refactor(sandbox): Unify host process management and refine tool enablement logic within the sandbox.

pkg/agent/instance.go

@@ -64,27 +64,35 @@ func NewAgentInstance(
 	toolsRegistry := tools.NewToolRegistry()
 
 	sandboxManager := sandbox.NewFromConfigWithAgent(workspace, restrict, cfg, agentID)
-	isSandboxAllowed := func(toolName string) bool {
+	// isToolEnabled determines if a base system tool should be registered for the LLM.
-		return sandboxManager == nil || sandbox.IsToolSandboxEnabled(cfg, toolName)
+	// SandboxManager *always* provides a Sandbox context (HostSandbox if mode="off").
+	isToolEnabled := func(toolName string) bool {
+		// If the user explicitly disables the sandbox mode (Mode = "off"), they are opting out of
+		// safety isolation. We grant the LLM access to all core tools via the HostSandbox.
+		if isSandboxModeOff(cfg) {
+			return true
+		}
+		// Otherwise, respect the fine-grained allow/deny policy defined in the Sandbox config.
+		return sandbox.IsToolSandboxEnabled(cfg, toolName)
 	}
 
-	if isSandboxAllowed("read_file") {
+	if isToolEnabled("read_file") {
 		toolsRegistry.Register(tools.NewReadFileTool(workspace, restrict))
 	}
-	if !roContainer && isSandboxAllowed("write_file") {
+	if !roContainer && isToolEnabled("write_file") {
 		toolsRegistry.Register(tools.NewWriteFileTool(workspace, restrict))
 	}
-	if isSandboxAllowed("list_dir") {
+	if isToolEnabled("list_dir") {
 		toolsRegistry.Register(tools.NewListDirTool(workspace, restrict))
 	}
-	if isSandboxAllowed("exec") {
+	if isToolEnabled("exec") {
 		toolsRegistry.Register(tools.NewExecToolWithConfig(workspace, restrict, cfg))
 	}
 	if !roContainer {
-		if isSandboxAllowed("edit_file") {
+		if isToolEnabled("edit_file") {
 			toolsRegistry.Register(tools.NewEditFileTool(workspace, restrict))
 		}
-		if isSandboxAllowed("append_file") {
+		if isToolEnabled("append_file") {
 			toolsRegistry.Register(tools.NewAppendFileTool(workspace, restrict))
 		}
 	}
@@ -170,8 +178,15 @@ func isContainerReadOnlySandbox(cfg *config.Config) bool {
 	if cfg == nil {
 		return false
 	}
-	return strings.EqualFold(strings.TrimSpace(cfg.Agents.Defaults.Sandbox.Mode), "all") &&
+	return cfg.Agents.Defaults.Sandbox.Mode == config.SandboxModeAll &&
-		strings.EqualFold(strings.TrimSpace(cfg.Agents.Defaults.Sandbox.WorkspaceAccess), "ro")
+		cfg.Agents.Defaults.Sandbox.WorkspaceAccess == config.WorkspaceAccessRO
+}
+
+func isSandboxModeOff(cfg *config.Config) bool {
+	if cfg == nil {
+		return false
+	}
+	return cfg.Agents.Defaults.Sandbox.Mode == config.SandboxModeOff
 }
 
 func expandHome(path string) string {

pkg/agent/instance_test.go

@@ -135,3 +135,36 @@ func TestNewAgentInstance_ReadOnlyContainerOmitsWriteTools(t *testing.T) {
 		t.Fatalf("write_file should be absent in ro sandbox, got: %+v", writeRes)
 	}
 }
+
+func TestNewAgentInstance_SandboxModeOffRegistersFullToolSet(t *testing.T) {
+	tmpDir, err := os.MkdirTemp("", "agent-instance-test-*")
+	if err != nil {
+		t.Fatalf("Failed to create temp dir: %v", err)
+	}
+	defer os.RemoveAll(tmpDir)
+
+	cfg := &config.Config{
+		Agents: config.AgentsConfig{
+			Defaults: config.AgentDefaults{
+				Workspace:         tmpDir,
+				Model:             "test-model",
+				MaxTokens:         1234,
+				MaxToolIterations: 5,
+				Sandbox: config.AgentSandboxConfig{
+					Mode: "off",
+				},
+			},
+		},
+	}
+	// Keep sandbox allowlist narrow to ensure mode=off bypasses sandbox policy gating.
+	cfg.Tools.Sandbox.Tools.Allow = []string{"exec"}
+
+	provider := &mockProvider{}
+	agent := NewAgentInstance(nil, &cfg.Agents.Defaults, cfg, provider)
+
+	for _, name := range []string{"read_file", "write_file", "list_dir", "exec", "edit_file", "append_file"} {
+		if _, ok := agent.Tools.Get(name); !ok {
+			t.Fatalf("%s should be registered when sandbox.mode=off", name)
+		}
+	}
+}

pkg/agent/loop.go

@@ -409,16 +409,17 @@ func (al *AgentLoop) runAgentLoop(ctx context.Context, agent *AgentInstance, opt
 	// it against the main session key for non-main mode.
 	ctx = sandbox.WithSessionKey(ctx, opts.SessionKey)
 
-	// Resolve sandbox environment for this run
+	// Resolve sandbox environment for this run.
-	if agent.SandboxManager != nil {
+	// We guarantee SandboxManager is non-nil (fallback to host manager) in NewAgentInstance.
+	ctx = sandbox.WithManager(ctx, agent.SandboxManager)
+
 	sb, err := agent.SandboxManager.Resolve(ctx)
 	if err != nil {
 		logger.ErrorCF("agent", "Failed to resolve sandbox", map[string]any{"error": err.Error()})
 		return "", fmt.Errorf("failed to resolve sandbox: %w", err)
 	}
-		// Add sandbox to context for thread-safe access in tools via registry
+	// Add pre-resolved sandbox to context for thread-safe access in tools
 	ctx = sandbox.WithSandbox(ctx, sb)
-	}
 
