fix: address 15 additional hang vectors found during deep audit (#45)

MCP server (5 fixes):
- Move semaphore acquisition inside goroutine so main loop stays
  responsive when all slots are taken
- Add 10s write timeout to sendResponse to prevent pipe deadlock
  when Claude Code pauses reading stdout
- Send fallback JSON-RPC error when json.Marshal fails instead of
  silently swallowing the error and leaving caller waiting forever
- Silence unknown notification methods (req.ID == nil) instead of
  sending unsolicited error responses that may desync the host
- Return MCP isError content for tool failures instead of top-level
  JSON-RPC error, matching the MCP specification

Vector/embedding (3 fixes):
- Move EmbedBatchWithContext call before writeMu.Lock in AddDocuments
  so ONNX inference runs outside the write lock
- Replace singleflight.Do with DoChan + ctx select in both
  getOrComputeEmbedding and UnifiedSearch so callers can bail out
  independently when their context expires
- Add activeQueries atomic counter; skip cache warming when user
  queries are in-flight; reduce warming timeout from 5s to 2s

Hooks (4 fixes):
- Cap EnsureWorkerRunning to 15s hard deadline with context; reduce
  StartupTimeout from 30s to 10s; reduce port-in-use retries
- Fix nil dereference panic in user-prompt hook when initResult is
  nil (non-JSON worker response); use comma-ok assertions
- Use package-level hookClient/healthClient with DisableKeepAlives
  to prevent FD leaks in short-lived hook processes
- Set SysProcAttr{Setpgid: true} to detach worker from hook process
  group, preventing kill-cascade from Claude Code

Worker/DB (3 fixes):
- Replace os.Exit(0) in MCP config watcher with context cancellation
  for clean protocol shutdown
- Add 60s context.WithTimeout around ProcessObservation calls in
  processAllSessions to prevent hung CLI subprocesses from blocking
  the queue processor forever
- Set explicit PRAGMA wal_autocheckpoint=1000 and add PASSIVE WAL
  checkpoint to Optimize() to prevent checkpoint stalls

Adds 20+ regression tests across all fix areas.

internal/db/gorm/store.go

@@ -99,8 +99,9 @@ func NewStore(cfg Config) (*Store, error) {
 		"PRAGMA synchronous=NORMAL",
 		"PRAGMA cache_size=-64000",   // 64MB cache (negative = KB)
 		"PRAGMA temp_store=MEMORY",   // Store temp tables in memory
-		"PRAGMA mmap_size=268435456", // 256MB memory-mapped I/O
-		"PRAGMA page_size=4096",      // 4KB pages (optimal for most systems)
+		"PRAGMA mmap_size=268435456",    // 256MB memory-mapped I/O
+		"PRAGMA page_size=4096",         // 4KB pages (optimal for most systems)
+		"PRAGMA wal_autocheckpoint=1000", // Explicit default; checkpoint every 1000 WAL frames
 	}
 	for _, pragma := range pragmas {
 		if _, err := sqlDB.Exec(pragma); err != nil {
@@ -192,6 +193,11 @@ func (s *Store) Optimize(ctx context.Context) error {
 		log.Warn().Err(err).Msg("PRAGMA optimize failed (non-fatal)")
 	}
 
+	// Passive WAL checkpoint — doesn't block readers/writers
+	if _, err := s.sqlDB.ExecContext(ctx, "PRAGMA wal_checkpoint(PASSIVE)"); err != nil {
+		log.Warn().Err(err).Msg("WAL checkpoint failed (non-fatal)")
+	}
+
 	log.Info().Dur("duration", time.Since(start)).Msg("Database optimization complete")
 	return nil
 }

internal/db/gorm/store_test.go

@@ -4,6 +4,7 @@
 package gorm
 
 import (
+	"context"
 	"os"
 	"path/filepath"
 	"testing"
@@ -150,3 +151,91 @@ func TestMigrationIdempotency(t *testing.T) {
 
