Not set** - Show interactive dialog each time (default)
+
+#### Command Line Override
+
+You can also set the behavior via command line:
+
+```bash
+# Switch once per query
+qwen --vlm-switch-mode once
+
+# Switch for entire session
+qwen --vlm-switch-mode session
+
+# Never switch automatically
+qwen --vlm-switch-mode persist
+```
+
+#### Disable Vision Models (Optional)
+
+To completely disable vision model support, add to your `.qwen/settings.json`:
+
+```json
+{
+  "experimental": {
+    "visionModelPreview": false
+  }
+}
+```
+
+> 💡 **Tip**: In YOLO mode (`--yolo`), vision switching happens automatically without prompts when images are detected.
+
 ### Authorization
 
 Choose your preferred authentication method based on your needs:

docs/cli/commands.md

@@ -124,6 +124,18 @@ Slash commands provide meta-level control over the CLI itself.
 - **`/auth`**
   - **Description:** Open a dialog that lets you change the authentication method.
 
+- **`/approval-mode`**
+  - **Description:** Change the approval mode for tool usage.
+  - **Usage:** `/approval-mode [mode] [--session|--project|--user]`
+  - **Available Modes:**
+    - **`plan`**: Analyze only; do not modify files or execute commands
+    - **`default`**: Require approval for file edits or shell commands
+    - **`auto-edit`**: Automatically approve file edits
+    - **`yolo`**: Automatically approve all tools
+  - **Examples:**
+    - `/approval-mode plan --project` (persist plan mode for this project)
+    - `/approval-mode yolo --user` (persist YOLO mode for this user across projects)
+
 - **`/about`**
   - **Description:** Show version info. Please share this information when filing issues.
 

docs/cli/configuration.md

@@ -362,6 +362,18 @@ If you are experiencing performance issues with file searching (e.g., with `@` c
     "skipLoopDetection": true
     ```
 
+- **`approvalMode`** (string):
+  - **Description:** Sets the default approval mode for tool usage. Accepted values are:
+    - `plan`: Analyze only, do not modify files or execute commands.
+    - `default`: Require approval before file edits or shell commands run.
+    - `auto-edit`: Automatically approve file edits.
+    - `yolo`: Automatically approve all tool calls.
+  - **Default:** `"default"`
+  - **Example:**
+    ```json
+    "approvalMode": "plan"
+    ```
+
 ### Example `settings.json`:
 
 ```json
@@ -486,12 +498,13 @@ Arguments passed directly when running the CLI can override other configurations
 - **`--yolo`**:
   - Enables YOLO mode, which automatically approves all tool calls.
 - **`--approval-mode <mode>`**:
-  - Sets the approval mode for tool calls. Available modes:
-    - `default`: Prompt for approval on each tool call (default behavior)
-    - `auto_edit`: Automatically approve edit tools (edit, write_file) while prompting for others
-    - `yolo`: Automatically approve all tool calls (equivalent to `--yolo`)
+  - Sets the approval mode for tool calls. Supported modes:
+    - `plan`: Analyze only—do not modify files or execute commands.
+    - `default`: Require approval for file edits or shell commands (default behavior).
+    - `auto-edit`: Automatically approve edit tools (edit, write_file) while prompting for others.
+    - `yolo`: Automatically approve all tool calls (equivalent to `--yolo`).
   - Cannot be used together with `--yolo`. Use `--approval-mode=yolo` instead of `--yolo` for the new unified approach.
-  - Example: `qwen --approval-mode auto_edit`
+  - Example: `qwen --approval-mode auto-edit`
 - **`--allowed-tools <tool1,tool2,...>`**:
   - A comma-separated list of tool names that will bypass the confirmation dialog.
   - Example: `qwen --allowed-tools "ShellTool(git status)"`

docs/keyboard-shortcuts.md

@@ -4,16 +4,16 @@ This document lists the available keyboard shortcuts in Qwen Code.
 
 ## General
 
-| Shortcut | Description                                                                                                           |
-| -------- | --------------------------------------------------------------------------------------------------------------------- |
-| `Esc`    | Close dialogs and suggestions.                                                                                        |
-| `Ctrl+C` | Cancel the ongoing request and clear the input. Press twice to exit the application.                                  |
-| `Ctrl+D` | Exit the application if the input is empty. Press twice to confirm.                                                   |
-| `Ctrl+L` | Clear the screen.                                                                                                     |
-| `Ctrl+O` | Toggle the display of the debug console.                                                                              |
-| `Ctrl+S` | Allows long responses to print fully, disabling truncation. Use your terminal's scrollback to view the entire output. |
-| `Ctrl+T` | Toggle the display of tool descriptions.                                                                              |
-| `Ctrl+Y` | Toggle auto-approval (YOLO mode) for all tool calls.                                                                  |
+| Shortcut    | Description                                                                                                           |
+| ----------- | --------------------------------------------------------------------------------------------------------------------- |
+| `Esc`       | Close dialogs and suggestions.                                                                                        |
+| `Ctrl+C`    | Cancel the ongoing request and clear the input. Press twice to exit the application.                                  |
+| `Ctrl+D`    | Exit the application if the input is empty. Press twice to confirm.                                                   |
+| `Ctrl+L`    | Clear the screen.                                                                                                     |
+| `Ctrl+O`    | Toggle the display of the debug console.                                                                              |
+| `Ctrl+S`    | Allows long responses to print fully, disabling truncation. Use your terminal's scrollback to view the entire output. |
+| `Ctrl+T`    | Toggle the display of tool descriptions.                                                                              |
+| `Shift+Tab` | Cycle approval modes (`plan` → `default` → `auto-edit` → `yolo`).                                                     |
 
 ## Input Prompt
 

docs/subagents.md

@@ -133,6 +133,28 @@ Focus on creating clear, comprehensive documentation that helps both
 new contributors and end users understand the project.
 ```
 
+## Using Subagents Effectively
+
+### Automatic Delegation
+
+Qwen Code proactively delegates tasks based on:
+
+- The task description in your request
+- The description field in subagent configurations
+- Current context and available tools
+
+To encourage more proactive subagent use, include phrases like "use PROACTIVELY" or "MUST BE USED" in your description field.
+
+### Explicit Invocation
+
+Request a specific subagent by mentioning it in your command:
+
+```
+> Let the testing-expert subagent create unit tests for the payment module
+> Have the documentation-writer subagent update the API reference
+> Get the react-specialist subagent to optimize this component's performance
+```
+
 ## Examples
 
 ### Development Workflow Agents

package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "@qwen-code/qwen-code",
-  "version": "0.0.11",
+  "version": "0.0.14",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "@qwen-code/qwen-code",
-      "version": "0.0.11",
+      "version": "0.0.14",
       "workspaces": [
         "packages/*"
       ],
@@ -13454,7 +13454,7 @@
     },
     "packages/cli": {
       "name": "@qwen-code/qwen-code",
-      "version": "0.0.11",
+      "version": "0.0.14",
       "dependencies": {
         "@google/genai": "1.9.0",
         "@iarna/toml": "^2.2.5",
@@ -13662,7 +13662,7 @@
     },
     "packages/core": {
       "name": "@qwen-code/qwen-code-core",
-      "version": "0.0.11",
+      "version": "0.0.14",
       "dependencies": {
         "@google/genai": "1.13.0",
         "@lvce-editor/ripgrep": "^1.6.0",
@@ -13788,7 +13788,7 @@
     },
     "packages/test-utils": {
       "name": "@qwen-code/qwen-code-test-utils",
-      "version": "0.0.11",
+      "version": "0.0.14",
       "dev": true,
       "license": "Apache-2.0",
       "devDependencies": {
@@ -13800,7 +13800,7 @@
     },
     "packages/vscode-ide-companion": {
       "name": "qwen-code-vscode-ide-companion",
-      "version": "0.0.11",
+      "version": "0.0.14",
       "license": "LICENSE",
       "dependencies": {
         "@modelcontextprotocol/sdk": "^1.15.1",

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@qwen-code/qwen-code",
-  "version": "0.0.11",
+  "version": "0.0.14",
   "engines": {
     "node": ">=20.0.0"
   },
@@ -13,7 +13,7 @@
     "url": "git+https://github.com/QwenLM/qwen-code.git"
   },
   "config": {
-    "sandboxImageUri": "ghcr.io/qwenlm/qwen-code:0.0.11"
+    "sandboxImageUri": "ghcr.io/qwenlm/qwen-code:0.0.14"
   },
   "scripts": {
     "start": "node scripts/start.js",

packages/cli/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@qwen-code/qwen-code",
-  "version": "0.0.11",
+  "version": "0.0.14",
   "description": "Qwen Code",
   "repository": {
     "type": "git",
@@ -25,7 +25,7 @@
     "dist"
   ],
   "config": {
-    "sandboxImageUri": "ghcr.io/qwenlm/qwen-code:0.0.11"
+    "sandboxImageUri": "ghcr.io/qwenlm/qwen-code:0.0.14"
   },
   "dependencies": {
     "@google/genai": "1.9.0",

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

@@ -269,7 +269,7 @@ describe('Configuration Integration Tests', () => {
       parseArguments = parseArgs;
     });
 
-    it('should parse --approval-mode=auto_edit correctly through the full argument parsing flow', async () => {
+    it('should parse --approval-mode=auto-edit correctly through the full argument parsing flow', async () => {
       const originalArgv = process.argv;
 
       try {
@@ -277,7 +277,7 @@ describe('Configuration Integration Tests', () => {
           'node',
           'script.js',
           '--approval-mode',
-          'auto_edit',
+          'auto-edit',
           '-p',
           'test',
         ];
@@ -285,7 +285,30 @@ describe('Configuration Integration Tests', () => {
         const argv = await parseArguments({} as Settings);
 
         // Verify that the argument was parsed correctly
-        expect(argv.approvalMode).toBe('auto_edit');
+        expect(argv.approvalMode).toBe('auto-edit');
+        expect(argv.prompt).toBe('test');
+        expect(argv.yolo).toBe(false);
+      } finally {
+        process.argv = originalArgv;
+      }
+    });
+
+    it('should parse --approval-mode=plan correctly through the full argument parsing flow', async () => {
+      const originalArgv = process.argv;
+
+      try {
+        process.argv = [
+          'node',
+          'script.js',
+          '--approval-mode',
+          'plan',
+          '-p',
+          'test',
+        ];
+
+        const argv = await parseArguments({} as Settings);
+
+        expect(argv.approvalMode).toBe('plan');
         expect(argv.prompt).toBe('test');
         expect(argv.yolo).toBe(false);
       } finally {

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

@@ -262,9 +262,9 @@ describe('parseArguments', () => {
   });
 
   it('should allow --approval-mode without --yolo', async () => {
-    process.argv = ['node', 'script.js', '--approval-mode', 'auto_edit'];
+    process.argv = ['node', 'script.js', '--approval-mode', 'auto-edit'];
     const argv = await parseArguments({} as Settings);
-    expect(argv.approvalMode).toBe('auto_edit');
+    expect(argv.approvalMode).toBe('auto-edit');
     expect(argv.yolo).toBe(false);
   });
 
@@ -1087,6 +1087,32 @@ describe('Approval mode tool exclusion logic', () => {
     expect(excludedTools).toContain(WriteFileTool.Name);
   });
 
+  it('should exclude all interactive tools in non-interactive mode with plan approval mode', async () => {
+    process.argv = [
+      'node',
+      'script.js',
+      '--approval-mode',
+      'plan',
+      '-p',
+      'test',
+    ];
+    const argv = await parseArguments({} as Settings);
+    const settings: Settings = {};
+    const extensions: Extension[] = [];
+
+    const config = await loadCliConfig(
+      settings,
+      extensions,
+      'test-session',
+      argv,
+    );
+
+    const excludedTools = config.getExcludeTools();
+    expect(excludedTools).toContain(ShellTool.Name);
+    expect(excludedTools).toContain(EditTool.Name);
+    expect(excludedTools).toContain(WriteFileTool.Name);
+  });
+
   it('should exclude all interactive tools in non-interactive mode with explicit default approval mode', async () => {
     process.argv = [
       'node',
@@ -1113,12 +1139,12 @@ describe('Approval mode tool exclusion logic', () => {
     expect(excludedTools).toContain(WriteFileTool.Name);
   });
 
-  it('should exclude only shell tools in non-interactive mode with auto_edit approval mode', async () => {
+  it('should exclude only shell tools in non-interactive mode with auto-edit approval mode', async () => {
     process.argv = [
       'node',
       'script.js',
       '--approval-mode',
-      'auto_edit',
+      'auto-edit',
       '-p',
       'test',
     ];
@@ -1189,8 +1215,9 @@ describe('Approval mode tool exclusion logic', () => {
 
     const testCases = [
       { args: ['node', 'script.js'] }, // default
+      { args: ['node', 'script.js', '--approval-mode', 'plan'] },
       { args: ['node', 'script.js', '--approval-mode', 'default'] },
-      { args: ['node', 'script.js', '--approval-mode', 'auto_edit'] },
+      { args: ['node', 'script.js', '--approval-mode', 'auto-edit'] },
       { args: ['node', 'script.js', '--approval-mode', 'yolo'] },
       { args: ['node', 'script.js', '--yolo'] },
     ];
@@ -1215,12 +1242,12 @@ describe('Approval mode tool exclusion logic', () => {
     }
   });
 
-  it('should merge approval mode exclusions with settings exclusions in auto_edit mode', async () => {
+  it('should merge approval mode exclusions with settings exclusions in auto-edit mode', async () => {
     process.argv = [
       'node',
       'script.js',
       '--approval-mode',
-      'auto_edit',
+      'auto-edit',
       '-p',
       'test',
     ];
@@ -1238,8 +1265,8 @@ describe('Approval mode tool exclusion logic', () => {
     const excludedTools = config.getExcludeTools();
     expect(excludedTools).toContain('custom_tool'); // From settings
     expect(excludedTools).toContain(ShellTool.Name); // From approval mode
-    expect(excludedTools).not.toContain(EditTool.Name); // Should be allowed in auto_edit
-    expect(excludedTools).not.toContain(WriteFileTool.Name); // Should be allowed in auto_edit
+    expect(excludedTools).not.toContain(EditTool.Name); // Should be allowed in auto-edit
+    expect(excludedTools).not.toContain(WriteFileTool.Name); // Should be allowed in auto-edit
   });
 
   it('should throw an error for invalid approval mode values in loadCliConfig', async () => {
@@ -1262,7 +1289,7 @@ describe('Approval mode tool exclusion logic', () => {
         invalidArgv as CliArgs,
       ),
     ).rejects.toThrow(
-      'Invalid approval mode: invalid_mode. Valid values are: yolo, auto_edit, default',
+      'Invalid approval mode: invalid_mode. Valid values are: plan, default, auto-edit, yolo',
     );
   });
 });
@@ -1514,7 +1541,7 @@ describe('loadCliConfig model selection', () => {
       argv,
     );
 
-    expect(config.getModel()).toBe('qwen3-coder-plus');
+    expect(config.getModel()).toBe('coder-model');
   });
 
   it('always prefers model from argvs', async () => {
@@ -1929,6 +1956,13 @@ describe('loadCliConfig approval mode', () => {
     expect(config.getApprovalMode()).toBe(ServerConfig.ApprovalMode.DEFAULT);
   });
 
+  it('should set PLAN approval mode when --approval-mode=plan', async () => {
+    process.argv = ['node', 'script.js', '--approval-mode', 'plan'];
+    const argv = await parseArguments({} as Settings);
+    const config = await loadCliConfig({}, [], 'test-session', argv);
+    expect(config.getApprovalMode()).toBe(ServerConfig.ApprovalMode.PLAN);
+  });
+
   it('should set YOLO approval mode when --yolo flag is used', async () => {
     process.argv = ['node', 'script.js', '--yolo'];
     const argv = await parseArguments({} as Settings);
@@ -1950,8 +1984,8 @@ describe('loadCliConfig approval mode', () => {
     expect(config.getApprovalMode()).toBe(ServerConfig.ApprovalMode.DEFAULT);
   });
 
-  it('should set AUTO_EDIT approval mode when --approval-mode=auto_edit', async () => {
-    process.argv = ['node', 'script.js', '--approval-mode', 'auto_edit'];
+  it('should set AUTO_EDIT approval mode when --approval-mode=auto-edit', async () => {
+    process.argv = ['node', 'script.js', '--approval-mode', 'auto-edit'];
     const argv = await parseArguments({} as Settings);
     const config = await loadCliConfig({}, [], 'test-session', argv);
     expect(config.getApprovalMode()).toBe(ServerConfig.ApprovalMode.AUTO_EDIT);
@@ -1964,6 +1998,33 @@ describe('loadCliConfig approval mode', () => {
     expect(config.getApprovalMode()).toBe(ServerConfig.ApprovalMode.YOLO);
   });
 
+  it('should use approval mode from settings when CLI flags are not provided', async () => {
+    process.argv = ['node', 'script.js'];
+    const argv = await parseArguments({} as Settings);
+    const settings: Settings = { approvalMode: 'plan' };
+    const config = await loadCliConfig(settings, [], 'test-session', argv);
+    expect(config.getApprovalMode()).toBe(ServerConfig.ApprovalMode.PLAN);
+  });
+
+  it('should normalize approval mode values from settings', async () => {
+    process.argv = ['node', 'script.js'];
+    const argv = await parseArguments({} as Settings);
+    const settings: Settings = { approvalMode: 'AutoEdit' };
+    const config = await loadCliConfig(settings, [], 'test-session', argv);
+    expect(config.getApprovalMode()).toBe(ServerConfig.ApprovalMode.AUTO_EDIT);
+  });
+
+  it('should throw when approval mode in settings is invalid', async () => {
+    process.argv = ['node', 'script.js'];
+    const argv = await parseArguments({} as Settings);
+    const settings: Settings = { approvalMode: 'invalid_mode' };
+    await expect(
+      loadCliConfig(settings, [], 'test-session', argv),
+    ).rejects.toThrow(
+      'Invalid approval mode: invalid_mode. Valid values are: plan, default, auto-edit, yolo',
+    );
+  });
+
   it('should prioritize --approval-mode over --yolo when both would be valid (but validation prevents this)', async () => {
     // Note: This test documents the intended behavior, but in practice the validation
     // prevents both flags from being used together
@@ -1995,8 +2056,8 @@ describe('loadCliConfig approval mode', () => {
       expect(config.getApprovalMode()).toBe(ServerConfig.ApprovalMode.DEFAULT);
     });
 
-    it('should override --approval-mode=auto_edit to DEFAULT', async () => {
-      process.argv = ['node', 'script.js', '--approval-mode', 'auto_edit'];
+    it('should override --approval-mode=auto-edit to DEFAULT', async () => {
+      process.argv = ['node', 'script.js', '--approval-mode', 'auto-edit'];
       const argv = await parseArguments({} as Settings);
       const config = await loadCliConfig({}, [], 'test-session', argv);
       expect(config.getApprovalMode()).toBe(ServerConfig.ApprovalMode.DEFAULT);
@@ -2015,6 +2076,13 @@ describe('loadCliConfig approval mode', () => {
       const config = await loadCliConfig({}, [], 'test-session', argv);
       expect(config.getApprovalMode()).toBe(ServerConfig.ApprovalMode.DEFAULT);
     });
+
+    it('should allow PLAN approval mode in untrusted folders', async () => {
+      process.argv = ['node', 'script.js', '--approval-mode', 'plan'];
+      const argv = await parseArguments({} as Settings);
+      const config = await loadCliConfig({}, [], 'test-session', argv);
+      expect(config.getApprovalMode()).toBe(ServerConfig.ApprovalMode.PLAN);
+    });
   });
 });
 

packages/cli/src/config/config.ts

@@ -52,6 +52,39 @@ const logger = {
   error: (...args: any[]) => console.error('[ERROR]', ...args),
 };
 
+const VALID_APPROVAL_MODE_VALUES = [
+  'plan',
+  'default',
+  'auto-edit',
+  'yolo',
+] as const;
+
+function formatApprovalModeError(value: string): Error {
+  return new Error(
+    `Invalid approval mode: ${value}. Valid values are: ${VALID_APPROVAL_MODE_VALUES.join(
+      ', ',
+    )}`,
+  );
+}
+
+function parseApprovalModeValue(value: string): ApprovalMode {
+  const normalized = value.trim().toLowerCase();
+  switch (normalized) {
+    case 'plan':
+      return ApprovalMode.PLAN;
+    case 'default':
+      return ApprovalMode.DEFAULT;
+    case 'yolo':
+      return ApprovalMode.YOLO;
+    case 'auto_edit':
+    case 'autoedit':
+    case 'auto-edit':
+      return ApprovalMode.AUTO_EDIT;
+    default:
+      throw formatApprovalModeError(value);
+  }
+}
+
 export interface CliArgs {
   model: string | undefined;
   sandbox: boolean | string | undefined;
@@ -82,6 +115,7 @@ export interface CliArgs {
   includeDirectories: string[] | undefined;
   tavilyApiKey: string | undefined;
   screenReader: boolean | undefined;
+  vlmSwitchMode: string | undefined;
 }
 
 export async function parseArguments(settings: Settings): Promise<CliArgs> {
@@ -146,9 +180,9 @@ export async function parseArguments(settings: Settings): Promise<CliArgs> {
         })
         .option('approval-mode', {
           type: 'string',
-          choices: ['default', 'auto_edit', 'yolo'],
+          choices: ['plan', 'default', 'auto-edit', 'yolo'],
           description:
-            'Set the approval mode: default (prompt for approval), auto_edit (auto-approve edit tools), yolo (auto-approve all tools)',
+            'Set the approval mode: plan (plan only), default (prompt for approval), auto-edit (auto-approve edit tools), yolo (auto-approve all tools)',
         })
         .option('telemetry', {
           type: 'boolean',
@@ -249,6 +283,13 @@ export async function parseArguments(settings: Settings): Promise<CliArgs> {
           description: 'Enable screen reader mode for accessibility.',
           default: false,
         })
+        .option('vlm-switch-mode', {
+          type: 'string',
+          choices: ['once', 'session', 'persist'],
+          description:
+            'Default behavior when images are detected in input. Values: once (one-time switch), session (switch for entire session), persist (continue with current model). Overrides settings files.',
+          default: process.env['VLM_SWITCH_MODE'],
+        })
         .check((argv) => {
           if (argv.prompt && argv['promptInteractive']) {
             throw new Error(
@@ -430,30 +471,21 @@ export async function loadCliConfig(
   // Determine approval mode with backward compatibility
   let approvalMode: ApprovalMode;
   if (argv.approvalMode) {
-    // New --approval-mode flag takes precedence
-    switch (argv.approvalMode) {
-      case 'yolo':
-        approvalMode = ApprovalMode.YOLO;
-        break;
-      case 'auto_edit':
-        approvalMode = ApprovalMode.AUTO_EDIT;
-        break;
-      case 'default':
-        approvalMode = ApprovalMode.DEFAULT;
-        break;
-      default:
-        throw new Error(
-          `Invalid approval mode: ${argv.approvalMode}. Valid values are: yolo, auto_edit, default`,
-        );
-    }
+    approvalMode = parseApprovalModeValue(argv.approvalMode);
+  } else if (argv.yolo) {
+    approvalMode = ApprovalMode.YOLO;
+  } else if (settings.approvalMode) {
+    approvalMode = parseApprovalModeValue(settings.approvalMode);
   } else {
-    // Fallback to legacy --yolo flag behavior
-    approvalMode =
-      argv.yolo || false ? ApprovalMode.YOLO : ApprovalMode.DEFAULT;
+    approvalMode = ApprovalMode.DEFAULT;
   }
 
   // Force approval mode to default if the folder is not trusted.
-  if (!trustedFolder && approvalMode !== ApprovalMode.DEFAULT) {
+  if (
+    !trustedFolder &&
+    approvalMode !== ApprovalMode.DEFAULT &&
+    approvalMode !== ApprovalMode.PLAN
+  ) {
     logger.warn(
       `Approval mode overridden to "default" because the current folder is not trusted.`,
     );
@@ -466,6 +498,7 @@ export async function loadCliConfig(
   const extraExcludes: string[] = [];
   if (!interactive && !argv.experimentalAcp) {
     switch (approvalMode) {
+      case ApprovalMode.PLAN:
       case ApprovalMode.DEFAULT:
         // In default non-interactive mode, all tools that require approval are excluded.
         extraExcludes.push(ShellTool.Name, EditTool.Name, WriteFileTool.Name);
@@ -524,6 +557,9 @@ export async function loadCliConfig(
     argv.screenReader !== undefined
       ? argv.screenReader
       : (settings.ui?.accessibility?.screenReader ?? false);
+
+  const vlmSwitchMode =
+    argv.vlmSwitchMode || settings.experimental?.vlmSwitchMode;
   return new Config({
     sessionId,
     embeddingModel: DEFAULT_GEMINI_EMBEDDING_MODEL,
@@ -630,6 +666,7 @@ export async function loadCliConfig(
     skipNextSpeakerCheck: settings.model?.skipNextSpeakerCheck,
     enablePromptCompletion: settings.general?.enablePromptCompletion ?? false,
     skipLoopDetection: settings.skipLoopDetection ?? false,
+    vlmSwitchMode,
   });
 }
 

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

@@ -69,7 +69,11 @@ const MOCK_WORKSPACE_SETTINGS_PATH = pathActual.join(
 );
 
 // A more flexible type for test data that allows arbitrary properties.
-type TestSettings = Settings & { [key: string]: unknown };
+type TestSettings = Settings & {
+  [key: string]: unknown;
+  nested?: { [key: string]: unknown };
+  nestedObj?: { [key: string]: unknown };
+};
 
 vi.mock('fs', async (importOriginal) => {
   // Get all the functions from the real 'fs' module
@@ -137,6 +141,9 @@ describe('Settings Loading and Merging', () => {
         advanced: {
           excludedEnvVars: [],
         },
+        experimental: {},
+        contentGenerator: {},
+        systemPromptMappings: {},
         extensions: {
           disabled: [],
           workspacesWithMigrationNudge: [],
@@ -197,6 +204,9 @@ describe('Settings Loading and Merging', () => {
         advanced: {
           excludedEnvVars: [],
         },
+        experimental: {},
+        contentGenerator: {},
+        systemPromptMappings: {},
         extensions: {
           disabled: [],
           workspacesWithMigrationNudge: [],
@@ -260,6 +270,9 @@ describe('Settings Loading and Merging', () => {
         advanced: {
           excludedEnvVars: [],
         },
+        experimental: {},
+        contentGenerator: {},
+        systemPromptMappings: {},
         extensions: {
           disabled: [],
           workspacesWithMigrationNudge: [],
@@ -320,6 +333,9 @@ describe('Settings Loading and Merging', () => {
         advanced: {
           excludedEnvVars: [],
         },
+        experimental: {},
+        contentGenerator: {},
+        systemPromptMappings: {},
         extensions: {
           disabled: [],
           workspacesWithMigrationNudge: [],
@@ -385,6 +401,9 @@ describe('Settings Loading and Merging', () => {
         advanced: {
           excludedEnvVars: [],
         },
+        experimental: {},
+        contentGenerator: {},
+        systemPromptMappings: {},
         extensions: {
           disabled: [],
           workspacesWithMigrationNudge: [],
@@ -477,6 +496,9 @@ describe('Settings Loading and Merging', () => {
         advanced: {
           excludedEnvVars: [],
         },
+        experimental: {},
+        contentGenerator: {},
+        systemPromptMappings: {},
         extensions: {
           disabled: [],
           workspacesWithMigrationNudge: [],
@@ -562,6 +584,9 @@ describe('Settings Loading and Merging', () => {
         advanced: {
           excludedEnvVars: [],
         },
+        experimental: {},
+        contentGenerator: {},
+        systemPromptMappings: {},
         extensions: {
           disabled: [],
           workspacesWithMigrationNudge: [],
@@ -691,6 +716,9 @@ describe('Settings Loading and Merging', () => {
             '/system/dir',
           ],
         },
+        experimental: {},
+        contentGenerator: {},
+        systemPromptMappings: {},
         extensions: {
           disabled: [],
           workspacesWithMigrationNudge: [],
@@ -1431,6 +1459,9 @@ describe('Settings Loading and Merging', () => {
         advanced: {
           excludedEnvVars: [],
         },
+        experimental: {},
+        contentGenerator: {},
+        systemPromptMappings: {},
         extensions: {
           disabled: [],
           workspacesWithMigrationNudge: [],
@@ -1516,7 +1547,11 @@ describe('Settings Loading and Merging', () => {
         'workspace_endpoint_from_env/api',
       );
       expect(
-        (settings.workspace.settings as TestSettings)['nested']['value'],
+        (
+          (settings.workspace.settings as TestSettings).nested as {
+            [key: string]: unknown;
+          }
+        )['value'],
       ).toBe('workspace_endpoint_from_env');
       expect((settings.merged as TestSettings)['endpoint']).toBe(
         'workspace_endpoint_from_env/api',
@@ -1766,19 +1801,39 @@ describe('Settings Loading and Merging', () => {
       ).toBeUndefined();
 
       expect(
-        (settings.user.settings as TestSettings)['nestedObj']['nestedNull'],
+        (
+          (settings.user.settings as TestSettings).nestedObj as {
+            [key: string]: unknown;
+          }
+        )['nestedNull'],
       ).toBeNull();
       expect(
-        (settings.user.settings as TestSettings)['nestedObj']['nestedBool'],
+        (
+          (settings.user.settings as TestSettings).nestedObj as {
+            [key: string]: unknown;
+          }
+        )['nestedBool'],
       ).toBe(true);
       expect(
-        (settings.user.settings as TestSettings)['nestedObj']['nestedNum'],
+        (
+          (settings.user.settings as TestSettings).nestedObj as {
+            [key: string]: unknown;
+          }
+        )['nestedNum'],
       ).toBe(0);
       expect(
-        (settings.user.settings as TestSettings)['nestedObj']['nestedString'],
+        (
+          (settings.user.settings as TestSettings).nestedObj as {
+            [key: string]: unknown;
+          }
+        )['nestedString'],
       ).toBe('literal');
       expect(
-        (settings.user.settings as TestSettings)['nestedObj']['anotherEnv'],
+        (
+          (settings.user.settings as TestSettings).nestedObj as {
+            [key: string]: unknown;
+          }
+        )['anotherEnv'],
       ).toBe('env_string_nested_value');
 
       delete process.env['MY_ENV_STRING'];
@@ -1864,6 +1919,9 @@ describe('Settings Loading and Merging', () => {
           advanced: {
             excludedEnvVars: [],
           },
+          experimental: {},
+          contentGenerator: {},
+          systemPromptMappings: {},
           extensions: {
             disabled: [],
             workspacesWithMigrationNudge: [],
@@ -2336,14 +2394,14 @@ describe('Settings Loading and Merging', () => {
           vimMode: false,
         },
         model: {
-          maxSessionTurns: 0,
+          maxSessionTurns: -1,
         },
         context: {
           includeDirectories: [],
         },
         security: {
           folderTrust: {
-            enabled: null,
+            enabled: false,
           },
         },
       };
@@ -2352,9 +2410,9 @@ describe('Settings Loading and Merging', () => {
 
       expect(v1Settings).toEqual({
         vimMode: false,
-        maxSessionTurns: 0,
+        maxSessionTurns: -1,
         includeDirectories: [],
-        folderTrust: null,
+        folderTrust: false,
       });
     });
 

packages/cli/src/config/settings.ts

@@ -396,6 +396,24 @@ function mergeSettings(
         ]),
       ],
     },
+    experimental: {
+      ...(systemDefaults.experimental || {}),
+      ...(user.experimental || {}),
+      ...(safeWorkspaceWithoutFolderTrust.experimental || {}),
+      ...(system.experimental || {}),
+    },
+    contentGenerator: {
+      ...(systemDefaults.contentGenerator || {}),
+      ...(user.contentGenerator || {}),
+      ...(safeWorkspaceWithoutFolderTrust.contentGenerator || {}),
+      ...(system.contentGenerator || {}),
+    },
+    systemPromptMappings: {
+      ...(systemDefaults.systemPromptMappings || {}),
+      ...(user.systemPromptMappings || {}),
+      ...(safeWorkspaceWithoutFolderTrust.systemPromptMappings || {}),
+      ...(system.systemPromptMappings || {}),
+    },
     extensions: {
       ...(systemDefaults.extensions || {}),
       ...(user.extensions || {}),

packages/cli/src/config/settingsSchema.ts

@@ -741,6 +741,26 @@ export const SETTINGS_SCHEMA = {
         description: 'Enable extension management features.',
         showInDialog: false,
       },
+      visionModelPreview: {
+        type: 'boolean',
+        label: 'Vision Model Preview',
+        category: 'Experimental',
+        requiresRestart: false,
+        default: true,
+        description:
+          'Enable vision model support and auto-switching functionality. When disabled, vision models like qwen-vl-max-latest will be hidden and auto-switching will not occur.',
+        showInDialog: true,
+      },
+      vlmSwitchMode: {
+        type: 'string',
+        label: 'VLM Switch Mode',
+        category: 'Experimental',
+        requiresRestart: false,
+        default: undefined as string | undefined,
+        description:
+          'Default behavior when images are detected in input. Values: once (one-time switch), session (switch for entire session), persist (continue with current model). If not set, user will be prompted each time. This is a temporary experimental feature.',
+        showInDialog: false,
+      },
     },
   },
 
@@ -872,6 +892,16 @@ export const SETTINGS_SCHEMA = {
     description: 'Disable all loop detection checks (streaming and LLM).',
     showInDialog: true,
   },
+  approvalMode: {
+    type: 'string',
+    label: 'Default Approval Mode',
+    category: 'General',
+    requiresRestart: false,
+    default: 'default',
+    description:
+      'Default approval mode for tool usage. Valid values: plan, default, auto-edit, yolo.',
+    showInDialog: true,
+  },
   enableWelcomeBack: {
     type: 'boolean',
     label: 'Enable Welcome Back',

packages/cli/src/services/BuiltinCommandLoader.test.ts

@@ -15,6 +15,14 @@ vi.mock('../ui/commands/aboutCommand.js', async () => {
   };
 });
 
+vi.mock('../ui/commands/approvalModeCommand.js', () => ({
+  approvalModeCommand: {
+    name: 'approval-mode',
+    description: 'Approval mode command',
+    kind: 'built-in',
+  },
+}));
+
 vi.mock('../ui/commands/ideCommand.js', () => ({ ideCommand: vi.fn() }));
 vi.mock('../ui/commands/restoreCommand.js', () => ({
   restoreCommand: vi.fn(),
@@ -56,6 +64,13 @@ vi.mock('../ui/commands/mcpCommand.js', () => ({
     kind: 'BUILT_IN',
   },
 }));
+vi.mock('../ui/commands/modelCommand.js', () => ({
+  modelCommand: {
+    name: 'model',
+    description: 'Model command',
+    kind: 'BUILT_IN',
+  },
+}));
 
 describe('BuiltinCommandLoader', () => {
   let mockConfig: Config;
@@ -121,10 +136,17 @@ describe('BuiltinCommandLoader', () => {
     expect(aboutCmd).toBeDefined();
     expect(aboutCmd?.kind).toBe(CommandKind.BUILT_IN);
 
+    const approvalModeCmd = commands.find((c) => c.name === 'approval-mode');
+    expect(approvalModeCmd).toBeDefined();
+    expect(approvalModeCmd?.kind).toBe(CommandKind.BUILT_IN);
+
     const ideCmd = commands.find((c) => c.name === 'ide');
     expect(ideCmd).toBeDefined();
 
     const mcpCmd = commands.find((c) => c.name === 'mcp');
     expect(mcpCmd).toBeDefined();
+
+    const modelCmd = commands.find((c) => c.name === 'model');
+    expect(modelCmd).toBeDefined();
   });
 });

packages/cli/src/services/BuiltinCommandLoader.ts

@@ -8,6 +8,8 @@ import type { ICommandLoader } from './types.js';
 import type { SlashCommand } from '../ui/commands/types.js';
 import type { Config } from '@qwen-code/qwen-code-core';
 import { aboutCommand } from '../ui/commands/aboutCommand.js';
+import { agentsCommand } from '../ui/commands/agentsCommand.js';
+import { approvalModeCommand } from '../ui/commands/approvalModeCommand.js';
 import { authCommand } from '../ui/commands/authCommand.js';
 import { bugCommand } from '../ui/commands/bugCommand.js';
 import { chatCommand } from '../ui/commands/chatCommand.js';
@@ -24,18 +26,18 @@ import { ideCommand } from '../ui/commands/ideCommand.js';
 import { initCommand } from '../ui/commands/initCommand.js';
 import { mcpCommand } from '../ui/commands/mcpCommand.js';
 import { memoryCommand } from '../ui/commands/memoryCommand.js';
+import { modelCommand } from '../ui/commands/modelCommand.js';
 import { privacyCommand } from '../ui/commands/privacyCommand.js';
 import { quitCommand, quitConfirmCommand } from '../ui/commands/quitCommand.js';
 import { restoreCommand } from '../ui/commands/restoreCommand.js';
+import { settingsCommand } from '../ui/commands/settingsCommand.js';
 import { statsCommand } from '../ui/commands/statsCommand.js';
 import { summaryCommand } from '../ui/commands/summaryCommand.js';
+import { terminalSetupCommand } from '../ui/commands/terminalSetupCommand.js';
 import { themeCommand } from '../ui/commands/themeCommand.js';
 import { toolsCommand } from '../ui/commands/toolsCommand.js';
-import { settingsCommand } from '../ui/commands/settingsCommand.js';
 import { vimCommand } from '../ui/commands/vimCommand.js';
 import { setupGithubCommand } from '../ui/commands/setupGithubCommand.js';
-import { terminalSetupCommand } from '../ui/commands/terminalSetupCommand.js';
-import { agentsCommand } from '../ui/commands/agentsCommand.js';
 
 /**
  * Loads the core, hard-coded slash commands that are an integral part
@@ -55,6 +57,7 @@ export class BuiltinCommandLoader implements ICommandLoader {
     const allDefinitions: Array<SlashCommand | null> = [
       aboutCommand,
       agentsCommand,
+      approvalModeCommand,
       authCommand,
       bugCommand,
       chatCommand,
@@ -71,6 +74,7 @@ export class BuiltinCommandLoader implements ICommandLoader {
       initCommand,
       mcpCommand,
       memoryCommand,
+      modelCommand,
       privacyCommand,
       quitCommand,
       quitConfirmCommand,

packages/cli/src/test-utils/mockCommandContext.ts

@@ -35,7 +35,10 @@ export const createMockCommandContext = (
     },
     services: {
       config: null,
-      settings: { merged: {} } as LoadedSettings,
+      settings: {
+        merged: {},
+        setValue: vi.fn(),
+      } as unknown as LoadedSettings,
       git: undefined as GitService | undefined,
       logger: {
         log: vi.fn(),

packages/cli/src/ui/App.tsx

@@ -53,6 +53,17 @@ import { FolderTrustDialog } from './components/FolderTrustDialog.js';
 import { ShellConfirmationDialog } from './components/ShellConfirmationDialog.js';
 import { QuitConfirmationDialog } from './components/QuitConfirmationDialog.js';
 import { RadioButtonSelect } from './components/shared/RadioButtonSelect.js';
+import { ModelSelectionDialog } from './components/ModelSelectionDialog.js';
+import {
+  ModelSwitchDialog,
+  type VisionSwitchOutcome,
+} from './components/ModelSwitchDialog.js';
+import {
+  getOpenAIAvailableModelFromEnv,
+  getFilteredQwenModels,
+  type AvailableModel,
+} from './models/availableModels.js';
+import { processVisionSwitchOutcome } from './hooks/useVisionAutoSwitch.js';
 import {
   AgentCreationWizard,
   AgentsManagerDialog,
@@ -248,6 +259,20 @@ const App = ({ config, settings, startupWarnings = [], version }: AppProps) => {
     onWorkspaceMigrationDialogClose,
   } = useWorkspaceMigration(settings);
 
+  // Model selection dialog states
+  const [isModelSelectionDialogOpen, setIsModelSelectionDialogOpen] =
+    useState(false);
+  const [isVisionSwitchDialogOpen, setIsVisionSwitchDialogOpen] =
+    useState(false);
+  const [visionSwitchResolver, setVisionSwitchResolver] = useState<{
+    resolve: (result: {
+      modelOverride?: string;
+      persistSessionModel?: string;
+      showGuidance?: boolean;
+    }) => void;
+    reject: () => void;
+  } | null>(null);
+
   useEffect(() => {
     const unsubscribe = ideContext.subscribeToIdeContext(setIdeContextState);
     // Set the initial value
@@ -541,7 +566,9 @@ const App = ({ config, settings, startupWarnings = [], version }: AppProps) => {
       }
 
       // Switch model for future use but return false to stop current retry
-      config.setModel(fallbackModel);
+      config.setModel(fallbackModel).catch((error) => {
+        console.error('Failed to switch to fallback model:', error);
+      });
       config.setFallbackMode(true);
       logFlashFallback(
         config,
@@ -590,6 +617,86 @@ const App = ({ config, settings, startupWarnings = [], version }: AppProps) => {
     openAuthDialog();
   }, [openAuthDialog, setAuthError]);
 
+  // Vision switch handler for auto-switch functionality
+  const handleVisionSwitchRequired = useCallback(
+    async (_query: unknown) =>
+      new Promise<{
+        modelOverride?: string;
+        persistSessionModel?: string;
+        showGuidance?: boolean;
+      }>((resolve, reject) => {
+        setVisionSwitchResolver({ resolve, reject });
+        setIsVisionSwitchDialogOpen(true);
+      }),
+    [],
+  );
+
+  const handleVisionSwitchSelect = useCallback(
+    (outcome: VisionSwitchOutcome) => {
+      setIsVisionSwitchDialogOpen(false);
+      if (visionSwitchResolver) {
+        const result = processVisionSwitchOutcome(outcome);
+        visionSwitchResolver.resolve(result);
+        setVisionSwitchResolver(null);
+      }
+    },
+    [visionSwitchResolver],
+  );
+
+  const handleModelSelectionOpen = useCallback(() => {
+    setIsModelSelectionDialogOpen(true);
+  }, []);
+
+  const handleModelSelectionClose = useCallback(() => {
+    setIsModelSelectionDialogOpen(false);
+  }, []);
+
+  const handleModelSelect = useCallback(
+    async (modelId: string) => {
+      try {
+        await config.setModel(modelId);
+        setCurrentModel(modelId);
+        setIsModelSelectionDialogOpen(false);
+        addItem(
+          {
+            type: MessageType.INFO,
+            text: `Switched model to \`${modelId}\` for this session.`,
+          },
+          Date.now(),
+        );
+      } catch (error) {
+        console.error('Failed to switch model:', error);
+        addItem(
+          {
+            type: MessageType.ERROR,
+            text: `Failed to switch to model \`${modelId}\`. Please try again.`,
+          },
+          Date.now(),
+        );
+      }
+    },
+    [config, setCurrentModel, addItem],
+  );
+
+  const getAvailableModelsForCurrentAuth = useCallback((): AvailableModel[] => {
+    const contentGeneratorConfig = config.getContentGeneratorConfig();
+    if (!contentGeneratorConfig) return [];
+
+    const visionModelPreviewEnabled =
+      settings.merged.experimental?.visionModelPreview ?? true;
+
+    switch (contentGeneratorConfig.authType) {
+      case AuthType.QWEN_OAUTH:
+        return getFilteredQwenModels(visionModelPreviewEnabled);
+      case AuthType.USE_OPENAI: {
+        const openAIModel = getOpenAIAvailableModelFromEnv();
+        return openAIModel ? [openAIModel] : [];
+      }
+      default:
+        return [];
+    }
+  }, [config, settings.merged.experimental?.visionModelPreview]);
+
   // Core hooks and processors
   const {
     vimEnabled: vimModeEnabled,
@@ -620,6 +727,7 @@ const App = ({ config, settings, startupWarnings = [], version }: AppProps) => {
     setQuittingMessages,
     openPrivacyNotice,
     openSettingsDialog,
+    handleModelSelectionOpen,
     openSubagentCreateDialog,
     openAgentsManagerDialog,
     toggleVimEnabled,
@@ -664,10 +772,18 @@ const App = ({ config, settings, startupWarnings = [], version }: AppProps) => {
     setModelSwitchedFromQuotaError,
     refreshStatic,
     () => cancelHandlerRef.current(),
+    settings.merged.experimental?.visionModelPreview ?? true,
+    handleVisionSwitchRequired,
   );
 
   const pendingHistoryItems = useMemo(
-    () => [...pendingSlashCommandHistoryItems, ...pendingGeminiHistoryItems],
+    () =>
+      [...pendingSlashCommandHistoryItems, ...pendingGeminiHistoryItems].map(
+        (item, index) => ({
+          ...item,
+          id: index,
+        }),
+      ),
     [pendingSlashCommandHistoryItems, pendingGeminiHistoryItems],
   );
 
@@ -1028,6 +1144,8 @@ const App = ({ config, settings, startupWarnings = [], version }: AppProps) => {
       !isAuthDialogOpen &&
       !isThemeDialogOpen &&
       !isEditorDialogOpen &&
+      !isModelSelectionDialogOpen &&
+      !isVisionSwitchDialogOpen &&
       !isSubagentCreateDialogOpen &&
       !showPrivacyNotice &&
       !showWelcomeBackDialog &&
@@ -1049,6 +1167,8 @@ const App = ({ config, settings, startupWarnings = [], version }: AppProps) => {
     showWelcomeBackDialog,
     welcomeBackChoice,
     geminiClient,
+    isModelSelectionDialogOpen,
+    isVisionSwitchDialogOpen,
   ]);
 
   if (quittingMessages) {
@@ -1121,16 +1241,14 @@ const App = ({ config, settings, startupWarnings = [], version }: AppProps) => {
         </Static>
         <OverflowProvider>
           <Box ref={pendingHistoryItemRef} flexDirection="column">
-            {pendingHistoryItems.map((item, i) => (
+            {pendingHistoryItems.map((item) => (
               <HistoryItemDisplay
-                key={i}
+                key={item.id}
                 availableTerminalHeight={
                   constrainHeight ? availableTerminalHeight : undefined
                 }
                 terminalWidth={mainAreaWidth}
-                // TODO(taehykim): It seems like references to ids aren't necessary in
-                // HistoryItemDisplay. Refactor later. Use a fake id for now.
-                item={{ ...item, id: 0 }}
+                item={item}
                 isPending={true}
                 config={config}
                 isFocused={!isEditorDialogOpen}
@@ -1318,6 +1436,15 @@ const App = ({ config, settings, startupWarnings = [], version }: AppProps) => {
                 onExit={exitEditorDialog}
               />
             </Box>
+          ) : isModelSelectionDialogOpen ? (
+            <ModelSelectionDialog
+              availableModels={getAvailableModelsForCurrentAuth()}
+              currentModel={currentModel}
+              onSelect={handleModelSelect}
+              onCancel={handleModelSelectionClose}
+            />
+          ) : isVisionSwitchDialogOpen ? (
+            <ModelSwitchDialog onSelect={handleVisionSwitchSelect} />
           ) : showPrivacyNotice ? (
             <PrivacyNotice
               onExit={() => setShowPrivacyNotice(false)}

