Crate motosan_agent_loop 

motosan-agent-loop

Standalone ReAct agent loop — LlmClient + Engine with no platform dependencies.

This crate provides the core loop that drives an LLM through iterative reasoning and tool execution, producing a final answer.

Extensions are registered via EngineBuilder::extension (errors fall back to default) or EngineBuilder::extension_with_policy (opt in to ErrorPolicy::Abort for fail-loud semantics). With the motosan-ai feature, wrap a motosan_ai::Client in MotosanAiClient to forward a max_tokens budget end-to-end (required for the token-budget extension to observe StopReason::MaxTokens).

Re-exports

pub use context::ContextProvider;

pub use core::engine::AgentMode;

pub use core::engine::AgentOp;

pub use core::engine::AgentResult;

pub use core::engine::AskUserOption;

pub use core::engine::AskUserQuestion;

pub use core::engine::BackpressurePolicy;

pub use core::engine::ChannelConfig;

pub use core::engine::Engine;

pub use core::engine::EngineBuilder;

pub use core::engine::NextStepSchedule;

pub use core::event::AgentEvent;

pub use core::event::CoreEvent;

pub use core::event::ExtensionEvent;

pub use delegate::DelegateAgentTool;

pub use error::AgentError;

pub use error::Result;

pub use extensions::ask_user::AskUserConfig;

pub use extensions::autocompact::compute_compaction_plan;

pub use extensions::autocompact::compute_compaction_plan_with_estimator;

pub use extensions::autocompact::ApproxCharsTokenEstimator;

pub use extensions::autocompact::AutocompactConfig;

pub use extensions::autocompact::AutocompactEvent;

pub use extensions::autocompact::AutocompactExtension;

pub use extensions::autocompact::ChainedCompactionPolicy;

pub use extensions::autocompact::CompactionPlan;

pub use extensions::autocompact::EvictOldImagesPolicy;

pub use extensions::autocompact::EvictOldReasoningPolicy;

pub use extensions::autocompact::EvictOldToolResultsPolicy;

pub use extensions::autocompact::NoopCompactionPolicy;

pub use extensions::autocompact::ReplayLastUserPolicy;

pub use extensions::autocompact::RequestCompactionPolicy;

pub use extensions::autocompact::TokenEstimator;

pub use extensions::config::build_extension_set;

pub use extensions::config::DelegationConfig;

pub use extensions::config::ExtensionBuildError;

pub use extensions::config::ExtensionWiring;

pub use extensions::config::ExtensionsConfig;

pub use extensions::config::PlanningConfig;

pub use extensions::follow_up::FollowUpEvent;

pub use extensions::follow_up::FollowUpExtension;

pub use extensions::follow_up::FollowUpFn;

pub use extensions::stuck_detection::StuckDetectionConfig;

pub use extensions::token_budget::TokenBudgetConfig;

pub use llm::ChatOutput;

pub use llm::LlmClient;

pub use llm::LlmResponse;

pub use llm::LlmResponseMeta;

pub use llm::StopReason;

pub use llm::StreamChunk;

pub use llm::TokenUsage;

pub use llm::ToolCallItem;

pub use message::find_cut_points;

pub use message::is_safe_cut;

pub use message::new_message_id;

pub use message::AssistantContent;

pub use message::ContentPart;

pub use message::DocumentSource;

pub use message::Message;

pub use message::MessageId;

pub use message::MessageMeta;

pub use message::Role;

pub use message::ToolCallRef;

pub use message::DOCUMENT_APPROX_CHAR_EQUIVALENT;

pub use message::IMAGE_APPROX_CHAR_EQUIVALENT;

pub use planning::Plan;

pub use planning::PlanStep;

pub use planning::PlanningEvent;

pub use planning::PlanningTool;

pub use planning::StepStatus;

pub use session::catalog::SessionCatalog;

pub use session::compaction::CompactionResult;

pub use session::compaction::CompactionStrategy;

pub use session::compaction::ThresholdStrategy;

pub use session::entry::EntryId;

pub use session::entry::SessionEntry;

pub use session::meta::SessionMeta;

pub use session::projection::default_projection;

pub use session::projection::DefaultProjection;

pub use session::projection::MessageProjection;

pub use session::session::AgentSession;

pub use session::session::TurnHandle;

pub use session::store::FileSessionStore;

pub use session::store::MemorySessionStore;

pub use session::store::SessionStore;

pub use stream::AgentStreamItem;

pub use stream::AgentTerminal;

pub use streaming_executor::StreamingToolExecutor;

Modules

context

core

Mechanism layer: the agent loop engine.

delegate

Delegation tool — wraps an Engine as a callable Tool so a main agent can delegate subtasks to a sub-agent (browser agent, SWE agent, etc.) via normal tool calling.

error

extensions

Policy layer: self-contained feature modules that plug into the core via the Extension trait (added in Phase 1).

llm

message

Conversation messages — the unit exchanged between the agent loop and the LLM.

planning

Built-in PlanningTool — lets agents create, manage, and track execution plans autonomously via tool calling.

session

Session layer: long-lived conversation state, persistence, and orchestration. See each submodule for its specific role.

stream

Stream item types for the Engine::run API.

streaming_executor

Streaming tool executor — starts executing tools as LLM streams tool_use blocks.