 	t.Logf("✅ Migrations are idempotent")
 }
+
+func TestWALAutocheckpoint(t *testing.T) {
+	tmpDir, err := os.MkdirTemp("", "gorm_wal_checkpoint_*")
+	if err != nil {
+		t.Fatalf("create temp dir: %v", err)
+	}
+	defer os.RemoveAll(tmpDir)
+
+	dbPath := filepath.Join(tmpDir, "test.db")
+	store, err := NewStore(Config{
+		Path:     dbPath,
+		MaxConns: 2,
+		LogLevel: logger.Silent,
+	})
+	if err != nil {
+		t.Fatalf("NewStore failed: %v", err)
+	}
+	defer store.Close()
+
+	// Verify wal_autocheckpoint is set to 1000
+	var checkpoint int
+	err = store.GetRawDB().QueryRow("PRAGMA wal_autocheckpoint").Scan(&checkpoint)
+	if err != nil {
+		t.Fatalf("query wal_autocheckpoint: %v", err)
+	}
+	if checkpoint != 1000 {
+		t.Errorf("expected wal_autocheckpoint=1000, got %d", checkpoint)
+	}
+}
+
+func TestOptimize_RunsWALCheckpoint(t *testing.T) {
+	tmpDir, err := os.MkdirTemp("", "gorm_optimize_*")
+	if err != nil {
+		t.Fatalf("create temp dir: %v", err)
+	}
+	defer os.RemoveAll(tmpDir)
+
+	dbPath := filepath.Join(tmpDir, "test.db")
+	store, err := NewStore(Config{
+		Path:     dbPath,
+		MaxConns: 2,
+		LogLevel: logger.Silent,
+	})
+	if err != nil {
+		t.Fatalf("NewStore failed: %v", err)
+	}
+	defer store.Close()
+
+	// Insert some data to generate WAL frames
+	_, err = store.GetRawDB().Exec("INSERT INTO observations (sdk_session_id, title, scope, project, type, created_at, created_at_epoch) VALUES ('test-sess', 'test data', 'project', '/tmp/test', 'decision', '2026-01-01T00:00:00Z', 1735689600)")
+	if err != nil {
+		t.Fatalf("insert test data: %v", err)
+	}
+
+	// Optimize should succeed (includes PASSIVE WAL checkpoint)
+	err = store.Optimize(context.Background())
+	if err != nil {
+		t.Fatalf("Optimize failed: %v", err)
+	}
+}
+
+func TestOptimize_RespectsContextCancellation(t *testing.T) {
+	tmpDir, err := os.MkdirTemp("", "gorm_optimize_cancel_*")
+	if err != nil {
+		t.Fatalf("create temp dir: %v", err)
+	}
+	defer os.RemoveAll(tmpDir)
+
+	dbPath := filepath.Join(tmpDir, "test.db")
+	store, err := NewStore(Config{
+		Path:     dbPath,
+		MaxConns: 2,
+		LogLevel: logger.Silent,
+	})
+	if err != nil {
+		t.Fatalf("NewStore failed: %v", err)
+	}
+	defer store.Close()
+
+	// Already-cancelled context should cause Optimize to fail
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+
+	err = store.Optimize(ctx)
+	if err == nil {
+		t.Error("expected error with cancelled context, got nil")
+	}
+}

internal/mcp/server.go

@@ -171,11 +171,22 @@ func (s *Server) Run(ctx context.Context) error {
 			}
 
 			// Dispatch request to its own goroutine.
+			// Semaphore is acquired inside the goroutine so the main
+			// loop never blocks on a full semaphore (Fix #1).
 			wg.Add(1)
-			sem <- struct{}{} // acquire semaphore slot
 			go func(r Request) {
 				defer wg.Done()
-				defer func() { <-sem }() // release semaphore slot
+				// Acquire semaphore inside goroutine, not blocking main loop
+				select {
+				case sem <- struct{}{}:
+					defer func() { <-sem }()
+				case <-ctx.Done():
+					// Server shutting down — send error so caller isn't left waiting
+					if r.ID != nil {
+						_ = s.sendError(r.ID, -32000, "Server shutting down", nil)
+					}
+					return
+				}
 
 				resp := s.handleRequest(ctx, &r)
 				if resp != nil {
@@ -203,6 +214,10 @@ func (s *Server) handleRequest(ctx context.Context, req *Request) *Response {
 	case "notifications/initialized", "notifications/cancelled":
 		return nil // Notifications don't get responses
 	default:
+		if req.ID == nil {
+			// Notifications must not receive responses per JSON-RPC 2.0
+			return nil
+		}
 		return &Response{
 			JSONRPC: "2.0",
 			ID:      req.ID,
@@ -775,13 +790,15 @@ func (s *Server) handleToolsCall(ctx context.Context, req *Request) *Response {
 
 	result, err := s.callTool(ctx, params.Name, params.Arguments)
 	if err != nil {
+		// MCP spec: tool failures use Result with isError, not top-level Error
 		return &Response{
 			JSONRPC: "2.0",
 			ID:      req.ID,
-			Error: &Error{
-				Code:    -32000,
-				Message: "Tool error",
-				Data:    err.Error(),
+			Result: map[string]any{
+				"content": []map[string]any{
+					{"type": "text", "text": "Error: " + err.Error()},
+				},
+				"isError": true,
 			},
 		}
 	}
@@ -1902,16 +1919,38 @@ func (s *Server) sendResponse(resp *Response) error {
 	data, err := json.Marshal(resp)
 	if err != nil {
 		log.Error().Err(err).Msg("Failed to marshal response")
+		// Send a fallback error so the caller isn't left waiting (Fix #3)
+		fallback, _ := json.Marshal(map[string]any{
+			"jsonrpc": "2.0",
+			"id":      resp.ID,
+			"error":   map[string]any{"code": -32603, "message": "internal marshal error"},
+		})
+		s.writeMu.Lock()
+		_, _ = fmt.Fprintln(s.stdout, string(fallback))
+		s.writeMu.Unlock()
+		return fmt.Errorf("marshal error: %w", err)
+	}
+
+	// Bound the write to prevent pipe deadlock (Fix #2)
+	done := make(chan error, 1)
+	go func() {
+		s.writeMu.Lock()
+		_, werr := fmt.Fprintln(s.stdout, string(data))
+		s.writeMu.Unlock()
+		done <- werr
+	}()
+
+	select {
+	case werr := <-done:
+		if werr != nil {
+			log.Error().Err(werr).Msg("Failed to write response to stdout")
+			return werr
+		}
 		return nil
+	case <-time.After(10 * time.Second):
+		log.Error().Msg("Stdout write timed out — pipe likely full")
+		return fmt.Errorf("write timeout: stdout pipe full")
 	}
-	s.writeMu.Lock()
-	_, err = fmt.Fprintln(s.stdout, string(data))
-	s.writeMu.Unlock()
-	if err != nil {
-		log.Error().Err(err).Msg("Failed to write response to stdout")
-		return err
-	}
-	return nil
 }
 
