Filehigh importancesource

interactiveHandler.ts

hooks/toolPermission/handlers/interactiveHandler.ts

537
Lines
20194
Bytes
1
Exports
19
Imports
10
Keywords

What this is

This page documents one file from the repository and includes its full source so you can read it without leaving the docs site.

Beginner explanation

This file is one piece of the larger system. Its name, directory, imports, and exports show where it fits. Start by reading the exports and related files first.

How it is used

Start from the exports list and related files. Those are the easiest clues for where this file fits into the system.

Expert explanation

Architecturally, this file intersects with tool-system, shell-safety, permissions. It contains 537 lines, 19 detected imports, and 1 detected exports.

Important relationships

Detected exports

  • handleInteractivePermission

Keywords

bridgecallbacksbridgerequestidclassifierchannelunsubscribetooluseidresponseresultpermissionpromptstarttimemssignaluser

Detected imports

  • bun:bundle
  • @anthropic-ai/sdk/resources/messages.mjs
  • crypto
  • src/utils/debug.js
  • ../../../bootstrap/state.js
  • ../../../bridge/bridgePermissionCallbacks.js
  • ../../../ink/terminal-focus-state.js
  • ../../../services/mcp/channelNotification.js
  • ../../../services/mcp/channelPermissions.js
  • ../../../services/mcp/channelPermissions.js
  • ../../../tools/BashTool/bashPermissions.js
  • ../../../tools/BashTool/toolName.js
  • ../../../utils/classifierApprovals.js
  • ../../../utils/errors.js
  • ../../../utils/permissions/PermissionResult.js
  • ../../../utils/permissions/PermissionUpdateSchema.js
  • ../../../utils/permissions/permissions.js
  • ../PermissionContext.js
  • ../PermissionContext.js

Source notes

This page embeds the full file contents. Small or leaf files are still indexed honestly instead of being over-explained.

Open parent directory

Full source

import { feature } from 'bun:bundle'
import type { ContentBlockParam } from '@anthropic-ai/sdk/resources/messages.mjs'
import { randomUUID } from 'crypto'
import { logForDebugging } from 'src/utils/debug.js'
import { getAllowedChannels } from '../../../bootstrap/state.js'
import type { BridgePermissionCallbacks } from '../../../bridge/bridgePermissionCallbacks.js'
import { getTerminalFocused } from '../../../ink/terminal-focus-state.js'
import {
  CHANNEL_PERMISSION_REQUEST_METHOD,
  type ChannelPermissionRequestParams,
  findChannelEntry,
} from '../../../services/mcp/channelNotification.js'
import type { ChannelPermissionCallbacks } from '../../../services/mcp/channelPermissions.js'
import {
  filterPermissionRelayClients,
  shortRequestId,
  truncateForPreview,
} from '../../../services/mcp/channelPermissions.js'
import { executeAsyncClassifierCheck } from '../../../tools/BashTool/bashPermissions.js'
import { BASH_TOOL_NAME } from '../../../tools/BashTool/toolName.js'
import {
  clearClassifierChecking,
  setClassifierApproval,
  setClassifierChecking,
  setYoloClassifierApproval,
} from '../../../utils/classifierApprovals.js'
import { errorMessage } from '../../../utils/errors.js'
import type { PermissionDecision } from '../../../utils/permissions/PermissionResult.js'
import type { PermissionUpdate } from '../../../utils/permissions/PermissionUpdateSchema.js'
import { hasPermissionsToUseTool } from '../../../utils/permissions/permissions.js'
import type { PermissionContext } from '../PermissionContext.js'
import { createResolveOnce } from '../PermissionContext.js'

type InteractivePermissionParams = {
  ctx: PermissionContext
  description: string
  result: PermissionDecision & { behavior: 'ask' }
  awaitAutomatedChecksBeforeDialog: boolean | undefined
  bridgeCallbacks?: BridgePermissionCallbacks
  channelCallbacks?: ChannelPermissionCallbacks
}

