Features

Core Extension Features

On Register(app) the extension registers constructors for:

All are registered eagerly, so constructor problems surface during startup instead of at first request.

State store:
- tries existing DI service (stateStore)
- else builds from config (memory, postgres, redis)
- on failure falls back to in-memory store with 24h TTL
Vector store:
- tries existing DI service (vectorStore)
- else builds from config
- on failure falls back to in-memory vector store

Each template includes a system prompt and three tools.

Type	Primary Scope
`cache_optimizer`	Cache hit/miss and eviction strategy
`scheduler`	Scheduling and resource allocation
`anomaly_detector`	Statistical anomaly analysis
`load_balancer`	Traffic and capacity balancing
`security_monitor`	Security anomaly and threat response
`resource_manager`	CPU, memory, and allocation optimization
`predictor`	Forecasting and trend prediction
`optimizer`	General system optimization

AgentController exposes:

Routes are explicit and only exist if you mount the controller.

These are available in the module, but not automatically bootstrapped by ai.Extension:

inference: high-throughput inference pipeline, batching, caching, scaling
training: model trainer, dataset manager, pipeline manager
middleware: intelligent rate limit, anomaly detection, personalization, response optimization, adaptive load balancing, security scanner
monitoring: health monitor, metrics collector, alert manager, dashboard

LLMConfiguration.Providers is not auto-registered into LLMManager; provider registration is manual.
CreateVectorStore only fully supports memory today.
- postgres, pinecone, and weaviate return not-implemented errors.
- extension-level constructor then falls back to memory vector store.
Training DI services are registered based on Training.Enabled.
EnableTraining in top-level config is not the registration switch by itself.
Agent definitions are stored in-memory by AgentManager.

Use type-based injection for app services (forge.InjectType[T]) and only use key aliases for compatibility.
Treat vector persistence as memory-only unless you deliberately replace the constructor with your own implementation.
If you need durability for agent definitions (not only conversation state), persist your own metadata outside AgentManager.