add local WebGPU provider support

.github/workflows/ci.yml

@@ -9,7 +9,7 @@ on:
 jobs:
   quality:
     name: Format, Lint, Typecheck, Test, Browser Test, Build
-    runs-on: blacksmith-4vcpu-ubuntu-2404
+    runs-on: ubuntu-24.04
     steps:
       - name: Checkout
         uses: actions/checkout@v4

README.md

@@ -1,8 +1,8 @@
 # T3 Code + Copilot
 
-This repo is a T3 Code fork that stays up to date with upstream and adds GitHub Copilot support.
+This repo is a T3 Code fork that stays up to date with upstream and adds GitHub Copilot plus browser-local WebGPU support.
 
-T3 Code is a minimal web GUI for coding agents. This fork supports both Codex and GitHub Copilot.
+T3 Code is a minimal web GUI for coding agents. This fork supports Codex, GitHub Copilot, and a browser-side Local WebGPU adapter powered by Hugging Face Transformers.js.
 
 ## Preview
 
@@ -13,17 +13,18 @@ T3 Code is a minimal web GUI for coding agents. This fork supports both Codex an
 
 - tracks upstream `pingdotgg/t3code`
 - adds GitHub Copilot provider support
+- adds a browser-side Local WebGPU provider for curated Hugging Face ONNX models
 - keeps Codex support working too
 
 ## How to use
 
 > [!WARNING]
-> You need to have either [Codex CLI](https://github.com/openai/codex) or GitHub Copilot available and authorized for T3 Code to work.
+> You need either [Codex CLI](https://github.com/openai/codex), GitHub Copilot, or a WebGPU-capable browser for the Local WebGPU adapter.
 
 The easiest way to use this fork is the desktop app.
 
 - Download it from the [releases page](https://github.com/zortos293/t3code-copilot/releases)
-- Launch the app and choose either `Codex` or `GitHub Copilot`
+- Launch the app and choose `Codex`, `GitHub Copilot`, or `Local WebGPU`
 
 You can also run it from source:
 
@@ -34,6 +35,13 @@ bun run dev
 
 Open the app, connect your provider, and start chatting.
 
+### Local WebGPU notes
+
+- Local WebGPU runs entirely in the browser and does not use the server provider runtime.
+- The first run downloads model files from Hugging Face/CDN endpoints and may take a while.
+- Use the Settings browser to search compatible Hugging Face models, then start with the smaller instruct variants for the best chance of fitting browser memory limits.
+- WebGPU availability depends on browser, OS, and GPU support. Recent Chromium-based browsers work best today.
+
 ## Some notes
 
 We are very very early in this project. Expect bugs.

apps/server/src/huggingFaceModelSearch.test.ts

@@ -0,0 +1,110 @@
+import { afterEach, describe, expect, it, vi } from "vitest";
+
+import {
+  clearHuggingFaceModelSearchCache,
+  searchHuggingFaceModels,
+} from "./huggingFaceModelSearch";
+
+describe("huggingFaceModelSearch", () => {
+  afterEach(() => {
+    clearHuggingFaceModelSearchCache();
+    vi.restoreAllMocks();
+  });
+
+  it("returns featured recommended models and caches repeated requests", async () => {
+    const fetchSpy = vi.spyOn(globalThis, "fetch").mockResolvedValue(
+      new Response(
+        JSON.stringify([
+          {
+            id: "onnx-community/Qwen2.5-1.5B-Instruct",
+            likes: 12,
+            downloads: 845,
+            private: false,
+            tags: ["transformers.js", "onnx", "text-generation", "license:apache-2.0"],
+            pipeline_tag: "text-generation",
+            library_name: "transformers.js",
+          },
+          {
+            id: "community/SmallLM-Instruct",
+            likes: 4,
+            downloads: 120,
+            private: false,
+            tags: ["transformers.js", "text-generation"],
+            pipeline_tag: "text-generation",
+            library_name: "transformers.js",
+          },
+          {
+            id: "onnx-community/Skip-Me",
+            likes: 1,
+            downloads: 2,
+            private: false,
+            tags: ["transformers.js", "custom_code"],
+            pipeline_tag: "text-generation",
+            library_name: "transformers.js",
+          },
+        ]),
+        { status: 200, headers: { "content-type": "application/json" } },
+      ),
+    );
+
+    const first = await searchHuggingFaceModels({ limit: 5 });
+    const second = await searchHuggingFaceModels({ limit: 5 });
+
+    expect(fetchSpy).toHaveBeenCalledTimes(1);
+    expect(first).toEqual({
+      mode: "featured",
+      models: [
+        {
+          id: "onnx-community/Qwen2.5-1.5B-Instruct",
+          author: "onnx-community",
+          name: "Qwen2.5-1.5B-Instruct",
+          downloads: 845,
+          likes: 12,
+          pipelineTag: "text-generation",
+          libraryName: "transformers.js",
+          license: "apache-2.0",
+          compatibility: "recommended",
+        },
+      ],
+      truncated: false,
+    });
+    expect(second).toEqual(first);
+  });
+
+  it("returns compatible search results for explicit queries", async () => {
+    vi.spyOn(globalThis, "fetch").mockResolvedValue(
+      new Response(
+        JSON.stringify([
+          {
+            id: "community/Phi-lite-Instruct",
+            likes: 14,
+            downloads: 912,
+            private: false,
+            tags: ["transformers.js", "text-generation", "license:mit"],
+            pipeline_tag: "text-generation",
+            library_name: "transformers.js",
+          },
+          {
+            id: "onnx-community/Phi-3.5-mini-instruct-onnx-web",
+            likes: 36,
+            downloads: 1500,
+            private: false,
+            tags: ["transformers.js", "onnx", "text-generation", "license:mit"],
+            pipeline_tag: "text-generation",
+            library_name: "transformers.js",
+          },
+        ]),
+        { status: 200, headers: { "content-type": "application/json" } },
+      ),
+    );
+
+    const result = await searchHuggingFaceModels({ query: "phi", limit: 10 });
+
+    expect(result.mode).toBe("search");
+    expect(result.query).toBe("phi");
+    expect(result.models.map((model) => model.id)).toEqual([
+      "onnx-community/Phi-3.5-mini-instruct-onnx-web",
+      "community/Phi-lite-Instruct",
+    ]);
+  });
+});