Module runtime_layer 

RuntimeLayer trait: middleware hooks for LLM calls and tool dispatch (#2286).

Provides interception points before and after each LLM chat call and each tool execution. Layers are composed in a stack: each layer is called in order, and any layer may short-circuit the actual call by returning Some(result) from before_chat or before_tool.

MVP

No layers are registered at bootstrap — the runtime_layers vec in Agent defaults to empty, making the hook loops zero-cost (no iteration, no allocation).

Future layers (rate limiting, guardrails, cost tracking, audit logging) add themselves to the vec at bootstrap without modifying the agent loop.

Implementation note

Default implementations return Box::pin(std::future::ready(...)). This allocates once per call per registered layer. For the MVP empty-vec case, no allocation occurs. Real layers should keep their work minimal to avoid blocking the agent loop.

Structs

LayerContext

Context available to runtime layers during interception.

NoopLayer

No-op layer that passes everything through unchanged.

Traits

RuntimeLayer

Middleware layer that wraps LLM calls and tool dispatch.

Type Aliases

BeforeToolResult

Short-circuit result type for before_tool: Some(result) bypasses tool execution.