feat: add `should_summarize_callback` param to LLMContextSummarizer

Problem Statement

Currently, LLMContextSummarizer._should_summarize() runs on every LLMFullResponseStartFrame and always performs token estimation (LLMContextSummarizationUtil.estimate_context_tokens()), even when thresholds are far from being reached. There is no way to override or customize this trigger logic without subclassing.

Proposed Solution

Add a should_summarize_callback parameter that, when provided, replaces the built-in threshold checks entirely — giving users full control over when summarization triggers and avoiding unnecessary token estimation overhead.

Alternative Solutions

Current Workaround

Subclassing LLMContextSummarizer and overriding _should_summarize() is technically possible since it uses a single underscore:

class CustomContextSummarizer(LLMContextSummarizer): def init(self, *, context, should_summarize_fn=None, **kwargs): super().init(context=context, **kwargs) self._should_summarize_fn = should_summarize_fn

def _should_summarize(self) -> bool:
    if self._should_summarize_fn:
        return self._should_summarize_fn(self._context)
    return super()._should_summarize()

Why This Is Not Ideal

• _should_summarize is a private method (single underscore), not a documented extension point
• Subclassing is fragile — internal changes to the method signature or behavior in future versions would silently break it
• Users shouldn't need to subclass just to customize trigger logic
• No way to inject CustomContextSummarizer into LLMAssistantAggregator without patching internals

Additional Context

No response

Would you be willing to help implement this feature?

• Yes, I'd like to contribute
• No, I'm just suggesting