refactor: replace Git LFS with runtime model download from Hugging Face

Remove ~170MB of model files from the repository (LFS + committed).
Models are now downloaded at runtime from Hugging Face on first use
and cached to the OS cache directory with progress reporting and retries.

- Add internal/models/download.go: runtime downloader with retry, progress bar, checksums
- Remove go:embed for ONNX models (keep tokenizers embedded)
- Use file-based ONNX session loading instead of byte-slice
- Add scripts/download-models.sh for dev/CI model setup
- Update Makefile with setup-models target
- Update workflow-prepare.sh to download models in CI
- Set lfs: false in all CI workflows
- SHA256: bge=828e14..., cross-encoder=5d3e70...

.gitattributes

@@ -1 +1,2 @@
-internal/embedding/assets/*.onnx filter=lfs diff=lfs merge=lfs -text
+# Models are now downloaded at runtime from GitHub Releases.
+# LFS tracking removed. See internal/models/download.go

.github/workflows/autoupdate.yaml

@@ -17,5 +17,5 @@ jobs:
     with:
       go-version: ">=1.24"
       release-workflow: "release.yaml"
-      lfs: true
+      lfs: false
     secrets: inherit

.github/workflows/ci.yaml

@@ -20,6 +20,6 @@ jobs:
     uses: lukaszraczylo/shared-actions/.github/workflows/go-pr.yaml@main
     with:
       go-version: ">=1.24"
-      lfs: true
+      lfs: false
       build-tags: "fts5"
     secrets: inherit

.github/workflows/release.yaml

@@ -25,7 +25,7 @@ jobs:
             node-cache-dependency-path: "ui/package-lock.json"
             node-output-path: "ui/dist"
             node-embed-path: "internal/worker/static"
-            lfs: true
+            lfs: false
         secrets: inherit
 
     commit-marketplace:

.gitignore

@@ -81,6 +81,12 @@ logs/
 *.db-wal
 *.db-shm
 
+# Model files (downloaded at runtime from GitHub Releases)
+# Test model data (downloaded from Hugging Face)
+testdata/models/
+internal/embedding/assets/model.onnx
+internal/reranking/assets/model.onnx
+
 # goreleaser
 dist/
 docs/dist

Makefile

@@ -23,6 +23,10 @@ all: build
 setup-libs:
 	@./scripts/download-onnx-libs.sh all
 
+# Download ONNX models from Hugging Face (for local dev, skips if present)
+setup-models:
+	@./scripts/download-models.sh
+
 # Update version in committed plugin metadata
 update-version:
 	@if [ -f .claude-plugin/plugin.json ]; then \
@@ -182,12 +186,14 @@ uninstall: stop-worker
 	@echo "Uninstallation complete!"
 
 # Run tests (with FTS5 support)
-test: setup-libs
-	go test $(BUILD_TAGS) -v -race ./...
+# CLAUDE_MNEMONIC_MODEL_DIR points to a common model directory for tests.
+# Run setup-models first to download models from Hugging Face.
+test: setup-libs setup-models
+	CLAUDE_MNEMONIC_MODEL_DIR=$$(pwd)/testdata/models go test $(BUILD_TAGS) -v -race ./...
 
 # Run tests with coverage (with FTS5 support)
-test-coverage: setup-libs
-	go test $(BUILD_TAGS) -v -race -coverprofile=coverage.out ./...
+test-coverage: setup-libs setup-models
+	CLAUDE_MNEMONIC_MODEL_DIR=$$(pwd)/testdata/models go test $(BUILD_TAGS) -v -race -coverprofile=coverage.out ./...
 	go tool cover -html=coverage.out -o coverage.html
 	@go tool cover -func=coverage.out | tail -1
 

internal/embedding/assets.go

@@ -1,14 +1,12 @@
-// Package embedding provides text embedding generation using all-MiniLM-L6-v2.
+// Package embedding provides text embedding generation using bge-small-en-v1.5.
 package embedding
 
 import (
 	_ "embed"
 )
 
