Distillation

Date: Sep 16, 2026

• 🔴 (23:52) User stated the WASM heap never shrinks: once the first oversized allocation grows the heap, every later attempt inherits it; recovery must respawn the worker with a fresh heap at a lower token cap.
• 🔴 (23:52) User stated native ONNX Runtime module/binary error classification is deliberately narrow so an unrelated resolution error, such as a missing model file, is never misclassified as a backend-load failure.
• 🔴 (23:52) User stated formatted-input objects and other later arguments must never match the relevant error signature; only the leading string is classified.
• 🔴 (23:52) User stated cache healing must avoid a purge→redownload→fail→purge loop that never converges and hammers the Hugging Face Hub; transient download, authentication, and network failures are excluded from on-disk-corruption classification.
• 🔴 (23:52) User stated the native embedding path uses __LORE_ORT_BINDING_PATH__ and “never sets this”; the referenced field was not visible in the supplied excerpt.
• 🔴 (23:52) User-provided runtime code treats ONNX/WASM out-of-memory failures as unrecoverable worker state and calls for worker respawn rather than reuse of the enlarged heap.
• 🔴 (23:52) User-provided runtime code validates cached ONNX files using sizeBytes >= MIN_ONNX_FILE_BYTES && firstByte === 0x08, where 0x08 is the expected leading ONNX/protobuf header byte.
• 🔴 (23:52) User-provided build configuration creates separate Node and Bun distributions plus CommonJS/ESM embedding and vector workers; it locally ships ort-wasm-simd-threaded.mjs and ort-wasm-simd-threaded.wasm, while the single-thread WASM path never invokes its pthread new Worker(...).
• 🔴 (23:52) In /home/byk/Code/opencode-lore-embedding-runtime/packages/gateway/script/ort-platform-package.ts, user defined an esbuild-style native ONNX Runtime packaging model: matching @loreai/onnxruntime-<os>-<arch> optional packages carry the native addon and shared libraries, use npm os/cpu gating, set preferUnplugged: true for Yarn PnP, contain no install script, and resolve the addon through require.resolve("@loreai/onnxruntime-<target>/onnxruntime_binding.node"); unavailable packages fall back to bundled WASM.
• 🔴 (23:52) User stated direct reliance on onnxruntime-node installation is unsuitable because its native binary is obtained by a GitHub-release postinstall download, npm 12 will stop running that automatically, and it fails in offline, air-gapped, or proxied installs. Bundled WASM works for dist-only installs but is single-threaded and 2.7–4.1× slower than native.
• 🔴 (23:52) /home/byk/Code/opencode-lore-embedding-runtime/packages/gateway/script/ort-platform-package.ts defines 5 native package targets in this order: 1. linux-x64 (linux/x64, OS linux, CPU x64), 2. linux-arm64 (linux/arm64, OS linux, CPU arm64), 3. darwin-arm64 (darwin/arm64, OS darwin, CPU arm64), 4. win32-x64 (win32/x64, OS win32, CPU x64), 5. win32-arm64 (win32/arm64, OS win32, CPU arm64). darwin-x64 is absent because onnxruntime-node 1.27.0 dropped its Intel macOS prebuilt; Intel Macs fall back to WASM through null binding-path resolution.
• 🔴 (23:52) /home/byk/Code/opencode-lore-embedding-runtime/packages/gateway/script/ort-platform-package.ts defines ortPackageName(target) as @loreai/onnxruntime-${target}, ortPlatformTarget(platform = process.platform, arch = process.arch) as ${platform}-${arch}, and ORT_PACKAGE_BINDING_SUBPATH = ORT_BINDING_FILE.
• 🔴 (23:52) buildOrtPlatformPackages(outDir, version, platforms = ORT_NPM_PLATFORMS) recreates outDir/onnxruntime-<target>/, copies native files flat into the package root, sorts file names, and writes package.json containing name, version, loreOrtNodeVersion, license: "FSL-1.1-Apache-2.0", repository metadata, os, cpu, preferUnplugged: true, files, and publishConfig: { access: "public" }; it intentionally omits an exports map so deep require.resolve("<pkg>/onnxruntime_binding.node") remains allowed.
• 🔴 (23:52) The native package generator CLI is tsx script/ort-platform-package.ts --out <dir> [--version x.y.z]; --out is mandatory, omitted --version defaults to the gateway package version, and failures exit with status 1.
• 🔴 (23:53) /home/byk/Code/opencode-lore-embedding-runtime/packages/gateway/package.json identifies @loreai/gateway version 0.40.0, type module, license FSL-1.1-Apache-2.0, Node engine >=22.15, main ./dist/index.cjs, types ./dist/index.d.cts, Bun export ./dist/index.bun.js, and binaries lore and lore-gateway, both mapped to ./dist/bin.cjs.
• 🔴 (23:53) /home/byk/Code/opencode-lore-embedding-runtime/packages/gateway/package.json scripts are: typecheck: "tsc --noEmit", build: "tsx script/build.ts", bundle: "tsx script/bundle.ts", build:binary: "tsx script/build.ts --binary", build:binary:sea: "tsx script/build-binary-sea.ts", and start: "tsx src/index.ts".
• 🔴 (23:53) The published files list for @loreai/gateway contains exactly 10 artifacts: 1. dist/bin.cjs, 2. dist/embedding-worker.cjs, 3. dist/embedding-worker.js, 4. dist/index.bun.js, 5. dist/index.cjs, 6. dist/index.d.cts, 7. dist/ort-wasm-simd-threaded.mjs, 8. dist/ort-wasm-simd-threaded.wasm, 9. dist/vector-worker.cjs, 10. dist/vector-worker.js.
• 🔴 (23:53) @loreai/gateway runtime dependencies are @streamparser/json 0.0.26, @stricli/auto-complete 1.2.8, @stricli/core 1.2.8, @supabase/supabase-js ^2.58.0, google-auth-library ^10.9.0, jsonc-parser 3.3.1, p-limit 7, qrcode-terminal ^0.12.0, semver ^7.7.3, and sqlite-vec 0.1.9.
• 🔴 (23:53) @loreai/gateway dev dependencies include @loreai/core workspace:*, @sentry/bun ^10.52.0, @types/bun ^1.2.0, @types/qrcode-terminal ^0.12.2, @types/semver ^7.7.1, ai-tokenizer ^1.0.6, binpatch ^0.3.1, fossilize ^0.10.1, onnxruntime-node 1.27.0, tar ^7.5.21, and undici ^8.9.0.