[BUG] Thinking stalls exceeding conversation cache TTL

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?

When Claude Code uses thinking (medium), the thinking process stalls repeatedly at round token counts (observed: 33.0k, 65.0k, 97.0k tokens). Each stall lasts long enough (more than 5 minutes) that the conversation cache expires. After a long wait I see tokens being spent again, but notice >20% jump in token usage for my 5-hour session! This makes using Claude Code unusable on Pro for any non-trivial task where you wait for Claude Code to finish its thinking process without interrupting.

I tried waiting it out but got hit with API Error: Claude's response exceeded the 32000 output token maximum. To configure this behaviour, set the CLAUDE_CODE_MAX_OUTPUT_TOKENS environment variable

I've been experiencing the same issue in Claude.ai Code on desktop where my usage for the 5 hour interval is used up without a single line of code output. (Not exaggerating here)

What Should Happen?

Thinking should progress continuously or resumes within cache TTL window.

Error Messages/Logs

Steps to Reproduce

1. Start Claude Code CLI in terminal
2. Give a complex implementation task
3. Observe thinking token counter pausing at round numbers
4. Stall duration exceeds cache TTL, see hits on session token usage

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.123

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Other

Additional Information

_No response_