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

@@ -0,0 +1,495 @@
+/**
+ * @license
+ * Copyright 2025 Qwen
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
+import { approvalModeCommand } from './approvalModeCommand.js';
+import {
+  type CommandContext,
+  CommandKind,
+  type MessageActionReturn,
+} from './types.js';
+import { createMockCommandContext } from '../../test-utils/mockCommandContext.js';
+import { ApprovalMode } from '@qwen-code/qwen-code-core';
+import { SettingScope, type LoadedSettings } from '../../config/settings.js';
+
+describe('approvalModeCommand', () => {
+  let mockContext: CommandContext;
+  let setApprovalModeMock: ReturnType<typeof vi.fn>;
+  let setSettingsValueMock: ReturnType<typeof vi.fn>;
+  const originalEnv = { ...process.env };
+  const userSettingsPath = '/mock/user/settings.json';
+  const projectSettingsPath = '/mock/project/settings.json';
+  const userSettingsFile = { path: userSettingsPath, settings: {} };
+  const projectSettingsFile = { path: projectSettingsPath, settings: {} };
+
+  const getModeSubCommand = (mode: ApprovalMode) =>
+    approvalModeCommand.subCommands?.find((cmd) => cmd.name === mode);
+
+  const getScopeSubCommand = (
+    mode: ApprovalMode,
+    scope: '--session' | '--user' | '--project',
+  ) => getModeSubCommand(mode)?.subCommands?.find((cmd) => cmd.name === scope);
+
+  beforeEach(() => {
+    setApprovalModeMock = vi.fn();
+    setSettingsValueMock = vi.fn();
+
+    mockContext = createMockCommandContext({
+      services: {
+        config: {
+          getApprovalMode: vi.fn().mockReturnValue(ApprovalMode.DEFAULT),
+          setApprovalMode: setApprovalModeMock,
+        },
+        settings: {
+          merged: {},
+          setValue: setSettingsValueMock,
+          forScope: vi
+            .fn()
+            .mockImplementation((scope: SettingScope) =>
+              scope === SettingScope.User
+                ? userSettingsFile
+                : scope === SettingScope.Workspace
+                  ? projectSettingsFile
+                  : { path: '', settings: {} },
+            ),
+        } as unknown as LoadedSettings,
+      },
+    } as unknown as CommandContext);
+  });
+
+  afterEach(() => {
+    process.env = { ...originalEnv };
+    vi.clearAllMocks();
+  });
+
+  it('should have the correct command properties', () => {
+    expect(approvalModeCommand.name).toBe('approval-mode');
+    expect(approvalModeCommand.kind).toBe(CommandKind.BUILT_IN);
+    expect(approvalModeCommand.description).toBe(
+      'View or change the approval mode for tool usage',
+    );
+  });
+
+  it('should show current mode, options, and usage when no arguments provided', async () => {
+    if (!approvalModeCommand.action) {
+      throw new Error('approvalModeCommand must have an action.');
+    }
+
+    const result = (await approvalModeCommand.action(
+      mockContext,
+      '',
+    )) as MessageActionReturn;
+
+    expect(result.type).toBe('message');
+    expect(result.messageType).toBe('info');
+    const expectedMessage = [
+      'Current approval mode: default',
+      '',
+      'Available approval modes:',
+      '  - plan: Plan mode - Analyze only, do not modify files or execute commands',
+      '  - default: Default mode - Require approval for file edits or shell commands',
+      '  - auto-edit: Auto-edit mode - Automatically approve file edits',
+      '  - yolo: YOLO mode - Automatically approve all tools',
+      '',
+      'Usage: /approval-mode <mode> [--session|--user|--project]',
+    ].join('\n');
+    expect(result.content).toBe(expectedMessage);
+  });
+
+  it('should display error when config is not available', async () => {
+    if (!approvalModeCommand.action) {
+      throw new Error('approvalModeCommand must have an action.');
+    }
+
+    const nullConfigContext = createMockCommandContext({
+      services: {
+        config: null,
+      },
+    } as unknown as CommandContext);
+
+    const result = (await approvalModeCommand.action(
+      nullConfigContext,
+      '',
+    )) as MessageActionReturn;
+
+    expect(result.type).toBe('message');
+    expect(result.messageType).toBe('error');
+    expect(result.content).toBe('Configuration not available.');
+  });
+
+  it('should change approval mode when valid mode is provided', async () => {
+    if (!approvalModeCommand.action) {
+      throw new Error('approvalModeCommand must have an action.');
+    }
+
+    const result = (await approvalModeCommand.action(
+      mockContext,
+      'plan',
+    )) as MessageActionReturn;
+
+    expect(setApprovalModeMock).toHaveBeenCalledWith(ApprovalMode.PLAN);
+    expect(setSettingsValueMock).not.toHaveBeenCalled();
+    expect(result.type).toBe('message');
+    expect(result.messageType).toBe('info');
+    expect(result.content).toBe('Approval mode changed to: plan');
+  });
+
+  it('should accept canonical auto-edit mode value', async () => {
+    if (!approvalModeCommand.action) {
+      throw new Error('approvalModeCommand must have an action.');
+    }
+
+    const result = (await approvalModeCommand.action(
+      mockContext,
+      'auto-edit',
+    )) as MessageActionReturn;
+
+    expect(setApprovalModeMock).toHaveBeenCalledWith(ApprovalMode.AUTO_EDIT);
+    expect(setSettingsValueMock).not.toHaveBeenCalled();
+    expect(result.type).toBe('message');
+    expect(result.messageType).toBe('info');
+    expect(result.content).toBe('Approval mode changed to: auto-edit');
+  });
+
+  it('should accept auto-edit alias for compatibility', async () => {
+    if (!approvalModeCommand.action) {
+      throw new Error('approvalModeCommand must have an action.');
+    }
+
+    const result = (await approvalModeCommand.action(
+      mockContext,
+      'auto-edit',
+    )) as MessageActionReturn;
+
+    expect(setApprovalModeMock).toHaveBeenCalledWith(ApprovalMode.AUTO_EDIT);
+    expect(setSettingsValueMock).not.toHaveBeenCalled();
+    expect(result.content).toBe('Approval mode changed to: auto-edit');
+  });
+
+  it('should display error when invalid mode is provided', async () => {
+    if (!approvalModeCommand.action) {
+      throw new Error('approvalModeCommand must have an action.');
+    }
+
+    const result = (await approvalModeCommand.action(
+      mockContext,
+      'invalid',
+    )) as MessageActionReturn;
+
+    expect(result.type).toBe('message');
+    expect(result.messageType).toBe('error');
+    expect(result.content).toContain('Invalid approval mode: invalid');
+    expect(result.content).toContain('Available approval modes:');
+    expect(result.content).toContain(
+      'Usage: /approval-mode <mode> [--session|--user|--project]',
+    );
+  });
+
+  it('should display error when setApprovalMode throws an error', async () => {
+    if (!approvalModeCommand.action) {
+      throw new Error('approvalModeCommand must have an action.');
+    }
+
+    const errorMessage = 'Failed to set approval mode';
+    mockContext.services.config!.setApprovalMode = vi
+      .fn()
+      .mockImplementation(() => {
+        throw new Error(errorMessage);
+      });
+
+    const result = (await approvalModeCommand.action(
+      mockContext,
+      'plan',
+    )) as MessageActionReturn;
+
+    expect(result.type).toBe('message');
+    expect(result.messageType).toBe('error');
+    expect(result.content).toBe(
+      `Failed to change approval mode: ${errorMessage}`,
+    );
+  });
+
+  it('should allow selecting auto-edit with user scope via nested subcommands', async () => {
+    if (!approvalModeCommand.subCommands) {
+      throw new Error('approvalModeCommand must have subCommands.');
+    }
+
+    const userSubCommand = getScopeSubCommand(ApprovalMode.AUTO_EDIT, '--user');
+    if (!userSubCommand?.action) {
+      throw new Error('--user scope subcommand must have an action.');
+    }
+
+    const result = (await userSubCommand.action(
+      mockContext,
+      '',
+    )) as MessageActionReturn;
+
+    expect(setApprovalModeMock).toHaveBeenCalledWith(ApprovalMode.AUTO_EDIT);
+    expect(setSettingsValueMock).toHaveBeenCalledWith(
+      SettingScope.User,
+      'approvalMode',
+      'auto-edit',
+    );
+    expect(result.content).toBe(
+      `Approval mode changed to: auto-edit (saved to user settings at ${userSettingsPath})`,
+    );
+  });
+
+  it('should allow selecting plan with project scope via nested subcommands', async () => {
+    if (!approvalModeCommand.subCommands) {
+      throw new Error('approvalModeCommand must have subCommands.');
+    }
+
+    const projectSubCommand = getScopeSubCommand(
+      ApprovalMode.PLAN,
+      '--project',
+    );
+    if (!projectSubCommand?.action) {
+      throw new Error('--project scope subcommand must have an action.');
+    }
+
+    const result = (await projectSubCommand.action(
+      mockContext,
+      '',
+    )) as MessageActionReturn;
+
+    expect(setApprovalModeMock).toHaveBeenCalledWith(ApprovalMode.PLAN);
+    expect(setSettingsValueMock).toHaveBeenCalledWith(
+      SettingScope.Workspace,
+      'approvalMode',
+      'plan',
+    );
+    expect(result.content).toBe(
+      `Approval mode changed to: plan (saved to project settings at ${projectSettingsPath})`,
+    );
+  });
+
+  it('should allow selecting plan with session scope via nested subcommands', async () => {
+    if (!approvalModeCommand.subCommands) {
+      throw new Error('approvalModeCommand must have subCommands.');
+    }
+
+    const sessionSubCommand = getScopeSubCommand(
+      ApprovalMode.PLAN,
+      '--session',
+    );
+    if (!sessionSubCommand?.action) {
+      throw new Error('--session scope subcommand must have an action.');
+    }
+
+    const result = (await sessionSubCommand.action(
+      mockContext,
+      '',
+    )) as MessageActionReturn;
+
+    expect(setApprovalModeMock).toHaveBeenCalledWith(ApprovalMode.PLAN);
+    expect(setSettingsValueMock).not.toHaveBeenCalled();
+    expect(result.content).toBe('Approval mode changed to: plan');
+  });
+
+  it('should allow providing a scope argument after selecting a mode subcommand', async () => {
+    if (!approvalModeCommand.subCommands) {
+      throw new Error('approvalModeCommand must have subCommands.');
+    }
+
+    const planSubCommand = getModeSubCommand(ApprovalMode.PLAN);
+    if (!planSubCommand?.action) {
+      throw new Error('plan subcommand must have an action.');
+    }
+
+    const result = (await planSubCommand.action(
+      mockContext,
+      '--user',
+    )) as MessageActionReturn;
+
+    expect(setApprovalModeMock).toHaveBeenCalledWith(ApprovalMode.PLAN);
+    expect(setSettingsValueMock).toHaveBeenCalledWith(
+      SettingScope.User,
+      'approvalMode',
+      'plan',
+    );
+    expect(result.content).toBe(
+      `Approval mode changed to: plan (saved to user settings at ${userSettingsPath})`,
+    );
+  });
+
+  it('should support --user plan pattern (scope first)', async () => {
+    if (!approvalModeCommand.action) {
+      throw new Error('approvalModeCommand must have an action.');
+    }
+
+    const result = (await approvalModeCommand.action(
+      mockContext,
+      '--user plan',
+    )) as MessageActionReturn;
+
+    expect(setApprovalModeMock).toHaveBeenCalledWith(ApprovalMode.PLAN);
+    expect(setSettingsValueMock).toHaveBeenCalledWith(
+      SettingScope.User,
+      'approvalMode',
+      'plan',
+    );
+    expect(result.content).toBe(
+      `Approval mode changed to: plan (saved to user settings at ${userSettingsPath})`,
+    );
+  });
+
+  it('should support plan --user pattern (mode first)', async () => {
+    if (!approvalModeCommand.action) {
+      throw new Error('approvalModeCommand must have an action.');
+    }
+
+    const result = (await approvalModeCommand.action(
+      mockContext,
+      'plan --user',
+    )) as MessageActionReturn;
+
+    expect(setApprovalModeMock).toHaveBeenCalledWith(ApprovalMode.PLAN);
+    expect(setSettingsValueMock).toHaveBeenCalledWith(
+      SettingScope.User,
+      'approvalMode',
+      'plan',
+    );
+    expect(result.content).toBe(
+      `Approval mode changed to: plan (saved to user settings at ${userSettingsPath})`,
+    );
+  });
+
+  it('should support --project auto-edit pattern', async () => {
+    if (!approvalModeCommand.action) {
+      throw new Error('approvalModeCommand must have an action.');
+    }
+
+    const result = (await approvalModeCommand.action(
+      mockContext,
+      '--project auto-edit',
+    )) as MessageActionReturn;
+
+    expect(setApprovalModeMock).toHaveBeenCalledWith(ApprovalMode.AUTO_EDIT);
+    expect(setSettingsValueMock).toHaveBeenCalledWith(
+      SettingScope.Workspace,
+      'approvalMode',
+      'auto-edit',
+    );
+    expect(result.content).toBe(
+      `Approval mode changed to: auto-edit (saved to project settings at ${projectSettingsPath})`,
+    );
+  });
+
+  it('should display error when only scope flag is provided', async () => {
+    if (!approvalModeCommand.action) {
+      throw new Error('approvalModeCommand must have an action.');
+    }
+
+    const result = (await approvalModeCommand.action(
+      mockContext,
+      '--user',
+    )) as MessageActionReturn;
+
+    expect(result.type).toBe('message');
+    expect(result.messageType).toBe('error');
+    expect(result.content).toContain('Missing approval mode');
+    expect(setApprovalModeMock).not.toHaveBeenCalled();
+    expect(setSettingsValueMock).not.toHaveBeenCalled();
+  });
+
+  it('should display error when multiple scope flags are provided', async () => {
+    if (!approvalModeCommand.action) {
+      throw new Error('approvalModeCommand must have an action.');
+    }
+
+    const result = (await approvalModeCommand.action(
+      mockContext,
+      '--user --project plan',
+    )) as MessageActionReturn;
+
+    expect(result.type).toBe('message');
+    expect(result.messageType).toBe('error');
+    expect(result.content).toContain('Multiple scope flags provided');
+    expect(setApprovalModeMock).not.toHaveBeenCalled();
+    expect(setSettingsValueMock).not.toHaveBeenCalled();
+  });
+
+  it('should surface a helpful error when scope subcommands receive extra arguments', async () => {
+    if (!approvalModeCommand.subCommands) {
+      throw new Error('approvalModeCommand must have subCommands.');
+    }
+
+    const userSubCommand = getScopeSubCommand(ApprovalMode.DEFAULT, '--user');
+    if (!userSubCommand?.action) {
+      throw new Error('--user scope subcommand must have an action.');
+    }
+
+    const result = (await userSubCommand.action(
+      mockContext,
+      'extra',
+    )) as MessageActionReturn;
+
+    expect(result.type).toBe('message');
+    expect(result.messageType).toBe('error');
+    expect(result.content).toBe(
+      'Scope subcommands do not accept additional arguments.',
+    );
+    expect(setApprovalModeMock).not.toHaveBeenCalled();
+    expect(setSettingsValueMock).not.toHaveBeenCalled();
+  });
+
+  it('should provide completion for approval modes', async () => {
+    if (!approvalModeCommand.completion) {
+      throw new Error('approvalModeCommand must have a completion function.');
+    }
+
+    // Test partial mode completion
+    const result = await approvalModeCommand.completion(mockContext, 'p');
+    expect(result).toEqual(['plan']);
+
+    const result2 = await approvalModeCommand.completion(mockContext, 'a');
+    expect(result2).toEqual(['auto-edit']);
+
+    // Test empty completion - should suggest available modes first
+    const result3 = await approvalModeCommand.completion(mockContext, '');
+    expect(result3).toEqual(['plan', 'default', 'auto-edit', 'yolo']);
+
+    const result4 = await approvalModeCommand.completion(mockContext, 'AUTO');
+    expect(result4).toEqual(['auto-edit']);
+
+    // Test mode first pattern: 'plan ' should suggest scope flags
+    const result5 = await approvalModeCommand.completion(mockContext, 'plan ');
+    expect(result5).toEqual(['--session', '--project', '--user']);
+
+    const result6 = await approvalModeCommand.completion(
+      mockContext,
+      'plan --u',
+    );
+    expect(result6).toEqual(['--user']);
+
+    // Test scope first pattern: '--user ' should suggest modes
+    const result7 = await approvalModeCommand.completion(
+      mockContext,
+      '--user ',
+    );
+    expect(result7).toEqual(['plan', 'default', 'auto-edit', 'yolo']);
+
+    const result8 = await approvalModeCommand.completion(
+      mockContext,
+      '--user p',
+    );
+    expect(result8).toEqual(['plan']);
+
+    // Test completed patterns should return empty
+    const result9 = await approvalModeCommand.completion(
+      mockContext,
+      'plan --user ',
+    );
+    expect(result9).toEqual([]);
+
+    const result10 = await approvalModeCommand.completion(
+      mockContext,
+      '--user plan ',
+    );
+    expect(result10).toEqual([]);
+  });
+});

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

@@ -0,0 +1,434 @@
+/**
+ * @license
+ * Copyright 2025 Qwen
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import type {
+  SlashCommand,
+  CommandContext,
+  MessageActionReturn,
+} from './types.js';
+import { CommandKind } from './types.js';
+import { ApprovalMode, APPROVAL_MODES } from '@qwen-code/qwen-code-core';
+import { SettingScope } from '../../config/settings.js';
+
+const USAGE_MESSAGE =
+  'Usage: /approval-mode <mode> [--session|--user|--project]';
+
+const normalizeInputMode = (value: string): string =>
+  value.trim().toLowerCase();
+
+const tokenizeArgs = (args: string): string[] => {
+  const matches = args.match(/(?:"[^"]*"|'[^']*'|[^\s"']+)/g);
+  if (!matches) {
+    return [];
+  }
+
+  return matches.map((token) => {
+    if (
+      (token.startsWith('"') && token.endsWith('"')) ||
+      (token.startsWith("'") && token.endsWith("'"))
+    ) {
+      return token.slice(1, -1);
+    }
+    return token;
+  });
+};
+
+const parseApprovalMode = (value: string | null): ApprovalMode | null => {
+  if (!value) {
+    return null;
+  }
+
+  const normalized = normalizeInputMode(value).replace(/_/g, '-');
+  const matchIndex = APPROVAL_MODES.findIndex(
+    (candidate) => candidate === normalized,
+  );
+
+  return matchIndex === -1 ? null : APPROVAL_MODES[matchIndex];
+};
+
+const formatModeDescription = (mode: ApprovalMode): string => {
+  switch (mode) {
+    case ApprovalMode.PLAN:
+      return 'Plan mode - Analyze only, do not modify files or execute commands';
+    case ApprovalMode.DEFAULT:
+      return 'Default mode - Require approval for file edits or shell commands';
+    case ApprovalMode.AUTO_EDIT:
+      return 'Auto-edit mode - Automatically approve file edits';
+    case ApprovalMode.YOLO:
+      return 'YOLO mode - Automatically approve all tools';
+    default:
+      return `${mode} mode`;
+  }
+};
+
+const parseApprovalArgs = (
+  args: string,
+): {
+  mode: string | null;
+  scope: 'session' | 'user' | 'project';
+  error?: string;
+} => {
+  const trimmedArgs = args.trim();
+  if (!trimmedArgs) {
+    return { mode: null, scope: 'session' };
+  }
+
+  const tokens = tokenizeArgs(trimmedArgs);
+  let mode: string | null = null;
+  let scope: 'session' | 'user' | 'project' = 'session';
+  let scopeFlag: string | null = null;
+
+  // Find scope flag and mode
+  for (const token of tokens) {
+    if (token === '--session' || token === '--user' || token === '--project') {
+      if (scopeFlag) {
+        return {
+          mode: null,
+          scope: 'session',
+          error: 'Multiple scope flags provided',
+        };
+      }
+      scopeFlag = token;
+      scope = token.substring(2) as 'session' | 'user' | 'project';
+    } else if (!mode) {
+      mode = token;
+    } else {
+      return {
+        mode: null,
+        scope: 'session',
+        error: 'Invalid arguments provided',
+      };
+    }
+  }
+
+  if (!mode) {
+    return { mode: null, scope: 'session', error: 'Missing approval mode' };
+  }
+
+  return { mode, scope };
+};
+
+const setApprovalModeWithScope = async (
+  context: CommandContext,
+  mode: ApprovalMode,
+  scope: 'session' | 'user' | 'project',
+): Promise<MessageActionReturn> => {
+  const { services } = context;
+  const { config } = services;
+
+  if (!config) {
+    return {
+      type: 'message',
+      messageType: 'error',
+      content: 'Configuration not available.',
+    };
+  }
+
+  try {
+    // Always set the mode in the current session
+    config.setApprovalMode(mode);
+
+    // If scope is not session, also persist to settings
+    if (scope !== 'session') {
+      const { settings } = context.services;
+      if (!settings || typeof settings.setValue !== 'function') {
+        return {
+          type: 'message',
+          messageType: 'error',
+          content:
+            'Settings service is not available; unable to persist the approval mode.',
+        };
+      }
+
+      const settingScope =
+        scope === 'user' ? SettingScope.User : SettingScope.Workspace;
+      const scopeLabel = scope === 'user' ? 'user' : 'project';
+      let settingsPath: string | undefined;
+
+      try {
+        if (typeof settings.forScope === 'function') {
+          settingsPath = settings.forScope(settingScope)?.path;
+        }
+      } catch (_error) {
+        settingsPath = undefined;
+      }
+
+      try {
+        settings.setValue(settingScope, 'approvalMode', mode);
+      } catch (error) {
+        return {
+          type: 'message',
+          messageType: 'error',
+          content: `Failed to save approval mode: ${(error as Error).message}`,
+        };
+      }
+
+      const locationSuffix = settingsPath ? ` at ${settingsPath}` : '';
+
+      const scopeSuffix = ` (saved to ${scopeLabel} settings${locationSuffix})`;
+
+      return {
+        type: 'message',
+        messageType: 'info',
+        content: `Approval mode changed to: ${mode}${scopeSuffix}`,
+      };
+    }
+
+    return {
+      type: 'message',
+      messageType: 'info',
+      content: `Approval mode changed to: ${mode}`,
+    };
+  } catch (error) {
+    return {
+      type: 'message',
+      messageType: 'error',
+      content: `Failed to change approval mode: ${(error as Error).message}`,
+    };
+  }
+};
+
+export const approvalModeCommand: SlashCommand = {
+  name: 'approval-mode',
+  description: 'View or change the approval mode for tool usage',
+  kind: CommandKind.BUILT_IN,
+  action: async (
+    context: CommandContext,
+    args: string,
+  ): Promise<MessageActionReturn> => {
+    const { config } = context.services;
+    if (!config) {
+      return {
+        type: 'message',
+        messageType: 'error',
+        content: 'Configuration not available.',
+      };
+    }
+
+    // If no arguments provided, show current mode and available options
+    if (!args || args.trim() === '') {
+      const currentMode =
+        typeof config.getApprovalMode === 'function'
+          ? config.getApprovalMode()
+          : null;
+
+      const messageLines: string[] = [];
+
+      if (currentMode) {
+        messageLines.push(`Current approval mode: ${currentMode}`);
+        messageLines.push('');
+      }
+
+      messageLines.push('Available approval modes:');
+      for (const mode of APPROVAL_MODES) {
+        messageLines.push(`  - ${mode}: ${formatModeDescription(mode)}`);
+      }
+      messageLines.push('');
+      messageLines.push(USAGE_MESSAGE);
+
+      return {
+        type: 'message',
+        messageType: 'info',
+        content: messageLines.join('\n'),
+      };
+    }
+
+    // Parse arguments flexibly
+    const parsed = parseApprovalArgs(args);
+
+    if (parsed.error) {
+      return {
+        type: 'message',
+        messageType: 'error',
+        content: `${parsed.error}. ${USAGE_MESSAGE}`,
+      };
+    }
+
+    if (!parsed.mode) {
+      return {
+        type: 'message',
+        messageType: 'info',
+        content: USAGE_MESSAGE,
+      };
+    }
+
+    const requestedMode = parseApprovalMode(parsed.mode);
+
+    if (!requestedMode) {
+      let message = `Invalid approval mode: ${parsed.mode}\n\n`;
+      message += 'Available approval modes:\n';
+      for (const mode of APPROVAL_MODES) {
+        message += `  - ${mode}: ${formatModeDescription(mode)}\n`;
+      }
+      message += `\n${USAGE_MESSAGE}`;
+      return {
+        type: 'message',
+        messageType: 'error',
+        content: message,
+      };
+    }
+
+    return setApprovalModeWithScope(context, requestedMode, parsed.scope);
+  },
+  subCommands: APPROVAL_MODES.map((mode) => ({
+    name: mode,
+    description: formatModeDescription(mode),
+    kind: CommandKind.BUILT_IN,
+    subCommands: [
+      {
+        name: '--session',
+        description: 'Apply to current session only (temporary)',
+        kind: CommandKind.BUILT_IN,
+        action: async (
+          context: CommandContext,
+          args: string,
+        ): Promise<MessageActionReturn> => {
+          if (args.trim().length > 0) {
+            return {
+              type: 'message',
+              messageType: 'error',
+              content: 'Scope subcommands do not accept additional arguments.',
+            };
+          }
+          return setApprovalModeWithScope(context, mode, 'session');
+        },
+      },
+      {
+        name: '--project',
+        description: 'Persist for this project/workspace',
+        kind: CommandKind.BUILT_IN,
+        action: async (
+          context: CommandContext,
+          args: string,
+        ): Promise<MessageActionReturn> => {
+          if (args.trim().length > 0) {
+            return {
+              type: 'message',
+              messageType: 'error',
+              content: 'Scope subcommands do not accept additional arguments.',
+            };
+          }
+          return setApprovalModeWithScope(context, mode, 'project');
+        },
+      },
+      {
+        name: '--user',
+        description: 'Persist for this user on this machine',
+        kind: CommandKind.BUILT_IN,
+        action: async (
+          context: CommandContext,
+          args: string,
+        ): Promise<MessageActionReturn> => {
+          if (args.trim().length > 0) {
+            return {
+              type: 'message',
+              messageType: 'error',
+              content: 'Scope subcommands do not accept additional arguments.',
+            };
+          }
+          return setApprovalModeWithScope(context, mode, 'user');
+        },
+      },
+    ],
+    action: async (
+      context: CommandContext,
+      args: string,
+    ): Promise<MessageActionReturn> => {
+      if (args.trim().length > 0) {
+        // Allow users who type `/approval-mode plan --user` via the subcommand path
+        const parsed = parseApprovalArgs(`${mode} ${args}`);
+        if (parsed.error) {
+          return {
+            type: 'message',
+            messageType: 'error',
+            content: `${parsed.error}. ${USAGE_MESSAGE}`,
+          };
+        }
+
+        const normalizedMode = parseApprovalMode(parsed.mode);
+        if (!normalizedMode) {
+          return {
+            type: 'message',
+            messageType: 'error',
+            content: `Invalid approval mode: ${parsed.mode}. ${USAGE_MESSAGE}`,
+          };
+        }
+
+        return setApprovalModeWithScope(context, normalizedMode, parsed.scope);
+      }
+
+      return setApprovalModeWithScope(context, mode, 'session');
+    },
+  })),
+  completion: async (_context: CommandContext, partialArg: string) => {
+    const tokens = tokenizeArgs(partialArg);
+    const hasTrailingSpace = /\s$/.test(partialArg);
+    const currentSegment = hasTrailingSpace
+      ? ''
+      : tokens.length > 0
+        ? tokens[tokens.length - 1]
+        : '';
+
+    const normalizedCurrent = normalizeInputMode(currentSegment).replace(
+      /_/g,
+      '-',
+    );
+
+    const scopeValues = ['--session', '--project', '--user'];
+
+    const normalizeToken = (token: string) =>
+      normalizeInputMode(token).replace(/_/g, '-');
+
+    const normalizedTokens = tokens.map(normalizeToken);
+
+    if (tokens.length === 0) {
+      if (currentSegment.startsWith('-')) {
+        return scopeValues.filter((scope) => scope.startsWith(currentSegment));
+      }
+      return APPROVAL_MODES;
+    }
+
+    if (tokens.length === 1 && !hasTrailingSpace) {
+      const originalToken = tokens[0];
+      if (originalToken.startsWith('-')) {
+        return scopeValues.filter((scope) =>
+          scope.startsWith(normalizedCurrent),
+        );
+      }
+      return APPROVAL_MODES.filter((mode) =>
+        mode.startsWith(normalizedCurrent),
+      );
+    }
+
+    if (tokens.length === 1 && hasTrailingSpace) {
+      const normalizedFirst = normalizedTokens[0];
+      if (scopeValues.includes(tokens[0])) {
+        return APPROVAL_MODES;
+      }
+      if (APPROVAL_MODES.includes(normalizedFirst as ApprovalMode)) {
+        return scopeValues;
+      }
+      return APPROVAL_MODES;
+    }
+
+    if (tokens.length === 2 && !hasTrailingSpace) {
+      const normalizedFirst = normalizedTokens[0];
+      if (scopeValues.includes(tokens[0])) {
+        return APPROVAL_MODES.filter((mode) =>
+          mode.startsWith(normalizedCurrent),
+        );
+      }
+      if (APPROVAL_MODES.includes(normalizedFirst as ApprovalMode)) {
+        return scopeValues.filter((scope) =>
+          scope.startsWith(normalizedCurrent),
+        );
+      }
+      return [];
+    }
+
+    return [];
+  },
+};

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

@@ -0,0 +1,179 @@
+/**
+ * @license
+ * Copyright 2025 Qwen
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import { describe, it, expect, beforeEach, vi } from 'vitest';
+import { modelCommand } from './modelCommand.js';
+import { type CommandContext } from './types.js';
+import { createMockCommandContext } from '../../test-utils/mockCommandContext.js';
+import {
+  AuthType,
+  type ContentGeneratorConfig,
+  type Config,
+} from '@qwen-code/qwen-code-core';
+import * as availableModelsModule from '../models/availableModels.js';
+
+// Mock the availableModels module
+vi.mock('../models/availableModels.js', () => ({
+  AVAILABLE_MODELS_QWEN: [
+    { id: 'qwen3-coder-plus', label: 'qwen3-coder-plus' },
+    { id: 'qwen-vl-max-latest', label: 'qwen-vl-max', isVision: true },
+  ],
+  getOpenAIAvailableModelFromEnv: vi.fn(),
+}));
+
+// Helper function to create a mock config
+function createMockConfig(
+  contentGeneratorConfig: ContentGeneratorConfig | null,
+): Partial<Config> {
+  return {
+    getContentGeneratorConfig: vi.fn().mockReturnValue(contentGeneratorConfig),
+  };
+}
+
+describe('modelCommand', () => {
+  let mockContext: CommandContext;
+  const mockGetOpenAIAvailableModelFromEnv = vi.mocked(
+    availableModelsModule.getOpenAIAvailableModelFromEnv,
+  );
+
+  beforeEach(() => {
+    mockContext = createMockCommandContext();
+    vi.clearAllMocks();
+  });
+
+  it('should have the correct name and description', () => {
+    expect(modelCommand.name).toBe('model');
+    expect(modelCommand.description).toBe('Switch the model for this session');
+  });
+
+  it('should return error when config is not available', async () => {
+    mockContext.services.config = null;
+
+    const result = await modelCommand.action!(mockContext, '');
+
+    expect(result).toEqual({
+      type: 'message',
+      messageType: 'error',
+      content: 'Configuration not available.',
+    });
+  });
+
+  it('should return error when content generator config is not available', async () => {
+    const mockConfig = createMockConfig(null);
+    mockContext.services.config = mockConfig as Config;
+
+    const result = await modelCommand.action!(mockContext, '');
+
+    expect(result).toEqual({
+      type: 'message',
+      messageType: 'error',
+      content: 'Content generator configuration not available.',
+    });
+  });
+
+  it('should return error when auth type is not available', async () => {
+    const mockConfig = createMockConfig({
+      model: 'test-model',
+      authType: undefined,
+    });
+    mockContext.services.config = mockConfig as Config;
+
+    const result = await modelCommand.action!(mockContext, '');
+
+    expect(result).toEqual({
+      type: 'message',
+      messageType: 'error',
+      content: 'Authentication type not available.',
+    });
+  });
+
+  it('should return dialog action for QWEN_OAUTH auth type', async () => {
+    const mockConfig = createMockConfig({
+      model: 'test-model',
+      authType: AuthType.QWEN_OAUTH,
+    });
+    mockContext.services.config = mockConfig as Config;
+
+    const result = await modelCommand.action!(mockContext, '');
+
+    expect(result).toEqual({
+      type: 'dialog',
+      dialog: 'model',
+    });
+  });
+
+  it('should return dialog action for USE_OPENAI auth type when model is available', async () => {
+    mockGetOpenAIAvailableModelFromEnv.mockReturnValue({
+      id: 'gpt-4',
+      label: 'gpt-4',
+    });
+
+    const mockConfig = createMockConfig({
+      model: 'test-model',
+      authType: AuthType.USE_OPENAI,
+    });
+    mockContext.services.config = mockConfig as Config;
+
+    const result = await modelCommand.action!(mockContext, '');
+
+    expect(result).toEqual({
+      type: 'dialog',
+      dialog: 'model',
+    });
+  });
+
+  it('should return error for USE_OPENAI auth type when no model is available', async () => {
+    mockGetOpenAIAvailableModelFromEnv.mockReturnValue(null);
+
+    const mockConfig = createMockConfig({
+      model: 'test-model',
+      authType: AuthType.USE_OPENAI,
+    });
+    mockContext.services.config = mockConfig as Config;
+
+    const result = await modelCommand.action!(mockContext, '');
+
+    expect(result).toEqual({
+      type: 'message',
+      messageType: 'error',
+      content:
+        'No models available for the current authentication type (openai).',
+    });
+  });
+
+  it('should return error for unsupported auth types', async () => {
+    const mockConfig = createMockConfig({
+      model: 'test-model',
+      authType: 'UNSUPPORTED_AUTH_TYPE' as AuthType,
+    });
+    mockContext.services.config = mockConfig as Config;
+
+    const result = await modelCommand.action!(mockContext, '');
+
+    expect(result).toEqual({
+      type: 'message',
+      messageType: 'error',
+      content:
+        'No models available for the current authentication type (UNSUPPORTED_AUTH_TYPE).',
+    });
+  });
+
+  it('should handle undefined auth type', async () => {
+    const mockConfig = createMockConfig({
+      model: 'test-model',
+      authType: undefined,
+    });
+    mockContext.services.config = mockConfig as Config;
+
+    const result = await modelCommand.action!(mockContext, '');
+
+    expect(result).toEqual({
+      type: 'message',
+      messageType: 'error',
+      content: 'Authentication type not available.',
+    });
+  });
+});

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

@@ -0,0 +1,88 @@
+/**
+ * @license
+ * Copyright 2025 Qwen
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import { AuthType } from '@qwen-code/qwen-code-core';
+import type {
+  SlashCommand,
+  CommandContext,
+  OpenDialogActionReturn,
+  MessageActionReturn,
+} from './types.js';
+import { CommandKind } from './types.js';
+import {
+  AVAILABLE_MODELS_QWEN,
+  getOpenAIAvailableModelFromEnv,
+  type AvailableModel,
+} from '../models/availableModels.js';
+
+function getAvailableModelsForAuthType(authType: AuthType): AvailableModel[] {
+  switch (authType) {
+    case AuthType.QWEN_OAUTH:
+      return AVAILABLE_MODELS_QWEN;
+    case AuthType.USE_OPENAI: {
+      const openAIModel = getOpenAIAvailableModelFromEnv();
+      return openAIModel ? [openAIModel] : [];
+    }
+    default:
+      // For other auth types, return empty array for now
+      // This can be expanded later according to the design doc
+      return [];
+  }
+}
+
+export const modelCommand: SlashCommand = {
+  name: 'model',
+  description: 'Switch the model for this session',
+  kind: CommandKind.BUILT_IN,
+  action: async (
+    context: CommandContext,
+  ): Promise<OpenDialogActionReturn | MessageActionReturn> => {
+    const { services } = context;
+    const { config } = services;
+
+    if (!config) {
+      return {
+        type: 'message',
+        messageType: 'error',
+        content: 'Configuration not available.',
+      };
+    }
+
+    const contentGeneratorConfig = config.getContentGeneratorConfig();
+    if (!contentGeneratorConfig) {
+      return {
+        type: 'message',
+        messageType: 'error',
+        content: 'Content generator configuration not available.',
+      };
+    }
+
+    const authType = contentGeneratorConfig.authType;
+    if (!authType) {
+      return {
+        type: 'message',
+        messageType: 'error',
+        content: 'Authentication type not available.',
+      };
+    }
+
+    const availableModels = getAvailableModelsForAuthType(authType);
+
+    if (availableModels.length === 0) {
+      return {
+        type: 'message',
+        messageType: 'error',
+        content: `No models available for the current authentication type (${authType}).`,
+      };
+    }
+
+    // Trigger model selection dialog
+    return {
+      type: 'dialog',
+      dialog: 'model',
+    };
+  },
+};

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

@@ -116,6 +116,7 @@ export interface OpenDialogActionReturn {
     | 'editor'
     | 'privacy'
     | 'settings'
+    | 'model'
     | 'subagent_create'
     | 'subagent_list';
 }

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

@@ -21,15 +21,20 @@ export const AutoAcceptIndicator: React.FC<AutoAcceptIndicatorProps> = ({
   let subText = '';
 
   switch (approvalMode) {
+    case ApprovalMode.PLAN:
+      textColor = Colors.AccentBlue;
+      textContent = 'plan mode';
+      subText = ' (shift + tab to cycle)';
+      break;
     case ApprovalMode.AUTO_EDIT:
       textColor = Colors.AccentGreen;
-      textContent = 'accepting edits';
-      subText = ' (shift + tab to toggle)';
+      textContent = 'auto-accept edits';
+      subText = ' (shift + tab to cycle)';
       break;
     case ApprovalMode.YOLO:
       textColor = Colors.AccentRed;
       textContent = 'YOLO mode';
-      subText = ' (ctrl + y to toggle)';
+      subText = ' (shift + tab to cycle)';
       break;
     case ApprovalMode.DEFAULT:
     default:

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

@@ -133,12 +133,6 @@ export const Help: React.FC<Help> = ({ commands }) => (
       </Text>{' '}
       - Open input in external editor
     </Text>
-    <Text color={Colors.Foreground}>
-      <Text bold color={Colors.AccentPurple}>
-        Ctrl+Y
-      </Text>{' '}
-      - Toggle YOLO mode
-    </Text>
     <Text color={Colors.Foreground}>
       <Text bold color={Colors.AccentPurple}>
         Enter
@@ -155,7 +149,7 @@ export const Help: React.FC<Help> = ({ commands }) => (
       <Text bold color={Colors.AccentPurple}>
         Shift+Tab
       </Text>{' '}
-      - Toggle auto-accepting edits
+      - Cycle approval modes
     </Text>
     <Text color={Colors.Foreground}>
       <Text bold color={Colors.AccentPurple}>

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

@@ -5,6 +5,7 @@
  */
 
 import type React from 'react';
+import { memo } from 'react';
 import type { HistoryItem } from '../types.js';
 import { UserMessage } from './messages/UserMessage.js';
 import { UserShellMessage } from './messages/UserShellMessage.js';
@@ -35,7 +36,7 @@ interface HistoryItemDisplayProps {
   commands?: readonly SlashCommand[];
 }
 
-export const HistoryItemDisplay: React.FC<HistoryItemDisplayProps> = ({
+const HistoryItemDisplayComponent: React.FC<HistoryItemDisplayProps> = ({
   item,
   availableTerminalHeight,
   terminalWidth,
@@ -101,3 +102,7 @@ export const HistoryItemDisplay: React.FC<HistoryItemDisplayProps> = ({
     {item.type === 'summary' && <SummaryMessage summary={item.summary} />}
   </Box>
 );
+
+HistoryItemDisplayComponent.displayName = 'HistoryItemDisplay';
+
+export const HistoryItemDisplay = memo(HistoryItemDisplayComponent);

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

@@ -0,0 +1,246 @@
+/**
+ * @license
+ * Copyright 2025 Qwen
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import React from 'react';
+import { render } from 'ink-testing-library';
+import { describe, it, expect, vi, beforeEach } from 'vitest';
+import { ModelSelectionDialog } from './ModelSelectionDialog.js';
+import type { AvailableModel } from '../models/availableModels.js';
+import type { RadioSelectItem } from './shared/RadioButtonSelect.js';
+
+// Mock the useKeypress hook
+const mockUseKeypress = vi.hoisted(() => vi.fn());
+vi.mock('../hooks/useKeypress.js', () => ({
+  useKeypress: mockUseKeypress,
+}));
+
+// Mock the RadioButtonSelect component
+const mockRadioButtonSelect = vi.hoisted(() => vi.fn());
+vi.mock('./shared/RadioButtonSelect.js', () => ({
+  RadioButtonSelect: mockRadioButtonSelect,
+}));
+
+describe('ModelSelectionDialog', () => {
+  const mockAvailableModels: AvailableModel[] = [
+    { id: 'qwen3-coder-plus', label: 'qwen3-coder-plus' },
+    { id: 'qwen-vl-max-latest', label: 'qwen-vl-max', isVision: true },
+    { id: 'gpt-4', label: 'GPT-4' },
+  ];
+
+  const mockOnSelect = vi.fn();
+  const mockOnCancel = vi.fn();
+
+  beforeEach(() => {
+    vi.clearAllMocks();
+
+    // Mock RadioButtonSelect to return a simple div
+    mockRadioButtonSelect.mockReturnValue(
+      React.createElement('div', { 'data-testid': 'radio-select' }),
+    );
+  });
+
+  it('should setup escape key handler to call onCancel', () => {
+    render(
+      <ModelSelectionDialog
+        availableModels={mockAvailableModels}
+        currentModel="qwen3-coder-plus"
+        onSelect={mockOnSelect}
+        onCancel={mockOnCancel}
+      />,
+    );
+
+    expect(mockUseKeypress).toHaveBeenCalledWith(expect.any(Function), {
+      isActive: true,
+    });
+
+    // Simulate escape key press
+    const keypressHandler = mockUseKeypress.mock.calls[0][0];
+    keypressHandler({ name: 'escape' });
+
+    expect(mockOnCancel).toHaveBeenCalled();
+  });
+
+  it('should not call onCancel for non-escape keys', () => {
+    render(
+      <ModelSelectionDialog
+        availableModels={mockAvailableModels}
+        currentModel="qwen3-coder-plus"
+        onSelect={mockOnSelect}
+        onCancel={mockOnCancel}
+      />,
+    );
+
+    const keypressHandler = mockUseKeypress.mock.calls[0][0];
+    keypressHandler({ name: 'enter' });
+
+    expect(mockOnCancel).not.toHaveBeenCalled();
+  });
+
+  it('should set correct initial index for current model', () => {
+    render(
+      <ModelSelectionDialog
+        availableModels={mockAvailableModels}
+        currentModel="qwen-vl-max-latest"
+        onSelect={mockOnSelect}
+        onCancel={mockOnCancel}
+      />,
+    );
+
+    const callArgs = mockRadioButtonSelect.mock.calls[0][0];
+    expect(callArgs.initialIndex).toBe(1); // qwen-vl-max-latest is at index 1
+  });
+
+  it('should set initial index to 0 when current model is not found', () => {
+    render(
+      <ModelSelectionDialog
+        availableModels={mockAvailableModels}
+        currentModel="non-existent-model"
+        onSelect={mockOnSelect}
+        onCancel={mockOnCancel}
+      />,
+    );
+
+    const callArgs = mockRadioButtonSelect.mock.calls[0][0];
+    expect(callArgs.initialIndex).toBe(0);
+  });
+
+  it('should call onSelect when a model is selected', () => {
+    render(
+      <ModelSelectionDialog
+        availableModels={mockAvailableModels}
+        currentModel="qwen3-coder-plus"
+        onSelect={mockOnSelect}
+        onCancel={mockOnCancel}
+      />,
+    );
+
+    const callArgs = mockRadioButtonSelect.mock.calls[0][0];
+    expect(typeof callArgs.onSelect).toBe('function');
+
+    // Simulate selection
+    const onSelectCallback = mockRadioButtonSelect.mock.calls[0][0].onSelect;
+    onSelectCallback('qwen-vl-max-latest');
+
+    expect(mockOnSelect).toHaveBeenCalledWith('qwen-vl-max-latest');
+  });
+
+  it('should handle empty models array', () => {
+    render(
+      <ModelSelectionDialog
+        availableModels={[]}
+        currentModel=""
+        onSelect={mockOnSelect}
+        onCancel={mockOnCancel}
+      />,
+    );
+
+    const callArgs = mockRadioButtonSelect.mock.calls[0][0];
+    expect(callArgs.items).toEqual([]);
+    expect(callArgs.initialIndex).toBe(0);
+  });
+
+  it('should create correct option items with proper labels', () => {
+    render(
+      <ModelSelectionDialog
+        availableModels={mockAvailableModels}
+        currentModel="qwen3-coder-plus"
+        onSelect={mockOnSelect}
+        onCancel={mockOnCancel}
+      />,
+    );
+
+    const expectedItems = [
+      {
+        label: 'qwen3-coder-plus (current)',
+        value: 'qwen3-coder-plus',
+      },
+      {
+        label: 'qwen-vl-max [Vision]',
+        value: 'qwen-vl-max-latest',
+      },
+      {
+        label: 'GPT-4',
+        value: 'gpt-4',
+      },
+    ];
+
+    const callArgs = mockRadioButtonSelect.mock.calls[0][0];
+    expect(callArgs.items).toEqual(expectedItems);
+  });
+
+  it('should show vision indicator for vision models', () => {
+    render(
+      <ModelSelectionDialog
+        availableModels={mockAvailableModels}
+        currentModel="gpt-4"
+        onSelect={mockOnSelect}
+        onCancel={mockOnCancel}
+      />,
+    );
+
+    const callArgs = mockRadioButtonSelect.mock.calls[0][0];
+    const visionModelItem = callArgs.items.find(
+      (item: RadioSelectItem<string>) => item.value === 'qwen-vl-max-latest',
+    );
+
+    expect(visionModelItem?.label).toContain('[Vision]');
+  });
+
+  it('should show current indicator for the current model', () => {
+    render(
+      <ModelSelectionDialog
+        availableModels={mockAvailableModels}
+        currentModel="qwen-vl-max-latest"
+        onSelect={mockOnSelect}
+        onCancel={mockOnCancel}
+      />,
+    );
+
+    const callArgs = mockRadioButtonSelect.mock.calls[0][0];
+    const currentModelItem = callArgs.items.find(
+      (item: RadioSelectItem<string>) => item.value === 'qwen-vl-max-latest',
+    );
+
+    expect(currentModelItem?.label).toContain('(current)');
+  });
+
+  it('should pass isFocused prop to RadioButtonSelect', () => {
+    render(
+      <ModelSelectionDialog
+        availableModels={mockAvailableModels}
+        currentModel="qwen3-coder-plus"
+        onSelect={mockOnSelect}
+        onCancel={mockOnCancel}
+      />,
+    );
+
+    const callArgs = mockRadioButtonSelect.mock.calls[0][0];
+    expect(callArgs.isFocused).toBe(true);
+  });
+
+  it('should handle multiple onSelect calls correctly', () => {
+    render(
+      <ModelSelectionDialog
+        availableModels={mockAvailableModels}
+        currentModel="qwen3-coder-plus"
+        onSelect={mockOnSelect}
+        onCancel={mockOnCancel}
+      />,
+    );
+
+    const onSelectCallback = mockRadioButtonSelect.mock.calls[0][0].onSelect;
+
+    // Call multiple times
+    onSelectCallback('qwen3-coder-plus');
+    onSelectCallback('qwen-vl-max-latest');
+    onSelectCallback('gpt-4');
+
+    expect(mockOnSelect).toHaveBeenCalledTimes(3);
+    expect(mockOnSelect).toHaveBeenNthCalledWith(1, 'qwen3-coder-plus');
+    expect(mockOnSelect).toHaveBeenNthCalledWith(2, 'qwen-vl-max-latest');
+    expect(mockOnSelect).toHaveBeenNthCalledWith(3, 'gpt-4');
+  });
+});

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

@@ -0,0 +1,87 @@
+/**
+ * @license
+ * Copyright 2025 Qwen
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import type React from 'react';
+import { Box, Text } from 'ink';
+import { Colors } from '../colors.js';
+import {
+  RadioButtonSelect,
+  type RadioSelectItem,
+} from './shared/RadioButtonSelect.js';
+import { useKeypress } from '../hooks/useKeypress.js';
+import type { AvailableModel } from '../models/availableModels.js';
+
+export interface ModelSelectionDialogProps {
+  availableModels: AvailableModel[];
+  currentModel: string;
+  onSelect: (modelId: string) => void;
+  onCancel: () => void;
+}
+
+export const ModelSelectionDialog: React.FC<ModelSelectionDialogProps> = ({
+  availableModels,
+  currentModel,
+  onSelect,
+  onCancel,
+}) => {
+  useKeypress(
+    (key) => {
+      if (key.name === 'escape') {
+        onCancel();
+      }
+    },
+    { isActive: true },
+  );
+
+  const options: Array<RadioSelectItem<string>> = availableModels.map(
+    (model) => {
+      const visionIndicator = model.isVision ? ' [Vision]' : '';
+      const currentIndicator = model.id === currentModel ? ' (current)' : '';
+      return {
+        label: `${model.label}${visionIndicator}${currentIndicator}`,
+        value: model.id,
+      };
+    },
+  );
+
+  const initialIndex = Math.max(
+    0,
+    availableModels.findIndex((model) => model.id === currentModel),
+  );
+
+  const handleSelect = (modelId: string) => {
+    onSelect(modelId);
+  };
+
+  return (
+    <Box
+      flexDirection="column"
+      borderStyle="round"
+      borderColor={Colors.AccentBlue}
+      padding={1}
+      width="100%"
+      marginLeft={1}
+    >
+      <Box flexDirection="column" marginBottom={1}>
+        <Text bold>Select Model</Text>
+        <Text>Choose a model for this session:</Text>
+      </Box>
+
+      <Box marginBottom={1}>
+        <RadioButtonSelect
+          items={options}
+          initialIndex={initialIndex}
+          onSelect={handleSelect}
+          isFocused
+        />
+      </Box>
+
+      <Box>
+        <Text color={Colors.Gray}>Press Enter to select, Esc to cancel</Text>
+      </Box>
+    </Box>
+  );
+};

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

@@ -0,0 +1,181 @@
+/**
+ * @license
+ * Copyright 2025 Qwen
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import React from 'react';
+import { render } from 'ink-testing-library';
+import { describe, it, expect, vi, beforeEach } from 'vitest';
+import { ModelSwitchDialog, VisionSwitchOutcome } from './ModelSwitchDialog.js';
+
+// Mock the useKeypress hook
+const mockUseKeypress = vi.hoisted(() => vi.fn());
+vi.mock('../hooks/useKeypress.js', () => ({
+  useKeypress: mockUseKeypress,
+}));
+
+// Mock the RadioButtonSelect component
+const mockRadioButtonSelect = vi.hoisted(() => vi.fn());
+vi.mock('./shared/RadioButtonSelect.js', () => ({
+  RadioButtonSelect: mockRadioButtonSelect,
+}));
+
+describe('ModelSwitchDialog', () => {
+  const mockOnSelect = vi.fn();
+
+  beforeEach(() => {
+    vi.clearAllMocks();
+
+    // Mock RadioButtonSelect to return a simple div
+    mockRadioButtonSelect.mockReturnValue(
+      React.createElement('div', { 'data-testid': 'radio-select' }),
+    );
+  });
+
+  it('should setup RadioButtonSelect with correct options', () => {
+    render(<ModelSwitchDialog onSelect={mockOnSelect} />);
+
+    const expectedItems = [
+      {
+        label: 'Switch for this request only',
+        value: VisionSwitchOutcome.SwitchOnce,
+      },
+      {
+        label: 'Switch session to vision model',
+        value: VisionSwitchOutcome.SwitchSessionToVL,
+      },
+      {
+        label: 'Continue with current model',
+        value: VisionSwitchOutcome.ContinueWithCurrentModel,
+      },
+    ];
+
+    const callArgs = mockRadioButtonSelect.mock.calls[0][0];
+    expect(callArgs.items).toEqual(expectedItems);
+    expect(callArgs.initialIndex).toBe(0);
+    expect(callArgs.isFocused).toBe(true);
+  });
+
+  it('should call onSelect when an option is selected', () => {
+    render(<ModelSwitchDialog onSelect={mockOnSelect} />);
+
+    const callArgs = mockRadioButtonSelect.mock.calls[0][0];
+    expect(typeof callArgs.onSelect).toBe('function');
+
+    // Simulate selection of "Switch for this request only"
+    const onSelectCallback = mockRadioButtonSelect.mock.calls[0][0].onSelect;
+    onSelectCallback(VisionSwitchOutcome.SwitchOnce);
+
+    expect(mockOnSelect).toHaveBeenCalledWith(VisionSwitchOutcome.SwitchOnce);
+  });
+
+  it('should call onSelect with SwitchSessionToVL when second option is selected', () => {
+    render(<ModelSwitchDialog onSelect={mockOnSelect} />);
+
+    const onSelectCallback = mockRadioButtonSelect.mock.calls[0][0].onSelect;
+    onSelectCallback(VisionSwitchOutcome.SwitchSessionToVL);
+
+    expect(mockOnSelect).toHaveBeenCalledWith(
+      VisionSwitchOutcome.SwitchSessionToVL,
+    );
+  });
+
+  it('should call onSelect with ContinueWithCurrentModel when third option is selected', () => {
+    render(<ModelSwitchDialog onSelect={mockOnSelect} />);
+
+    const onSelectCallback = mockRadioButtonSelect.mock.calls[0][0].onSelect;
+    onSelectCallback(VisionSwitchOutcome.ContinueWithCurrentModel);
+
+    expect(mockOnSelect).toHaveBeenCalledWith(
+      VisionSwitchOutcome.ContinueWithCurrentModel,
+    );
+  });
+
+  it('should setup escape key handler to call onSelect with ContinueWithCurrentModel', () => {
+    render(<ModelSwitchDialog onSelect={mockOnSelect} />);
+
+    expect(mockUseKeypress).toHaveBeenCalledWith(expect.any(Function), {
+      isActive: true,
+    });
+
+    // Simulate escape key press
+    const keypressHandler = mockUseKeypress.mock.calls[0][0];
+    keypressHandler({ name: 'escape' });
+
+    expect(mockOnSelect).toHaveBeenCalledWith(
+      VisionSwitchOutcome.ContinueWithCurrentModel,
+    );
+  });
+
+  it('should not call onSelect for non-escape keys', () => {
+    render(<ModelSwitchDialog onSelect={mockOnSelect} />);
+
+    const keypressHandler = mockUseKeypress.mock.calls[0][0];
+    keypressHandler({ name: 'enter' });
+
+    expect(mockOnSelect).not.toHaveBeenCalled();
+  });
+
+  it('should set initial index to 0 (first option)', () => {
+    render(<ModelSwitchDialog onSelect={mockOnSelect} />);
+
+    const callArgs = mockRadioButtonSelect.mock.calls[0][0];
+    expect(callArgs.initialIndex).toBe(0);
+  });
+
+  describe('VisionSwitchOutcome enum', () => {
+    it('should have correct enum values', () => {
+      expect(VisionSwitchOutcome.SwitchOnce).toBe('once');
+      expect(VisionSwitchOutcome.SwitchSessionToVL).toBe('session');
+      expect(VisionSwitchOutcome.ContinueWithCurrentModel).toBe('persist');
+    });
+  });
+
+  it('should handle multiple onSelect calls correctly', () => {
+    render(<ModelSwitchDialog onSelect={mockOnSelect} />);
+
+    const onSelectCallback = mockRadioButtonSelect.mock.calls[0][0].onSelect;
+
+    // Call multiple times
+    onSelectCallback(VisionSwitchOutcome.SwitchOnce);
+    onSelectCallback(VisionSwitchOutcome.SwitchSessionToVL);
+    onSelectCallback(VisionSwitchOutcome.ContinueWithCurrentModel);
+
+    expect(mockOnSelect).toHaveBeenCalledTimes(3);
+    expect(mockOnSelect).toHaveBeenNthCalledWith(
+      1,
+      VisionSwitchOutcome.SwitchOnce,
+    );
+    expect(mockOnSelect).toHaveBeenNthCalledWith(
+      2,
+      VisionSwitchOutcome.SwitchSessionToVL,
+    );
+    expect(mockOnSelect).toHaveBeenNthCalledWith(
+      3,
+      VisionSwitchOutcome.ContinueWithCurrentModel,
+    );
+  });
+
+  it('should pass isFocused prop to RadioButtonSelect', () => {
+    render(<ModelSwitchDialog onSelect={mockOnSelect} />);
+
+    const callArgs = mockRadioButtonSelect.mock.calls[0][0];
+    expect(callArgs.isFocused).toBe(true);
+  });
+
+  it('should handle escape key multiple times', () => {
+    render(<ModelSwitchDialog onSelect={mockOnSelect} />);
+
+    const keypressHandler = mockUseKeypress.mock.calls[0][0];
+
+    // Call escape multiple times
+    keypressHandler({ name: 'escape' });
+    keypressHandler({ name: 'escape' });
+
+    expect(mockOnSelect).toHaveBeenCalledTimes(2);
+    expect(mockOnSelect).toHaveBeenCalledWith(
+      VisionSwitchOutcome.ContinueWithCurrentModel,
+    );
+  });
+});

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

@@ -0,0 +1,89 @@
+/**
+ * @license
+ * Copyright 2025 Qwen
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import type React from 'react';
+import { Box, Text } from 'ink';
+import { Colors } from '../colors.js';
+import {
+  RadioButtonSelect,
+  type RadioSelectItem,
+} from './shared/RadioButtonSelect.js';
+import { useKeypress } from '../hooks/useKeypress.js';
+
+export enum VisionSwitchOutcome {
+  SwitchOnce = 'once',
+  SwitchSessionToVL = 'session',
+  ContinueWithCurrentModel = 'persist',
+}
+
+export interface ModelSwitchDialogProps {
+  onSelect: (outcome: VisionSwitchOutcome) => void;
+}
+
+export const ModelSwitchDialog: React.FC<ModelSwitchDialogProps> = ({
+  onSelect,
+}) => {
+  useKeypress(
+    (key) => {
+      if (key.name === 'escape') {
+        onSelect(VisionSwitchOutcome.ContinueWithCurrentModel);
+      }
+    },
+    { isActive: true },
+  );
+
+  const options: Array<RadioSelectItem<VisionSwitchOutcome>> = [
+    {
+      label: 'Switch for this request only',
+      value: VisionSwitchOutcome.SwitchOnce,
+    },
+    {
+      label: 'Switch session to vision model',
+      value: VisionSwitchOutcome.SwitchSessionToVL,
+    },
+    {
+      label: 'Continue with current model',
+      value: VisionSwitchOutcome.ContinueWithCurrentModel,
+    },
+  ];
+
+  const handleSelect = (outcome: VisionSwitchOutcome) => {
+    onSelect(outcome);
+  };
+
+  return (
+    <Box
+      flexDirection="column"
+      borderStyle="round"
+      borderColor={Colors.AccentYellow}
+      padding={1}
+      width="100%"
+      marginLeft={1}
+    >
+      <Box flexDirection="column" marginBottom={1}>
+        <Text bold>Vision Model Switch Required</Text>
+        <Text>
+          Your message contains an image, but the current model doesn&apos;t
+          support vision.
+        </Text>
+        <Text>How would you like to proceed?</Text>
+      </Box>
+
+      <Box marginBottom={1}>
+        <RadioButtonSelect
+          items={options}
+          initialIndex={0}
+          onSelect={handleSelect}
+          isFocused
+        />
+      </Box>
+
+      <Box>
+        <Text color={Colors.Gray}>Press Enter to select, Esc to cancel</Text>
+      </Box>
+    </Box>
+  );
+};

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

@@ -0,0 +1,41 @@
+/**
+ * @license
+ * Copyright 2025 Qwen
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import type React from 'react';
+import { Box, Text } from 'ink';
+import { MarkdownDisplay } from '../utils/MarkdownDisplay.js';
+import { Colors } from '../colors.js';
+import type { PlanResultDisplay } from '@qwen-code/qwen-code-core';
+
+interface PlanSummaryDisplayProps {
+  data: PlanResultDisplay;
+  availableHeight?: number;
+  childWidth: number;
+}
+
+export const PlanSummaryDisplay: React.FC<PlanSummaryDisplayProps> = ({
+  data,
+  availableHeight,
+  childWidth,
+}) => {
+  const { message, plan } = data;
+
+  return (
+    <Box flexDirection="column">
+      <Box marginBottom={1}>
+        <Text color={Colors.AccentGreen} wrap="wrap">
+          {message}
+        </Text>
+      </Box>
+      <MarkdownDisplay
+        text={plan}
+        isPending={false}
+        availableTerminalHeight={availableHeight}
+        terminalWidth={childWidth}
+      />
+    </Box>
+  );
+};

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

@@ -5,6 +5,7 @@
  */
 
 import { describe, it, expect, vi } from 'vitest';
+import { EOL } from 'node:os';
 import { ToolConfirmationMessage } from './ToolConfirmationMessage.js';
 import type {
   ToolCallConfirmationDetails,
@@ -66,6 +67,30 @@ describe('ToolConfirmationMessage', () => {
     );
   });
 
+  it('should render plan confirmation with markdown plan content', () => {
+    const confirmationDetails: ToolCallConfirmationDetails = {
+      type: 'plan',
+      title: 'Would you like to proceed?',
+      plan: '# Implementation Plan\n- Step one\n- Step two'.replace(/\n/g, EOL),
+      onConfirm: vi.fn(),
+    };
+
+    const { lastFrame } = renderWithProviders(
+      <ToolConfirmationMessage
+        confirmationDetails={confirmationDetails}
+        config={mockConfig}
+        availableTerminalHeight={30}
+        terminalWidth={80}
+      />,
+    );
+
+    expect(lastFrame()).toContain('Yes, and auto-accept edits');
+    expect(lastFrame()).toContain('Yes, and manually approve edits');
+    expect(lastFrame()).toContain('No, keep planning');
+    expect(lastFrame()).toContain('Implementation Plan');
+    expect(lastFrame()).toContain('Step one');
+  });
+
   describe('with folder trust', () => {
     const editConfirmationDetails: ToolCallConfirmationDetails = {
       type: 'edit',

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

@@ -9,6 +9,7 @@ import { Box, Text } from 'ink';
 import { DiffRenderer } from './DiffRenderer.js';
 import { Colors } from '../../colors.js';
 import { RenderInline } from '../../utils/InlineMarkdownRenderer.js';
+import { MarkdownDisplay } from '../../utils/MarkdownDisplay.js';
 import type {
   ToolCallConfirmationDetails,
   ToolExecuteConfirmationDetails,
@@ -27,6 +28,7 @@ export interface ToolConfirmationMessageProps {
   isFocused?: boolean;
   availableTerminalHeight?: number;
   terminalWidth: number;
+  compactMode?: boolean;
 }
 
 export const ToolConfirmationMessage: React.FC<
@@ -37,6 +39,7 @@ export const ToolConfirmationMessage: React.FC<
   isFocused = true,
   availableTerminalHeight,
   terminalWidth,
+  compactMode = false,
 }) => {
   const { onConfirm } = confirmationDetails;
   const childWidth = terminalWidth - 2; // 2 for padding
@@ -70,6 +73,40 @@ export const ToolConfirmationMessage: React.FC<
 
   const handleSelect = (item: ToolConfirmationOutcome) => handleConfirm(item);
 
+  // Compact mode: return simple 3-option display
+  if (compactMode) {
+    const compactOptions: Array<RadioSelectItem<ToolConfirmationOutcome>> = [
+      {
+        label: 'Yes, allow once',
+        value: ToolConfirmationOutcome.ProceedOnce,
+      },
+      {
+        label: 'Allow always',
+        value: ToolConfirmationOutcome.ProceedAlways,
+      },
+      {
+        label: 'No',
+        value: ToolConfirmationOutcome.Cancel,
+      },
+    ];
+
+    return (
+      <Box flexDirection="column">
+        <Box>
+          <Text wrap="truncate">Do you want to proceed?</Text>
+        </Box>
+        <Box>
+          <RadioButtonSelect
+            items={compactOptions}
+            onSelect={handleSelect}
+            isFocused={isFocused}
+          />
+        </Box>
+      </Box>
+    );
+  }
+
+  // Original logic continues unchanged below
   let bodyContent: React.ReactNode | null = null; // Removed contextDisplay here
   let question: string;
 
@@ -199,6 +236,33 @@ export const ToolConfirmationMessage: React.FC<
         </Box>
       </Box>
     );
+  } else if (confirmationDetails.type === 'plan') {
+    const planProps = confirmationDetails;
+
+    question = planProps.title;
+    options.push({
+      label: 'Yes, and auto-accept edits',
+      value: ToolConfirmationOutcome.ProceedAlways,
+    });
+    options.push({
+      label: 'Yes, and manually approve edits',
+      value: ToolConfirmationOutcome.ProceedOnce,
+    });
+    options.push({
+      label: 'No, keep planning (esc)',
+      value: ToolConfirmationOutcome.Cancel,
+    });
+
+    bodyContent = (
+      <Box flexDirection="column" paddingX={1} marginLeft={1}>
+        <MarkdownDisplay
+          text={planProps.plan}
+          isPending={false}
+          availableTerminalHeight={availableBodyContentHeight()}
+          terminalWidth={childWidth}
+        />
+      </Box>
+    );
   } else if (confirmationDetails.type === 'info') {
     const infoProps = confirmationDetails;
     const displayUrls =

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

@@ -18,9 +18,11 @@ import { TOOL_STATUS } from '../../constants.js';
 import type {
   TodoResultDisplay,
   TaskResultDisplay,
+  PlanResultDisplay,
   Config,
 } from '@qwen-code/qwen-code-core';
 import { AgentExecutionDisplay } from '../subagents/index.js';
+import { PlanSummaryDisplay } from '../PlanSummaryDisplay.js';
 
 const STATIC_HEIGHT = 1;
 const RESERVED_LINE_COUNT = 5; // for tool name, status, padding etc.
@@ -35,6 +37,7 @@ export type TextEmphasis = 'high' | 'medium' | 'low';
 type DisplayRendererResult =
   | { type: 'none' }
   | { type: 'todo'; data: TodoResultDisplay }
+  | { type: 'plan'; data: PlanResultDisplay }
   | { type: 'string'; data: string }
   | { type: 'diff'; data: { fileDiff: string; fileName: string } }
   | { type: 'task'; data: TaskResultDisplay };
@@ -63,6 +66,18 @@ const useResultDisplayRenderer = (
       };
     }
 
+    if (
+      typeof resultDisplay === 'object' &&
+      resultDisplay !== null &&
+      'type' in resultDisplay &&
+      resultDisplay.type === 'plan_summary'
+    ) {
+      return {
+        type: 'plan',
+        data: resultDisplay as PlanResultDisplay,
+      };
+    }
+
     // Check for SubagentExecutionResultDisplay (for non-task tools)
     if (
       typeof resultDisplay === 'object' &&
@@ -102,6 +117,18 @@ const TodoResultRenderer: React.FC<{ data: TodoResultDisplay }> = ({
   data,
 }) => <TodoDisplay todos={data.todos} />;
 
+const PlanResultRenderer: React.FC<{
+  data: PlanResultDisplay;
+  availableHeight?: number;
+  childWidth: number;
+}> = ({ data, availableHeight, childWidth }) => (
+  <PlanSummaryDisplay
+    data={data}
+    availableHeight={availableHeight}
+    childWidth={childWidth}
+  />
+);
+
 /**
  * Component to render subagent execution results
  */
@@ -229,6 +256,13 @@ export const ToolMessage: React.FC<ToolMessageProps> = ({
             {displayRenderer.type === 'todo' && (
               <TodoResultRenderer data={displayRenderer.data} />
             )}
+            {displayRenderer.type === 'plan' && (
+              <PlanResultRenderer
+                data={displayRenderer.data}
+                availableHeight={availableHeight}
+                childWidth={childWidth}
+              />
+            )}
             {displayRenderer.type === 'task' && (
               <SubagentExecutionRenderer
                 data={displayRenderer.data}

packages/cli/src/ui/components/subagents/create/AgentCreationWizard.tsx

@@ -5,7 +5,7 @@
  */
 
 import { useReducer, useCallback, useMemo } from 'react';
-import { Box, Text, useInput } from 'ink';
+import { Box, Text } from 'ink';
 import { wizardReducer, initialWizardState } from '../reducers.js';
 import { LocationSelector } from './LocationSelector.js';
 import { GenerationMethodSelector } from './GenerationMethodSelector.js';
@@ -20,6 +20,7 @@ import type { Config } from '@qwen-code/qwen-code-core';
 import { Colors } from '../../../colors.js';
 import { theme } from '../../../semantic-colors.js';
 import { TextEntryStep } from './TextEntryStep.js';
+import { useKeypress } from '../../../hooks/useKeypress.js';
 
 interface AgentCreationWizardProps {
   onClose: () => void;
@@ -49,8 +50,12 @@ export function AgentCreationWizard({
   }, [onClose]);
 
   // Centralized ESC key handling for the entire wizard
-  useInput((input, key) => {
-    if (key.escape) {
+  useKeypress(
+    (key) => {
+      if (key.name !== 'escape') {
+        return;
+      }
+
       // LLM DescriptionInput handles its own ESC logic when generating
       const kind = getStepKind(state.generationMethod, state.currentStep);
       if (kind === 'LLM_DESC' && state.isGenerating) {
@@ -64,8 +69,9 @@ export function AgentCreationWizard({
         // On other steps, ESC goes back to previous step
         handlePrevious();
       }
-    }
-  });
+    },
+    { isActive: true },
+  );
 
   const stepProps: WizardStepProps = useMemo(
     () => ({

packages/cli/src/ui/components/subagents/manage/AgentSelectionStep.tsx

@@ -227,7 +227,7 @@ export const AgentSelectionStep = ({
     const textColor = isSelected ? theme.text.accent : theme.text.primary;
 
     return (
-      <Box key={agent.name} alignItems="center">
+      <Box key={`${agent.name}-${agent.level}`} alignItems="center">
         <Box minWidth={2} flexShrink={0}>
           <Text color={isSelected ? theme.text.accent : theme.text.primary}>
             {isSelected ? '●' : ' '}

packages/cli/src/ui/components/subagents/manage/AgentsManagerDialog.tsx

@@ -5,7 +5,7 @@
  */
 
 import { useState, useCallback, useMemo, useEffect } from 'react';
-import { Box, Text, useInput } from 'ink';
+import { Box, Text } from 'ink';
 import { AgentSelectionStep } from './AgentSelectionStep.js';
 import { ActionSelectionStep } from './ActionSelectionStep.js';
 import { AgentViewerStep } from './AgentViewerStep.js';
@@ -17,7 +17,8 @@ import { MANAGEMENT_STEPS } from '../types.js';
 import { Colors } from '../../../colors.js';
 import { theme } from '../../../semantic-colors.js';
 import { getColorForDisplay, shouldShowColor } from '../utils.js';
-import type { Config, SubagentConfig } from '@qwen-code/qwen-code-core';
+import type { SubagentConfig, Config } from '@qwen-code/qwen-code-core';
+import { useKeypress } from '../../../hooks/useKeypress.js';
 
 interface AgentsManagerDialogProps {
   onClose: () => void;
@@ -52,18 +53,7 @@ export function AgentsManagerDialog({
     const manager = config.getSubagentManager();
 
     // Load agents from all levels separately to show all agents including conflicts
-    const [projectAgents, userAgents, builtinAgents] = await Promise.all([
-      manager.listSubagents({ level: 'project' }),
-      manager.listSubagents({ level: 'user' }),
-      manager.listSubagents({ level: 'builtin' }),
-    ]);
-
-    // Combine all agents (project, user, and builtin level)
-    const allAgents = [
-      ...(projectAgents || []),
-      ...(userAgents || []),
-      ...(builtinAgents || []),
-    ];
+    const allAgents = await manager.listSubagents({ force: true });
 
     setAvailableAgents(allAgents);
   }, [config]);
@@ -122,8 +112,12 @@ export function AgentsManagerDialog({
   );
 
   // Centralized ESC key handling for the entire dialog
-  useInput((input, key) => {
-    if (key.escape) {
+  useKeypress(
+    (key) => {
+      if (key.name !== 'escape') {
+        return;
+      }
+
       const currentStep = getCurrentStep();
       if (currentStep === MANAGEMENT_STEPS.AGENT_SELECTION) {
         // On first step, ESC cancels the entire dialog
@@ -132,8 +126,9 @@ export function AgentsManagerDialog({
         // On other steps, ESC goes back to previous step in navigation stack
         handleNavigateBack();
       }
-    }
-  });
+    },
+    { isActive: true },
+  );
 
   // Props for child components - now using direct state and callbacks
   const commonProps = useMemo(

packages/cli/src/ui/components/subagents/runtime/AgentExecutionDisplay.tsx

@@ -18,12 +18,12 @@ import { COLOR_OPTIONS } from '../constants.js';
 import { fmtDuration } from '../utils.js';
 import { ToolConfirmationMessage } from '../../messages/ToolConfirmationMessage.js';
 
-export type DisplayMode = 'default' | 'verbose';
+export type DisplayMode = 'compact' | 'default' | 'verbose';
 
 export interface AgentExecutionDisplayProps {
   data: TaskResultDisplay;
   availableHeight?: number;
-  childWidth?: number;
+  childWidth: number;
   config: Config;
 }
 
@@ -80,7 +80,7 @@ export const AgentExecutionDisplay: React.FC<AgentExecutionDisplayProps> = ({
   childWidth,
   config,
 }) => {
-  const [displayMode, setDisplayMode] = React.useState<DisplayMode>('default');
+  const [displayMode, setDisplayMode] = React.useState<DisplayMode>('compact');
 
   const agentColor = useMemo(() => {
     const colorOption = COLOR_OPTIONS.find(
@@ -93,8 +93,6 @@ export const AgentExecutionDisplay: React.FC<AgentExecutionDisplayProps> = ({
     // This component only listens to keyboard shortcut events when the subagent is running
     if (data.status !== 'running') return '';
 
-    if (displayMode === 'verbose') return 'Press ctrl+r to show less.';
-
     if (displayMode === 'default') {
       const hasMoreLines =
         data.taskPrompt.split('\n').length > MAX_TASK_PROMPT_LINES;
@@ -102,17 +100,28 @@ export const AgentExecutionDisplay: React.FC<AgentExecutionDisplayProps> = ({
         data.toolCalls && data.toolCalls.length > MAX_TOOL_CALLS;
 
       if (hasMoreToolCalls || hasMoreLines) {
-        return 'Press ctrl+r to show more.';
+        return 'Press ctrl+r to show less, ctrl+e to show more.';
       }
-      return '';
+      return 'Press ctrl+r to show less.';
     }
-    return '';
-  }, [displayMode, data.toolCalls, data.taskPrompt, data.status]);
 
-  // Handle ctrl+r keypresses to control display mode
+    if (displayMode === 'verbose') {
+      return 'Press ctrl+e to show less.';
+    }
+
+    return '';
+  }, [displayMode, data]);
+
+  // Handle keyboard shortcuts to control display mode
   useKeypress(
     (key) => {
       if (key.ctrl && key.name === 'r') {
+        // ctrl+r toggles between compact and default
+        setDisplayMode((current) =>
+          current === 'compact' ? 'default' : 'compact',
+        );
+      } else if (key.ctrl && key.name === 'e') {
+        // ctrl+e toggles between default and verbose
         setDisplayMode((current) =>
           current === 'default' ? 'verbose' : 'default',
         );
@@ -121,6 +130,82 @@ export const AgentExecutionDisplay: React.FC<AgentExecutionDisplayProps> = ({
     { isActive: true },
   );
 
+  if (displayMode === 'compact') {
+    return (
+      <Box flexDirection="column">
+        {/* Header: Agent name and status */}
+        {!data.pendingConfirmation && (
+          <Box flexDirection="row">
+            <Text bold color={agentColor}>
+              {data.subagentName}
+            </Text>
+            <StatusDot status={data.status} />
+            <StatusIndicator status={data.status} />
+          </Box>
+        )}
+
+        {/* Running state: Show current tool call and progress */}
+        {data.status === 'running' && (
+          <>
+            {/* Current tool call */}
+            {data.toolCalls && data.toolCalls.length > 0 && (
+              <Box flexDirection="column">
+                <ToolCallItem
+                  toolCall={data.toolCalls[data.toolCalls.length - 1]}
+                  compact={true}
+                />
+                {/* Show count of additional tool calls if there are more than 1 */}
+                {data.toolCalls.length > 1 && !data.pendingConfirmation && (
+                  <Box flexDirection="row" paddingLeft={4}>
+                    <Text color={Colors.Gray}>
+                      +{data.toolCalls.length - 1} more tool calls (ctrl+r to
+                      expand)
+                    </Text>
+                  </Box>
+                )}
+              </Box>
+            )}
+
+            {/* Inline approval prompt when awaiting confirmation */}
+            {data.pendingConfirmation && (
+              <Box flexDirection="column" marginTop={1} paddingLeft={1}>
+                <ToolConfirmationMessage
+                  confirmationDetails={data.pendingConfirmation}
+                  isFocused={true}
+                  availableTerminalHeight={availableHeight}
+                  terminalWidth={childWidth}
+                  compactMode={true}
+                  config={config}
+                />
+              </Box>
+            )}
+          </>
+        )}
+
+        {/* Completed state: Show summary line */}
+        {data.status === 'completed' && data.executionSummary && (
+          <Box flexDirection="row" marginTop={1}>
+            <Text color={theme.text.secondary}>
+              Execution Summary: {data.executionSummary.totalToolCalls} tool
+              uses · {data.executionSummary.totalTokens.toLocaleString()} tokens
+              · {fmtDuration(data.executionSummary.totalDurationMs)}
+            </Text>
+          </Box>
+        )}
+
+        {/* Failed/Cancelled state: Show error reason */}
+        {data.status === 'failed' && (
+          <Box flexDirection="row" marginTop={1}>
+            <Text color={theme.status.error}>
+              Failed: {data.terminateReason}
+            </Text>
+          </Box>
+        )}
+      </Box>
+    );
+  }
+
+  // Default and verbose modes use normal layout
   return (
     <Box flexDirection="column" paddingX={1} gap={1}>
       {/* Header with subagent name and status */}
@@ -158,7 +243,8 @@ export const AgentExecutionDisplay: React.FC<AgentExecutionDisplayProps> = ({
             config={config}
             isFocused={true}
             availableTerminalHeight={availableHeight}
-            terminalWidth={childWidth ?? 80}
+            terminalWidth={childWidth}
+            compactMode={true}
           />
         </Box>
       )}
@@ -280,7 +366,8 @@ const ToolCallItem: React.FC<{
     resultDisplay?: string;
     description?: string;
   };
-}> = ({ toolCall }) => {
+  compact?: boolean;
+}> = ({ toolCall, compact = false }) => {
   const STATUS_INDICATOR_WIDTH = 3;
 
   // Map subagent status to ToolCallStatus-like display
@@ -335,8 +422,8 @@ const ToolCallItem: React.FC<{
         </Text>
       </Box>
 
-      {/* Second line: truncated returnDisplay output */}
-      {truncatedOutput && (
+      {/* Second line: truncated returnDisplay output - hidden in compact mode */}
+      {!compact && truncatedOutput && (
         <Box flexDirection="row" paddingLeft={STATUS_INDICATOR_WIDTH}>
           <Text color={Colors.Gray}>{truncatedOutput}</Text>
         </Box>

packages/cli/src/ui/contexts/KeypressContext.test.tsx

@@ -526,7 +526,7 @@ describe('KeypressContext - Kitty Protocol', () => {
         });
 
         await waitFor(() => {
-          expect(keyHandler).toHaveBeenCalledTimes(2); // 1 paste event + 1 paste event for 'after'
+          expect(keyHandler).toHaveBeenCalledTimes(6); // 1 paste event + 5 individual chars for 'after'
         });
 
         // Should emit paste event first
@@ -538,12 +538,40 @@ describe('KeypressContext - Kitty Protocol', () => {
           }),
         );
 
-        // Then process 'after' as a paste event (since it's > 2 chars)
+        // Then process 'after' as individual characters (since it doesn't contain return)
         expect(keyHandler).toHaveBeenNthCalledWith(
           2,
           expect.objectContaining({
-            paste: true,
-            sequence: 'after',
+            name: 'a',
+            paste: false,
+          }),
+        );
+        expect(keyHandler).toHaveBeenNthCalledWith(
+          3,
+          expect.objectContaining({
+            name: 'f',
+            paste: false,
+          }),
+        );
+        expect(keyHandler).toHaveBeenNthCalledWith(
+          4,
+          expect.objectContaining({
+            name: 't',
+            paste: false,
+          }),
+        );
+        expect(keyHandler).toHaveBeenNthCalledWith(
+          5,
+          expect.objectContaining({
+            name: 'e',
+            paste: false,
+          }),
+        );
+        expect(keyHandler).toHaveBeenNthCalledWith(
+          6,
+          expect.objectContaining({
+            name: 'r',
+            paste: false,
           }),
         );
       });
@@ -571,7 +599,7 @@ describe('KeypressContext - Kitty Protocol', () => {
         });
 
         await waitFor(() => {
-          expect(keyHandler).toHaveBeenCalledTimes(14); // Adjusted based on actual behavior
+          expect(keyHandler).toHaveBeenCalledTimes(16); // 5 + 1 + 6 + 1 + 3 = 16 calls
         });
 
         // Check the sequence: 'start' (5 chars) + paste1 + 'middle' (6 chars) + paste2 + 'end' (3 chars as paste)
@@ -643,13 +671,18 @@ describe('KeypressContext - Kitty Protocol', () => {
           }),
         );
 
-        // 'end' as paste event (since it's > 2 chars)
+        // 'end' as individual characters (since it doesn't contain return)
         expect(keyHandler).toHaveBeenNthCalledWith(
           callIndex++,
-          expect.objectContaining({
-            paste: true,
-            sequence: 'end',
-          }),
+          expect.objectContaining({ name: 'e' }),
+        );
+        expect(keyHandler).toHaveBeenNthCalledWith(
+          callIndex++,
+          expect.objectContaining({ name: 'n' }),
+        );
+        expect(keyHandler).toHaveBeenNthCalledWith(
+          callIndex++,
+          expect.objectContaining({ name: 'd' }),
         );
       });
 
@@ -738,16 +771,18 @@ describe('KeypressContext - Kitty Protocol', () => {
         });
 
         await waitFor(() => {
-          // With the current implementation, fragmented data gets processed differently
-          // The first fragment '\x1b[20' gets processed as individual characters
-          // The second fragment '0~content\x1b[2' gets processed as paste + individual chars
-          // The third fragment '01~' gets processed as individual characters
-          expect(keyHandler).toHaveBeenCalled();
+          // With the current implementation, fragmented paste markers get reconstructed
+          // into a single paste event for 'content'
+          expect(keyHandler).toHaveBeenCalledTimes(1);
         });
 
-        // The current implementation processes fragmented paste markers as separate events
-        // rather than reconstructing them into a single paste event
-        expect(keyHandler.mock.calls.length).toBeGreaterThan(1);
+        // Should reconstruct the fragmented paste markers into a single paste event
+        expect(keyHandler).toHaveBeenCalledWith(
+          expect.objectContaining({
+            paste: true,
+            sequence: 'content',
+          }),
+        );
       });
     });
 
@@ -851,19 +886,38 @@ describe('KeypressContext - Kitty Protocol', () => {
           stdin.emit('data', Buffer.from('lo'));
         });
 
-        // With the current implementation, data is processed as it arrives
-        // First chunk 'hel' is treated as paste (multi-character)
+        // With the current implementation, data is processed as individual characters
+        // since 'hel' doesn't contain return (0x0d)
         expect(keyHandler).toHaveBeenNthCalledWith(
           1,
           expect.objectContaining({
-            paste: true,
-            sequence: 'hel',
+            name: 'h',
+            sequence: 'h',
+            paste: false,
           }),
         );
 
-        // Second chunk 'lo' is processed as individual characters
         expect(keyHandler).toHaveBeenNthCalledWith(
           2,
+          expect.objectContaining({
+            name: 'e',
+            sequence: 'e',
+            paste: false,
+          }),
+        );
+
+        expect(keyHandler).toHaveBeenNthCalledWith(
+          3,
+          expect.objectContaining({
+            name: 'l',
+            sequence: 'l',
+            paste: false,
+          }),
+        );
+
+        // Second chunk 'lo' is also processed as individual characters
+        expect(keyHandler).toHaveBeenNthCalledWith(
+          4,
           expect.objectContaining({
             name: 'l',
             sequence: 'l',
@@ -872,7 +926,7 @@ describe('KeypressContext - Kitty Protocol', () => {
         );
 
         expect(keyHandler).toHaveBeenNthCalledWith(
-          3,
+          5,
           expect.objectContaining({
             name: 'o',
             sequence: 'o',
@@ -880,7 +934,7 @@ describe('KeypressContext - Kitty Protocol', () => {
           }),
         );
 
-        expect(keyHandler).toHaveBeenCalledTimes(3);
+        expect(keyHandler).toHaveBeenCalledTimes(5);
       } finally {
         vi.useRealTimers();
       }
@@ -907,14 +961,20 @@ describe('KeypressContext - Kitty Protocol', () => {
         });
 
         // Should flush immediately without waiting for timeout
-        // Large data gets treated as paste event
-        expect(keyHandler).toHaveBeenCalledTimes(1);
-        expect(keyHandler).toHaveBeenCalledWith(
-          expect.objectContaining({
-            paste: true,
-            sequence: largeData,
-          }),
-        );
+        // Large data without return gets treated as individual characters
+        expect(keyHandler).toHaveBeenCalledTimes(65);
+
+        // Each character should be processed individually
+        for (let i = 0; i < 65; i++) {
+          expect(keyHandler).toHaveBeenNthCalledWith(
+            i + 1,
+            expect.objectContaining({
+              name: 'x',
+              sequence: 'x',
+              paste: false,
+            }),
+          );
+        }
 
         // Advancing timer should not cause additional calls
         const callCountBefore = keyHandler.mock.calls.length;

packages/cli/src/ui/contexts/KeypressContext.tsx

@@ -407,7 +407,11 @@ export function KeypressProvider({
         return;
       }
 
-      if (rawDataBuffer.length <= 2 || isPaste) {
+      if (
+        (rawDataBuffer.length <= 2 && rawDataBuffer.includes(0x0d)) ||
+        !rawDataBuffer.includes(0x0d) ||
+        isPaste
+      ) {
         keypressStream.write(rawDataBuffer);
       } else {
         // Flush raw data buffer as a paste event

packages/cli/src/ui/hooks/slashCommandProcessor.test.ts

@@ -106,6 +106,7 @@ describe('useSlashCommandProcessor', () => {
   const mockLoadHistory = vi.fn();
   const mockOpenThemeDialog = vi.fn();
   const mockOpenAuthDialog = vi.fn();
+  const mockOpenModelSelectionDialog = vi.fn();
   const mockSetQuittingMessages = vi.fn();
 
   const mockConfig = makeFakeConfig({});
@@ -122,6 +123,7 @@ describe('useSlashCommandProcessor', () => {
     mockBuiltinLoadCommands.mockResolvedValue([]);
     mockFileLoadCommands.mockResolvedValue([]);
     mockMcpLoadCommands.mockResolvedValue([]);
+    mockOpenModelSelectionDialog.mockClear();
   });
 
   const setupProcessorHook = (
@@ -150,11 +152,13 @@ describe('useSlashCommandProcessor', () => {
         mockSetQuittingMessages,
         vi.fn(), // openPrivacyNotice
         vi.fn(), // openSettingsDialog
+        mockOpenModelSelectionDialog,
         vi.fn(), // openSubagentCreateDialog
         vi.fn(), // openAgentsManagerDialog
         vi.fn(), // toggleVimEnabled
         setIsProcessing,
         vi.fn(), // setGeminiMdFileCount
+        vi.fn(), // _showQuitConfirmation
       ),
     );
 
@@ -395,6 +399,21 @@ describe('useSlashCommandProcessor', () => {
       expect(mockOpenThemeDialog).toHaveBeenCalled();
     });
 
+    it('should handle "dialog: model" action', async () => {
+      const command = createTestCommand({
+        name: 'modelcmd',
+        action: vi.fn().mockResolvedValue({ type: 'dialog', dialog: 'model' }),
+      });
+      const result = setupProcessorHook([command]);
+      await waitFor(() => expect(result.current.slashCommands).toHaveLength(1));
+
+      await act(async () => {
+        await result.current.handleSlashCommand('/modelcmd');
+      });
+
+      expect(mockOpenModelSelectionDialog).toHaveBeenCalled();
+    });
+
     it('should handle "load_history" action', async () => {
       const command = createTestCommand({
         name: 'load',
@@ -904,11 +923,13 @@ describe('useSlashCommandProcessor', () => {
           mockSetQuittingMessages,
           vi.fn(), // openPrivacyNotice
           vi.fn(), // openSettingsDialog
+          vi.fn(), // openModelSelectionDialog
           vi.fn(), // openSubagentCreateDialog
           vi.fn(), // openAgentsManagerDialog
           vi.fn(), // toggleVimEnabled
           vi.fn(), // setIsProcessing
           vi.fn(), // setGeminiMdFileCount
+          vi.fn(), // _showQuitConfirmation
         ),
       );
 

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

@@ -53,6 +53,7 @@ export const useSlashCommandProcessor = (
   setQuittingMessages: (message: HistoryItem[]) => void,
   openPrivacyNotice: () => void,
   openSettingsDialog: () => void,
+  openModelSelectionDialog: () => void,
   openSubagentCreateDialog: () => void,
   openAgentsManagerDialog: () => void,
   toggleVimEnabled: () => Promise<boolean>,
@@ -404,6 +405,9 @@ export const useSlashCommandProcessor = (
                     case 'settings':
                       openSettingsDialog();
                       return { type: 'handled' };
+                    case 'model':
+                      openModelSelectionDialog();
+                      return { type: 'handled' };
                     case 'subagent_create':
                       openSubagentCreateDialog();
                       return { type: 'handled' };
@@ -663,6 +667,7 @@ export const useSlashCommandProcessor = (
       setSessionShellAllowlist,
       setIsProcessing,
       setConfirmationRequest,
+      openModelSelectionDialog,
       session.stats,
     ],
   );

packages/cli/src/ui/hooks/useAutoAcceptIndicator.test.ts

@@ -158,7 +158,19 @@ describe('useAutoAcceptIndicator', () => {
     expect(mockConfigInstance.getApprovalMode).toHaveBeenCalledTimes(1);
   });
 
-  it('should toggle the indicator and update config when Shift+Tab or Ctrl+Y is pressed', () => {
+  it('should initialize with ApprovalMode.PLAN if config.getApprovalMode returns ApprovalMode.PLAN', () => {
+    mockConfigInstance.getApprovalMode.mockReturnValue(ApprovalMode.PLAN);
+    const { result } = renderHook(() =>
+      useAutoAcceptIndicator({
+        config: mockConfigInstance as unknown as ActualConfigType,
+        addItem: vi.fn(),
+      }),
+    );
+    expect(result.current).toBe(ApprovalMode.PLAN);
+    expect(mockConfigInstance.getApprovalMode).toHaveBeenCalledTimes(1);
+  });
+
+  it('should cycle approval modes when Shift+Tab is pressed', () => {
     mockConfigInstance.getApprovalMode.mockReturnValue(ApprovalMode.DEFAULT);
     const { result } = renderHook(() =>
       useAutoAcceptIndicator({
@@ -180,23 +192,10 @@ describe('useAutoAcceptIndicator', () => {
     expect(result.current).toBe(ApprovalMode.AUTO_EDIT);
 
     act(() => {
-      capturedUseKeypressHandler({ name: 'y', ctrl: true } as Key);
-    });
-    expect(mockConfigInstance.setApprovalMode).toHaveBeenCalledWith(
-      ApprovalMode.YOLO,
-    );
-    expect(result.current).toBe(ApprovalMode.YOLO);
-
-    act(() => {
-      capturedUseKeypressHandler({ name: 'y', ctrl: true } as Key);
-    });
-    expect(mockConfigInstance.setApprovalMode).toHaveBeenCalledWith(
-      ApprovalMode.DEFAULT,
-    );
-    expect(result.current).toBe(ApprovalMode.DEFAULT);
-
-    act(() => {
-      capturedUseKeypressHandler({ name: 'y', ctrl: true } as Key);
+      capturedUseKeypressHandler({
+        name: 'tab',
+        shift: true,
+      } as Key);
     });
     expect(mockConfigInstance.setApprovalMode).toHaveBeenCalledWith(
       ApprovalMode.YOLO,
@@ -210,9 +209,9 @@ describe('useAutoAcceptIndicator', () => {
       } as Key);
     });
     expect(mockConfigInstance.setApprovalMode).toHaveBeenCalledWith(
-      ApprovalMode.AUTO_EDIT,
+      ApprovalMode.PLAN,
     );
-    expect(result.current).toBe(ApprovalMode.AUTO_EDIT);
+    expect(result.current).toBe(ApprovalMode.PLAN);
 
     act(() => {
       capturedUseKeypressHandler({
@@ -314,118 +313,10 @@ describe('useAutoAcceptIndicator', () => {
       mockConfigInstance.isTrustedFolder.mockReturnValue(false);
     });
 
-    it('should not enable YOLO mode when Ctrl+Y is pressed', () => {
-      mockConfigInstance.getApprovalMode.mockReturnValue(ApprovalMode.DEFAULT);
-      mockConfigInstance.setApprovalMode.mockImplementation(() => {
-        throw new Error(
-          'Cannot enable privileged approval modes in an untrusted folder.',
-        );
-      });
-      const mockAddItem = vi.fn();
-      const { result } = renderHook(() =>
-        useAutoAcceptIndicator({
-          config: mockConfigInstance as unknown as ActualConfigType,
-          addItem: mockAddItem,
-        }),
-      );
-
-      expect(result.current).toBe(ApprovalMode.DEFAULT);
-
-      act(() => {
-        capturedUseKeypressHandler({ name: 'y', ctrl: true } as Key);
-      });
-
-      // We expect setApprovalMode to be called, and the error to be caught.
-      expect(mockConfigInstance.setApprovalMode).toHaveBeenCalledWith(
-        ApprovalMode.YOLO,
-      );
-      expect(mockAddItem).toHaveBeenCalled();
-      // Verify the underlying config value was not changed
-      expect(mockConfigInstance.getApprovalMode()).toBe(ApprovalMode.DEFAULT);
-    });
-
-    it('should not enable AUTO_EDIT mode when Shift+Tab is pressed', () => {
-      mockConfigInstance.getApprovalMode.mockReturnValue(ApprovalMode.DEFAULT);
-      mockConfigInstance.setApprovalMode.mockImplementation(() => {
-        throw new Error(
-          'Cannot enable privileged approval modes in an untrusted folder.',
-        );
-      });
-      const mockAddItem = vi.fn();
-      const { result } = renderHook(() =>
-        useAutoAcceptIndicator({
-          config: mockConfigInstance as unknown as ActualConfigType,
-          addItem: mockAddItem,
-        }),
-      );
-
-      expect(result.current).toBe(ApprovalMode.DEFAULT);
-
-      act(() => {
-        capturedUseKeypressHandler({
-          name: 'tab',
-          shift: true,
-        } as Key);
-      });
-
-      // We expect setApprovalMode to be called, and the error to be caught.
-      expect(mockConfigInstance.setApprovalMode).toHaveBeenCalledWith(
-        ApprovalMode.AUTO_EDIT,
-      );
-      expect(mockAddItem).toHaveBeenCalled();
-      // Verify the underlying config value was not changed
-      expect(mockConfigInstance.getApprovalMode()).toBe(ApprovalMode.DEFAULT);
-    });
-
-    it('should disable YOLO mode when Ctrl+Y is pressed', () => {
-      mockConfigInstance.getApprovalMode.mockReturnValue(ApprovalMode.YOLO);
-      const mockAddItem = vi.fn();
-      renderHook(() =>
-        useAutoAcceptIndicator({
-          config: mockConfigInstance as unknown as ActualConfigType,
-          addItem: mockAddItem,
-        }),
-      );
-
-      act(() => {
-        capturedUseKeypressHandler({ name: 'y', ctrl: true } as Key);
-      });
-
-      expect(mockConfigInstance.setApprovalMode).toHaveBeenCalledWith(
-        ApprovalMode.DEFAULT,
-      );
-      expect(mockConfigInstance.getApprovalMode()).toBe(ApprovalMode.DEFAULT);
-    });
-
-    it('should disable AUTO_EDIT mode when Shift+Tab is pressed', () => {
-      mockConfigInstance.getApprovalMode.mockReturnValue(
-        ApprovalMode.AUTO_EDIT,
-      );
-      const mockAddItem = vi.fn();
-      renderHook(() =>
-        useAutoAcceptIndicator({
-          config: mockConfigInstance as unknown as ActualConfigType,
-          addItem: mockAddItem,
-        }),
-      );
-
-      act(() => {
-        capturedUseKeypressHandler({
-          name: 'tab',
-          shift: true,
-        } as Key);
-      });
-
-      expect(mockConfigInstance.setApprovalMode).toHaveBeenCalledWith(
-        ApprovalMode.DEFAULT,
-      );
-      expect(mockConfigInstance.getApprovalMode()).toBe(ApprovalMode.DEFAULT);
-    });
-
-    it('should show a warning when trying to enable privileged modes', () => {
-      // Mock the error thrown by setApprovalMode
+    it('should show a warning when cycling from DEFAULT to AUTO_EDIT', () => {
       const errorMessage =
         'Cannot enable privileged approval modes in an untrusted folder.';
+      mockConfigInstance.getApprovalMode.mockReturnValue(ApprovalMode.DEFAULT);
       mockConfigInstance.setApprovalMode.mockImplementation(() => {
         throw new Error(errorMessage);
       });
@@ -438,11 +329,13 @@ describe('useAutoAcceptIndicator', () => {
         }),
       );
 
-      // Try to enable YOLO mode
       act(() => {
-        capturedUseKeypressHandler({ name: 'y', ctrl: true } as Key);
+        capturedUseKeypressHandler({ name: 'tab', shift: true } as Key);
       });
 
+      expect(mockConfigInstance.setApprovalMode).toHaveBeenCalledWith(
+        ApprovalMode.AUTO_EDIT,
+      );
       expect(mockAddItem).toHaveBeenCalledWith(
         {
           type: MessageType.INFO,
@@ -450,15 +343,33 @@ describe('useAutoAcceptIndicator', () => {
         },
         expect.any(Number),
       );
+    });
 
-      // Try to enable AUTO_EDIT mode
-      act(() => {
-        capturedUseKeypressHandler({
-          name: 'tab',
-          shift: true,
-        } as Key);
+    it('should show a warning when cycling from AUTO_EDIT to YOLO', () => {
+      const errorMessage =
+        'Cannot enable privileged approval modes in an untrusted folder.';
+      mockConfigInstance.getApprovalMode.mockReturnValue(
+        ApprovalMode.AUTO_EDIT,
+      );
+      mockConfigInstance.setApprovalMode.mockImplementation(() => {
+        throw new Error(errorMessage);
       });
 
+      const mockAddItem = vi.fn();
+      renderHook(() =>
+        useAutoAcceptIndicator({
+          config: mockConfigInstance as unknown as ActualConfigType,
+          addItem: mockAddItem,
+        }),
+      );
+
+      act(() => {
+        capturedUseKeypressHandler({ name: 'tab', shift: true } as Key);
+      });
+
+      expect(mockConfigInstance.setApprovalMode).toHaveBeenCalledWith(
+        ApprovalMode.YOLO,
+      );
       expect(mockAddItem).toHaveBeenCalledWith(
         {
           type: MessageType.INFO,
@@ -466,8 +377,27 @@ describe('useAutoAcceptIndicator', () => {
         },
         expect.any(Number),
       );
+    });
 
-      expect(mockAddItem).toHaveBeenCalledTimes(2);
+    it('should cycle from YOLO to PLAN when Shift+Tab is pressed', () => {
+      mockConfigInstance.getApprovalMode.mockReturnValue(ApprovalMode.YOLO);
+      const mockAddItem = vi.fn();
+      renderHook(() =>
+        useAutoAcceptIndicator({
+          config: mockConfigInstance as unknown as ActualConfigType,
+          addItem: mockAddItem,
+        }),
+      );
+
+      act(() => {
+        capturedUseKeypressHandler({ name: 'tab', shift: true } as Key);
+      });
+
+      expect(mockConfigInstance.setApprovalMode).toHaveBeenCalledWith(
+        ApprovalMode.PLAN,
+      );
+      expect(mockConfigInstance.getApprovalMode()).toBe(ApprovalMode.PLAN);
+      expect(mockAddItem).not.toHaveBeenCalled();
     });
   });
 });

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

@@ -4,7 +4,11 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-import { ApprovalMode, type Config } from '@qwen-code/qwen-code-core';
+import {
+  type ApprovalMode,
+  APPROVAL_MODES,
+  type Config,
+} from '@qwen-code/qwen-code-core';
 import { useEffect, useState } from 'react';
 import { useKeypress } from './useKeypress.js';
 import type { HistoryItemWithoutId } from '../types.js';
@@ -29,34 +33,28 @@ export function useAutoAcceptIndicator({
 
   useKeypress(
     (key) => {
-      let nextApprovalMode: ApprovalMode | undefined;
-
-      if (key.ctrl && key.name === 'y') {
-        nextApprovalMode =
-          config.getApprovalMode() === ApprovalMode.YOLO
-            ? ApprovalMode.DEFAULT
-            : ApprovalMode.YOLO;
-      } else if (key.shift && key.name === 'tab') {
-        nextApprovalMode =
-          config.getApprovalMode() === ApprovalMode.AUTO_EDIT
-            ? ApprovalMode.DEFAULT
-            : ApprovalMode.AUTO_EDIT;
+      if (!(key.shift && key.name === 'tab')) {
+        return;
       }
 
-      if (nextApprovalMode) {
-        try {
-          config.setApprovalMode(nextApprovalMode);
-          // Update local state immediately for responsiveness
-          setShowAutoAcceptIndicator(nextApprovalMode);
-        } catch (e) {
-          addItem(
-            {
-              type: MessageType.INFO,
-              text: (e as Error).message,
-            },
-            Date.now(),
-          );
-        }
+      const currentMode = config.getApprovalMode();
+      const currentIndex = APPROVAL_MODES.indexOf(currentMode);
+      const nextIndex =
+        currentIndex === -1 ? 0 : (currentIndex + 1) % APPROVAL_MODES.length;
+      const nextApprovalMode = APPROVAL_MODES[nextIndex];
+
+      try {
+        config.setApprovalMode(nextApprovalMode);
+        // Update local state immediately for responsiveness
+        setShowAutoAcceptIndicator(nextApprovalMode);
+      } catch (e) {
+        addItem(
+          {
+            type: MessageType.INFO,
+            text: (e as Error).message,
+          },
+          Date.now(),
+        );
       }
     },
     { isActive: true },

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

@@ -56,6 +56,14 @@ const MockedUserPromptEvent = vi.hoisted(() =>
 );
 const mockParseAndFormatApiError = vi.hoisted(() => vi.fn());
 
+// Vision auto-switch mocks (hoisted)
+const mockHandleVisionSwitch = vi.hoisted(() =>
+  vi.fn().mockResolvedValue({ shouldProceed: true }),
+);
+const mockRestoreOriginalModel = vi.hoisted(() =>
+  vi.fn().mockResolvedValue(undefined),
+);
+
 vi.mock('@qwen-code/qwen-code-core', async (importOriginal) => {
   const actualCoreModule = (await importOriginal()) as any;
   return {
@@ -76,6 +84,13 @@ vi.mock('./useReactToolScheduler.js', async (importOriginal) => {
   };
 });
 
+vi.mock('./useVisionAutoSwitch.js', () => ({
+  useVisionAutoSwitch: vi.fn(() => ({
+    handleVisionSwitch: mockHandleVisionSwitch,
+    restoreOriginalModel: mockRestoreOriginalModel,
+  })),
+}));
+
 vi.mock('./useKeypress.js', () => ({
   useKeypress: vi.fn(),
 }));
@@ -199,6 +214,7 @@ describe('useGeminiStream', () => {
       getContentGeneratorConfig: vi
         .fn()
         .mockReturnValue(contentGeneratorConfig),
+      getMaxSessionTurns: vi.fn(() => 50),
     } as unknown as Config;
     mockOnDebugMessage = vi.fn();
     mockHandleSlashCommand = vi.fn().mockResolvedValue(false);
@@ -287,6 +303,8 @@ describe('useGeminiStream', () => {
           () => {},
           () => {},
           () => {},
+          false, // visionModelPreviewEnabled
+          undefined, // onVisionSwitchRequired (optional)
         );
       },
       {
@@ -448,6 +466,8 @@ describe('useGeminiStream', () => {
         () => {},
         () => {},
         () => {},
+        false, // visionModelPreviewEnabled
+        undefined, // onVisionSwitchRequired (optional)
       ),
     );
 
