Struct TurnFactMiddleware 

pub struct TurnFactMiddleware { /* private fields */ }

Turn fact summary middleware — injects structured facts from the previous turn’s tool results into the next user message.

This prevents long-conversation attention drift: the LLM sees deterministic facts (tool call names) instead of relying on fuzzy memory of tool outputs buried 20+ turns ago.

Design

• After each LLM turn with tool calls, collects which tools were called and stores them in a buffer.
• At the start of the next user message, prepends the buffered facts as a structured prefix, then clears the buffer.
• Facts are derived from tool names only (not parsing output text), keeping the logic simple and model-agnostic.
• The prefix language can be configured via Language; defaults to Language::Zh for backward compatibility.

Implementations

impl TurnFactMiddleware

pub fn new() -> Self

Create a new middleware with the default language (Chinese).

pub fn with_language(language: Language) -> Self

Create a new middleware with the specified language for the prefix text.

Trait Implementations

impl Default for TurnFactMiddleware

fn default() -> Self

Returns the “default value” for a type.

impl Middleware for TurnFactMiddleware

fn on_user_message<'life0, 'life1, 'async_trait>( &'life0 self, ctx: &'life1 mut UserMessageCtx, ) -> Pin<Box<dyn Future<Output = AgentResult<()>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait,

fn on_post_llm<'life0, 'life1, 'async_trait>( &'life0 self, ctx: &'life1 mut PostLlmCtx, ) -> Pin<Box<dyn Future<Output = AgentResult<()>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait,

fn on_pre_llm<'life0, 'life1, 'async_trait>( &'life0 self, _ctx: &'life1 mut PreLlmCtx, ) -> Pin<Box<dyn Future<Output = AgentResult<()>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait,