[BUG] System prompts and tool descriptions cause frequent stoppages, time estimates

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?

On long-running tasks, both Sonnet 4.5 and Opus 4.5 exhibit the following:

• Stopping when they appear concerned about remaining context space, providing an unwanted summary to the user and often a question amounting to "should i carry on?"
• Persistent time-based estimates despite a) no such request from the user and b) no grounding in any reality such as team size, human vs agent resources, business priorities, etc

In both cases, strenuous instructions to the contrary in claude.md are not given sufficient weight to overrule whatever is in the system prompts that causes this behaviour.

I'm classifying this as a bug because:

• As advertised by Anthropic, Claude is intended to optionally be a long-running autonomous tool. When a plan is researched and accepted by the user, and if the user has even expressly said "don't stop until all phases are complete", it is incorrect for Claude to have cause to unilaterally stop for any other reason than it hit a technical brick wall
• Many times the presence of time estimates also highlight that the plans generated by Claude are skewed towards options that take less time or are subjectively less complicated. This undermines the quality of the plan and research itself, and it's objectively a waste of billable tokens to have to converse back and forth with the tool in order to nudge it back into the direction already mandated in the claude.md file

What Should Happen?

Instructions in the only configurable area - claude.md - should be adhered to over suggestions in system prompts

Error Messages/Logs

Steps to Reproduce

Ask Claude Code to make any plan and it'll tell you how many hours days or weeks it'll take, anchored in anything but reality

Claude Model

None

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

Any

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_