/**
 * Handles the interactive (main-agent) permission flow.
 *
 * Pushes a ToolUseConfirm entry to the confirm queue with callbacks:
 * onAbort, onAllow, onReject, recheckPermission, onUserInteraction.
 *
 * Runs permission hooks and bash classifier checks asynchronously in the
 * background, racing them against user interaction. Uses a resolve-once
 * guard and `userInteracted` flag to prevent multiple resolutions.
 *
 * This function does NOT return a Promise -- it sets up callbacks that
 * eventually call `resolve()` to resolve the outer promise owned by
 * the caller.
 */
function handleInteractivePermission(
  params: InteractivePermissionParams,
  resolve: (decision: PermissionDecision) => void,
): void {
  const {
    ctx,
    description,
    result,
    awaitAutomatedChecksBeforeDialog,
    bridgeCallbacks,
    channelCallbacks,
  } = params

  const { resolve: resolveOnce, isResolved, claim } = createResolveOnce(resolve)
  let userInteracted = false
  let checkmarkTransitionTimer: ReturnType<typeof setTimeout> | undefined
  // Hoisted so onDismissCheckmark (Esc during checkmark window) can also
  // remove the abort listener — not just the timer callback.
  let checkmarkAbortHandler: (() => void) | undefined
  const bridgeRequestId = bridgeCallbacks ? randomUUID() : undefined
  // Hoisted so local/hook/classifier wins can remove the pending channel
  // entry. No "tell remote to dismiss" equivalent — the text sits in your
  // phone, and a stale "yes abc123" after local-resolve falls through
  // tryConsumeReply (entry gone) and gets enqueued as normal chat.
  let channelUnsubscribe: (() => void) | undefined

  const permissionPromptStartTimeMs = Date.now()
  const displayInput = result.updatedInput ?? ctx.input

  function clearClassifierIndicator(): void {
    if (feature('BASH_CLASSIFIER')) {
      ctx.updateQueueItem({ classifierCheckInProgress: false })
    }
  }

  ctx.pushToQueue({
    assistantMessage: ctx.assistantMessage,
    tool: ctx.tool,
    description,
    input: displayInput,
    toolUseContext: ctx.toolUseContext,
    toolUseID: ctx.toolUseID,
    permissionResult: result,
    permissionPromptStartTimeMs,
    ...(feature('BASH_CLASSIFIER')
      ? {
          classifierCheckInProgress:
            !!result.pendingClassifierCheck &&
            !awaitAutomatedChecksBeforeDialog,
        }
      : {}),
    onUserInteraction() {
      // Called when user starts interacting with the permission dialog
      // (e.g., arrow keys, tab, typing feedback)
      // Hide the classifier indicator since auto-approve is no longer possible
      //
      // Grace period: ignore interactions in the first 200ms to prevent
      // accidental keypresses from canceling the classifier prematurely
      const GRACE_PERIOD_MS = 200
      if (Date.now() - permissionPromptStartTimeMs < GRACE_PERIOD_MS) {
        return
      }
      userInteracted = true
      clearClassifierChecking(ctx.toolUseID)
      clearClassifierIndicator()
    },
    onDismissCheckmark() {
      if (checkmarkTransitionTimer) {
        clearTimeout(checkmarkTransitionTimer)
        checkmarkTransitionTimer = undefined
        if (checkmarkAbortHandler) {
          ctx.toolUseContext.abortController.signal.removeEventListener(
            'abort',
            checkmarkAbortHandler,
          )
          checkmarkAbortHandler = undefined
        }
        ctx.removeFromQueue()
      }
    },
    onAbort() {
      if (!claim()) return
      if (bridgeCallbacks && bridgeRequestId) {
        bridgeCallbacks.sendResponse(bridgeRequestId, {
          behavior: 'deny',
          message: 'User aborted',
        })
        bridgeCallbacks.cancelRequest(bridgeRequestId)
      }
      channelUnsubscribe?.()
      ctx.logCancelled()
      ctx.logDecision(
        { decision: 'reject', source: { type: 'user_abort' } },
        { permissionPromptStartTimeMs },
      )
      resolveOnce(ctx.cancelAndAbort(undefined, true))
    },
    async onAllow(
      updatedInput,
      permissionUpdates: PermissionUpdate[],
      feedback?: string,
      contentBlocks?: ContentBlockParam[],
    ) {
      if (!claim()) return // atomic check-and-mark before await

      if (bridgeCallbacks && bridgeRequestId) {
        bridgeCallbacks.sendResponse(bridgeRequestId, {
          behavior: 'allow',
          updatedInput,
          updatedPermissions: permissionUpdates,
        })
        bridgeCallbacks.cancelRequest(bridgeRequestId)
      }
      channelUnsubscribe?.()

