feat: add permission approval hooks

.changeset/permission-approval-hooks.md

@@ -0,0 +1,6 @@
+---
+"@moonshot-ai/agent-core": minor
+"@moonshot-ai/kimi-code": minor
+---
+
+Add approval lifecycle hook events for observing pending and completed permission prompts.

docs/en/customization/hooks.md

@@ -74,7 +74,7 @@ If stdout is JSON and `hookSpecificOutput.permissionDecision` is `deny`, the res
 }
 ```
 
-Blocking only applies to events that participate in control flow. For example, `PreToolUse` can block a tool call, and `Stop` can append one continuation message to the current turn. Observer events (such as `PostToolUse`, `PostToolUseFailure`, `PostCompact`, `SubagentStop`, `StopFailure`, and `Notification`) are dispatched asynchronously in a fire-and-forget fashion; their return values are ignored and do not change the main flow. `PreCompact` is invoked with `trigger` (not `triggerBlock`); its return value is likewise completely ignored, and it is not a blockable event.
+Blocking only applies to events that participate in control flow. For example, `PreToolUse` can block a tool call, and `Stop` can append one continuation message to the current turn. Observer events (such as `PostToolUse`, `PostToolUseFailure`, `PermissionRequest`, `PermissionResult`, `PostCompact`, `SubagentStop`, `StopFailure`, and `Notification`) are dispatched asynchronously in a fire-and-forget fashion; their return values are ignored and do not change the main flow. `PreCompact` is invoked with `trigger` (not `triggerBlock`); its return value is likewise completely ignored, and it is not a blockable event.
 
 When a block takes effect, if the script does not provide a reason through stderr or JSON output, the CLI falls back to `Blocked by <event> hook` as a placeholder reason. A `PreToolUse` block is written back into context as a failed tool result, so the model can choose an alternative based on the reason.
 
@@ -88,6 +88,8 @@ The following events are triggered automatically today:
 | `PreToolUse` | Tool name | `tool_name`, `tool_input`, `tool_call_id` | Fires before permission checks. If blocked, the tool does not run |
 | `PostToolUse` | Tool name | `tool_name`, `tool_input`, `tool_call_id`, `tool_output` | Fires after a successful tool call. `tool_output` is truncated to the first 2000 characters |
 | `PostToolUseFailure` | Tool name | `tool_name`, `tool_input`, `tool_call_id`, `error` | Fires after a tool call fails or is blocked by a hook |
+| `PermissionRequest` | Tool name | `turn_id`, `tool_call_id`, `tool_name`, `action`, `tool_input`, `display` | Fires asynchronously immediately before the CLI waits for user approval |
+| `PermissionResult` | Tool name | `turn_id`, `tool_call_id`, `tool_name`, `action`, `decision`, `scope`, `feedback`, `selected_label`, `error` | Fires asynchronously after the approval request resolves or fails |
 | `Stop` | Empty string | `stop_hook_active` | Fires when the model is about to stop. If blocked, the reason is appended directly to context as a system-triggered user message, and the turn may continue once |
 | `StopFailure` | Error type | `error_type`, `error_message` | Fires after the current turn fails with a non-cancellation error |
 | `SessionStart` | `startup` or `resume` | `source` | Fires after the main agent is created for a new session, or after a historical session is resumed |

docs/zh/customization/hooks.md

@@ -74,7 +74,7 @@ Hook 命令的退出码和 stdout 会被解释为以下结果：
 }
 ```
 
-阻断只对支持控制流的事件生效。例如 `PreToolUse` 可以阻断工具调用，`Stop` 可以让当前轮次追加一次继续消息。观察型事件（例如 `PostToolUse`、`PostToolUseFailure`、`PostCompact`、`SubagentStop`、`StopFailure`、`Notification`）以「即发即忘（fire-and-forget）」方式异步触发，返回值被忽略，不会改变主流程。`PreCompact` 使用 `trigger`（而非 `triggerBlock`）调用，返回值同样被完全忽略，不属于可阻断事件。
+阻断只对支持控制流的事件生效。例如 `PreToolUse` 可以阻断工具调用，`Stop` 可以让当前轮次追加一次继续消息。观察型事件（例如 `PostToolUse`、`PostToolUseFailure`、`PermissionRequest`、`PermissionResult`、`PostCompact`、`SubagentStop`、`StopFailure`、`Notification`）以「即发即忘（fire-and-forget）」方式异步触发，返回值被忽略，不会改变主流程。`PreCompact` 使用 `trigger`（而非 `triggerBlock`）调用，返回值同样被完全忽略，不属于可阻断事件。
 
