Added
- The stack follows the idea. With "Choose stack" on Automatic, the home
screen picks Next.js, Express, FastAPI or a background worker when the idea
mentions them (React / Vite otherwise) and says why under the prompt. Ideas
that name a framework without a ready-made stack (Vue, Svelte, Astro, Flask,
Django…) or a language the workspace does not have (PHP, Ruby, Java, .NET,
Go, Rust, native mobile) get a plain warning before anything is built. - The preview follows a change of framework. When a task edits
sandbox.yaml, runtimed applies it right away (new start command, health and
build check) instead of at the next workspace start, and the chat says what
happened. Changes that would break the preview — an invalid or removed file,
another port, switching between a web app and a background worker — are
undone and the app keeps running. A Restore applies the restored settings too.
The agent is told how to change frameworks safely.
Changed
- The console and desktop focus on building new apps from a prompt. Removed the
App library of ready-made apps, the starter templates, Import from Git and
Blank project from the home screen; the engine API still supports Git import
and empty apps.
Fixed
- Desktop setup no longer runs Windows out of memory. On a busy 16 GB
laptop, unpacking the engine images filled the engine's file cache until
Windows ran out of memory and stopped the engine mid-download. Setup now keeps
that cache small, starts the download over if a previous one was cut off
(which left empty files and no app preview), and a stopped setup says in
plain words to check the connection, close some apps and click Retry setup. - Restore after a change of framework or dependencies brings the app back.
Installed packages are not part of a restore point, so restoring a React app
after a switch to Astro left Astro's packages and the preview stayed down.
Restore now reinstalls dependencies when the files that define them changed. - A used-up free model no longer leaves the agent stuck. OpenCode only logged
its free usage limit and waited until the task timed out. The task now ends
right away with: "The free AI model has reached its usage limit for now. Try
again later, or connect your own account in Settings to keep building." - An agent can no longer take over its own workspace. A coding agent that ran
runtimed --helpstarted a second supervisor, which failed the running task
and cut the workspace off. runtimed now refuses arguments and runs only once. - A background worker restarted after a task is no longer reported as down.