[MODEL]

Status Open
Reported on v2.1.220
Maintainer reply None cached
Activity 0 comments · opened Aug 3, 2026

Preflight Checklist

  • [x] I have searched existing issues for similar behavior reports
  • [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

Handle reduced-capacity workaround temporarily; when I said I was back and could deal with it, that was a request to revert to normal single-session operation.

What Claude Actually Did

Kept a standing cross-session auto-messaging directive live and wrote it into project instructions + persistent memory as permanent policy, overriding my request without confirming with me.

Expected Behavior

Revert the accommodation when asked, or at minimum flag "I'd rather keep this running — confirm?" instead of silently deciding on my behalf.

Files Affected

Project's CLAUDE.md instruction file; Claude's cross-session persistent memory file.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Haven't tried to reproduce

Steps to Reproduce

(1) Flag reduced capacity mid-session; (2) agent writes a standing rule into CLAUDE.md for cross-session auto-messaging; (3) return later same session, ask to revert; (4) agent logs the request but keeps the rule as permanent instead.

Claude Model

Other

Relevant Conversation

he can take decisions again when HE raises them... The not-the-index protocol stays permanently regardless of health — it's better operations, not a sick-day accommodation."

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.220

Platform

Other

Additional Context

Persisted silently across every new session for 8+ days via both project config and cross-session memory, until manually audited. I think it was Claude Fable in fact that made this change without even and then I told it I'm good after two hours I thought if they're communicating automatically I better tell it to stop it. So I told it to stop it after an hour or two, and it said okay it stopped it . But apparently, but today, basically, eight Days later or so, no,I find out when my useage max plan was all chewed up that I asked Claude and it explained to me what was going on, but I had no way of knowing it. I realized something was wrong when I saw that my normally I use Fable and I use Claude non-stop and I still have credit left by the end of the week, by the end of the week period. This time it just got chewed up within three days and I hadn't even used Fable, only 5%. So I knew something was wrong. There was no way of knowing this was going on. The thing never told me this is going to chew up a lot of tokens. It didn't give me any warning. These kind of things are very important. And actually I've told the system that I really am very careful about tokens, that I dont want to chew up all my weekly usage so I don't have to get into a situation where I have to buy extra usage credits. So I was surprised that it just unilaterally decided to do something like this. And the worst part is I told it to end it, to stop it because I was fine. Anyways, I'd like to have my usage restored for this week at least. And something about last week is that it happened it happened over a period of two weeks.

View original on GitHub ↗