 阻断生效时，如果脚本未通过 stderr 或 JSON 输出提供原因，CLI 会回退到 `Blocked by <event> hook` 作为占位原因。`PreToolUse` 阻断会作为工具失败结果写回上下文，模型可以根据原因选择替代方案。
 
@@ -88,6 +88,8 @@ Hook 命令的退出码和 stdout 会被解释为以下结果：
 | `PreToolUse` | 工具名 | `tool_name`、`tool_input`、`tool_call_id` | 在权限检查前触发；阻断后工具不会执行 |
 | `PostToolUse` | 工具名 | `tool_name`、`tool_input`、`tool_call_id`、`tool_output` | 工具成功后触发；`tool_output` 被截断至前 2000 个字符 |
 | `PostToolUseFailure` | 工具名 | `tool_name`、`tool_input`、`tool_call_id`、`error` | 工具失败或被 hook 阻断后触发 |
+| `PermissionRequest` | 工具名 | `turn_id`、`tool_call_id`、`tool_name`、`action`、`tool_input`、`display` | CLI 即将等待用户审批前异步触发 |
+| `PermissionResult` | 工具名 | `turn_id`、`tool_call_id`、`tool_name`、`action`、`decision`、`scope`、`feedback`、`selected_label`、`error` | 审批请求结束或失败后异步触发 |
 | `Stop` | 空字符串 | `stop_hook_active` | 模型准备停止时触发；阻断后会把原因直接作为系统触发的 User 消息追加进上下文，并最多继续一次 |
 | `StopFailure` | 错误类型 | `error_type`、`error_message` | 当前轮次因非取消错误失败后触发 |
 | `SessionStart` | `startup` 或 `resume` | `source` | 新会话主 Agent 创建后，或历史会话恢复完成后触发 |

packages/agent-core/src/agent/permission/index.ts

@@ -131,7 +131,20 @@ export class PermissionManager {
     const startedAt = Date.now();
 
     let response: ApprovalResponse;
+    let requestedApproval = false;
     if (this.agent.rpc?.requestApproval) {
+      requestedApproval = true;
+      void this.agent.hooks?.fireAndForgetTrigger?.('PermissionRequest', {
+        matcherValue: name,
+        inputData: {
+          turnId: Number(context.turnId),
+          toolCallId: id,
+          toolName: name,
+          action,
+          toolInput: context.args,
+          display,
+        },
+      });
       try {
         response = await this.agent.rpc.requestApproval(
           {
@@ -154,6 +167,17 @@ export class PermissionManager {
           session_cache_written: false,
           has_feedback: false,
         });
+        void this.agent.hooks?.fireAndForgetTrigger?.('PermissionResult', {
+          matcherValue: name,
+          inputData: {
+            turnId: Number(context.turnId),
+            toolCallId: id,
+            toolName: name,
+            action,
+            decision: 'error',
+            error: error instanceof Error ? error.message : String(error),
+          },
+        });
         const resolved = result.resolveError?.(error);
         return resolved === undefined
           ? Promise.reject(error)
@@ -170,6 +194,22 @@ export class PermissionManager {
         ? context.execution.approvalRule
         : undefined;
 