      resolveOnce(
        await ctx.handleUserAllow(
          updatedInput,
          permissionUpdates,
          feedback,
          permissionPromptStartTimeMs,
          contentBlocks,
          result.decisionReason,
        ),
      )
    },
    onReject(feedback?: string, contentBlocks?: ContentBlockParam[]) {
      if (!claim()) return

      if (bridgeCallbacks && bridgeRequestId) {
        bridgeCallbacks.sendResponse(bridgeRequestId, {
          behavior: 'deny',
          message: feedback ?? 'User denied permission',
        })
        bridgeCallbacks.cancelRequest(bridgeRequestId)
      }
      channelUnsubscribe?.()

      ctx.logDecision(
        {
          decision: 'reject',
          source: { type: 'user_reject', hasFeedback: !!feedback },
        },
        { permissionPromptStartTimeMs },
      )
      resolveOnce(ctx.cancelAndAbort(feedback, undefined, contentBlocks))
    },
    async recheckPermission() {
      if (isResolved()) return
      const freshResult = await hasPermissionsToUseTool(
        ctx.tool,
        ctx.input,
        ctx.toolUseContext,
        ctx.assistantMessage,
        ctx.toolUseID,
      )
      if (freshResult.behavior === 'allow') {
        // claim() (atomic check-and-mark), not isResolved() — the async
        // hasPermissionsToUseTool call above opens a window where CCR
        // could have responded in flight. Matches onAllow/onReject/hook
        // paths. cancelRequest tells CCR to dismiss its prompt — without
        // it, the web UI shows a stale prompt for a tool that's already
        // executing (particularly visible when recheck is triggered by
        // a CCR-initiated mode switch, the very case this callback exists
        // for after useReplBridge started calling it).
        if (!claim()) return
        if (bridgeCallbacks && bridgeRequestId) {
          bridgeCallbacks.cancelRequest(bridgeRequestId)
        }
        channelUnsubscribe?.()
        ctx.removeFromQueue()
        ctx.logDecision({ decision: 'accept', source: 'config' })
        resolveOnce(ctx.buildAllow(freshResult.updatedInput ?? ctx.input))
      }
    },
  })

  // Race 4: Bridge permission response from CCR (claude.ai)
  // When the bridge is connected, send the permission request to CCR and
  // subscribe for a response. Whichever side (CLI or CCR) responds first
  // wins via claim().
  //
  // All tools are forwarded — CCR's generic allow/deny modal handles any
  // tool, and can return `updatedInput` when it has a dedicated renderer
  // (e.g. plan edit). Tools whose local dialog injects fields (ReviewArtifact
  // `selected`, AskUserQuestion `answers`) tolerate the field being missing
  // so generic remote approval degrades gracefully instead of throwing.
  if (bridgeCallbacks && bridgeRequestId) {
    bridgeCallbacks.sendRequest(
      bridgeRequestId,
      ctx.tool.name,
      displayInput,
      ctx.toolUseID,
      description,
      result.suggestions,
      result.blockedPath,
    )

    const signal = ctx.toolUseContext.abortController.signal
    const unsubscribe = bridgeCallbacks.onResponse(
      bridgeRequestId,
      response => {
        if (!claim()) return // Local user/hook/classifier already responded
        signal.removeEventListener('abort', unsubscribe)
        clearClassifierChecking(ctx.toolUseID)
        clearClassifierIndicator()
        ctx.removeFromQueue()
        channelUnsubscribe?.()

        if (response.behavior === 'allow') {
          if (response.updatedPermissions?.length) {
            void ctx.persistPermissions(response.updatedPermissions)
          }
          ctx.logDecision(
            {
              decision: 'accept',
              source: {
                type: 'user',
                permanent: !!response.updatedPermissions?.length,
              },
            },
            { permissionPromptStartTimeMs },
          )
          resolveOnce(ctx.buildAllow(response.updatedInput ?? displayInput))
        } else {
          ctx.logDecision(
            {
              decision: 'reject',
              source: {
                type: 'user_reject',
                hasFeedback: !!response.message,
              },
            },
            { permissionPromptStartTimeMs },
          )
          resolveOnce(ctx.cancelAndAbort(response.message))
        }
      },
    )