-// Model and tokenizer files - embedded for all platforms
+// Tokenizer file - embedded for all platforms (small, not in LFS).
+// The ONNX model is downloaded at runtime to ~/.claude-mnemonic/models/.
 //
-//go:embed assets/model.onnx
-var modelData []byte
-
 //go:embed assets/tokenizer.json
 var tokenizerData []byte

internal/embedding/service.go

@@ -12,6 +12,7 @@ import (
 	"strings"
 	"sync"
 
+	"github.com/lukaszraczylo/claude-mnemonic/internal/models"
 	"github.com/sugarme/tokenizer"
 	"github.com/sugarme/tokenizer/pretrained"
 	ort "github.com/yalue/onnxruntime_go"
@@ -29,6 +30,13 @@ const (
 	BGEModelName = "bge-small-en-v1.5"
 	// DefaultModelVersion is the default model to use
 	DefaultModelVersion = BGEModelVersion
+
+	// BGEAssetName is the local filename for the BGE model.
+	BGEAssetName = "bge-small-en-v1.5-model.onnx"
+	// BGEAssetURL is the Hugging Face URL for the BGE ONNX model.
+	BGEAssetURL = "https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/onnx/model.onnx"
+	// BGEModelSHA256 is the expected SHA-256 of the BGE model file.
+	BGEModelSHA256 = "828e1496d7fabb79cfa4dcd84fa38625c0d3d21da474a00f08db0f559940cf35"
 )
 
 // MaxSequenceLength is the maximum token sequence length for the model.
@@ -60,6 +68,7 @@ var _ EmbeddingModel = (*bgeModel)(nil)
 var _ ONNXConfigurer = (*bgeModel)(nil)
 
 // newBGEModel creates a new BGE embedding model using bundled ONNX runtime and model.
+// The ONNX model is downloaded at runtime from GitHub Releases and cached to disk.
 func newBGEModel() (EmbeddingModel, error) {
 	// Extract ONNX runtime library to temp directory
 	libDir, err := extractONNXLibrary()
@@ -84,9 +93,15 @@ func newBGEModel() (EmbeddingModel, error) {
 		return nil, fmt.Errorf("load tokenizer: %w", err)
 	}
 
-	// Create ONNX session using model-specific configuration
+	// Download model on first use (cached to ~/.claude-mnemonic/models/)
+	modelPath, err := models.EnsureModel(BGEAssetName, BGEAssetURL, BGEModelSHA256)
+	if err != nil {
+		return nil, fmt.Errorf("ensure embedding model: %w", err)
+	}
+
+	// Create ONNX session from file path using model-specific configuration
 	config := bgeONNXConfig
-	session, err := ort.NewDynamicAdvancedSessionWithONNXData(modelData, config.InputNames, config.OutputNames, nil)
+	session, err := ort.NewDynamicAdvancedSession(modelPath, config.InputNames, config.OutputNames, nil)
 	if err != nil {
 		return nil, fmt.Errorf("create ONNX session: %w", err)
 	}

internal/models/download.go

@@ -0,0 +1,259 @@
+// Package models provides runtime model download and caching.
+// Models are downloaded from Hugging Face on first use and cached
+// to the OS cache directory (~/Library/Caches/claude-mnemonic/models/ on macOS,
+// ~/.cache/claude-mnemonic/models/ on Linux, %LocalAppData%/claude-mnemonic/models/ on Windows).
+// This replaces the previous approach of embedding models in the binary
+// via go:embed and tracking them with Git LFS.
+package models
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"errors"
+	"fmt"
+	"io"
+	"math"
+	"net/http"
+	"os"
+	"path/filepath"
+	"time"
+
+	"github.com/rs/zerolog/log"
+)
+
+const (
+	// ModelVersion is incremented to force re-download of all cached models.
+	ModelVersion = "1"
+
+	maxRetries      = 3
+	retryBaseDelay  = 2 * time.Second
+	progressEvery   = 3 * time.Second
+	downloadTimeout = 15 * time.Minute
+)
+
+// ModelsDir returns the path to the model cache directory.
+// Uses the OS cache dir for cross-platform compatibility:
+//
+//	macOS:   ~/Library/Caches/claude-mnemonic/models
+//	Linux:   ~/.cache/claude-mnemonic/models
+//	Windows: %LocalAppData%/claude-mnemonic/models
+func ModelsDir() string {
+	dir, err := os.UserCacheDir()
+	if err != nil {
+		home, _ := os.UserHomeDir()
+		dir = filepath.Join(home, ".cache")
+	}
+	return filepath.Join(dir, "claude-mnemonic", "models")
+}
+
+// EnsureModel ensures the model file exists and returns its path.
+//
+// Resolution order:
+//  1. CLAUDE_MNEMONIC_MODEL_DIR env var — if set, load models from this directory directly.
+//     This is intended for development and CI where models are already on disk.
+//  2. User cache directory — if the model is already cached and its checksum matches,
+//     return it immediately.
+//  3. url — download the model from the given URL with retries and cache it.
+//
+// assetName is the local filename (e.g. "bge-small-en-v1.5-model.onnx").
+// expectedSHA256 is the hex-encoded SHA-256 of the file.
+func EnsureModel(assetName, url, expectedSHA256 string) (string, error) {
+	if localDir := os.Getenv("CLAUDE_MNEMONIC_MODEL_DIR"); localDir != "" {
+		localPath := filepath.Join(localDir, assetName)
+		if _, err := os.Stat(localPath); err == nil {
+			if valid, err := verifyChecksum(localPath, expectedSHA256); err != nil || !valid {
+				return "", fmt.Errorf("local model %s checksum mismatch", assetName)
+			}
+			return localPath, nil
+		}
+		return "", fmt.Errorf("local model %s not found in %s", assetName, localDir)
+	}
+
+	dir := ModelsDir()
+	versionFile := filepath.Join(dir, ".model_version")
+	modelPath := filepath.Join(dir, assetName)
+
+	if err := os.MkdirAll(dir, 0700); err != nil {
+		return "", fmt.Errorf("create model cache dir: %w", err)
+	}
+
+	currentVersion, _ := os.ReadFile(versionFile)
+
+	if string(currentVersion) == ModelVersion {
+		if fi, err := os.Stat(modelPath); err == nil && fi.Size() > 0 {
+			if valid, _ := verifyChecksum(modelPath, expectedSHA256); valid {
+				return modelPath, nil
+			}
+			log.Warn().Str("model", assetName).Msg("Cached model corrupted, re-downloading")
+			os.Remove(modelPath)
+		}
+	} else if len(currentVersion) > 0 {
+		log.Info().Str("old", string(currentVersion)).Str("new", ModelVersion).Msg("Model version updated, clearing cache")
+		os.RemoveAll(dir)
+		if err := os.MkdirAll(dir, 0700); err != nil {
+			return "", fmt.Errorf("create model cache dir: %w", err)
+		}
+	}
+
+	if err := downloadWithRetries(url, modelPath); err != nil {
+		os.Remove(modelPath)
+		return "", fmt.Errorf("download %s: %w", assetName, err)
+	}
+
+	if valid, err := verifyChecksum(modelPath, expectedSHA256); err != nil || !valid {
+		os.Remove(modelPath)
+		return "", fmt.Errorf("downloaded model %s failed checksum verification — the file at %s may have changed", assetName, url)
+	}
+