@@ -527,6 +547,8 @@ describe('useGeminiStream', () => {
         () => {},
         () => {},
         () => {},
+        false, // visionModelPreviewEnabled
+        undefined, // onVisionSwitchRequired (optional)
       ),
     );
 
@@ -635,6 +657,8 @@ describe('useGeminiStream', () => {
         () => {},
         () => {},
         () => {},
+        false, // visionModelPreviewEnabled
+        undefined, // onVisionSwitchRequired (optional)
       ),
     );
 
@@ -744,6 +768,8 @@ describe('useGeminiStream', () => {
         () => {},
         () => {},
         () => {},
+        false, // visionModelPreviewEnabled
+        undefined, // onVisionSwitchRequired (optional)
       ),
     );
 
@@ -873,6 +899,8 @@ describe('useGeminiStream', () => {
           () => {},
           () => {},
           cancelSubmitSpy,
+          false, // visionModelPreviewEnabled
+          undefined, // onVisionSwitchRequired (optional)
         ),
       );
 
@@ -1184,6 +1212,8 @@ describe('useGeminiStream', () => {
           () => {},
           () => {},
           () => {},
+          false, // visionModelPreviewEnabled
+          undefined, // onVisionSwitchRequired (optional)
         ),
       );
 
@@ -1237,6 +1267,8 @@ describe('useGeminiStream', () => {
           () => {},
           () => {},
           () => {},
+          false, // visionModelPreviewEnabled
+          undefined, // onVisionSwitchRequired (optional)
         ),
       );
 
@@ -1287,6 +1319,8 @@ describe('useGeminiStream', () => {
           () => {},
           () => {},
           () => {},
+          false, // visionModelPreviewEnabled
+          undefined, // onVisionSwitchRequired (optional)
         ),
       );
 
@@ -1335,6 +1369,8 @@ describe('useGeminiStream', () => {
           () => {},
           () => {},
           () => {},
+          false, // visionModelPreviewEnabled
+          undefined, // onVisionSwitchRequired (optional)
         ),
       );
 
@@ -1384,6 +1420,8 @@ describe('useGeminiStream', () => {
           () => {},
           () => {},
           () => {},
+          false, // visionModelPreviewEnabled
+          undefined, // onVisionSwitchRequired (optional)
         ),
       );
 
@@ -1473,6 +1511,8 @@ describe('useGeminiStream', () => {
             () => {},
             () => {},
             () => {},
+            false, // visionModelPreviewEnabled
+            undefined, // onVisionSwitchRequired (optional)
           ),
         );
 
@@ -1523,6 +1563,8 @@ describe('useGeminiStream', () => {
         vi.fn(), // setModelSwitched
         vi.fn(), // onEditorClose
         vi.fn(), // onCancelSubmit
+        false, // visionModelPreviewEnabled
+        undefined, // onVisionSwitchRequired (optional)
       ),
     );
 
@@ -1551,6 +1593,7 @@ describe('useGeminiStream', () => {
       expect.any(String), // Argument 3: The prompt_id string
     );
   });
+
   describe('Thought Reset', () => {
     it('should reset thought to null when starting a new prompt', async () => {
       // First, simulate a response with a thought
@@ -1587,6 +1630,8 @@ describe('useGeminiStream', () => {
           () => {},
           () => {},
           () => {},
+          false, // visionModelPreviewEnabled
+          undefined, // onVisionSwitchRequired (optional)
         ),
       );
 
@@ -1665,6 +1710,8 @@ describe('useGeminiStream', () => {
           () => {},
           () => {},
           () => {},
+          false, // visionModelPreviewEnabled
+          undefined, // onVisionSwitchRequired (optional)
         ),
       );
 
@@ -1719,6 +1766,8 @@ describe('useGeminiStream', () => {
           () => {},
           () => {},
           () => {},
+          false, // visionModelPreviewEnabled
+          undefined, // onVisionSwitchRequired (optional)
         ),
       );
 
@@ -1900,4 +1949,174 @@ describe('useGeminiStream', () => {
       );
     });
   });
+
+  // --- New tests focused on recent modifications ---
+  describe('Vision Auto Switch Integration', () => {
+    it('should call handleVisionSwitch and proceed to send when allowed', async () => {
+      mockHandleVisionSwitch.mockResolvedValueOnce({ shouldProceed: true });
+      mockSendMessageStream.mockReturnValue(
+        (async function* () {
+          yield { type: ServerGeminiEventType.Content, value: 'ok' };
+          yield { type: ServerGeminiEventType.Finished, value: 'STOP' };
+        })(),
+      );
+
+      const { result } = renderHook(() =>
+        useGeminiStream(
+          new MockedGeminiClientClass(mockConfig),
+          [],
+          mockAddItem,
+          mockConfig,
+          mockOnDebugMessage,
+          mockHandleSlashCommand,
+          false,
+          () => 'vscode' as EditorType,
+          () => {},
+          () => Promise.resolve(),
+          false,
+          () => {},
+          () => {},
+          () => {},
+          false, // visionModelPreviewEnabled
+          undefined, // onVisionSwitchRequired (optional)
+        ),
+      );
+
+      await act(async () => {
+        await result.current.submitQuery('image prompt');
+      });
+
+      await waitFor(() => {
+        expect(mockHandleVisionSwitch).toHaveBeenCalled();
+        expect(mockSendMessageStream).toHaveBeenCalled();
+      });
+    });
+
+    it('should gate submission when handleVisionSwitch returns shouldProceed=false', async () => {
+      mockHandleVisionSwitch.mockResolvedValueOnce({ shouldProceed: false });
+
+      const { result } = renderHook(() =>
+        useGeminiStream(
+          new MockedGeminiClientClass(mockConfig),
+          [],
+          mockAddItem,
+          mockConfig,
+          mockOnDebugMessage,
+          mockHandleSlashCommand,
+          false,
+          () => 'vscode' as EditorType,
+          () => {},
+          () => Promise.resolve(),
+          false,
+          () => {},
+          () => {},
+          () => {},
+          false, // visionModelPreviewEnabled
+          undefined, // onVisionSwitchRequired (optional)
+        ),
+      );
+
+      await act(async () => {
+        await result.current.submitQuery('vision-gated');
+      });
+
+      // No call to API, no restoreOriginalModel needed since no override occurred
+      expect(mockSendMessageStream).not.toHaveBeenCalled();
+      expect(mockRestoreOriginalModel).not.toHaveBeenCalled();
+
+      // Next call allowed (flag reset path)
+      mockHandleVisionSwitch.mockResolvedValueOnce({ shouldProceed: true });
+      mockSendMessageStream.mockReturnValue(
+        (async function* () {
+          yield { type: ServerGeminiEventType.Content, value: 'ok' };
+          yield { type: ServerGeminiEventType.Finished, value: 'STOP' };
+        })(),
+      );
+      await act(async () => {
+        await result.current.submitQuery('after-gate');
+      });
+      await waitFor(() => {
+        expect(mockSendMessageStream).toHaveBeenCalled();
+      });
+    });
+  });
+
+  describe('Model restore on completion and errors', () => {
+    it('should restore model after successful stream completion', async () => {
+      mockSendMessageStream.mockReturnValue(
+        (async function* () {
+          yield { type: ServerGeminiEventType.Content, value: 'content' };
+          yield { type: ServerGeminiEventType.Finished, value: 'STOP' };
+        })(),
+      );
+
+      const { result } = renderHook(() =>
+        useGeminiStream(
+          new MockedGeminiClientClass(mockConfig),
+          [],
+          mockAddItem,
+          mockConfig,
+          mockOnDebugMessage,
+          mockHandleSlashCommand,
+          false,
+          () => 'vscode' as EditorType,
+          () => {},
+          () => Promise.resolve(),
+          false,
+          () => {},
+          () => {},
+          () => {},
+          false, // visionModelPreviewEnabled
+          undefined, // onVisionSwitchRequired (optional)
+        ),
+      );
+
+      await act(async () => {
+        await result.current.submitQuery('restore-success');
+      });
+
+      await waitFor(() => {
+        expect(mockRestoreOriginalModel).toHaveBeenCalledTimes(1);
+      });
+    });
+
+    it('should restore model when an error occurs during streaming', async () => {
+      const testError = new Error('stream failure');
+      mockSendMessageStream.mockReturnValue(
+        (async function* () {
+          yield { type: ServerGeminiEventType.Content, value: 'content' };
+          throw testError;
+        })(),
+      );
+
+      const { result } = renderHook(() =>
+        useGeminiStream(
+          new MockedGeminiClientClass(mockConfig),
+          [],
+          mockAddItem,
+          mockConfig,
+          mockOnDebugMessage,
+          mockHandleSlashCommand,
+          false,
+          () => 'vscode' as EditorType,
+          () => {},
+          () => Promise.resolve(),
+          false,
+          () => {},
+          () => {},
+          () => {},
+          false, // visionModelPreviewEnabled
+          undefined, // onVisionSwitchRequired (optional)
+        ),
+      );
+
+      await act(async () => {
+        await result.current.submitQuery('restore-error');
+      });
+
+      await waitFor(() => {
+        expect(mockRestoreOriginalModel).toHaveBeenCalledTimes(1);
+      });
+    });
+  });
 });

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

@@ -42,6 +42,7 @@ import type {
 import { StreamingState, MessageType, ToolCallStatus } from '../types.js';
 import { isAtCommand, isSlashCommand } from '../utils/commandUtils.js';
 import { useShellCommandProcessor } from './shellCommandProcessor.js';
+import { useVisionAutoSwitch } from './useVisionAutoSwitch.js';
 import { handleAtCommand } from './atCommandProcessor.js';
 import { findLastSafeSplitPoint } from '../utils/markdownUtilities.js';
 import { useStateAndRef } from './useStateAndRef.js';
@@ -88,6 +89,12 @@ export const useGeminiStream = (
   setModelSwitchedFromQuotaError: React.Dispatch<React.SetStateAction<boolean>>,
   onEditorClose: () => void,
   onCancelSubmit: () => void,
+  visionModelPreviewEnabled: boolean,
+  onVisionSwitchRequired?: (query: PartListUnion) => Promise<{
+    modelOverride?: string;
+    persistSessionModel?: string;
+    showGuidance?: boolean;
+  }>,
 ) => {
   const [initError, setInitError] = useState<string | null>(null);
   const abortControllerRef = useRef<AbortController | null>(null);
@@ -155,6 +162,13 @@ export const useGeminiStream = (
     geminiClient,
   );
 
+  const { handleVisionSwitch, restoreOriginalModel } = useVisionAutoSwitch(
+    config,
+    addItem,
+    visionModelPreviewEnabled,
+    onVisionSwitchRequired,
+  );
+
   const streamingState = useMemo(() => {
     if (toolCalls.some((tc) => tc.status === 'awaiting_approval')) {
       return StreamingState.WaitingForConfirmation;
@@ -715,6 +729,20 @@ export const useGeminiStream = (
         return;
       }
 
+      // Handle vision switch requirement
+      const visionSwitchResult = await handleVisionSwitch(
+        queryToSend,
+        userMessageTimestamp,
+        options?.isContinuation || false,
+      );
+
+      if (!visionSwitchResult.shouldProceed) {
+        isSubmittingQueryRef.current = false;
+        return;
+      }
+
+      const finalQueryToSend = queryToSend;
+
       if (!options?.isContinuation) {
         startNewPrompt();
         setThought(null); // Reset thought when starting a new prompt
@@ -725,7 +753,7 @@ export const useGeminiStream = (
 
       try {
         const stream = geminiClient.sendMessageStream(
-          queryToSend,
+          finalQueryToSend,
           abortSignal,
           prompt_id!,
         );
@@ -736,6 +764,10 @@ export const useGeminiStream = (
         );
 
         if (processingStatus === StreamProcessingStatus.UserCancelled) {
+          // Restore original model if it was temporarily overridden
+          restoreOriginalModel().catch((error) => {
+            console.error('Failed to restore original model:', error);
+          });
           isSubmittingQueryRef.current = false;
           return;
         }
@@ -748,7 +780,17 @@ export const useGeminiStream = (
           loopDetectedRef.current = false;
           handleLoopDetectedEvent();
         }
+
+        // Restore original model if it was temporarily overridden
+        restoreOriginalModel().catch((error) => {
+          console.error('Failed to restore original model:', error);
+        });
       } catch (error: unknown) {
+        // Restore original model if it was temporarily overridden
+        restoreOriginalModel().catch((error) => {
+          console.error('Failed to restore original model:', error);
+        });
+
         if (error instanceof UnauthorizedError) {
           onAuthError();
         } else if (!isNodeError(error) || error.name !== 'AbortError') {
@@ -786,6 +828,8 @@ export const useGeminiStream = (
       startNewPrompt,
       getPromptCount,
       handleLoopDetectedEvent,
+      handleVisionSwitch,
+      restoreOriginalModel,
     ],
   );
 
@@ -911,10 +955,13 @@ export const useGeminiStream = (
     ],
   );
 
-  const pendingHistoryItems = [
-    pendingHistoryItemRef.current,
-    pendingToolCallGroupDisplay,
-  ].filter((i) => i !== undefined && i !== null);
+  const pendingHistoryItems = useMemo(
+    () =>
+      [pendingHistoryItemRef.current, pendingToolCallGroupDisplay].filter(
+        (i) => i !== undefined && i !== null,
+      ),
+    [pendingHistoryItemRef, pendingToolCallGroupDisplay],
+  );
 
   useEffect(() => {
     const saveRestorableToolCalls = async () => {

packages/cli/src/ui/hooks/useVisionAutoSwitch.test.ts

@@ -0,0 +1,874 @@
+/**
+ * @license
+ * Copyright 2025 Qwen
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/* eslint-disable @typescript-eslint/no-explicit-any */
+import { describe, it, expect, vi, beforeEach } from 'vitest';
+import { renderHook, act } from '@testing-library/react';
+import type { Part, PartListUnion } from '@google/genai';
+import { AuthType, type Config, ApprovalMode } from '@qwen-code/qwen-code-core';
+
+// Mock the image format functions from core package
+vi.mock('@qwen-code/qwen-code-core', async (importOriginal) => {
+  const actual = (await importOriginal()) as Record<string, unknown>;
+  return {
+    ...actual,
+    isSupportedImageMimeType: vi.fn((mimeType: string) =>
+      [
+        'image/png',
+        'image/jpeg',
+        'image/jpg',
+        'image/gif',
+        'image/webp',
+      ].includes(mimeType),
+    ),
+    getUnsupportedImageFormatWarning: vi.fn(
+      () =>
+        'Only the following image formats are supported: BMP, JPEG, JPG, PNG, TIFF, WEBP, HEIC. Other formats may not work as expected.',
+    ),
+  };
+});
+import {
+  shouldOfferVisionSwitch,
+  processVisionSwitchOutcome,
+  getVisionSwitchGuidanceMessage,
+  useVisionAutoSwitch,
+} from './useVisionAutoSwitch.js';
+import { VisionSwitchOutcome } from '../components/ModelSwitchDialog.js';
+import { MessageType } from '../types.js';
+import { getDefaultVisionModel } from '../models/availableModels.js';
+
+describe('useVisionAutoSwitch helpers', () => {
+  describe('shouldOfferVisionSwitch', () => {
+    it('returns false when authType is not QWEN_OAUTH', () => {
+      const parts: PartListUnion = [
+        { inlineData: { mimeType: 'image/png', data: '...' } },
+      ];
+      const result = shouldOfferVisionSwitch(
+        parts,
+        AuthType.USE_GEMINI,
+        'qwen3-coder-plus',
+        true,
+      );
+      expect(result).toBe(false);
+    });
+
+    it('returns false when current model is already a vision model', () => {
+      const parts: PartListUnion = [
+        { inlineData: { mimeType: 'image/png', data: '...' } },
+      ];
+      const result = shouldOfferVisionSwitch(
+        parts,
+        AuthType.QWEN_OAUTH,
+        'vision-model',
+        true,
+      );
+      expect(result).toBe(false);
+    });
+
+    it('returns true when image parts exist, QWEN_OAUTH, and model is not vision', () => {
+      const parts: PartListUnion = [
+        { text: 'hello' },
+        { inlineData: { mimeType: 'image/jpeg', data: '...' } },
+      ];
+      const result = shouldOfferVisionSwitch(
+        parts,
+        AuthType.QWEN_OAUTH,
+        'qwen3-coder-plus',
+        true,
+      );
+      expect(result).toBe(true);
+    });
+
+    it('detects image when provided as a single Part object (non-array)', () => {
+      const singleImagePart: PartListUnion = {
+        fileData: { mimeType: 'image/gif', fileUri: 'file://image.gif' },
+      } as Part;
+      const result = shouldOfferVisionSwitch(
+        singleImagePart,
+        AuthType.QWEN_OAUTH,
+        'qwen3-coder-plus',
+        true,
+      );
+      expect(result).toBe(true);
+    });
+
+    it('returns false when parts contain no images', () => {
+      const parts: PartListUnion = [{ text: 'just text' }];
+      const result = shouldOfferVisionSwitch(
+        parts,
+        AuthType.QWEN_OAUTH,
+        'qwen3-coder-plus',
+        true,
+      );
+      expect(result).toBe(false);
+    });
+
+    it('returns false when parts is a plain string', () => {
+      const parts: PartListUnion = 'plain text';
+      const result = shouldOfferVisionSwitch(
+        parts,
+        AuthType.QWEN_OAUTH,
+        'qwen3-coder-plus',
+        true,
+      );
+      expect(result).toBe(false);
+    });
+
+    it('returns false when visionModelPreviewEnabled is false', () => {
+      const parts: PartListUnion = [
+        { inlineData: { mimeType: 'image/png', data: '...' } },
+      ];
+      const result = shouldOfferVisionSwitch(
+        parts,
+        AuthType.QWEN_OAUTH,
+        'qwen3-coder-plus',
+        false,
+      );
+      expect(result).toBe(false);
+    });
+
+    it('returns true when image parts exist in YOLO mode context', () => {
+      const parts: PartListUnion = [
+        { inlineData: { mimeType: 'image/png', data: '...' } },
+      ];
+      const result = shouldOfferVisionSwitch(
+        parts,
+        AuthType.QWEN_OAUTH,
+        'qwen3-coder-plus',
+        true,
+      );
+      expect(result).toBe(true);
+    });
+
+    it('returns false when no image parts exist in YOLO mode context', () => {
+      const parts: PartListUnion = [{ text: 'just text' }];
+      const result = shouldOfferVisionSwitch(
+        parts,
+        AuthType.QWEN_OAUTH,
+        'qwen3-coder-plus',
+        true,
+      );
+      expect(result).toBe(false);
+    });
+
+    it('returns false when already using vision model in YOLO mode context', () => {
+      const parts: PartListUnion = [
+        { inlineData: { mimeType: 'image/png', data: '...' } },
+      ];
+      const result = shouldOfferVisionSwitch(
+        parts,
+        AuthType.QWEN_OAUTH,
+        'vision-model',
+        true,
+      );
+      expect(result).toBe(false);
+    });
+
+    it('returns false when authType is not QWEN_OAUTH in YOLO mode context', () => {
+      const parts: PartListUnion = [
+        { inlineData: { mimeType: 'image/png', data: '...' } },
+      ];
+      const result = shouldOfferVisionSwitch(
+        parts,
+        AuthType.USE_GEMINI,
+        'qwen3-coder-plus',
+        true,
+      );
+      expect(result).toBe(false);
+    });
+  });
+
+  describe('processVisionSwitchOutcome', () => {
+    it('maps SwitchOnce to a one-time model override', () => {
+      const vl = getDefaultVisionModel();
+      const result = processVisionSwitchOutcome(VisionSwitchOutcome.SwitchOnce);
+      expect(result).toEqual({ modelOverride: vl });
+    });
+
+    it('maps SwitchSessionToVL to a persistent session model', () => {
+      const vl = getDefaultVisionModel();
+      const result = processVisionSwitchOutcome(
+        VisionSwitchOutcome.SwitchSessionToVL,
+      );
+      expect(result).toEqual({ persistSessionModel: vl });
+    });
+
+    it('maps ContinueWithCurrentModel to empty result', () => {
+      const result = processVisionSwitchOutcome(
+        VisionSwitchOutcome.ContinueWithCurrentModel,
+      );
+      expect(result).toEqual({});
+    });
+  });
+
+  describe('getVisionSwitchGuidanceMessage', () => {
+    it('returns the expected guidance message', () => {
+      const vl = getDefaultVisionModel();
+      const expected =
+        'To use images with your query, you can:\n' +
+        `• Use /model set ${vl} to switch to a vision-capable model\n` +
+        '• Or remove the image and provide a text description instead';
+      expect(getVisionSwitchGuidanceMessage()).toBe(expected);
+    });
+  });
+});
+
+describe('useVisionAutoSwitch hook', () => {
+  type AddItemFn = (
+    item: { type: MessageType; text: string },
+    ts: number,
+  ) => any;
+
+  const createMockConfig = (
+    authType: AuthType,
+    initialModel: string,
+    approvalMode: ApprovalMode = ApprovalMode.DEFAULT,
+    vlmSwitchMode?: string,
+  ) => {
+    let currentModel = initialModel;
+    const mockConfig: Partial<Config> = {
+      getModel: vi.fn(() => currentModel),
+      setModel: vi.fn(async (m: string) => {
+        currentModel = m;
+      }),
+      getApprovalMode: vi.fn(() => approvalMode),
+      getVlmSwitchMode: vi.fn(() => vlmSwitchMode),
+      getContentGeneratorConfig: vi.fn(() => ({
+        authType,
+        model: currentModel,
+        apiKey: 'test-key',
+        vertexai: false,
+      })),
+    };
+    return mockConfig as Config;
+  };
+
+  let addItem: AddItemFn;
+
+  beforeEach(() => {
+    vi.clearAllMocks();
+    addItem = vi.fn();
+  });
+
+  it('returns shouldProceed=true immediately for continuations', async () => {
+    const config = createMockConfig(AuthType.QWEN_OAUTH, 'qwen3-coder-plus');
+    const { result } = renderHook(() =>
+      useVisionAutoSwitch(config, addItem as any, true, vi.fn()),
+    );
+
+    const parts: PartListUnion = [
+      { inlineData: { mimeType: 'image/png', data: '...' } },
+    ];
+    let res: any;
+    await act(async () => {
+      res = await result.current.handleVisionSwitch(parts, Date.now(), true);
+    });
+    expect(res).toEqual({ shouldProceed: true });
+    expect(addItem).not.toHaveBeenCalled();
+  });
+
+  it('does nothing when authType is not QWEN_OAUTH', async () => {
+    const config = createMockConfig(AuthType.USE_GEMINI, 'qwen3-coder-plus');
+    const onVisionSwitchRequired = vi.fn();
+    const { result } = renderHook(() =>
+      useVisionAutoSwitch(config, addItem as any, true, onVisionSwitchRequired),
+    );
+
+    const parts: PartListUnion = [
+      { inlineData: { mimeType: 'image/png', data: '...' } },
+    ];
+    let res: any;
+    await act(async () => {
+      res = await result.current.handleVisionSwitch(parts, 123, false);
+    });
+    expect(res).toEqual({ shouldProceed: true });
+    expect(onVisionSwitchRequired).not.toHaveBeenCalled();
+  });
+
+  it('does nothing when there are no image parts', async () => {
+    const config = createMockConfig(AuthType.QWEN_OAUTH, 'qwen3-coder-plus');
+    const onVisionSwitchRequired = vi.fn();
+    const { result } = renderHook(() =>
+      useVisionAutoSwitch(config, addItem as any, true, onVisionSwitchRequired),
+    );
+
+    const parts: PartListUnion = [{ text: 'no images here' }];
+    let res: any;
+    await act(async () => {
+      res = await result.current.handleVisionSwitch(parts, 456, false);
+    });
+    expect(res).toEqual({ shouldProceed: true });
+    expect(onVisionSwitchRequired).not.toHaveBeenCalled();
+  });
+
+  it('continues with current model when dialog returns empty result', async () => {
+    const config = createMockConfig(AuthType.QWEN_OAUTH, 'qwen3-coder-plus');
+    const onVisionSwitchRequired = vi.fn().mockResolvedValue({}); // Empty result for ContinueWithCurrentModel
+    const { result } = renderHook(() =>
+      useVisionAutoSwitch(config, addItem as any, true, onVisionSwitchRequired),
+    );
+
+    const parts: PartListUnion = [
+      { inlineData: { mimeType: 'image/png', data: '...' } },
+    ];
+
+    const userTs = 1010;
+    let res: any;
+    await act(async () => {
+      res = await result.current.handleVisionSwitch(parts, userTs, false);
+    });
+
+    // Should not add any guidance message
+    expect(addItem).not.toHaveBeenCalledWith(
+      { type: MessageType.INFO, text: getVisionSwitchGuidanceMessage() },
+      userTs,
+    );
+    expect(res).toEqual({ shouldProceed: true });
+    expect(config.setModel).not.toHaveBeenCalled();
+  });
+
+  it('applies a one-time override and returns originalModel, then restores', async () => {
+    const initialModel = 'qwen3-coder-plus';
+    const config = createMockConfig(AuthType.QWEN_OAUTH, initialModel);
+    const onVisionSwitchRequired = vi
+      .fn()
+      .mockResolvedValue({ modelOverride: 'coder-model' });
+    const { result } = renderHook(() =>
+      useVisionAutoSwitch(config, addItem as any, true, onVisionSwitchRequired),
+    );
+
+    const parts: PartListUnion = [
+      { inlineData: { mimeType: 'image/png', data: '...' } },
+    ];
+
+    let res: any;
+    await act(async () => {
+      res = await result.current.handleVisionSwitch(parts, 2020, false);
+    });
+
+    expect(res).toEqual({ shouldProceed: true, originalModel: initialModel });
+    expect(config.setModel).toHaveBeenCalledWith('coder-model', {
+      reason: 'vision_auto_switch',
+      context: 'User-prompted vision switch (one-time override)',
+    });
+
+    // Now restore
+    await act(async () => {
+      await result.current.restoreOriginalModel();
+    });
+    expect(config.setModel).toHaveBeenLastCalledWith(initialModel, {
+      reason: 'vision_auto_switch',
+      context: 'Restoring original model after vision switch',
+    });
+  });
+
+  it('persists session model when dialog requests persistence', async () => {
+    const config = createMockConfig(AuthType.QWEN_OAUTH, 'qwen3-coder-plus');
+    const onVisionSwitchRequired = vi
+      .fn()
+      .mockResolvedValue({ persistSessionModel: 'coder-model' });
+    const { result } = renderHook(() =>
+      useVisionAutoSwitch(config, addItem as any, true, onVisionSwitchRequired),
+    );
+
+    const parts: PartListUnion = [
+      { inlineData: { mimeType: 'image/png', data: '...' } },
+    ];
+
+    let res: any;
+    await act(async () => {
+      res = await result.current.handleVisionSwitch(parts, 3030, false);
+    });
+
+    expect(res).toEqual({ shouldProceed: true });
+    expect(config.setModel).toHaveBeenCalledWith('coder-model', {
+      reason: 'vision_auto_switch',
+      context: 'User-prompted vision switch (session persistent)',
+    });
+
+    // Restore should be a no-op since no one-time override was used
+    await act(async () => {
+      await result.current.restoreOriginalModel();
+    });
+    // Last call should still be the persisted model set
+    expect((config.setModel as any).mock.calls.pop()?.[0]).toBe('coder-model');
+  });
+
+  it('returns shouldProceed=true when dialog returns no special flags', async () => {
+    const config = createMockConfig(AuthType.QWEN_OAUTH, 'qwen3-coder-plus');
+    const onVisionSwitchRequired = vi.fn().mockResolvedValue({});
+    const { result } = renderHook(() =>
+      useVisionAutoSwitch(config, addItem as any, true, onVisionSwitchRequired),
+    );
+
+    const parts: PartListUnion = [
+      { inlineData: { mimeType: 'image/png', data: '...' } },
+    ];
+    let res: any;
+    await act(async () => {
+      res = await result.current.handleVisionSwitch(parts, 4040, false);
+    });
+    expect(res).toEqual({ shouldProceed: true });
+    expect(config.setModel).not.toHaveBeenCalled();
+  });
+
+  it('blocks when dialog throws or is cancelled', async () => {
+    const config = createMockConfig(AuthType.QWEN_OAUTH, 'qwen3-coder-plus');
+    const onVisionSwitchRequired = vi.fn().mockRejectedValue(new Error('x'));
+    const { result } = renderHook(() =>
+      useVisionAutoSwitch(config, addItem as any, true, onVisionSwitchRequired),
+    );
+
+    const parts: PartListUnion = [
+      { inlineData: { mimeType: 'image/png', data: '...' } },
+    ];
+    let res: any;
+    await act(async () => {
+      res = await result.current.handleVisionSwitch(parts, 5050, false);
+    });
+    expect(res).toEqual({ shouldProceed: false });
+    expect(config.setModel).not.toHaveBeenCalled();
+  });
+
+  it('does nothing when visionModelPreviewEnabled is false', async () => {
+    const config = createMockConfig(AuthType.QWEN_OAUTH, 'qwen3-coder-plus');
+    const onVisionSwitchRequired = vi.fn();
+    const { result } = renderHook(() =>
+      useVisionAutoSwitch(
+        config,
+        addItem as any,
+        false,
+        onVisionSwitchRequired,
+      ),
+    );
+
+    const parts: PartListUnion = [
+      { inlineData: { mimeType: 'image/png', data: '...' } },
+    ];
+    let res: any;
+    await act(async () => {
+      res = await result.current.handleVisionSwitch(parts, 6060, false);
+    });
+    expect(res).toEqual({ shouldProceed: true });
+    expect(onVisionSwitchRequired).not.toHaveBeenCalled();
+  });
+
+  describe('YOLO mode behavior', () => {
+    it('automatically switches to vision model in YOLO mode without showing dialog', async () => {
+      const initialModel = 'qwen3-coder-plus';
+      const config = createMockConfig(
+        AuthType.QWEN_OAUTH,
+        initialModel,
+        ApprovalMode.YOLO,
+      );
+      const onVisionSwitchRequired = vi.fn(); // Should not be called in YOLO mode
+      const { result } = renderHook(() =>
+        useVisionAutoSwitch(
+          config,
+          addItem as any,
+          true,
+          onVisionSwitchRequired,
+        ),
+      );
+
+      const parts: PartListUnion = [
+        { inlineData: { mimeType: 'image/png', data: '...' } },
+      ];
+
+      let res: any;
+      await act(async () => {
+        res = await result.current.handleVisionSwitch(parts, 7070, false);
+      });
+
+      // Should automatically switch without calling the dialog
+      expect(onVisionSwitchRequired).not.toHaveBeenCalled();
+      expect(res).toEqual({
+        shouldProceed: true,
+        originalModel: initialModel,
+      });
+      expect(config.setModel).toHaveBeenCalledWith(getDefaultVisionModel(), {
+        reason: 'vision_auto_switch',
+        context: 'YOLO mode auto-switch for image content',
+      });
+    });
+
+    it('does not switch in YOLO mode when no images are present', async () => {
+      const config = createMockConfig(
+        AuthType.QWEN_OAUTH,
+        'qwen3-coder-plus',
+        ApprovalMode.YOLO,
+      );
+      const onVisionSwitchRequired = vi.fn();
+      const { result } = renderHook(() =>
+        useVisionAutoSwitch(
+          config,
+          addItem as any,
+          true,
+          onVisionSwitchRequired,
+        ),
+      );
+
+      const parts: PartListUnion = [{ text: 'no images here' }];
+
+      let res: any;
+      await act(async () => {
+        res = await result.current.handleVisionSwitch(parts, 8080, false);
+      });
+
+      expect(res).toEqual({ shouldProceed: true });
+      expect(onVisionSwitchRequired).not.toHaveBeenCalled();
+      expect(config.setModel).not.toHaveBeenCalled();
+    });
+
+    it('does not switch in YOLO mode when already using vision model', async () => {
+      const config = createMockConfig(
+        AuthType.QWEN_OAUTH,
+        'vision-model',
+        ApprovalMode.YOLO,
+      );
+      const onVisionSwitchRequired = vi.fn();
+      const { result } = renderHook(() =>
+        useVisionAutoSwitch(
+          config,
+          addItem as any,
+          true,
+          onVisionSwitchRequired,
+        ),
+      );
+
+      const parts: PartListUnion = [
+        { inlineData: { mimeType: 'image/png', data: '...' } },
+      ];
+
+      let res: any;
+      await act(async () => {
+        res = await result.current.handleVisionSwitch(parts, 9090, false);
+      });
+
+      expect(res).toEqual({ shouldProceed: true });
+      expect(onVisionSwitchRequired).not.toHaveBeenCalled();
+      expect(config.setModel).not.toHaveBeenCalled();
+    });
+
+    it('restores original model after YOLO mode auto-switch', async () => {
+      const initialModel = 'qwen3-coder-plus';
+      const config = createMockConfig(
+        AuthType.QWEN_OAUTH,
+        initialModel,
+        ApprovalMode.YOLO,
+      );
+      const onVisionSwitchRequired = vi.fn();
+      const { result } = renderHook(() =>
+        useVisionAutoSwitch(
+          config,
+          addItem as any,
+          true,
+          onVisionSwitchRequired,
+        ),
+      );
+
+      const parts: PartListUnion = [
+        { inlineData: { mimeType: 'image/png', data: '...' } },
+      ];
+
+      // First, trigger the auto-switch
+      await act(async () => {
+        await result.current.handleVisionSwitch(parts, 10100, false);
+      });
+
+      // Verify model was switched
+      expect(config.setModel).toHaveBeenCalledWith(getDefaultVisionModel(), {
+        reason: 'vision_auto_switch',
+        context: 'YOLO mode auto-switch for image content',
+      });
+
+      // Now restore the original model
+      await act(async () => {
+        await result.current.restoreOriginalModel();
+      });
+
+      // Verify model was restored
+      expect(config.setModel).toHaveBeenLastCalledWith(initialModel, {
+        reason: 'vision_auto_switch',
+        context: 'Restoring original model after vision switch',
+      });
+    });
+
+    it('does not switch in YOLO mode when authType is not QWEN_OAUTH', async () => {
+      const config = createMockConfig(
+        AuthType.USE_GEMINI,
+        'qwen3-coder-plus',
+        ApprovalMode.YOLO,
+      );
+      const onVisionSwitchRequired = vi.fn();
+      const { result } = renderHook(() =>
+        useVisionAutoSwitch(
+          config,
+          addItem as any,
+          true,
+          onVisionSwitchRequired,
+        ),
+      );
+
+      const parts: PartListUnion = [
+        { inlineData: { mimeType: 'image/png', data: '...' } },
+      ];
+
+      let res: any;
+      await act(async () => {
+        res = await result.current.handleVisionSwitch(parts, 11110, false);
+      });
+
+      expect(res).toEqual({ shouldProceed: true });
+      expect(onVisionSwitchRequired).not.toHaveBeenCalled();
+      expect(config.setModel).not.toHaveBeenCalled();
+    });
+
+    it('does not switch in YOLO mode when visionModelPreviewEnabled is false', async () => {
+      const config = createMockConfig(
+        AuthType.QWEN_OAUTH,
+        'qwen3-coder-plus',
+        ApprovalMode.YOLO,
+      );
+      const onVisionSwitchRequired = vi.fn();
+      const { result } = renderHook(() =>
+        useVisionAutoSwitch(
+          config,
+          addItem as any,
+          false,
+          onVisionSwitchRequired,
+        ),
+      );
+
+      const parts: PartListUnion = [
+        { inlineData: { mimeType: 'image/png', data: '...' } },
+      ];
+
+      let res: any;
+      await act(async () => {
+        res = await result.current.handleVisionSwitch(parts, 12120, false);
+      });
+
+      expect(res).toEqual({ shouldProceed: true });
+      expect(onVisionSwitchRequired).not.toHaveBeenCalled();
+      expect(config.setModel).not.toHaveBeenCalled();
+    });
+
+    it('handles multiple image formats in YOLO mode', async () => {
+      const initialModel = 'qwen3-coder-plus';
+      const config = createMockConfig(
+        AuthType.QWEN_OAUTH,
+        initialModel,
+        ApprovalMode.YOLO,
+      );
+      const onVisionSwitchRequired = vi.fn();
+      const { result } = renderHook(() =>
+        useVisionAutoSwitch(
+          config,
+          addItem as any,
+          true,
+          onVisionSwitchRequired,
+        ),
+      );
+
+      const parts: PartListUnion = [
+        { text: 'Here are some images:' },
+        { inlineData: { mimeType: 'image/jpeg', data: '...' } },
+        { fileData: { mimeType: 'image/png', fileUri: 'file://image.png' } },
+        { text: 'Please analyze them.' },
+      ];
+
+      let res: any;
+      await act(async () => {
+        res = await result.current.handleVisionSwitch(parts, 13130, false);
+      });
+
+      expect(res).toEqual({
+        shouldProceed: true,
+        originalModel: initialModel,
+      });
+      expect(config.setModel).toHaveBeenCalledWith(getDefaultVisionModel(), {
+        reason: 'vision_auto_switch',
+        context: 'YOLO mode auto-switch for image content',
+      });
+      expect(onVisionSwitchRequired).not.toHaveBeenCalled();
+    });
+  });
+
+  describe('VLM switch mode default behavior', () => {
+    it('should automatically switch once when vlmSwitchMode is "once"', async () => {
+      const config = createMockConfig(
+        AuthType.QWEN_OAUTH,
+        'qwen3-coder-plus',
+        ApprovalMode.DEFAULT,
+        'once',
+      );
+      const onVisionSwitchRequired = vi.fn(); // Should not be called
+      const { result } = renderHook(() =>
+        useVisionAutoSwitch(
+          config,
+          addItem as any,
+          true,
+          onVisionSwitchRequired,
+        ),
+      );
+
+      const parts: PartListUnion = [
+        { inlineData: { mimeType: 'image/jpeg', data: 'base64data' } },
+      ];
+
+      const switchResult = await result.current.handleVisionSwitch(
+        parts,
+        Date.now(),
+        false,
+      );
+
+      expect(switchResult.shouldProceed).toBe(true);
+      expect(switchResult.originalModel).toBe('qwen3-coder-plus');
+      expect(config.setModel).toHaveBeenCalledWith('vision-model', {
+        reason: 'vision_auto_switch',
+        context: 'Default VLM switch mode: once (one-time override)',
+      });
+      expect(onVisionSwitchRequired).not.toHaveBeenCalled();
+    });
+
+    it('should switch session when vlmSwitchMode is "session"', async () => {
+      const config = createMockConfig(
+        AuthType.QWEN_OAUTH,
+        'qwen3-coder-plus',
+        ApprovalMode.DEFAULT,
+        'session',
+      );
+      const onVisionSwitchRequired = vi.fn(); // Should not be called
+      const { result } = renderHook(() =>
+        useVisionAutoSwitch(
+          config,
+          addItem as any,
+          true,
+          onVisionSwitchRequired,
+        ),
+      );
+
+      const parts: PartListUnion = [
+        { inlineData: { mimeType: 'image/jpeg', data: 'base64data' } },
+      ];
+
+      const switchResult = await result.current.handleVisionSwitch(
+        parts,
+        Date.now(),
+        false,
+      );
+
+      expect(switchResult.shouldProceed).toBe(true);
+      expect(switchResult.originalModel).toBeUndefined(); // No original model for session switch
+      expect(config.setModel).toHaveBeenCalledWith('vision-model', {
+        reason: 'vision_auto_switch',
+        context: 'Default VLM switch mode: session (session persistent)',
+      });
+      expect(onVisionSwitchRequired).not.toHaveBeenCalled();
+    });
+
+    it('should continue with current model when vlmSwitchMode is "persist"', async () => {
+      const config = createMockConfig(
+        AuthType.QWEN_OAUTH,
+        'qwen3-coder-plus',
+        ApprovalMode.DEFAULT,
+        'persist',
+      );
+      const onVisionSwitchRequired = vi.fn(); // Should not be called
+      const { result } = renderHook(() =>
+        useVisionAutoSwitch(
+          config,
+          addItem as any,
+          true,
+          onVisionSwitchRequired,
+        ),
+      );
+
+      const parts: PartListUnion = [
+        { inlineData: { mimeType: 'image/jpeg', data: 'base64data' } },
+      ];
+
+      const switchResult = await result.current.handleVisionSwitch(
+        parts,
+        Date.now(),
+        false,
+      );
+
+      expect(switchResult.shouldProceed).toBe(true);
+      expect(switchResult.originalModel).toBeUndefined();
+      expect(config.setModel).not.toHaveBeenCalled();
+      expect(onVisionSwitchRequired).not.toHaveBeenCalled();
+    });
+
+    it('should fall back to user prompt when vlmSwitchMode is not set', async () => {
+      const config = createMockConfig(
+        AuthType.QWEN_OAUTH,
+        'qwen3-coder-plus',
+        ApprovalMode.DEFAULT,
+        undefined, // No default mode
+      );
+      const onVisionSwitchRequired = vi
+        .fn()
+        .mockResolvedValue({ modelOverride: 'vision-model' });
+      const { result } = renderHook(() =>
+        useVisionAutoSwitch(
+          config,
+          addItem as any,
+          true,
+          onVisionSwitchRequired,
+        ),
+      );
+
+      const parts: PartListUnion = [
+        { inlineData: { mimeType: 'image/jpeg', data: 'base64data' } },
+      ];
+
+      const switchResult = await result.current.handleVisionSwitch(
+        parts,
+        Date.now(),
+        false,
+      );
+
+      expect(switchResult.shouldProceed).toBe(true);
+      expect(onVisionSwitchRequired).toHaveBeenCalledWith(parts);
+    });
+
+    it('should fall back to persist behavior when vlmSwitchMode has invalid value', async () => {
+      const config = createMockConfig(
+        AuthType.QWEN_OAUTH,
+        'qwen3-coder-plus',
+        ApprovalMode.DEFAULT,
+        'invalid-value',
+      );
+      const onVisionSwitchRequired = vi.fn(); // Should not be called
+      const { result } = renderHook(() =>
+        useVisionAutoSwitch(
+          config,
+          addItem as any,
+          true,
+          onVisionSwitchRequired,
+        ),
+      );
+
+      const parts: PartListUnion = [
+        { inlineData: { mimeType: 'image/jpeg', data: 'base64data' } },
+      ];
+
+      const switchResult = await result.current.handleVisionSwitch(
+        parts,
+        Date.now(),
+        false,
+      );
+
+      expect(switchResult.shouldProceed).toBe(true);
+      expect(switchResult.originalModel).toBeUndefined();
+      // For invalid values, it should continue with current model (persist behavior)
+      expect(config.setModel).not.toHaveBeenCalled();
+      expect(onVisionSwitchRequired).not.toHaveBeenCalled();
+    });
+  });
+});

