Add switch_llm function to manager for runtime model switching

[obigroup]

Description

This PR adds a switch_llm function to the manager that allows changing the model during execution.

Changes

• Added switch_llm() function in the manager
• The function resets tools and nodes but preserves state
• Enables runtime model switching without losing conversation context

Known Issue

However, there's a persistent problem:

When TextFrame objects are received by LLMUserContextAggregator or LLMAssistantContextAggregator, a loop is created and text gets added multiple times to the context.

Example:

• Input: Hello my name is joe.
• Context records: Hello Hello my Hello my name Hello Hello Hello my name is

Help Needed

If someone could help me progress on this issue, I would appreciate it. I've tried adding logs everywhere in the Aggregators but I still don't understand where this is coming from.

Testing

• Basic functionality works
• State preservation confirmed
• Tools and nodes reset correctly
• Text duplication issue needs investigation

---

Note: This PR is functional but the text duplication issue in the context aggregators needs to be resolved before merging.

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Updated (UTC)
pipecat-flows	✅ Ready (Inspect)	Visit Preview	Jul 16, 2025 4:33pm

[markbackman]

This is being handled in Pipecat. We now have an llm_switching.py example using that in Flows. More LLMs will be supported in Pipecat very soon.