rootcx-agent

name	rootcx-agent
description	Building a RootCX AI agent — project structure, backend/agent.json config, LangGraph createReactAgent, IPC bridge, tools & RBAC, invoke SSE endpoint. Use when scaffolding or editing an agent app with a LangGraph backend.
version	0.1.0

RootCX AI Agents

Agents are apps with a backend/ containing a LangGraph agent. Same manifest, same deploy, same RBAC. Core passes LLM credentials via IPC.

Project structure

my-agent/
├── manifest.json                  # Data contract (no agent field)
├── .rootcx/launch.json
├── src/App.tsx                    # Chat UI
└── backend/
    ├── agent.json                 # Agent config (limits, memory, supervision)
    ├── agent/system.md            # System prompt
    ├── index.ts                   # LangGraph agent + IPC bridge
    └── package.json               # @langchain/langgraph, @langchain/openai, zod

backend/agent.json

Agent config — read by Core at deploy time, separate from manifest:

{
  "name": "<name>",
  "description": "<description>",
  "systemPrompt": "./agent/system.md",
  "memory": { "enabled": true },
  "limits": { "maxTurns": 50, "maxContextTokens": 100000, "keepRecentMessages": 10 },
  "supervision": { "mode": "autonomous" }
}

LLM provider selected at scaffold time. Platform secrets (ANTHROPIC_API_KEY, OPENAI_API_KEY, AWS_BEARER_TOKEN_BEDROCK) set via dashboard. Core passes credentials to agent at startup.

See templates/agent.json and templates/system.md for starting files.

Tools & permissions

All registered tools available via IPC. RBAC permissions declared in manifest.json permissions.permissions[] as { "key": "<entity>.<action>", "description": "..." } strings.

Backend code

The scaffold generates a single index.ts — the developer owns the code:

• LangGraph createReactAgent handles the ReAct loop and streaming
• Provider-specific LangChain SDK (ChatAnthropic, ChatOpenAI, ChatBedrockConverse)
• IPC bridge (JSON-lines stdin/stdout) connects to Core for tool calls and supervision
• Dependencies: @langchain/langgraph, provider package, @langchain/core, zod

Invocation

POST /api/v1/apps/{app_id}/agent/invoke
{ "message": "...", "session_id": "optional-uuid" }

Response: SSE stream (chunk, tool_call_started, tool_call_completed, approval_required, done, error events).

Other endpoints:

• GET /api/v1/apps/{app_id}/agent — agent config
• GET /api/v1/apps/{app_id}/agent/sessions — list sessions
• GET /api/v1/apps/{app_id}/agent/sessions/{session_id} — get session