One retrieval engine; every local engine picks a GPU that works and recovers from faults

.github/workflows/publish-performance-pack.yml

@@ -70,7 +70,12 @@ jobs:
         run: ./scripts/fetch-win-binaries.ps1
       - name: Build Kev server runtime (model weights excluded)
         shell: pwsh
-        run: ./scripts/build-kev-runtime.ps1
+        run: |
+          ./scripts/build-kev-runtime.ps1
+          . ./scripts/config/windows-vc-runtime.ps1
+          Get-ChildItem resources/bin/kev-runtime -Recurse -File -Filter '*.exe' |
+            Select-Object -ExpandProperty DirectoryName -Unique |
+            ForEach-Object { Copy-VCRuntime $_ }
       - name: Check staged Windows runtimes
         shell: bash
         run: |
@@ -160,6 +165,13 @@ jobs:
           rm -rf "$BIN_ROOT/llama-cuda" "$BIN_ROOT/cuda-runtime"
           cp -a "$RUNNER_TEMP/engine/llama-cuda" "$RUNNER_TEMP/engine/cuda-runtime" "$BIN_ROOT/"
           cat "$RUNNER_TEMP/engine/engine.json"
+          if [ "$PLATFORM" = win32 ]; then
+            # llama-cuda was staged on Linux; reuse the VC redist from the Windows job.
+            for dll in MSVCP140.dll MSVCP140_1.dll MSVCP140_2.dll MSVCP140_ATOMIC_WAIT.dll \
+              MSVCP140_CODECVT_IDS.dll VCRUNTIME140.dll VCRUNTIME140_1.dll CONCRT140.dll VCOMP140.dll; do
+              cp "$BIN_ROOT/whisper/$dll" "$BIN_ROOT/llama-cuda/$dll"
+            done
+          fi
       - name: Check the Linux llama engines' libraries resolve
         if: inputs.platform == 'linux'
         run: |
@@ -183,7 +195,8 @@ jobs:
         run: |
           python3 -m pip install --no-cache-dir --no-deps --target "$RUNNER_TEMP/onnx-nvidia" \
             'nvidia-cuda-runtime==13.2.86' 'nvidia-cublas==13.2.2.2' \
-            'nvidia-curand==10.4.2.66' 'nvidia-cudnn-cu13==9.23.1.3'
+            'nvidia-curand==10.4.2.66' 'nvidia-cudnn-cu13==9.23.1.3' \
+            'nvidia-cufft==12.2.0.57' 'nvidia-cuda-nvrtc==13.2.86'
           mkdir -p "$BIN_ROOT/onnx-cuda"
           for folder in cu13 cudnn; do
             find "$RUNNER_TEMP/onnx-nvidia/nvidia/$folder/lib" -maxdepth 1 -type f -name '*.so.*' \

