[BUG] Plan mode exit → agent launch: denial of ExitPlanMode incorrectly cascades as denial of subsequent Task tool call

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

CLI version: 2.1.59 (behavior not observed in 2.1.37)

Steps to reproduce:

1. Enter plan mode (/plan or via EnterPlanMode)
2. Claude calls ExitPlanMode to request approval
3. User denies ExitPlanMode with feedback: "Good plan, but launch an agent to

handle this instead"

1. Claude immediately calls the Task tool to launch a general-purpose agent
2. User never sees a prompt to approve/deny the Task tool call
3. Claude's context receives a "tool rejected" signal for the Task call — but the

agent appears to have launched and executed work anyway

What I observed:

• I saw a few approval prompts (for sketch edits and git operations) appearing to

come from an agent, with no preceding prompt asking me to approve launching an
agent at all

• After approving those, the agent continued running silently with no visible

progress — I could not distinguish "agent is working" from "agent is stuck"

• Claude reported the Task call as rejected and believed no work had been done

Root cause hypothesis:

When the user denies ExitPlanMode and includes text saying "launch an agent
instead," the CLI may be recycling/cascading that denial event to cover the next
tool call (the Task launch), skipping the approval prompt entirely. This creates
two problems:

1. The agent launches without explicit user approval of the Task tool call
2. Claude's context receives a rejection signal even though the agent is actually

running, causing Claude to incorrectly believe nothing happened

Expected behavior:

• Denial of ExitPlanMode should only apply to ExitPlanMode
• The subsequent Task tool call should surface its own fresh approval prompt
• If the agent is running, Claude's context should not receive a rejection signal

for that call

Additional UX issue (separate but related):
Once a background agent is approved and running, there is no real-time visibility
into its progress in 2.1.59. The UI shows only a spinner with no indication of
which files are being edited or which tool calls are being made. This makes it
impossible to distinguish an active agent from a stalled one.

What Should Happen?

It should be clear when Claude is proposing to launch an agent or not, especially when it is presenting a numbered prompt. If a numbered prompt is rejected with a message, that message should be evaluated and a new prompt or discussion initiated to clarify the user's intended behavior. Sub-agents should not be launched without their work being visible to the main agent and the user.

Error Messages/Logs

Steps to Reproduce

1. Enter plan mode (/plan or via EnterPlanMode)
2. Claude calls ExitPlanMode to request approval
3. User denies ExitPlanMode with feedback: "Good plan, but launch an agent to

handle this instead"

1. Claude immediately calls the Task tool to launch a general-purpose agent
2. User never sees a prompt to approve/deny the Task tool call
3. Claude's context receives a "tool rejected" signal for the Task call — but the

agent appears to have launched and executed work anyway

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.59

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_