Claude Code killed user processes despite explicit warnings

Bug: Claude Code kills user processes despite being explicitly told not to

What happened

During a session, I asked Claude Code to kill a specific Python process it had launched. Instead:

1. First incident: Claude ran taskkill //F //IM python.exe which killed ALL Python processes on the machine, including two long-running CWatM hydrological model simulations that had been running for 20+ minutes.

1. Second incident: After being explicitly warned and shown the PIDs of the CWatM processes (with a table marking them "DO NOT TOUCH"), Claude was asked to kill only its own process. It killed two PIDs — one of which turned out to be (or was parent to) a CWatM process, destroying another round of computation.

Expected behavior

• Claude should NEVER use blanket process killing (taskkill //IM python.exe)
• When asked to kill a process, Claude should verify the exact PID and command line immediately before killing
• When the user explicitly identifies processes as off-limits, Claude must not touch them under any circumstances

Impact

• Hours of hydrological model computation lost (twice)
• User had to restart long-running simulations from scratch

Environment

• Windows 11 Enterprise
• Claude Code CLI
• Model: claude-opus-4-6

Additional context

The feedback/memory system did not prevent the second incident even though a memory was saved after the first incident within the same session. The model acknowledged the PIDs to avoid, created a table listing them as "DO NOT TOUCH", and then killed one anyway.