AI-Powered Operations Agent + Dashboard Chatbot

[agamatlab]

@yeasy @YoungHypo ,

Opening this to track my mentorship project and lay out how I want to approach
it. I've got Cello running locally with make local and I've been through the
existing POCs to see how the API behaves. Would love a review on the plan below
before I start creating PRs.

What I want to build

I want an operator to be able to manage their Fabric deployment in plain language
instead of clicking through the Dashboard. Two parts:

1. A Python agent backend that takes a natural-language request, calls
   Cello's REST API through an LLM with tool-calling, and returns a structured
   answer. I also want it to summarize Fabric fault logs for common failures.
2. A chat panel in the Dashboard (React) with streaming replies and
   multi-turn conversation.

I want to start read-only (nodes, channels, chaincodes, orgs, health) and add
write actions later, once RBAC is in.

How I'm thinking of building it

Operator types in the Dashboard chat panel (React / Umi / AntD)
  -> POST /api/v1/agent/chat            (Django + DRF, JWT-auth'd, streaming)
    -> LLM tool-calling loop
      -> read-only tools over the existing REST / service layer
        -> nodes / channels / chaincodes / organizations

• I'd add a new Django app src/api-engine/agent/, next to node/, channel/,
  etc. I want the tools to reuse the same querysets and serializers the REST
  viewsets already use, so there's one source of truth.
• I'm planning to make the chat endpoint a normal JWT-protected DRF view, so the
  agent acts as the logged-in user and inherits their permissions. That way I can
  leave per-org API keys for the write phase.
• I'll keep the LLM provider configurable, defaulting to Claude, no keys in the
  repo.

PRs I want to push, in order

1. Backend read-only skeleton - the agent app: read-only tools (list
   nodes / channels / chaincodes / orgs + a deployment summary), the
   POST /api/v1/agent/chat endpoint, and unit tests with the LLM mocked.
2. Streaming - switch the chat endpoint to SSE so replies stream in.
3. Fault-log summarization - a tool that pulls a failing node's logs and
   returns an LLM summary of the likely cause.
4. Dashboard chat panel - the React component that consumes the stream,
   multi-turn, with en-US and zh-CN strings.
5. Docs - how to configure the provider, run the agent, and add a tool.

Before I get started, having answers to these could really help me think about the PRs more clearly:

• Does this match what you had in mind for the project, or would you shape it
  differently?
• Any preference on the LLM provider, or is keeping it swappable fine?
• Are you okay with the JWT-as-logged-in-user approach for now, and handling API
  keys in the write phase?