[BUG]

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?

Title: Claude enters plan mode unnecessarily, burns 200K+ tokens over 20 minutes with zero output

Description:
After I verbally approved an implementation approach in conversation, Claude called EnterPlanMode and spent over 20 minutes silently running expensive subagent research (~73K + ~66K tokens) and repeated file reads, duplicating
information already available in the conversation context. All ~200,000–300,000 tokens were consumed during this silent phase before I ever had a chance to intervene. No progress feedback. No prompts. No output.

When I finally pressed escape to ask what was happening, Claude then took 4 attempts to call ExitPlanMode, burning more tokens on each try. The actual implementation (creating a new file, refactoring a function, testing) took ~5
minutes once plan mode was exited — the entire 20 minutes of plan mode was pure waste.

Timeline:

1. User verbally approved the implementation approach in conversation
2. Claude called EnterPlanMode instead of just coding
3. Claude silently launched Explore agent (~73,000 tokens)
4. Claude silently launched Plan agent (~66,000 tokens)
5. Claude read multiple files already available in context
6. Claude wrote a plan file — 20+ minutes elapsed, zero code written
7. User pressed escape to ask what was taking so long
8. Claude called ExitPlanMode — rejected/interrupted 3 times as user tried to get answers
9. 4th ExitPlanMode attempt finally approved
10. Actual implementation completed in ~5 minutes

Problems:

1. Entered plan mode when user had already verbally approved the approach — completely unnecessary
2. 20+ minutes of silent expensive research with zero feedback to user
3. ~200,000–300,000 tokens consumed with zero code written
4. Subagent research duplicated information already in the conversation context
5. No progress indicator — user had no idea what was happening or how long it would take
6. 4 attempts to exit plan mode, each burning full context resends
7. ExitPlanMode approval prompt unclear — user didn't know what to approve or that they were blocking the exit
8. No way for user to cancel plan mode or see token consumption in real time

Expected behaviour:

• Do not enter plan mode when user has already approved the approach in conversation
• Show progress feedback and estimated time during long-running operations
• Do not run expensive research agents for information already available in context
• Make ExitPlanMode approval prompt obvious and explain what it does
• First rejection should trigger a clear explanation, not a silent retry loop
• Provide real-time token usage visibility so users can intervene before costs escalate
• Allow users to cancel plan mode without going through the approval flow

Token cost: Estimated 200,000–300,000 tokens wasted over 20+ minutes with zero productive output.

Environment: Claude Code CLI, Opus 4.6, Windows 10 Pro

What Should Happen?

not what happened here

Error Messages/Logs

Steps to Reproduce

Steps to reproduce:
This is not a deterministic bug — it depends on Claude choosing to enter plan mode. However the conditions that triggered it:

1. Have a long conversation with substantial context already built up
2. Discuss and verbally agree on an implementation approach
3. Tell Claude to go ahead and implement it (e.g. "ok lets do it")
4. Claude may call EnterPlanMode instead of coding, despite the approach already being agreed
5. Once in plan mode, Claude launches expensive research agents and reads files silently with no user feedback
6. User has no visibility into what's happening or how many tokens are being consumed
7. User's only option is to wait or press escape — pressing escape triggers a confusing ExitPlanMode approval loop

The core issue is systemic, not a specific reproduction case:

• Plan mode has no progress feedback
• Plan mode has no token budget or time limit
• There is no user-initiated way to cancel plan mode
• ExitPlanMode approval UX is unclear
• Claude can enter plan mode even when it's unnecessary

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

Claude Code version 2.1.58

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

i want my tokens back!