.github/workflows/release.yml

@@ -122,6 +122,8 @@ jobs:
   build-mac:
     needs: version
     runs-on: macos-15 # Standard arm64 runner; matches the pinned native payload.
+    env:
+      MACOSX_DEPLOYMENT_TARGET: '13.0'
     steps:
       - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
         with:
@@ -139,13 +141,12 @@ jobs:
             'https://github.com/off-grid-ai/OGAD/releases/download/native-deps-2026-09/ogad-mac-native-2026-09-arm64-v1.tar.gz'
           echo 'ef1880daac727595bf825a254b1ebbc2a0c15c0cecd8da422c499b1568db0ad0  '"$RUNNER_TEMP/mac-native.tar.gz" | shasum -a 256 -c -
           tar -xzf "$RUNNER_TEMP/mac-native.tar.gz" -C resources/bin
+      - name: Build macOS image engine for macOS 13
+        run: bash scripts/build-sd-mac.sh
       - name: Verify cached macOS engines
         run: |
           for variant in llama llama-prism; do
             test -x "resources/bin/$variant/llama-server"
-            minos=$(otool -l "resources/bin/$variant/llama-server" | awk '/LC_BUILD_VERSION/{f=1} f&&/minos/{print $2; exit}')
-            test -n "$minos"
-            case "$minos" in 1[0-3].*|[1-9].*) ;; *) echo "Unsupported $variant minos $minos" >&2; exit 1 ;; esac
             DYLD_LIBRARY_PATH="resources/bin/$variant" "resources/bin/$variant/llama-server" --help >/dev/null
           done
           test -x resources/bin/whisper/whisper-cli
@@ -283,6 +284,7 @@ jobs:
           token: ${{ secrets.CI_CROSS_REPO_TOKEN }}
           path: _executorch_speech
           lfs: true # Mac releases bundle the default English voices for offline first use.
+          submodules: recursive
           persist-credentials: false
       - name: Put the speech runtime beside this checkout
         run: |
@@ -292,6 +294,17 @@ jobs:
           fi
           rm -rf ../executorch-speech
           mv _executorch_speech ../executorch-speech
+      - name: Build macOS speech runtime for macOS 13
+        run: |
+          npm --prefix ../executorch-speech ci
+          # Use the bundled Python and its existing torchgen, YAML, and Jinja dependencies.
+          speech_python="$GITHUB_WORKSPACE/resources/bin/kev-runtime/python/bin/python3"
+          "$speech_python" -c 'import torchgen, yaml, jinja2'
+          cmake -S ../executorch-speech/native -B ../executorch-speech/native/build \
+            -DCMAKE_BUILD_TYPE=Release -DCMAKE_OSX_DEPLOYMENT_TARGET=13.0 \
+            -DFLATCC_ALLOW_WERROR=OFF -DPYTHON_EXECUTABLE="$speech_python"
+          cmake --build ../executorch-speech/native/build --config Release --parallel 4 \
+            --target executorch-speech
       - name: Let npm read the private UI package (core and Pro)
         # @offgrid/operator-ui is a private Wednesday package that core and Pro both render with
         # (skin, primitives, motion). It is listed as optional, so npm would skip it silently
@@ -382,14 +395,15 @@ jobs:
           printf '%s' "$APPLE_API_KEY_CONTENT" > "$RUNNER_TEMP/AuthKey.p8"
           export APPLE_API_KEY="$RUNNER_TEMP/AuthKey.p8"
           npm run prepare:speech-defaults
+          node scripts/check-macos-minos.mjs 13.0
           # Build every artifact with publication disabled. The artifact hook verifies
           # both the updater ZIP and DMG before this command can complete.
           npx electron-builder --mac \
-            -c.dmg.artifactName='OffGrid-${version}.${ext}' \
+            -c.dmg.artifactName='OffGrid-${version}-${arch}.${ext}' \
             --publish never
           # electron-builder notarizes the app before it creates the DMG. Submit the
           # completed, signed DMG too, then staple its ticket for offline Gatekeeper.
-          DMG="dist/OffGrid-$VERSION.dmg"
+          DMG="dist/OffGrid-$VERSION-arm64.dmg"
           xcrun notarytool submit "$DMG" \
             --key "$APPLE_API_KEY" \
             --key-id "$APPLE_API_KEY_ID" \
@@ -485,7 +499,7 @@ jobs:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           VERSION: ${{ needs.version.outputs.version }}
         run: |
-          cp "dist/OffGrid-$VERSION.dmg" dist/OffGrid-latest.dmg
+          cp "dist/OffGrid-$VERSION-arm64.dmg" dist/OffGrid-latest.dmg
           gh release upload "v$VERSION" dist/OffGrid-latest.dmg --clobber
       # Constant nightly download link. /releases/latest EXCLUDES pre-releases, so it
       # can never point at a beta build (that is exactly what keeps the STABLE link
@@ -499,7 +513,7 @@ jobs:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           VERSION: ${{ needs.version.outputs.version }}
         run: |
-          cp "dist/OffGrid-$VERSION.dmg" dist/OffGrid-nightly.dmg
+          cp "dist/OffGrid-$VERSION-arm64.dmg" dist/OffGrid-nightly.dmg
           # Ensure the rolling 'nightly' pre-release exists (kept separate from the
           # versioned -beta.N releases electron-builder makes for the updater feed),
           # then clobber its dmg so the link always serves the freshest nightly.
@@ -678,6 +692,14 @@ jobs:
           fi
       - name: Install dependencies
         run: npm ci
+      - name: Install Vulkan SDK for Whisper
+        uses: humbletim/install-vulkan-sdk@v1.2
+        with:
+          version: '1.4.321.1'
+          cache: true
+      - name: Build Windows Vulkan Whisper
+        shell: pwsh
+        run: ./scripts/build-whisper-vulkan-win.ps1
       - name: Fetch Windows native binaries (llama/whisper/sd/ffmpeg)
         shell: pwsh
         env:
@@ -694,7 +716,7 @@ jobs:
           GOOGLE_CLIENT_ID: ${{ secrets.GOOGLE_CLIENT_ID }}
           GOOGLE_CLIENT_SECRET: ${{ secrets.GOOGLE_CLIENT_SECRET }}
           MICROSOFT_CLIENT_ID: ${{ vars.MICROSOFT_CLIENT_ID }}
-        run: node scripts/stage-native.mjs && npx electron-vite build && node scripts/check-windows-onnx-preload.mjs
+        run: node scripts/stage-native.mjs && npx electron-vite build
