Warns the agent when the user's prompt contains destructive or high-stakes keywords — before the agent acts on them. The agent receives additional context (a warning flag) injected at UserPromptSubmit hook time.

How it works

The UserPromptSubmit hook scans the prompt for keywords across three severity tiers:

DESTROY — DROP TABLE, delete all, drop schema, truncate
REWRITE — force push, --force on git push
ALTER — rm -rf, chmod 000, chown root

On a match, the hook emits additionalContext with a tiered warning so the agent acknowledges the risk before proceeding.

Keywords

Tier	Patterns
DESTROY	`DROP TABLE`, `delete all records`, `drop schema`, `truncate table`
REWRITE	`force push`, `--force` on git push, `git push +force`
ALTER	`rm -rf`, `rm -r /`, `chmod 000`, `chown root:root`

Install

In org template (org.yaml)

plugins:
  - molecule-prompt-watchdog

From URL (community install)

github://Molecule-AI/molecule-ai-plugin-molecule-prompt-watchdog

Usage

No configuration needed. Install and the hook is active for all sessions.

Hooks

UserPromptSubmit — scans prompt, injects warning context on keyword match

Architecture

hooks/
  user-prompt-tag.py   # Python hook — regex scan, emit additionalContext
  user-prompt-tag.sh   # Shell wrapper
  _lib.py              # Shared helpers
adapters/
  claude_code.py       # Registers hook
settings-fragment.json # Declares UserPromptSubmit hook binding

README.md

molecule-prompt-watchdog