v0.0.14

Reverted

• Temporarily reverted the deprecation of the tts parameter in
  FlowManager.__init__(). This feature will be deprecated in a future release
  after the required Pipecat changes are completed.

v0.0.13

Added

• Added context update strategies to control how context is managed during node
  transitions:
  • APPEND: Add new messages to existing context (default behavior)
  • RESET: Clear and replace context with new messages and most recent
    function call results
  • RESET_WITH_SUMMARY: Reset context but include an LLM-generated summary
    along with the new messages
  • Strategies can be set globally or per-node
  • Includes automatic fallback to RESET if summary generation fails

Example usage:

# Global strategy
flow_manager = FlowManager(
    context_strategy=ContextStrategyConfig(
        strategy=ContextStrategy.RESET
    )
)

# Per-node strategy
node_config = {
    "task_messages": [...],
    "functions": [...],
    "context_strategy": ContextStrategyConfig(
        strategy=ContextStrategy.RESET_WITH_SUMMARY,
        summary_prompt="Summarize the key points discussed so far."
    )
}

• Added a new function called get_current_context which provides access to
  the LLM context.

Example usage:

# Access current conversation context
context = flow_manager.get_current_context()

• Added a new dynamic example called restaurant_reservation.py.

Changed

• Transition callbacks now receive function results directly as a second argument:
  async def handle_transition(args: Dict, result: FlowResult, flow_manager: FlowManager).
  This enables direct access to typed function results for making routing decisions.
  For backwards compatibility, the two-argument signature
  (args: Dict, flow_manager: FlowManager) is still supported.

• Updated dynamic examples to use the new result argument.

Deprecated

• The tts parameter in FlowManager.__init__() is now deprecated and will
  be removed in a future version. The tts_say action now pushes a
  TTSSpeakFrame.

v0.0.12

Added

• Support for inline action handlers in flow configuration:
  • Actions can now be registered via handler field in config
  • Maintains backwards compatibility with manual registration
  • Built-in actions (tts_say, end_conversation) work without changes

Example of the new pattern:

"pre_actions": [
    {
        "type": "check_status",
        "handler": check_status_handler
    }
]

Changed

• Updated dynamic flows to use per-function, inline transition callbacks:
  • Removed global transition_callback from FlowManager initialization
  • Transition handlers are now specified directly in function definitions
  • Dynamic transitions are now specified similarly to the static flows'
    transition_to field
  • Breaking change: Dynamic flows must now specify transition callbacks in
    function configuration

Example of the new pattern:

# Before - global transition callback
flow_manager = FlowManager(
    transition_callback=handle_transition
)

# After - inline transition callbacks
def create_node() -> NodeConfig:
    return {
        "functions": [{
            "type": "function",
            "function": {
                "name": "collect_age",
                "handler": collect_age,
                "description": "Record user's age",
                "parameters": {...},
                "transition_callback": handle_age_collection
            }
        }]
    }

• Updated dynamic flow examples to use the new transition_callback pattern.

Fixed

• Fixed an issue where multiple, consecutive function calls could result in two completions.

v0.0.11

Changed

• Updated FlowManager to more predictably handle function calls:

  • Edge functions (which transition to a new node) now result in an LLM
    completion after both the function call and messages are added to the
    LLM's context.
  • Node functions (which execute a function call without transitioning nodes)
    result in an LLM completion upon the function call result returning.
  • This change also improves the reliability of the pre- and post-action
    execution timing.

• Breaking changes:

  • The FlowManager has a new required arg, context_aggregator.
  • Pipecat's minimum version has been updated to 0.0.53 in order to use the
    new FunctionCallResultProperties frame.

• Updated all examples to align with the new changes.

v0.0.10

Changed

• Nodes now have two message types to better delineate defining the role or
  persona of the bot from the task it needs to accomplish. The message types are:

  • role_messages, which defines the personality or role of the bot
  • task_messages, which defines the task to be completed for a given node

• role_messages can be defined for the initial node and then inherited by
  subsequent nodes. You can treat this as an LLM "system" message.

• Simplified FlowManager initialization by removing the need for manual context
  setup in both static and dynamic flows. Now, you need to create a FlowManager
  and initialize it to start the flow.

• All examples have been updated to align with the API changes.

Fixed

• Fixed an issue where importing the Flows module would require OpenAI,
  Anthropic, and Google LLM modules.

v0.0.9

Changed

• Fixed function handler registration in FlowManager to handle __function__: tokens
  • Previously, the handler string was used directly, causing "not callable" errors
  • Now correctly looks up and uses the actual function object from the main module
  • Supports both direct function references and function names exported from the Flows editor

v0.0.8

Changed

• Improved type safety in FlowManager by requiring keyword arguments for initialization
• Enhanced error messages for LLM service type validation

v0.0.7

Added

• New transition_to field for static flows
  • Combines function handlers with state transitions
  • Supports all LLM providers (OpenAI, Anthropic, Gemini)
  • Static examples updated to use this new transition

Changed

• Static flow transitions now use transition_to instead of matching function names
  • Before: Function name had to match target node name
  • After: Function explicitly declares target via transition_to

Fixed

• Duplicate LLM responses during transitions

v0.0.6

Added

• New FlowManager supporting both static and dynamic conversation flows
  • Static flows: Configuration-driven with predefined paths
  • Dynamic flows: Runtime-determined conversation paths
  • Documentation: Guide and Reference
• Provider-specific examples demonstrating dynamic flows:
  • OpenAI: insurance_openai.py
  • Anthropic: insurance_anthropic.py
  • Gemini: insurance_gemini.py
• Type safety improvements:
  • FlowArgs: Type-safe function arguments
  • FlowResult: Type-safe function returns

Changed

• Simplified function handling:
  • Automatic LLM function registration
  • Optional handlers for edge nodes
• Updated all examples to use unified FlowManager interface

v0.0.5

Added

• Added LLM support for:

  • Anthropic
  • Google Gemini

• Added LLMFormatParser, a format parser to handle LLM provider-specific
  messages and function call formats

• Added new examples:

  • movie_explorer_anthropic.py (Claude 3.5)
  • movie_explorer_gemini.py (Gemini 1.5 Flash)
  • travel_planner_gemini.py (Gemini 1.5 Flash)