summarization

Default prompt used to summarize conversation history.

The <messages> marker (on its own line) and the {messages} placeholder are part of this constant's public contract, not just cosmetic formatting. Downstream consumers depend on them: for example, deep agents' SummarizationMiddleware splices an extra instruction block in immediately before the <messages> marker via str.replace. Removing, renaming, or reformatting the marker (or the {messages} placeholder) is a breaking change for those consumers even though it does not alter any function signature, so treat edits to it accordingly.

Fraction of model's maximum input tokens.

Absolute number of tokens.

Absolute number of messages.

Return metadata that marks a model call as internal to middleware.

Initialize a chat model from any supported provider using a unified interface.

Two main use cases:

1. Fixed model – specify the model upfront and get a ready-to-use chat model.
2. Configurable model – choose to specify parameters (including model name) at runtime via config. Makes it easy to switch between models/providers without changing your code

Keep internal model calls out of run.messages and the raw event log.

Used by middleware that makes internal model calls and runs before built-in transformers.

For tagged events, streamed message-start events are marked as tool-role and whole-AIMessage payloads are cleared so MessagesTransformer ignores them. The mutated events are then dropped from the raw log.

Only events within this transformer's scope are modified.

Base middleware class for an agent.

Subclass this and implement any of the defined methods to customize agent behavior between steps in the main agent loop.

State schema for the agent.

Dictionary-based trigger specification for AND conditions.

All specified thresholds in a single TriggerClause must be met for the clause to trigger summarization (AND semantics). When multiple clauses are provided in a list, summarization triggers if any clause is met (OR semantics).

Summarizes conversation history when token limits are approached.

This middleware monitors message token counts and automatically summarizes older messages when a threshold is reached, preserving recent messages and maintaining context continuity by ensuring AI/Tool message pairs remain together.

Transient summary-generation errors (rate limits, timeouts) are retried in-process, up to 3 attempts total, via Runnable.with_retry. If a summary call still fails after those attempts, the underlying error propagates rather than fabricating a summary.

Union type for context size specifications.

Can be either:

• ContextFraction: A fraction of the model's maximum input tokens.
• ContextTokens: An absolute number of tokens.
• ContextMessages: An absolute number of messages.

Depending on use with trigger or keep parameters, this type indicates either when to trigger summarization or how much context to retain.