Agent Chat

Agent Chat is the chat front end for the RobOS app suite. Today it is a working interface over scripted behavior.

chat:send-message in main.js does not call a model. It streams a fixed list of “thoughts”, attaches sample tool results chosen by keywords in your message (such as sql, kube, pr, petshop), and returns canned text. The tool cards show names like robos_db_execute_sql but nothing is executed. For a prompt that runs a real agent, use AI Prompt Studio.

What you do in it

• The left column lists Active Threads. The plus button starts a new session.
• Choose a model in the header dropdown (Claude 3.7 Sonnet, OpenAI o3-mini, Gemini 2.5 Flash, Antigravity Harness, Local Ollama) and a mode with the Agent, Chat, Edit and Review buttons. The choice is passed along with the message and echoed in the sample reply; it does not change which model answers.
• Type in the box at the bottom. Enter sends, Shift+Enter adds a line. The chips /explain, /fix, /test, /plan and /skills prepend that text to your message.
• Code blocks in replies have Copy (copies to the clipboard) and Insert buttons. Insert has no handler yet.
• Pop Out Prompt and Prompt Popup start the AI Prompt app (packages/ai-prompt) in a separate window.
• The wrench button opens a list of twelve sample MCP tool names for apps such as DB Manager, Kube Studio and PR Review Theater.

The bundled sample session: tool cards above the reply, model and mode selectors on top, session list on the left.

Data and security

Sessions are saved to ~/.config/robos/agent-chat-sessions.json. On first run, and whenever that file is missing, the app shows two built-in sample sessions (an Acme Petshop vaccination feature and an auth blast-radius question). The workspace pill (acme-petshop, feature/rabies-tracking) and the “32 MCP Tools” label are static text.

Each outgoing message is scanned with the shared PromptSecurityGuard from robos-lib before it is stored. Depending on the configured policy, a message is blocked with a “Security Policy Violation” error or has sensitive strings redacted.

Running it

electron packages/agent-chat

Only one instance runs at a time. No Docker or API keys are needed because no model is called.