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

@@ -0,0 +1,363 @@
+/**
+ * @license
+ * Copyright 2025 Qwen
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import { type PartListUnion, type Part } from '@google/genai';
+import { AuthType, type Config, ApprovalMode } from '@qwen-code/qwen-code-core';
+import { useCallback, useRef } from 'react';
+import { VisionSwitchOutcome } from '../components/ModelSwitchDialog.js';
+import {
+  getDefaultVisionModel,
+  isVisionModel,
+} from '../models/availableModels.js';
+import { MessageType } from '../types.js';
+import type { UseHistoryManagerReturn } from './useHistoryManager.js';
+import {
+  isSupportedImageMimeType,
+  getUnsupportedImageFormatWarning,
+} from '@qwen-code/qwen-code-core';
+
+/**
+ * Checks if a PartListUnion contains image parts
+ */
+function hasImageParts(parts: PartListUnion): boolean {
+  if (typeof parts === 'string') {
+    return false;
+  }
+
+  if (Array.isArray(parts)) {
+    return parts.some((part) => {
+      // Skip string parts
+      if (typeof part === 'string') return false;
+      return isImagePart(part);
+    });
+  }
+
+  // If it's a single Part (not a string), check if it's an image
+  if (typeof parts === 'object') {
+    return isImagePart(parts);
+  }
+
+  return false;
+}
+
+/**
+ * Checks if a single Part is an image part
+ */
+function isImagePart(part: Part): boolean {
+  // Check for inlineData with image mime type
+  if ('inlineData' in part && part.inlineData?.mimeType?.startsWith('image/')) {
+    return true;
+  }
+
+  // Check for fileData with image mime type
+  if ('fileData' in part && part.fileData?.mimeType?.startsWith('image/')) {
+    return true;
+  }
+
+  return false;
+}
+
+/**
+ * Checks if image parts have supported formats and returns unsupported ones
+ */
+function checkImageFormatsSupport(parts: PartListUnion): {
+  hasImages: boolean;
+  hasUnsupportedFormats: boolean;
+  unsupportedMimeTypes: string[];
+} {
+  const unsupportedMimeTypes: string[] = [];
+  let hasImages = false;
+
+  if (typeof parts === 'string') {
+    return {
+      hasImages: false,
+      hasUnsupportedFormats: false,
+      unsupportedMimeTypes: [],
+    };
+  }
+
+  const partsArray = Array.isArray(parts) ? parts : [parts];
+
+  for (const part of partsArray) {
+    if (typeof part === 'string') continue;
+
+    let mimeType: string | undefined;
+
+    // Check inlineData
+    if (
+      'inlineData' in part &&
+      part.inlineData?.mimeType?.startsWith('image/')
+    ) {
+      hasImages = true;
+      mimeType = part.inlineData.mimeType;
+    }
+
+    // Check fileData
+    if ('fileData' in part && part.fileData?.mimeType?.startsWith('image/')) {
+      hasImages = true;
+      mimeType = part.fileData.mimeType;
+    }
+
+    // Check if the mime type is supported
+    if (mimeType && !isSupportedImageMimeType(mimeType)) {
+      unsupportedMimeTypes.push(mimeType);
+    }
+  }
+
+  return {
+    hasImages,
+    hasUnsupportedFormats: unsupportedMimeTypes.length > 0,
+    unsupportedMimeTypes,
+  };
+}
+
+/**
+ * Determines if we should offer vision switch for the given parts, auth type, and current model
+ */
+export function shouldOfferVisionSwitch(
+  parts: PartListUnion,
+  authType: AuthType,
+  currentModel: string,
+  visionModelPreviewEnabled: boolean = true,
+): boolean {
+  // Only trigger for qwen-oauth
+  if (authType !== AuthType.QWEN_OAUTH) {
+    return false;
+  }
+
+  // If vision model preview is disabled, never offer vision switch
+  if (!visionModelPreviewEnabled) {
+    return false;
+  }
+
+  // If current model is already a vision model, no need to switch
+  if (isVisionModel(currentModel)) {
+    return false;
+  }
+
+  // Check if the current message contains image parts
+  return hasImageParts(parts);
+}
+
+/**
+ * Interface for vision switch result
+ */
+export interface VisionSwitchResult {
+  modelOverride?: string;
+  persistSessionModel?: string;
+  showGuidance?: boolean;
+}
+
+/**
+ * Processes the vision switch outcome and returns the appropriate result
+ */
+export function processVisionSwitchOutcome(
+  outcome: VisionSwitchOutcome,
+): VisionSwitchResult {
+  const vlModelId = getDefaultVisionModel();
+
+  switch (outcome) {
+    case VisionSwitchOutcome.SwitchOnce:
+      return { modelOverride: vlModelId };
+
+    case VisionSwitchOutcome.SwitchSessionToVL:
+      return { persistSessionModel: vlModelId };
+
+    case VisionSwitchOutcome.ContinueWithCurrentModel:
+      return {}; // Continue with current model, no changes needed
+
+    default:
+      return {}; // Default to continuing with current model
+  }
+}
+
+/**
+ * Gets the guidance message for when vision switch is disallowed
+ */
+export function getVisionSwitchGuidanceMessage(): string {
+  const vlModelId = getDefaultVisionModel();
+  return `To use images with your query, you can:
+• Use /model set ${vlModelId} to switch to a vision-capable model
+• Or remove the image and provide a text description instead`;
+}
+
+/**
+ * Interface for vision switch handling result
+ */
+export interface VisionSwitchHandlingResult {
+  shouldProceed: boolean;
+  originalModel?: string;
+}
+
+/**
+ * Custom hook for handling vision model auto-switching
+ */
+export function useVisionAutoSwitch(
+  config: Config,
+  addItem: UseHistoryManagerReturn['addItem'],
+  visionModelPreviewEnabled: boolean = true,
+  onVisionSwitchRequired?: (query: PartListUnion) => Promise<{
+    modelOverride?: string;
+    persistSessionModel?: string;
+    showGuidance?: boolean;
+  }>,
+) {
+  const originalModelRef = useRef<string | null>(null);
+
+  const handleVisionSwitch = useCallback(
+    async (
+      query: PartListUnion,
+      userMessageTimestamp: number,
+      isContinuation: boolean,
+    ): Promise<VisionSwitchHandlingResult> => {
+      // Skip vision switch handling for continuations or if no handler provided
+      if (isContinuation || !onVisionSwitchRequired) {
+        return { shouldProceed: true };
+      }
+
+      const contentGeneratorConfig = config.getContentGeneratorConfig();
+
+      // Only handle qwen-oauth auth type
+      if (contentGeneratorConfig?.authType !== AuthType.QWEN_OAUTH) {
+        return { shouldProceed: true };
+      }
+
+      // Check image format support first
+      const formatCheck = checkImageFormatsSupport(query);
+
+      // If there are unsupported image formats, show warning
+      if (formatCheck.hasUnsupportedFormats) {
+        addItem(
+          {
+            type: MessageType.INFO,
+            text: getUnsupportedImageFormatWarning(),
+          },
+          userMessageTimestamp,
+        );
+        // Continue processing but with warning shown
+      }
+
+      // Check if vision switch is needed
+      if (
+        !shouldOfferVisionSwitch(
+          query,
+          contentGeneratorConfig.authType,
+          config.getModel(),
+          visionModelPreviewEnabled,
+        )
+      ) {
+        return { shouldProceed: true };
+      }
+
+      // In YOLO mode, automatically switch to vision model without user interaction
+      if (config.getApprovalMode() === ApprovalMode.YOLO) {
+        const vlModelId = getDefaultVisionModel();
+        originalModelRef.current = config.getModel();
+        await config.setModel(vlModelId, {
+          reason: 'vision_auto_switch',
+          context: 'YOLO mode auto-switch for image content',
+        });
+        return {
+          shouldProceed: true,
+          originalModel: originalModelRef.current,
+        };
+      }
+
+      // Check if there's a default VLM switch mode configured
+      const defaultVlmSwitchMode = config.getVlmSwitchMode();
+      if (defaultVlmSwitchMode) {
+        // Convert string value to VisionSwitchOutcome enum
+        let outcome: VisionSwitchOutcome;
+        switch (defaultVlmSwitchMode) {
+          case 'once':
+            outcome = VisionSwitchOutcome.SwitchOnce;
+            break;
+          case 'session':
+            outcome = VisionSwitchOutcome.SwitchSessionToVL;
+            break;
+          case 'persist':
+            outcome = VisionSwitchOutcome.ContinueWithCurrentModel;
+            break;
+          default:
+            // Invalid value, fall back to prompting user
+            outcome = VisionSwitchOutcome.ContinueWithCurrentModel;
+        }
+
+        // Process the default outcome
+        const visionSwitchResult = processVisionSwitchOutcome(outcome);
+
+        if (visionSwitchResult.modelOverride) {
+          // One-time model override
+          originalModelRef.current = config.getModel();
+          await config.setModel(visionSwitchResult.modelOverride, {
+            reason: 'vision_auto_switch',
+            context: `Default VLM switch mode: ${defaultVlmSwitchMode} (one-time override)`,
+          });
+          return {
+            shouldProceed: true,
+            originalModel: originalModelRef.current,
+          };
+        } else if (visionSwitchResult.persistSessionModel) {
+          // Persistent session model change
+          await config.setModel(visionSwitchResult.persistSessionModel, {
+            reason: 'vision_auto_switch',
+            context: `Default VLM switch mode: ${defaultVlmSwitchMode} (session persistent)`,
+          });
+          return { shouldProceed: true };
+        }
+
+        // For ContinueWithCurrentModel or any other case, proceed with current model
+        return { shouldProceed: true };
+      }
+
+      try {
+        const visionSwitchResult = await onVisionSwitchRequired(query);
+
+        if (visionSwitchResult.modelOverride) {
+          // One-time model override
+          originalModelRef.current = config.getModel();
+          await config.setModel(visionSwitchResult.modelOverride, {
+            reason: 'vision_auto_switch',
+            context: 'User-prompted vision switch (one-time override)',
+          });
+          return {
+            shouldProceed: true,
+            originalModel: originalModelRef.current,
+          };
+        } else if (visionSwitchResult.persistSessionModel) {
+          // Persistent session model change
+          await config.setModel(visionSwitchResult.persistSessionModel, {
+            reason: 'vision_auto_switch',
+            context: 'User-prompted vision switch (session persistent)',
+          });
+          return { shouldProceed: true };
+        }
+
+        // For ContinueWithCurrentModel or any other case, proceed with current model
+        return { shouldProceed: true };
+      } catch (_error) {
+        // If vision switch dialog was cancelled or errored, don't proceed
+        return { shouldProceed: false };
+      }
+    },
+    [config, addItem, visionModelPreviewEnabled, onVisionSwitchRequired],
+  );
+
+  const restoreOriginalModel = useCallback(async () => {
+    if (originalModelRef.current) {
+      await config.setModel(originalModelRef.current, {
+        reason: 'vision_auto_switch',
+        context: 'Restoring original model after vision switch',
+      });
+      originalModelRef.current = null;
+    }
+  }, [config]);
+
+  return {
+    handleVisionSwitch,
+    restoreOriginalModel,
+  };
+}

packages/cli/src/ui/models/availableModels.ts

@@ -0,0 +1,55 @@
+/**
+ * @license
+ * Copyright 2025 Qwen
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+export type AvailableModel = {
+  id: string;
+  label: string;
+  isVision?: boolean;
+};
+
+export const MAINLINE_VLM = 'vision-model';
+export const MAINLINE_CODER = 'coder-model';
+
+export const AVAILABLE_MODELS_QWEN: AvailableModel[] = [
+  { id: MAINLINE_CODER, label: MAINLINE_CODER },
+  { id: MAINLINE_VLM, label: MAINLINE_VLM, isVision: true },
+];
+
+/**
+ * Get available Qwen models filtered by vision model preview setting
+ */
+export function getFilteredQwenModels(
+  visionModelPreviewEnabled: boolean,
+): AvailableModel[] {
+  if (visionModelPreviewEnabled) {
+    return AVAILABLE_MODELS_QWEN;
+  }
+  return AVAILABLE_MODELS_QWEN.filter((model) => !model.isVision);
+}
+
+/**
+ * Currently we use the single model of `OPENAI_MODEL` in the env.
+ * In the future, after settings.json is updated, we will allow users to configure this themselves.
+ */
+export function getOpenAIAvailableModelFromEnv(): AvailableModel | null {
+  const id = process.env['OPENAI_MODEL']?.trim();
+  return id ? { id, label: id } : null;
+}
+
+/**
+/**
+ * Hard code the default vision model as a string literal,
+ * until our coding model supports multimodal.
+ */
+export function getDefaultVisionModel(): string {
+  return MAINLINE_VLM;
+}
+
+export function isVisionModel(modelId: string): boolean {
+  return AVAILABLE_MODELS_QWEN.some(
+    (model) => model.id === modelId && model.isVision,
+  );
+}

packages/cli/src/ui/utils/MarkdownDisplay.test.tsx

@@ -9,7 +9,6 @@ import { describe, it, expect, vi, beforeEach } from 'vitest';
 import { MarkdownDisplay } from './MarkdownDisplay.js';
 import { LoadedSettings } from '../../config/settings.js';
 import { SettingsContext } from '../contexts/SettingsContext.js';
-import { EOL } from 'node:os';
 
 describe('<MarkdownDisplay />', () => {
   const baseProps = {
@@ -57,7 +56,7 @@ describe('<MarkdownDisplay />', () => {
 ## Header 2
 ### Header 3
 #### Header 4
-`.replace(/\n/g, EOL);
+`;
     const { lastFrame } = render(
       <SettingsContext.Provider value={mockSettings}>
         <MarkdownDisplay {...baseProps} text={text} />
@@ -67,10 +66,7 @@ describe('<MarkdownDisplay />', () => {
   });
 
   it('renders a fenced code block with a language', () => {
-    const text = '```javascript\nconst x = 1;\nconsole.log(x);\n```'.replace(
-      /\n/g,
-      EOL,
-    );
+    const text = '```javascript\nconst x = 1;\nconsole.log(x);\n```';
     const { lastFrame } = render(
       <SettingsContext.Provider value={mockSettings}>
         <MarkdownDisplay {...baseProps} text={text} />
@@ -80,7 +76,7 @@ describe('<MarkdownDisplay />', () => {
   });
 
   it('renders a fenced code block without a language', () => {
-    const text = '```\nplain text\n```'.replace(/\n/g, EOL);
+    const text = '```\nplain text\n```';
     const { lastFrame } = render(
       <SettingsContext.Provider value={mockSettings}>
         <MarkdownDisplay {...baseProps} text={text} />
@@ -90,7 +86,7 @@ describe('<MarkdownDisplay />', () => {
   });
 
   it('handles unclosed (pending) code blocks', () => {
-    const text = '```typescript\nlet y = 2;'.replace(/\n/g, EOL);
+    const text = '```typescript\nlet y = 2;';
     const { lastFrame } = render(
       <SettingsContext.Provider value={mockSettings}>
         <MarkdownDisplay {...baseProps} text={text} isPending={true} />
@@ -104,7 +100,7 @@ describe('<MarkdownDisplay />', () => {
 - item A
 * item B
 + item C
-`.replace(/\n/g, EOL);
+`;
     const { lastFrame } = render(
       <SettingsContext.Provider value={mockSettings}>
         <MarkdownDisplay {...baseProps} text={text} />
@@ -118,7 +114,7 @@ describe('<MarkdownDisplay />', () => {
 * Level 1
   * Level 2
     * Level 3
-`.replace(/\n/g, EOL);
+`;
     const { lastFrame } = render(
       <SettingsContext.Provider value={mockSettings}>
         <MarkdownDisplay {...baseProps} text={text} />
@@ -131,7 +127,7 @@ describe('<MarkdownDisplay />', () => {
     const text = `
 1. First item
 2. Second item
-`.replace(/\n/g, EOL);
+`;
     const { lastFrame } = render(
       <SettingsContext.Provider value={mockSettings}>
         <MarkdownDisplay {...baseProps} text={text} />
@@ -147,7 +143,7 @@ Hello
 World
 ***
 Test
-`.replace(/\n/g, EOL);
+`;
     const { lastFrame } = render(
       <SettingsContext.Provider value={mockSettings}>
         <MarkdownDisplay {...baseProps} text={text} />
@@ -162,7 +158,7 @@ Test
 |----------|:--------:|
 | Cell 1   | Cell 2   |
 | Cell 3   | Cell 4   |
-`.replace(/\n/g, EOL);
+`;
     const { lastFrame } = render(
       <SettingsContext.Provider value={mockSettings}>
         <MarkdownDisplay {...baseProps} text={text} />
@@ -176,7 +172,7 @@ Test
 Some text before.
 | A | B |
 |---|
-| 1 | 2 |`.replace(/\n/g, EOL);
+| 1 | 2 |`;
     const { lastFrame } = render(
       <SettingsContext.Provider value={mockSettings}>
         <MarkdownDisplay {...baseProps} text={text} />
@@ -188,7 +184,7 @@ Some text before.
   it('inserts a single space between paragraphs', () => {
     const text = `Paragraph 1.
 
-Paragraph 2.`.replace(/\n/g, EOL);
+Paragraph 2.`;
     const { lastFrame } = render(
       <SettingsContext.Provider value={mockSettings}>
         <MarkdownDisplay {...baseProps} text={text} />
@@ -211,7 +207,7 @@ some code
 \`\`\`
 
 Another paragraph.
-`.replace(/\n/g, EOL);
+`;
     const { lastFrame } = render(
       <SettingsContext.Provider value={mockSettings}>
         <MarkdownDisplay {...baseProps} text={text} />
@@ -221,7 +217,7 @@ Another paragraph.
   });
 
   it('hides line numbers in code blocks when showLineNumbers is false', () => {
-    const text = '```javascript\nconst x = 1;\n```'.replace(/\n/g, EOL);
+    const text = '```javascript\nconst x = 1;\n```';
     const settings = new LoadedSettings(
       { path: '', settings: {} },
       { path: '', settings: {} },
@@ -242,7 +238,7 @@ Another paragraph.
   });
 
   it('shows line numbers in code blocks by default', () => {
-    const text = '```javascript\nconst x = 1;\n```'.replace(/\n/g, EOL);
+    const text = '```javascript\nconst x = 1;\n```';
     const { lastFrame } = render(
       <SettingsContext.Provider value={mockSettings}>
         <MarkdownDisplay {...baseProps} text={text} />
@@ -251,4 +247,21 @@ Another paragraph.
     expect(lastFrame()).toMatchSnapshot();
     expect(lastFrame()).toContain(' 1 ');
   });
+
+  it('correctly splits lines using \\n regardless of platform EOL', () => {
+    // Test that the component uses \n for splitting, not EOL
+    const textWithUnixLineEndings = 'Line 1\nLine 2\nLine 3';
+
+    const { lastFrame } = render(
+      <SettingsContext.Provider value={mockSettings}>
+        <MarkdownDisplay {...baseProps} text={textWithUnixLineEndings} />
+      </SettingsContext.Provider>,
+    );
+
+    const output = lastFrame();
+    expect(output).toContain('Line 1');
+    expect(output).toContain('Line 2');
+    expect(output).toContain('Line 3');
+    expect(output).toMatchSnapshot();
+  });
 });

packages/cli/src/ui/utils/MarkdownDisplay.tsx

@@ -6,7 +6,6 @@
 
 import React from 'react';
 import { Text, Box } from 'ink';
-import { EOL } from 'node:os';
 import { Colors } from '../colors.js';
 import { colorizeCode } from './CodeColorizer.js';
 import { TableRenderer } from './TableRenderer.js';
@@ -35,7 +34,7 @@ const MarkdownDisplayInternal: React.FC<MarkdownDisplayProps> = ({
 }) => {
   if (!text) return <></>;
 
-  const lines = text.split(EOL);
+  const lines = text.split(`\n`);
   const headerRegex = /^ *(#{1,4}) +(.*)/;
   const codeFenceRegex = /^ *(`{3,}|~{3,}) *(\w*?) *$/;
   const ulItemRegex = /^([ \t]*)([-*+]) +(.*)/;

packages/cli/src/ui/utils/__snapshots__/MarkdownDisplay.test.tsx.snap

@@ -14,6 +14,12 @@ Another paragraph.
 "
 `;
 
+exports[`<MarkdownDisplay /> > correctly splits lines using \\n regardless of platform EOL 1`] = `
+"Line 1
+Line 2
+Line 3"
+`;
+
 exports[`<MarkdownDisplay /> > handles a table at the end of the input 1`] = `
 "Some text before.
 | A | B |

packages/cli/src/validateNonInterActiveAuth.test.ts

@@ -126,6 +126,18 @@ describe('validateNonInterActiveAuth', () => {
     expect(refreshAuthMock).toHaveBeenCalledWith(AuthType.USE_OPENAI);
   });
 
+  it('uses configured QWEN_OAUTH if provided', async () => {
+    const nonInteractiveConfig: NonInteractiveConfig = {
+      refreshAuth: refreshAuthMock,
+    };
+    await validateNonInteractiveAuth(
+      AuthType.QWEN_OAUTH,
+      undefined,
+      nonInteractiveConfig,
+    );
+    expect(refreshAuthMock).toHaveBeenCalledWith(AuthType.QWEN_OAUTH);
+  });
+
   it('uses USE_VERTEX_AI if GOOGLE_GENAI_USE_VERTEXAI is true (with GOOGLE_CLOUD_PROJECT and GOOGLE_CLOUD_LOCATION)', async () => {
     process.env['GOOGLE_GENAI_USE_VERTEXAI'] = 'true';
     process.env['GOOGLE_CLOUD_PROJECT'] = 'test-project';

packages/cli/src/zed-integration/zedIntegration.ts

@@ -97,6 +97,18 @@ class GeminiAgent {
         name: 'Vertex AI',
         description: null,
       },
+      {
+        id: AuthType.USE_OPENAI,
+        name: 'Use OpenAI API key',
+        description:
+          'Requires setting the `OPENAI_API_KEY` environment variable',
+      },
+      {
+        id: AuthType.QWEN_OAUTH,
+        name: 'Qwen OAuth',
+        description:
+          'OAuth authentication for Qwen models with 2000 daily requests',
+      },
     ];
 
     return {
@@ -871,6 +883,16 @@ function toToolCallContent(toolResult: ToolResult): acp.ToolCallContent | null {
         type: 'content',
         content: { type: 'text', text: todoText },
       };
+    } else if (
+      'type' in toolResult.returnDisplay &&
+      toolResult.returnDisplay.type === 'plan_summary'
+    ) {
+      const planDisplay = toolResult.returnDisplay;
+      const planText = `${planDisplay.message}\n\n${planDisplay.plan}`;
+      return {
+        type: 'content',
+        content: { type: 'text', text: planText },
+      };
     } else if ('fileDiff' in toolResult.returnDisplay) {
       // Handle FileDiff
       return {
@@ -942,6 +964,15 @@ function toPermissionOptions(
         },
         ...basicPermissionOptions,
       ];
+    case 'plan':
+      return [
+        {
+          optionId: ToolConfirmationOutcome.ProceedAlways,
+          name: `Always Allow Plans`,
+          kind: 'allow_always',
+        },
+        ...basicPermissionOptions,
+      ];
     default: {
       const unreachable: never = confirmation;
       throw new Error(`Unexpected: ${unreachable}`);

packages/core/index.ts

@@ -19,3 +19,4 @@ export {
 } from './src/telemetry/types.js';
 export { makeFakeConfig } from './src/test-utils/config.js';
 export * from './src/utils/pathReader.js';
+export * from './src/utils/request-tokenizer/supportedImageFormats.js';

packages/core/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@qwen-code/qwen-code-core",
-  "version": "0.0.11",
+  "version": "0.0.14",
   "description": "Qwen Code Core",
   "repository": {
     "type": "git",

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

@@ -710,6 +710,18 @@ describe('setApprovalMode with folder trust', () => {
     expect(() => config.setApprovalMode(ApprovalMode.DEFAULT)).not.toThrow();
   });
 
+  it('should NOT throw an error when setting PLAN mode in an untrusted folder', () => {
+    const config = new Config({
+      sessionId: 'test',
+      targetDir: '.',
+      debugMode: false,
+      model: 'test-model',
+      cwd: '.',
+      trustedFolder: false, // Untrusted
+    });
+    expect(() => config.setApprovalMode(ApprovalMode.PLAN)).not.toThrow();
+  });
+
   it('should NOT throw an error when setting any mode in a trusted folder', () => {
     const config = new Config({
       sessionId: 'test',
@@ -722,6 +734,7 @@ describe('setApprovalMode with folder trust', () => {
     expect(() => config.setApprovalMode(ApprovalMode.YOLO)).not.toThrow();
     expect(() => config.setApprovalMode(ApprovalMode.AUTO_EDIT)).not.toThrow();
     expect(() => config.setApprovalMode(ApprovalMode.DEFAULT)).not.toThrow();
+    expect(() => config.setApprovalMode(ApprovalMode.PLAN)).not.toThrow();
   });
 
   it('should NOT throw an error when setting any mode if trustedFolder is undefined', () => {
@@ -736,5 +749,87 @@ describe('setApprovalMode with folder trust', () => {
     expect(() => config.setApprovalMode(ApprovalMode.YOLO)).not.toThrow();
     expect(() => config.setApprovalMode(ApprovalMode.AUTO_EDIT)).not.toThrow();
     expect(() => config.setApprovalMode(ApprovalMode.DEFAULT)).not.toThrow();
+    expect(() => config.setApprovalMode(ApprovalMode.PLAN)).not.toThrow();
+  });
+
+  describe('Model Switch Logging', () => {
+    it('should log model switch when setModel is called with different model', async () => {
+      const config = new Config({
+        sessionId: 'test-model-switch',
+        targetDir: '.',
+        debugMode: false,
+        model: 'qwen3-coder-plus',
+        cwd: '.',
+      });
+
+      // Initialize the config to set up content generator
+      await config.initialize();
+
+      // Mock the logger's logModelSwitch method
+      const logModelSwitchSpy = vi.spyOn(config['logger']!, 'logModelSwitch');
+
+      // Change the model
+      await config.setModel('qwen-vl-max-latest', {
+        reason: 'vision_auto_switch',
+        context: 'Test model switch',
+      });
+
+      // Verify that logModelSwitch was called with correct parameters
+      expect(logModelSwitchSpy).toHaveBeenCalledWith({
+        fromModel: 'qwen3-coder-plus',
+        toModel: 'qwen-vl-max-latest',
+        reason: 'vision_auto_switch',
+        context: 'Test model switch',
+      });
+    });
+
+    it('should not log when setModel is called with same model', async () => {
+      const config = new Config({
+        sessionId: 'test-same-model',
+        targetDir: '.',
+        debugMode: false,
+        model: 'qwen3-coder-plus',
+        cwd: '.',
+      });
+
+      // Initialize the config to set up content generator
+      await config.initialize();
+
+      // Mock the logger's logModelSwitch method
+      const logModelSwitchSpy = vi.spyOn(config['logger']!, 'logModelSwitch');
+
+      // Set the same model
+      await config.setModel('qwen3-coder-plus');
+
+      // Verify that logModelSwitch was not called
+      expect(logModelSwitchSpy).not.toHaveBeenCalled();
+    });
+
+    it('should use default reason when no options provided', async () => {
+      const config = new Config({
+        sessionId: 'test-default-reason',
+        targetDir: '.',
+        debugMode: false,
+        model: 'qwen3-coder-plus',
+        cwd: '.',
+      });
+
+      // Initialize the config to set up content generator
+      await config.initialize();
+
+      // Mock the logger's logModelSwitch method
+      const logModelSwitchSpy = vi.spyOn(config['logger']!, 'logModelSwitch');
+
+      // Change the model without options
+      await config.setModel('qwen-vl-max-latest');
+
+      // Verify that logModelSwitch was called with default reason
+      expect(logModelSwitchSpy).toHaveBeenCalledWith({
+        fromModel: 'qwen3-coder-plus',
+        toModel: 'qwen-vl-max-latest',
+        reason: 'manual',
+        context: undefined,
+      });
+    });
   });
 });

packages/core/src/config/config.ts

@@ -33,6 +33,7 @@ import {
 import { logCliConfiguration, logIdeConnection } from '../telemetry/loggers.js';
 import { IdeConnectionEvent, IdeConnectionType } from '../telemetry/types.js';
 import { EditTool } from '../tools/edit.js';
+import { ExitPlanModeTool } from '../tools/exitPlanMode.js';
 import { GlobTool } from '../tools/glob.js';
 import { GrepTool } from '../tools/grep.js';
 import { LSTool } from '../tools/ls.js';
@@ -56,16 +57,20 @@ import {
   DEFAULT_GEMINI_FLASH_MODEL,
 } from './models.js';
 import { Storage } from './storage.js';
+import { Logger, type ModelSwitchEvent } from '../core/logger.js';
 
 // Re-export OAuth config type
 export type { AnyToolInvocation, MCPOAuthConfig };
 
 export enum ApprovalMode {
+  PLAN = 'plan',
   DEFAULT = 'default',
-  AUTO_EDIT = 'autoEdit',
+  AUTO_EDIT = 'auto-edit',
   YOLO = 'yolo',
 }
 
+export const APPROVAL_MODES = Object.values(ApprovalMode);
+
 export interface AccessibilitySettings {
   disableLoadingPhrases?: boolean;
   screenReader?: boolean;
@@ -239,6 +244,7 @@ export interface ConfigParameters {
   extensionManagement?: boolean;
   enablePromptCompletion?: boolean;
   skipLoopDetection?: boolean;
+  vlmSwitchMode?: string;
 }
 
 export class Config {
@@ -330,9 +336,11 @@ export class Config {
   private readonly extensionManagement: boolean;
   private readonly enablePromptCompletion: boolean = false;
   private readonly skipLoopDetection: boolean;
+  private readonly vlmSwitchMode: string | undefined;
   private initialized: boolean = false;
   readonly storage: Storage;
   private readonly fileExclusions: FileExclusions;
+  private logger: Logger | null = null;
 
   constructor(params: ConfigParameters) {
     this.sessionId = params.sessionId;
@@ -424,8 +432,15 @@ export class Config {
     this.extensionManagement = params.extensionManagement ?? false;
     this.storage = new Storage(this.targetDir);
     this.enablePromptCompletion = params.enablePromptCompletion ?? false;
+    this.vlmSwitchMode = params.vlmSwitchMode;
     this.fileExclusions = new FileExclusions(this);
 
+    // Initialize logger asynchronously
+    this.logger = new Logger(this.sessionId, this.storage);
+    this.logger.initialize().catch((error) => {
+      console.debug('Failed to initialize logger:', error);
+    });
+
     if (params.contextFileName) {
       setGeminiMdFilename(params.contextFileName);
     }
@@ -517,10 +532,48 @@ export class Config {
     return this.contentGeneratorConfig?.model || this.model;
   }
 
-  setModel(newModel: string): void {
+  async setModel(
+    newModel: string,
+    options?: {
+      reason?: ModelSwitchEvent['reason'];
+      context?: string;
+    },
+  ): Promise<void> {
+    const oldModel = this.getModel();
+
     if (this.contentGeneratorConfig) {
       this.contentGeneratorConfig.model = newModel;
     }
+
+    // Log the model switch if the model actually changed
+    if (oldModel !== newModel && this.logger) {
+      const switchEvent: ModelSwitchEvent = {
+        fromModel: oldModel,
+        toModel: newModel,
+        reason: options?.reason || 'manual',
+        context: options?.context,
+      };
+
+      // Log asynchronously to avoid blocking
+      this.logger.logModelSwitch(switchEvent).catch((error) => {
+        console.debug('Failed to log model switch:', error);
+      });
+    }
+
+    // Reinitialize chat with updated configuration while preserving history
+    const geminiClient = this.getGeminiClient();
+    if (geminiClient && geminiClient.isInitialized()) {
+      // Now await the reinitialize operation to ensure completion
+      try {
+        await geminiClient.reinitialize();
+      } catch (error) {
+        console.error(
+          'Failed to reinitialize chat with updated config:',
+          error,
+        );
+        throw error; // Re-throw to let callers handle the error
+      }
+    }
   }
 
   isInFallbackMode(): boolean {
@@ -651,7 +704,11 @@ export class Config {
   }
 
   setApprovalMode(mode: ApprovalMode): void {
-    if (this.isTrustedFolder() === false && mode !== ApprovalMode.DEFAULT) {
+    if (
+      this.isTrustedFolder() === false &&
+      mode !== ApprovalMode.DEFAULT &&
+      mode !== ApprovalMode.PLAN
+    ) {
       throw new Error(
         'Cannot enable privileged approval modes in an untrusted folder.',
       );
@@ -926,6 +983,10 @@ export class Config {
     return this.skipLoopDetection;
   }
 
+  getVlmSwitchMode(): string | undefined {
+    return this.vlmSwitchMode;
+  }
+
   async getGitService(): Promise<GitService> {
     if (!this.gitService) {
       this.gitService = new GitService(this.targetDir, this.storage);
@@ -990,11 +1051,12 @@ export class Config {
     registerCoreTool(GlobTool, this);
     registerCoreTool(EditTool, this);
     registerCoreTool(WriteFileTool, this);
-    registerCoreTool(WebFetchTool, this);
     registerCoreTool(ReadManyFilesTool, this);
     registerCoreTool(ShellTool, this);
     registerCoreTool(MemoryTool);
     registerCoreTool(TodoWriteTool, this);
+    registerCoreTool(ExitPlanModeTool, this);
+    registerCoreTool(WebFetchTool, this);
     // Conditionally register web search tool only if Tavily API key is set
     if (this.getTavilyApiKey()) {
       registerCoreTool(WebSearchTool, this);

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

@@ -41,7 +41,7 @@ describe('Flash Model Fallback Configuration', () => {
   // with the fallback mechanism. This will be necessary we introduce more
   // intelligent model routing.
   describe('setModel', () => {
-    it('should only mark as switched if contentGeneratorConfig exists', () => {
+    it('should only mark as switched if contentGeneratorConfig exists', async () => {
       // Create config without initializing contentGeneratorConfig
       const newConfig = new Config({
         sessionId: 'test-session-2',
@@ -52,15 +52,15 @@ describe('Flash Model Fallback Configuration', () => {
       });
 
       // Should not crash when contentGeneratorConfig is undefined
-      newConfig.setModel(DEFAULT_GEMINI_FLASH_MODEL);
+      await newConfig.setModel(DEFAULT_GEMINI_FLASH_MODEL);
       expect(newConfig.isInFallbackMode()).toBe(false);
     });
   });
 
   describe('getModel', () => {
-    it('should return contentGeneratorConfig model if available', () => {
+    it('should return contentGeneratorConfig model if available', async () => {
       // Simulate initialized content generator config
-      config.setModel(DEFAULT_GEMINI_FLASH_MODEL);
+      await config.setModel(DEFAULT_GEMINI_FLASH_MODEL);
       expect(config.getModel()).toBe(DEFAULT_GEMINI_FLASH_MODEL);
     });
 
@@ -88,8 +88,8 @@ describe('Flash Model Fallback Configuration', () => {
       expect(config.isInFallbackMode()).toBe(false);
     });
 
-    it('should persist switched state throughout session', () => {
-      config.setModel(DEFAULT_GEMINI_FLASH_MODEL);
+    it('should persist switched state throughout session', async () => {
+      await config.setModel(DEFAULT_GEMINI_FLASH_MODEL);
       // Setting state for fallback mode as is expected of clients
       config.setFallbackMode(true);
       expect(config.isInFallbackMode()).toBe(true);

packages/core/src/config/models.ts

@@ -4,11 +4,10 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-export const DEFAULT_QWEN_MODEL = 'qwen3-coder-plus';
-// We do not have a fallback model for now, but note it here anyway.
-export const DEFAULT_QWEN_FLASH_MODEL = 'qwen3-coder-flash';
+export const DEFAULT_QWEN_MODEL = 'coder-model';
+export const DEFAULT_QWEN_FLASH_MODEL = 'coder-model';
 
-export const DEFAULT_GEMINI_MODEL = 'qwen3-coder-plus';
+export const DEFAULT_GEMINI_MODEL = 'coder-model';
 export const DEFAULT_GEMINI_FLASH_MODEL = 'gemini-2.5-flash';
 export const DEFAULT_GEMINI_FLASH_LITE_MODEL = 'gemini-2.5-flash-lite';
 

packages/core/src/core/__tests__/openaiTimeoutHandling.test.ts

@@ -5,9 +5,10 @@
  */
 
 import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
-import { OpenAIContentGenerator } from '../openaiContentGenerator.js';
+import { OpenAIContentGenerator } from '../openaiContentGenerator/openaiContentGenerator.js';
 import type { Config } from '../../config/config.js';
 import { AuthType } from '../contentGenerator.js';
+import type { OpenAICompatibleProvider } from '../openaiContentGenerator/provider/index.js';
 import OpenAI from 'openai';
 
 // Mock OpenAI
@@ -30,6 +31,7 @@ describe('OpenAIContentGenerator Timeout Handling', () => {
   let mockConfig: Config;
   // eslint-disable-next-line @typescript-eslint/no-explicit-any
   let mockOpenAIClient: any;
+  let mockProvider: OpenAICompatibleProvider;
 
   beforeEach(() => {
     // Reset mocks
@@ -42,6 +44,7 @@ describe('OpenAIContentGenerator Timeout Handling', () => {
     mockConfig = {
       getContentGeneratorConfig: vi.fn().mockReturnValue({
         authType: 'openai',
+        enableOpenAILogging: false,
       }),
       getCliVersion: vi.fn().mockReturnValue('1.0.0'),
     } as unknown as Config;
@@ -53,17 +56,34 @@ describe('OpenAIContentGenerator Timeout Handling', () => {
           create: vi.fn(),
         },
       },
+      embeddings: {
+        create: vi.fn(),
+      },
     };
 
     vi.mocked(OpenAI).mockImplementation(() => mockOpenAIClient);
 
+    // Create mock provider
+    mockProvider = {
+      buildHeaders: vi.fn().mockReturnValue({
+        'User-Agent': 'QwenCode/1.0.0 (test; test)',
+      }),
+      buildClient: vi.fn().mockReturnValue(mockOpenAIClient),
+      buildRequest: vi.fn().mockImplementation((req) => req),
+    };
+
     // Create generator instance
     const contentGeneratorConfig = {
       model: 'gpt-4',
       apiKey: 'test-key',
       authType: AuthType.USE_OPENAI,
+      enableOpenAILogging: false,
     };
-    generator = new OpenAIContentGenerator(contentGeneratorConfig, mockConfig);
+    generator = new OpenAIContentGenerator(
+      contentGeneratorConfig,
+      mockConfig,
+      mockProvider,
+    );
   });
 
   afterEach(() => {
@@ -209,7 +229,7 @@ describe('OpenAIContentGenerator Timeout Handling', () => {
       await expect(
         generator.generateContentStream(request, 'test-prompt-id'),
       ).rejects.toThrow(
-        /Streaming setup timeout after \d+s\. Try reducing input length or increasing timeout in config\./,
+        /Streaming request timeout after \d+s\. Try reducing input length or increasing timeout in config\./,
       );
     });
 
@@ -227,12 +247,8 @@ describe('OpenAIContentGenerator Timeout Handling', () => {
       } catch (error: unknown) {
         const errorMessage =
           error instanceof Error ? error.message : String(error);
-        expect(errorMessage).toContain(
-          'Streaming setup timeout troubleshooting:',
-        );
-        expect(errorMessage).toContain(
-          'Check network connectivity and firewall settings',
-        );
+        expect(errorMessage).toContain('Streaming timeout troubleshooting:');
+        expect(errorMessage).toContain('Check network connectivity');
         expect(errorMessage).toContain('Consider using non-streaming mode');
       }
     });
@@ -246,23 +262,21 @@ describe('OpenAIContentGenerator Timeout Handling', () => {
         authType: AuthType.USE_OPENAI,
         baseUrl: 'http://localhost:8080',
       };
-      new OpenAIContentGenerator(contentGeneratorConfig, mockConfig);
+      new OpenAIContentGenerator(
+        contentGeneratorConfig,
+        mockConfig,
+        mockProvider,
+      );
 
-      // Verify OpenAI client was created with timeout config
-      expect(OpenAI).toHaveBeenCalledWith({
-        apiKey: 'test-key',
-        baseURL: 'http://localhost:8080',
-        timeout: 120000,
-        maxRetries: 3,
-        defaultHeaders: {
-          'User-Agent': expect.stringMatching(/^QwenCode/),
-        },
-      });
+      // Verify provider buildClient was called
+      expect(mockProvider.buildClient).toHaveBeenCalled();
     });
 
     it('should use custom timeout from config', () => {
       const customConfig = {
-        getContentGeneratorConfig: vi.fn().mockReturnValue({}),
+        getContentGeneratorConfig: vi.fn().mockReturnValue({
+          enableOpenAILogging: false,
+        }),
         getCliVersion: vi.fn().mockReturnValue('1.0.0'),
       } as unknown as Config;
 
@@ -274,22 +288,31 @@ describe('OpenAIContentGenerator Timeout Handling', () => {
         timeout: 300000,
         maxRetries: 5,
       };
-      new OpenAIContentGenerator(contentGeneratorConfig, customConfig);
 
-      expect(OpenAI).toHaveBeenCalledWith({
-        apiKey: 'test-key',
-        baseURL: 'http://localhost:8080',
-        timeout: 300000,
-        maxRetries: 5,
-        defaultHeaders: {
-          'User-Agent': expect.stringMatching(/^QwenCode/),
-        },
-      });
+      // Create a custom mock provider for this test
+      const customMockProvider: OpenAICompatibleProvider = {
+        buildHeaders: vi.fn().mockReturnValue({
+          'User-Agent': 'QwenCode/1.0.0 (test; test)',
+        }),
+        buildClient: vi.fn().mockReturnValue(mockOpenAIClient),
+        buildRequest: vi.fn().mockImplementation((req) => req),
+      };
+
+      new OpenAIContentGenerator(
+        contentGeneratorConfig,
+        customConfig,
+        customMockProvider,
+      );
+
+      // Verify provider buildClient was called
+      expect(customMockProvider.buildClient).toHaveBeenCalled();
     });
 
     it('should handle missing timeout config gracefully', () => {
       const noTimeoutConfig = {
-        getContentGeneratorConfig: vi.fn().mockReturnValue({}),
+        getContentGeneratorConfig: vi.fn().mockReturnValue({
+          enableOpenAILogging: false,
+        }),
         getCliVersion: vi.fn().mockReturnValue('1.0.0'),
       } as unknown as Config;
 
@@ -299,17 +322,24 @@ describe('OpenAIContentGenerator Timeout Handling', () => {
         authType: AuthType.USE_OPENAI,
         baseUrl: 'http://localhost:8080',
       };
-      new OpenAIContentGenerator(contentGeneratorConfig, noTimeoutConfig);
 
-      expect(OpenAI).toHaveBeenCalledWith({
-        apiKey: 'test-key',
-        baseURL: 'http://localhost:8080',
-        timeout: 120000, // default
-        maxRetries: 3, // default
-        defaultHeaders: {
-          'User-Agent': expect.stringMatching(/^QwenCode/),
-        },
-      });
+      // Create a custom mock provider for this test
+      const noTimeoutMockProvider: OpenAICompatibleProvider = {
+        buildHeaders: vi.fn().mockReturnValue({
+          'User-Agent': 'QwenCode/1.0.0 (test; test)',
+        }),
+        buildClient: vi.fn().mockReturnValue(mockOpenAIClient),
+        buildRequest: vi.fn().mockImplementation((req) => req),
+      };
+
+      new OpenAIContentGenerator(
+        contentGeneratorConfig,
+        noTimeoutConfig,
+        noTimeoutMockProvider,
+      );
+
+      // Verify provider buildClient was called
+      expect(noTimeoutMockProvider.buildClient).toHaveBeenCalled();
     });
   });
 

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

@@ -23,6 +23,7 @@ import type {
 } from '@google/genai';
 import { GoogleGenAI } from '@google/genai';
 import { findIndexAfterFraction, GeminiClient } from './client.js';
+import { getPlanModeSystemReminder } from './prompts.js';
 import {
   AuthType,
   type ContentGenerator,
@@ -50,6 +51,10 @@ const mockGenerateContentFn = vi.fn();
 const mockEmbedContentFn = vi.fn();
 const mockTurnRunFn = vi.fn();
 
+let ApprovalModeEnum: typeof import('../config/config.js').ApprovalMode;
+// eslint-disable-next-line @typescript-eslint/no-explicit-any
+let mockConfigObject: any;
+
 vi.mock('@google/genai');
 vi.mock('./turn', async (importOriginal) => {
   const actual = await importOriginal<typeof import('./turn.js')>();
@@ -178,6 +183,12 @@ describe('Gemini Client (client.ts)', () => {
   beforeEach(async () => {
     vi.resetAllMocks();
 
+    ApprovalModeEnum = (
+      await vi.importActual<typeof import('../config/config.js')>(
+        '../config/config.js',
+      )
+    ).ApprovalMode;
+
     // Disable 429 simulation for tests
     setSimulate429(false);
 
@@ -226,7 +237,11 @@ describe('Gemini Client (client.ts)', () => {
       vertexai: false,
       authType: AuthType.USE_GEMINI,
     };
-    const mockConfigObject = {
+    const mockSubagentManager = {
+      listSubagents: vi.fn().mockResolvedValue([]),
+      addChangeListener: vi.fn().mockReturnValue(() => {}),
+    };
+    mockConfigObject = {
       getContentGeneratorConfig: vi
         .fn()
         .mockReturnValue(contentGeneratorConfig),
@@ -249,6 +264,7 @@ describe('Gemini Client (client.ts)', () => {
       getNoBrowser: vi.fn().mockReturnValue(false),
       getSystemPromptMappings: vi.fn().mockReturnValue(undefined),
       getUsageStatisticsEnabled: vi.fn().mockReturnValue(true),
+      getApprovalMode: vi.fn().mockReturnValue(ApprovalModeEnum.DEFAULT),
       getIdeModeFeature: vi.fn().mockReturnValue(false),
       getIdeMode: vi.fn().mockReturnValue(true),
       getDebugMode: vi.fn().mockReturnValue(false),
@@ -260,6 +276,7 @@ describe('Gemini Client (client.ts)', () => {
       getCliVersion: vi.fn().mockReturnValue('1.0.0'),
       getChatCompression: vi.fn().mockReturnValue(undefined),
       getSkipNextSpeakerCheck: vi.fn().mockReturnValue(false),
+      getSubagentManager: vi.fn().mockReturnValue(mockSubagentManager),
       getSkipLoopDetection: vi.fn().mockReturnValue(false),
     };
     const MockedConfig = vi.mocked(Config, true);
@@ -417,8 +434,6 @@ describe('Gemini Client (client.ts)', () => {
           config: {
             abortSignal,
             systemInstruction: getCoreSystemPrompt(''),
-            temperature: 0,
-            topP: 1,
             tools: [
               {
                 functionDeclarations: [
@@ -437,7 +452,8 @@ describe('Gemini Client (client.ts)', () => {
       );
     });
 
-    it('should allow overriding model and config', async () => {
+    /* We now use model in contentGeneratorConfig in most cases. */
+    it.skip('should allow overriding model and config', async () => {
       const contents: Content[] = [
         { role: 'user', parts: [{ text: 'hello' }] },
       ];
@@ -468,7 +484,6 @@ describe('Gemini Client (client.ts)', () => {
             abortSignal,
             systemInstruction: getCoreSystemPrompt(''),
             temperature: 0.9,
-            topP: 1, // from default
             topK: 20,
             tools: [
               {
@@ -943,6 +958,42 @@ describe('Gemini Client (client.ts)', () => {
   });
 
   describe('sendMessageStream', () => {
+    it('injects a plan mode reminder before user queries when approval mode is PLAN', async () => {
+      const mockStream = (async function* () {})();
+      mockTurnRunFn.mockReturnValue(mockStream);
+
+      mockConfigObject.getApprovalMode.mockReturnValue(ApprovalModeEnum.PLAN);
+
+      const mockChat: Partial<GeminiChat> = {
+        addHistory: vi.fn(),
+        getHistory: vi.fn().mockReturnValue([]),
+      };
+      client['chat'] = mockChat as GeminiChat;
+
+      const mockGenerator: Partial<ContentGenerator> = {
+        countTokens: vi.fn().mockResolvedValue({ totalTokens: 0 }),
+        generateContent: mockGenerateContentFn,
+      };
+      client['contentGenerator'] = mockGenerator as ContentGenerator;
+
+      const stream = client.sendMessageStream(
+        'Plan mode test',
+        new AbortController().signal,
+        'prompt-plan-1',
+      );
+
+      await fromAsync(stream);
+
+      expect(mockTurnRunFn).toHaveBeenCalledWith(
+        [getPlanModeSystemReminder(), 'Plan mode test'],
+        expect.any(Object),
+      );
+
+      mockConfigObject.getApprovalMode.mockReturnValue(
+        ApprovalModeEnum.DEFAULT,
+      );
+    });
+
     it('emits a compression event when the context was automatically compressed', async () => {
       // Arrange
       const mockStream = (async function* () {
@@ -1171,10 +1222,7 @@ ${JSON.stringify(
 
       // Assert
       expect(ideContext.getIdeContext).toHaveBeenCalled();
-      expect(mockTurnRunFn).toHaveBeenCalledWith(
-        initialRequest,
-        expect.any(Object),
-      );
+      expect(mockTurnRunFn).toHaveBeenCalledWith(['Hi'], expect.any(Object));
     });
 
     it('should add context if ideMode is enabled and there is one active file', async () => {
@@ -2410,7 +2458,6 @@ ${JSON.stringify(
             abortSignal,
             systemInstruction: getCoreSystemPrompt(''),
             temperature: 0.5,
-            topP: 1,
           },
           contents,
         },
@@ -2545,4 +2592,82 @@ ${JSON.stringify(
       expect(mockChat.setHistory).toHaveBeenCalledWith(historyWithThoughts);
     });
   });
+
+  describe('initialize', () => {
+    it('should accept extraHistory parameter and pass it to startChat', async () => {
+      const mockStartChat = vi.fn().mockResolvedValue({});
+      client['startChat'] = mockStartChat;
+
+      const extraHistory = [
+        { role: 'user', parts: [{ text: 'Previous message' }] },
+        { role: 'model', parts: [{ text: 'Previous response' }] },
+      ];
+
+      const contentGeneratorConfig = {
+        model: 'test-model',
+        apiKey: 'test-key',
+        vertexai: false,
+        authType: AuthType.USE_GEMINI,
+      };
+
+      await client.initialize(contentGeneratorConfig, extraHistory);
+
+      expect(mockStartChat).toHaveBeenCalledWith(extraHistory, 'test-model');
+    });
+
+    it('should use empty array when no extraHistory is provided', async () => {
+      const mockStartChat = vi.fn().mockResolvedValue({});
+      client['startChat'] = mockStartChat;
+
+      const contentGeneratorConfig = {
+        model: 'test-model',
+        apiKey: 'test-key',
+        vertexai: false,
+        authType: AuthType.USE_GEMINI,
+      };
+
+      await client.initialize(contentGeneratorConfig);
+
+      expect(mockStartChat).toHaveBeenCalledWith([], 'test-model');
+    });
+  });
+
+  describe('reinitialize', () => {
+    it('should reinitialize with preserved user history', async () => {
+      // Mock the initialize method
+      const mockInitialize = vi.fn().mockResolvedValue(undefined);
+      client['initialize'] = mockInitialize;
+
+      // Set up initial history with environment context + user messages
+      const mockHistory = [
+        { role: 'user', parts: [{ text: 'Environment context' }] },
+        { role: 'model', parts: [{ text: 'Got it. Thanks for the context!' }] },
+        { role: 'user', parts: [{ text: 'User message 1' }] },
+        { role: 'model', parts: [{ text: 'Model response 1' }] },
+      ];
+
+      const mockChat = {
+        getHistory: vi.fn().mockReturnValue(mockHistory),
+      };
+      client['chat'] = mockChat as unknown as GeminiChat;
+      client['getHistory'] = vi.fn().mockReturnValue(mockHistory);
+
+      await client.reinitialize();
+
+      // Should call initialize with preserved user history (excluding first 2 env messages)
+      expect(mockInitialize).toHaveBeenCalledWith(
+        expect.any(Object), // contentGeneratorConfig
+        [
+          { role: 'user', parts: [{ text: 'User message 1' }] },
+          { role: 'model', parts: [{ text: 'Model response 1' }] },
+        ],
+      );
+    });
+
+    it('should not throw error when chat is not initialized', async () => {
+      client['chat'] = undefined;
+
+      await expect(client.reinitialize()).resolves.not.toThrow();
+    });
+  });
 });

packages/core/src/core/client.ts

@@ -17,6 +17,7 @@ import type {
 import { ProxyAgent, setGlobalDispatcher } from 'undici';
 import type { UserTierId } from '../code_assist/types.js';
 import type { Config } from '../config/config.js';
+import { ApprovalMode } from '../config/config.js';
 import { DEFAULT_GEMINI_FLASH_MODEL } from '../config/models.js';
 import type { File, IdeContext } from '../ide/ideContext.js';
 import { ideContext } from '../ide/ideContext.js';
@@ -29,6 +30,7 @@ import {
   makeChatCompressionEvent,
   NextSpeakerCheckEvent,
 } from '../telemetry/types.js';
+import { TaskTool } from '../tools/task.js';
 import {
   getDirectoryContextString,
   getEnvironmentContext,
@@ -39,6 +41,7 @@ import { getFunctionCalls } from '../utils/generateContentResponseUtilities.js';
 import { isFunctionResponse } from '../utils/messageInspectors.js';
 import { checkNextSpeaker } from '../utils/nextSpeakerChecker.js';
 import { retryWithBackoff } from '../utils/retry.js';
+import { flatMapTextParts } from '../utils/partUtils.js';
 import type {
   ContentGenerator,
   ContentGeneratorConfig,
@@ -49,6 +52,8 @@ import {
   getCompressionPrompt,
   getCoreSystemPrompt,
   getCustomSystemPrompt,
+  getPlanModeSystemReminder,
+  getSubagentSystemReminder,
 } from './prompts.js';
 import { tokenLimit } from './tokenLimits.js';
 import type { ChatCompressionInfo, ServerGeminiStreamEvent } from './turn.js';
@@ -110,10 +115,7 @@ export class GeminiClient {
   private chat?: GeminiChat;
   private contentGenerator?: ContentGenerator;
   private readonly embeddingModel: string;
-  private readonly generateContentConfig: GenerateContentConfig = {
-    temperature: 0,
-    topP: 1,
-  };
+  private readonly generateContentConfig: GenerateContentConfig = {};
   private sessionTurnCount = 0;
 
   private readonly loopDetector: LoopDetectionService;
@@ -137,13 +139,24 @@ export class GeminiClient {
     this.lastPromptId = this.config.getSessionId();
   }
 
-  async initialize(contentGeneratorConfig: ContentGeneratorConfig) {
+  async initialize(
+    contentGeneratorConfig: ContentGeneratorConfig,
+    extraHistory?: Content[],
+  ) {
     this.contentGenerator = await createContentGenerator(
       contentGeneratorConfig,
       this.config,
       this.config.getSessionId(),
     );
-    this.chat = await this.startChat();
+    /**
+     * Always take the model from contentGeneratorConfig to initialize,
+     * despite the `this.config.contentGeneratorConfig` is not updated yet because in
+     * `Config` it will not be updated until the initialization is successful.
+     */
+    this.chat = await this.startChat(
+      extraHistory || [],
+      contentGeneratorConfig.model,
+    );
   }
 
   getContentGenerator(): ContentGenerator {
@@ -216,6 +229,28 @@ export class GeminiClient {
     this.chat = await this.startChat();
   }
 
+  /**
+   * Reinitializes the chat with the current contentGeneratorConfig while preserving chat history.
+   * This creates a new chat object using the existing history and updated configuration.
+   * Should be called when configuration changes (model, auth, etc.) to ensure consistency.
+   */
+  async reinitialize(): Promise<void> {
+    if (!this.chat) {
+      return;
+    }
+
+    // Preserve the current chat history (excluding environment context)
+    const currentHistory = this.getHistory();
+    // Remove the initial environment context (first 2 messages: user env + model acknowledgment)
+    const userHistory = currentHistory.slice(2);
+
+    // Get current content generator config and reinitialize with preserved history
+    const contentGeneratorConfig = this.config.getContentGeneratorConfig();
+    if (contentGeneratorConfig) {
+      await this.initialize(contentGeneratorConfig, userHistory);
+    }
+  }
+
   async addDirectoryContext(): Promise<void> {
     if (!this.chat) {
       return;
@@ -227,7 +262,10 @@ export class GeminiClient {
     });
   }
 
-  async startChat(extraHistory?: Content[]): Promise<GeminiChat> {
+  async startChat(
+    extraHistory?: Content[],
+    model?: string,
+  ): Promise<GeminiChat> {
     this.forceFullIdeContext = true;
     this.hasFailedCompressionAttempt = false;
     const envParts = await getEnvironmentContext(this.config);
@@ -247,9 +285,13 @@ export class GeminiClient {
     ];
     try {
       const userMemory = this.config.getUserMemory();
-      const systemInstruction = getCoreSystemPrompt(userMemory);
+      const systemInstruction = getCoreSystemPrompt(
+        userMemory,
+        {},
+        model || this.config.getModel(),
+      );
       const generateContentConfigWithThinking = isThinkingSupported(
-        this.config.getModel(),
+        model || this.config.getModel(),
       )
         ? {
             ...this.generateContentConfig,
@@ -455,7 +497,8 @@ export class GeminiClient {
     turns: number = MAX_TURNS,
     originalModel?: string,
   ): AsyncGenerator<ServerGeminiStreamEvent, Turn> {
-    if (this.lastPromptId !== prompt_id) {
+    const isNewPrompt = this.lastPromptId !== prompt_id;
+    if (isNewPrompt) {
       this.loopDetector.reset(prompt_id);
       this.lastPromptId = prompt_id;
     }
@@ -488,7 +531,11 @@ export class GeminiClient {
       // Get all the content that would be sent in an API call
       const currentHistory = this.getChat().getHistory(true);
       const userMemory = this.config.getUserMemory();
-      const systemPrompt = getCoreSystemPrompt(userMemory);
+      const systemPrompt = getCoreSystemPrompt(
+        userMemory,
+        {},
+        this.config.getModel(),
+      );
       const environment = await getEnvironmentContext(this.config);
 