 // sendError sends a JSON-RPC error response. Returns an error if writing fails.

internal/mcp/server_test.go

@@ -1665,6 +1665,7 @@ func TestToolListContainsExpectedSchemas(t *testing.T) {
 }
 
 // TestHandleToolsCall_UnknownTool tests tools/call with unknown tool name.
+// After Fix #5: tool errors use Result with isError, not top-level Error.
 func TestHandleToolsCall_UnknownTool(t *testing.T) {
 	t.Parallel()
 
@@ -1679,9 +1680,13 @@ func TestHandleToolsCall_UnknownTool(t *testing.T) {
 	}
 
 	resp := server.handleToolsCall(ctx, req)
-	require.NotNil(t, resp.Error)
-	assert.Equal(t, -32000, resp.Error.Code)
-	assert.Contains(t, resp.Error.Data, "unknown tool")
+	assert.Nil(t, resp.Error, "tool errors must not use top-level Error (MCP spec)")
+	require.NotNil(t, resp.Result)
+	result, ok := resp.Result.(map[string]any)
+	require.True(t, ok)
+	assert.Equal(t, true, result["isError"])
+	content := result["content"].([]map[string]any)
+	assert.Contains(t, content[0]["text"], "unknown tool")
 }
 
 // TestCallTool_ToolNameRecognition tests that valid tool names are recognized.
@@ -1982,6 +1987,7 @@ func TestTimelineParamsStruct_Validation(t *testing.T) {
 }
 
 // TestHandleToolsCall_EmptyParams tests tools/call with empty params.
+// After Fix #5: tool errors use Result with isError, not top-level Error.
 func TestHandleToolsCall_EmptyParams(t *testing.T) {
 	t.Parallel()
 
@@ -1997,8 +2003,12 @@ func TestHandleToolsCall_EmptyParams(t *testing.T) {
 
 	resp := server.handleToolsCall(ctx, req)
 
-	// Should error due to missing name
-	require.NotNil(t, resp.Error)
+	// Empty name goes through callTool default branch -> isError
+	assert.Nil(t, resp.Error, "tool errors must use isError in Result")
+	require.NotNil(t, resp.Result)
+	result, ok := resp.Result.(map[string]any)
+	require.True(t, ok)
+	assert.Equal(t, true, result["isError"])
 }
 
 // TestSendResponse_WithError tests sendResponse with an error response.
@@ -2062,6 +2072,7 @@ func TestToolCallParamsWithComplexArgs(t *testing.T) {
 }
 
 // TestHandleToolsCall_UnknownToolNameError tests tools/call with unknown tool returns error.
+// After Fix #5: tool errors use Result with isError, not top-level Error.
 func TestHandleToolsCall_UnknownToolNameError(t *testing.T) {
 	t.Parallel()
 
@@ -2077,12 +2088,13 @@ func TestHandleToolsCall_UnknownToolNameError(t *testing.T) {
 
 	resp := server.handleToolsCall(ctx, req)
 
-	// Should get an error response
 	assert.Equal(t, "2.0", resp.JSONRPC)
 	assert.Equal(t, 1, resp.ID)
-	require.NotNil(t, resp.Error)
-	// Error is "Tool error" with message containing "unknown tool"
-	assert.True(t, resp.Error.Code != 0)
+	assert.Nil(t, resp.Error, "tool errors must use isError in Result, not top-level Error")
+	require.NotNil(t, resp.Result)
+	result, ok := resp.Result.(map[string]any)
+	require.True(t, ok)
+	assert.Equal(t, true, result["isError"])
 }
 
