fix(embedding): load transformers via CJS so Intel Mac onnxruntime pin applies (#210)

.github/workflows/embedding-backend.yml

@@ -15,6 +15,7 @@ on:
       - "src/services/embedding.ts"
       - "src/services/onnxruntime-resolve.ts"
       - "scripts/verify-embedding-backend.mjs"
+      - "scripts/verify-nested-onnxruntime-fixture.mjs"
       - ".github/workflows/embedding-backend.yml"
   workflow_dispatch:
 
@@ -24,9 +25,18 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        # macos-15-intel = Intel (darwin/x64); macos-15 = Apple Silicon (darwin/arm64).
-        # Intel must keep resolving onnxruntime-node@1.22.0 (#184).
-        os: [ubuntu-latest, macos-15-intel, macos-15, windows-latest]
+        # Intel: macos-15-intel / macos-26-intel (darwin/x64).
+        # Apple Silicon: macos-15 / macos-26 (darwin/arm64).
+        # Intel must keep resolving onnxruntime-node@1.22.0 (#184 / #210).
+        os:
+          [
+            ubuntu-latest,
+            macos-15-intel,
+            macos-26-intel,
+            macos-15,
+            macos-26,
+            windows-latest,
+          ]
     runs-on: ${{ matrix.os }}
     steps:
       - uses: actions/checkout@v5
@@ -72,3 +82,36 @@ jobs:
 
       - name: Embedding smoke (Node)
         run: node scripts/verify-embedding-backend.mjs
+
+  # Targeted #210 regression: OpenCode-shaped nested install on the oldest
+  # available standard Intel runner, using the reporter's Bun/Node versions.
+  nested-intel-regression:
+    name: macos-15-intel / nested OpenCode fixture
+    runs-on: macos-15-intel
+    steps:
+      - uses: actions/checkout@v5
+
+      - uses: oven-sh/setup-bun@v2
+        with:
+          bun-version: "1.3.14"
+
+      - uses: actions/setup-node@v5
+        with:
+          node-version: "22"
+          package-manager-cache: false
+
+      - name: Install source dependencies
+        run: bun install
+
+      - name: Install web dependencies
+        working-directory: web
+        run: bun install
+
+      - name: Build package
+        run: bun run build
+
+      - name: Nested fixture + embedding (Bun 1.3.14)
+        run: bun scripts/verify-nested-onnxruntime-fixture.mjs
+
+      - name: Nested fixture + embedding (Node 22)
+        run: node scripts/verify-nested-onnxruntime-fixture.mjs

.github/workflows/platform-smoke.yml

@@ -9,6 +9,7 @@ on:
       - "src/**"
       - "scripts/native-deps-smoke.mjs"
       - "scripts/verify-libsql-vector.mjs"
+      - "scripts/verify-nested-onnxruntime-fixture.mjs"
       - "scripts/smoke-test.mjs"
       - "web/**"
       - ".github/workflows/platform-smoke.yml"
@@ -21,11 +22,15 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
+        # Intel: macos-15-intel / macos-26-intel (darwin/x64).
+        # Apple Silicon: macos-15 / macos-26 (darwin/arm64).
         os:
           - ubuntu-latest
           - windows-latest
           - macos-15-intel
+          - macos-26-intel
           - macos-15
+          - macos-26
 
     steps:
       - uses: actions/checkout@v5

README.md

@@ -33,6 +33,8 @@ This plugin uses embedded Turso/libSQL with native vector indexes (`F32_BLOB`, `
 - Internet access on first use if you use the default local embedding model, because the model is downloaded by `@huggingface/transformers`.
 - For source/development installs, run `bun install` before building or testing. The published plugin package installs its runtime dependencies automatically through OpenCode.
 
+**CI-tested platforms:** Linux, Windows, macOS 15 and macOS 26 on both Intel (`darwin/x64`) and Apple Silicon (`darwin/arm64`). Older macOS releases are not excluded by that matrix; they are simply outside the current GitHub-hosted runner set.
+
 **Notes:**
 
 - Vector embeddings are stored and searched directly in Turso/libSQL; inserts update the vector index automatically.
@@ -228,7 +230,7 @@ Example — remote OpenAI embeddings:
 
 Changing `embeddingModel` (or dimensions) can trigger re-embedding of stored memories on next startup. Prefer picking a model once and sticking with it for a given data directory.
 
-**Intel Mac (`darwin/x64`):** newer `onnxruntime-node` builds may ship without an x64 native binding, so local embedding init can fail. Use a remote endpoint via `embeddingApiUrl` + `embeddingApiKey` (example above). If you stay on local embeddings after a plugin upgrade, clear OpenCode's nested plugin cache (`~/.cache/opencode/packages/opencode-mem@*`) so the install picks up the pinned runtime.
+**Intel Mac (`darwin/x64`):** newer `onnxruntime-node` builds may ship without an x64 native binding, so local embedding init can fail. `opencode-mem` pins `onnxruntime-node@1.22.0` and loads transformers through a CJS resolve shim so OpenCode nested installs keep that binding. If init still fails after a plugin upgrade, clear OpenCode's nested plugin cache (`~/.cache/opencode/packages/opencode-mem@*`) and reinstall, or use a remote endpoint via `embeddingApiUrl` + `embeddingApiKey` (example above).
 