+	if err := os.WriteFile(versionFile, []byte(ModelVersion), 0600); err != nil {
+		log.Warn().Err(err).Msg("Failed to write model version file")
+	}
+
+	log.Info().Str("model", assetName).Str("cache", dir).Msg("Model ready")
+	return modelPath, nil
+}
+
+func downloadWithRetries(url, dest string) error {
+	var lastErr error
+	for attempt := 0; attempt < maxRetries; attempt++ {
+		if attempt > 0 {
+			delay := retryBaseDelay * time.Duration(1<<(attempt-1))
+			log.Warn().Int("attempt", attempt+1).Int("max", maxRetries).Dur("delay", delay).Msg("Retrying download")
+			time.Sleep(delay)
+		}
+		if err := downloadFile(url, dest); err != nil {
+			lastErr = err
+			continue
+		}
+		return nil
+	}
+	return fmt.Errorf("failed after %d attempts: %w", maxRetries, lastErr)
+}
+
+func downloadFile(url, dest string) error {
+	client := &http.Client{Timeout: downloadTimeout}
+
+	resp, err := client.Get(url)
+	if err != nil {
+		return fmt.Errorf("HTTP GET: %w", err)
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode < 200 || resp.StatusCode >= 400 {
+		return fmt.Errorf("HTTP %d", resp.StatusCode)
+	}
+
+	tmp := dest + ".tmp"
+	f, err := os.Create(tmp)
+	if err != nil {
+		return fmt.Errorf("create temp file: %w", err)
+	}
+	cleanup := func() { f.Close(); os.Remove(tmp) }
+
+	name := filepath.Base(dest)
+	reader := &progressReader{
+		reader: resp.Body,
+		total:  resp.ContentLength,
+		name:   name,
+	}
+
+	if _, err := io.Copy(f, reader); err != nil {
+		cleanup()
+		return fmt.Errorf("write file: %w", err)
+	}
+
+	if err := f.Sync(); err != nil {
+		cleanup()
+		return fmt.Errorf("sync file: %w", err)
+	}
+	f.Close()
+
+	if err := os.Rename(tmp, dest); err != nil {
+		os.Remove(tmp)
+		return fmt.Errorf("commit file: %w", err)
+	}
+
+	return nil
+}
+
+// progressReader wraps an io.Reader and logs progress for large downloads.
+type progressReader struct {
+	reader    io.Reader
+	total     int64
+	current   int64
+	name      string
+	lastLog   time.Time
+	loggedPct int
+}
+
+func (pr *progressReader) Read(p []byte) (n int, err error) {
+	n, err = pr.reader.Read(p)
+	pr.current += int64(n)
+
+	if time.Since(pr.lastLog) >= progressEvery || err == io.EOF {
+		if pr.total > 0 {
+			pct := int(math.Round(float64(pr.current) / float64(pr.total) * 100))
+			if pct != pr.loggedPct {
+				pr.loggedPct = pct
+				mibCurrent := float64(pr.current) / (1024 * 1024)
+				mibTotal := float64(pr.total) / (1024 * 1024)
+				log.Info().
+					Str("model", pr.name).
+					Int("pct", pct).
+					Str("progress", fmt.Sprintf("%.0f/%.0f MiB", mibCurrent, mibTotal)).
+					Msg("Downloading")
+			}
+		} else {
+			mibCurrent := float64(pr.current) / (1024 * 1024)
+			log.Info().
+				Str("model", pr.name).
+				Str("progress", fmt.Sprintf("%.0f MiB", mibCurrent)).
+				Msg("Downloading (size unknown)")
+		}
+		pr.lastLog = time.Now()
+	}
+
+	return n, err
+}
+
+func verifyChecksum(path, expectedHex string) (bool, error) {
+	f, err := os.Open(path)
+	if err != nil {
+		return false, err
+	}
+	defer f.Close()
+
+	h := sha256.New()
+	if _, err := io.Copy(h, f); err != nil {
+		return false, err
+	}
+
+	actual := hex.EncodeToString(h.Sum(nil))
+	return actual == expectedHex, nil
+}
+
+// CleanStale removes any temporary download files that may have been left behind
+// by a previous interrupted download.
+func CleanStale() error {
+	dir := ModelsDir()
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		if errors.Is(err, os.ErrNotExist) {
+			return nil
+		}
+		return err
+	}
+
+	for _, e := range entries {
+		if filepath.Ext(e.Name()) == ".tmp" {
+			tmpPath := filepath.Join(dir, e.Name())
+			if info, err := e.Info(); err == nil {
+				if time.Since(info.ModTime()) > time.Hour {
+					os.Remove(tmpPath)
+					log.Debug().Str("file", tmpPath).Msg("Cleaned stale temp file")
+				}
+			}
+		}
+	}
+	return nil
+}

