Sentinel for VS Code

Sentinel is a coding agent that runs entirely on your machine: a local server runs the agent, keeps an audit trail of everything it does and enforces its guardrails; a web UI talks to it. This extension puts that UI in VS Code's side bar, starts the server for you, and works with the editor: the file and lines in front of you are the chat's context, and the chat's files, diffs and proposed changes open in VS Code's own editors.

It is a thin client. The extension edits no file and calls no model; everything happens in the server, which the extension carries (its Python source and prebuilt UI, about 12 MB) and runs on the Python already on your machine.

What you need

• VS Code 1.100 or newer.
• Python 3.11 or newer on the machine. The extension finds it (the py launcher, then python3 / python on PATH, or the path in sentinel.pythonPath); it never installs it.
• pip able to install packages. On a managed machine that is the package index already in your pip configuration; the extension carries no index address of its own and downloads nothing itself. If pip is not configured for you, set sentinel.pipIndexUrl.
• An open folder: Sentinel works on the folder open in the window.

First run

1. Open a folder.
2. Click the Sentinel shield in the activity bar (or run Sentinel: Open). Drag the view to the secondary side bar if you want the chat on the right.
3. The first time, the runtime is prepared: a virtual environment, then pip install of the server's packages. A few minutes once; seconds afterwards. The Sentinel output channel shows every line.
4. The chat opens on its Settings panel: enter your model endpoint and key, pick a model, save. The key is stored in the operating system's credential store, never a file.

In the editor

• Context follows you. Click in a file or select lines: the chat's context chip shows them, and "what does this do?" is about them. Only the path and line numbers are sent; the agent reads the lines itself, audited.
• Add to Chat and Explain This on the editor's right-click menu; Fix with Sentinel in the light bulb on an error or warning.
• Files and diffs open in VS Code. File links open at their line; a changed file opens in the diff editor against its checked-in copy.
• Proposed changes are reviewed in the diff editor, with Accept and Reject in its title bar. The decision goes through Sentinel's own approval (recorded, checkpointed, verified, audited).
• Unsaved work is safe. The agent will not write a file you have unsaved changes in; it asks you to save or discard first.
• The agent sees your Problems. After it edits, the errors VS Code's language servers report in those files go back to it once, to fix or explain.
• TFVC in Source Control. In a TFVC workspace, the Source Control view lists the pending changes, with gutter markers and a diff per file. Check In goes through Sentinel's own check-in (the work item your team requires, an audit record), after you confirm.
• The chat follows VS Code's theme.

The full app in your browser

The same Sentinel also runs as the full application in your browser (file tree, split viewer, a wide window), with the same chats: click the globe in the chat. Each folder keeps its address. A browser's sign-in lasts until the browser closes or VS Code closes the window, whichever comes first; the globe signs it in again.

Where things live

Everything the extension creates is under one folder, %LOCALAPPDATA%\Sentinel on Windows (~/.local/share/sentinel elsewhere, or sentinel.dataFolder):

Nothing is written into the extension's folder or into your project. Deleting runtime\ is always safe; the next start rebuilds it (Sentinel: Reinstall Python Runtime).

Settings

Commands

Sentinel: Open, Open in Editor Tab, Open in Browser, New Chat, Add to Chat, Explain This, Accept Change / Reject Change (in a proposal's diff), Sentinel TFVC: Check In / Refresh (in Source Control), Restart Server, Stop Server, Show Server Log, Open Data Folder, Reinstall Python Runtime, and Menu (the status bar item).

Privacy and security

• The server listens on 127.0.0.1 only, on a port chosen for this window, and every call to it needs a session token the extension mints per start; the panel and Open in Browser sign in through one-time links, so the bare address shows a sign-in wall. The browser's sign-in is a session cookie: it ends when the browser closes or the server stops.
• The only network traffic is the model calls you configure in Sentinel's Settings panel. The extension itself sends nothing anywhere and has no telemetry.
• The UI is shown in a webview that may frame only the local server, and the server allows only that webview to frame it; by default the server refuses to be framed at all.
• Between the editor and the chat pass paths, line numbers and short prompts, never your files' text. The webview's one script relays only a fixed list of message types, on the local server's origin only, and the extension accepts only paths inside your open folder.
• When the window closes, the server is told to stop; if VS Code dies without a word, the server notices and exits on its own.

Troubleshooting

The Sentinel output channel (Sentinel: Show Server Log) has pip's and the server's output; the last lines say what went wrong. No Python found: install one or set sentinel.pythonPath. pip install failed with a connection or certificate error: pip is not configured for your package index; set sentinel.pipIndexUrl or ask whoever manages the machine. The server did not come up: the traceback is at the end of the log; a broken agent.config.yaml in the data folder is the usual cause.

Limitations

Local windows only (no Remote-SSH, WSL or container windows). A link inside the UI to another site may not open from inside the frame; use Open in Browser.

Documentation, source and issues: https://github.com/jamesshca/coding-agent.