Issues / #3369
#3369 [GUI] Optional Local Model Extension That Survives App Updates
open · @medking82 · 0 commentaires · Sur GitHub
Setup & installModels & quantsWindows
Description
### Area GUI ### What do you want to do? Use an already running local Strata model in the desktop App without copying its GGUF files into the App's prepared-model library. Keep the official Model Catalog, recommendations and updater while an optional companion adds Current Model, Connected Local Models tiles in Manage Models/Storage, and explicit Use In Chat. ### Proposed scope A separate opt-in tool under `tools/local-model-extension/`, outside the installed App. Its versioned, integrity-checked UI package can be enabled, disabled and rolled back. A bounded Node preload supports the 0.2.x host only and leaves the official frontend/API/updater path unchanged when absent or incompatible. The optional managed Tray in #3365 remains the only server owner. A user-enabled local file protocol supplies its confirmed model/context/idle state and admits explicit context, idle, preload and unload requests through the Tray's existing drain boundary. The Extension does not create a second server. It takes arbitrary configured physical model IDs and loopback ports, without a personal model allowlist, private paths or a Codex dependency. Reasoning Effort remains the official composer control; #3363 addresses the direct API forwarding bug separately. ### Scope agreement and validation limits This is a request for maintainer scope agreement, accompanied by a draft candidate for inspection at the user's request. No maintainer approval is assumed. Default App, server and engine files stay byte-identical; native prepared counts and directories are not fabricated or rewritten. The existing private implementation was exercised in the Windows 0.2.0-beta App for connected tiles and actual Chat model selection. The portable candidate will have its own anonymous fixture screenshots and tests. Such screenshots represent simulated state, not GPU inference, capacity or timing measurements. Real GPU behavior and generic App/Tray hardware acceptance remain separate checks.
Sur le site
Liens install, modèles, releases.