 // =============================================================================
@@ -3172,3 +3184,237 @@ func TestRun_GracefulDrainOnCancel(t *testing.T) {
 		assert.Equal(t, "2.0", resp["jsonrpc"], "any completed response must be valid JSON-RPC 2.0")
 	}
 }
+
+// =============================================================================
+// REGRESSION TESTS — Fix #1-#5
+// =============================================================================
+
+// blockingWriter is an io.Writer that blocks forever on Write.
+type blockingWriter struct {
+	blocked chan struct{} // closed when Write is entered
+}
+
+func (bw *blockingWriter) Write(p []byte) (int, error) {
+	if bw.blocked != nil {
+		close(bw.blocked)
+	}
+	select {} // block forever
+}
+
+// TestRun_SemaphoreDoesNotBlockMainLoop (Fix #1 regression) fills all semaphore
+// slots with blocked requests, then sends a notification. The main loop must
+// stay responsive and not hang on semaphore acquisition.
+func TestRun_SemaphoreDoesNotBlockMainLoop(t *testing.T) {
+	t.Parallel()
+
+	const maxConcurrent = 10
+
+	// Mock worker that blocks until test context is cancelled
+	handlerCtx, handlerCancel := context.WithCancel(context.Background())
+	defer handlerCancel()
+
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		<-handlerCtx.Done() // block until test cleanup
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = io.WriteString(w, `{}`)
+	}))
+	defer func() {
+		handlerCancel()          // unblock all handlers first
+		ts.CloseClientConnections()
+		ts.Close()
+	}()
+
+	stdinR, stdinW := io.Pipe()
+	var stdout bytes.Buffer
+
+	server := &Server{
+		client:    ts.Client(),
+		workerURL: ts.URL,
+		project:   "test",
+		version:   "1.0.0",
+		stdin:     stdinR,
+		stdout:    &stdout,
+	}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	runDone := make(chan error, 1)
+	go func() {
+		runDone <- server.Run(ctx)
+	}()
+
+	// Send maxConcurrent+2 requests to fill all semaphore slots
+	for i := 0; i < maxConcurrent+2; i++ {
+		req := Request{JSONRPC: "2.0", ID: i + 1, Method: "tools/call",
+			Params: json.RawMessage(`{"name":"get_memory_stats","arguments":{}}`)}
+		data, _ := json.Marshal(req)
+		_, err := io.WriteString(stdinW, string(data)+"\n")
+		require.NoError(t, err)
+	}
+
+	// Give goroutines time to start and fill semaphore
+	time.Sleep(100 * time.Millisecond)
+
+	// Now send a notification — this must not hang the main loop
+	notifSent := make(chan struct{})
+	go func() {
+		notif := `{"jsonrpc":"2.0","method":"notifications/initialized"}` + "\n"
+		_, _ = io.WriteString(stdinW, notif)
+		close(notifSent)
+	}()
+
+	select {
+	case <-notifSent:
+		// Main loop accepted the notification write (stdin is a pipe, so
+		// the write completing means the reader consumed it)
+	case <-time.After(3 * time.Second):
+		t.Fatal("main loop blocked — semaphore acquisition is blocking the main goroutine")
+	}
+
+	// Clean up: cancel server context, close stdin
+	cancel()
+	_ = stdinW.Close()
+
+	// Wait for Run to finish
+	select {
+	case <-runDone:
+	case <-time.After(5 * time.Second):
+		// Acceptable — some goroutines may still be draining
+	}
+}
+
+// TestSendResponse_WriteTimeout (Fix #2 regression) verifies that sendResponse
+// returns an error within a bounded time when the writer blocks forever.
+func TestSendResponse_WriteTimeout(t *testing.T) {
+	t.Parallel()
+
+	bw := &blockingWriter{blocked: make(chan struct{})}
+	server := &Server{stdout: bw}
+
+	resp := &Response{
+		JSONRPC: "2.0",
+		ID:      1,
+		Result:  "ok",
+	}
+
+	done := make(chan error, 1)
+	go func() {
+		done <- server.sendResponse(resp)
+	}()
+
+	select {
+	case err := <-done:
+		require.Error(t, err, "sendResponse must return error on write timeout")
+		assert.Contains(t, err.Error(), "write timeout")
+	case <-time.After(15 * time.Second):
+		t.Fatal("sendResponse hung forever — write timeout not working")
+	}
+}
+
+// TestSendResponse_MarshalError (Fix #3 regression) verifies that when
+// json.Marshal fails, sendResponse sends a fallback error response and
+// returns an error (instead of silently returning nil).
+func TestSendResponse_MarshalError(t *testing.T) {
+	t.Parallel()
+
+	var buf bytes.Buffer
+	server := &Server{stdout: &buf}
+
+	// Channels are not JSON-serializable — this will cause json.Marshal to fail
+	resp := &Response{
+		JSONRPC: "2.0",
+		ID:      42,
+		Result:  make(chan int), // unserializable
+	}
+
+	err := server.sendResponse(resp)
+
+	// (a) Must return an error
+	require.Error(t, err, "sendResponse must return error when marshal fails")
+	assert.Contains(t, err.Error(), "marshal error")
+
+	// (b) Must have written a fallback JSON-RPC error to stdout
+	output := buf.String()
+	require.NotEmpty(t, output, "fallback response must be written to stdout")
+
+	var fallback map[string]any
+	require.NoError(t, json.Unmarshal([]byte(strings.TrimSpace(output)), &fallback),
+		"fallback must be valid JSON")
+	assert.Equal(t, "2.0", fallback["jsonrpc"])
+	assert.Equal(t, float64(42), fallback["id"], "fallback must preserve original request ID")
+
+	errObj, ok := fallback["error"].(map[string]any)
+	require.True(t, ok, "fallback must contain error object")
+	assert.Equal(t, float64(-32603), errObj["code"])
+	assert.Equal(t, "internal marshal error", errObj["message"])
+}
+
+// TestHandleRequest_UnknownNotification (Fix #4 regression) verifies that
+// unknown notification methods (ID == nil) get no response, while unknown
+// methods with an ID still get a -32601 error.
+func TestHandleRequest_UnknownNotification(t *testing.T) {
+	t.Parallel()
+
+	server := NewServer(nil, "", "", "1.0.0")
+	ctx := context.Background()
+
+	// Case 1: Unknown notification (no ID) — must return nil
+	notifReq := &Request{
+		JSONRPC: "2.0",
+		ID:      nil,
+		Method:  "notifications/roots/list_changed",
+	}
+	resp := server.handleRequest(ctx, notifReq)
+	assert.Nil(t, resp, "unknown notification must not produce a response")
+
+	// Case 2: Unknown method WITH an ID — must return error response
+	methodReq := &Request{
+		JSONRPC: "2.0",
+		ID:      99,
+		Method:  "some/unknown/method",
+	}
+	resp = server.handleRequest(ctx, methodReq)
+	require.NotNil(t, resp, "unknown method with ID must produce an error response")
+	require.NotNil(t, resp.Error)
+	assert.Equal(t, -32601, resp.Error.Code)
+	assert.Equal(t, "Method not found", resp.Error.Message)
+	assert.Equal(t, 99, resp.ID)
+}
+
+// TestHandleToolsCall_ErrorUsesIsError (Fix #5 regression) verifies that when
+// callTool returns an error, the response uses Result with isError:true instead
+// of top-level Error field (per MCP spec).
+func TestHandleToolsCall_ErrorUsesIsError(t *testing.T) {
+	t.Parallel()
+
+	server := NewServer(nil, "", "", "1.0.0")
+	ctx := context.Background()
+
+	req := &Request{
+		JSONRPC: "2.0",
+		ID:      7,
+		Method:  "tools/call",
+		Params:  json.RawMessage(`{"name":"nonexistent_tool","arguments":{}}`),
+	}
+
+	resp := server.handleToolsCall(ctx, req)
+
+	// (a) Response must NOT have top-level Error
+	assert.Nil(t, resp.Error, "tool errors must not use top-level JSON-RPC Error")
+
+	// (b) Response must have Result with isError: true
+	require.NotNil(t, resp.Result, "tool error response must have Result")
+	result, ok := resp.Result.(map[string]any)
+	require.True(t, ok, "Result must be a map")
+	assert.Equal(t, true, result["isError"], "Result must contain isError: true")
+
+	// (c) Result.content[0].text must contain the error message
+	content, ok := result["content"].([]map[string]any)
+	require.True(t, ok, "Result.content must be []map[string]any")
+	require.Len(t, content, 1)
+	assert.Equal(t, "text", content[0]["type"])
+	errText, ok := content[0]["text"].(string)
+	require.True(t, ok)
+	assert.Contains(t, errText, "unknown tool: nonexistent_tool")
+}

