Crate ares_server

Expand description

ARES agent library facade.

Downstream crates inject Execute, Tools, and Llm on a Cordis Context and run an agent with no HTTP stack on the graph.

use ares_server::{Context, Execute, Tools, Llm};

The ares-server binary (same package) binds the Axum router only when the http feature is on, and ships only with http and cli. Default features are the server defaults: postgres, openai, ares-vector, mcp, inventory, rhai-policy, http, cli, script-tools. For an embed-only build use default-features = false and pick the providers you need (e.g. features = ["openai"]); no axum, clap, sqlx, ollama, boa, or rhai then enters the graph.

Public surface: Context, Execute, Tools, Llm, Store (postgres), Plugin, Loader, Dispatch, register_plugins. Construct in-memory Llm with Llm::from_client for the library proof.

Modules§

supervisor: Daemon-side supervised-worker protocol: restart loop and child spawn. Daemon-side half of the supervised-worker protocol.

Structs§

AgentConfig: Agent configuration binding a model to tools and behavior.
AgentRegistry: Registry for managing agent configurations and creating agent instances
AgentRequest: Request for unified agent execution.
AgentResponse: Response from agent execution, including content and optional token usage
Calculator: Calculator tool for basic arithmetic operations.
CalculatorConfig: Empty config for CalculatorService. Default suffices; Serialize and Deserialize are required for Plugin Config via RegistryService.
CalculatorService: CalculatorService wraps Calculator for Cordis dependency injection.
Context
ConversationMessage: A message in a tool-calling conversation.
Entry: A single declarative loader entry.
EventsService
Execute: Unified agent execution service — the single place handling:
ExecutionMetadata: Metadata about the execution of an agent
ExecutionResult
LLMResponse: Response from an LLM generation call
Llm: Unified LLM capability composing provider registry, catalog, factory, and pool.
Loader: Declarative loader — diffs EntryTrees incrementally.
Message: A single message in a conversation.
PluginRegistry: Name-keyed directory of PluginFactory closures.
ProviderRegistry: Registry for managing multiple named LLM providers
TenantContext
TokenUsage: Token usage statistics from an LLM generation call
ToolCall: A request to call a tool.
ToolDefinition: Definition of a tool that can be called by an LLM.
Tools: Tenant-aware tool capability.

Enums§

AgentSource: Result of Execute::run including resolution metadata.
AppError: Application-wide error type.
CordisError
Dispatch
ProviderConfig: LLM provider configuration.
TenantTier

Traits§

ContextProvider: Trait for injecting external context into agent calls.
Disposable
LLMClient: Generic LLM client trait for provider abstraction
Plugin: A plugin is a declarative unit of configuration that can provide a service into a context. The registry calls apply to create the service value and then inserts it into the context with single-source discipline. The returned Arc is the live service instance that callers retrieve through ctx.get. Plain prose is used here to keep the docs readable and to avoid clever abstractions that would hide the simple flow of create then provide.
RunTracker: Trait for tracking active agent runs. Implemented by the root crate’s ActiveRuns and injected into Execute via the Context.
Service
Tool: Trait for implementing tools that agents can invoke.
TypedEvent: Compile-time binding between a catalog event and its payload type.

Functions§

register_plugins: Register capability-crate loader factories on reg.

Type Aliases§

FiberId
ServiceInitFuture
Store: Tenant database. Gated so --no-default-features builds do not enable postgres.