Add back file that was removed, when it should've just been deprecated.
Also, fix version numbers in deprecation messages to match the next expected release.
Update `create_context_aggregator()` (which we're keeping around for backward compatibility) to create a `LLMContextAggregatorPair` rather than OpenAI-Realtime-specific aggregators.
Implement sending tool call results to the OpenAI server based on reading context updates. This lets us use the normal assistant context aggregator and not a special OpenAI Realtime subclass that pushes up a special frame for function call results.
Receiving a new context (via a context frame) no longer serves as a signal to reset the conversation. That’s because we’re now receiving new contexts from the user aggregator every time new messages are added, and from the assistant aggregator when function call results come in. The code pattern we're heading towards, of “diffing” each new context with the previous on, sets us up for doing more sophisticated things in the future, like sending specific messages to OpenAI to edit its internally-tracked context.
Also, remove code that was directly modifying context.
Push `TranscriptionFrame`s upstream, to be handled by the user context aggregator. This will require at least a couple of other changes:
- Updating examples to put transcript processors upstream from `OpenAIRealtimeLLMService`
- Maybe figuring out a way to preserve backward compatibility with existing pipelines that put transcript processors downstream from `OpenAIRealtimeLLMService`
- Updating `OpenAIRealtimeLLMService` to ignore new received context frames, since the upstream user context aggregator will generate those after each newly-added user message; hopefully nobody was reliant on the old behavior of resetting the conversation upon receiving a new context!
Avoid pushing `LLMTextFrame` when `OpenAIRealtimeLLMService` is configured to output audio. This avoids duplicate text in assistant messages in context. Conceptually, a speech-to-speech service encapsulates TTS behavior; in a "traditional" pipeline, `LLMTextFrames` are swallowed by the TTS service, so they should similarly not be pushed by a speech-to-speech service. Only. `TTSTextFrame`s should be pushed.
Properly bookended responses now work with:
- AUDIO modality (validated with 26b example)
- TEXT modality (validated with 26d example)
- AUDIO modality with Vertex AI (validated with 26h example)
It doesn't seem that TEXT modality is supported with Vertex AI, hence the missing "quadrant" of validation.
