Bring the system prompt back PLEASE!!!

What version of Kimi Code CLI is running?

kimi, version 1.41.0

Which open platform/subscription were you using?

Kimi code

Which model were you using?

2.6

What platform is your computer?

Linux x64 ubuntu 24.04, macOS 26

What issue are you seeing?

the Kimi code completely forget about the coding rule, skill both in project and global.
WTF???? I regret purchase the kimi coding plan subscription. for what? because kimi changes my code without even asking.

why there is no system prompt level coding rules??
don't tell me "We have skills". your kimi code will forget about your skills completely after 10 conversations and getting changes things crazy without permission, without asking. it changes the code using shell. WHY?

PLEASE. bring the system prompt back. I promise, I will not break your kimi code. just I want kimi code to follow my coding rules. asking things before taking any code action. that is it.

and there are another issue: after approve code changes, there is no way to revoke the permission... What the hell are you thinking about? are your guys use your things during coding??

What steps can reproduce the bug?

tell your kimi code some coding rules and ask to do something just after 10 conversations.
the kimi code completely forget the global skills.

What is the expected behavior?

I WANT KIMI CODE TO:

1. ALWAYS ask me anything if there are anything unclear or there is no very precise plan instead of random guessing what to do.
2. ALWAYS share the plan (or ask) before taking ANY code changes. even between frequent tool calls. sometimes kimi code execute 30+ shell command without a single word. I don't know what the hell is going on in the background. the crazy tool calls are continued even after Ctrl + C and ask "hello?".

Additional information

No response

No file or test is named. Start by reproducing the reported loss of global and project coding rules after roughly 10 conversations, then trace system-prompt or skill handling and the approval, shell-execution, and cancellation paths. Done should include persistent rule adherence, a visible plan and confirmation before code actions, and a way to revoke permission or stop ongoing tool calls.