openclaw - 💡(How to fix) Fix [Feature]: Add a blocking pre-execution hook to prevent unsafe model/tool actions [1 participants]

Summary

Introduce a hook that runs right before a model or tool executes, enabling final validation and the ability to block harmful actions.

Problem to solve

OpenClaw lacks a critical safety checkpoint: the capability to stop a dangerous action at the last responsible moment.

Currently, before_tool_call triggers too early—before we know the tool’s final inputs—making meaningful risk assessment nearly impossible. after_tool_call, on the other hand, is observation-only and executes once it’s too late to intervene. That’s fine for auditing, but useless for preventing damage.

For real-world scenarios—agents touching file systems, hitting external APIs, running shell commands—this gap is alarming. In sensitive environments or when dealing with untrusted inputs, we need a synchronous (or asynchronous) blocking hook with full context right before execution.

Proposed solution

A pre_execution or validate_action hook—able to inspect resolved arguments and issue a definitive go/no-go—would close this safety gap. It wouldn’t replace existing hooks, but it would give teams the final decision point they need.

Alternatives considered

No response

Impact

This is urgent. As automation grows more powerful, the cost of a single misfire skyrockets. Without a final validation layer that can say “stop” once all parameters are known, we’re effectively running without brakes.

Evidence/examples

No response

Additional information

No response