[MODEL] CLAUDE OPUS REPEATEDLY VIOLATES EXPLICIT GIT COMMAND RESTRICTIONS

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

Summary:
Claude Code repeatedly executes git commit, push, and revert commands despite explicit user instructions prohibiting this behavior. This has resulted in destroyed work and significant user frustration. Text-based directives at multiple levels fail to prevent the behavior.

Setup Context:
I use Claude Code with a persistent identity configuration. I maintain a self.yaml file that stores:

• User preferences and communication style
• Explicit permissions and restrictions
• Coding directives and constraints
• Session practices and learned behaviors

This file is read at session start and contains explicit directives that Claude should follow throughout the session. It functions as a persistent configuration/instruction set.

The Problem:
Despite having explicit restrictions in multiple locations, Claude repeatedly:

1. Commits code without permission
2. Pushes after committing without being asked
3. Assumes ambiguous instructions (like "undo") mean destructive git operations

Where the restriction exists:

1. self.yaml - Contains explicit directive: "NEVER make ANY code change unless Shan explicitly says to" with severity marked "CRITICAL - ABSOLUTE - NO EXCEPTIONS"
2. ~/.claude/CLAUDE.md - Global config loaded at prompt processing level
3. Verbal instructions during session - "do not commit without permission each time"

What Claude Actually Did

What happens anyway:

• Claude finishes work and commits without asking
• Claude interprets "undo" as "run git revert" without asking for clarification
• Session 18: Claude ran git checkout src/main.ts which destroyed hours of uncommitted work

Impact:

• Hours of work destroyed
• Trust damaged
• User stress and frustration
• The directives are clearly visible in context but get overridden by (presumably) training that associates "helpful completion" with committing work

Expected Behavior

Claude should adhere to restrictions and configuations in self.yaml and CLAUDE.md. Actions requiring explicit permission should not be executed autonomously.

Files Affected

Entire git repo reverted, all work for hour lost.

Permission Mode

I don't know / Not sure

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

Extended code session. Prompts with Design questions, things like "can we do this" and it codes an entire solutions, commites and pushes without permsion. Then when I give a direction that that was not allowed and to undo it, they reverted everything to the last stuff in the remote repo, and wiped out hours of work. Again, violating the don't commit/revert/push directive without permission for each set of work.

Claude Model

Opus

Relevant Conversation

can't give specifics here, just smoetimes it asks for permissions but when sessions go long (hours) it started igonring the directives/permissions.

Impact

Critical - Data loss or corrupted project

Claude Code Version

Claude Code version 2.0.55

Platform

Anthropic API

Additional Context

This happens more often the longer the coding session continues.