Modules
Each module page below is based on a scan of the corresponding Go package and periodically audited against the source tree (source is authoritative when a page lags). Every page includes a responsibility summary, a Mermaid architecture diagram, the exported interfaces and key methods, module collaboration relationships, concrete extension points, and a maturity annotation (Production / Beta / Experimental).
Maturity criteria:
- Production — covered by tests, no experimental markers, integrated into the SDK entry points.
- Beta — functional and tested but still evolving its public API.
- Experimental — new or undergoing significant rework; APIs may change.
- sdk
Top-level unified entry point that wires every ARES subsystem behind a production-friendly API.
Production - llm
Provider-agnostic LLM client, failover, and the routing service that picks Chat vs. plain generate.
Production - tools
Tool registry with 9 builtins, a capability planner, and an intent-driven execution bridge.
Production - core
Shared DTOs, value types, and service/repository interfaces that every ARES module depends on.
Production - ares_memory
Unified memory management coordinating session memory, task memory, distilled experiences, and RAG context retrieval.
Production - knowledge
ARES Knowledge Fabric (AKF) - universal knowledge representation, distillation bridge, and intent-driven retrieval.
Experimental - evidence
Universal data primitive for ARES - structured evidence collection, storage, and aggregation across all subsystems.
Beta - storage
PostgreSQL repositories, connection pooling, tenant isolation, and vector search for the ARES persistence layer.
Production - retrievalservice
Tenant-scoped knowledge base service - CRUD lifecycle and search over a pluggable retrieval repository.
Beta - ares_mcp
MCP client, server manager, and stdio/SSE transports that auto-register remote tools into the registry.
Production - agents
Leader/sub agent implementations, the base Agent interface, agent service, and shared strategy contracts.
Production - workflow
Unified DAG runner with serializable IR, scheduler, retry/recovery, HITL, loops, and idempotent checkpoint resume.
Production - kernel
ARES scheduling kernel and system control plane: the sole site of scheduling decisions, plus the component Registry/Orchestrator unified from system_runtime.
Production - ares_skills
Capability Fabric: SkillCatalog over declared sources, progressive disclosure, lazy MCP activation, and the experience relevance prior.
Production - system_runtime
System-level control plane: component registry, dependency-aware topological ordering, lifecycle orchestration, and status snapshots.
Production - agentipc
Peer-to-peer agent message bus: Send / Request / Reply / Delegate / Handoff / Subscribe / Broadcast, plus dual-track dispatch policy.
Production - agentfabric
Disposable-execution agent substrate: spawn provenance, lifecycle state machine, three-layer context isolation, and P5 resource admission.
Production - taskfabric
Durable-intent task substrate: lease fencing, cooperative preemption, quantum-boundary yields, and capability-aware work stealing.
Production - agentloop
ReAct (reason+act) execution engine: iterate→generate→tool-call→feed-back cycle with token/timeout budgets, human-in-the-loop approval, runtime tool discovery, and structured memory persistence.
Production - ares_archive
Round archive for the closed-loop agent: per-round RoundRecord persisted as round_N.json, P0–P5 retention, compaction-independent persistence, and the recall CLI.
Production - ares_shutdown
Graceful shutdown: phase-coordinated teardown (pre-shutdown → graceful → force → done), priority-sorted callback registry, signal handling, and per-phase executors with retries and rollback.
Production - aresrecovery
Kernel Recovery P5: durable task survival across agent deaths — lease expiry requeue, checkpoint recovery, and agent restart with restart budget/backoff and evolution-aware spawn gates.
Production - api
Public HTTP and library entry points that wire every ARES module into a single runtime.
Production - ares_runtime
Agent lifecycle manager: registration, supervised start/stop, crash resurrection, snapshots, and workflow plugins.
Production - ares_protocol
AHP messaging protocol: typed messages, per-agent queues, heartbeats, codecs, and a dead-letter queue.
Production - ares_callbacks
Lifecycle event hooks for LLM calls, agent execution, and tool invocations: a thread-safe registry, an Emitter interface, and an EventStore bridge.
Beta - ares_events
Event sourcing for ARES - append-only event store with optimistic concurrency, subscriptions, and stream compaction.
Production - ares_bootstrap
Single wiring hub that assembles all ARES components from config and optional dependencies.
Production - ares_config
YAML configuration loading, defaults, and validation for the ARES server — ares.yaml is the single config entry point.
Production - ares_observability
Tracing, metrics, structured logging, and cost tracking for ARES agent and evolution operations.
Beta - ares_ratelimit
Pluggable rate limiters (token bucket, sliding window, semaphore) with per-key limiting and a factory registry.
Beta - ares_security
Dependency-free sensitive-data sanitizer with 9 field types, JSON-aware redaction, and a safe logger wrapper.
Production - ares_evolution
Autonomous strategy evolution: GA dream cycle, genome/diff/patch coordinator, and 7-source runtime patch pipeline.
Beta - ares_eval
Evaluation framework: test suites, runners, LLM-as-judge, concurrent exec, comparison, and persistence.
Beta - ares_flight
Flight recorder for multi-agent systems: execution timeline, call graph, decision log, diagnostics, genealogy, memory pipeline, and step-by-step replay.
Beta - ares_arena
Chaos engineering arena: fault injection, survival mode, scenario execution, resilience scoring, and A/B regression testing of agent strategies.
Beta - discovery
Pluggable service discovery engine: provider-based scanning, identity merge, health verification, passive registration, and event emission.
Production - plugins
Plugin system for ARES: the resurrection plugin monitors any agent type and automatically recreates it on failure with snapshot or event-replay state recovery.
Production