Crash: SIGBUS/SIGTRAP in onnxruntime-node ReleaseIoBinding during local embedding load (concurrent opencode sessions, macOS arm64) #21

Description

Both opencode TUI sessions crash with a native SIGBUS / SIGTRAP inside onnxruntime-node while the magic-context plugin loads the local embedding model (Xenova/all-MiniLM-L6-v2). The crashes occurred in two concurrently-running opencode processes (separate sessions, separate cwd) within seconds of each other — strongly suggests a concurrency / double-free bug in the local embedding loader when two processes touch the same cached model at the same time.

Environment

• Plugin: @cortexkit/opencode-magic-context@0.9.1
• OpenCode: 1.4.7
• OS: macOS 26.4.1 (build 25E253)
• Arch: arm64 (Apple Silicon, M2 Max — Mac14,6)
• Node: v24.15.0
• onnxruntime-node: 1.21.0 (npm) — bundles libonnxruntime.1.14.0.dylib (native C++ lib)
• Transitive dep via: @huggingface/transformers@~3.7.6

Configuration

~/.config/opencode/magic-context.jsonc:

{
  "$schema": "https://raw.githubusercontent.com/cortexkit/opencode-magic-context/master/assets/magic-context.schema.json",
  "enabled": true,
  "historian": {
    "model": "anthropic/claude-haiku-4-5",
    "fallback_models": ["opencode-go/glm-5"]
  },
  "dreamer": {
    "enabled": true,
    "model": "anthropic/claude-sonnet-4-6"
  },
  "sidekick": {
    "model": "anthropic/claude-haiku-4-5"
  },
  "experimental": {
    "user_memories": { "enabled": false },
    "pin_key_files": { "enabled": false }
  },
  "compaction_markers": false
}

No explicit embedding block, so the default (provider: "local", model Xenova/all-MiniLM-L6-v2) is used.

Reproduction

1. macOS 26.4.1 arm64, opencode 1.4.7, magic-context 0.9.1 (default local embedding config).
2. Cached model already present at ~/.local/share/opencode/storage/plugin/magic-context/models/Xenova/all-MiniLM-L6-v2/onnx/model.onnx (90 MB, md5 = 84f837de2a0f667784facf2ba0f36b22).
3. Start two opencode sessions in separate terminals, in different project directories:
   • opencode -s ses_267e0c216ffeI6nQIAWBnt6G20
   • opencode -s ses_26616a332ffenMsmj6S59ueIH1
4. Work in both simultaneously (both sessions presumably trigger embedding/ctx_search/dreamer init near the same time).
5. Within ~5 minutes, both processes die silently, no TUI error, parent shell prints nothing.

Three crash reports were generated in 17 minutes (10:56, 11:12:07, 11:12:12), all with the same native-frame signature.

Crash signatures (from ~/Library/Logs/DiagnosticReports/opencode-*.ips)

Crash 1 — opencode-2026-04-17-105636.ips

exception  : EXC_BAD_ACCESS (SIGBUS)
subtype    : KERN_PROTECTION_FAILURE at 0x00000004138ed418
termination: Bus error: 10
faultingThread: "Worker"
frames (top):
  OrtApis::ReleaseIoBinding(OrtIoBinding*) + 24
  InferenceSessionWrap::LoadModel(Napi::CallbackInfo const&) + 3220
  Napi::InstanceWrap<InferenceSessionWrap>::InstanceMethodCallbackWrapper::lambda()
  Napi::InstanceWrap<InferenceSessionWrap>::InstanceMethodCallbackWrapper

Crash 2 — opencode-2026-04-17-111207.ips

exception  : EXC_BREAKPOINT (SIGTRAP)
termination: Trace/BPT trap: 5
faultingThread: "Worker"
frames (top):
  (native, unsymbolicated + _sigtramp)
  InferenceSessionWrap::LoadModel(Napi::CallbackInfo const&) + 3184
  Napi::InstanceWrap<InferenceSessionWrap>::InstanceMethodCallbackWrapper::lambda()
  Napi::InstanceWrap<InferenceSessionWrap>::InstanceMethodCallbackWrapper

Crash 3 — opencode-2026-04-17-111212.ips

exception  : EXC_BREAKPOINT (SIGTRAP)
termination: Trace/BPT trap: 5
faultingThread: "Worker"
frames (top):
  (native, unsymbolicated + _sigtramp)
  OrtApis::ReleaseIoBinding(OrtIoBinding*) + 36   ← double frame!
  OrtApis::ReleaseIoBinding(OrtIoBinding*) + 36   ← double-free signature
  InferenceSessionWrap::LoadModel(Napi::CallbackInfo const&) + 3220
  Napi::InstanceWrap<InferenceSessionWrap>::InstanceMethodCallbackWrapper::lambda()

Loaded native modules at crash time (from usedImages):

/…/onnxruntime_binding.node
/…/libonnxruntime.1.14.0.dylib

Root-cause hypothesis

All three crashes share:

1. The same faulting thread name ("Worker") — likely a transformers.js / onnxruntime-node worker pool thread.
2. The same top-of-stack function: OrtApis::ReleaseIoBinding inside InferenceSessionWrap::LoadModel.
3. Crash 3 shows the ReleaseIoBinding frame twice → double-free; Crash 1 shows KERN_PROTECTION_FAILURE on a write → use-after-free.

The most plausible cause: when two opencode processes share the same on-disk ONNX model cache and both initialize InferenceSession at roughly the same wall-clock time, ONNX Runtime 1.14.0's session-initialization path releases/cleans up an IoBinding that the other process (or its own abort path) also releases — double-free → ORT_ENFORCE → brk 1 → SIGTRAP. The 1.14.0 native lib is quite old (Feb 2023) relative to macOS 26.4.1.

Suggested fixes (in order of effort)

1. Lazy + singleton-initialized embedding session — ensure only one InferenceSession per process, behind an async-locked init. If there's already a lock, it may not be covering the full LoadModel → ready path.
2. Per-process model cache path (or file-lock the first-time extraction) — avoid two processes mmap-ing / opening the same .onnx file during a warm-start window.
3. Bump @huggingface/transformers → newer versions pull onnxruntime-node ≥ 1.18 which carries native lib ≥ 1.18 with many post-1.14 stability fixes on Apple Silicon + newer macOS.
4. Graceful fallback — wrap await pipeline(...) / InferenceSession.create in try/catch; on native crash signal or init error, auto-disable embeddings for this process instead of letting it SIGTRAP-kill the entire opencode TUI. (Currently there's zero user-visible error — the whole TUI just vanishes.)
5. Document the workaround — embedding: { provider: "off" } in magic-context.jsonc prevents the crash (ctx_search still works via FTS5 fallback). Worth noting in README until the root fix lands.

Workaround applied locally

Setting embedding.provider: "off" in ~/.config/opencode/magic-context.jsonc bypasses the crash entirely. FTS5 full-text fallback keeps ctx_search functional; only semantic ranking is lost.

Additional context

Happy to attach the full .ips files, run with any DEBUG=* env var, or test a patched build. Let me know if you want the full thread dumps — I can upload them as a gist.