       // Create a mock request content to count total tokens
@@ -562,7 +609,30 @@ export class GeminiClient {
       }
     }
 
-    const resultStream = turn.run(request, signal);
+    // append system reminders to the request
+    let requestToSent = await flatMapTextParts(request, async (text) => [text]);
+    if (isNewPrompt) {
+      const systemReminders = [];
+
+      // add subagent system reminder if there are subagents
+      const hasTaskTool = this.config.getToolRegistry().getTool(TaskTool.Name);
+      const subagents = (await this.config.getSubagentManager().listSubagents())
+        .filter((subagent) => subagent.level !== 'builtin')
+        .map((subagent) => subagent.name);
+
+      if (hasTaskTool && subagents.length > 0) {
+        systemReminders.push(getSubagentSystemReminder(subagents));
+      }
+
+      // add plan mode system reminder if approval mode is plan
+      if (this.config.getApprovalMode() === ApprovalMode.PLAN) {
+        systemReminders.push(getPlanModeSystemReminder());
+      }
+
+      requestToSent = [...systemReminders, ...requestToSent];
+    }
+
+    const resultStream = turn.run(requestToSent, signal);
     for await (const event of resultStream) {
       if (!this.config.getSkipLoopDetection()) {
         if (this.loopDetector.addAndCheck(event)) {
@@ -624,14 +694,18 @@ export class GeminiClient {
     model?: string,
     config: GenerateContentConfig = {},
   ): Promise<Record<string, unknown>> {
-    // Use current model from config instead of hardcoded Flash model
-    const modelToUse =
-      model || this.config.getModel() || DEFAULT_GEMINI_FLASH_MODEL;
+    /**
+     * TODO: ensure `model` consistency among GeminiClient, GeminiChat, and ContentGenerator
+     * `model` passed to generateContent is not respected as we always use contentGenerator
+     * We should ignore model for now because some calls use `DEFAULT_GEMINI_FLASH_MODEL`
+     * which is not available as `qwen3-coder-flash`
+     */
+    const modelToUse = this.config.getModel() || DEFAULT_GEMINI_FLASH_MODEL;
     try {
       const userMemory = this.config.getUserMemory();
       const finalSystemInstruction = config.systemInstruction
         ? getCustomSystemPrompt(config.systemInstruction, userMemory)
-        : getCoreSystemPrompt(userMemory);
+        : getCoreSystemPrompt(userMemory, {}, modelToUse);
 
       const requestConfig = {
         abortSignal,
@@ -722,7 +796,7 @@ export class GeminiClient {
       const userMemory = this.config.getUserMemory();
       const finalSystemInstruction = generationConfig.systemInstruction
         ? getCustomSystemPrompt(generationConfig.systemInstruction, userMemory)
-        : getCoreSystemPrompt(userMemory);
+        : getCoreSystemPrompt(userMemory, {}, this.config.getModel());
 
       const requestConfig: GenerateContentConfig = {
         abortSignal,
@@ -985,7 +1059,7 @@ export class GeminiClient {
           error,
         );
         if (accepted !== false && accepted !== null) {
-          this.config.setModel(fallbackModel);
+          await this.config.setModel(fallbackModel);
           this.config.setFallbackMode(true);
           return fallbackModel;
         }

packages/core/src/core/coreToolScheduler.test.ts

@@ -10,11 +10,13 @@ import { describe, expect, it, vi } from 'vitest';
 import type {
   Config,
   ToolCallConfirmationDetails,
+  ToolCallRequestInfo,
   ToolConfirmationPayload,
   ToolInvocation,
   ToolResult,
   ToolResultDisplay,
   ToolRegistry,
+  SuccessfulToolCall,
 } from '../index.js';
 import {
   ApprovalMode,
@@ -24,11 +26,16 @@ import {
   ToolConfirmationOutcome,
 } from '../index.js';
 import { MockModifiableTool, MockTool } from '../test-utils/tools.js';
-import type { ToolCall, WaitingToolCall } from './coreToolScheduler.js';
+import type {
+  ToolCall,
+  WaitingToolCall,
+  ErroredToolCall,
+} from './coreToolScheduler.js';
 import {
   CoreToolScheduler,
   convertToFunctionResponse,
 } from './coreToolScheduler.js';
+import { getPlanModeSystemReminder } from './prompts.js';
 
 class TestApprovalTool extends BaseDeclarativeTool<{ id: string }, ToolResult> {
   static readonly Name = 'testApprovalTool';
@@ -101,6 +108,49 @@ class TestApprovalInvocation extends BaseToolInvocation<
   }
 }
 
+class SimpleToolInvocation extends BaseToolInvocation<
+  Record<string, unknown>,
+  ToolResult
+> {
+  constructor(
+    params: Record<string, unknown>,
+    private readonly executeImpl: () => Promise<ToolResult> | ToolResult,
+  ) {
+    super(params);
+  }
+
+  getDescription(): string {
+    return 'simple tool invocation';
+  }
+
+  async execute(): Promise<ToolResult> {
+    return await Promise.resolve(this.executeImpl());
+  }
+}
+
+class SimpleTool extends BaseDeclarativeTool<
+  Record<string, unknown>,
+  ToolResult
+> {
+  constructor(
+    name: string,
+    kind: Kind,
+    private readonly executeImpl: () => Promise<ToolResult> | ToolResult,
+  ) {
+    super(name, name, 'Simple test tool', kind, {
+      type: 'object',
+      properties: {},
+      additionalProperties: true,
+    });
+  }
+
+  protected createInvocation(
+    params: Record<string, unknown>,
+  ): ToolInvocation<Record<string, unknown>, ToolResult> {
+    return new SimpleToolInvocation(params, this.executeImpl);
+  }
+}
+
 async function waitForStatus(
   onToolCallsUpdate: Mock,
   status: 'awaiting_approval' | 'executing' | 'success' | 'error' | 'cancelled',
@@ -197,6 +247,249 @@ describe('CoreToolScheduler', () => {
     expect(completedCalls[0].status).toBe('cancelled');
   });
 
+  describe('plan mode enforcement', () => {
+    it('returns plan reminder and skips execution for edit tools', async () => {
+      const executeSpy = vi.fn().mockResolvedValue({
+        llmContent: 'should not execute',
+        returnDisplay: 'should not execute',
+      });
+      // Use MockTool with shouldConfirm=true to simulate a tool that requires confirmation
+      const tool = new MockTool('write_file');
+      tool.shouldConfirm = true;
+      tool.executeFn = executeSpy;
+
+      const mockToolRegistry = {
+        getTool: vi.fn().mockReturnValue(tool),
+        getAllToolNames: vi.fn().mockReturnValue([tool.name]),
+      } as unknown as ToolRegistry;
+
+      const onAllToolCallsComplete = vi.fn();
+      const onToolCallsUpdate = vi.fn();
+
+      const mockConfig = {
+        getSessionId: () => 'plan-session',
+        getUsageStatisticsEnabled: () => true,
+        getDebugMode: () => false,
+        getApprovalMode: vi.fn().mockReturnValue(ApprovalMode.PLAN),
+        getAllowedTools: () => [],
+        getContentGeneratorConfig: () => ({
+          model: 'test-model',
+          authType: 'oauth-personal',
+        }),
+        getToolRegistry: () => mockToolRegistry,
+      } as unknown as Config;
+
+      const scheduler = new CoreToolScheduler({
+        config: mockConfig,
+        onAllToolCallsComplete,
+        onToolCallsUpdate,
+        getPreferredEditor: () => 'vscode',
+        onEditorClose: vi.fn(),
+      });
+
+      const request: ToolCallRequestInfo = {
+        callId: 'plan-1',
+        name: 'write_file',
+        args: {},
+        isClientInitiated: false,
+        prompt_id: 'prompt-plan',
+      };
+
+      await scheduler.schedule([request], new AbortController().signal);
+
+      const errorCall = (await waitForStatus(
+        onToolCallsUpdate,
+        'error',
+      )) as ErroredToolCall;
+
+      expect(executeSpy).not.toHaveBeenCalled();
+      expect(
+        errorCall.response.responseParts[0]?.functionResponse?.response?.[
+          'output'
+        ],
+      ).toBe(getPlanModeSystemReminder());
+      expect(errorCall.response.resultDisplay).toContain('Plan mode');
+    });
+
+    it('allows read tools to execute in plan mode', async () => {
+      const executeSpy = vi.fn().mockResolvedValue({
+        llmContent: 'read ok',
+        returnDisplay: 'read ok',
+      });
+      const tool = new SimpleTool('read_file', Kind.Read, executeSpy);
+
+      const mockToolRegistry = {
+        getTool: vi.fn().mockReturnValue(tool),
+        getAllToolNames: vi.fn().mockReturnValue([tool.name]),
+      } as unknown as ToolRegistry;
+
+      const onAllToolCallsComplete = vi.fn();
+      const onToolCallsUpdate = vi.fn();
+
+      const mockConfig = {
+        getSessionId: () => 'plan-session',
+        getUsageStatisticsEnabled: () => true,
+        getDebugMode: () => false,
+        getApprovalMode: vi.fn().mockReturnValue(ApprovalMode.PLAN),
+        getAllowedTools: () => [],
+        getContentGeneratorConfig: () => ({
+          model: 'test-model',
+          authType: 'oauth-personal',
+        }),
+        getToolRegistry: () => mockToolRegistry,
+      } as unknown as Config;
+
+      const scheduler = new CoreToolScheduler({
+        config: mockConfig,
+        onAllToolCallsComplete,
+        onToolCallsUpdate,
+        getPreferredEditor: () => 'vscode',
+        onEditorClose: vi.fn(),
+      });
+
+      const request: ToolCallRequestInfo = {
+        callId: 'plan-2',
+        name: tool.name,
+        args: {},
+        isClientInitiated: false,
+        prompt_id: 'prompt-plan',
+      };
+
+      await scheduler.schedule([request], new AbortController().signal);
+
+      const successCall = (await waitForStatus(
+        onToolCallsUpdate,
+        'success',
+      )) as SuccessfulToolCall;
+
+      expect(executeSpy).toHaveBeenCalledTimes(1);
+      expect(
+        successCall.response.responseParts[0]?.functionResponse?.response?.[
+          'output'
+        ],
+      ).toBe('read ok');
+    });
+
+    it('enforces shell command restrictions in plan mode', async () => {
+      const executeSpyAllowed = vi.fn().mockResolvedValue({
+        llmContent: 'shell ok',
+        returnDisplay: 'shell ok',
+      });
+      const allowedTool = new SimpleTool(
+        'run_shell_command',
+        Kind.Execute,
+        executeSpyAllowed,
+      );
+
+      const allowedToolRegistry = {
+        getTool: vi.fn().mockReturnValue(allowedTool),
+        getAllToolNames: vi.fn().mockReturnValue([allowedTool.name]),
+      } as unknown as ToolRegistry;
+
+      const allowedConfig = {
+        getSessionId: () => 'plan-session',
+        getUsageStatisticsEnabled: () => true,
+        getDebugMode: () => false,
+        getApprovalMode: vi.fn().mockReturnValue(ApprovalMode.PLAN),
+        getAllowedTools: () => [],
+        getContentGeneratorConfig: () => ({
+          model: 'test-model',
+          authType: 'oauth-personal',
+        }),
+        getToolRegistry: () => allowedToolRegistry,
+      } as unknown as Config;
+
+      const allowedUpdates = vi.fn();
+      const allowedScheduler = new CoreToolScheduler({
+        config: allowedConfig,
+        onAllToolCallsComplete: vi.fn(),
+        onToolCallsUpdate: allowedUpdates,
+        getPreferredEditor: () => 'vscode',
+        onEditorClose: vi.fn(),
+      });
+
+      const allowedRequest: ToolCallRequestInfo = {
+        callId: 'plan-shell-allowed',
+        name: allowedTool.name,
+        args: { command: 'ls -la' },
+        isClientInitiated: false,
+        prompt_id: 'prompt-plan',
+      };
+
+      await allowedScheduler.schedule(
+        [allowedRequest],
+        new AbortController().signal,
+      );
+
+      await waitForStatus(allowedUpdates, 'success');
+      expect(executeSpyAllowed).toHaveBeenCalledTimes(1);
+
+      const executeSpyBlocked = vi.fn().mockResolvedValue({
+        llmContent: 'blocked',
+        returnDisplay: 'blocked',
+      });
+      // Use MockTool with shouldConfirm=true to simulate a shell tool that requires confirmation
+      const blockedTool = new MockTool('run_shell_command');
+      blockedTool.shouldConfirm = true;
+      blockedTool.executeFn = executeSpyBlocked;
+
+      const blockedToolRegistry = {
+        getTool: vi.fn().mockReturnValue(blockedTool),
+        getAllToolNames: vi.fn().mockReturnValue([blockedTool.name]),
+      } as unknown as ToolRegistry;
+
+      const blockedConfig = {
+        getSessionId: () => 'plan-session',
+        getUsageStatisticsEnabled: () => true,
+        getDebugMode: () => false,
+        getApprovalMode: vi.fn().mockReturnValue(ApprovalMode.PLAN),
+        getAllowedTools: () => [],
+        getContentGeneratorConfig: () => ({
+          model: 'test-model',
+          authType: 'oauth-personal',
+        }),
+        getToolRegistry: () => blockedToolRegistry,
+      } as unknown as Config;
+
+      const blockedUpdates = vi.fn();
+      const blockedScheduler = new CoreToolScheduler({
+        config: blockedConfig,
+        onAllToolCallsComplete: vi.fn(),
+        onToolCallsUpdate: blockedUpdates,
+        getPreferredEditor: () => 'vscode',
+        onEditorClose: vi.fn(),
+      });
+
+      const blockedRequest: ToolCallRequestInfo = {
+        callId: 'plan-shell-blocked',
+        name: 'run_shell_command',
+        args: { command: 'rm -rf tmp' },
+        isClientInitiated: false,
+        prompt_id: 'prompt-plan',
+      };
+
+      await blockedScheduler.schedule(
+        [blockedRequest],
+        new AbortController().signal,
+      );
+
+      const blockedCall = (await waitForStatus(
+        blockedUpdates,
+        'error',
+      )) as ErroredToolCall;
+      expect(executeSpyBlocked).not.toHaveBeenCalled();
+      expect(
+        blockedCall.response.responseParts[0]?.functionResponse?.response?.[
+          'output'
+        ],
+      ).toBe(getPlanModeSystemReminder());
+      const observedStatuses = blockedUpdates.mock.calls
+        .flatMap((call) => call[0] as ToolCall[])
+        .map((tc) => tc.status);
+      expect(observedStatuses).not.toContain('awaiting_approval');
+    });
+  });
+
   describe('getToolSuggestion', () => {
     it('should suggest the top N closest tool names for a typo', () => {
       // Create mocked tool registry

packages/core/src/core/coreToolScheduler.ts

@@ -34,6 +34,7 @@ import {
 import * as Diff from 'diff';
 import { doesToolInvocationMatch } from '../utils/tool-utils.js';
 import levenshtein from 'fast-levenshtein';
+import { getPlanModeSystemReminder } from './prompts.js';
 
 export type ValidatingToolCall = {
   status: 'validating';
@@ -674,7 +675,27 @@ export class CoreToolScheduler {
           }
 
           const allowedTools = this.config.getAllowedTools() || [];
-          if (
+          const isPlanMode =
+            this.config.getApprovalMode() === ApprovalMode.PLAN;
+          const isExitPlanModeTool = reqInfo.name === 'exit_plan_mode';
+
+          if (isPlanMode && !isExitPlanModeTool) {
+            if (confirmationDetails) {
+              this.setStatusInternal(reqInfo.callId, 'error', {
+                callId: reqInfo.callId,
+                responseParts: convertToFunctionResponse(
+                  reqInfo.name,
+                  reqInfo.callId,
+                  getPlanModeSystemReminder(),
+                ),
+                resultDisplay: 'Plan mode blocked a non-read-only tool call.',
+                error: undefined,
+                errorType: undefined,
+              });
+            } else {
+              this.setStatusInternal(reqInfo.callId, 'scheduled');
+            }
+          } else if (
             this.config.getApprovalMode() === ApprovalMode.YOLO ||
             doesToolInvocationMatch(toolCall.tool, invocation, allowedTools)
           ) {

packages/core/src/core/geminiChat.ts

@@ -224,7 +224,7 @@ export class GeminiChat {
           error,
         );
         if (accepted !== false && accepted !== null) {
-          this.config.setModel(fallbackModel);
+          await this.config.setModel(fallbackModel);
           this.config.setFallbackMode(true);
           return fallbackModel;
         }
@@ -500,7 +500,7 @@ export class GeminiChat {
         if (error instanceof Error && error.message) {
           if (isSchemaDepthError(error.message)) return false;
           if (error.message.includes('429')) return true;
-          if (error.message.match(/5\d{2}/)) return true;
+          if (error.message.match(/^5\d{2}/)) return true;
         }
         return false;
       },

packages/core/src/core/logger.test.ts

@@ -755,4 +755,84 @@ describe('Logger', () => {
       expect(logger['messageId']).toBe(0);
     });
   });
+
+  describe('Model Switch Logging', () => {
+    it('should log model switch events correctly', async () => {
+      const testSessionId = 'test-session-model-switch';
+      const logger = new Logger(testSessionId, new Storage(process.cwd()));
+      await logger.initialize();
+
+      const modelSwitchEvent = {
+        fromModel: 'qwen3-coder-plus',
+        toModel: 'qwen-vl-max-latest',
+        reason: 'vision_auto_switch' as const,
+        context: 'YOLO mode auto-switch for image content',
+      };
+
+      await logger.logModelSwitch(modelSwitchEvent);
+
+      // Read the log file to verify the entry was written
+      const logContent = await fs.readFile(TEST_LOG_FILE_PATH, 'utf-8');
+      const logs: LogEntry[] = JSON.parse(logContent);
+
+      const modelSwitchLog = logs.find(
+        (log) =>
+          log.sessionId === testSessionId &&
+          log.type === MessageSenderType.MODEL_SWITCH,
+      );
+
+      expect(modelSwitchLog).toBeDefined();
+      expect(modelSwitchLog!.type).toBe(MessageSenderType.MODEL_SWITCH);
+
+      const loggedEvent = JSON.parse(modelSwitchLog!.message);
+      expect(loggedEvent.fromModel).toBe('qwen3-coder-plus');
+      expect(loggedEvent.toModel).toBe('qwen-vl-max-latest');
+      expect(loggedEvent.reason).toBe('vision_auto_switch');
+      expect(loggedEvent.context).toBe(
+        'YOLO mode auto-switch for image content',
+      );
+    });
+
+    it('should handle multiple model switch events', async () => {
+      const testSessionId = 'test-session-multiple-switches';
+      const logger = new Logger(testSessionId, new Storage(process.cwd()));
+      await logger.initialize();
+
+      // Log first switch
+      await logger.logModelSwitch({
+        fromModel: 'qwen3-coder-plus',
+        toModel: 'qwen-vl-max-latest',
+        reason: 'vision_auto_switch',
+        context: 'Auto-switch for image',
+      });
+
+      // Log second switch (restore)
+      await logger.logModelSwitch({
+        fromModel: 'qwen-vl-max-latest',
+        toModel: 'qwen3-coder-plus',
+        reason: 'vision_auto_switch',
+        context: 'Restoring original model',
+      });
+
+      // Read the log file to verify both entries were written
+      const logContent = await fs.readFile(TEST_LOG_FILE_PATH, 'utf-8');
+      const logs: LogEntry[] = JSON.parse(logContent);
+
+      const modelSwitchLogs = logs.filter(
+        (log) =>
+          log.sessionId === testSessionId &&
+          log.type === MessageSenderType.MODEL_SWITCH,
+      );
+
+      expect(modelSwitchLogs).toHaveLength(2);
+
+      const firstSwitch = JSON.parse(modelSwitchLogs[0].message);
+      expect(firstSwitch.fromModel).toBe('qwen3-coder-plus');
+      expect(firstSwitch.toModel).toBe('qwen-vl-max-latest');
+
+      const secondSwitch = JSON.parse(modelSwitchLogs[1].message);
+      expect(secondSwitch.fromModel).toBe('qwen-vl-max-latest');
+      expect(secondSwitch.toModel).toBe('qwen3-coder-plus');
+    });
+  });
 });

packages/core/src/core/logger.ts

@@ -13,6 +13,7 @@ const LOG_FILE_NAME = 'logs.json';
 
 export enum MessageSenderType {
   USER = 'user',
+  MODEL_SWITCH = 'model_switch',
 }
 
 export interface LogEntry {
@@ -23,6 +24,13 @@ export interface LogEntry {
   message: string;
 }
 
+export interface ModelSwitchEvent {
+  fromModel: string;
+  toModel: string;
+  reason: 'vision_auto_switch' | 'manual' | 'fallback' | 'other';
+  context?: string;
+}
+
 // This regex matches any character that is NOT a letter (a-z, A-Z),
 // a number (0-9), a hyphen (-), an underscore (_), or a dot (.).
 
@@ -270,6 +278,17 @@ export class Logger {
     }
   }
 
+  async logModelSwitch(event: ModelSwitchEvent): Promise<void> {
+    const message = JSON.stringify({
+      fromModel: event.fromModel,
+      toModel: event.toModel,
+      reason: event.reason,
+      context: event.context,
+    });
+
+    await this.logMessage(MessageSenderType.MODEL_SWITCH, message);
+  }
+
   private _checkpointPath(tag: string): string {
     if (!tag.length) {
       throw new Error('No checkpoint tag specified.');

packages/core/src/core/openaiContentGenerator/converter.ts

@@ -376,28 +376,22 @@ export class OpenAIContentConverter {
     parsedParts: Pick<ParsedParts, 'textParts' | 'mediaParts'>,
   ): OpenAI.Chat.ChatCompletionMessageParam | null {
     const { textParts, mediaParts } = parsedParts;
-    const combinedText = textParts.join('');
+    const content = textParts.map((text) => ({ type: 'text' as const, text }));
 
     // If no media parts, return simple text message
     if (mediaParts.length === 0) {
-      return combinedText ? { role, content: combinedText } : null;
+      return content.length > 0 ? { role, content } : null;
     }
 
     // For assistant messages with media, convert to text only
     // since OpenAI assistant messages don't support media content arrays
     if (role === 'assistant') {
-      return combinedText
-        ? { role: 'assistant' as const, content: combinedText }
+      return content.length > 0
+        ? { role: 'assistant' as const, content }
         : null;
     }
 
-    // Create multimodal content array for user messages
-    const contentArray: OpenAI.Chat.ChatCompletionContentPart[] = [];
-
-    // Add text content
-    if (combinedText) {
-      contentArray.push({ type: 'text', text: combinedText });
-    }
+    const contentArray: OpenAI.Chat.ChatCompletionContentPart[] = [...content];
 
     // Add media content
     for (const mediaPart of mediaParts) {
@@ -405,14 +399,14 @@ export class OpenAIContentConverter {
         if (mediaPart.fileUri) {
           // For file URIs, use the URI directly
           contentArray.push({
-            type: 'image_url',
+            type: 'image_url' as const,
             image_url: { url: mediaPart.fileUri },
           });
         } else if (mediaPart.data) {
           // For inline data, create data URL
           const dataUrl = `data:${mediaPart.mimeType};base64,${mediaPart.data}`;
           contentArray.push({
-            type: 'image_url',
+            type: 'image_url' as const,
             image_url: { url: dataUrl },
           });
         }
@@ -421,7 +415,7 @@ export class OpenAIContentConverter {
         const format = this.getAudioFormat(mediaPart.mimeType);
         if (format) {
           contentArray.push({
-            type: 'input_audio',
+            type: 'input_audio' as const,
             input_audio: {
               data: mediaPart.data,
               format: format as 'wav' | 'mp3',

packages/core/src/core/openaiContentGenerator/index.ts

@@ -12,6 +12,7 @@ import type { Config } from '../../config/config.js';
 import { OpenAIContentGenerator } from './openaiContentGenerator.js';
 import {
   DashScopeOpenAICompatibleProvider,
+  DeepSeekOpenAICompatibleProvider,
   OpenRouterOpenAICompatibleProvider,
   type OpenAICompatibleProvider,
   DefaultOpenAICompatibleProvider,
@@ -23,6 +24,7 @@ export { ContentGenerationPipeline, type PipelineConfig } from './pipeline.js';
 export {
   type OpenAICompatibleProvider,
   DashScopeOpenAICompatibleProvider,
+  DeepSeekOpenAICompatibleProvider,
   OpenRouterOpenAICompatibleProvider,
 } from './provider/index.js';
 
@@ -61,6 +63,13 @@ export function determineProvider(
     );
   }
 
+  if (DeepSeekOpenAICompatibleProvider.isDeepSeekProvider(config)) {
+    return new DeepSeekOpenAICompatibleProvider(
+      contentGeneratorConfig,
+      cliConfig,
+    );
+  }
+
   // Check for OpenRouter provider
   if (OpenRouterOpenAICompatibleProvider.isOpenRouterProvider(config)) {
     return new OpenRouterOpenAICompatibleProvider(

packages/core/src/core/openaiContentGenerator/openaiContentGenerator.test.ts

@@ -5,6 +5,37 @@
  */
 
 import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
+
+// Mock the request tokenizer module BEFORE importing the class that uses it
+const mockTokenizer = {
+  calculateTokens: vi.fn().mockResolvedValue({
+    totalTokens: 50,
+    breakdown: {
+      textTokens: 50,
+      imageTokens: 0,
+      audioTokens: 0,
+      otherTokens: 0,
+    },
+    processingTime: 1,
+  }),
+  dispose: vi.fn(),
+};
+
+vi.mock('../../../utils/request-tokenizer/index.js', () => ({
+  getDefaultTokenizer: vi.fn(() => mockTokenizer),
+  DefaultRequestTokenizer: vi.fn(() => mockTokenizer),
+  disposeDefaultTokenizer: vi.fn(),
+}));
+
+// Mock tiktoken as well for completeness
+vi.mock('tiktoken', () => ({
+  get_encoding: vi.fn(() => ({
+    encode: vi.fn(() => new Array(50)), // Mock 50 tokens
+    free: vi.fn(),
+  })),
+}));
+
+// Now import the modules that depend on the mocked modules
 import { OpenAIContentGenerator } from './openaiContentGenerator.js';
 import type { Config } from '../../config/config.js';
 import { AuthType } from '../contentGenerator.js';
@@ -15,14 +46,6 @@ import type {
 import type { OpenAICompatibleProvider } from './provider/index.js';
 import type OpenAI from 'openai';
 
-// Mock tiktoken
-vi.mock('tiktoken', () => ({
-  get_encoding: vi.fn().mockReturnValue({
-    encode: vi.fn().mockReturnValue(new Array(50)), // Mock 50 tokens
-    free: vi.fn(),
-  }),
-}));
-
 describe('OpenAIContentGenerator (Refactored)', () => {
   let generator: OpenAIContentGenerator;
   let mockConfig: Config;

packages/core/src/core/openaiContentGenerator/openaiContentGenerator.ts

@@ -13,6 +13,7 @@ import type { PipelineConfig } from './pipeline.js';
 import { ContentGenerationPipeline } from './pipeline.js';
 import { DefaultTelemetryService } from './telemetryService.js';
 import { EnhancedErrorHandler } from './errorHandler.js';
+import { getDefaultTokenizer } from '../../utils/request-tokenizer/index.js';
 import type { ContentGeneratorConfig } from '../contentGenerator.js';
 
 export class OpenAIContentGenerator implements ContentGenerator {
@@ -71,27 +72,30 @@ export class OpenAIContentGenerator implements ContentGenerator {
   async countTokens(
     request: CountTokensParameters,
   ): Promise<CountTokensResponse> {
-    // Use tiktoken for accurate token counting
-    const content = JSON.stringify(request.contents);
-    let totalTokens = 0;
-
     try {
-      const { get_encoding } = await import('tiktoken');
-      const encoding = get_encoding('cl100k_base'); // GPT-4 encoding, but estimate for qwen
-      totalTokens = encoding.encode(content).length;
-      encoding.free();
+      // Use the new high-performance request tokenizer
+      const tokenizer = getDefaultTokenizer();
+      const result = await tokenizer.calculateTokens(request, {
+        textEncoding: 'cl100k_base', // Use GPT-4 encoding for consistency
+      });
+
+      return {
+        totalTokens: result.totalTokens,
+      };
     } catch (error) {
       console.warn(
-        'Failed to load tiktoken, falling back to character approximation:',
+        'Failed to calculate tokens with new tokenizer, falling back to simple method:',
         error,
       );
-      // Fallback: rough approximation using character count
-      totalTokens = Math.ceil(content.length / 4); // Rough estimate: 1 token ≈ 4 characters
-    }
 
-    return {
-      totalTokens,
-    };
+      // Fallback to original simple method
+      const content = JSON.stringify(request.contents);
+      const totalTokens = Math.ceil(content.length / 4); // Rough estimate: 1 token ≈ 4 characters
+
+      return {
+        totalTokens,
+      };
+    }
   }
 
   async embedContent(

packages/core/src/core/openaiContentGenerator/pipeline.test.ts

@@ -1105,5 +1105,164 @@ describe('ContentGenerationPipeline', () => {
         expect.any(Array),
       );
     });
+
+    it('should collect all OpenAI chunks for logging even when Gemini responses are filtered', async () => {
+      // Create chunks that would produce empty Gemini responses (partial tool calls)
+      const partialToolCallChunk1: OpenAI.Chat.ChatCompletionChunk = {
+        id: 'chunk-1',
+        object: 'chat.completion.chunk',
+        created: Date.now(),
+        model: 'test-model',
+        choices: [
+          {
+            index: 0,
+            delta: {
+              tool_calls: [
+                {
+                  index: 0,
+                  id: 'call_123',
+                  type: 'function',
+                  function: { name: 'test_function', arguments: '{"par' },
+                },
+              ],
+            },
+            finish_reason: null,
+          },
+        ],
+      };
+
+      const partialToolCallChunk2: OpenAI.Chat.ChatCompletionChunk = {
+        id: 'chunk-2',
+        object: 'chat.completion.chunk',
+        created: Date.now(),
+        model: 'test-model',
+        choices: [
+          {
+            index: 0,
+            delta: {
+              tool_calls: [
+                {
+                  index: 0,
+                  function: { arguments: 'am": "value"}' },
+                },
+              ],
+            },
+            finish_reason: null,
+          },
+        ],
+      };
+
+      const finishChunk: OpenAI.Chat.ChatCompletionChunk = {
+        id: 'chunk-3',
+        object: 'chat.completion.chunk',
+        created: Date.now(),
+        model: 'test-model',
+        choices: [
+          {
+            index: 0,
+            delta: {},
+            finish_reason: 'tool_calls',
+          },
+        ],
+      };
+
+      // Mock empty Gemini responses for partial chunks (they get filtered)
+      const emptyGeminiResponse1 = new GenerateContentResponse();
+      emptyGeminiResponse1.candidates = [
+        {
+          content: { parts: [], role: 'model' },
+          index: 0,
+          safetyRatings: [],
+        },
+      ];
+
+      const emptyGeminiResponse2 = new GenerateContentResponse();
+      emptyGeminiResponse2.candidates = [
+        {
+          content: { parts: [], role: 'model' },
+          index: 0,
+          safetyRatings: [],
+        },
+      ];
+
+      // Mock final Gemini response with tool call
+      const finalGeminiResponse = new GenerateContentResponse();
+      finalGeminiResponse.candidates = [
+        {
+          content: {
+            parts: [
+              {
+                functionCall: {
+                  id: 'call_123',
+                  name: 'test_function',
+                  args: { param: 'value' },
+                },
+              },
+            ],
+            role: 'model',
+          },
+          finishReason: FinishReason.STOP,
+          index: 0,
+          safetyRatings: [],
+        },
+      ];
+
+      // Setup converter mocks
+      (mockConverter.convertGeminiRequestToOpenAI as Mock).mockReturnValue([
+        { role: 'user', content: 'test' },
+      ]);
+      (mockConverter.convertOpenAIChunkToGemini as Mock)
+        .mockReturnValueOnce(emptyGeminiResponse1) // First partial chunk -> empty response
+        .mockReturnValueOnce(emptyGeminiResponse2) // Second partial chunk -> empty response
+        .mockReturnValueOnce(finalGeminiResponse); // Finish chunk -> complete response
+
+      // Mock stream
+      const mockStream = {
+        async *[Symbol.asyncIterator]() {
+          yield partialToolCallChunk1;
+          yield partialToolCallChunk2;
+          yield finishChunk;
+        },
+      };
+
+      (mockClient.chat.completions.create as Mock).mockResolvedValue(
+        mockStream,
+      );
+
+      const request: GenerateContentParameters = {
+        model: 'test-model',
+        contents: [{ role: 'user', parts: [{ text: 'test' }] }],
+      };
+
+      // Collect responses
+      const responses: GenerateContentResponse[] = [];
+      const resultGenerator = await pipeline.executeStream(
+        request,
+        'test-prompt-id',
+      );
+      for await (const response of resultGenerator) {
+        responses.push(response);
+      }
+
+      // Should only yield the final response (empty ones are filtered)
+      expect(responses).toHaveLength(1);
+      expect(responses[0]).toBe(finalGeminiResponse);
+
+      // Verify telemetry was called with ALL OpenAI chunks, including the filtered ones
+      expect(mockTelemetryService.logStreamingSuccess).toHaveBeenCalledWith(
+        expect.objectContaining({
+          model: 'test-model',
+          duration: expect.any(Number),
+          userPromptId: 'test-prompt-id',
+          authType: 'openai',
+        }),
+        [finalGeminiResponse], // Only the non-empty Gemini response
+        expect.objectContaining({
+          model: 'test-model',
+          messages: [{ role: 'user', content: 'test' }],
+        }),
+        [partialToolCallChunk1, partialToolCallChunk2, finishChunk], // ALL OpenAI chunks
+      );
+    });
   });
 });

packages/core/src/core/openaiContentGenerator/pipeline.ts

@@ -10,14 +10,11 @@ import {
   GenerateContentResponse,
 } from '@google/genai';
 import type { Config } from '../../config/config.js';
-import { type ContentGeneratorConfig } from '../contentGenerator.js';
-import { type OpenAICompatibleProvider } from './provider/index.js';
+import type { ContentGeneratorConfig } from '../contentGenerator.js';
+import type { OpenAICompatibleProvider } from './provider/index.js';
 import { OpenAIContentConverter } from './converter.js';
-import {
-  type TelemetryService,
-  type RequestContext,
-} from './telemetryService.js';
-import { type ErrorHandler } from './errorHandler.js';
+import type { TelemetryService, RequestContext } from './telemetryService.js';
+import type { ErrorHandler } from './errorHandler.js';
 
 export interface PipelineConfig {
   cliConfig: Config;
@@ -101,7 +98,7 @@ export class ContentGenerationPipeline {
    * 2. Filter empty responses
    * 3. Handle chunk merging for providers that send finishReason and usageMetadata separately
    * 4. Collect both formats for logging
-   * 5. Handle success/error logging with original OpenAI format
+   * 5. Handle success/error logging
    */
   private async *processStreamWithLogging(
     stream: AsyncIterable<OpenAI.Chat.ChatCompletionChunk>,
@@ -121,6 +118,9 @@ export class ContentGenerationPipeline {
     try {
       // Stage 2a: Convert and yield each chunk while preserving original
       for await (const chunk of stream) {
+        // Always collect OpenAI chunks for logging, regardless of Gemini conversion result
+        collectedOpenAIChunks.push(chunk);
+
         const response = this.converter.convertOpenAIChunkToGemini(chunk);
 
         // Stage 2b: Filter empty responses to avoid downstream issues
@@ -135,9 +135,7 @@ export class ContentGenerationPipeline {
         // Stage 2c: Handle chunk merging for providers that send finishReason and usageMetadata separately
         const shouldYield = this.handleChunkMerging(
           response,
-          chunk,
           collectedGeminiResponses,
-          collectedOpenAIChunks,
           (mergedResponse) => {
             pendingFinishResponse = mergedResponse;
           },
@@ -169,19 +167,11 @@ export class ContentGenerationPipeline {
         collectedOpenAIChunks,
       );
     } catch (error) {
-      // Stage 2e: Stream failed - handle error and logging
-      context.duration = Date.now() - context.startTime;
-
       // Clear streaming tool calls on error to prevent data pollution
       this.converter.resetStreamingToolCalls();
 
-      await this.config.telemetryService.logError(
-        context,
-        error,
-        openaiRequest,
-      );
-
-      this.config.errorHandler.handle(error, context, request);
+      // Use shared error handling logic
+      await this.handleError(error, context, request);
     }
   }
 
@@ -193,17 +183,13 @@ export class ContentGenerationPipeline {
    * finishReason and the most up-to-date usage information from any provider pattern.
    *
    * @param response Current Gemini response
-   * @param chunk Current OpenAI chunk
    * @param collectedGeminiResponses Array to collect responses for logging
-   * @param collectedOpenAIChunks Array to collect chunks for logging
    * @param setPendingFinish Callback to set pending finish response
    * @returns true if the response should be yielded, false if it should be held for merging
    */
   private handleChunkMerging(
     response: GenerateContentResponse,
-    chunk: OpenAI.Chat.ChatCompletionChunk,
     collectedGeminiResponses: GenerateContentResponse[],
-    collectedOpenAIChunks: OpenAI.Chat.ChatCompletionChunk[],
     setPendingFinish: (response: GenerateContentResponse) => void,
   ): boolean {
     const isFinishChunk = response.candidates?.[0]?.finishReason;
@@ -217,7 +203,6 @@ export class ContentGenerationPipeline {
     if (isFinishChunk) {
       // This is a finish reason chunk
       collectedGeminiResponses.push(response);
-      collectedOpenAIChunks.push(chunk);
       setPendingFinish(response);
       return false; // Don't yield yet, wait for potential subsequent chunks to merge
     } else if (hasPendingFinish) {
@@ -239,7 +224,6 @@ export class ContentGenerationPipeline {
       // Update the collected responses with the merged response
       collectedGeminiResponses[collectedGeminiResponses.length - 1] =
         mergedResponse;
-      collectedOpenAIChunks.push(chunk);
 
       setPendingFinish(mergedResponse);
       return true; // Yield the merged response
@@ -247,7 +231,6 @@ export class ContentGenerationPipeline {
 
     // Normal chunk - collect and yield
     collectedGeminiResponses.push(response);
-    collectedOpenAIChunks.push(chunk);
     return true;
   }
 
@@ -265,26 +248,23 @@ export class ContentGenerationPipeline {
       ...this.buildSamplingParameters(request),
     };
 
-    // Let provider enhance the request (e.g., add metadata, cache control)
-    const enhancedRequest = this.config.provider.buildRequest(
-      baseRequest,
-      userPromptId,
-    );
+    // Add streaming options if present
+    if (streaming) {
+      (
+        baseRequest as unknown as OpenAI.Chat.ChatCompletionCreateParamsStreaming
+      ).stream = true;
+      baseRequest.stream_options = { include_usage: true };
+    }
 
     // Add tools if present
     if (request.config?.tools) {
-      enhancedRequest.tools = await this.converter.convertGeminiToolsToOpenAI(
+      baseRequest.tools = await this.converter.convertGeminiToolsToOpenAI(
         request.config.tools,
       );
     }
 
-    // Add streaming options if needed
-    if (streaming) {
-      enhancedRequest.stream = true;
-      enhancedRequest.stream_options = { include_usage: true };
-    }
-
-    return enhancedRequest;
+    // Let provider enhance the request (e.g., add metadata, cache control)
+    return this.config.provider.buildRequest(baseRequest, userPromptId);
   }
 
   private buildSamplingParameters(
@@ -322,9 +302,9 @@ export class ContentGenerationPipeline {
     };
 
     const params = {
-      // Parameters with request fallback and defaults
-      temperature: getParameterValue('temperature', 'temperature', 0.0),
-      top_p: getParameterValue('top_p', 'topP', 1.0),
+      // Parameters with request fallback but no defaults
+      ...addParameterIfDefined('temperature', 'temperature', 'temperature'),
+      ...addParameterIfDefined('top_p', 'top_p', 'topP'),
 
       // Max tokens (special case: different property names)
       ...addParameterIfDefined('max_tokens', 'max_tokens', 'maxOutputTokens'),
@@ -365,25 +345,59 @@ export class ContentGenerationPipeline {
       context.duration = Date.now() - context.startTime;
       return result;
     } catch (error) {
-      context.duration = Date.now() - context.startTime;
-
-      // Log error
-      const openaiRequest = await this.buildRequest(
+      // Use shared error handling logic
+      return await this.handleError(
+        error,
+        context,
         request,
         userPromptId,
         isStreaming,
       );
-      await this.config.telemetryService.logError(
-        context,
-        error,
-        openaiRequest,
-      );
-
-      // Handle and throw enhanced error
-      this.config.errorHandler.handle(error, context, request);
     }
   }
 
+  /**
+   * Shared error handling logic for both executeWithErrorHandling and processStreamWithLogging
+   * This centralizes the common error processing steps to avoid duplication
+   */
+  private async handleError(
+    error: unknown,
+    context: RequestContext,
+    request: GenerateContentParameters,
+    userPromptId?: string,
+    isStreaming?: boolean,
+  ): Promise<never> {
+    context.duration = Date.now() - context.startTime;
+
+    // Build request for logging (may fail, but we still want to log the error)
+    let openaiRequest: OpenAI.Chat.ChatCompletionCreateParams;
+    try {
+      if (userPromptId !== undefined && isStreaming !== undefined) {
+        openaiRequest = await this.buildRequest(
+          request,
+          userPromptId,
+          isStreaming,
+        );
+      } else {
+        // For processStreamWithLogging, we don't have userPromptId/isStreaming,
+        // so create a minimal request
+        openaiRequest = {
+          model: this.contentGeneratorConfig.model,
+          messages: [],
+        };
+      }
+    } catch (_buildError) {
+      // If we can't build the request, create a minimal one for logging
+      openaiRequest = {
+        model: this.contentGeneratorConfig.model,
+        messages: [],
+      };
+    }
+
+    await this.config.telemetryService.logError(context, error, openaiRequest);
+    this.config.errorHandler.handle(error, context, request);
+  }
+
   /**
    * Create request context with common properties
    */

packages/core/src/core/openaiContentGenerator/provider/dashscope.test.ts

@@ -17,6 +17,7 @@ import { DashScopeOpenAICompatibleProvider } from './dashscope.js';
 import type { Config } from '../../../config/config.js';
 import type { ContentGeneratorConfig } from '../../contentGenerator.js';
 import { AuthType } from '../../contentGenerator.js';
+import type { ChatCompletionToolWithCache } from './types.js';
 import { DEFAULT_TIMEOUT, DEFAULT_MAX_RETRIES } from '../constants.js';
 
 // Mock OpenAI
@@ -253,17 +254,110 @@ describe('DashScopeOpenAICompatibleProvider', () => {
         },
       ]);
 
-      // Last message should NOT have cache control for non-streaming
+      // Last message should NOT have cache control for non-streaming requests
       const lastMessage = result.messages[1];
       expect(lastMessage.role).toBe('user');
       expect(lastMessage.content).toBe('Hello!');
     });
 
-    it('should add cache control to both system and last messages for streaming requests', () => {
-      const request = { ...baseRequest, stream: true };
-      const result = provider.buildRequest(request, 'test-prompt-id');
+    it('should add cache control to system message only for non-streaming requests with tools', () => {
+      const requestWithTool: OpenAI.Chat.ChatCompletionCreateParams = {
+        ...baseRequest,
+        messages: [
+          { role: 'system', content: 'You are a helpful assistant.' },
+          {
+            role: 'tool',
+            content: 'First tool output',
+            tool_call_id: 'call_1',
+          },
+          {
+            role: 'tool',
+            content: 'Second tool output',
+            tool_call_id: 'call_2',
+          },
+          { role: 'user', content: 'Hello!' },
+        ],
+        tools: [
+          {
+            type: 'function',
+            function: {
+              name: 'mockTool',
+              parameters: { type: 'object', properties: {} },
+            },
+          },
+        ],
+        stream: false,
+      };
 
-      expect(result.messages).toHaveLength(2);
+      const result = provider.buildRequest(requestWithTool, 'test-prompt-id');
+
+      expect(result.messages).toHaveLength(4);
+
+      const systemMessage = result.messages[0];
+      expect(systemMessage.content).toEqual([
+        {
+          type: 'text',
+          text: 'You are a helpful assistant.',
+          cache_control: { type: 'ephemeral' },
+        },
+      ]);
+
+      // Tool messages should remain unchanged
+      const firstToolMessage = result.messages[1];
+      expect(firstToolMessage.role).toBe('tool');
+      expect(firstToolMessage.content).toBe('First tool output');
+
+      const secondToolMessage = result.messages[2];
+      expect(secondToolMessage.role).toBe('tool');
+      expect(secondToolMessage.content).toBe('Second tool output');
+
+      // Last message should NOT have cache control for non-streaming requests
+      const lastMessage = result.messages[3];
+      expect(lastMessage.role).toBe('user');
+      expect(lastMessage.content).toBe('Hello!');
+
+      // Tools should NOT have cache control for non-streaming requests
+      const tools = result.tools as ChatCompletionToolWithCache[];
+      expect(tools).toBeDefined();
+      expect(tools).toHaveLength(1);
+      expect(tools[0].cache_control).toBeUndefined();
+    });
+
+    it('should add cache control to system, last history message, and last tool definition for streaming requests', () => {
+      const request = { ...baseRequest, stream: true };
+      const requestWithToolMessage: OpenAI.Chat.ChatCompletionCreateParams = {
+        ...request,
+        messages: [
+          { role: 'system', content: 'You are a helpful assistant.' },
+          {
+            role: 'tool',
+            content: 'First tool output',
+            tool_call_id: 'call_1',
+          },
+          {
+            role: 'tool',
+            content: 'Second tool output',
+            tool_call_id: 'call_2',
+          },
+          { role: 'user', content: 'Hello!' },
+        ],
+        tools: [
+          {
+            type: 'function',
+            function: {
+              name: 'mockTool',
+              parameters: { type: 'object', properties: {} },
+            },
+          },
+        ],
+      };
+
+      const result = provider.buildRequest(
+        requestWithToolMessage,
+        'test-prompt-id',
+      );
+
+      expect(result.messages).toHaveLength(4);
 
       // System message should have cache control
       const systemMessage = result.messages[0];
@@ -275,8 +369,17 @@ describe('DashScopeOpenAICompatibleProvider', () => {
         },
       ]);
 