internal/reranking/assets.go

@@ -5,10 +5,8 @@ import (
 	_ "embed"
 )
 
-// Cross-encoder model and tokenizer files - embedded for all platforms
+// Tokenizer file - embedded for all platforms (small, not in LFS).
+// The ONNX model is downloaded at runtime to ~/.claude-mnemonic/models/.
 //
-//go:embed assets/model.onnx
-var crossEncoderModelData []byte
-
 //go:embed assets/tokenizer.json
 var crossEncoderTokenizerData []byte

internal/reranking/service.go

@@ -9,6 +9,7 @@ import (
 	"sort"
 	"sync"
 
+	"github.com/lukaszraczylo/claude-mnemonic/internal/models"
 	"github.com/rs/zerolog/log"
 	"github.com/sugarme/tokenizer"
 	"github.com/sugarme/tokenizer/pretrained"
@@ -26,6 +27,13 @@ const (
 	DefaultCandidateLimit = 100
 	// DefaultResultLimit is the default number of results to return after reranking
 	DefaultResultLimit = 10
+
+	// CrossEncoderAssetName is the local filename for the cross-encoder model.
+	CrossEncoderAssetName = "ms-marco-MiniLM-L6-v2-model.onnx"
+	// CrossEncoderAssetURL is the Hugging Face URL for the cross-encoder ONNX model.
+	CrossEncoderAssetURL = "https://huggingface.co/cross-encoder/ms-marco-MiniLM-L6-v2/resolve/main/onnx/model.onnx"
+	// CrossEncoderModelSHA256 is the expected SHA-256 of the cross-encoder model file.
+	CrossEncoderModelSHA256 = "5d3e70fd0c9ff14b9b5169a51e957b7a9c74897afd0a35ce4bd318150c1d4d4a"
 )
 
 // Candidate represents a search result candidate for reranking.
@@ -91,12 +99,18 @@ func NewService(cfg Config) (*Service, error) {
 		Stride:    0,
 	})
 
+	// Download model on first use (cached to ~/.claude-mnemonic/models/)
+	modelPath, err := models.EnsureModel(CrossEncoderAssetName, CrossEncoderAssetURL, CrossEncoderModelSHA256)
+	if err != nil {
+		return nil, fmt.Errorf("ensure cross-encoder model: %w", err)
+	}
+
 	// Cross-encoder outputs a single logit for relevance scoring
 	inputNames := []string{"input_ids", "attention_mask", "token_type_ids"}
 	outputNames := []string{"logits"}
 
-	session, err := ort.NewDynamicAdvancedSessionWithONNXData(
-		crossEncoderModelData,
+	session, err := ort.NewDynamicAdvancedSession(
+		modelPath,
 		inputNames,
 		outputNames,
 		nil,