internal/search/manager.go

@@ -46,7 +46,7 @@ const (
 	cacheWarmingInterval     = 20 * time.Second // Run warming cycle every 20 seconds
 	frequencyCleanupInterval = 5 * time.Minute  // Cleanup stale entries every 5 minutes
 	cacheCleanupInterval     = time.Minute      // Cleanup expired cache every minute
-	warmingQueryTimeout      = 5 * time.Second  // Timeout for warming queries
+	warmingQueryTimeout      = 2 * time.Second  // Timeout for warming queries (kept short to minimize mutex hold)
 	warmingBatchSize         = 5                // Warm top 5 queries per cycle
 	minRecencyFactor         = 0.1              // Minimum recency factor for scoring
 
@@ -127,6 +127,7 @@ type Manager struct {
 	queryFrequency   map[string]*queryFrequencyInfo
 	cacheTTL         time.Duration
 	cacheMaxSize     int
+	activeQueries    atomic.Int32 // tracks in-flight user queries to yield embedding mutex
 	resultCacheMu    sync.RWMutex
 	queryFrequencyMu sync.RWMutex
 }
@@ -256,7 +257,14 @@ func (m *Manager) cleanupStaleFrequencyEntries() {
 }
 
 // warmFrequentQueries pre-executes frequently used queries to warm the cache.
+// Skips warming entirely when user queries are in-flight to avoid competing
+// for the embedding model mutex on throttled hardware.
 func (m *Manager) warmFrequentQueries() {
+	if m.activeQueries.Load() > 0 {
+		log.Debug().Msg("Skipping cache warming: user queries in flight")
+		return
+	}
+
 	m.queryFrequencyMu.RLock()
 	// Find top N most frequent queries that aren't recently cached
 	type queryScore struct {
@@ -687,30 +695,40 @@ func (m *Manager) UnifiedSearch(ctx context.Context, params SearchParams) (*Unif
 		params.OrderBy = defaultOrderBy
 	}
 
+	// Track active user queries so cache warming can yield the embedding mutex
+	m.activeQueries.Add(1)
+	defer m.activeQueries.Add(-1)
+
 	// Check cache first
 	cacheKey := m.getCacheKey(params)
 	if cached, ok := m.getFromCache(cacheKey); ok {
 		return cached, nil
 	}
 
