FlowCraft

Go SDK for building AI agents with long-term memory, knowledge retrieval, runtime orchestration, and voice. Source on github.com/GizClaw/flowcraft.

Guides

Runtime

• Graph Runtime — core/graph: declarative DAG engine, node I/O roles, parallel branches, custom node types.
• Tool System — core/tool: LLM function-calling contract, Registry / Catalog / Executor split, middleware chain, built-in tool adapters and the MCP bridge.
• Event Bus — core/event: subject-routed publish/subscribe, in-process MemoryBus, host capability wiring, backpressure policies.
• Application Runtime — core/runtime + core/runtime/session: process-level services and leased, interruptible streaming sessions above a built deployment.
• Prompt Lifecycle Events — the agent.run.<id>.prompt.* lifecycle events UI consumers subscribe to.

State and execution boundary

• Workspace — core/workspace: per-run filesystem abstraction, backends, capabilities, the state vs policy split vs Sandbox.
• Sandbox — core/sandbox: agent execution boundary, env / net / resources policy, runners (local / seatbelt / bwrap), decorators and approval.
• Memory Stack — the three-layer memory stack: core/memory contracts, deploy/runtime glue, and app-registered implementations.

Assembly

• Inference Runtime — unified Generate / Embed / Transcription (Realtime reserved): deployment config, routing, extensions, streaming, media intents, hot reload.
• Deployment Assembly — core/deploy: one YAML document + one Build call to wire shared resources, named agents, engines, and lifecycle hooks.
• Resource Protocol — core/resource: the provider-neutral factory, dependency DAG, loader, and lifecycle phases that every deployment resource uses.

Delegation

• Delegation — core/delegation: backend-neutral target discovery, sync / async execution, and the session-bound delegation lifecycle.

Migrations

• core/v0.1.0 — the breaking cut from sdk/sdkx to the core platform module, provider driver/* modules, and platform-specific backends/*.
• One driver per wire family — the consolidation that folded the Azure, DeepSeek and Kimi drivers into driver/openai, moved MiniMax's Messages surface to driver/anthropic, and removed driver/qwen.

Older sdk/sdkx migration notes remain in docs/migrations/ as historical reference and are not part of the current core migration path.

Layered architecture

The repository is organised as independently released Go modules:

Layer	Package	Responsibility
Execution contracts	core/agent	Engine / Board / Run / Host / Interrupt / Checkpoint contracts
DAG executor	core/graph	Declarative graph runtime (*Graph implements agent.Engine)
Agent runtime	core/agent	Agents, observers, referees, board seeders, and execution lifecycle
Delegation contracts	core/delegation	Backend-neutral target discovery, sync / handoff / async requests, service and host contracts
Async delegation	core/delegation/kanban	In-memory AsyncBackend / WorkSource implementation and operational views
Adapters	driver/*, backends/*	Concrete provider / protocol bindings layered on core contracts

Repository layout

core/            Platform module (contracts, deploy, runtime, built-in resources)
driver/          Provider inference adapters
backends/        SQLite checkpoints and plugin shell (sandbox backends live in core/)
examples/        Reference assemblies

Getting started

go get github.com/GizClaw/flowcraft/core@latest

See the package-level doc.go files for runnable usage snippets: core/agent/doc.go, core/graph/doc.go, and the focused packages under core/.