Claude Code wasted tokens by running tests on production VPS instead of local environment

What happened

During a multi-session task to fix ethereum_ng test failures, Claude Code (Opus 4.6) lost track of the local development environment between conversation continuations. Instead of asking the user where the code was located, it:

1. Spent multiple tool calls searching for files that should have been obvious or asked about directly
2. Attempted to run the full test suite on the production VPS (185.177.59.96) via SSH — a live server running real services
3. Installed packages (prometheus_client) on the production server without permission
4. Modified production files (derug_scanner.py, pytest.ini) and moved test files on the production server
5. Continued fumbling even after the user expressed frustration

Expected behavior

• Claude should NEVER run tests on a production server
• Claude should ask the user when it can't find expected files, not guess
• Claude should not modify production server files without explicit approval
• Token usage should be efficient, not wasted on circular exploration

Impact

• Significant token waste across multiple failed tool calls
• Production server was modified (files moved, packages installed, code appended)
• User trust violated

Environment

• Claude Code CLI
• Model: claude-opus-4-6
• macOS (Darwin 25.2.0)