Merge pull request #1453 from trheyi/main

Enhance sandbox environment setup for VNC and Python compatibility

agent/sandbox/claude/command.go

@@ -121,6 +121,10 @@ func BuildCommandWithContinuation(messages []agentContext.Message, opts *Options
 	// System prompt may contain quotes, newlines, special characters that break shell quoting
 	var bashCmd strings.Builder
 
+	// Ensure $HOME/.Xauthority exists for PyAutoGUI/Xlib (HOME=/workspace).
+	// Xvfb runs without auth, but Xlib requires the file to exist.
+	bashCmd.WriteString("touch /home/sandbox/.Xauthority 2>/dev/null; touch \"$HOME/.Xauthority\" 2>/dev/null\n")
+
 	// If we have a system prompt (first request only), write it to a temp file via heredoc first
 	// then use --append-system-prompt-file
 	if systemPrompt != "" {
@@ -306,10 +310,30 @@ func buildEnvironment(opts *Options, systemPrompt string) map[string]string {
 	// Session data is stored in $HOME/.claude/ (i.e., /workspace/.claude/)
 	env["HOME"] = "/workspace"
 
+	// Fix Python user-site-packages: changing HOME from /home/sandbox to /workspace
+	// breaks Python's ability to find packages installed via pip --user (e.g., playwright,
+	// pyautogui, playwright-stealth) which live in /home/sandbox/.local/lib/pythonX.Y/site-packages/
+	env["PYTHONPATH"] = "/home/sandbox/.local/lib/python3.12/site-packages"
+
+	// Fix X11 auth: PyAutoGUI/Xlib looks for $HOME/.Xauthority, but HOME=/workspace
+	// so it fails to find /home/sandbox/.Xauthority created during image build.
+	// Explicitly set XAUTHORITY to the correct path.
+	env["XAUTHORITY"] = "/home/sandbox/.Xauthority"
+
 	// claude-proxy runs on localhost:3456, Claude CLI connects to it
 	env["ANTHROPIC_BASE_URL"] = "http://127.0.0.1:3456"
 	env["ANTHROPIC_API_KEY"] = "dummy" // Proxy doesn't verify this
 
+	// Pass secrets as environment variables for Claude CLI to use
+	// These are configured in package.yao sandbox.secrets (e.g., LLM_API_KEY, GITHUB_TOKEN)
+	// start-claude-proxy also exports them for the proxy process, but Claude CLI
+	// is launched via a separate docker exec, so it needs them passed explicitly here.
+	if len(opts.Secrets) > 0 {
+		for k, v := range opts.Secrets {
+			env[k] = v
+		}
+	}
+
 	// Note: System prompt and max_turns are passed via CLI flags in BuildCommand
 	// CLAUDE_SYSTEM_PROMPT environment variable is NOT supported by Claude CLI
 	// --append-system-prompt or --system-prompt flags must be used instead

agent/sandbox/claude/executor.go

@@ -1103,14 +1103,13 @@ func (e *Executor) GetVNCUrl() string {
 		return ""
 	}
 
-	// Check if the image supports VNC (playwright or desktop variants)
 	imageName := e.opts.Image
 	if imageName == "" {
 		return ""
 	}
 
-	// VNC is only available for playwright and desktop images
+	// Check if the image supports VNC using the shared keyword list in sandbox package
-	if !strings.Contains(imageName, "playwright") && !strings.Contains(imageName, "desktop") {
+	if !infraSandbox.IsVNCImage(imageName) {
 		return ""
 	}
 

sandbox/manager.go

@@ -22,6 +22,10 @@ import (
 	"github.com/yaoapp/yao/sandbox/ipc"
 )
 
+// vncImageKeywords lists image name keywords that indicate VNC support.
+// Add new keywords here when adding new VNC-capable sandbox images.
+var vncImageKeywords = []string{"playwright", "desktop", "chrome"}
+
 // execReadCloser wraps a Reader with a Closer
 type execReadCloser struct {
 	*bufio.Reader
@@ -324,9 +328,15 @@ func (m *Manager) createContainer(ctx context.Context, opts CreateOptions) (*Con
 		CapDrop:     []string{"ALL"},
 	}
 
+	// Chrome/browser images need SYS_ADMIN for namespace-based process isolation.
+	// Without it, Chrome renderer/GPU processes crash with error code 5.
+	if IsVNCImage(image) {
+		hostConfig.CapAdd = []string{"SYS_ADMIN"}
+	}
+
 	// VNC port mapping for Docker Desktop (macOS/Windows)
 	// Only enable for VNC-capable images (playwright/desktop) when config is enabled
-	if m.config.VNCPortMapping && isVNCImage(image) {
+	if m.config.VNCPortMapping && IsVNCImage(image) {
 		// Expose VNC ports in container config
 		containerConfig.ExposedPorts = nat.PortSet{
 			"6080/tcp": struct{}{}, // noVNC websockify
@@ -952,9 +962,14 @@ func (m *Manager) fixIPCSocketPermissions(ctx context.Context, containerID strin
 	time.Sleep(50 * time.Millisecond)
 }
 
-// isVNCImage checks if the image is VNC-capable (playwright or desktop variants)
+// IsVNCImage checks if the image is VNC-capable based on vncImageKeywords.
-func isVNCImage(imageName string) bool {
+func IsVNCImage(imageName string) bool {
-	return strings.Contains(imageName, "playwright") || strings.Contains(imageName, "desktop")
+	for _, kw := range vncImageKeywords {
+		if strings.Contains(imageName, kw) {
+			return true
+		}
+	}
+	return false
 }
 
 // findAvailablePort finds an available port on the host