perf(embedding): embed gzipped ONNX runtime, gunzip at first use

.gitignore

@@ -95,3 +95,6 @@ docs/dist
 plugin/.claude-plugin/plugin.json
 plugin/.claude-plugin/marketplace.json
 user-prompt
+
+# Compressed ONNX runtime libs are produced by scripts/download-onnx-libs.sh
+internal/embedding/assets/lib/**/*.gz

experiments.md

@@ -10,3 +10,9 @@ Mutate: Go code and build flags. Frozen: public API, DB schema, plugin manifest.
 |----|----------|-------------|---------------|-------|
 | 01 | baseline | 102698416 | - | base |
 | 02 | pkg/hooks: raw HTTP/1.0 client replaces net/http | 87603952 | -14.7% | KEEP |
+| 03 | statusline: shared client, drop net/http | 84531056 | -3.5% vs 02 | KEEP |
+| 04 | embed gzipped ONNX runtime libs, gunzip at first use | 57216880 | -32.3% vs 03 | KEEP |
+
+Notes:
+- Trial 04 needs `scripts/download-onnx-libs.sh` to produce `.gz` files; existing raw libs trigger a re-download.
+- Extracted library verified byte-identical to the raw dylib (`cmp`).

internal/embedding/lib_darwin_arm64.go

@@ -6,7 +6,7 @@ import (
 	_ "embed"
 )
 
-//go:embed assets/lib/darwin-arm64/libonnxruntime.dylib
+//go:embed assets/lib/darwin-arm64/libonnxruntime.dylib.gz
 var onnxRuntimeLib []byte
 
 const onnxRuntimeLibName = "libonnxruntime.dylib"

internal/embedding/lib_linux_amd64.go

@@ -6,10 +6,10 @@ import (
 	_ "embed"
 )
 
-//go:embed assets/lib/linux-amd64/libonnxruntime.so
+//go:embed assets/lib/linux-amd64/libonnxruntime.so.gz
 var onnxRuntimeLib []byte
 
-//go:embed assets/lib/linux-amd64/libonnxruntime_providers_shared.so
+//go:embed assets/lib/linux-amd64/libonnxruntime_providers_shared.so.gz
 var onnxRuntimeProvidersLib []byte
 
 const onnxRuntimeLibName = "libonnxruntime.so"

internal/embedding/lib_linux_arm64.go

@@ -6,10 +6,10 @@ import (
 	_ "embed"
 )
 
-//go:embed assets/lib/linux-arm64/libonnxruntime.so
+//go:embed assets/lib/linux-arm64/libonnxruntime.so.gz
 var onnxRuntimeLib []byte
 
-//go:embed assets/lib/linux-arm64/libonnxruntime_providers_shared.so
+//go:embed assets/lib/linux-arm64/libonnxruntime_providers_shared.so.gz
 var onnxRuntimeProvidersLib []byte
 
 const onnxRuntimeLibName = "libonnxruntime.so"

internal/embedding/lib_windows_amd64.go

@@ -6,7 +6,7 @@ import (
 	_ "embed"
 )
 
-//go:embed assets/lib/windows-amd64/onnxruntime.dll
+//go:embed assets/lib/windows-amd64/onnxruntime.dll.gz
 var onnxRuntimeLib []byte
 
 const onnxRuntimeLibName = "onnxruntime.dll"

internal/embedding/service.go

@@ -3,10 +3,12 @@ package embedding
 
 import (
 	"bytes"
+	"compress/gzip"
 	"context"
 	"crypto/sha256"
 	"encoding/hex"
 	"fmt"
+	"io"
 	"os"
 	"path/filepath"
 	"strings"
@@ -134,7 +136,7 @@ func (m *bgeModel) Dimensions() int {
 	return EmbeddingDim
 }
 