       - name: Package Windows web installer, payload, and update metadata
         shell: bash
         env:
@@ -708,10 +730,10 @@ jobs:
         shell: bash
         run: |
           test -s dist/win-unpacked/resources/performance-packs.json
-          for binary in llama/llama-server.exe llama-cpu/llama-server.exe llama-prism/llama-server.exe llama-prism-cpu/llama-server.exe whisper-cpu/whisper-cli.exe sd/sd-cli.exe sd-cpu/sd-cli.exe ffmpeg.exe; do
+          for binary in llama/llama-server.exe llama-cpu/llama-server.exe llama-prism/llama-server.exe llama-prism-cpu/llama-server.exe whisper-cpu/whisper-cli.exe whisper-vulkan/whisper-cli.exe whisper-vulkan/whisper-server.exe sd/sd-cli.exe sd-cpu/sd-cli.exe ffmpeg.exe; do
             test -s "dist/win-unpacked/resources/bin/$binary"
           done
-          for folder in llama-cuda llama-prism-cuda cuda-runtime sd-cuda whisper kev-runtime; do
+          for folder in llama-cuda llama-prism-cuda cuda-runtime sd-cuda whisper kev-runtime onnx-cuda; do
             test ! -e "dist/win-unpacked/resources/bin/$folder"
           done
       - name: Report Windows native payload sizes
@@ -925,6 +947,14 @@ jobs:
         run: npm ci
       - name: Install Linux GPU runtime dependencies
         run: sudo apt-get update -qq && sudo apt-get install -y libgomp1 libvulkan1 libvulkan-dev glslc spirv-headers unzip
+      - name: Stage Vulkan loaders beside cached Linux engines
+        run: |
+          source="$(ldconfig -p | awk '$1 == "libvulkan.so.1" && /x86-64/ && !found { found=$NF } END { print found }')"
+          test -n "$source"
+          for folder in llama llama-prism whisper; do
+            install -m 755 "$(readlink -f "$source")" "build/linux-bin/$folder/libvulkan.so.1"
+            test -s "build/linux-bin/$folder/libvulkan.so.1"
+          done
       - name: Typecheck Linux main process
         env:
           OFFGRID_FORCE_CORE: '0'
@@ -953,8 +983,21 @@ jobs:
       - name: Verify Linux package inputs
         run: |
           test -s dist/linux-unpacked/resources/performance-packs.json
+          provider="$(find dist/linux-unpacked/resources/app.asar.unpacked/node_modules \
+            -path '*/onnxruntime-node/*' -name libonnxruntime_providers_cuda.so -type f -print -quit)"
+          test -n "$provider"
+          test -s "$provider"
+          for folder in llama llama-prism whisper; do
+            test -s "dist/linux-unpacked/resources/bin/$folder/libvulkan.so.1"
+          done
+          while IFS= read -r -d '' binary; do
+            if file "$binary" | grep -q 'ELF.*executable'; then
+              glibc="$(objdump -T "$binary" | grep -o 'GLIBC_[0-9.]*' | sort -V | tail -1 || true)"
+              printf '[ci] %s highest GLIBC: %s\n' "$binary" "${glibc:-none (static binary)}"
+            fi
+          done < <(find dist/linux-unpacked/resources/bin -type f -print0)
           cmp resources/bin/kev-local-server.py dist/linux-unpacked/resources/bin/kev-local-server.py
-          for folder in llama-cuda llama-prism-cuda cuda-runtime sd-cuda whisper-cuda kev-runtime; do
+          for folder in llama-cuda llama-prism-cuda cuda-runtime sd-cuda whisper-cuda kev-runtime onnx-cuda; do
             test ! -e "dist/linux-unpacked/resources/bin/$folder"
           done
           for variant in llama llama-cpu llama-prism llama-prism-cpu; do
@@ -1007,9 +1050,6 @@ jobs:
             const root = path.resolve("dist/linux-unpacked/resources/app.asar/node_modules")
             const Database = require(path.join(root, "better-sqlite3-multiple-ciphers"))
             new Database(":memory:").close()
-            if (typeof require(path.join(root, "@lancedb/lancedb")).connect !== "function") {
-              throw new Error("LanceDB native package did not load")
-            }
             if (typeof require(path.join(root, "sharp")) !== "function") {
               throw new Error("Sharp native package did not load")
             }
@@ -1032,7 +1072,7 @@ jobs:
             test -n "$bundled"
             cmp "dist/linux-unpacked/resources/bin/$variant/llama-server" "$bundled"
           done
-          for helper in whisper/whisper-cli whisper-cpu/whisper-cli whisper/LICENSE ffmpeg licenses/ffmpeg.txt sd/sd-cli sd/sd-server sd/libggml-vulkan.so sd/libgomp.so.1 sd/libvulkan.so.1 executorch-speech licenses/libgomp1.txt licenses/libvulkan1.txt; do
+          for helper in llama/libvulkan.so.1 llama-prism/libvulkan.so.1 whisper/libvulkan.so.1 whisper/whisper-cli whisper-cpu/whisper-cli whisper/LICENSE ffmpeg licenses/ffmpeg.txt sd/sd-cli sd/sd-server sd/libggml-vulkan.so sd/libgomp.so.1 sd/libvulkan.so.1 executorch-speech licenses/libgomp1.txt licenses/libvulkan1.txt; do
             bundled="$(find "$extraction/squashfs-root" -path "*/resources/bin/$helper" -type f -print -quit)"
             test -n "$bundled"
             cmp "dist/linux-unpacked/resources/bin/$helper" "$bundled"
@@ -1049,7 +1089,7 @@ jobs:
             test -n "$bundled"
             cmp "dist/linux-unpacked/resources/bin/$variant/llama-server" "$bundled"
           done
-          for helper in whisper/whisper-cli whisper-cpu/whisper-cli whisper/LICENSE ffmpeg licenses/ffmpeg.txt sd/sd-cli sd/sd-server sd/libggml-vulkan.so sd/libgomp.so.1 sd/libvulkan.so.1 executorch-speech licenses/libgomp1.txt licenses/libvulkan1.txt; do
+          for helper in llama/libvulkan.so.1 llama-prism/libvulkan.so.1 whisper/libvulkan.so.1 whisper/whisper-cli whisper-cpu/whisper-cli whisper/LICENSE ffmpeg licenses/ffmpeg.txt sd/sd-cli sd/sd-server sd/libggml-vulkan.so sd/libgomp.so.1 sd/libvulkan.so.1 executorch-speech licenses/libgomp1.txt licenses/libvulkan1.txt; do
             bundled="$(find "$extraction/deb" -path "*/resources/bin/$helper" -type f -print -quit)"
             test -n "$bundled"
             cmp "dist/linux-unpacked/resources/bin/$helper" "$bundled"

