diff --git a/docs/agents/agent-types/copilot-cli.md b/docs/agents/agent-types/copilot-cli.md index 464ff207c1..af4ec819f5 100644 --- a/docs/agents/agent-types/copilot-cli.md +++ b/docs/agents/agent-types/copilot-cli.md @@ -68,7 +68,7 @@ If you want the changes from the Copilot CLI session to be applied directly to y Copilot CLI sessions support the same [permission levels](/docs/agents/approvals.md#permission-levels) as local agents. The available permission levels depend on the isolation mode you choose: * **Worktree isolation**: the permission level is automatically set to **Bypass Approvals** and can't be changed. Because the agent operates on an isolated copy of your codebase (Git worktree), all tool calls are auto-approved without confirmation dialogs. -* **Folder isolation**: all three permission levels are available (**Default Approvals**, **Bypass Approvals**, and **Autopilot**), just like local agent sessions. Select a level from the permissions picker in the chat input area. +* **Folder isolation**: you can select the permission level from the permissions picker in the chat input area. **Default Approvals** and **Bypass Approvals** are always available, and **Assisted permissions** is available when you enable `setting(chat.assistedPermissions.enabled)`. ## Create a Copilot CLI session diff --git a/docs/agents/approvals.md b/docs/agents/approvals.md index 7255ebda0a..b904b6788e 100644 --- a/docs/agents/approvals.md +++ b/docs/agents/approvals.md @@ -39,11 +39,14 @@ The permission level applies to the current chat session, and can be changed at | Permission level | Description | |---|---| | **Default Approvals** (default) | Uses your configured approval settings. Tools that require approval show a confirmation dialog before they run. When in doubt, the agent asks clarifying questions. | +| **Assisted permissions** | Uses an LLM judge to evaluate each tool call. Tool calls the judge does not approve still require your approval. | | **Bypass Approvals** | Auto-approves all tool calls without showing confirmation dialogs. When in doubt, the agent asks clarifying questions. | | **Autopilot** | Auto-approves all tool calls without showing confirmation dialogs. When questions arise, the agent automatically responds to clarifying questions. | The permission level determines whether your finer-grained settings apply. **Default Approvals** respects the per-tool, URL, terminal, and sandbox settings you configure in the following sections. **Bypass Approvals** and **Autopilot** override those settings and approve everything automatically. +Some agent-host sessions, such as Copilot CLI sessions, can show **Assisted permissions** in the permissions picker. To show this option, enable `setting(chat.assistedPermissions.enabled)`. + > [!CAUTION] > **Bypass Approvals** and **Autopilot** bypass manual approval prompts, including for potentially destructive actions like file edits, terminal commands, and external tool calls. The first time you enable either level, a warning dialog asks you to confirm. Only use these levels if you understand the security implications. See the [Security considerations](/docs/agents/security.md) for more details. diff --git a/docs/agents/reference/ai-settings.md b/docs/agents/reference/ai-settings.md index a5d2f73496..7dbd5f8b8e 100644 --- a/docs/agents/reference/ai-settings.md +++ b/docs/agents/reference/ai-settings.md @@ -114,7 +114,8 @@ The team is continuously working on improving the AI features in VS Code and add | `setting(chat.tools.terminal.enforceTimeoutFromModel)` _(Experimental)_
Control whether to enforce the timeout value that the agent specifies for terminal commands. When enabled, the agent stops tracking the command after the specified duration and returns the output collected so far. | `true` | | `setting(chat.tools.terminal.ignoreDefaultAutoApproveRules)`
Ignore the default auto-approve rules for terminal commands. | `false` | | `setting(chat.tools.global.autoApprove)`
Automatically approve all tools - this setting [disables critical security protections](/docs/agents/security.md). | `false` | -| `setting(chat.permissions.default)` _(Experimental)_
Set the default [permission level](/docs/agents/approvals.md#permission-levels) for new chat sessions. Options: `default` (Default Approvals), `autoApprove` (Bypass Approvals), `autopilot` (Autopilot). You can still change the permission level per session. If enterprise policy disables auto-approval, new sessions use Default Approvals. | `"default"` | +| `setting(chat.permissions.default)` _(Experimental)_
Set the default [permission level](/docs/agents/approvals.md#permission-levels) for new local chat sessions. Options: `default` (Default Approvals), `autoApprove` (Bypass Approvals), `autopilot` (Autopilot). You can still change the permission level per session. If enterprise policy disables auto-approval, new sessions use Default Approvals. | `"default"` | +| `setting(chat.assistedPermissions.enabled)` _(Experimental)_
Control whether **Assisted permissions** is shown in Agent Host approval pickers, such as Copilot CLI sessions. | `true` (Insiders)
`false` (Stable) | | `setting(chat.autopilot.advanced.enabled)` _(Experimental)_
Enable [Advanced Autopilot](/docs/agents/approvals.md#advanced-autopilot), where a separate model evaluates whether your request is complete after each Autopilot turn and guides the next turn, instead of relying on the agent to signal completion. | `false` | | `setting(chat.tools.urls.autoApprove)`
Control which [URL requests and responses are auto-approved](/docs/agents/approvals.md#url-approval). | `[]` | | `setting(chat.agent.thinking.collapsedTools)` _(Experimental)_
Configure how tool calls are grouped with thinking content: `off` keeps tool calls separate, `withThinking` groups them only when thinking is present, and `always` always groups tool calls in collapsible sections. | `always` |