-      // Last message should also have cache control for streaming
-      const lastMessage = result.messages[1];
+      // Tool messages should remain unchanged
+      const firstToolMessage = result.messages[1];
+      expect(firstToolMessage.role).toBe('tool');
+      expect(firstToolMessage.content).toBe('First tool output');
+
+      const secondToolMessage = result.messages[2];
+      expect(secondToolMessage.role).toBe('tool');
+      expect(secondToolMessage.content).toBe('Second tool output');
+
+      // Last message should also have cache control
+      const lastMessage = result.messages[3];
       expect(lastMessage.content).toEqual([
         {
           type: 'text',
@@ -284,6 +387,40 @@ describe('DashScopeOpenAICompatibleProvider', () => {
           cache_control: { type: 'ephemeral' },
         },
       ]);
+
+      const tools = result.tools as ChatCompletionToolWithCache[];
+      expect(tools).toBeDefined();
+      expect(tools).toHaveLength(1);
+      expect(tools[0].cache_control).toEqual({ type: 'ephemeral' });
+    });
+
+    it('should not add cache control to tool messages when request.tools is undefined', () => {
+      const requestWithoutConfiguredTools: OpenAI.Chat.ChatCompletionCreateParams =
+        {
+          ...baseRequest,
+          messages: [
+            { role: 'system', content: 'You are a helpful assistant.' },
+            {
+              role: 'tool',
+              content: 'Tool output',
+              tool_call_id: 'call_1',
+            },
+            { role: 'user', content: 'Hello!' },
+          ],
+        };
+
+      const result = provider.buildRequest(
+        requestWithoutConfiguredTools,
+        'test-prompt-id',
+      );
+
+      expect(result.messages).toHaveLength(3);
+
+      const toolMessage = result.messages[1];
+      expect(toolMessage.role).toBe('tool');
+      expect(toolMessage.content).toBe('Tool output');
+
+      expect(result.tools).toBeUndefined();
     });
 
     it('should include metadata in the request', () => {
@@ -560,4 +697,200 @@ describe('DashScopeOpenAICompatibleProvider', () => {
       ]);
     });
   });
+
+  describe('output token limits', () => {
+    it('should limit max_tokens when it exceeds model limit for qwen3-coder-plus', () => {
+      const request: OpenAI.Chat.ChatCompletionCreateParams = {
+        model: 'qwen3-coder-plus',
+        messages: [{ role: 'user', content: 'Hello' }],
+        max_tokens: 100000, // Exceeds the 65536 limit
+      };
+
+      const result = provider.buildRequest(request, 'test-prompt-id');
+
+      expect(result.max_tokens).toBe(65536); // Should be limited to model's output limit
+    });
+
+    it('should limit max_tokens when it exceeds model limit for qwen-vl-max-latest', () => {
+      const request: OpenAI.Chat.ChatCompletionCreateParams = {
+        model: 'qwen-vl-max-latest',
+        messages: [{ role: 'user', content: 'Hello' }],
+        max_tokens: 20000, // Exceeds the 8192 limit
+      };
+
+      const result = provider.buildRequest(request, 'test-prompt-id');
+
+      expect(result.max_tokens).toBe(8192); // Should be limited to model's output limit
+    });
+
+    it('should not modify max_tokens when it is within model limit', () => {
+      const request: OpenAI.Chat.ChatCompletionCreateParams = {
+        model: 'qwen3-coder-plus',
+        messages: [{ role: 'user', content: 'Hello' }],
+        max_tokens: 1000, // Within the 65536 limit
+      };
+
+      const result = provider.buildRequest(request, 'test-prompt-id');
+
+      expect(result.max_tokens).toBe(1000); // Should remain unchanged
+    });
+
+    it('should not add max_tokens when not present in request', () => {
+      const request: OpenAI.Chat.ChatCompletionCreateParams = {
+        model: 'qwen3-coder-plus',
+        messages: [{ role: 'user', content: 'Hello' }],
+        // No max_tokens parameter
+      };
+
+      const result = provider.buildRequest(request, 'test-prompt-id');
+
+      expect(result.max_tokens).toBeUndefined(); // Should remain undefined
+    });
+
+    it('should handle null max_tokens parameter', () => {
+      const request: OpenAI.Chat.ChatCompletionCreateParams = {
+        model: 'qwen3-coder-plus',
+        messages: [{ role: 'user', content: 'Hello' }],
+        max_tokens: null,
+      };
+
+      const result = provider.buildRequest(request, 'test-prompt-id');
+
+      expect(result.max_tokens).toBeNull(); // Should remain null
+    });
+
+    it('should use default output limit for unknown models', () => {
+      const request: OpenAI.Chat.ChatCompletionCreateParams = {
+        model: 'unknown-model',
+        messages: [{ role: 'user', content: 'Hello' }],
+        max_tokens: 10000, // Exceeds the default 4096 limit
+      };
+
+      const result = provider.buildRequest(request, 'test-prompt-id');
+
+      expect(result.max_tokens).toBe(4096); // Should be limited to default output limit
+    });
+
+    it('should preserve other request parameters when limiting max_tokens', () => {
+      const request: OpenAI.Chat.ChatCompletionCreateParams = {
+        model: 'qwen3-coder-plus',
+        messages: [{ role: 'user', content: 'Hello' }],
+        max_tokens: 100000, // Will be limited
+        temperature: 0.8,
+        top_p: 0.9,
+        frequency_penalty: 0.1,
+        presence_penalty: 0.2,
+        stop: ['END'],
+        user: 'test-user',
+      };
+
+      const result = provider.buildRequest(request, 'test-prompt-id');
+
+      // max_tokens should be limited
+      expect(result.max_tokens).toBe(65536);
+
+      // Other parameters should be preserved
+      expect(result.temperature).toBe(0.8);
+      expect(result.top_p).toBe(0.9);
+      expect(result.frequency_penalty).toBe(0.1);
+      expect(result.presence_penalty).toBe(0.2);
+      expect(result.stop).toEqual(['END']);
+      expect(result.user).toBe('test-user');
+    });
+
+    it('should work with vision models and output token limits', () => {
+      const request: OpenAI.Chat.ChatCompletionCreateParams = {
+        model: 'qwen-vl-max-latest',
+        messages: [
+          {
+            role: 'user',
+            content: [
+              { type: 'text', text: 'Look at this image:' },
+              {
+                type: 'image_url',
+                image_url: { url: 'https://example.com/image.jpg' },
+              },
+            ],
+          },
+        ],
+        max_tokens: 20000, // Exceeds the 8192 limit
+      };
+
+      const result = provider.buildRequest(request, 'test-prompt-id');
+
+      expect(result.max_tokens).toBe(8192); // Should be limited
+      expect(
+        (result as { vl_high_resolution_images?: boolean })
+          .vl_high_resolution_images,
+      ).toBe(true); // Vision-specific parameter should be preserved
+    });
+
+    it('should set high resolution flag for qwen3-vl-plus', () => {
+      const request: OpenAI.Chat.ChatCompletionCreateParams = {
+        model: 'qwen3-vl-plus',
+        messages: [
+          {
+            role: 'user',
+            content: [
+              { type: 'text', text: 'Please inspect the image.' },
+              {
+                type: 'image_url',
+                image_url: { url: 'https://example.com/vl.jpg' },
+              },
+            ],
+          },
+        ],
+        max_tokens: 50000,
+      };
+
+      const result = provider.buildRequest(request, 'test-prompt-id');
+
+      expect(result.max_tokens).toBe(32768);
+      expect(
+        (result as { vl_high_resolution_images?: boolean })
+          .vl_high_resolution_images,
+      ).toBe(true);
+    });
+
+    it('should set high resolution flag for the vision-model alias', () => {
+      const request: OpenAI.Chat.ChatCompletionCreateParams = {
+        model: 'vision-model',
+        messages: [
+          {
+            role: 'user',
+            content: [
+              { type: 'text', text: 'Alias payload' },
+              {
+                type: 'image_url',
+                image_url: { url: 'https://example.com/alias.png' },
+              },
+            ],
+          },
+        ],
+        max_tokens: 9000,
+      };
+
+      const result = provider.buildRequest(request, 'test-prompt-id');
+
+      expect(result.max_tokens).toBe(8192);
+      expect(
+        (result as { vl_high_resolution_images?: boolean })
+          .vl_high_resolution_images,
+      ).toBe(true);
+    });
+
+    it('should handle streaming requests with output token limits', () => {
+      const request: OpenAI.Chat.ChatCompletionCreateParams = {
+        model: 'qwen3-coder-plus',
+        messages: [{ role: 'user', content: 'Hello' }],
+        max_tokens: 100000, // Exceeds the 65536 limit
+        stream: true,
+      };
+
+      const result = provider.buildRequest(request, 'test-prompt-id');
+
+      expect(result.max_tokens).toBe(65536); // Should be limited
+      expect(result.stream).toBe(true); // Streaming should be preserved
+    });
+  });
 });

packages/core/src/core/openaiContentGenerator/provider/dashscope.ts

@@ -3,11 +3,13 @@ import type { Config } from '../../../config/config.js';
 import type { ContentGeneratorConfig } from '../../contentGenerator.js';
 import { AuthType } from '../../contentGenerator.js';
 import { DEFAULT_TIMEOUT, DEFAULT_MAX_RETRIES } from '../constants.js';
+import { tokenLimit } from '../../tokenLimits.js';
 import type {
   OpenAICompatibleProvider,
   DashScopeRequestMetadata,
   ChatCompletionContentPartTextWithCache,
   ChatCompletionContentPartWithCache,
+  ChatCompletionToolWithCache,
 } from './types.js';
 
 export class DashScopeOpenAICompatibleProvider
@@ -65,25 +67,62 @@ export class DashScopeOpenAICompatibleProvider
     });
   }
 
+  /**
+   * Build and configure the request for DashScope API.
+   *
+   * This method applies DashScope-specific configurations including:
+   * - Cache control for the system message, last tool message (when tools are configured),
+   *   and the latest history message
+   * - Output token limits based on model capabilities
+   * - Vision model specific parameters (vl_high_resolution_images)
+   * - Request metadata for session tracking
+   *
+   * @param request - The original chat completion request parameters
+   * @param userPromptId - Unique identifier for the user prompt for session tracking
+   * @returns Configured request with DashScope-specific parameters applied
+   */
   buildRequest(
     request: OpenAI.Chat.ChatCompletionCreateParams,
     userPromptId: string,
   ): OpenAI.Chat.ChatCompletionCreateParams {
     let messages = request.messages;
+    let tools = request.tools;
 
     // Apply DashScope cache control only if not disabled
     if (!this.shouldDisableCacheControl()) {
-      // Add cache control to system and last messages for DashScope providers
-      // Only add cache control to system message for non-streaming requests
-      const cacheTarget = request.stream ? 'both' : 'system';
-      messages = this.addDashScopeCacheControl(messages, cacheTarget);
+      const { messages: updatedMessages, tools: updatedTools } =
+        this.addDashScopeCacheControl(
+          request,
+          request.stream ? 'all' : 'system_only',
+        );
+      messages = updatedMessages;
+      tools = updatedTools;
+    }
+
+    // Apply output token limits based on model capabilities
+    // This ensures max_tokens doesn't exceed the model's maximum output limit
+    const requestWithTokenLimits = this.applyOutputTokenLimit(
+      request,
+      request.model,
+    );
+
+    if (this.isVisionModel(request.model)) {
+      return {
+        ...requestWithTokenLimits,
+        messages,
+        ...(tools ? { tools } : {}),
+        ...(this.buildMetadata(userPromptId) || {}),
+        /* @ts-expect-error dashscope exclusive */
+        vl_high_resolution_images: true,
+      } as OpenAI.Chat.ChatCompletionCreateParams;
     }
 
     return {
-      ...request, // Preserve all original parameters including sampling params
+      ...requestWithTokenLimits, // Preserve all original parameters including sampling params and adjusted max_tokens
       messages,
+      ...(tools ? { tools } : {}),
       ...(this.buildMetadata(userPromptId) || {}),
-    };
+    } as OpenAI.Chat.ChatCompletionCreateParams;
   }
 
   buildMetadata(userPromptId: string): DashScopeRequestMetadata {
@@ -99,75 +138,67 @@ export class DashScopeOpenAICompatibleProvider
    * Add cache control flag to specified message(s) for DashScope providers
    */
   private addDashScopeCacheControl(
-    messages: OpenAI.Chat.ChatCompletionMessageParam[],
-    target: 'system' | 'last' | 'both' = 'both',
-  ): OpenAI.Chat.ChatCompletionMessageParam[] {
-    if (messages.length === 0) {
-      return messages;
-    }
+    request: OpenAI.Chat.ChatCompletionCreateParams,
+    cacheControl: 'system_only' | 'all',
+  ): {
+    messages: OpenAI.Chat.ChatCompletionMessageParam[];
+    tools?: ChatCompletionToolWithCache[];
+  } {
+    const messages = request.messages;
 
-    let updatedMessages = [...messages];
+    const systemIndex = messages.findIndex((msg) => msg.role === 'system');
+    const lastIndex = messages.length - 1;
 
-    // Add cache control to system message if requested
-    if (target === 'system' || target === 'both') {
-      updatedMessages = this.addCacheControlToMessage(
-        updatedMessages,
-        'system',
-      );
-    }
+    const updatedMessages =
+      messages.length === 0
+        ? messages
+        : messages.map((message, index) => {
+            const shouldAddCacheControl = Boolean(
+              (index === systemIndex && systemIndex !== -1) ||
+                (index === lastIndex && cacheControl === 'all'),
+            );
 
-    // Add cache control to last message if requested
-    if (target === 'last' || target === 'both') {
-      updatedMessages = this.addCacheControlToMessage(updatedMessages, 'last');
-    }
+            if (
+              !shouldAddCacheControl ||
+              !('content' in message) ||
+              message.content === null ||
+              message.content === undefined
+            ) {
+              return message;
+            }
 
-    return updatedMessages;
+            return {
+              ...message,
+              content: this.addCacheControlToContent(message.content),
+            } as OpenAI.Chat.ChatCompletionMessageParam;
+          });
+
+    const updatedTools =
+      cacheControl === 'all' && request.tools?.length
+        ? this.addCacheControlToTools(request.tools)
+        : (request.tools as ChatCompletionToolWithCache[] | undefined);
+
+    return {
+      messages: updatedMessages,
+      tools: updatedTools,
+    };
   }
 
-  /**
-   * Helper method to add cache control to a specific message
-   */
-  private addCacheControlToMessage(
-    messages: OpenAI.Chat.ChatCompletionMessageParam[],
-    target: 'system' | 'last',
-  ): OpenAI.Chat.ChatCompletionMessageParam[] {
-    const updatedMessages = [...messages];
-    const messageIndex = this.findTargetMessageIndex(messages, target);
-
-    if (messageIndex === -1) {
-      return updatedMessages;
+  private addCacheControlToTools(
+    tools: OpenAI.Chat.ChatCompletionTool[],
+  ): ChatCompletionToolWithCache[] {
+    if (tools.length === 0) {
+      return tools as ChatCompletionToolWithCache[];
     }
 
-    const message = updatedMessages[messageIndex];
+    const updatedTools = [...tools] as ChatCompletionToolWithCache[];
+    const lastToolIndex = tools.length - 1;
+    updatedTools[lastToolIndex] = {
+      ...updatedTools[lastToolIndex],
+      cache_control: { type: 'ephemeral' },
+    };
 
-    // Only process messages that have content
-    if (
-      'content' in message &&
-      message.content !== null &&
-      message.content !== undefined
-    ) {
-      const updatedContent = this.addCacheControlToContent(message.content);
-      updatedMessages[messageIndex] = {
-        ...message,
-        content: updatedContent,
-      } as OpenAI.Chat.ChatCompletionMessageParam;
-    }
-
-    return updatedMessages;
-  }
-
-  /**
-   * Find the index of the target message (system or last)
-   */
-  private findTargetMessageIndex(
-    messages: OpenAI.Chat.ChatCompletionMessageParam[],
-    target: 'system' | 'last',
-  ): number {
-    if (target === 'system') {
-      return messages.findIndex((msg) => msg.role === 'system');
-    } else {
-      return messages.length - 1;
-    }
+    return updatedTools;
   }
 
   /**
@@ -236,6 +267,63 @@ export class DashScopeOpenAICompatibleProvider
     return contentArray;
   }
 
+  private isVisionModel(model: string | undefined): boolean {
+    if (!model) {
+      return false;
+    }
+
+    const normalized = model.toLowerCase();
+
+    if (normalized === 'vision-model') {
+      return true;
+    }
+
+    if (normalized.startsWith('qwen-vl')) {
+      return true;
+    }
+
+    if (normalized.startsWith('qwen3-vl-plus')) {
+      return true;
+    }
+
+    return false;
+  }
+
+  /**
+   * Apply output token limit to a request's max_tokens parameter.
+   *
+   * Ensures that existing max_tokens parameters don't exceed the model's maximum output
+   * token limit. Only modifies max_tokens when already present in the request.
+   *
+   * @param request - The chat completion request parameters
+   * @param model - The model name to get the output token limit for
+   * @returns The request with max_tokens adjusted to respect the model's limits (if present)
+   */
+  private applyOutputTokenLimit<T extends { max_tokens?: number | null }>(
+    request: T,
+    model: string,
+  ): T {
+    const currentMaxTokens = request.max_tokens;
+
+    // Only process if max_tokens is already present in the request
+    if (currentMaxTokens === undefined || currentMaxTokens === null) {
+      return request; // No max_tokens parameter, return unchanged
+    }
+
+    const modelLimit = tokenLimit(model, 'output');
+
+    // If max_tokens exceeds the model limit, cap it to the model's limit
+    if (currentMaxTokens > modelLimit) {
+      return {
+        ...request,
+        max_tokens: modelLimit,
+      };
+    }
+
+    // If max_tokens is within the limit, return the request unchanged
+    return request;
+  }
+
   /**
    * Check if cache control should be disabled based on configuration.
    *

packages/core/src/core/openaiContentGenerator/provider/deepseek.test.ts

@@ -0,0 +1,132 @@
+/**
+ * @license
+ * Copyright 2025 Qwen
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import { describe, it, expect, vi, beforeEach } from 'vitest';
+import type OpenAI from 'openai';
+import { DeepSeekOpenAICompatibleProvider } from './deepseek.js';
+import type { ContentGeneratorConfig } from '../../contentGenerator.js';
+import type { Config } from '../../../config/config.js';
+
+// Mock OpenAI client to avoid real network calls
+vi.mock('openai', () => ({
+  default: vi.fn().mockImplementation((config) => ({
+    config,
+  })),
+}));
+
+describe('DeepSeekOpenAICompatibleProvider', () => {
+  let provider: DeepSeekOpenAICompatibleProvider;
+  let mockContentGeneratorConfig: ContentGeneratorConfig;
+  let mockCliConfig: Config;
+
+  beforeEach(() => {
+    vi.clearAllMocks();
+
+    mockContentGeneratorConfig = {
+      apiKey: 'test-api-key',
+      baseUrl: 'https://api.deepseek.com/v1',
+      model: 'deepseek-chat',
+    } as ContentGeneratorConfig;
+
+    mockCliConfig = {
+      getCliVersion: vi.fn().mockReturnValue('1.0.0'),
+    } as unknown as Config;
+
+    provider = new DeepSeekOpenAICompatibleProvider(
+      mockContentGeneratorConfig,
+      mockCliConfig,
+    );
+  });
+
+  describe('isDeepSeekProvider', () => {
+    it('returns true when baseUrl includes deepseek', () => {
+      const result = DeepSeekOpenAICompatibleProvider.isDeepSeekProvider(
+        mockContentGeneratorConfig,
+      );
+      expect(result).toBe(true);
+    });
+
+    it('returns false for non deepseek baseUrl', () => {
+      const config = {
+        ...mockContentGeneratorConfig,
+        baseUrl: 'https://api.example.com/v1',
+      } as ContentGeneratorConfig;
+
+      const result =
+        DeepSeekOpenAICompatibleProvider.isDeepSeekProvider(config);
+      expect(result).toBe(false);
+    });
+  });
+
+  describe('buildRequest', () => {
+    const userPromptId = 'prompt-123';
+
+    it('converts array content into a string', () => {
+      const originalRequest: OpenAI.Chat.ChatCompletionCreateParams = {
+        model: 'deepseek-chat',
+        messages: [
+          {
+            role: 'user',
+            content: [
+              { type: 'text', text: 'Hello' },
+              { type: 'text', text: ' world' },
+            ],
+          },
+        ],
+      };
+
+      const result = provider.buildRequest(originalRequest, userPromptId);
+
+      expect(result.messages).toHaveLength(1);
+      expect(result.messages?.[0]).toEqual({
+        role: 'user',
+        content: 'Hello world',
+      });
+      expect(originalRequest.messages?.[0].content).toEqual([
+        { type: 'text', text: 'Hello' },
+        { type: 'text', text: ' world' },
+      ]);
+    });
+
+    it('leaves string content unchanged', () => {
+      const originalRequest: OpenAI.Chat.ChatCompletionCreateParams = {
+        model: 'deepseek-chat',
+        messages: [
+          {
+            role: 'user',
+            content: 'Hello world',
+          },
+        ],
+      };
+
+      const result = provider.buildRequest(originalRequest, userPromptId);
+
+      expect(result.messages?.[0].content).toBe('Hello world');
+    });
+
+    it('throws when encountering non-text multimodal parts', () => {
+      const originalRequest: OpenAI.Chat.ChatCompletionCreateParams = {
+        model: 'deepseek-chat',
+        messages: [
+          {
+            role: 'user',
+            content: [
+              { type: 'text', text: 'Hello' },
+              {
+                type: 'image_url',
+                image_url: { url: 'https://example.com/image.png' },
+              },
+            ],
+          },
+        ],
+      };
+
+      expect(() =>
+        provider.buildRequest(originalRequest, userPromptId),
+      ).toThrow(/only supports text content/i);
+    });
+  });
+});

packages/core/src/core/openaiContentGenerator/provider/deepseek.ts

@@ -0,0 +1,79 @@
+/**
+ * @license
+ * Copyright 2025 Qwen
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import type OpenAI from 'openai';
+import type { Config } from '../../../config/config.js';
+import type { ContentGeneratorConfig } from '../../contentGenerator.js';
+import { DefaultOpenAICompatibleProvider } from './default.js';
+
+export class DeepSeekOpenAICompatibleProvider extends DefaultOpenAICompatibleProvider {
+  constructor(
+    contentGeneratorConfig: ContentGeneratorConfig,
+    cliConfig: Config,
+  ) {
+    super(contentGeneratorConfig, cliConfig);
+  }
+
+  static isDeepSeekProvider(
+    contentGeneratorConfig: ContentGeneratorConfig,
+  ): boolean {
+    const baseUrl = contentGeneratorConfig.baseUrl ?? '';
+
+    return baseUrl.toLowerCase().includes('api.deepseek.com');
+  }
+
+  override buildRequest(
+    request: OpenAI.Chat.ChatCompletionCreateParams,
+    userPromptId: string,
+  ): OpenAI.Chat.ChatCompletionCreateParams {
+    const baseRequest = super.buildRequest(request, userPromptId);
+    if (!baseRequest.messages?.length) {
+      return baseRequest;
+    }
+
+    const messages = baseRequest.messages.map((message) => {
+      if (!('content' in message)) {
+        return message;
+      }
+
+      const { content } = message;
+
+      if (
+        typeof content === 'string' ||
+        content === null ||
+        content === undefined
+      ) {
+        return message;
+      }
+
+      if (!Array.isArray(content)) {
+        return message;
+      }
+
+      const text = content
+        .map((part) => {
+          if (part.type !== 'text') {
+            throw new Error(
+              `DeepSeek provider only supports text content. Found non-text part of type '${part.type}' in message with role '${message.role}'.`,
+            );
+          }
+
+          return part.text ?? '';
+        })
+        .join('');
+
+      return {
+        ...message,
+        content: text,
+      } as OpenAI.Chat.ChatCompletionMessageParam;
+    });
+
+    return {
+      ...baseRequest,
+      messages,
+    };
+  }
+}

packages/core/src/core/openaiContentGenerator/provider/index.ts

@@ -1,4 +1,5 @@
 export { DashScopeOpenAICompatibleProvider } from './dashscope.js';
+export { DeepSeekOpenAICompatibleProvider } from './deepseek.js';
 export { OpenRouterOpenAICompatibleProvider } from './openrouter.js';
 export { DefaultOpenAICompatibleProvider } from './default.js';
 export type {

packages/core/src/core/openaiContentGenerator/provider/types.ts

@@ -11,6 +11,10 @@ export type ChatCompletionContentPartWithCache =
   | OpenAI.Chat.ChatCompletionContentPartImage
   | OpenAI.Chat.ChatCompletionContentPartRefusal;
 
+export type ChatCompletionToolWithCache = OpenAI.Chat.ChatCompletionTool & {
+  cache_control?: { type: 'ephemeral' };
+};
+
 export interface OpenAICompatibleProvider {
   buildHeaders(): Record<string, string | undefined>;
   buildClient(): OpenAI;

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

@@ -5,7 +5,12 @@
  */
 
 import { describe, it, expect, vi, beforeEach } from 'vitest';
-import { getCoreSystemPrompt, getCustomSystemPrompt } from './prompts.js';
+import {
+  getCoreSystemPrompt,
+  getCustomSystemPrompt,
+  getSubagentSystemReminder,
+  getPlanModeSystemReminder,
+} from './prompts.js';
 import { isGitRepository } from '../utils/gitUtils.js';
 import fs from 'node:fs';
 import os from 'node:os';
@@ -364,6 +369,120 @@ describe('URL matching with trailing slash compatibility', () => {
   });
 });
 
+describe('Model-specific tool call formats', () => {
+  beforeEach(() => {
+    vi.resetAllMocks();
+    vi.stubEnv('SANDBOX', undefined);
+  });
+
+  it('should use XML format for qwen3-coder model', () => {
+    vi.mocked(isGitRepository).mockReturnValue(false);
+    const prompt = getCoreSystemPrompt(undefined, undefined, 'qwen3-coder-7b');
+
+    // Should contain XML-style tool calls
+    expect(prompt).toContain('<tool_call>');
+    expect(prompt).toContain('<function=run_shell_command>');
+    expect(prompt).toContain('<parameter=command>');
+    expect(prompt).toContain('</function>');
+    expect(prompt).toContain('</tool_call>');
+
+    // Should NOT contain bracket-style tool calls
+    expect(prompt).not.toContain('[tool_call: run_shell_command for');
+
+    // Should NOT contain JSON-style tool calls
+    expect(prompt).not.toContain('{"name": "run_shell_command"');
+
+    expect(prompt).toMatchSnapshot();
+  });
+
+  it('should use JSON format for qwen-vl model', () => {
+    vi.mocked(isGitRepository).mockReturnValue(false);
+    const prompt = getCoreSystemPrompt(undefined, undefined, 'qwen-vl-max');
+
+    // Should contain JSON-style tool calls
+    expect(prompt).toContain('<tool_call>');
+    expect(prompt).toContain('{"name": "run_shell_command"');
+    expect(prompt).toContain('"arguments": {"command": "node server.js &"}');
+    expect(prompt).toContain('</tool_call>');
+
+    // Should NOT contain bracket-style tool calls
+    expect(prompt).not.toContain('[tool_call: run_shell_command for');
+
+    // Should NOT contain XML-style tool calls with parameters
+    expect(prompt).not.toContain('<function=run_shell_command>');
+    expect(prompt).not.toContain('<parameter=command>');
+
+    expect(prompt).toMatchSnapshot();
+  });
+
+  it('should use bracket format for generic models', () => {
+    vi.mocked(isGitRepository).mockReturnValue(false);
+    const prompt = getCoreSystemPrompt(undefined, undefined, 'gpt-4');
+
+    // Should contain bracket-style tool calls
+    expect(prompt).toContain('[tool_call: run_shell_command for');
+    expect(prompt).toContain('because it must run in the background]');
+
+    // Should NOT contain XML-style tool calls
+    expect(prompt).not.toContain('<function=run_shell_command>');
+    expect(prompt).not.toContain('<parameter=command>');
+
+    // Should NOT contain JSON-style tool calls
+    expect(prompt).not.toContain('{"name": "run_shell_command"');
+
+    expect(prompt).toMatchSnapshot();
+  });
+
+  it('should use bracket format when no model is specified', () => {
+    vi.mocked(isGitRepository).mockReturnValue(false);
+    const prompt = getCoreSystemPrompt();
+
+    // Should contain bracket-style tool calls (default behavior)
+    expect(prompt).toContain('[tool_call: run_shell_command for');
+    expect(prompt).toContain('because it must run in the background]');
+
+    // Should NOT contain XML or JSON formats
+    expect(prompt).not.toContain('<function=run_shell_command>');
+    expect(prompt).not.toContain('{"name": "run_shell_command"');
+
+    expect(prompt).toMatchSnapshot();
+  });
+
+  it('should preserve model-specific formats with user memory', () => {
+    vi.mocked(isGitRepository).mockReturnValue(false);
+    const userMemory = 'User prefers concise responses.';
+    const prompt = getCoreSystemPrompt(
+      userMemory,
+      undefined,
+      'qwen3-coder-14b',
+    );
+
+    // Should contain XML-style tool calls
+    expect(prompt).toContain('<tool_call>');
+    expect(prompt).toContain('<function=run_shell_command>');
+
+    // Should contain user memory with separator
+    expect(prompt).toContain('---');
+    expect(prompt).toContain('User prefers concise responses.');
+
+    expect(prompt).toMatchSnapshot();
+  });
+
+  it('should preserve model-specific formats with sandbox environment', () => {
+    vi.stubEnv('SANDBOX', 'true');
+    vi.mocked(isGitRepository).mockReturnValue(false);
+    const prompt = getCoreSystemPrompt(undefined, undefined, 'qwen-vl-plus');
+
+    // Should contain JSON-style tool calls
+    expect(prompt).toContain('{"name": "run_shell_command"');
+
+    // Should contain sandbox instructions
+    expect(prompt).toContain('# Sandbox');
+
+    expect(prompt).toMatchSnapshot();
+  });
+});
+
 describe('getCustomSystemPrompt', () => {
   it('should handle string custom instruction without user memory', () => {
     const customInstruction =
@@ -405,3 +524,53 @@ describe('getCustomSystemPrompt', () => {
     expect(result).toContain('---');
   });
 });
+
+describe('getSubagentSystemReminder', () => {
+  it('should format single agent type correctly', () => {
+    const result = getSubagentSystemReminder(['python']);
+
+    expect(result).toMatch(/^<system-reminder>.*<\/system-reminder>$/);
+    expect(result).toContain('available agent types are: python');
+    expect(result).toContain('PROACTIVELY use the');
+  });
+
+  it('should join multiple agent types with commas', () => {
+    const result = getSubagentSystemReminder(['python', 'web', 'analysis']);
+
+    expect(result).toContain(
+      'available agent types are: python, web, analysis',
+    );
+  });
+
+  it('should handle empty array', () => {
+    const result = getSubagentSystemReminder([]);
+
+    expect(result).toContain('available agent types are: ');
+    expect(result).toContain('<system-reminder>');
+  });
+});
+
+describe('getPlanModeSystemReminder', () => {
+  it('should return plan mode system reminder with proper structure', () => {
+    const result = getPlanModeSystemReminder();
+
+    expect(result).toMatch(/^<system-reminder>[\s\S]*<\/system-reminder>$/);
+    expect(result).toContain('Plan mode is active');
+    expect(result).toContain('MUST NOT make any edits');
+  });
+
+  it('should include workflow instructions', () => {
+    const result = getPlanModeSystemReminder();
+
+    expect(result).toContain("1. Answer the user's query comprehensively");
+    expect(result).toContain("2. When you're done researching");
+    expect(result).toContain('exit_plan_mode tool');
+  });
+
+  it('should be deterministic', () => {
+    const result1 = getPlanModeSystemReminder();
+    const result2 = getPlanModeSystemReminder();
+
+    expect(result1).toBe(result2);
+  });
+});

packages/core/src/core/prompts.ts

@@ -7,18 +7,10 @@
 import path from 'node:path';
 import fs from 'node:fs';
 import os from 'node:os';
-import { EditTool } from '../tools/edit.js';
-import { GlobTool } from '../tools/glob.js';
-import { GrepTool } from '../tools/grep.js';
-import { ReadFileTool } from '../tools/read-file.js';
-import { ReadManyFilesTool } from '../tools/read-many-files.js';
-import { ShellTool } from '../tools/shell.js';
-import { WriteFileTool } from '../tools/write-file.js';
+import { ToolNames } from '../tools/tool-names.js';
 import process from 'node:process';
 import { isGitRepository } from '../utils/gitUtils.js';
-import { MemoryTool, GEMINI_CONFIG_DIR } from '../tools/memoryTool.js';
-import { TodoWriteTool } from '../tools/todoWrite.js';
-import { TaskTool } from '../tools/task.js';
+import { GEMINI_CONFIG_DIR } from '../tools/memoryTool.js';
 import type { GenerateContentConfig } from '@google/genai';
 
 export interface ModelTemplateMapping {
@@ -91,6 +83,7 @@ export function getCustomSystemPrompt(
 export function getCoreSystemPrompt(
   userMemory?: string,
   config?: SystemPromptConfig,
+  model?: string,
 ): string {
   // if GEMINI_SYSTEM_MD is set (and not 0|false), override system prompt from file
   // default path is .gemini/system.md but can be modified via custom path in GEMINI_SYSTEM_MD
@@ -177,11 +170,11 @@ You are Qwen Code, an interactive CLI agent developed by Alibaba Group, speciali
 - **Proactiveness:** Fulfill the user's request thoroughly, including reasonable, directly implied follow-up actions.
 - **Confirm Ambiguity/Expansion:** Do not take significant actions beyond the clear scope of the request without confirming with the user. If asked *how* to do something, explain first, don't just do it.
 - **Explaining Changes:** After completing a code modification or file operation *do not* provide summaries unless asked.
-- **Path Construction:** Before using any file system tool (e.g., ${ReadFileTool.Name}' or '${WriteFileTool.Name}'), you must construct the full absolute path for the file_path argument. Always combine the absolute path of the project's root directory with the file's path relative to the root. For example, if the project root is /path/to/project/ and the file is foo/bar/baz.txt, the final path you must use is /path/to/project/foo/bar/baz.txt. If the user provides a relative path, you must resolve it against the root directory to create an absolute path.
+- **Path Construction:** Before using any file system tool (e.g., ${ToolNames.READ_FILE}' or '${ToolNames.WRITE_FILE}'), you must construct the full absolute path for the file_path argument. Always combine the absolute path of the project's root directory with the file's path relative to the root. For example, if the project root is /path/to/project/ and the file is foo/bar/baz.txt, the final path you must use is /path/to/project/foo/bar/baz.txt. If the user provides a relative path, you must resolve it against the root directory to create an absolute path.
 - **Do Not revert changes:** Do not revert changes to the codebase unless asked to do so by the user. Only revert changes made by you if they have resulted in an error or if the user has explicitly asked you to revert the changes.
 
 # Task Management
-You have access to the ${TodoWriteTool.Name} tool to help you manage and plan tasks. Use these tools VERY frequently to ensure that you are tracking your tasks and giving the user visibility into your progress.
+You have access to the ${ToolNames.TODO_WRITE} tool to help you manage and plan tasks. Use these tools VERY frequently to ensure that you are tracking your tasks and giving the user visibility into your progress.
 These tools are also EXTREMELY helpful for planning tasks, and for breaking down larger complex tasks into smaller steps. If you do not use this tool when planning, you may forget to do important tasks - and that is unacceptable.
 
 It is critical that you mark todos as completed as soon as you are done with a task. Do not batch up multiple tasks before marking them as completed.
@@ -190,13 +183,13 @@ Examples:
 
 <example>
 user: Run the build and fix any type errors
-assistant: I'm going to use the ${TodoWriteTool.Name} tool to write the following items to the todo list: 
+assistant: I'm going to use the ${ToolNames.TODO_WRITE} tool to write the following items to the todo list: 
 - Run the build
 - Fix any type errors
 
 I'm now going to run the build using Bash.
 
-Looks like I found 10 type errors. I'm going to use the ${TodoWriteTool.Name} tool to write 10 items to the todo list.
+Looks like I found 10 type errors. I'm going to use the ${ToolNames.TODO_WRITE} tool to write 10 items to the todo list.
 
 marking the first todo as in_progress
 
@@ -211,7 +204,7 @@ In the above example, the assistant completes all the tasks, including the 10 er
 <example>
 user: Help me write a new feature that allows users to track their usage metrics and export them to various formats
 
-A: I'll help you implement a usage metrics tracking and export feature. Let me first use the ${TodoWriteTool.Name} tool to plan this task.
+A: I'll help you implement a usage metrics tracking and export feature. Let me first use the ${ToolNames.TODO_WRITE} tool to plan this task.
 Adding the following todos to the todo list:
 1. Research existing metrics tracking in the codebase
 2. Design the metrics collection system
@@ -232,8 +225,8 @@ I've found some existing telemetry code. Let me mark the first todo as in_progre
 
 ## Software Engineering Tasks
 When requested to perform tasks like fixing bugs, adding features, refactoring, or explaining code, follow this iterative approach:
-- **Plan:** After understanding the user's request, create an initial plan based on your existing knowledge and any immediately obvious context. Use the '${TodoWriteTool.Name}' tool to capture this rough plan for complex or multi-step work. Don't wait for complete understanding - start with what you know.
-- **Implement:** Begin implementing the plan while gathering additional context as needed. Use '${GrepTool.Name}', '${GlobTool.Name}', '${ReadFileTool.Name}', and '${ReadManyFilesTool.Name}' tools strategically when you encounter specific unknowns during implementation. Use the available tools (e.g., '${EditTool.Name}', '${WriteFileTool.Name}' '${ShellTool.Name}' ...) to act on the plan, strictly adhering to the project's established conventions (detailed under 'Core Mandates').
+- **Plan:** After understanding the user's request, create an initial plan based on your existing knowledge and any immediately obvious context. Use the '${ToolNames.TODO_WRITE}' tool to capture this rough plan for complex or multi-step work. Don't wait for complete understanding - start with what you know.
+- **Implement:** Begin implementing the plan while gathering additional context as needed. Use '${ToolNames.GREP}', '${ToolNames.GLOB}', '${ToolNames.READ_FILE}', and '${ToolNames.READ_MANY_FILES}' tools strategically when you encounter specific unknowns during implementation. Use the available tools (e.g., '${ToolNames.EDIT}', '${ToolNames.WRITE_FILE}' '${ToolNames.SHELL}' ...) to act on the plan, strictly adhering to the project's established conventions (detailed under 'Core Mandates').
 - **Adapt:** As you discover new information or encounter obstacles, update your plan and todos accordingly. Mark todos as in_progress when starting and completed when finishing each task. Add new todos if the scope expands. Refine your approach based on what you learn.
 - **Verify (Tests):** If applicable and feasible, verify the changes using the project's testing procedures. Identify the correct test commands and frameworks by examining 'README' files, build/package configuration (e.g., 'package.json'), or existing test execution patterns. NEVER assume standard test commands.
 - **Verify (Standards):** VERY IMPORTANT: After making code changes, execute the project-specific build, linting and type-checking commands (e.g., 'tsc', 'npm run lint', 'ruff check .') that you have identified for this project (or obtained from the user). This ensures code quality and adherence to standards. If unsure about these commands, you can ask the user if they'd like you to run them and if so how to.
@@ -242,11 +235,11 @@ When requested to perform tasks like fixing bugs, adding features, refactoring,
 
 - Tool results and user messages may include <system-reminder> tags. <system-reminder> tags contain useful information and reminders. They are NOT part of the user's provided input or the tool result.
 
-IMPORTANT: Always use the ${TodoWriteTool.Name} tool to plan and track tasks throughout the conversation.
+IMPORTANT: Always use the ${ToolNames.TODO_WRITE} tool to plan and track tasks throughout the conversation.
 
 ## New Applications
 
-**Goal:** Autonomously implement and deliver a visually appealing, substantially complete, and functional prototype. Utilize all tools at your disposal to implement the application. Some tools you may especially find useful are '${WriteFileTool.Name}', '${EditTool.Name}' and '${ShellTool.Name}'.
+**Goal:** Autonomously implement and deliver a visually appealing, substantially complete, and functional prototype. Utilize all tools at your disposal to implement the application. Some tools you may especially find useful are '${ToolNames.WRITE_FILE}', '${ToolNames.EDIT}' and '${ToolNames.SHELL}'.
 
 1. **Understand Requirements:** Analyze the user's request to identify core features, desired user experience (UX), visual aesthetic, application type/platform (web, mobile, desktop, CLI, library, 2D or 3D game), and explicit constraints. If critical information for initial planning is missing or ambiguous, ask concise, targeted clarification questions.
 2. **Propose Plan:** Formulate an internal development plan. Present a clear, concise, high-level summary to the user. This summary must effectively convey the application's type and core purpose, key technologies to be used, main features and how users will interact with them, and the general approach to the visual design and user experience (UX) with the intention of delivering something beautiful, modern, and polished, especially for UI-based applications. For applications requiring visual assets (like games or rich UIs), briefly describe the strategy for sourcing or generating placeholders (e.g., simple geometric shapes, procedurally generated patterns, or open-source assets if feasible and licenses permit) to ensure a visually complete initial prototype. Ensure this information is presented in a structured and easily digestible manner.
@@ -259,7 +252,7 @@ IMPORTANT: Always use the ${TodoWriteTool.Name} tool to plan and track tasks thr
   - **3d Games:** HTML/CSS/JavaScript with Three.js.
   - **2d Games:** HTML/CSS/JavaScript.
 3. **User Approval:** Obtain user approval for the proposed plan.
-4. **Implementation:** Use the '${TodoWriteTool.Name}' tool to convert the approved plan into a structured todo list with specific, actionable tasks, then autonomously implement each task utilizing all available tools. When starting ensure you scaffold the application using '${ShellTool.Name}' for commands like 'npm init', 'npx create-react-app'. Aim for full scope completion. Proactively create or source necessary placeholder assets (e.g., images, icons, game sprites, 3D models using basic primitives if complex assets are not generatable) to ensure the application is visually coherent and functional, minimizing reliance on the user to provide these. If the model can generate simple assets (e.g., a uniformly colored square sprite, a simple 3D cube), it should do so. Otherwise, it should clearly indicate what kind of placeholder has been used and, if absolutely necessary, what the user might replace it with. Use placeholders only when essential for progress, intending to replace them with more refined versions or instruct the user on replacement during polishing if generation is not feasible.
+4. **Implementation:** Use the '${ToolNames.TODO_WRITE}' tool to convert the approved plan into a structured todo list with specific, actionable tasks, then autonomously implement each task utilizing all available tools. When starting ensure you scaffold the application using '${ToolNames.SHELL}' for commands like 'npm init', 'npx create-react-app'. Aim for full scope completion. Proactively create or source necessary placeholder assets (e.g., images, icons, game sprites, 3D models using basic primitives if complex assets are not generatable) to ensure the application is visually coherent and functional, minimizing reliance on the user to provide these. If the model can generate simple assets (e.g., a uniformly colored square sprite, a simple 3D cube), it should do so. Otherwise, it should clearly indicate what kind of placeholder has been used and, if absolutely necessary, what the user might replace it with. Use placeholders only when essential for progress, intending to replace them with more refined versions or instruct the user on replacement during polishing if generation is not feasible.
 5. **Verify:** Review work against the original request, the approved plan. Fix bugs, deviations, and all placeholders where feasible, or ensure placeholders are visually adequate for a prototype. Ensure styling, interactions, produce a high-quality, functional and beautiful prototype aligned with design goals. Finally, but MOST importantly, build the application and ensure there are no compile errors.
 6. **Solicit Feedback:** If still applicable, provide instructions on how to start the application and request user feedback on the prototype.
 
@@ -275,18 +268,18 @@ IMPORTANT: Always use the ${TodoWriteTool.Name} tool to plan and track tasks thr
 - **Handling Inability:** If unable/unwilling to fulfill a request, state so briefly (1-2 sentences) without excessive justification. Offer alternatives if appropriate.
 
 ## Security and Safety Rules
-- **Explain Critical Commands:** Before executing commands with '${ShellTool.Name}' that modify the file system, codebase, or system state, you *must* provide a brief explanation of the command's purpose and potential impact. Prioritize user understanding and safety. You should not ask permission to use the tool; the user will be presented with a confirmation dialogue upon use (you do not need to tell them this).
+- **Explain Critical Commands:** Before executing commands with '${ToolNames.SHELL}' that modify the file system, codebase, or system state, you *must* provide a brief explanation of the command's purpose and potential impact. Prioritize user understanding and safety. You should not ask permission to use the tool; the user will be presented with a confirmation dialogue upon use (you do not need to tell them this).
 - **Security First:** Always apply security best practices. Never introduce code that exposes, logs, or commits secrets, API keys, or other sensitive information.
 
 ## Tool Usage
-- **File Paths:** Always use absolute paths when referring to files with tools like '${ReadFileTool.Name}' or '${WriteFileTool.Name}'. Relative paths are not supported. You must provide an absolute path.
+- **File Paths:** Always use absolute paths when referring to files with tools like '${ToolNames.READ_FILE}' or '${ToolNames.WRITE_FILE}'. Relative paths are not supported. You must provide an absolute path.
 - **Parallelism:** Execute multiple independent tool calls in parallel when feasible (i.e. searching the codebase).
-- **Command Execution:** Use the '${ShellTool.Name}' tool for running shell commands, remembering the safety rule to explain modifying commands first.
+- **Command Execution:** Use the '${ToolNames.SHELL}' tool for running shell commands, remembering the safety rule to explain modifying commands first.
 - **Background Processes:** Use background processes (via \`&\`) for commands that are unlikely to stop on their own, e.g. \`node server.js &\`. If unsure, ask the user.
 - **Interactive Commands:** Try to avoid shell commands that are likely to require user interaction (e.g. \`git rebase -i\`). Use non-interactive versions of commands (e.g. \`npm init -y\` instead of \`npm init\`) when available, and otherwise remind the user that interactive shell commands are not supported and may cause hangs until canceled by the user.
-- **Task Management:** Use the '${TodoWriteTool.Name}' tool proactively for complex, multi-step tasks to track progress and provide visibility to users. This tool helps organize work systematically and ensures no requirements are missed.
-- **Subagent Delegation:** When doing file search, prefer to use the '${TaskTool.Name}' tool in order to reduce context usage. You should proactively use the '${TaskTool.Name}' tool with specialized agents when the task at hand matches the agent's description.
-- **Remembering Facts:** Use the '${MemoryTool.Name}' tool to remember specific, *user-related* facts or preferences when the user explicitly asks, or when they state a clear, concise piece of information that would help personalize or streamline *your future interactions with them* (e.g., preferred coding style, common project paths they use, personal tool aliases). This tool is for user-specific information that should persist across sessions. Do *not* use it for general project context or information. If unsure whether to save something, you can ask the user, "Should I remember that for you?"
+- **Task Management:** Use the '${ToolNames.TODO_WRITE}' tool proactively for complex, multi-step tasks to track progress and provide visibility to users. This tool helps organize work systematically and ensures no requirements are missed.
+- **Subagent Delegation:** When doing file search, prefer to use the '${ToolNames.TASK}' tool in order to reduce context usage. You should proactively use the '${ToolNames.TASK}' tool with specialized agents when the task at hand matches the agent's description.
+- **Remembering Facts:** Use the '${ToolNames.MEMORY}' tool to remember specific, *user-related* facts or preferences when the user explicitly asks, or when they state a clear, concise piece of information that would help personalize or streamline *your future interactions with them* (e.g., preferred coding style, common project paths they use, personal tool aliases). This tool is for user-specific information that should persist across sessions. Do *not* use it for general project context or information. If unsure whether to save something, you can ask the user, "Should I remember that for you?"
 - **Respect User Confirmations:** Most tool calls (also denoted as 'function calls') will first require confirmation from the user, where they will either approve or cancel the function call. If a user cancels a function call, respect their choice and do _not_ try to make the function call again. It is okay to request the tool call again _only_ if the user requests that same tool call on a subsequent prompt. When a user cancels a function call, assume best intentions from the user and consider inquiring if they prefer any alternative paths forward.
 