-	// Use singleflight to coalesce concurrent identical requests
-	result, err, _ := m.searchGroup.Do(cacheKey, func() (any, error) {
+	// Use singleflight DoChan to coalesce concurrent identical requests.
+	// DoChan + select allows per-caller context cancellation: waiting callers
+	// can bail out when their context expires without blocking on a slow first call.
+	ch := m.searchGroup.DoChan(cacheKey, func() (any, error) {
 		return m.executeSearch(ctx, params)
 	})
 
-	if err != nil {
-		return nil, err
+	select {
+	case res := <-ch:
+		if res.Err != nil {
+			return nil, res.Err
+		}
+		searchResult := res.Val.(*UnifiedSearchResult)
+
+		// Cache the result
+		m.putInCache(cacheKey, searchResult)
+
+		// Track query frequency for cache warming
+		m.trackQueryFrequency(params)
+
+		return searchResult, nil
+	case <-ctx.Done():
+		return nil, ctx.Err()
 	}
-
-	searchResult := result.(*UnifiedSearchResult)
-
-	// Cache the result
-	m.putInCache(cacheKey, searchResult)
-
-	// Track query frequency for cache warming
-	m.trackQueryFrequency(params)
-
-	return searchResult, nil
 }
 
 // executeSearch performs the actual search without caching/coalescing.

internal/search/manager_test.go

@@ -1009,6 +1009,91 @@ func TestSearchParams_FormatValues(t *testing.T) {
 	}
 }
 
+// =============================================================================
+// REGRESSION TESTS FOR Fix #3: Cache warming skipped during active queries
+// =============================================================================
+
+// TestCacheWarming_SkippedDuringActiveQueries verifies that warmFrequentQueries
+// returns immediately without doing work when activeQueries > 0, and actually
+// warms when activeQueries == 0.
+func TestCacheWarming_SkippedDuringActiveQueries(t *testing.T) {
+	m := NewManager(nil, nil, nil, nil)
+	defer m.Close()
+
+	// Seed a frequent query so warming has something to consider
+	params := SearchParams{
+		Query:   "test warming query",
+		Project: "test-project",
+		Limit:   10,
+	}
+	m.trackQueryFrequency(params)
+	// Track it multiple times to boost its score
+	for range 5 {
+		m.trackQueryFrequency(params)
+	}
+
+	// Case 1: activeQueries > 0 — warming should be skipped (fast return)
+	m.activeQueries.Store(1)
+
+	start := time.Now()
+	m.warmFrequentQueries()
+	elapsed := time.Since(start)
+
+	// Warming should return almost immediately when skipped
+	assert.Less(t, elapsed, 50*time.Millisecond,
+		"warmFrequentQueries should return quickly when user queries are in flight")
+
+	// Case 2: activeQueries == 0 — warming should proceed past the guard.
+	// With nil stores, executeSearch would panic, so we verify the guard was
+	// bypassed by checking that the frequency map's lastCached is NOT updated
+	// (warming fails silently on nil stores). The key assertion is that Case 1
+	// returns immediately (the guard works) while Case 2 enters the body.
+	m.activeQueries.Store(0)
+
+	// Clear frequency data to prevent executeSearch from being called
+	m.queryFrequencyMu.Lock()
+	m.queryFrequency = make(map[string]*queryFrequencyInfo)
+	m.queryFrequencyMu.Unlock()
+
+	start2 := time.Now()
+	m.warmFrequentQueries()
+	elapsed2 := time.Since(start2)
+
+	// With empty frequency map, warming enters the function body (past guard)
+	// but finds no candidates — fast return without calling executeSearch.
+	assert.Less(t, elapsed2, 200*time.Millisecond,
+		"warmFrequentQueries should complete quickly with no candidates")
+}
+
+// TestActiveQueries_IncrementDecrement verifies that the activeQueries counter
+// is correctly incremented and decremented around search operations.
+func TestActiveQueries_IncrementDecrement(t *testing.T) {
+	m := NewManager(nil, nil, nil, nil)
+	defer m.Close()
+
+	// Before any search, counter should be 0
+	assert.Equal(t, int32(0), m.activeQueries.Load())
+
+	// Directly test the atomic increment/decrement pattern used in UnifiedSearch
+	// (can't call UnifiedSearch with nil stores without panicking on filterSearch)
+	m.activeQueries.Add(1)
+	assert.Equal(t, int32(1), m.activeQueries.Load(), "should be 1 after increment")
+
+	m.activeQueries.Add(1)
+	assert.Equal(t, int32(2), m.activeQueries.Load(), "should be 2 after second increment")
+
+	m.activeQueries.Add(-1)
+	assert.Equal(t, int32(1), m.activeQueries.Load(), "should be 1 after decrement")
+
+	m.activeQueries.Add(-1)
+	assert.Equal(t, int32(0), m.activeQueries.Load(), "should be 0 after final decrement")
+
+	// Verify the field is accessible from warmFrequentQueries guard
+	m.activeQueries.Store(3)
+	assert.Equal(t, int32(3), m.activeQueries.Load())
+	m.activeQueries.Store(0)
+}
+
 // TestUnifiedSearchResult_MultipleResults tests result with multiple items.
 func TestUnifiedSearchResult_MultipleResults(t *testing.T) {
 	results := []SearchResult{

internal/vector/sqlitevec/client.go

@@ -145,20 +145,24 @@ func (c *Client) AddDocuments(ctx context.Context, docs []Document) error {
 		return nil
 	}
 
