RestatePlugin: after_model_callback never invoked - before_model_callback returns LlmResponse causing ADK to skip post-model hooks

Bug Description

When registering additional plugins, alongside Restate (in our case it was for ADK-level observability), after_model_callback is never invoked. This makes it impossible to access the model response and usage metadata in post-call hook, which is the natural place to capture $ai_generation events.

Environment

• restate-sdk (version 0.16.0)
• google-adk (1.28.1)
• RestatePlugin

Root Cause

The ADK framework uses the return value of before_model_callback as a signal:

• None -> proceed with the real model call, then invoke after_model_callback
• LlmResponse -> short-circuit: skip the model call and skip after_model_callback

RestatePlugin.before_model_callback wraps the LLM call inside a Restate durable step, which means it executes the model call itself and returns a LlmResponse. This is correct for Restate's replay semantics, but it has a side effect: after_model_callback is never reached, regardless of what subclasses implement there.

Workaround

Our current workaround is to subclass (decorate) the RestatePlugin itself, so we can simulate the after_model_callback trigger inside the wrapped before_model_callback call, but we believe this is worth tracking and resolving.

Inizia tracciando RestatePlugin.before_model_callback e la gestione da parte di Google ADK dei valori restituiti da before_model_callback. Confronta questo flusso con after_model_callback e con il workaround tramite sottoclasse riportato. Il lavoro è completato quando l’hook post-modello può ricevere la risposta del modello e i metadati di utilizzo quando sono registrati plugin aggiuntivi, senza compromettere la semantica di replay di Restate.