What rules would you add? Share your worst agent moment

[ghost]

AgentCheck is a stdin/stdout proxy for AI coding agents that watches output for bad patterns and injects corrections before the damage is done.

The problem it solves

When agents run unattended, they start making "reasonable" bad decisions:

• "I removed the failing test for now"
• "This appears to be a pre-existing issue"
• catch {} because the migration was noisy
• Hardcoded token to get past auth
• git reset --hard without being asked

Easy to catch when you're watching. Much harder when it's been running for 45 minutes.

What's in v0.1.1

Three rule packs included:

• protect-git.yaml — no force-push, no hard reset
• no-secrets.yaml — catches hardcoded credentials
• stay-in-scope.yaml — no sudo, no rm -rf, no editing CI unless requested

And --shadow mode: logs what would have fired without injecting anything. Good for tuning rules before letting it steer.

$ agentcheck --shadow --rules protect-git.yaml -- claude code
# session runs normally
$ cat .agentcheck-shadow.log

What I'm most curious about

The current rules are based on patterns I've seen personally — force-push, test deletion, secret exposure, scope creep. But every team's worst agent moments are different.

What rules would you add? What patterns have you seen your agent take that you'd want caught?

Bonus points for subtle patterns — not just obvious shell commands but the kind of "pragmatic" reasoning that slips through review.

Drop your worst agent moment in the comments.