Add from Agui method #3068
Add from Agui method #3068
Conversation
| messages.append( | ||
| AssistantMessage( | ||
| id=str(uuid.uuid4()), | ||
| content=' '.join(content_parts) if content_parts else None, |
There was a problem hiding this comment.
This isn't quite right: consecutive text parts should be concatenated without a space in between, and if there are parts in between (like built-in tool calls), the text to either side should be joined by \n\n. See the example here: https://github.com/pydantic/pydantic-ai/pull/2970/files#diff-2eb561c8eaa8a723f1017556cce8006c42e504997c187b8b394b5e8634f91283R1148
| # Create separate ToolMessages for builtin tool returns | ||
| for builtin_return in builtin_returns: | ||
| prefixed_id = ( | ||
| f'{_BUILTIN_TOOL_CALL_ID_PREFIX}|{builtin_return.provider_name or ""}|{builtin_return.tool_call_id}' |
There was a problem hiding this comment.
Use a helper method as I suggested above
| tool_call_id=prefixed_id, | ||
| content=builtin_return.content | ||
| if isinstance(builtin_return.content, str) | ||
| else str(builtin_return.content), |
There was a problem hiding this comment.
Use builtin_return.model_response_str()
| result = messages_to_ag_ui(messages) | ||
|
|
||
| # Check structure and count | ||
| assert len(result) == 10 |
There was a problem hiding this comment.
Let's use assert result == snapshot() here and below so we can see the entire thing in the test. The first time you run the test it'll be filled in.
| return messages, builtin_returns | ||
|
|
||
|
|
||
| def messages_to_ag_ui(messages: list[ModelMessage]) -> list[Message]: |
There was a problem hiding this comment.
Let's make messages_from_ag_ui public as well
|
@DouweM Great, thanks for input and nice that you pointed out to model_response_str. Will get to them shortly. One thing I noticed while doing this was that Logfire won't work with AGUI loaded from history as the chat history will be loaded in as json strings when running run_ag_ui (as required by RunAgentInput). This is a separate issue and PR not related to this, but adding the option to pass history highlights this issue |
Add reverse conversion function
messages_to_ag_ui()Motivation
The AG-UI integration currently supports converting AG-UI messages to Pydantic AI's internal format via
_messages_from_ag_ui(), but there's no public function to perform the reverse conversion. This creates an asymmetry in the API and makes it difficult for users to:Changes
New Function:
messages_to_ag_ui()Added a public function that converts Pydantic AI's
ModelMessageformat to AG-UI'sMessageformat:Implementation details:
UserPromptPart,SystemPromptPart,ToolReturnPart,RetryPromptPart,TextPart,ToolCallPart,BuiltinToolCallPart, andBuiltinToolReturnPartpyd_ai_builtin|{provider_name}|{tool_call_id}to match the format expected by_messages_from_ag_ui()TextPartand tool calls from a singleModelResponseinto oneAssistantMessageThinkingPartas it's not part of the conversational message historyComplexity reduction:
To maintain code quality (cyclomatic complexity < 15), the implementation is split into helper functions:
_convert_request_part(): Converts individualModelRequestparts to AG-UI messages_convert_response_parts(): ProcessesModelResponseparts and collects builtin tool returnsComprehensive Test Coverage
Added
test_messages_to_ag_ui()that validates:The test covers the same complex scenario as
test_messages_from_ag_ui()to ensure bidirectional conversion compatibility.Checklist