I think we also need to consider the case if c.parsed_repr is a string (I believe that's what a MOT will return after a completed CBlock action (?) )

quick way to test:

mot = ModelOutputThunk(value='reply text')
mot.parsed_repr = 'reply text'
as_generic_chat_history(ChatContext().add(Message('user','hi')).add(mot))
# → WARNING: Unknown component type str ...

Yes, it was an issue. I fix it. Thanks!

should we add a test case for it too?

Minor / style: Small typing issue here — both _default_formatter and the formatter parameter use Any, which the project convention asks us to avoid unless an external library forces it (nothing does here). object is the right type: it accepts everything, is covariant, and keeps mypy happy with callers that pass typed formatters.

def _default_formatter(obj: object) -> str: ...
formatter: Callable[[object], str] | None = None

The test helpers already use obj: object, so it is just the library code that needs updating.

Two things worth noting on this line — one minor, one more significant.

Minor / style: When parsed_repr is a dict or similar, the default formatter will log "Unknown component type dict" — but dict is not an unexpected component here, it is a parsed form of a ModelOutputThunk we already recognise. That message will be confusing to anyone monitoring logs. A dedicated log line would be clearer:

_logger.warning(
    "ModelOutputThunk.parsed_repr is %s, not a Message; falling back to value.",
    type(c.parsed_repr).__name__,
)

More significant: If both parsed_repr and value are None (a ModelOutputThunk that has not been evaluated yet), str(None) silently produces the literal string "None" in the chat history with no warning logged. A backend receiving that will have no indication anything went wrong. Worth guarding explicitly:

if c.value is None:
    raise ValueError("ModelOutputThunk has no value and no parsed_repr — was it evaluated?")

More significant: One tricky edge case: ImageBlock inherits from CBlock (see core/base.py:79), so it will fall into this arm and produce Message(role="user", content=str(image_block)) — meaning a raw base64 PNG string ends up in the chat history as a text message. This is a reachable code path if a context with image content is passed in, and the corruption would be invisible downstream.

A simple guard fixes it:

case CBlock():
    if type(c) is not CBlock:
        return Message(role="user", content=formatter(c))
    return Message(role="user", content=str(c))

That way any CBlock subclasses not explicitly handled go through the formatter instead.