 ### Memory Scope
 
@@ -349,7 +351,7 @@ Troubleshooting:
 - If auto-capture reports that a provider is not connected, confirm the provider name with `opencode providers list` and configure that provider in opencode first.
 - If a proxy or custom provider returns plain text instead of structured/tool output, choose another model/provider or use one of the manual provider modes above.
 - For models that reject `temperature`, add `"memoryTemperature": false` when using manual API configuration.
-- **Intel Mac (darwin/x64) local embedding:** if embedding init fails (missing onnxruntime x64 binding), switch to a remote embedding endpoint via `embeddingApiUrl` + `embeddingApiKey`, or clear `~/.cache/opencode/packages/opencode-mem@*` after upgrading so the nested install picks up the pinned `onnxruntime-node`. See [Choosing / configuring embeddings](#choosing-configuring-embeddings). MLX is not supported.
+- **Intel Mac (darwin/x64) local embedding:** if embedding init fails, clear `~/.cache/opencode/packages/opencode-mem@*` after upgrading so the nested install picks up the pinned `onnxruntime-node@1.22.0`, or switch to a remote embedding endpoint via `embeddingApiUrl` + `embeddingApiKey`. See [Choosing / configuring embeddings](#choosing-configuring-embeddings). MLX is not supported.
 
 ## Public Subpath Exports
 

scripts/verify-embedding-backend.mjs

@@ -4,10 +4,10 @@
  * feature-extraction path loads and runs the native ONNX runtime without
  * crashing on the host platform.
  *
- * This is the reproducible form of the manual checks requested when migrating
- * off @xenova/transformers: the prior revert (8fb0836) was motivated by native
- * ONNX runtime crashes under Windows + Bun, so this runs in CI across
- * ubuntu / macOS / windows to catch a regression before merge.
+ * Mirrors the production loader: prefer the CJS export so OpenCode nested
+ * installs can pin onnxruntime-node@1.22.0 via Module._resolveFilename (#210).
+ * This script deliberately does not import prepareOnnxruntimeForTransformers()
+ * because the embedding-backend workflow runs without a TypeScript build.
  *
  * Uses a tiny model (all-MiniLM-L6-v2, ~25 MB) — the goal is to exercise the
  * runtime load + a real embedding call, not to validate any specific model.
@@ -16,15 +16,59 @@
  * `node scripts/verify-embedding-backend.mjs`.
  */
 
+import { createRequire } from "node:module";
+import { dirname, join } from "node:path";
+import { existsSync, readFileSync } from "node:fs";
+
 const MODEL = "Xenova/all-MiniLM-L6-v2";
 const EXPECTED_DIMS = 384;
+const PINNED_ONNX_VERSION = "1.22.0";
 
 const runtime = typeof globalThis.Bun !== "undefined" ? "bun" : "node";
 console.log(
   `[verify-embedding] runtime=${runtime} platform=${process.platform} arch=${process.arch}`
 );
 
-const { pipeline, env } = await import("@huggingface/transformers");
+function readPackageJsonNear(entry) {
+  let dir = dirname(entry);
+  for (let i = 0; i < 6; i++) {
+    const candidate = join(dir, "package.json");
+    if (existsSync(candidate)) {
+      const parsed = JSON.parse(readFileSync(candidate, "utf8"));
+      // onnxruntime-common ships helper package.json files under dist/* without version.
+      if (typeof parsed.version === "string" && parsed.version.length > 0) {
+        return parsed;
+      }
+    }
+    dir = dirname(dir);
+  }
+  throw new Error(`versioned package.json not found near ${entry}`);
+}
+
+const requireFromHere = createRequire(import.meta.url);
+const transformersSpecifier = ["@huggingface", "transformers"].join("/");
+const { pipeline, env } = requireFromHere(transformersSpecifier);
+
+// Assert production-shaped CJS load resolved the pinned onnxruntime stack.
+const onnxEntry = requireFromHere.resolve("onnxruntime-node");
+const onnxPkg = readPackageJsonNear(onnxEntry);
+if (onnxPkg.name !== "onnxruntime-node" || onnxPkg.version !== PINNED_ONNX_VERSION) {
+  console.error(
+    `[verify-embedding] FAIL: expected onnxruntime-node@${PINNED_ONNX_VERSION}, got ${onnxPkg.name}@${onnxPkg.version} at ${onnxEntry}`
+  );
+  process.exit(1);
+}
+console.log(`[verify-embedding] onnxruntime-node@${onnxPkg.version} at ${onnxEntry}`);
+
+const commonEntry = createRequire(onnxEntry).resolve("onnxruntime-common");
+const commonPkg = readPackageJsonNear(commonEntry);
+if (commonPkg.name !== "onnxruntime-common" || commonPkg.version !== PINNED_ONNX_VERSION) {
+  console.error(
+    `[verify-embedding] FAIL: expected onnxruntime-common@${PINNED_ONNX_VERSION}, got ${commonPkg.name}@${commonPkg.version} at ${commonEntry}`
+  );
+  process.exit(1);
+}
+console.log(`[verify-embedding] onnxruntime-common@${commonPkg.version} at ${commonEntry}`);
 
 // Mirror the plugin's runtime configuration.
 env.allowLocalModels = true;