+    if (requestedApproval) {
+      void this.agent.hooks?.fireAndForgetTrigger?.('PermissionResult', {
+        matcherValue: name,
+        inputData: {
+          turnId: Number(context.turnId),
+          toolCallId: id,
+          toolName: name,
+          action,
+          decision: response.decision,
+          scope: response.scope,
+          feedback: response.feedback,
+          selectedLabel: response.selectedLabel,
+        },
+      });
+    }
+
     this.recordApprovalResult({
       turnId: Number(context.turnId),
       toolCallId: id,

packages/agent-core/src/session/hooks/types.ts

@@ -4,6 +4,8 @@ export const HOOK_EVENT_TYPES = [
   'PreToolUse',
   'PostToolUse',
   'PostToolUseFailure',
+  'PermissionRequest',
+  'PermissionResult',
   'UserPromptSubmit',
   'Stop',
   'StopFailure',

packages/agent-core/test/agent/permission.test.ts

@@ -2293,6 +2293,82 @@ describe('Agent-local approve for session', () => {
 });
 
 describe('Approval telemetry', () => {
+  it('fires observer hooks while waiting for user approval', async () => {
+    const triggerBlock = vi.fn(async () => undefined);
+    const fireAndForgetTrigger = vi.fn(async () => []);
+    const { manager, requestApproval } = makePermissionManager(
+      async () => {
+        expect(fireAndForgetTrigger).toHaveBeenCalledWith('PermissionRequest', {
+          matcherValue: 'Bash',
+          inputData: {
+            turnId: 0,
+            toolCallId: 'call_approval_hooks',
+            toolName: 'Bash',
+            action: 'run command',
+            toolInput: { command: 'printf first', timeout: 60 },
+            display: expect.objectContaining({ kind: 'command' }),
+          },
+        });
+        expect(fireAndForgetTrigger).not.toHaveBeenCalledWith(
+          'PermissionResult',
+          expect.anything(),
+        );
+        return {
+          decision: 'approved',
+          selectedLabel: 'Approve once',
+        };
+      },
+      {
+        hooks: { triggerBlock, fireAndForgetTrigger } as unknown as Agent['hooks'],
+      },
+    );
+
+    await expect(manager.beforeToolCall(hookContext({ id: 'call_approval_hooks' }))).resolves
+      .toBeUndefined();
+
+    expect(requestApproval).toHaveBeenCalledTimes(1);
+    expect(fireAndForgetTrigger).toHaveBeenCalledWith('PermissionResult', {
+      matcherValue: 'Bash',
+      inputData: {
+        turnId: 0,
+        toolCallId: 'call_approval_hooks',
+        toolName: 'Bash',
+        action: 'run command',
+        decision: 'approved',
+        selectedLabel: 'Approve once',
+        scope: undefined,
+        feedback: undefined,
+      },
+    });
+  });
+
+  it('does not fire approval observer hooks without an approval request', async () => {
+    const triggerBlock = vi.fn(async () => undefined);
+    const fireAndForgetTrigger = vi.fn(async () => []);
+    const { manager, requestApproval } = makePermissionManager(
+      async () => ({
+        decision: 'approved',
+      }),
+      {
+        approvalRpc: false,
+        hooks: { triggerBlock, fireAndForgetTrigger } as unknown as Agent['hooks'],
+      },
+    );
+
+    await expect(manager.beforeToolCall(hookContext({ id: 'call_no_approval_rpc' }))).resolves
+      .toBeUndefined();
+
+    expect(requestApproval).not.toHaveBeenCalled();
+    expect(fireAndForgetTrigger).not.toHaveBeenCalledWith(
+      'PermissionRequest',
+      expect.anything(),
+    );
+    expect(fireAndForgetTrigger).not.toHaveBeenCalledWith(
+      'PermissionResult',
+      expect.anything(),
+    );
+  });
+
   it('tracks cancelled approval requests', async () => {
     const { manager, telemetryTrack } = makePermissionManager(async () => ({
       decision: 'cancelled',
@@ -3562,6 +3638,7 @@ function makePermissionManager(
     readonly cwd?: string;
     readonly agentType?: Agent['type'];
     readonly hooks?: Agent['hooks'];
+    readonly approvalRpc?: boolean;
   } = {},
 ): {
   manager: PermissionManager;
@@ -3580,7 +3657,7 @@ function makePermissionManager(
     emitStatusUpdated: vi.fn(),
     records: { logRecord: record },
     replayBuilder: { push: vi.fn() },
-    rpc: { requestApproval },
+    rpc: options.approvalRpc === false ? undefined : { requestApproval },
     hooks: options.hooks,
     telemetry: { track: telemetryTrack },
     planMode: {