Four guardrail patterns that fit before_tool_callback (free MIT repo)

[stackyardweekstart]

This is a small MIT repo we maintain (disclosure: it's ours, from Stackyard). It collects ten guardrail patterns for agents that touch real files, money or APIs, and several map neatly onto a before_tool_callback in ADK:

• Ask before you delete. Deletes and payments always wait for a person, whatever the model decided.
• When in doubt, stop. If a tool call has a field or permission the agent can't resolve, block it instead of guessing.
• Retry with a key. Each real-world write gets one idempotency key that's reused on every retry, and a changed payload under the same key raises a conflict. There are regression tests for that case.
• Session budget guard. A hard token ceiling per run that ends it with a plain message.

They're patterns to read and adapt, not a library: https://github.com/stackyardweekstart/weekstart-starter-pack

If you've wired something like the delete gate into ADK callbacks, I'd like to know where you put the check.