The pipecat, cartesia, and agent-framework packages passed search results
into dedup/format helpers written for plain dicts, but typed SDK results
are pydantic models (snake_case attributes, no dict interface). pipecat
and cartesia crashed with AttributeError — swallowed upstream, so no
memories were ever injected for users with search results — while
agent-framework silently dropped every search-result memory.
Extract memory fields through a tolerant accessor that accepts both the
camelCase dicts returned by the profile endpoint and SDK result models,
so the helpers work regardless of installed SDK version.
The pipecat/cartesia dependency stubs move from test_empty_profile.py
into a shared tests/conftest.py so the new utils tests can reuse them.
Fixes#1266
#### RE-RAISING Pipecat live speech PR
### Added native speech-to-speech model support
### Summary:
- Speech-to-speech support - Auto-detect audio frames and inject memories to system prompt for native audio models (Gemini Live, etc.)
- Fix memory bloating - Replace memories each turn using XML tags instead of accumulating
- Add temporal context - Show recency on search results ([2d ago], [15 Jan])
- New inject_mode param - auto (default), system, or user
### Docs update
- Update the docs for native speech-2-speech models