docs/RELEASE_DESKTOP.md

@@ -240,3 +240,9 @@ quickly. No EV needed.
 
 - [ ] Windows: bundled llama-server doesn't start — fix in progress separately.
 - [ ] Azure Trusted Signing validated + `azureSignOptions` wired (§4) — resumes with Windows.
+
+### Linux AppImage
+
+The AppImage and deb packages are x64 only. On Ubuntu 24.04, install
+`libfuse2t64` before you start the AppImage. Headless servers and some sandboxes
+need the `--no-sandbox` option.

e2e/app068-project-knowledge-isolation.spec.ts

@@ -125,9 +125,11 @@ async function createProject(project: string, document: string, content: string)
   fs.writeFileSync(filePath, content)
   await chooseFile(filePath)
   await page.getByRole('button', { name: 'Add files', exact: true }).click()
-  await expect(page.getByText(`${document}: indexed`, { exact: true })).toBeVisible({
+  // Indexed: the row's progress bar is gone and it shows the ready line ("text  ·  62 B").
+  await expect(page.getByRole('progressbar', { name: `Indexing ${document}` })).toBeHidden({
     timeout: 60_000
   })
+  await expect(page.getByText(/^text\s+·\s+\d+ B$/).first()).toBeVisible({ timeout: 60_000 })
   await expect(page.getByText(document, { exact: true })).toBeVisible()
 }
 