scripts/download-models.sh

@@ -0,0 +1,65 @@
+#!/bin/bash
+# Download ONNX models from Hugging Face for local development and CI.
+# Usage: ./scripts/download-models.sh [--force]
+#
+# Downloads models to internal/*/assets/ for go:embed and to testdata/models/
+# for Go tests (CLAUDE_MNEMONIC_MODEL_DIR points there).
+
+set -e
+
+ASSETS_EMB="internal/embedding/assets"
+ASSETS_RERANK="internal/reranking/assets"
+TESTDATA="testdata/models"
+FORCE_DOWNLOAD=false
+
+for arg in "$@"; do
+	if [ "$arg" = "--force" ]; then
+		FORCE_DOWNLOAD=true
+	fi
+done
+
+download_if_needed() {
+	local url="$1"
+	local dest="$2"
+	local name="$3"
+	local expected_sha="$4"
+
+	if [ "$FORCE_DOWNLOAD" = false ] && [ -f "$dest" ]; then
+		local actual_sha
+		actual_sha=$(shasum -a 256 "$dest" | awk '{print $1}')
+		if [ "$actual_sha" = "$expected_sha" ]; then
+			echo "[skip] $name"
+			return
+		fi
+		echo "[mismatch] $name checksum mismatch, re-downloading"
+	fi
+
+	echo "[download] $name ($(basename "$url"))"
+	curl -fsSL --retry 3 --retry-delay 2 "$url" -o "$dest"
+	echo "[ok] $name"
+}
+
+echo "=== Downloading models from Hugging Face ==="
+
+# BGE-small-en-v1.5 embedding model (~127 MB)
+download_if_needed \
+	"https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/onnx/model.onnx" \
+	"${ASSETS_EMB}/model.onnx" \
+	"embedding (bge-small-en-v1.5)" \
+	"828e1496d7fabb79cfa4dcd84fa38625c0d3d21da474a00f08db0f559940cf35"
+
+# MS-MARCO MiniLM-L6-v2 cross-encoder model (~91 MB)
+download_if_needed \
+	"https://huggingface.co/cross-encoder/ms-marco-MiniLM-L6-v2/resolve/main/onnx/model.onnx" \
+	"${ASSETS_RERANK}/model.onnx" \
+	"cross-encoder (ms-marco-MiniLM-L6-v2)" \
+	"5d3e70fd0c9ff14b9b5169a51e957b7a9c74897afd0a35ce4bd318150c1d4d4a"
+
+echo ""
+echo "=== Staging models for tests ==="
+mkdir -p "${TESTDATA}"
+cp "${ASSETS_EMB}/model.onnx" "${TESTDATA}/bge-small-en-v1.5-model.onnx"
+cp "${ASSETS_RERANK}/model.onnx" "${TESTDATA}/ms-marco-MiniLM-L6-v2-model.onnx"
+echo "[ok] Test models staged to ${TESTDATA}/"
+
+echo "Done!"

workflow-prepare.sh

@@ -10,46 +10,65 @@ HOST_OS=$(uname -s | tr '[:upper:]' '[:lower:]')
 # Determine target platform for ONNX library download
 # Use TARGET_GOOS/TARGET_GOARCH from CI matrix if available, otherwise auto-detect
 if [ -n "$TARGET_GOOS" ] && [ -n "$TARGET_GOARCH" ]; then
-    ONNX_PLATFORM="${TARGET_GOOS}-${TARGET_GOARCH}"
-    echo "Target platform from CI matrix: $ONNX_PLATFORM"
+	ONNX_PLATFORM="${TARGET_GOOS}-${TARGET_GOARCH}"
+	echo "Target platform from CI matrix: $ONNX_PLATFORM"
 else
-    ONNX_PLATFORM="auto"
+	ONNX_PLATFORM="auto"
 fi
 
 # On Windows, install SQLite development headers for CGO
 if [[ "$HOST_OS" == mingw* ]] || [[ "$HOST_OS" == msys* ]] || [[ "$HOST_OS" == cygwin* ]]; then
-    echo "Installing SQLite for Windows..."
+	echo "Installing SQLite for Windows..."
 
-    # Download SQLite amalgamation and set up for CGO
-    SQLITE_VERSION="3470200"
-    SQLITE_YEAR="2024"
-    SQLITE_DIR="/c/sqlite"
-    SQLITE_URL="https://www.sqlite.org/${SQLITE_YEAR}/sqlite-amalgamation-${SQLITE_VERSION}.zip"
+	# Download SQLite amalgamation and set up for CGO
+	SQLITE_VERSION="3470200"
+	SQLITE_YEAR="2024"
+	SQLITE_DIR="/c/sqlite"
+	SQLITE_URL="https://www.sqlite.org/${SQLITE_YEAR}/sqlite-amalgamation-${SQLITE_VERSION}.zip"
 
