Local embeddings fail on Alpine (musl): ERR_DLOPEN_FAILED — extend WASM fallback

[kir-ka]

Short description

onnxruntime-node native binding fails to load on Alpine (musl). Plugin should fall back to the WASM runtime it already uses for Electron instead of disabling local embeddings.

What happened?

On Alpine Linux (musl libc), embedding.provider: "local" fails because onnxruntime-node ships only glibc prebuilt binaries. The plugin already detects this exact failure via isNativeRuntimeMissingError() and — instead of falling back to the WASM runtime it already uses for Electron — simply disables local embeddings.

Environment:

• OS: Alpine Linux (Docker), official image ghcr.io/anomalyco/opencode:latest
• libc: musl (no __vsnprintf_chk)
• Node.js: v24.18.1
• opencode: 1.18.19
• magic-context: v0.40.0
• Note: the attached diagnostic report was captured with embedding.provider: "off" (the current workaround). The bug reproduces by setting embedding.provider: "local".

Error (doctor --harness opencode):

Embedding provider: local — onnxruntime-node native binding missing —
binding failed to load: ERR_DLOPEN_FAILED: Error relocating
.../onnxruntime-node/bin/napi-v6/linux/x64/libonnxruntime.so.1:
__vsnprintf_chk: symbol not found

Root cause:
__vsnprintf_chk is glibc-specific; Alpine's musl does not provide it, so dlopen() of the native libonnxruntime.so.1 fails. Upstream has no musl prebuilt: microsoft/onnxruntime#9483.

What the code already does:
embedding-local.ts already injects the WASM runtime via injectWasmOrtForElectron() (guarded by process.versions?.electron) and detects this exact failure via isNativeRuntimeMissingError(). So detection is present, but the response is "disable" rather than "use WASM".

Steps to reproduce

1. docker run -it ghcr.io/anomalyco/opencode:latest
2. Set embedding.provider: "local" in magic-context.jsonc
3. Run: npx @cortexkit/magic-context@latest doctor --harness opencode
4. Observe ERR_DLOPEN_FAILED: __vsnprintf_chk: symbol not found

Proposed change

When isNativeRuntimeMissingError() fires on a non-Electron runtime (musl/Alpine), fall back to the existing WASM injection instead of disabling local embeddings.

Related

• Embedding model fails to load in Electron - transformers.node.mjs ERR_MODULE_NOT_FOUND (v0.19.0+) #78 — Electron WASM injection path
• PR fix(embedding): guard Windows Bun and support trusted headers #342 — related embedding guard (not merged)
• Javascript Node Alpine microsoft/onnxruntime#9483 — Node.js on Alpine, no prebuilt

Diagnostics

magic-context-issue-20260824-182207-full.md

Plugin version

v0.40.0

OpenCode version

1.18.19

Platform

Alpine Linux (Docker) — musl libc

Client

OpenCode web UI (browser, via serve on Docker)

[magic-alfonso]

Implemented on master, riding the next release: the local provider now tries the native ONNX runtime first and, when the binding cannot load (your Alpine/musl ERR_DLOPEN_FAILED case), retries once with the bundled onnxruntime-web/WASM runtime instead of disabling itself. The fallback decision is per-process sticky and logged. Honest caveat: WASM inference is noticeably slower than the native CPU path — for larger memory pools an openai-compatible remote embedding provider will be faster; the local WASM path is there so musl environments work out of the box. doctor now reports which runtime state you're in (native / WASM fallback / both unavailable), so if anything still misbehaves after the release, a fresh doctor --issue bundle will show it directly.