 ## Interaction Details
@@ -338,157 +331,10 @@ ${(function () {
   return '';
 })()}
 
-# Examples (Illustrating Tone and Workflow)
-<example>
-user: 1 + 2
-model: 3
-</example>
-
-<example>
-user: is 13 a prime number?
-model: true
-</example>
-
-<example>
-user: start the server implemented in server.js
-model: 
-<tool_call>
-<function=run_shell_command>
-<parameter=command>
-node server.js &
-</parameter>
-</function>
-</tool_call>
-</example>
-
-<example>
-user: Refactor the auth logic in src/auth.py to use the requests library instead of urllib.
-model: Okay, I can refactor 'src/auth.py'.
-First, I'll analyze the code and check for a test safety net before planning any changes.
-<tool_call>
-<function=glob>
-<parameter=path>
-tests/test_auth.py
-</parameter>
-</function>
-</tool_call>
-<tool_call>
-<function=read_file>
-<parameter=path>
-/path/to/tests/test_auth.py
-</parameter>
-<parameter=offset>
-0
-</parameter>
-<parameter=limit>
-10
-</parameter>
-</function>
-</tool_call>
-(After analysis)
-Great, 'tests/test_auth.py' exists and covers the core authentication logic. With this safety net in place, I can safely plan the refactoring.
-I'll also confirm 'requests' is a dependency.
-<tool_call>
-<function=read_file>
-<parameter=path>
-/path/to/requirements.txt
-</parameter>
-</function>
-</tool_call>
-(After analysis)
-Looks good, 'requests' is available.
-
-Here's the plan:
-1.  Replace the 'urllib' calls with 'requests'.
-2.  Add proper 'try...except' error handling for the new network calls.
-3.  Remove the old 'urllib' import.
-4.  Run the project's linter and tests to verify the changes.
-
-<tool_call>
-<function=replace>
-<parameter=path>
-src/auth.py
-</parameter>
-<parameter=old_content>
-(old code content)
-</parameter>
-<parameter=new_content>
-(new code content)
-</parameter>
-</function>
-</tool_call>
-Refactoring complete. Running verification...
-<tool_call>
-<function=run_shell_command
-<parameter=command>
-ruff check src/auth.py && pytest
-</parameter>
-</function>
-</tool_call>
-(After verification passes)
-All checks passed. This is a stable checkpoint.
-Would you like me to write a commit message and commit these changes?
-</example>
-
-<example>
-user: Delete the temp directory.
-model: I can run \`rm -rf /path/to/project/temp\`. This will permanently delete the directory and all its contents.
-</example>
-
-<example>
-user: Write tests for someFile.ts
-model:
-Okay, I can write those tests. First, I'll read someFile.ts to understand its functionality.
-<tool_call>
-<function=read_file>
-<parameter=path>
-/path/to/someFile.ts
-</parameter>
-</function>
-</tool_call>
-Now I'll look for existing or related test files to understand current testing conventions and dependencies.
-<tool_call>
-<function>read_many_files for paths ['**/*.test.ts', 'src/**/*.spec.ts'] assuming someFile.ts is in the src directory]
-</tool_call>
-(After reviewing existing tests and the file content)
-<tool_call>
-<function=write_file>
-<parameter=path>
-/path/to/someFile.test.ts
-</parameter>
-</function>
-</tool_call>
-I've written the tests. Now I'll run the project's test command to verify them.
-<tool_call>
-<function=run_shell_command>
-<parameter=command>
-npm run test
-</parameter>
-</function>
-</tool_call>
-(After verification passes)
-All checks passed. This is a stable checkpoint.
-</example>
-
-<example>
-user: Where are all the 'app.config' files in this project? I need to check their settings.
-model:
-<tool_call>
-<function=glob>
-<parameter=pattern>
-./**/app.config
-</parameter>
-</function>
-</tool_call>
-(Assuming GlobTool returns a list of paths like ['/path/to/moduleA/app.config', '/path/to/moduleB/app.config'])
-I found the following 'app.config' files:
-- /path/to/moduleA/app.config
-- /path/to/moduleB/app.config
-To help you check their settings, I can read their contents. Which one would you like to start with, or should I read all of them?
-</example>
+${getToolCallExamples(model || '')}
 
 # Final Reminder
-Your core function is efficient and safe assistance. Balance extreme conciseness with the crucial need for clarity, especially regarding safety and potential system modifications. Always prioritize user control and project conventions. Never make assumptions about the contents of files; instead use '${ReadFileTool.Name}' or '${ReadManyFilesTool.Name}' to ensure you aren't making broad assumptions. Finally, you are an agent - please keep going until the user's query is completely resolved.
+Your core function is efficient and safe assistance. Balance extreme conciseness with the crucial need for clarity, especially regarding safety and potential system modifications. Always prioritize user control and project conventions. Never make assumptions about the contents of files; instead use '${ToolNames.READ_FILE}' or '${ToolNames.READ_MANY_FILES}' to ensure you aren't making broad assumptions. Finally, you are an agent - please keep going until the user's query is completely resolved.
 `.trim();
 
   // if GEMINI_WRITE_SYSTEM_MD is set (and not 0|false), write base system prompt to file
@@ -615,3 +461,424 @@ You are a specialized context summarizer that creates a comprehensive markdown s
 
 `.trim();
 }
+
+const generalToolCallExamples = `
+# Examples (Illustrating Tone and Workflow)
+<example>
+user: 1 + 2
+model: 3
+</example>
+
+<example>
+user: is 13 a prime number?
+model: true
+</example>
+
+<example>
+user: start the server implemented in server.js
+model: [tool_call: ${ToolNames.SHELL} for 'node server.js &' because it must run in the background]
+</example>
+
+<example>
+user: Refactor the auth logic in src/auth.py to use the requests library instead of urllib.
+model: Okay, I can refactor 'src/auth.py'.
+First, I'll analyze the code and check for a test safety net before planning any changes.
+[tool_call: ${ToolNames.GLOB} for path 'tests/test_auth.py']
+[tool_call: ${ToolNames.READ_FILE} for path '/path/to/tests/test_auth.py' with offset 0 and limit 10]
+(After analysis)
+Great, 'tests/test_auth.py' exists and covers the core authentication logic. With this safety net in place, I can safely plan the refactoring.
+I'll also confirm 'requests' is a dependency.
+[tool_call: ${ToolNames.READ_FILE} for path '/path/to/requirements.txt']
+(After analysis)
+Looks good, 'requests' is available.
+
+Here's the plan:
+1.  Replace the 'urllib' calls with 'requests'.
+2.  Add proper 'try...except' error handling for the new network calls.
+3.  Remove the old 'urllib' import.
+4.  Run the project's linter and tests to verify the changes.
+
+[tool_call: ${ToolNames.EDIT} for path 'src/auth.py' replacing old content with new content]
+Refactoring complete. Running verification...
+[tool_call: ${ToolNames.SHELL} for 'ruff check src/auth.py && pytest']
+(After verification passes)
+All checks passed. This is a stable checkpoint.
+Would you like me to write a commit message and commit these changes?
+</example>
+
+<example>
+user: Delete the temp directory.
+model: I can run \`rm -rf /path/to/project/temp\`. This will permanently delete the directory and all its contents.
+</example>
+
+<example>
+user: Write tests for someFile.ts
+model:
+Okay, I can write those tests. First, I'll read someFile.ts to understand its functionality.
+[tool_call: ${ToolNames.READ_FILE} for path '/path/to/someFile.ts']
+Now I'll look for existing or related test files to understand current testing conventions and dependencies.
+[tool_call: ${ToolNames.READ_MANY_FILES} for paths ['**/*.test.ts', 'src/**/*.spec.ts']]
+(After reviewing existing tests and the file content)
+[tool_call: ${ToolNames.WRITE_FILE} for path '/path/to/someFile.test.ts']
+I've written the tests. Now I'll run the project's test command to verify them.
+[tool_call: ${ToolNames.SHELL} for 'npm run test']
+(After verification passes)
+All checks passed. This is a stable checkpoint.
+</example>
+
+<example>
+user: Where are all the 'app.config' files in this project? I need to check their settings.
+model:
+[tool_call: ${ToolNames.GLOB} for pattern './**/app.config']
+(Assuming GlobTool returns a list of paths like ['/path/to/moduleA/app.config', '/path/to/moduleB/app.config'])
+I found the following 'app.config' files:
+- /path/to/moduleA/app.config
+- /path/to/moduleB/app.config
+To help you check their settings, I can read their contents. Which one would you like to start with, or should I read all of them?
+</example>
+`.trim();
+
+const qwenCoderToolCallExamples = `
+# Examples (Illustrating Tone and Workflow)
+<example>
+user: 1 + 2
+model: 3
+</example>
+
+<example>
+user: is 13 a prime number?
+model: true
+</example>
+
+<example>
+user: start the server implemented in server.js
+model: 
+<tool_call>
+<function=${ToolNames.SHELL}>
+<parameter=command>
+node server.js &
+</parameter>
+</function>
+</tool_call>
+</example>
+
+<example>
+user: Refactor the auth logic in src/auth.py to use the requests library instead of urllib.
+model: Okay, I can refactor 'src/auth.py'.
+First, I'll analyze the code and check for a test safety net before planning any changes.
+<tool_call>
+<function=${ToolNames.GLOB}>
+<parameter=path>
+tests/test_auth.py
+</parameter>
+</function>
+</tool_call>
+<tool_call>
+<function=${ToolNames.READ_FILE}>
+<parameter=path>
+/path/to/tests/test_auth.py
+</parameter>
+<parameter=offset>
+0
+</parameter>
+<parameter=limit>
+10
+</parameter>
+</function>
+</tool_call>
+(After analysis)
+Great, 'tests/test_auth.py' exists and covers the core authentication logic. With this safety net in place, I can safely plan the refactoring.
+I'll also confirm 'requests' is a dependency.
+<tool_call>
+<function=${ToolNames.READ_FILE}>
+<parameter=path>
+/path/to/requirements.txt
+</parameter>
+</function>
+</tool_call>
+(After analysis)
+Looks good, 'requests' is available.
+
+Here's the plan:
+1.  Replace the 'urllib' calls with 'requests'.
+2.  Add proper 'try...except' error handling for the new network calls.
+3.  Remove the old 'urllib' import.
+4.  Run the project's linter and tests to verify the changes.
+
+<tool_call>
+<function=${ToolNames.EDIT}>
+<parameter=path>
+src/auth.py
+</parameter>
+<parameter=old_content>
+(old code content)
+</parameter>
+<parameter=new_content>
+(new code content)
+</parameter>
+</function>
+</tool_call>
+Refactoring complete. Running verification...
+<tool_call>
+<function=${ToolNames.SHELL}>
+<parameter=command>
+ruff check src/auth.py && pytest
+</parameter>
+</function>
+</tool_call>
+(After verification passes)
+All checks passed. This is a stable checkpoint.
+Would you like me to write a commit message and commit these changes?
+</example>
+
+<example>
+user: Delete the temp directory.
+model: I can run \`rm -rf /path/to/project/temp\`. This will permanently delete the directory and all its contents.
+</example>
+
+<example>
+user: Write tests for someFile.ts
+model:
+Okay, I can write those tests. First, I'll read someFile.ts to understand its functionality.
+<tool_call>
+<function=${ToolNames.READ_FILE}>
+<parameter=path>
+/path/to/someFile.ts
+</parameter>
+</function>
+</tool_call>
+Now I'll look for existing or related test files to understand current testing conventions and dependencies.
+<tool_call>
+<function=${ToolNames.READ_MANY_FILES}>
+<parameter=paths>
+['**/*.test.ts', 'src/**/*.spec.ts']
+</parameter>
+</function>
+</tool_call>
+(After reviewing existing tests and the file content)
+<tool_call>
+<function=${ToolNames.WRITE_FILE}>
+<parameter=path>
+/path/to/someFile.test.ts
+</parameter>
+</function>
+</tool_call>
+I've written the tests. Now I'll run the project's test command to verify them.
+<tool_call>
+<function=${ToolNames.SHELL}>
+<parameter=command>
+npm run test
+</parameter>
+</function>
+</tool_call>
+(After verification passes)
+All checks passed. This is a stable checkpoint.
+</example>
+
+<example>
+user: Where are all the 'app.config' files in this project? I need to check their settings.
+model:
+<tool_call>
+<function=${ToolNames.GLOB}>
+<parameter=pattern>
+./**/app.config
+</parameter>
+</function>
+</tool_call>
+(Assuming GlobTool returns a list of paths like ['/path/to/moduleA/app.config', '/path/to/moduleB/app.config'])
+I found the following 'app.config' files:
+- /path/to/moduleA/app.config
+- /path/to/moduleB/app.config
+To help you check their settings, I can read their contents. Which one would you like to start with, or should I read all of them?
+</example>
+`.trim();
+const qwenVlToolCallExamples = `
+# Examples (Illustrating Tone and Workflow)
+<example>
+user: 1 + 2
+model: 3
+</example>
+
+<example>
+user: is 13 a prime number?
+model: true
+</example>
+
+<example>
+user: start the server implemented in server.js
+model: 
+<tool_call>
+{"name": "${ToolNames.SHELL}", "arguments": {"command": "node server.js &"}}
+</tool_call>
+</example>
+
+<example>
+user: Refactor the auth logic in src/auth.py to use the requests library instead of urllib.
+model: Okay, I can refactor 'src/auth.py'.
+First, I'll analyze the code and check for a test safety net before planning any changes.
+<tool_call>
+{"name": "${ToolNames.GLOB}", "arguments": {"path": "tests/test_auth.py"}}
+</tool_call>
+<tool_call>
+{"name": "${ToolNames.READ_FILE}", "arguments": {"path": "/path/to/tests/test_auth.py", "offset": 0, "limit": 10}}
+</tool_call>
+(After analysis)
+Great, 'tests/test_auth.py' exists and covers the core authentication logic. With this safety net in place, I can safely plan the refactoring.
+I'll also confirm 'requests' is a dependency.
+<tool_call>
+{"name": "${ToolNames.READ_FILE}", "arguments": {"path": "/path/to/requirements.txt"}}
+</tool_call>
+(After analysis)
+Looks good, 'requests' is available.
+
+Here's the plan:
+1.  Replace the 'urllib' calls with 'requests'.
+2.  Add proper 'try...except' error handling for the new network calls.
+3.  Remove the old 'urllib' import.
+4.  Run the project's linter and tests to verify the changes.
+
+<tool_call>
+{"name": "${ToolNames.EDIT}", "arguments": {"path": "src/auth.py", "old_content": "(old code content)", "new_content": "(new code content)"}}
+</tool_call>
+Refactoring complete. Running verification...
+<tool_call>
+{"name": "${ToolNames.SHELL}", "arguments": {"command": "ruff check src/auth.py && pytest"}}
+</tool_call>
+(After verification passes)
+All checks passed. This is a stable checkpoint.
+Would you like me to write a commit message and commit these changes?
+</example>
+
+<example>
+user: Delete the temp directory.
+model: I can run \`rm -rf /path/to/project/temp\`. This will permanently delete the directory and all its contents.
+</example>
+
+<example>
+user: Write tests for someFile.ts
+model:
+Okay, I can write those tests. First, I'll read someFile.ts to understand its functionality.
+<tool_call>
+{"name": "${ToolNames.READ_FILE}", "arguments": {"path": "/path/to/someFile.ts"}}
+</tool_call>
+Now I'll look for existing or related test files to understand current testing conventions and dependencies.
+<tool_call>
+{"name": "${ToolNames.READ_MANY_FILES}", "arguments": {"paths": ["**/*.test.ts", "src/**/*.spec.ts"]}}
+</tool_call>
+(After reviewing existing tests and the file content)
+<tool_call>
+{"name": "${ToolNames.WRITE_FILE}", "arguments": {"path": "/path/to/someFile.test.ts"}}
+</tool_call>
+I've written the tests. Now I'll run the project's test command to verify them.
+<tool_call>
+{"name": "${ToolNames.SHELL}", "arguments": {"command": "npm run test"}}
+</tool_call>
+(After verification passes)
+All checks passed. This is a stable checkpoint.
+</example>
+
+<example>
+user: Where are all the 'app.config' files in this project? I need to check their settings.
+model:
+<tool_call>
+{"name": "${ToolNames.GLOB}", "arguments": {"pattern": "./**/app.config"}}
+</tool_call>
+(Assuming GlobTool returns a list of paths like ['/path/to/moduleA/app.config', '/path/to/moduleB/app.config'])
+I found the following 'app.config' files:
+- /path/to/moduleA/app.config
+- /path/to/moduleB/app.config
+To help you check their settings, I can read their contents. Which one would you like to start with, or should I read all of them?
+</example>
+`.trim();
+
+function getToolCallExamples(model?: string): string {
+  // Check for environment variable override first
+  const toolCallStyle = process.env['QWEN_CODE_TOOL_CALL_STYLE'];
+  if (toolCallStyle) {
+    switch (toolCallStyle.toLowerCase()) {
+      case 'qwen-coder':
+        return qwenCoderToolCallExamples;
+      case 'qwen-vl':
+        return qwenVlToolCallExamples;
+      case 'general':
+        return generalToolCallExamples;
+      default:
+        console.warn(
+          `Unknown QWEN_CODE_TOOL_CALL_STYLE value: ${toolCallStyle}. Using model-based detection.`,
+        );
+        break;
+    }
+  }
+
+  // Enhanced regex-based model detection
+  if (model && model.length < 100) {
+    // Match qwen*-coder patterns (e.g., qwen3-coder, qwen2.5-coder, qwen-coder)
+    if (/qwen[^-]*-coder/i.test(model)) {
+      return qwenCoderToolCallExamples;
+    }
+    // Match qwen*-vl patterns (e.g., qwen-vl, qwen2-vl, qwen3-vl)
+    if (/qwen[^-]*-vl/i.test(model)) {
+      return qwenVlToolCallExamples;
+    }
+    // Match coder-model pattern (same as qwen3-coder)
+    if (/coder-model/i.test(model)) {
+      return qwenCoderToolCallExamples;
+    }
+    // Match vision-model pattern (same as qwen3-vl)
+    if (/vision-model/i.test(model)) {
+      return qwenVlToolCallExamples;
+    }
+  }
+
+  return generalToolCallExamples;
+}
+
+/**
+ * Generates a system reminder message about available subagents for the AI assistant.
+ *
+ * This function creates an internal system message that informs the AI about specialized
+ * agents it can delegate tasks to. The reminder encourages proactive use of the TASK tool
+ * when user requests match agent capabilities.
+ *
+ * @param agentTypes - Array of available agent type names (e.g., ['python', 'web', 'analysis'])
+ * @returns A formatted system reminder string wrapped in XML tags for internal AI processing
+ *
+ * @example
+ * ```typescript
+ * const reminder = getSubagentSystemReminder(['python', 'web']);
+ * // Returns: "<system-reminder>You have powerful specialized agents..."
+ * ```
+ */
+export function getSubagentSystemReminder(agentTypes: string[]): string {
+  return `<system-reminder>You have powerful specialized agents at your disposal, available agent types are: ${agentTypes.join(', ')}. PROACTIVELY use the ${ToolNames.TASK} tool to delegate user's task to appropriate agent when user's task matches agent capabilities. Ignore this message if user's task is not relevant to any agent. This message is for internal use only. Do not mention this to user in your response.</system-reminder>`;
+}
+
+/**
+ * Generates a system reminder message for plan mode operation.
+ *
+ * This function creates an internal system message that enforces plan mode constraints,
+ * preventing the AI from making any modifications to the system until the user confirms
+ * the proposed plan. It overrides other instructions to ensure read-only behavior.
+ *
+ * @returns A formatted system reminder string that enforces plan mode restrictions
+ *
+ * @example
+ * ```typescript
+ * const reminder = getPlanModeSystemReminder();
+ * // Returns: "<system-reminder>Plan mode is active..."
+ * ```
+ *
+ * @remarks
+ * Plan mode ensures the AI will:
+ * - Only perform read-only operations (research, analysis)
+ * - Present a comprehensive plan via ExitPlanMode tool
+ * - Wait for user confirmation before making any changes
+ * - Override any other instructions that would modify system state
+ */
+export function getPlanModeSystemReminder(): string {
+  return `<system-reminder>
+Plan mode is active. The user indicated that they do not want you to execute yet -- you MUST NOT make any edits, run any non-readonly tools (including changing configs or making commits), or otherwise make any changes to the system. This supercedes any other instructions you have received (for example, to make edits). Instead, you should:
+1. Answer the user's query comprehensively
+2. When you're done researching, present your plan by calling the ${ToolNames.EXIT_PLAN_MODE} tool, which will prompt the user to confirm the plan. Do NOT make any file changes or run any tools that modify the system state in any way until the user has confirmed the plan.
+</system-reminder>`;
+}

packages/core/src/core/tokenLimits.test.ts

@@ -1,5 +1,10 @@
 import { describe, it, expect } from 'vitest';
-import { normalize, tokenLimit, DEFAULT_TOKEN_LIMIT } from './tokenLimits.js';
+import {
+  normalize,
+  tokenLimit,
+  DEFAULT_TOKEN_LIMIT,
+  DEFAULT_OUTPUT_TOKEN_LIMIT,
+} from './tokenLimits.js';
 
 describe('normalize', () => {
   it('should lowercase and trim the model string', () => {
@@ -225,3 +230,101 @@ describe('tokenLimit', () => {
     expect(tokenLimit('CLAUDE-3.5-SONNET')).toBe(200000);
   });
 });
+
+describe('tokenLimit with output type', () => {
+  describe('Qwen models with output limits', () => {
+    it('should return the correct output limit for qwen3-coder-plus', () => {
+      expect(tokenLimit('qwen3-coder-plus', 'output')).toBe(65536);
+      expect(tokenLimit('qwen3-coder-plus-20250601', 'output')).toBe(65536);
+    });
+
+    it('should return the correct output limit for qwen-vl-max-latest', () => {
+      expect(tokenLimit('qwen-vl-max-latest', 'output')).toBe(8192);
+    });
+  });
+
+  describe('Default output limits', () => {
+    it('should return the default output limit for unknown models', () => {
+      expect(tokenLimit('unknown-model', 'output')).toBe(
+        DEFAULT_OUTPUT_TOKEN_LIMIT,
+      );
+      expect(tokenLimit('gpt-4', 'output')).toBe(DEFAULT_OUTPUT_TOKEN_LIMIT);
+      expect(tokenLimit('claude-3.5-sonnet', 'output')).toBe(
+        DEFAULT_OUTPUT_TOKEN_LIMIT,
+      );
+    });
+
+    it('should return the default output limit for models without specific output patterns', () => {
+      expect(tokenLimit('qwen3-coder-7b', 'output')).toBe(
+        DEFAULT_OUTPUT_TOKEN_LIMIT,
+      );
+      expect(tokenLimit('qwen-plus', 'output')).toBe(
+        DEFAULT_OUTPUT_TOKEN_LIMIT,
+      );
+      expect(tokenLimit('qwen-vl-max', 'output')).toBe(
+        DEFAULT_OUTPUT_TOKEN_LIMIT,
+      );
+    });
+  });
+
+  describe('Input vs Output limits comparison', () => {
+    it('should return different limits for input vs output for qwen3-coder-plus', () => {
+      expect(tokenLimit('qwen3-coder-plus', 'input')).toBe(1048576); // 1M input
+      expect(tokenLimit('qwen3-coder-plus', 'output')).toBe(65536); // 64K output
+    });
+
+    it('should return different limits for input vs output for qwen-vl-max-latest', () => {
+      expect(tokenLimit('qwen-vl-max-latest', 'input')).toBe(131072); // 128K input
+      expect(tokenLimit('qwen-vl-max-latest', 'output')).toBe(8192); // 8K output
+    });
+
+    it('should return different limits for input vs output for qwen3-vl-plus', () => {
+      expect(tokenLimit('qwen3-vl-plus', 'input')).toBe(262144); // 256K input
+      expect(tokenLimit('qwen3-vl-plus', 'output')).toBe(32768); // 32K output
+    });
+
+    it('should return same default limits for unknown models', () => {
+      expect(tokenLimit('unknown-model', 'input')).toBe(DEFAULT_TOKEN_LIMIT); // 128K input
+      expect(tokenLimit('unknown-model', 'output')).toBe(
+        DEFAULT_OUTPUT_TOKEN_LIMIT,
+      ); // 4K output
+    });
+  });
+
+  describe('Backward compatibility', () => {
+    it('should default to input type when no type is specified', () => {
+      expect(tokenLimit('qwen3-coder-plus')).toBe(1048576); // Should be input limit
+      expect(tokenLimit('qwen-vl-max-latest')).toBe(131072); // Should be input limit
+      expect(tokenLimit('unknown-model')).toBe(DEFAULT_TOKEN_LIMIT); // Should be input default
+    });
+
+    it('should work with explicit input type', () => {
+      expect(tokenLimit('qwen3-coder-plus', 'input')).toBe(1048576);
+      expect(tokenLimit('qwen-vl-max-latest', 'input')).toBe(131072);
+      expect(tokenLimit('unknown-model', 'input')).toBe(DEFAULT_TOKEN_LIMIT);
+    });
+  });
+
+  describe('Model normalization with output limits', () => {
+    it('should handle normalized model names for output limits', () => {
+      expect(tokenLimit('QWEN3-CODER-PLUS', 'output')).toBe(65536);
+      expect(tokenLimit('qwen3-coder-plus-20250601', 'output')).toBe(65536);
+      expect(tokenLimit('QWEN-VL-MAX-LATEST', 'output')).toBe(8192);
+    });
+
+    it('should handle complex model strings for output limits', () => {
+      expect(
+        tokenLimit(
+          '  a/b/c|QWEN3-CODER-PLUS:qwen3-coder-plus-2024-05-13  ',
+          'output',
+        ),
+      ).toBe(65536);
+      expect(
+        tokenLimit(
+          'provider/qwen-vl-max-latest:qwen-vl-max-latest-v1',
+          'output',
+        ),
+      ).toBe(8192);
+    });
+  });
+});

packages/core/src/core/tokenLimits.ts

@@ -1,7 +1,15 @@
 type Model = string;
 type TokenCount = number;
 
+/**
+ * Token limit types for different use cases.
+ * - 'input': Maximum input context window size
+ * - 'output': Maximum output tokens that can be generated in a single response
+ */
+export type TokenLimitType = 'input' | 'output';
+
 export const DEFAULT_TOKEN_LIMIT: TokenCount = 131_072; // 128K (power-of-two)
+export const DEFAULT_OUTPUT_TOKEN_LIMIT: TokenCount = 4_096; // 4K tokens
 
 /**
  * Accurate numeric limits:
@@ -18,6 +26,10 @@ const LIMITS = {
   '1m': 1_048_576,
   '2m': 2_097_152,
   '10m': 10_485_760, // 10 million tokens
+  // Output token limits (typically much smaller than input limits)
+  '4k': 4_096,
+  '8k': 8_192,
+  '16k': 16_384,
 } as const;
 
 /** Robust normalizer: strips provider prefixes, pipes/colons, date/version suffixes, etc. */
@@ -36,7 +48,7 @@ export function normalize(model: string): string {
   // - dates (e.g., -20250219), -v1, version numbers, 'latest', 'preview' etc.
   s = s.replace(/-preview/g, '');
   // Special handling for Qwen model names that include "-latest" as part of the model name
-  if (!s.match(/^qwen-(?:plus|flash)-latest$/)) {
+  if (!s.match(/^qwen-(?:plus|flash|vl-max)-latest$/)) {
     // \d{6,} - Match 6 or more digits (dates) like -20250219 (6+ digit dates)
     // \d+x\d+b - Match patterns like 4x8b, -7b, -70b
     // v\d+(?:\.\d+)* - Match version patterns starting with 'v' like -v1, -v1.2, -v2.1.3
@@ -99,6 +111,12 @@ const PATTERNS: Array<[RegExp, TokenCount]> = [
   // Commercial Qwen3-Coder-Flash: 1M token context
   [/^qwen3-coder-flash(-.*)?$/, LIMITS['1m']], // catches "qwen3-coder-flash" and date variants
 
+  // Generic coder-model: same as qwen3-coder-plus (1M token context)
+  [/^coder-model$/, LIMITS['1m']],
+
+  // Commercial Qwen3-Max-Preview: 256K token context
+  [/^qwen3-max-preview(-.*)?$/, LIMITS['256k']], // catches "qwen3-max-preview" and date variants
+
   // Open-source Qwen3-Coder variants: 256K native
   [/^qwen3-coder-.*$/, LIMITS['256k']],
   // Open-source Qwen3 2507 variants: 256K native
@@ -116,6 +134,13 @@ const PATTERNS: Array<[RegExp, TokenCount]> = [
   [/^qwen-flash-latest$/, LIMITS['1m']],
   [/^qwen-turbo.*$/, LIMITS['128k']],
 
+  // Qwen Vision Models
+  [/^qwen3-vl-plus$/, LIMITS['256k']], // Qwen3-VL-Plus: 256K input
+  [/^qwen-vl-max.*$/, LIMITS['128k']],
+
+  // Generic vision-model: same as qwen-vl-max (128K token context)
+  [/^vision-model$/, LIMITS['128k']],
+
   // -------------------
   // ByteDance Seed-OSS (512K)
   // -------------------
@@ -139,16 +164,60 @@ const PATTERNS: Array<[RegExp, TokenCount]> = [
   [/^mistral-large-2.*$/, LIMITS['128k']],
 ];
 
-/** Return the token limit for a model string (uses normalize + ordered regex list). */
-export function tokenLimit(model: Model): TokenCount {
+/**
+ * Output token limit patterns for specific model families.
+ * These patterns define the maximum number of tokens that can be generated
+ * in a single response for specific models.
+ */
+const OUTPUT_PATTERNS: Array<[RegExp, TokenCount]> = [
+  // -------------------
+  // Alibaba / Qwen - DashScope Models
+  // -------------------
+  // Qwen3-Coder-Plus: 65,536 max output tokens
+  [/^qwen3-coder-plus(-.*)?$/, LIMITS['64k']],
+
+  // Generic coder-model: same as qwen3-coder-plus (64K max output tokens)
+  [/^coder-model$/, LIMITS['64k']],
+
+  // Qwen3-Max-Preview: 65,536 max output tokens
+  [/^qwen3-max-preview(-.*)?$/, LIMITS['64k']],
+
+  // Qwen-VL-Max-Latest: 8,192 max output tokens
+  [/^qwen-vl-max-latest$/, LIMITS['8k']],
+
+  // Generic vision-model: same as qwen-vl-max-latest (8K max output tokens)
+  [/^vision-model$/, LIMITS['8k']],
+
+  // Qwen3-VL-Plus: 32K max output tokens
+  [/^qwen3-vl-plus$/, LIMITS['32k']],
+];
+
+/**
+ * Return the token limit for a model string based on the specified type.
+ *
+ * This function determines the maximum number of tokens for either input context
+ * or output generation based on the model and token type. It uses the same
+ * normalization logic for consistency across both input and output limits.
+ *
+ * @param model - The model name to get the token limit for
+ * @param type - The type of token limit ('input' for context window, 'output' for generation)
+ * @returns The maximum number of tokens allowed for this model and type
+ */
+export function tokenLimit(
+  model: Model,
+  type: TokenLimitType = 'input',
+): TokenCount {
   const norm = normalize(model);
 
-  for (const [regex, limit] of PATTERNS) {
+  // Choose the appropriate patterns based on token type
+  const patterns = type === 'output' ? OUTPUT_PATTERNS : PATTERNS;
+
+  for (const [regex, limit] of patterns) {
     if (regex.test(norm)) {
       return limit;
     }
   }
 
-  // final fallback: DEFAULT_TOKEN_LIMIT (power-of-two 128K)
-  return DEFAULT_TOKEN_LIMIT;
+  // Return appropriate default based on token type
+  return type === 'output' ? DEFAULT_OUTPUT_TOKEN_LIMIT : DEFAULT_TOKEN_LIMIT;
 }

packages/core/src/core/turn.test.ts

@@ -242,7 +242,7 @@ describe('Turn', () => {
       expect(turn.getDebugResponses().length).toBe(0);
       expect(reportError).toHaveBeenCalledWith(
         error,
-        'Error when talking to Gemini API',
+        'Error when talking to API',
         [...historyContent, reqParts],
         'Turn.run-sendMessageStream',
       );

packages/core/src/core/turn.ts

@@ -310,7 +310,7 @@ export class Turn {
       const contextForReport = [...this.chat.getHistory(/*curated*/ true), req];
       await reportError(
         error,
-        'Error when talking to Gemini API',
+        'Error when talking to API',
         contextForReport,
         'Turn.run-sendMessageStream',
       );

packages/core/src/qwen/qwenContentGenerator.test.ts

@@ -401,11 +401,9 @@ describe('QwenContentGenerator', () => {
       expect(mockQwenClient.getAccessToken).toHaveBeenCalled();
     });
 
-    it('should count tokens with valid token', async () => {
-      vi.mocked(mockQwenClient.getAccessToken).mockResolvedValue({
-        token: 'valid-token',
-      });
-      vi.mocked(mockQwenClient.getCredentials).mockReturnValue(mockCredentials);
+    it('should count tokens without requiring authentication', async () => {
+      // Clear any previous mock calls
+      vi.clearAllMocks();
 
       const request: CountTokensParameters = {
         model: 'qwen-turbo',
@@ -415,7 +413,8 @@ describe('QwenContentGenerator', () => {
       const result = await qwenContentGenerator.countTokens(request);
 
       expect(result.totalTokens).toBe(15);
-      expect(mockQwenClient.getAccessToken).toHaveBeenCalled();
+      // countTokens is a local operation and should not require OAuth credentials
+      expect(mockQwenClient.getAccessToken).not.toHaveBeenCalled();
     });
 
     it('should embed content with valid token', async () => {
@@ -1652,7 +1651,7 @@ describe('QwenContentGenerator', () => {
       SharedTokenManager.getInstance = originalGetInstance;
     });
 
-    it('should handle all method types with token failure', async () => {
+    it('should handle method types with token failure (except countTokens)', async () => {
       const mockTokenManager = {
         getValidCredentials: vi
           .fn()
@@ -1685,7 +1684,7 @@ describe('QwenContentGenerator', () => {
         contents: [{ parts: [{ text: 'Embed' }] }],
       };
 
-      // All methods should fail with the same error
+      // Methods requiring authentication should fail
       await expect(
         newGenerator.generateContent(generateRequest, 'test-id'),
       ).rejects.toThrow('Failed to obtain valid Qwen access token');
@@ -1694,14 +1693,14 @@ describe('QwenContentGenerator', () => {
         newGenerator.generateContentStream(generateRequest, 'test-id'),
       ).rejects.toThrow('Failed to obtain valid Qwen access token');
 
-      await expect(newGenerator.countTokens(countRequest)).rejects.toThrow(
-        'Failed to obtain valid Qwen access token',
-      );
-
       await expect(newGenerator.embedContent(embedRequest)).rejects.toThrow(
         'Failed to obtain valid Qwen access token',
       );
 
+      // countTokens should succeed as it's a local operation
+      const countResult = await newGenerator.countTokens(countRequest);
+      expect(countResult.totalTokens).toBe(15);
+
       SharedTokenManager.getInstance = originalGetInstance;
     });
   });

packages/core/src/qwen/qwenContentGenerator.ts

@@ -180,9 +180,7 @@ export class QwenContentGenerator extends OpenAIContentGenerator {
   override async countTokens(
     request: CountTokensParameters,
   ): Promise<CountTokensResponse> {
-    return this.executeWithCredentialManagement(() =>
-      super.countTokens(request),
-    );
+    return super.countTokens(request);
   }
 
   /**

packages/core/src/qwen/qwenOAuth2.ts

@@ -712,8 +712,6 @@ async function authWithQwenDeviceFlow(
             `Polling... (attempt ${attempt + 1}/${maxAttempts})`,
           );
 
-          process.stdout.write('.');
-
           // Wait with cancellation check every 100ms
           await new Promise<void>((resolve) => {
             const checkInterval = 100; // Check every 100ms

packages/core/src/qwen/sharedTokenManager.test.ts

@@ -901,5 +901,37 @@ describe('SharedTokenManager', () => {
         );
       }
     });
+
+    it('should properly clean up timeout when file operation completes before timeout', async () => {
+      const tokenManager = SharedTokenManager.getInstance();
+      tokenManager.clearCache();
+
+      const mockClient = {
+        getCredentials: vi.fn().mockReturnValue(null),
+        setCredentials: vi.fn(),
+        getAccessToken: vi.fn(),
+        requestDeviceAuthorization: vi.fn(),
+        pollDeviceToken: vi.fn(),
+        refreshAccessToken: vi.fn(),
+      };
+
+      // Mock clearTimeout to verify it's called
+      const clearTimeoutSpy = vi.spyOn(global, 'clearTimeout');
+
+      // Mock file stat to resolve quickly (before timeout)
+      mockFs.stat.mockResolvedValue({ mtimeMs: 12345 } as Stats);
+
+      // Call checkAndReloadIfNeeded which uses withTimeout internally
+      const checkMethod = getPrivateProperty(
+        tokenManager,
+        'checkAndReloadIfNeeded',
+      ) as (client?: IQwenOAuth2Client) => Promise<void>;
+      await checkMethod.call(tokenManager, mockClient);
+
+      // Verify that clearTimeout was called to clean up the timer
+      expect(clearTimeoutSpy).toHaveBeenCalled();
+
+      clearTimeoutSpy.mockRestore();
+    });
   });
 });

packages/core/src/qwen/sharedTokenManager.ts

@@ -290,6 +290,36 @@ export class SharedTokenManager {
     }
   }
 
+  /**
+   * Utility method to add timeout to any promise operation
+   * Properly cleans up the timeout when the promise completes
+   */
+  private withTimeout<T>(
+    promise: Promise<T>,
+    timeoutMs: number,
+    operationType = 'Operation',
+  ): Promise<T> {
+    let timeoutId: NodeJS.Timeout;
+
+    return Promise.race([
+      promise.finally(() => {
+        // Clear timeout when main promise completes (success or failure)
+        if (timeoutId) {
+          clearTimeout(timeoutId);
+        }
+      }),
+      new Promise<never>((_, reject) => {
+        timeoutId = setTimeout(
+          () =>
+            reject(
+              new Error(`${operationType} timed out after ${timeoutMs}ms`),
+            ),
+          timeoutMs,
+        );
+      }),
+    ]);
+  }
+
   /**
    * Perform the actual file check and reload operation
    * This is separated to enable proper promise-based synchronization
@@ -303,25 +333,12 @@ export class SharedTokenManager {
 
     try {
       const filePath = this.getCredentialFilePath();
-      // Add timeout to file stat operation
-      const withTimeout = async <T>(
-        promise: Promise<T>,
-        timeoutMs: number,
-      ): Promise<T> =>
-        Promise.race([
-          promise,
-          new Promise<never>((_, reject) =>
-            setTimeout(
-              () =>
-                reject(
-                  new Error(`File operation timed out after ${timeoutMs}ms`),
-                ),
-              timeoutMs,
-            ),
-          ),
-        ]);
 
-      const stats = await withTimeout(fs.stat(filePath), 3000);
+      const stats = await this.withTimeout(
+        fs.stat(filePath),
+        3000,
+        'File operation',
+      );
       const fileModTime = stats.mtimeMs;
 
       // Reload credentials if file has been modified since last cache
@@ -451,7 +468,7 @@ export class SharedTokenManager {
       // Check if we have a refresh token before attempting refresh
       const currentCredentials = qwenClient.getCredentials();
       if (!currentCredentials.refresh_token) {
-        console.debug('create a NO_REFRESH_TOKEN error');
+        // console.debug('create a NO_REFRESH_TOKEN error');
         throw new TokenManagerError(
           TokenError.NO_REFRESH_TOKEN,
           'No refresh token available for token refresh',
@@ -589,26 +606,12 @@ export class SharedTokenManager {
     const dirPath = path.dirname(filePath);
     const tempPath = `${filePath}.tmp.${randomUUID()}`;
 
-    // Add timeout wrapper for file operations
-    const withTimeout = async <T>(
-      promise: Promise<T>,
-      timeoutMs: number,
-    ): Promise<T> =>
-      Promise.race([
-        promise,
-        new Promise<never>((_, reject) =>
-          setTimeout(
-            () => reject(new Error(`Operation timed out after ${timeoutMs}ms`)),
-            timeoutMs,
-          ),
-        ),
-      ]);
-
     // Create directory with restricted permissions
     try {
-      await withTimeout(
+      await this.withTimeout(
         fs.mkdir(dirPath, { recursive: true, mode: 0o700 }),
         5000,
+        'File operation',
       );
     } catch (error) {
       throw new TokenManagerError(
@@ -622,21 +625,30 @@ export class SharedTokenManager {
 
     try {
       // Write to temporary file first with restricted permissions
-      await withTimeout(
+      await this.withTimeout(
         fs.writeFile(tempPath, credString, { mode: 0o600 }),
         5000,
+        'File operation',
       );
 
       // Atomic move to final location
-      await withTimeout(fs.rename(tempPath, filePath), 5000);
+      await this.withTimeout(
+        fs.rename(tempPath, filePath),
+        5000,
+        'File operation',
+      );
 
       // Update cached file modification time atomically after successful write
-      const stats = await withTimeout(fs.stat(filePath), 5000);
+      const stats = await this.withTimeout(
+        fs.stat(filePath),
+        5000,
+        'File operation',
+      );
       this.memoryCache.fileModTime = stats.mtimeMs;
     } catch (error) {
       // Clean up temp file if it exists
       try {
-        await withTimeout(fs.unlink(tempPath), 1000);
+        await this.withTimeout(fs.unlink(tempPath), 1000, 'File operation');
       } catch (_cleanupError) {
         // Ignore cleanup errors - temp file might not exist
       }

packages/core/src/subagents/subagent-manager.test.ts

@@ -185,6 +185,7 @@ You are a helpful assistant.
       const config = manager.parseSubagentContent(
         validMarkdown,
         validConfig.filePath,
+        'project',
       );
 
       expect(config.name).toBe('test-agent');
@@ -209,6 +210,7 @@ You are a helpful assistant.
       const config = manager.parseSubagentContent(
         markdownWithTools,
         validConfig.filePath,
+        'project',
       );
 
       expect(config.tools).toEqual(['read_file', 'write_file']);
@@ -229,6 +231,7 @@ You are a helpful assistant.
       const config = manager.parseSubagentContent(
         markdownWithModel,
         validConfig.filePath,
+        'project',
       );
 
       expect(config.modelConfig).toEqual({ model: 'custom-model', temp: 0.5 });
@@ -249,6 +252,7 @@ You are a helpful assistant.
       const config = manager.parseSubagentContent(
         markdownWithRun,
         validConfig.filePath,
+        'project',
       );
 
       expect(config.runConfig).toEqual({ max_time_minutes: 5, max_turns: 10 });
@@ -266,6 +270,7 @@ You are a helpful assistant.
       const config = manager.parseSubagentContent(
         markdownWithNumeric,
         validConfig.filePath,
+        'project',
       );
 
       expect(config.name).toBe('11');
@@ -286,6 +291,7 @@ You are a helpful assistant.
       const config = manager.parseSubagentContent(
         markdownWithBoolean,
         validConfig.filePath,
+        'project',
       );
 
       expect(config.name).toBe('true');
@@ -301,8 +307,13 @@ You are a helpful assistant.
       const projectConfig = manager.parseSubagentContent(
         validMarkdown,
         projectPath,
+        'project',
+      );
+      const userConfig = manager.parseSubagentContent(
+        validMarkdown,
+        userPath,
+        'user',
       );
-      const userConfig = manager.parseSubagentContent(validMarkdown, userPath);
 
       expect(projectConfig.level).toBe('project');
       expect(userConfig.level).toBe('user');
@@ -313,7 +324,11 @@ You are a helpful assistant.
 Just content`;
 
       expect(() =>
-        manager.parseSubagentContent(invalidMarkdown, validConfig.filePath),
+        manager.parseSubagentContent(
+          invalidMarkdown,
+          validConfig.filePath,
+          'project',
+        ),
       ).toThrow(SubagentError);
     });
 
@@ -326,7 +341,11 @@ You are a helpful assistant.
 `;
 
       expect(() =>
-        manager.parseSubagentContent(markdownWithoutName, validConfig.filePath),
+        manager.parseSubagentContent(
+          markdownWithoutName,
+          validConfig.filePath,
+          'project',
+        ),
       ).toThrow(SubagentError);
     });
 
@@ -342,39 +361,20 @@ You are a helpful assistant.
         manager.parseSubagentContent(
           markdownWithoutDescription,
           validConfig.filePath,
+          'project',
         ),
       ).toThrow(SubagentError);
     });
 
-    it('should warn when filename does not match subagent name', () => {
-      const consoleSpy = vi.spyOn(console, 'warn').mockImplementation(() => {});
-      const mismatchedPath = '/test/project/.qwen/agents/wrong-filename.md';
-
-      const config = manager.parseSubagentContent(
-        validMarkdown,
-        mismatchedPath,
-      );
-
-      expect(config.name).toBe('test-agent');
-      expect(consoleSpy).toHaveBeenCalledWith(
-        expect.stringContaining(
-          'Warning: Subagent file "wrong-filename.md" contains name "test-agent"',
-        ),
-      );
-      expect(consoleSpy).toHaveBeenCalledWith(
-        expect.stringContaining(
-          'Consider renaming the file to "test-agent.md"',
-        ),
-      );
-
-      consoleSpy.mockRestore();
-    });
-
     it('should not warn when filename matches subagent name', () => {
       const consoleSpy = vi.spyOn(console, 'warn').mockImplementation(() => {});
       const matchingPath = '/test/project/.qwen/agents/test-agent.md';
 
-      const config = manager.parseSubagentContent(validMarkdown, matchingPath);
+      const config = manager.parseSubagentContent(
+        validMarkdown,
+        matchingPath,
+        'project',
+      );
 
       expect(config.name).toBe('test-agent');
       expect(consoleSpy).not.toHaveBeenCalled();

packages/core/src/subagents/subagent-manager.ts

@@ -39,11 +39,30 @@ const AGENT_CONFIG_DIR = 'agents';
  */
 export class SubagentManager {
   private readonly validator: SubagentValidator;
+  private subagentsCache: Map<SubagentLevel, SubagentConfig[]> | null = null;
+  private readonly changeListeners: Set<() => void> = new Set();
 
   constructor(private readonly config: Config) {
     this.validator = new SubagentValidator();
   }
 
+  addChangeListener(listener: () => void): () => void {
+    this.changeListeners.add(listener);
+    return () => {
+      this.changeListeners.delete(listener);
+    };
+  }
+
+  private notifyChangeListeners(): void {
+    for (const listener of this.changeListeners) {
+      try {
+        listener();
+      } catch (error) {
+        console.warn('Subagent change listener threw an error:', error);
+      }
+    }
+  }
+
   /**
    * Creates a new subagent configuration.
    *
@@ -92,6 +111,8 @@ export class SubagentManager {
 
     try {
       await fs.writeFile(filePath, content, 'utf8');
+      // Refresh cache after successful creation
+      await this.refreshCache();
     } catch (error) {
       throw new SubagentError(
         `Failed to write subagent file: ${error instanceof Error ? error.message : 'Unknown error'}`,
@@ -180,6 +201,8 @@ export class SubagentManager {
 
     try {
       await fs.writeFile(existing.filePath, content, 'utf8');
+      // Refresh cache after successful update
+      await this.refreshCache();
     } catch (error) {
       throw new SubagentError(
         `Failed to update subagent file: ${error instanceof Error ? error.message : 'Unknown error'}`,
@@ -236,6 +259,9 @@ export class SubagentManager {
         name,
       );
     }
+
+    // Refresh cache after successful deletion
+    await this.refreshCache();
   }
 
   /**
@@ -254,9 +280,17 @@ export class SubagentManager {
       ? [options.level]
       : ['project', 'user', 'builtin'];
 
+    // Check if we should use cache or force refresh
+    const shouldUseCache = !options.force && this.subagentsCache !== null;
+
+    // Initialize cache if it doesn't exist or we're forcing a refresh
+    if (!shouldUseCache) {
+      await this.refreshCache();
+    }
+
     // Collect subagents from each level (project takes precedence over user, user takes precedence over builtin)
     for (const level of levelsToCheck) {
-      const levelSubagents = await this.listSubagentsAtLevel(level);
+      const levelSubagents = this.subagentsCache?.get(level) || [];
 
       for (const subagent of levelSubagents) {
         // Skip if we've already seen this name (precedence: project > user > builtin)
@@ -304,6 +338,26 @@ export class SubagentManager {
     return subagents;
   }
 
+  /**
+   * Refreshes the subagents cache by loading all subagents from disk.
+   * This method is called automatically when cache is null or when force=true.
+   *
+   * @private
+   */
+  private async refreshCache(): Promise<void> {
+    const subagentsCache = new Map();
+
+    const levels: SubagentLevel[] = ['project', 'user', 'builtin'];
+
+    for (const level of levels) {
+      const levelSubagents = await this.listSubagentsAtLevel(level);
+      subagentsCache.set(level, levelSubagents);
+    }
+
+    this.subagentsCache = subagentsCache;
+    this.notifyChangeListeners();
+  }
+
   /**
    * Finds a subagent by name and returns its metadata.
    *
@@ -329,7 +383,10 @@ export class SubagentManager {
    * @returns SubagentConfig
    * @throws SubagentError if parsing fails
    */
-  async parseSubagentFile(filePath: string): Promise<SubagentConfig> {
+  async parseSubagentFile(
+    filePath: string,
+    level: SubagentLevel,
+  ): Promise<SubagentConfig> {
     let content: string;
 
     try {
@@ -341,7 +398,7 @@ export class SubagentManager {
       );
     }
 
-    return this.parseSubagentContent(content, filePath);
+    return this.parseSubagentContent(content, filePath, level);
   }
 
   /**
@@ -352,7 +409,11 @@ export class SubagentManager {
    * @returns SubagentConfig
    * @throws SubagentError if parsing fails
    */
-  parseSubagentContent(content: string, filePath: string): SubagentConfig {
+  parseSubagentContent(
+    content: string,
+    filePath: string,
+    level: SubagentLevel,
+  ): SubagentConfig {
     try {
       // Split frontmatter and content
       const frontmatterRegex = /^---\n([\s\S]*?)\n---\n([\s\S]*)$/;
@@ -393,31 +454,16 @@ export class SubagentManager {
         | undefined;
       const color = frontmatter['color'] as string | undefined;
 
-      // Determine level from file path using robust, cross-platform check
-      // A project-level agent lives under <projectRoot>/.qwen/agents
-      const projectAgentsDir = path.join(
-        this.config.getProjectRoot(),
-        QWEN_CONFIG_DIR,
-        AGENT_CONFIG_DIR,
-      );
-      const rel = path.relative(
-        path.normalize(projectAgentsDir),
-        path.normalize(filePath),
-      );
-      const isProjectLevel =
-        rel !== '' && !rel.startsWith('..') && !path.isAbsolute(rel);
-      const level: SubagentLevel = isProjectLevel ? 'project' : 'user';
-
       const config: SubagentConfig = {
         name,
         description,
         tools,
         systemPrompt: systemPrompt.trim(),
-        level,
         filePath,
         modelConfig: modelConfig as Partial<ModelConfig>,
         runConfig: runConfig as Partial<RunConfig>,
         color,
+        level,
       };
 
       // Validate the parsed configuration
@@ -426,16 +472,6 @@ export class SubagentManager {
         throw new Error(`Validation failed: ${validation.errors.join(', ')}`);
       }
 
-      // Warn if filename doesn't match subagent name (potential issue)
-      const expectedFilename = `${config.name}.md`;
-      const actualFilename = path.basename(filePath);
-      if (actualFilename !== expectedFilename) {
-        console.warn(
-          `Warning: Subagent file "${actualFilename}" contains name "${config.name}" but filename suggests "${path.basename(actualFilename, '.md')}". ` +
-            `Consider renaming the file to "${expectedFilename}" for consistency.`,
-        );
-      }
-
       return config;
     } catch (error) {
       throw new SubagentError(
@@ -678,14 +714,18 @@ export class SubagentManager {
       return BuiltinAgentRegistry.getBuiltinAgents();
     }
 
-    const baseDir =
-      level === 'project'
-        ? path.join(
-            this.config.getProjectRoot(),
-            QWEN_CONFIG_DIR,
-            AGENT_CONFIG_DIR,
-          )
-        : path.join(os.homedir(), QWEN_CONFIG_DIR, AGENT_CONFIG_DIR);
+    const projectRoot = this.config.getProjectRoot();
+    const homeDir = os.homedir();
+    const isHomeDirectory = path.resolve(projectRoot) === path.resolve(homeDir);
+
+    // If project level is requested but project root is same as home directory,
+    // return empty array to avoid conflicts between project and global agents
+    if (level === 'project' && isHomeDirectory) {
+      return [];
+    }
+
+    let baseDir = level === 'project' ? projectRoot : homeDir;
+    baseDir = path.join(baseDir, QWEN_CONFIG_DIR, AGENT_CONFIG_DIR);
 
     try {
       const files = await fs.readdir(baseDir);
@@ -697,7 +737,7 @@ export class SubagentManager {
         const filePath = path.join(baseDir, file);
 
         try {
-          const config = await this.parseSubagentFile(filePath);
+          const config = await this.parseSubagentFile(filePath, level);
           subagents.push(config);
         } catch (_error) {
           // Ignore invalid files

packages/core/src/subagents/subagent.test.ts

@@ -23,7 +23,11 @@ import {
 } from 'vitest';
 import { Config, type ConfigParameters } from '../config/config.js';
 import { DEFAULT_GEMINI_MODEL } from '../config/models.js';
-import { createContentGenerator } from '../core/contentGenerator.js';
+import {
+  createContentGenerator,
+  createContentGeneratorConfig,
+  AuthType,
+} from '../core/contentGenerator.js';
 import { GeminiChat } from '../core/geminiChat.js';
 import { executeToolCall } from '../core/nonInteractiveToolExecutor.js';
 import type { ToolRegistry } from '../tools/tool-registry.js';
@@ -37,12 +41,14 @@ import type {
   ToolConfig,
 } from './types.js';
 import { SubagentTerminateMode } from './types.js';
+import { GeminiClient } from '../core/client.js';
 
 vi.mock('../core/geminiChat.js');
 vi.mock('../core/contentGenerator.js');
 vi.mock('../utils/environmentContext.js');
 vi.mock('../core/nonInteractiveToolExecutor.js');
 vi.mock('../ide/ide-client.js');
+vi.mock('../core/client.js');
 
 async function createMockConfig(
   toolRegistryMocks = {},
@@ -56,8 +62,7 @@ async function createMockConfig(
   };
   const config = new Config(configParams);
   await config.initialize();
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
-  await config.refreshAuth('test-auth' as any);
+  await config.refreshAuth(AuthType.USE_GEMINI);
 
   // Mock ToolRegistry
   const mockToolRegistry = {
@@ -69,6 +74,19 @@ async function createMockConfig(
   } as unknown as ToolRegistry;
 
   vi.spyOn(config, 'getToolRegistry').mockReturnValue(mockToolRegistry);
+
+  // Mock getContentGeneratorConfig to return a valid config
+  vi.spyOn(config, 'getContentGeneratorConfig').mockReturnValue({
+    model: DEFAULT_GEMINI_MODEL,
+    authType: AuthType.USE_GEMINI,
+  });
+
+  // Mock setModel method
+  vi.spyOn(config, 'setModel').mockResolvedValue();
+
+  // Mock getSessionId method
+  vi.spyOn(config, 'getSessionId').mockReturnValue('test-session');
+
   return { config, toolRegistry: mockToolRegistry };
 }
 
@@ -164,6 +182,10 @@ describe('subagent.ts', () => {
         getGenerativeModel: vi.fn(),
         // eslint-disable-next-line @typescript-eslint/no-explicit-any
       } as any);
+      vi.mocked(createContentGeneratorConfig).mockReturnValue({
+        model: DEFAULT_GEMINI_MODEL,
+        authType: undefined,
+      });
 
       mockSendMessageStream = vi.fn();
       // We mock the implementation of the constructor.
@@ -174,6 +196,28 @@ describe('subagent.ts', () => {
           }) as unknown as GeminiChat,
       );
 
+      // Mock GeminiClient constructor to return a properly mocked client
+      const mockGeminiChat = {
+        setTools: vi.fn(),
+        getHistory: vi.fn().mockReturnValue([]),
+        setHistory: vi.fn(),
+        sendMessageStream: vi.fn(),
+      };
+
+      const mockGeminiClient = {
+        getChat: vi.fn().mockReturnValue(mockGeminiChat),
+        setTools: vi.fn().mockResolvedValue(undefined),
+        isInitialized: vi.fn().mockReturnValue(true),
+        getHistory: vi.fn().mockReturnValue([]),
+        initialize: vi.fn().mockResolvedValue(undefined),
+        setHistory: vi.fn(),
+      };
+
+      // Mock the GeminiClient constructor
+      vi.mocked(GeminiClient).mockImplementation(
+        () => mockGeminiClient as unknown as GeminiClient,
+      );
+
       // Default mock for executeToolCall
       vi.mocked(executeToolCall).mockResolvedValue({
         callId: 'default-call',

packages/core/src/subagents/subagent.ts

@@ -826,7 +826,7 @@ export class SubAgentScope {
       );
 
       if (this.modelConfig.model) {
-        this.runtimeContext.setModel(this.modelConfig.model);
+        await this.runtimeContext.setModel(this.modelConfig.model);
       }
 
       return new GeminiChat(