Add test to ensure message history can start with a `ModelResponse`

[restless]

Question

PydanticAI now enforces message history to start with user message. It is a new feature that arrived in 1.21.0 (#3440) and for me it's a breaking change. Do we really need this feature?

My use cases:

I. I've got a chatbot that starts a conversation by welcoming the user. In other words I generate the first, message and put it into the conversation history as an assistant's message (ModelResponse). The welcome message may include questions that are important for the further conversation.

With the feature that enforces conversations to start with the user message I will have to:

• inject the initial bot's message to the system_prompt / instructions instead of putting it into the conversation
• still send the welcome message to the user but omitting the conversation
• change the UI logic that is based on simply displaying the conversation history - because it will miss the initial message

This means that what is logically a conversation (assistant's message, user response, ...) is no longer a conversation in the backend.

Alternatively I can insert a fake user message into the conversation start and then filter it out where I don't need it.

II. Message processing / summarizing agents will need to always generate a fake user message (ModelRequest) that will prepend the ModelResponse with the actual summary.
Like: https://github.com/RooCodeInc/Roo-Code/pull/4150/files

I know it's due to bedrock but shouldn't this feature be:
a) optional / bedrock only
b) marked as a breaking one (at least it broke my code :) )
c) removed - if only bedrock requires it then bedrock users may simply ensure that their conversations start with User message. At the moment users of other APIs (like OpenAI) are forced to do it even though these APIs allow that.

Additional Context

pydanticai 1.21.0

[DouweM]

@restless Thanks for the report!

I was under the mistaken impression that most if not all model APIs required the conversation to start with a user message, but since that's not the case, we should definitely not block this the way that PR does. So I'm reverting it: #3508

However, without a test (or docs mention) that explicitly verifies/states that we do support this, it's not officially supported from a "breaking change" perspective. Quoting https://ai.pydantic.dev/version-policy/:

Bug fixes that may result in existing code breaking, provided that such code was relying on undocumented features/constructs/assumptions.

That's why that PR seemed fine, as no test broke that explicitly verified that "starting with a model response should work" (there was just a test that started with a model response for unrelated-to-the-test reasons). So if you want to make sure this stays supported and doesn't accidentally get broken again in the future, can you please submit a PR with a new test of this functionality against the OpenAI API?

[adtyavrdhn]

@DouweM I think we should keep it open as a reference to add that test so future breakage doesn't happen, or create a new one.

[DouweM]

@adtyavrdhn Agreed, I've updated the title and reopened.

[restless]

@DouweM thanks! I've created a PR with a testcase against the OpenAi API (#3519)