CHANGELOG 0.2.3 - Cartesia Docs

Detailed Changelog: v0.2.2 → v0.2.3

Feature: History Management API

LlmAgent now exposes a self.history attribute of type History, providing structured control over the conversation history that the LLM sees. history.add_entry(content, role, *, before, after) Inserts a CustomHistoryEntry into the conversation history. The entry appears as a message with the given role ("system" or "user") in the LLM conversation. Positioning is controlled via anchors:

No anchor: appends to the current end of history.
before=event: inserts immediately before the specified event.
after=event: inserts immediately after the specified event.

Raises ValueError if both before and after are specified, or if the anchor event is not found in the current history.

# Append a user note at the end (role="user" is the default)
agent.history.add_entry("The user prefers formal language.")

# Insert before a specific event
agent.history.add_entry("Context about the caller.", before=some_event)

history.update(events, *, start, end) Replaces a segment of history with new events. Behavior depends on which anchors are provided:

Neither start nor end: replaces the entire history as it currently exists.
start only: replaces everything from start through the current end of history.
end only: replaces everything from the beginning through end with events.
Both start and end: replaces the segment [start..end] inclusive with events.

Raises ValueError if anchors are not found or end appears before start.

Feature: Per-Turn Overrides on `process()`

LlmAgent.process() now accepts four arguments that apply to just that turn: config: Optional[LlmConfig] Merged onto the agent’s stored config for this turn only. Uses a layered merge where the last explicitly-set value wins (agent default < stored config < per-turn override).

# Use a higher temperature for just this turn
await agent.process(env, event, config=LlmConfig(temperature=0.9))

Only the fields you explicitly set in the override LlmConfig take effect. Unset fields fall through to the agent’s stored config. tools: Optional[List[ToolSpec]] Tools with matching names replace those in the agent’s tool list; unmatched base tools are preserved. This lets you swap out a specific tool’s implementation for one turn without losing the rest.

# Replace just the "lookup" tool for this turn
await agent.process(env, event, tools=[custom_lookup_tool])

context: Union[str, List[HistoryEvent], None] Extra context appended to the end of history for message building. If a string, it is automatically wrapped as a user message. If a list of events, they are appended as-is. Not persisted — only affects this single turn.

# Inject ephemeral context
await agent.process(env, event, context="The user is a VIP customer.")

history: Optional[List[HistoryEvent]] Completely overrides the managed history for this turn. The managed history still receives normal updates (so subsequent turns without the override see the full history).

Feature: Configurable `EndCallTool`

end_call now accepts an description argument.

from line.llm_agent.tools.system import end_call

# Custom — control when the LLM ends the call
LlmAgent(tools=[end_call(description="End when the customer confirms their order.")])

Feature: Multilingual STT/TTS Language Support

AgentUpdateCall now has a language field:

yield AgentUpdateCall(voice_id="some-voice", language="es")

to specify TTS/STT language. Use language="multilingual" to enable automatic language detection.

Feature: Custom Events

Two new event types enable arbitrary metadata exchange between agent code and the harness. Agent → Harness: AgentSendCustom

yield AgentSendCustom(metadata={"action": "show_form", "form_id": "checkout"})

Harness → Agent: UserCustomSent Received when the call side sends a custom event to the agent.

Feature: Agent-as-Handoff Update Parameters

agent_as_handoff() now accepts an update_call parameter to change voice/STT/TTS settings during a handoff:

from line.llm_agent.tools.system import agent_as_handoff, UpdateCallConfig

spanish_agent = LlmAgent(
    model="gemini/gemini-2.5-flash-preview-09-2025",
    config=LlmConfig(system_prompt="You speak only in Spanish."),
)

main_agent = LlmAgent(
    model="gemini/gemini-2.5-flash-preview-09-2025",
    tools=[
        agent_as_handoff(
            spanish_agent,
            handoff_message="Transferring you to our Spanish-speaking agent...",
            update_call=UpdateCallConfig(
                voice_id="spanish-voice-id",
                language="es",
            ),
            name="transfer_to_spanish",
            description="Transfer to a Spanish-speaking agent when requested.",
        ),
    ],
)

UpdateCallConfig supports voice_id, pronunciation_dict_id, and language.

Feature: LLM Agent Timing Metrics

Three LogMetric events are now automatically emitted on every turn.

Metric	Description
`llm_first_chunk_ms`	Time from start of response generation to first chunk (text or tool call) from the LLM
`llm_first_text_ms`	Time from start of response generation to first text chunk.
`agent_turn_ms`	Total agent processing time for the turn, from start to completion.

​Detailed Changelog: v0.2.2 → v0.2.3

​Feature: History Management API

​Feature: Per-Turn Overrides on process()

​Feature: Configurable EndCallTool

​Feature: Multilingual STT/TTS Language Support

​Feature: Custom Events

​Feature: Agent-as-Handoff Update Parameters

​Feature: LLM Agent Timing Metrics