@albertvillanova one potential difficulty : many people will have step checks that rely on particular attributes of ActionStep, such as the observations, so this would break their workflows by trying to use attributes that do not exist on PlanningStep. Maybe we should differentiate callbacks for planning and for action steps.

@aymeric-roucher, thanks for your comment! You're absolutely right about the potential issue with the callbacks assuming just ActionStep.

To address this, I have tried to implement a typed callback registry:

• this allows registering callbacks per step type

Benefits:

• Backward compatible: Existing action-step callbacks remain unaffected.
• Flexible: Advanced users can register callbacks for planning, action, or all steps.
• Type-safe and extensible: Easy to add support for new step types in the future without modifying core agent logic.
• Encourages separation of concerns: Callbacks for planning and action steps remain clearly scoped.

This also removes the need to guess whether a callback is safe to run after planning, since users explicitly declare the expected step type when registering the callback.

Let me know what you think.

I really like the solution you propose @albertvillanova!