    signal.addEventListener('abort', unsubscribe, { once: true })
  }

  // Channel permission relay — races alongside the bridge block above. Send a
  // permission prompt to every active channel (Telegram, iMessage, etc.) via
  // its MCP send_message tool, then race the reply against local/bridge/hook/
  // classifier. The inbound "yes abc123" is intercepted in the notification
  // handler (useManageMCPConnections.ts) BEFORE enqueue, so it never reaches
  // Claude as a conversation turn.
  //
  // Unlike the bridge block, this still guards on `requiresUserInteraction` —
  // channel replies are pure yes/no with no `updatedInput` path. In practice
  // the guard is dead code today: all three `requiresUserInteraction` tools
  // (ExitPlanMode, AskUserQuestion, ReviewArtifact) return `isEnabled()===false`
  // when channels are configured, so they never reach this handler.
  //
  // Fire-and-forget send: if callTool fails (channel down, tool missing),
  // the subscription never fires and another racer wins. Graceful degradation
  // — the local dialog is always there as the floor.
  if (
    (feature('KAIROS') || feature('KAIROS_CHANNELS')) &&
    channelCallbacks &&
    !ctx.tool.requiresUserInteraction?.()
  ) {
    const channelRequestId = shortRequestId(ctx.toolUseID)
    const allowedChannels = getAllowedChannels()
    const channelClients = filterPermissionRelayClients(
      ctx.toolUseContext.getAppState().mcp.clients,
      name => findChannelEntry(name, allowedChannels) !== undefined,
    )

    if (channelClients.length > 0) {
      // Outbound is structured too (Kenneth's symmetry ask) — server owns
      // message formatting for its platform (Telegram markdown, iMessage
      // rich text, Discord embed). CC sends the RAW parts; server composes.
      // The old callTool('send_message', {text,content,message}) triple-key
      // hack is gone — no more guessing which arg name each plugin takes.
      const params: ChannelPermissionRequestParams = {
        request_id: channelRequestId,
        tool_name: ctx.tool.name,
        description,
        input_preview: truncateForPreview(displayInput),
      }

      for (const client of channelClients) {
        if (client.type !== 'connected') continue // refine for TS
        void client.client
          .notification({
            method: CHANNEL_PERMISSION_REQUEST_METHOD,
            params,
          })
          .catch(e => {
            logForDebugging(
              `Channel permission_request failed for ${client.name}: ${errorMessage(e)}`,
              { level: 'error' },
            )
          })
      }

      const channelSignal = ctx.toolUseContext.abortController.signal
      // Wrap so BOTH the map delete AND the abort-listener teardown happen
      // at every call site. The 6 channelUnsubscribe?.() sites after local/
      // hook/classifier wins previously only deleted the map entry — the
      // dead closure stayed registered on the session-scoped abort signal
      // until the session ended. Not a functional bug (Map.delete is
      // idempotent), but it held the closure alive.
      const mapUnsub = channelCallbacks.onResponse(
        channelRequestId,
        response => {
          if (!claim()) return // Another racer won
          channelUnsubscribe?.() // both: map delete + listener remove
          clearClassifierChecking(ctx.toolUseID)
          clearClassifierIndicator()
          ctx.removeFromQueue()
          // Bridge is the other remote — tell it we're done.
          if (bridgeCallbacks && bridgeRequestId) {
            bridgeCallbacks.cancelRequest(bridgeRequestId)
          }

          if (response.behavior === 'allow') {
            ctx.logDecision(
              {
                decision: 'accept',
                source: { type: 'user', permanent: false },
              },
              { permissionPromptStartTimeMs },
            )
            resolveOnce(ctx.buildAllow(displayInput))
          } else {
            ctx.logDecision(
              {
                decision: 'reject',
                source: { type: 'user_reject', hasFeedback: false },
              },
              { permissionPromptStartTimeMs },
            )
            resolveOnce(
              ctx.cancelAndAbort(`Denied via channel ${response.fromServer}`),
            )
          }
        },
      )
      channelUnsubscribe = () => {
        mapUnsub()
        channelSignal.removeEventListener('abort', channelUnsubscribe!)
      }

      channelSignal.addEventListener('abort', channelUnsubscribe, {
        once: true,
      })
    }
  }

