[BUG] Nested Plan Mode in Task Tool Creates Problematic Control Flow

Environment

• Platform (select one):
• [x] Anthropic API
• [ ] AWS Bedrock
• [ ] Google Vertex AI
• [ ] Other: <!-- specify -->
• Claude CLI version: 1.0.35
• Operating System: macOS Darwin 24.3.0
• Terminal: iTerm2

## Bug Description
When the parent assistant is in plan mode and invokes the Task tool, the
Task sub-agent inherits plan mode and has access to exit_plan_mode.
This creates nested plan modes with problematic control flow - approved
plans may execute in the sub-agent scope without returning control to the
parent, while denied plans waste all sub-agent work without returning
results.

## Steps to Reproduce

1. Enter plan mode (e.g., type "ultrathink" or another plan mode trigger)
2. Invoke the Task tool with any request
3. Task sub-agent performs research and calls exit_plan_mode
4. A plan modal appears to the user
5. Approve or deny the plan
6. Observe the problematic behavior (implementation in wrong scope or

total cancellation)

## Expected Behavior

• Task tool should complete its work and return results to the parent

assistant

• Parent assistant should remain in control of plan presentation and

implementation

• Sub-agents should not have independent access to exit_plan_mode

## Actual Behavior

• Task sub-agent inherits plan mode and can call exit_plan_mode

independently

• If plan is approved: implementation may occur within Task tool scope,

parent never regains control

• If plan is denied: Task tool is cancelled, parent receives only an

error message, all sub-agent work is lost

## Additional Context

• This is a known issue documented in user CLAUDE.md files as rule: "Do

not use Task tool when in plan mode, they don't work together"

• The issue persists after v1.0.33 which included "plan mode

improvements"

• Root cause appears to be plan mode inheritance in nested agent contexts
• Creates confusing UX where users may not know which agent's plan

they're approving