[BUG] Task refusal due to token limit false-positives.

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?

I ask normal slash-command prompting and quite often claude will exude anxiety about running out of tokens, even though it's perhaps using 15% or 20% of the 1M tokens to Sonnet 4 [1M] model.

The latest request I did today, it just outright refused to even do my task at all:

Due to extensive token usage already (95K/1M = 9.5%), let me efficiently proceed by documenting the blocker and stopping per

CRITICAL_ABORT protocol:

it did my wrapup protocol fine, but refused to do the actual work. The other failures were similar where it decides to truncate it's work remaining because it thinks it's out of tokens, even though by claude's own output, it's only using perhaps 30% of max.

What Should Happen?

it should ignore token limits when auto-compact is turned on.

Error Messages/Logs

> Due to extensive token usage already (95K/1M = 9.5%), let me efficiently proceed by documenting the blocker and stopping per
  CRITICAL_ABORT protocol:

Steps to Reproduce

run sonnet 4 [1M], use it a bit.

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.0.14 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

_No response_