  // Skip hooks if they were already awaited in the coordinator branch above
  if (!awaitAutomatedChecksBeforeDialog) {
    // Execute PermissionRequest hooks asynchronously
    // If hook returns a decision before user responds, apply it
    void (async () => {
      if (isResolved()) return
      const currentAppState = ctx.toolUseContext.getAppState()
      const hookDecision = await ctx.runHooks(
        currentAppState.toolPermissionContext.mode,
        result.suggestions,
        result.updatedInput,
        permissionPromptStartTimeMs,
      )
      if (!hookDecision || !claim()) return
      if (bridgeCallbacks && bridgeRequestId) {
        bridgeCallbacks.cancelRequest(bridgeRequestId)
      }
      channelUnsubscribe?.()
      ctx.removeFromQueue()
      resolveOnce(hookDecision)
    })()
  }

  // Execute bash classifier check asynchronously (if applicable)
  if (
    feature('BASH_CLASSIFIER') &&
    result.pendingClassifierCheck &&
    ctx.tool.name === BASH_TOOL_NAME &&
    !awaitAutomatedChecksBeforeDialog
  ) {
    // UI indicator for "classifier running" — set here (not in
    // toolExecution.ts) so commands that auto-allow via prefix rules
    // don't flash the indicator for a split second before allow returns.
    setClassifierChecking(ctx.toolUseID)
    void executeAsyncClassifierCheck(
      result.pendingClassifierCheck,
      ctx.toolUseContext.abortController.signal,
      ctx.toolUseContext.options.isNonInteractiveSession,
      {
        shouldContinue: () => !isResolved() && !userInteracted,
        onComplete: () => {
          clearClassifierChecking(ctx.toolUseID)
          clearClassifierIndicator()
        },
        onAllow: decisionReason => {
          if (!claim()) return
          if (bridgeCallbacks && bridgeRequestId) {
            bridgeCallbacks.cancelRequest(bridgeRequestId)
          }
          channelUnsubscribe?.()
          clearClassifierChecking(ctx.toolUseID)

          const matchedRule =
            decisionReason.type === 'classifier'
              ? (decisionReason.reason.match(
                  /^Allowed by prompt rule: "(.+)"$/,
                )?.[1] ?? decisionReason.reason)
              : undefined

          // Show auto-approved transition with dimmed options
          if (feature('TRANSCRIPT_CLASSIFIER')) {
            ctx.updateQueueItem({
              classifierCheckInProgress: false,
              classifierAutoApproved: true,
              classifierMatchedRule: matchedRule,
            })
          }

          if (
            feature('TRANSCRIPT_CLASSIFIER') &&
            decisionReason.type === 'classifier'
          ) {
            if (decisionReason.classifier === 'auto-mode') {
              setYoloClassifierApproval(ctx.toolUseID, decisionReason.reason)
            } else if (matchedRule) {
              setClassifierApproval(ctx.toolUseID, matchedRule)
            }
          }

          ctx.logDecision(
            { decision: 'accept', source: { type: 'classifier' } },
            { permissionPromptStartTimeMs },
          )
          resolveOnce(ctx.buildAllow(ctx.input, { decisionReason }))

          // Keep checkmark visible, then remove dialog.
          // 3s if terminal is focused (user can see it), 1s if not.
          // User can dismiss early with Esc via onDismissCheckmark.
          const signal = ctx.toolUseContext.abortController.signal
          checkmarkAbortHandler = () => {
            if (checkmarkTransitionTimer) {
              clearTimeout(checkmarkTransitionTimer)
              checkmarkTransitionTimer = undefined
              // Sibling Bash error can fire this (StreamingToolExecutor
              // cascades via siblingAbortController) — must drop the
              // cosmetic ✓ dialog or it blocks the next queued item.
              ctx.removeFromQueue()
            }
          }
          const checkmarkMs = getTerminalFocused() ? 3000 : 1000
          checkmarkTransitionTimer = setTimeout(() => {
            checkmarkTransitionTimer = undefined
            if (checkmarkAbortHandler) {
              signal.removeEventListener('abort', checkmarkAbortHandler)
              checkmarkAbortHandler = undefined
            }
            ctx.removeFromQueue()
          }, checkmarkMs)
          signal.addEventListener('abort', checkmarkAbortHandler, {
            once: true,
          })
        },
      },
    ).catch(error => {
      // Log classifier API errors for debugging but don't propagate them as interruptions
      // These errors can be network failures, rate limits, or model issues - not user cancellations
      logForDebugging(`Async classifier check failed: ${errorMessage(error)}`, {
        level: 'error',
      })
    })
  }
}

// --

export { handleInteractivePermission }
export type { InteractivePermissionParams }