-	c.writeMu.Lock()
-	defer c.writeMu.Unlock()
-
-	// Generate embeddings for all documents
+	// Prepare texts for embedding
 	texts := make([]string, len(docs))
 	for i, doc := range docs {
 		texts[i] = doc.Content
 	}
 
-	embeddings, err := c.embedSvc.EmbedBatch(texts)
+	// Compute embeddings OUTSIDE the write lock for better concurrency.
+	// Embedding is ONNX inference (slow, mutex-protected internally) — holding
+	// writeMu during inference blocks all concurrent writes and reads.
+	embeddings, err := c.embedSvc.EmbedBatchWithContext(ctx, texts)
 	if err != nil {
 		return fmt.Errorf("generate embeddings: %w", err)
 	}
 
+	// Acquire write lock for DB operations only
+	c.writeMu.Lock()
+	defer c.writeMu.Unlock()
+
 	// Insert into vectors table with model version tracking
 	const insertQuery = `
 		INSERT OR REPLACE INTO vectors (doc_id, embedding, sqlite_id, doc_type, field_type, project, scope, model_version)
@@ -906,8 +910,10 @@ func (c *Client) getOrComputeEmbedding(ctx context.Context, query string) ([]flo
 	}
 	c.queryCacheMu.RUnlock()
 
-	// Cache miss - use singleflight to deduplicate concurrent embedding requests
-	result, err, _ := c.embeddingGroup.Do(query, func() (any, error) {
+	// Cache miss - use singleflight DoChan to deduplicate concurrent embedding requests.
+	// DoChan + select allows per-caller context cancellation: if a caller's context
+	// expires it can bail out without waiting for the shared computation to finish.
+	ch := c.embeddingGroup.DoChan(query, func() (any, error) {
 		// Double-check cache inside singleflight (another goroutine may have just cached it)
 		c.queryCacheMu.RLock()
 		if entry, ok := c.queryCache[query]; ok {
@@ -921,8 +927,10 @@ func (c *Client) getOrComputeEmbedding(ctx context.Context, query string) ([]flo
 		// Record cache miss
 		c.stats.embeddingMisses.Add(1)
 
-		// Compute embedding with context-aware lock acquisition
-		emb, err := c.embedSvc.EmbedWithContext(ctx, query)
+		// Compute embedding — use non-context Embed here because the singleflight
+		// result is shared across callers with different contexts. Per-caller
+		// cancellation is handled by the select below.
+		emb, err := c.embedSvc.Embed(query)
 		if err != nil {
 			return nil, err
 		}
@@ -969,10 +977,15 @@ func (c *Client) getOrComputeEmbedding(ctx context.Context, query string) ([]flo
 		return emb, nil
 	})
 
-	if err != nil {
-		return nil, err
+	select {
+	case res := <-ch:
+		if res.Err != nil {
+			return nil, res.Err
+		}
+		return res.Val.([]float32), nil
+	case <-ctx.Done():
+		return nil, ctx.Err()
 	}
-	return result.([]float32), nil
 }
 
 // ClearCache clears the embedding cache.