e2e/fixtures/gpu-fault-llama-server.mjs

@@ -0,0 +1,124 @@
+#!/usr/bin/env node
+
+/* eslint-disable @typescript-eslint/explicit-function-return-type -- executable JavaScript fixture */
+
+// The GPU fault journey's only fake: the native llama-server boundary. Everything that decides what
+// happens after a fault (the HTTP client's classifier, the chat fallback policy, the same-model
+// restart, the backend ladder, the CPU-only launch flags, the runtime status and Settings) is
+// production code. Like a real engine, this one reports where it put the model in its load log,
+// and like a GPU that has lost its context it answers "500 Compute error" for as many questions as
+// faults.json beside it still lists. A fault is spent when it is answered, across restarts.
+
+import fs from 'node:fs'
+import http from 'node:http'
+import path from 'node:path'
+import { fileURLToPath } from 'node:url'
+
+const args = process.argv.slice(2)
+const portFlag = Math.max(args.indexOf('--port'), args.indexOf('-p'))
+const port = portFlag >= 0 ? Number(args[portFlag + 1]) : 8439
+const deviceFlag = args.indexOf('--device')
+const cpuOnly = deviceFlag >= 0 && args[deviceFlag + 1] === 'none'
+const here = path.dirname(fileURLToPath(import.meta.url))
+const FAULTS = path.join(here, 'faults.json')
+const LOG = path.join(here, 'launches.jsonl')
+
+fs.appendFileSync(LOG, `${JSON.stringify({ args })}\n`)
+
+// The load log the app reads to name the backend and the layers on the GPU.
+if (cpuOnly) {
+  process.stderr.write('load_tensors: CPU model buffer size = 120.00 MiB\n')
+} else {
+  process.stderr.write('load_tensors: offloaded 28/28 layers to GPU\n')
+  process.stderr.write('load_tensors: Metal model buffer size = 120.00 MiB\n')
+}
+
+function takeFault() {
+  try {
+    const plan = JSON.parse(fs.readFileSync(FAULTS, 'utf8'))
+    if (!(plan.remaining > 0)) return false
+    fs.writeFileSync(FAULTS, JSON.stringify({ remaining: plan.remaining - 1 }))
+    return true
+  } catch {
+    return false
+  }
+}
+
+function lastUser(messages) {
+  const user = [...(messages ?? [])].reverse().find((message) => message.role === 'user')
+  const content = user?.content
+  return typeof content === 'string'
+    ? content
+    : Array.isArray(content)
+      ? content.map((part) => part?.text ?? '').join('')
+      : ''
+}
+
+const delta = (payload) => `data: ${JSON.stringify({ choices: [{ delta: payload }] })}\n\n`
+
+const server = http.createServer((request, response) => {
+  if (request.method === 'GET' && request.url === '/health') {
+    response.writeHead(200, { 'Content-Type': 'application/json' })
+    response.end(JSON.stringify({ status: 'ok' }))
+    return
+  }
+  if (request.method === 'GET' && request.url === '/v1/models') {
+    response.writeHead(200, { 'Content-Type': 'application/json' })
+    response.end(JSON.stringify({ data: [{ id: 'gpu-fault-model' }] }))
+    return
+  }
+  if (request.method !== 'POST' || !String(request.url).includes('/chat/completions')) {
+    response.writeHead(404)
+    response.end()
+    return
+  }
+  let body = ''
+  request.on('data', (chunk) => {
+    body += chunk
+  })
+  request.on('end', () => {
+    let payload = {}
+    try {
+      payload = JSON.parse(body)
+    } catch {
+      /* answer the default */
+    }
+    const question = lastUser(payload.messages)
+    // Only the journey's own questions fault; follow-up suggestions and titles answer normally.
+    const faulted = question.startsWith('GPU check') && takeFault()
+    fs.appendFileSync(
+      LOG,
+      `${JSON.stringify({ question: question.slice(0, 80), stream: payload.stream !== false, faulted, cpuOnly })}\n`
+    )
+    if (faulted) {
+      response.writeHead(500, { 'Content-Type': 'application/json' })
+      response.end(
+        JSON.stringify({ error: { code: 500, message: 'Compute error', type: 'server_error' } })
+      )
+      return
+    }
+    const text = question.startsWith('GPU check')
+      ? `Answered on the ${cpuOnly ? 'CPU' : 'GPU'}.`
+      : 'Okay.'
+    if (payload.stream === false) {
+      response.writeHead(200, { 'Content-Type': 'application/json' })
+      response.end(
+        JSON.stringify({
+          choices: [
+            { index: 0, finish_reason: 'stop', message: { role: 'assistant', content: text } }
+          ]
+        })
+      )
+      return
+    }
+    response.writeHead(200, { 'Content-Type': 'text/event-stream', 'Cache-Control': 'no-cache' })
+    response.write(delta({ content: text }))
+    response.write(
+      `data: ${JSON.stringify({ choices: [{ delta: {}, finish_reason: 'stop' }] })}\n\n`
+    )
+    response.write('data: [DONE]\n\n')
+    response.end()
+  })
+})
+
+server.listen(port, '127.0.0.1')