🤖 AI Agent & Chat

Overview

The Chat panel is your primary interface to Oryn's AI agent. Type a task in natural language — "add email validation to the signup form", "scaffold a new React app with Docker", "find all usages of deprecated API X and replace them" — and the agent handles the rest. It reads your project files, reasons about what needs to change, applies edits through purpose-built tools, and delivers a completion summary with copy-paste verification commands. Every action that modifies a file pauses and shows you a diff so you can approve or reject it before anything is written to disk.

How it works

When you send a message the backend AgentOrchestrator runs an iterative loop: it builds a full conversation history (including your current workspace context), streams the AI model's response token-by-token to the chat panel over SignalR, and collects any tool calls the model wants to make. Read-only tools — file_read, codebase_search, list_files, file_search — execute immediately. Destructive tools — file_write, patch_apply, file_move, file_delete — are held in the ApprovalCoordinator, which suspends the loop for up to five minutes while the UI shows you the proposed diff in a modal; clicking Approve or Reject resumes or cancels that step, and the loop continues until the model signals it is done (no tool calls in the response) or the 200-iteration safety ceiling is reached. Before any prompt reaches the model, PromptGuardrails scans it for credential patterns (AWS keys, GitHub tokens, private key blocks, etc.) and either redacts them or blocks the request entirely.

What you can do

How to use it

1. Open a workspace
   Use File > Open Folder to open your project directory. The agent is workspace-scoped — it reads and writes files relative to this root, and its file search tools will not traverse outside it.
2. Type your task
   Click the compose box at the bottom of the Chat panel and describe what you want done in plain language. You can write in English or any of the supported languages listed in the model's system prompt (Hindi, Spanish, French, Japanese, and many others). Press Enter or click Send.
3. Attach context if needed
   Before or during typing you can attach more context: type @ to pick a file or folder from the workspace, type @codebase followed by a description to pull semantically relevant snippets, drag-drop an image, paste a screenshot, or click the paperclip to browse files. Attached context is included in the prompt so the agent skips redundant file reads.
4. Use a slash command for common tasks
   Type / to browse built-in commands (such as code review or MR review templates) and any commands your team has defined in .oryn/commands/*.md. Press Enter or Tab to expand the selected command into the compose box.
5. Watch the agent think and act
   The agent's reply streams in real time. As it works, tool call cards appear in the message stream showing which tools are running (file_read, codebase_search, patch_apply, etc.) and their status. The progress indicator in the panel header shows the current iteration and activity.
6. Approve or reject file changes
   When the agent wants to write, patch, move, or delete a file, the loop pauses and an approval modal appears showing a side-by-side diff of the original and proposed content. Click Approve to let the change proceed or Reject to skip it. The agent records your decision and continues. If you do not respond within five minutes the step times out and is skipped.
7. Review the completion summary
   When the task finishes, the agent sends a structured summary: a bullet list of every file it changed, a 'How to verify' section with copy-paste commands or URLs, and optional next steps. A token and iteration count is shown in the system message below.
8. Undo an edit if needed
   If you want to roll back the last agent file change, click the Undo button (↶) in the panel toolbar. A confirmation dialog shows the file name and the tool that made the change; confirming restores the pre-edit content.

Example

Admin notes

Related