-    mkdir -p "$SQLITE_DIR"
-    curl -sSL "$SQLITE_URL" -o /tmp/sqlite.zip
-    unzip -q /tmp/sqlite.zip -d /tmp/
-    cp /tmp/sqlite-amalgamation-${SQLITE_VERSION}/* "$SQLITE_DIR/"
-    rm -rf /tmp/sqlite.zip /tmp/sqlite-amalgamation-${SQLITE_VERSION}
+	mkdir -p "$SQLITE_DIR"
+	curl -sSL "$SQLITE_URL" -o /tmp/sqlite.zip
+	unzip -q /tmp/sqlite.zip -d /tmp/
+	cp /tmp/sqlite-amalgamation-${SQLITE_VERSION}/* "$SQLITE_DIR/"
+	rm -rf /tmp/sqlite.zip /tmp/sqlite-amalgamation-${SQLITE_VERSION}
 
-    # Download Go modules first so we can patch sqlite-vec
-    echo "Downloading Go modules..."
-    go mod download
+	# Download Go modules first so we can patch sqlite-vec
+	echo "Downloading Go modules..."
+	go mod download
 
-    # Find the sqlite-vec module and copy sqlite3.h there
-    SQLITE_VEC_PATH=$(go list -m -f '{{.Dir}}' github.com/asg017/sqlite-vec-go-bindings 2>/dev/null || true)
-    if [ -n "$SQLITE_VEC_PATH" ] && [ -d "$SQLITE_VEC_PATH/cgo" ]; then
-        # Make module writable (it's read-only by default)
-        chmod -R u+w "$SQLITE_VEC_PATH"
-        cp "$SQLITE_DIR/sqlite3.h" "$SQLITE_VEC_PATH/cgo/"
-        cp "$SQLITE_DIR/sqlite3.c" "$SQLITE_VEC_PATH/cgo/"
-        echo "SQLite headers copied to $SQLITE_VEC_PATH/cgo/"
-    fi
+	# Find the sqlite-vec module and copy sqlite3.h there
+	SQLITE_VEC_PATH=$(go list -m -f '{{.Dir}}' github.com/asg017/sqlite-vec-go-bindings 2>/dev/null || true)
+	if [ -n "$SQLITE_VEC_PATH" ] && [ -d "$SQLITE_VEC_PATH/cgo" ]; then
+		# Make module writable (it's read-only by default)
+		chmod -R u+w "$SQLITE_VEC_PATH"
+		cp "$SQLITE_DIR/sqlite3.h" "$SQLITE_VEC_PATH/cgo/"
+		cp "$SQLITE_DIR/sqlite3.c" "$SQLITE_VEC_PATH/cgo/"
+		echo "SQLite headers copied to $SQLITE_VEC_PATH/cgo/"
+	fi
 
-    # Tell linker to allow multiple definitions (both go-sqlite3 and sqlite-vec embed SQLite)
-    echo "CGO_LDFLAGS=-Wl,--allow-multiple-definition" >> "$GITHUB_ENV"
-    echo "SQLite setup complete"
+	# Tell linker to allow multiple definitions (both go-sqlite3 and sqlite-vec embed SQLite)
+	echo "CGO_LDFLAGS=-Wl,--allow-multiple-definition" >>"$GITHUB_ENV"
+	echo "SQLite setup complete"
 fi
 
 # Download ONNX runtime libraries for target platform
 ./scripts/download-onnx-libs.sh "$ONNX_PLATFORM"
+
+# Download ONNX models from Hugging Face and stage for tests
+# Non-fatal: if download fails, tests will attempt runtime download themselves
+set +e
+./scripts/download-models.sh
+DOWNLOAD_EXIT=$?
+set -e
+
+if [ $DOWNLOAD_EXIT -eq 0 ]; then
+    # Export model directory so tests can find models without network access
+    if [ -n "$GITHUB_ENV" ]; then
+        echo "CLAUDE_MNEMONIC_MODEL_DIR=$GITHUB_WORKSPACE/testdata/models" >> "$GITHUB_ENV"
+    fi
+else
+    echo "Warning: Model download failed — tests will try to download models at runtime"
+fi
+else
+	echo "Warning: Model download failed — tests will try to download models at runtime"
+fi