GoWind Admin Open-Source Project Deconstructs AI Capability Layer Architecture

On November 11, 2026 at 11:56, Juejin published an article breaking down the AI capability layer architecture of GoWind Admin, an open source project. It is the third installment in a series on the project's AI Agent module, following two earlier pieces that covered the canvas and the orchestration engine. This article focuses on the infrastructure layer that sits between the engine and the LLM: the tools that run in loops, the knowledge that gets retrieved, the tokens that get metered, and the ways an Agent is exposed to external callers.

Four components are dissected. The tool calling loop is built on the function calling multi-turn protocol, with aiToolLoop.run (in ai_tools.go) bounded by aiToolMaxRounds = 4. The counterintuitive part: the fourth round sends no tools in the request, so the model has nothing left to call and must converge on a text answer. That is more graceful than raising an error when a loop is detected, since the user always gets a response, and the hard cap also neutralizes prompt injection attempts that try to trigger repeated calls. Inside the loop, tool errors do not break the conversation. They are fed back to the model as tool results, letting it correct parameters or switch tools on the next round. Coexistence with streaming SSE is handled by frame routing, with content deltas carried on AIChatChunk frames. Knowledge retrieval uses a pgvector vector store, usage is metered through five checkpoints in AI_TOKENS, and external access is guarded by AK/SK machine credentials.

This is the only report published so far. Design decisions in the capability layer are now public, but details on runtime behavior and future changes remain to be covered.