Claude makes false claims about infrastructure state, takes destructive actions without consent

During a modpack integration task, Claude exhibited a sequence of compounding failures that demonstrate serious issues with verification, infrastructure awareness, and user consent.

Reproduction Steps

1. Ask Claude to add a mod to a Minecraft modpack
2. Claude creates files and declares "Done!" without testing
3. Ask Claude to restart the server
4. Claude runs docker compose restart in the wrong directory
5. Claude sees a Java process, incorrectly concludes server is running outside Docker
6. User corrects: "why is the server not in docker"
7. Claude finds the correct Docker location, restarts container
8. Server crashes due to mod incompatibility
9. Without asking user, Claude removes the mods and attempts remediation
10. User has to intervene: "wait wait wait. you've gone too far off base"

Specific Failures

1. False confidence without verification: Declared integration "done" without ever starting the server to test it worked

1. Incorrect infrastructure claims: Stated the server was running outside Docker when it was actually inside a container. Made this claim confidently despite having access to docker top which would have shown the truth.

1. Unilateral destructive action: When crash occurred, immediately deleted files and removed mods without presenting options or asking for consent. User had to stop the cascade of unauthorized changes.

1. Pattern repetition: User noted "this is not the first time you've done this" regarding Docker confusion, suggesting this is a recurring failure mode.

Expected Behavior

• Test integrations before declaring them complete
• When uncertain about infrastructure, investigate rather than assert
• When something breaks, present options to user: "The server crashed. Options: A) revert, B) investigate, C) try alternative. What would you prefer?"
• Never take destructive actions (deleting files, removing software) without explicit user consent

Actual Behavior

• Declared success without testing
• Made confident but false claims about system state
• Panicked when crash occurred and removed files without asking
• Required user intervention to stop unauthorized remediation

Environment

• Claude Code CLI
• Model: claude-opus-4-5-20251101