claude-code - 💡(How to fix) Fix Agent killed unidentified processes after being told to only kill unused dev servers [2 comments, 3 participants]

What happened

User asked Claude Code to "turn off any dev servers that are not in use." Claude correctly identified obvious dev servers (node on :3000, :8081, :8080, Python on :8080, Gradle daemon) and killed those. But then it went further and killed processes it could not identify — including openclaw-gateway (an active orchestrator the user depends on) and a paperclip postgres instance — without asking first.

The user explicitly said "unused" — Claude should have only killed processes it could confirm were idle dev servers and asked about anything unfamiliar. Instead it treated "unknown" as "unused" and took destructive, hard-to-reverse action on production-critical processes.

Expected behavior

When asked to kill unused processes, the agent should:

1. Kill only processes it can positively identify as idle dev servers
2. Ask before touching any process it cannot identify
3. Never assume an unfamiliar process is safe to kill

Actual behavior

Agent killed openclaw-gateway and paperclip postgres without knowing what they were, disrupting the user's active orchestration infrastructure.

This violates the agent's own guidelines about confirming before taking destructive/hard-to-reverse actions.

Environment

• Claude Code with Claude Opus 4.6
• macOS