One app, several processes and threads
The app is not a web page in a window. Its work is spread out so that no part can hold up another:
| Where | What runs there |
|---|---|
| The app’s main process | the window, the menu, the list of environments. It does no work of its own |
A process of its own (OpenKumi Server) |
the server, when the local environment is on. Each bot’s Chrome, and each Claude Code or Codex turn, is a further process the server starts |
| The window | the interface, loaded from the app itself (openkumi://app/), talking straight to the environment it shows |
| A worker thread in the window | the connection to the environment: it reads what arrives, hands changes to the window at most once per frame, and draws the pictures of a bot’s browser straight onto the page |
A program of its own (dictation), while you dictate |
listening, and writing down what is said, on the Mac; the app passes its words to the window |
Two things follow for speed. Text that streams in is collected on the worker and applied once per frame, and a reply being written is drawn piece by piece, so only its last paragraph is worked out again as it grows. Pictures of a bot’s browser arrive as bytes and are decoded and drawn by the worker; they never pass through the window’s own thread. Measured on the packaged app (M-series Mac): the app is usable about 0.6 s after it starts, server included, and while a 6,000-word reply streamed in beside a browser page that changed on every frame, the window kept drawing 60 frames a second with no frame longer than 19 ms.