diff --git a/chrome_flags.go b/chrome_flags.go
new file mode 100644
index 0000000..143a89e
--- /dev/null
+++ b/chrome_flags.go
@@ -0,0 +1,56 @@
+package main
+
+import (
+	"runtime"
+
+	"github.com/go-rod/rod/lib/launcher"
+)
+
+// singleProcessSupported reports whether --single-process is safe to pass on
+// this platform.
+//
+// The flag collapses Chromium's renderer, GPU and utility services into the
+// browser process. It is there because the multi-process compositor hangs under
+// gVisor and screenshots time out (notes/gvisor-screenshots/README.md), but on
+// macOS the collapse is fatal. media/capture's Apple backend CHECKs that it owns
+// a CFRunLoop-enabled thread, which only holds when video capture has a utility
+// process of its own, so any page that touches navigator.mediaDevices aborts the
+// whole browser:
+//
+//	FATAL:video_capture_device_factory_apple.mm(37)] Check failed: mode.
+//	The MacOS video capture code must be run on a CFRunLoop-enabled thread
+//
+// Device enumeration is routine in the fingerprinting scripts commercial sites
+// ship, which is what made this look like "heavy sites crash the browser".
+// Losing --single-process on macOS costs nothing: gVisor is a Linux sandbox.
+func singleProcessSupported() bool {
+	return runtime.GOOS != "darwin"
+}
+
+// configureExperiments stops Chromium from running features that are still
+// being written.
+//
+// rodney's pinned browser (128.0.6568.0) is a development snapshot, and those
+// builds apply testing/variations/fieldtrial_testing_config.json by default,
+// which force-enables in-development features. One of them, HistoryEmbeddings,
+// computes passage embeddings on the browser's main thread after a navigation
+// and CHECK-fails when its cache lookup misses, killing the browser process a
+// few seconds after a text-heavy page loads:
+//
+//	FATAL:history_embeddings_service.cc(597)] Check failed:
+//	cached_embedding != embedding_cache.end()
+//
+// It needs the on-device model the optimization guide downloads, so it only
+// starts biting once a profile has been used for a while — a long-lived
+// ~/.rodney profile crashes where a throwaway one does not.
+//
+// --disable-field-trial-config drops that config wholesale, so rodney drives a
+// browser with shipped defaults rather than whatever happened to be mid-flight
+// when the snapshot was cut. HistoryEmbeddings is named explicitly as well
+// because a browser supplied through ROD_CHROME_BIN can enable it from a
+// server-side variations seed, which that switch does not cover.
+func configureExperiments(l *launcher.Launcher) *launcher.Launcher {
+	// Append, not Set: rod already disables features of its own.
+	return l.Set("disable-field-trial-config").
+		Append("disable-features", "HistoryEmbeddings")
+}
diff --git a/chrome_flags_test.go b/chrome_flags_test.go
new file mode 100644
index 0000000..350b8ca
--- /dev/null
+++ b/chrome_flags_test.go
@@ -0,0 +1,75 @@
+package main
+
+import (
+	"runtime"
+	"strings"
+	"testing"
+
+	"github.com/go-rod/rod/lib/launcher"
+)
+
+// features returns the --disable-features values as rod would render them.
+func features(l *launcher.Launcher) string {
+	values, ok := l.GetFlags("disable-features")
+	if !ok {
+		return ""
+	}
+	return strings.Join(values, ",")
+}
+
+// HistoryEmbeddings runs in the browser process, so its CHECK failure takes the
+// whole browser down rather than one tab. It is enabled by the field-trial
+// testing config baked into the pinned Chromium snapshot.
+func TestConfigureExperiments_DisablesFieldTrialConfig(t *testing.T) {
+	l := configureExperiments(launcher.New())
+
+	if !l.Has("disable-field-trial-config") {
+		t.Error("--disable-field-trial-config missing; the snapshot will force-enable in-development features")
+	}
+}
+
+func TestConfigureExperiments_DisablesHistoryEmbeddings(t *testing.T) {
+	l := configureExperiments(launcher.New())
+
+	if !strings.Contains(features(l), "HistoryEmbeddings") {
+		t.Errorf("disable-features = %q, want it to contain HistoryEmbeddings", features(l))
+	}
+}
+
+// rod disables features of its own, and configureExtensions appends one more.
+// Setting rather than appending would silently re-enable them.
+func TestConfigureExperiments_KeepsExistingDisabledFeatures(t *testing.T) {
+	l := configureExperiments(launcher.New().Set("disable-features", "site-per-process", "TranslateUI"))
+
+	got := features(l)
+	for _, want := range []string{"site-per-process", "TranslateUI", "HistoryEmbeddings"} {
+		if !strings.Contains(got, want) {
+			t.Errorf("disable-features = %q, want it to contain %q", got, want)
+		}
+	}
+}
+
+// Later callers append to the same flag; the values must accumulate rather than
+// replace each other.
+func TestConfigureExperiments_LeavesRoomForLaterAppends(t *testing.T) {
+	l := configureExperiments(launcher.New().Set("disable-features", "site-per-process"))
+	l = l.Append("disable-features", "SomeLaterFeature")
+
+	got := features(l)
+	for _, want := range []string{"site-per-process", "HistoryEmbeddings", "SomeLaterFeature"} {
+		if !strings.Contains(got, want) {
+			t.Errorf("disable-features = %q, want it to contain %q", got, want)
+		}
+	}
+}
+
+// On macOS --single-process makes any navigator.mediaDevices call abort the
+// browser; on Linux it is what makes screenshots work under gVisor.
+func TestSingleProcessSupported_SkipsMacOSOnly(t *testing.T) {
+	got := singleProcessSupported()
+	want := runtime.GOOS != "darwin"
+
+	if got != want {
+		t.Errorf("singleProcessSupported() = %v on %s, want %v", got, runtime.GOOS, want)
+	}
+}
diff --git a/main.go b/main.go
index a5188db..4700c56 100644
--- a/main.go
+++ b/main.go
@@ -377,11 +377,16 @@ func cmdStart(args []string) {
 	l := launcher.New().
 		Set("no-sandbox").
 		Set("disable-gpu").
-		Set("single-process"). // Required for screenshots in gVisor/container environments
-		Leakless(false).        // Keep Chrome alive after CLI exits
+		Leakless(false). // Keep Chrome alive after CLI exits
 		UserDataDir(dataDir).
 		Headless(headless)
 
