content_block_delta event not deserialized correctly during streaming

[sahibpandori]

We got an exception that the following event couldn't be deserialized even though it looks correct.

Event -

{
  delta: {
    text: "something something",
    type: "text_delta"
  },
    index: 0,
    type: "content_block_delta"
}

Error

Unexpected event runtime type, after deserialising twice - {'type': 'content_block_delta', 'index': 0, 'delta': {'type': 'text_delta', 'text': 'something something'}} - <class 'dict'>

[RobertCraigie]

@sahibpandori thanks for the report, can you share any more details about your setup? e.g. dependencies, OS, how you're instantiating the client, if you're doing any monkeypaching

[RobertCraigie]

@sahibpandori did you ever run into this again? we've had a very hard time reproducing

[Ashutosh0x]

@RobertCraigie Hey! I came across this issue while working on a streaming project and wanted to help out.

I've been digging into the codebase - specifically the �ccumulate_event() function in _messages.py and the construct_type_unchecked() path in _models.py.

Looking at the error, it seems like the discriminated union resolution for RawMessageStreamEvent is failing silently and returning the raw dict instead of a proper Pydantic model. The construct_type() function falls back to returning the original value when it can't resolve the discriminator mapping.

I'd love to try and reproduce this. A few questions:

1. Was this happening with a specific model (Opus, Sonnet, Haiku)?
2. Any chance it's related to high-throughput/concurrent streaming scenarios?
3. I noticed this was from a Vertex AI context - is that potentially relevant?

Happy to set up some reproduction tests if you have any leads on what conditions might trigger this. Also, if there's appetite for a PR that adds more defensive handling in the deserialization path, I'd be glad to take a stab at it.