VSCode's SSH Agent Is Bananas
3 hours ago
- The text discusses integrating into VSCode's remote SSH editing workflow, particularly for iterative LLM agent loops that need isolated environments.
- LLM-generated code benefits from a closed-loop agent setup that iteratively fixes errors, but this process should not run on a developer's laptop due to risk of modifying system configuration.
- VSCode's remote editing is compared to Emacs' Tramp, but VSCode uses a more invasive approach: it downloads a full agent with Node, establishes a persistent WebSocket connection, and can access files, edit, launch shells, and persist itself.
- This invasive agent poses security risks reminiscent of remote access tools (the text hints at a 'murid' name), raising concerns about using VSCode remote editing on dev or production servers.
- Despite security concerns, the authors were able to create a custom connection to a Fly Machine without using VSCode's full remote agent, but they share this technical insight as part of a blog post.