+	if singleProcessSupported() {
+		l = l.Set("single-process") // Required for screenshots in gVisor/container environments
+	}
+
+	l = configureExperiments(l)
+
 	// When in non-headless mode, make sure that we show the startup window immediately
 	// (instead of showing a window only after calling "rodney open")
 	if !headless {
diff --git a/notes/gvisor-screenshots/README.md b/notes/gvisor-screenshots/README.md
index 0b2a8b6..19060f7 100644
--- a/notes/gvisor-screenshots/README.md
+++ b/notes/gvisor-screenshots/README.md
@@ -120,6 +120,32 @@ For a CLI automation tool that controls one page at a time, these trade-offs
 are acceptable. The stability concern is mitigated by the fact that `rodney stop`
 and `rodney start` can quickly restart Chrome.
 
+## Not applied on macOS
+
+The "a crash takes down the whole browser" trade-off turned out to be more than
+theoretical on macOS, so `rodney start` skips `--single-process` there
+(`singleProcessSupported` in `chrome_flags.go`).
+
+`media/capture`'s Apple backend CHECKs that it owns a CFRunLoop-enabled thread,
+which only holds when video capture runs in a utility process of its own. Under
+`--single-process` that CHECK fails and aborts the browser:
+
+```
+FATAL:video_capture_device_factory_apple.mm(37)] Check failed: mode.
+The MacOS video capture code must be run on a CFRunLoop-enabled thread
+```
+
+Any page calling `navigator.mediaDevices.enumerateDevices()` triggers it, which
+the device-fingerprinting scripts on most commercial sites do:
+
+```bash
+rodney open https://example.com/
+rodney js "navigator.mediaDevices.enumerateDevices().then(d=>d.length)"
+```
+
+gVisor is a Linux sandbox, so nothing in this investigation is lost by skipping
+the flag on Darwin. Linux and Windows keep it.
+
 ## Reproducing
 
 The test script `screenshot_test.go` in this directory reproduces the investigation.