v1.12.4

Added

• You can now save a GitHub token in Settings → Models & loading. GitHub allows only 60 unauthenticated API requests per hour per IP, shared across everything TurboLLM asks it — listing branches for a build-from-source engine, checking engines for updates, and downloading new engine builds. Once that budget is spent, all of those quietly stop working until the hour rolls over. Adding a token raises the limit to 5,000 requests/hour and applies to every one of those calls. The token is write-only: it is never shown again after saving, and the app only ever reports whether one is configured. It takes effect immediately, with no restart.

Fixed

• Prebuilt llama.cpp engines could never be updated. "Manage GPU builds" reported "Update status unavailable" no matter how far behind the installed build was. llama.cpp tags every build b##### but also cuts occasional versioned releases (v0.4.0) and marks those as GitHub's "latest release" — so a build installed as b10455 was being compared against v0.4.0, two formats that cannot be ordered, and the answer was always "can't tell". Updates now compare against the newest build tag, so an out-of-date engine is correctly offered as b10455 → b10818 and installs.
• The branch dropdown offered only "main" for every build-from-source engine. The repository name was double-encoded before being sent to GitHub, so the lookup always failed and the list fell back to a single hardcoded "main" — which was also wrong for engines whose default branch is something else. The dropdown now lists the repository's real branches, and falls back to the engine's own default when GitHub cannot be reached: llama.cpp is master, KoboldCpp is concedo, Prism is prism, TurboQuant is master.
• Repositories with more than 100 branches could not be fully listed or searched. Only the first page of results was ever fetched, so the search box and "load more" could never reach the rest — llama.cpp alone has 771 branches. All of them are now listed and searchable.
• A used-up GitHub rate limit was reported as "offline". Update checks said "Couldn't check for updates (offline)" when the machine's network was perfectly fine and the real cause was the hourly API budget being spent — sending you to debug a connection that was never broken. It now says the rate limit was reached and links to where you can add a token to raise it.

Changed

• The branch list is fetched when you open the dropdown, not when the Engines screen loads. Previously, merely visiting Engines queried GitHub for every build-from-source engine listed, spending part of the hourly API budget whether or not you intended to pick a branch.