How LangAlpha Works

You run on a computer: one sandbox machine that belongs to the user. Each of the user's workspaces on it is a folder, side by side at the computer root; yours is your working directory, so the root is ... A thread is one conversation in one workspace, and a turn is one user message and everything done to answer it. Workspaces on one computer share its disk, OS user, CPUs, /tmp and installed packages, and can read each other's folders.

Reference files

• The user points back at an earlier conversation or at work done in another thread or workspace; a compaction summary; a background task's steps; a tool result saved to a file: .agents/skills/langalpha-doc/references/history.md.
• Where things live, how a path resolves, why Grep or Glob missed a file, memory and memo storage, what the user sees in the file panel: .agents/skills/langalpha-doc/references/files.md.
• Processes, output and timeouts, CPUs and disk, installed packages, what survives a stop or a rebuild, data-server tools from code, secrets: .agents/skills/langalpha-doc/references/computer.md.
• Installing, writing or changing a skill, here or for every workspace: .agents/skills/langalpha-doc/references/skills.md.
• The user wants a new data source, MCP server, brokerage connection or plugin, a server needs a key, or a server you expected is missing: .agents/skills/langalpha-doc/references/plugins.md.
• Talking with the user in a chat app, or sending a message or file to Slack, Discord, Telegram, iMessage or Feishu: .agents/skills/langalpha-doc/references/chat.md.

Traps