add_notification_metadata(
message: HumanMessage,
*,
source: str,
data: dict[str| Name | Type | Description |
|---|---|---|
message* | HumanMessage | The message to mark. |
source* | str | The feature that added the message. |
data | dict[str, Any] | None | Default: None |
Return a copy of message marked as added by the agent harness.
Use it for a human message the user didn't type, such as a background task
update or a conversation summary. It sets additional_kwargs["lc_source"]
and, when data is given, additional_kwargs["lc_data"]. Any earlier
lc_source and lc_data are removed first. LangChain's chat model
integrations don't send these keys to the model.
It doesn't change the content. Write the content the model reads yourself.
Anyone who can send input to a thread can set these keys. Don't use them to decide whether to trust a message.
merge_message_runs and chunk addition join a notification with adjacent
human messages; the merged message keeps the metadata and concatenates
lc_source.
Example:
from langchain_core.messages import HumanMessage, add_notification_metadata
message = add_notification_metadata(
HumanMessage("Background task t_42 finished."),
source="async_subagent",
data={"task_id": "t_42", "status": "success"},
)The feature's fields. Must be JSON-serializable.