internal/vector/sqlitevec/client_test.go

@@ -3,6 +3,7 @@ package sqlitevec
 import (
 	"context"
 	"database/sql"
+	"fmt"
 	"os"
 	"path/filepath"
 	"sync"
@@ -2013,3 +2014,141 @@ func TestAcquireRLockWithContext_CleanupOnCancel(t *testing.T) {
 		t.Fatal("write lock acquisition timed out: cleanup goroutine may have leaked an RLock")
 	}
 }
+
+// =============================================================================
+// REGRESSION TESTS FOR Fix #1: Embedding outside writeMu in AddDocuments
+// =============================================================================
+
+// TestAddDocuments_EmbeddingOutsideWriteLock verifies that AddDocuments does NOT
+// hold the write lock during embedding computation. A concurrent Query call
+// should complete quickly while AddDocuments is computing embeddings.
+func TestAddDocuments_EmbeddingOutsideWriteLock(t *testing.T) {
+	db, dbCleanup := testDB(t)
+	defer dbCleanup()
+
+	embedSvc, embedCleanup := testEmbeddingService(t)
+	defer embedCleanup()
+
+	client, err := NewClient(Config{DB: db}, embedSvc)
+	require.NoError(t, err)
+
+	// Seed the DB with one document so Query has something to search
+	seedDocs := []Document{
+		{ID: "seed-1", Content: "Seed document for concurrency test"},
+	}
+	err = client.AddDocuments(context.Background(), seedDocs)
+	require.NoError(t, err)
+
+	// Pre-warm the embedding cache for the query text so the Query call
+	// itself doesn't need the embedding mutex — it only needs the DB read lock.
+	_, err = client.Query(context.Background(), "concurrency test", 5, nil)
+	require.NoError(t, err)
+
+	// Prepare a batch of documents to trigger a slow AddDocuments call
+	batchDocs := make([]Document, 10)
+	for i := range batchDocs {
+		batchDocs[i] = Document{
+			ID:      fmt.Sprintf("batch-%d", i),
+			Content: fmt.Sprintf("Batch document number %d for write lock test with unique content", i),
+		}
+	}
+
+	// Launch AddDocuments in background — embedding will take time
+	addDone := make(chan error, 1)
+	go func() {
+		addDone <- client.AddDocuments(context.Background(), batchDocs)
+	}()
+
+	// Give AddDocuments a moment to start embedding computation
+	time.Sleep(10 * time.Millisecond)
+
+	// Invalidate result cache so Query must go through to DB (tests read lock)
+	client.InvalidateResultCache()
+
+	// A concurrent Query should NOT be blocked by AddDocuments.
+	// If the old code held writeMu during embedding, this would block until
+	// embedding finishes. With the fix, it should complete quickly.
+	queryCtx, queryCancel := context.WithTimeout(context.Background(), 2*time.Second)
+	defer queryCancel()
+
+	start := time.Now()
+	_, err = client.Query(queryCtx, "concurrency test", 5, nil)
+	queryDuration := time.Since(start)
+
+	require.NoError(t, err, "Query should succeed while AddDocuments is embedding")
+	// The query should complete well within the timeout if writeMu is not held
+	assert.Less(t, queryDuration, 1*time.Second,
+		"Query should complete quickly when writeMu is not held during embedding")
+
+	// Wait for AddDocuments to finish
+	err = <-addDone
+	require.NoError(t, err, "AddDocuments should succeed")
+}
+
+// =============================================================================
+// REGRESSION TESTS FOR Fix #2a: DoChan + context select in getOrComputeEmbedding
+// =============================================================================
+
+// TestGetOrComputeEmbedding_ContextCancelDuringSingleflight verifies that when
+// a singleflight embedding computation is in progress, a second caller with a
+// short-lived context returns context.DeadlineExceeded promptly rather than
+// waiting for the slow first call to finish.
+func TestGetOrComputeEmbedding_ContextCancelDuringSingleflight(t *testing.T) {
+	db, dbCleanup := testDB(t)
+	defer dbCleanup()
+
+	embedSvc, embedCleanup := testEmbeddingService(t)
+	defer embedCleanup()
+
+	client, err := NewClient(Config{DB: db}, embedSvc)
+	require.NoError(t, err)
+
+	// Add a document so we can query
+	docs := []Document{
+		{ID: "sf-test-1", Content: "Singleflight context cancellation test document"},
+	}
+	err = client.AddDocuments(context.Background(), docs)
+	require.NoError(t, err)
+
+	// Clear cache to force embedding computation
+	client.ClearCache()
+	client.InvalidateResultCache()
+
+	queryText := "unique singleflight context test query"
+
+	// First call: start a normal query in background (will trigger singleflight)
+	firstDone := make(chan struct{})
+	go func() {
+		defer close(firstDone)
+		_, _ = client.Query(context.Background(), queryText, 5, nil)
+	}()
+
+	// Give the first call a moment to start the singleflight computation
+	time.Sleep(5 * time.Millisecond)
+
+	// Second call: use a very short context that should expire quickly
+	shortCtx, shortCancel := context.WithTimeout(context.Background(), 20*time.Millisecond)
+	defer shortCancel()
+
+	// Clear result cache again so the second call hits getOrComputeEmbedding
+	client.InvalidateResultCache()
+
+	start := time.Now()
+	_, err = client.Query(shortCtx, queryText, 5, nil)
+	elapsed := time.Since(start)
+
+	// With DoChan + select, the second caller should return quickly with context error.
+	// Note: If the embedding completes fast enough, the second call may succeed
+	// via singleflight sharing. That's also valid — the test primarily checks
+	// that it doesn't BLOCK for the full embedding duration on context cancel.
+	if err != nil {
+		assert.ErrorIs(t, err, context.DeadlineExceeded,
+			"Should return DeadlineExceeded when context expires during singleflight")
+		assert.Less(t, elapsed, 500*time.Millisecond,
+			"Should return promptly on context cancellation, not wait for slow computation")
+	}
+	// If err == nil, the singleflight completed before the context expired — also valid.
+
+	// Wait for first call to finish
+	<-firstDone
+}

internal/worker/service.go

@@ -1651,11 +1651,15 @@ func (s *Service) processAllSessions() {
 				defer wg.Done()
 				defer func() { <-sem }()
 
+				// Timeout prevents a hung subprocess from blocking the queue processor forever
+				procCtx, procCancel := context.WithTimeout(s.ctx, 60*time.Second)
+				defer procCancel()
+
 				switch msg.Type {
 				case session.MessageTypeObservation:
 					if msg.Observation != nil {
 						err := proc.ProcessObservation(
-							s.ctx,
+							procCtx,
 							sess.SDKSessionID,
 							sess.Project,
 							msg.Observation.ToolName,
@@ -1674,7 +1678,7 @@ func (s *Service) processAllSessions() {
 				case session.MessageTypeSummarize:
 					if msg.Summarize != nil {
 						err := proc.ProcessSummary(
-							s.ctx,
+							procCtx,
 							sess.SessionDBID,
 							sess.SDKSessionID,
 							sess.Project,