domain-model

To structure an AI agent's domain model, define Tools, ToolUseContext, AppState, Message, Task, and Command to bridge LLM tool calls with the local runtime using typed Zod schemas and immutable state transitions.

The best way to implement permission checks is by using a ToolUseContext to inject per-call permission callbacks into every tool execution, ensuring capabilities are validated independently without relying on global singletons.

Manage concurrency and state transitions by implementing a reducer-based AppState store that updates atomically and tracks background Tasks, ensuring safe execution of concurrent operations and immutable state changes.

Yes, you can use Zod schemas to define typed input contracts for agent tools, producing structured tool calls and results for the LLM while ensuring capabilities remain typed and testable.

Pass dependencies into an LLM tool call by utilizing contextual dependency injection through a ToolUseContext, which carries state, abort signals, and permissions directly into each function execution.

Yes, this domain model approach works for spawning long-running background tasks by tracking them within the AppState store and utilizing abort signals passed through the ToolUseContext for safe monitoring.