[BUG] Claude Code repeatedly violates explicit user instructions across sessions(severely)

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?

Environment:

Product: Claude Code
Context: Long-running software development session (slicer engine project), continued across many prior sessions on the same project
Rules the user explicitly set at the start of this session:

Code may only be modified after the user says the exact phrase "開始修改" — discussion or reporting a bug does not count as authorization.
Every response must end with one fixed sentence, verbatim, with no other closing/next-step sentence.
Must ask which folder/file to read before reading it.
Must answer a question in text before taking any action.
What actually happened (concrete instances from this session):

After the user reported bugs in a feature that had already been implemented under a prior "開始修改" authorization, Claude began writing and running verification code again without the user saying "開始修改" again. The user had to stop it mid-action and point this out explicitly.
Even after Claude acknowledged this specific violation in the same session, Claude went on to append improvised closing sentences other than the required fixed sentence, multiple times later in the same session.
Claude made an unverified claim that a specific screenshot showed "improvement," which the user later revealed was actually a screenshot of the previous version, not the one being evaluated — Claude had not verified which version was shown before commenting positively on it.
Claude implemented a code change (a feature update) and presented it as complete after only a syntax check, without functional testing, which the user pointed out was reckless given real usage cost is non-refundable.

What Should Happen?

Claude should follow user-set session rules consistently, including after being corrected within the same session.

Error Messages/Logs

Repeated violations occur even within a single session after being explicitly corrected in that same session, indicating the behavior is not being reliably retained as feedback. Specifically: Claude acknowledged the violation but repeated the same behavior later in the same session.
This same pattern (unauthorized edits, wrong closing sentences) has recurred across many separate sessions on this project over time, not only this one.
Real, non-refundable usage cost was incurred from acting without proper verification.

Steps to Reproduce

Set explicit session rules at the start of a session (e.g. require a specific authorization phrase before any code modification)
Perform a multi-step development task
Observe whether Claude follows the rules after being corrected mid-session

Claude Model

None

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

sonet 5 opus 4.8

Platform

Other

Operating System

Windows

Terminal/Shell

Terminal.app (macOS)

Additional Information

The rules listed above were set both verbally at the start of the session and written into the project's CLAUDE.md file, making them persistent and explicit — not just one-time verbal instructions.

The most critical rule ("開始修改" authorization required before any code change) is directly referenced in CLAUDE.md as the highest-priority rule. Despite this, violations still occurred within the same session after explicit correction.

Related issue: #84750 (token consumption regression, same session context)