-// extractONNXLibrary extracts the embedded ONNX runtime library to a temp directory.
+// extractONNXLibrary gunzips the embedded ONNX runtime library to a temp directory.
 // Uses content hash to avoid re-extracting if already present.
 func extractONNXLibrary() (string, error) {
 	// Create a hash of the library content for cache key
@@ -157,16 +159,14 @@ func extractONNXLibrary() (string, error) {
 	}
 
 	// Write main library
-	// #nosec G306 -- Shared library needs executable permission (0755) for dynamic linker
-	if err := os.WriteFile(libPath, onnxRuntimeLib, 0755); err != nil {
+	if err := writeGunzipped(libPath, onnxRuntimeLib); err != nil {
 		return "", fmt.Errorf("write library: %w", err)
 	}
 
 	// Write providers library if present (Linux only)
 	if len(onnxRuntimeProvidersLib) > 0 && onnxRuntimeProvidersLibName != "" {
 		providersPath := filepath.Join(cacheDir, onnxRuntimeProvidersLibName)
-		// #nosec G306 -- Shared library needs executable permission (0755) for dynamic linker
-		if err := os.WriteFile(providersPath, onnxRuntimeProvidersLib, 0755); err != nil {
+		if err := writeGunzipped(providersPath, onnxRuntimeProvidersLib); err != nil {
 			return "", fmt.Errorf("write providers library: %w", err)
 		}
 	}
@@ -626,3 +626,23 @@ func (s *Service) EmbedBatchWithContext(ctx context.Context, texts []string) ([]
 func (s *Service) Close() error {
 	return s.model.Close()
 }
+
+// writeGunzipped decompresses gz into path with the executable bit the dynamic linker needs.
+func writeGunzipped(path string, gz []byte) error {
+	zr, err := gzip.NewReader(bytes.NewReader(gz))
+	if err != nil {
+		return err
+	}
+	defer func() { _ = zr.Close() }()
+
+	// #nosec G302 G304 -- shared library needs 0755; path is under our own cache dir
+	f, err := os.OpenFile(path, os.O_CREATE|os.O_WRONLY|os.O_TRUNC, 0755)
+	if err != nil {
+		return err
+	}
+	if _, err := io.Copy(f, zr); err != nil {
+		_ = f.Close()
+		return err
+	}
+	return f.Close()
+}

scripts/download-onnx-libs.sh

@@ -51,9 +51,20 @@ get_installed_version() {
     fi
 }
 
+# Gzip the libraries: Go embeds the .gz and gunzips once at first use, which
+# shrinks the worker binary by roughly 25 MB.
+compress_libs() {
+    local plat="$1"
+    local f
+    for f in "${ASSETS_DIR}/${plat}"/*.dylib "${ASSETS_DIR}/${plat}"/*.so "${ASSETS_DIR}/${plat}"/*.dll; do
+        [ -f "$f" ] && gzip -9 -n -f "$f"
+    done
+}
+
 # Write version file after successful download
 write_version_file() {
     local plat="$1"
+    compress_libs "$plat"
     echo "${ONNX_VERSION}" > "${ASSETS_DIR}/${plat}/.version"
 }
 
@@ -110,9 +121,9 @@ download_windows_amd64() {
 lib_exists() {
     local plat="$1"
     case "$plat" in
-        darwin-*) [ -f "${ASSETS_DIR}/${plat}/libonnxruntime.dylib" ] ;;
-        linux-*) [ -f "${ASSETS_DIR}/${plat}/libonnxruntime.so" ] ;;
-        windows-*) [ -f "${ASSETS_DIR}/${plat}/onnxruntime.dll" ] ;;
+        darwin-*) [ -f "${ASSETS_DIR}/${plat}/libonnxruntime.dylib.gz" ] ;;
+        linux-*) [ -f "${ASSETS_DIR}/${plat}/libonnxruntime.so.gz" ] ;;
+        windows-*) [ -f "${ASSETS_DIR}/${plat}/onnxruntime.dll.gz" ] ;;
         *) return 1 ;;
     esac
 }