 	// 2. Build messages (skip history for heartbeat)
 	var history []providers.Message

pkg/agent/sandbox/container.go

@@ -7,6 +7,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"io"
+	"io/fs"
 	"math"
 	"os"
 	"path"
@@ -101,7 +102,7 @@ func NewContainerSandbox(cfg ContainerSandboxConfig) *ContainerSandbox {
 	if cfg.Env == nil {
 		cfg.Env = map[string]string{"LANG": "C.UTF-8"}
 	}
-	cfg.WorkspaceAccess = normalizeWorkspaceAccess(cfg.WorkspaceAccess)
+	cfg.WorkspaceAccess = string(normalizeWorkspaceAccess(config.WorkspaceAccess(cfg.WorkspaceAccess)))
 	cfg.WorkspaceRoot = strings.TrimSpace(cfg.WorkspaceRoot)
 	sb := &ContainerSandbox{cfg: cfg}
 	sb.hash = computeContainerConfigHash(cfg)
@@ -125,7 +126,7 @@ func (c *ContainerSandbox) Start(ctx context.Context) error {
 		c.startErr = err
 		return err
 	}
-	if strings.TrimSpace(c.cfg.Workspace) != "" && c.cfg.WorkspaceAccess == "none" {
+	if strings.TrimSpace(c.cfg.Workspace) != "" && c.cfg.WorkspaceAccess == string(config.WorkspaceAccessNone) {
 		if err := os.MkdirAll(c.cfg.Workspace, 0o755); err != nil {
 			c.startErr = fmt.Errorf("sandbox workspace init failed: %w", err)
 			return c.startErr
@@ -255,6 +256,13 @@ func (c *ContainerSandbox) ExecStream(
 	if err != nil {
 		return nil, fmt.Errorf("docker exec attach failed: %w", err)
 	}
+
+	// Prevent stdcopy.StdCopy from blocking indefinitely if the container hangs.
+	// We force close the hijacked connection when the context times out.
+	go func() {
+		<-execCtx.Done()
+		attach.Close()
+	}()
 	defer attach.Close()
 
 	var stdout, stderr bytes.Buffer
@@ -404,16 +412,16 @@ func (c *ContainerSandbox) binds() []string {
 	}
 
 	if hostDir != "" {
-		if c.cfg.WorkspaceAccess == "none" {
+		if c.cfg.WorkspaceAccess == string(config.WorkspaceAccessNone) {
 			// Ensure the isolated directory exists on the host so Docker doesn't create it as root
 			_ = os.MkdirAll(hostDir, 0o755)
 		}
 		// Add :Z flag for SELinux (Podman) to label the content with a private unshared label.
 		// This fixes errors like: "crun: getcwd: Operation not permitted: OCI permission denied"
-		switch c.cfg.WorkspaceAccess {
+		switch config.WorkspaceAccess(c.cfg.WorkspaceAccess) {
-		case "ro":
+		case config.WorkspaceAccessRO:
 			binds = append(binds, fmt.Sprintf("%s:%s:ro,Z", hostDir, c.cfg.Workdir))
-		case "rw", "none":
+		case config.WorkspaceAccessRW, config.WorkspaceAccessNone:
 			binds = append(binds, fmt.Sprintf("%s:%s:rw,Z", hostDir, c.cfg.Workdir))
 		default:
 			// Default to no mount for unknown access types
@@ -574,7 +582,7 @@ func (f *containerFS) ReadFile(ctx context.Context, p string) ([]byte, error) {
 			return content, nil
 		}
 	}
-	return nil, fmt.Errorf("file not found in container: %s", containerPath)
+	return nil, fmt.Errorf("file not found in container %s: %w", containerPath, fs.ErrNotExist)
 }
 
 func (f *containerFS) WriteFile(ctx context.Context, p string, data []byte, mkdir bool) error {
@@ -630,8 +638,94 @@ func (f *containerFS) WriteFile(ctx context.Context, p string, data []byte, mkdi
 	return nil
 }
 
+func (f *containerFS) ReadDir(ctx context.Context, p string) ([]os.DirEntry, error) {
+	if err := f.sb.ensureContainer(ctx); err != nil {
+		return nil, err
+	}
+	containerPath, err := resolveContainerPathWithRoot(f.sb.cfg.Workdir, p)
+	if err != nil {
+		return nil, err
+	}
+
+	rc, _, err := f.sb.cli.CopyFromContainer(ctx, f.sb.cfg.ContainerName, containerPath)
+	if err != nil {
+		return nil, fmt.Errorf("docker copy from container failed: %w", err)
+	}
+	defer rc.Close()
+
+	var entries []os.DirEntry
+	tr := tar.NewReader(rc)
+	entries, err = parseTopLevelDirEntriesFromTar(tr)
+	if err != nil {
+		return nil, err
+	}
+	return entries, nil
+}
+
+func parseTopLevelDirEntriesFromTar(tr *tar.Reader) ([]os.DirEntry, error) {
+	var entries []os.DirEntry
+	seen := make(map[string]struct{})
+	rootName := ""
+	first := true
+
+	for {
+		hdr, err := tr.Next()
+		if err == io.EOF {
+			break
+		}
+		if err != nil {
+			return nil, fmt.Errorf("tar read failed: %w", err)
+		}
+
+		cleanName := path.Clean(hdr.Name)
+		if first {
+			first = false
+			rootName = cleanName
+			continue
+		}
+
+		relName := cleanName
+		if rootName != "" && rootName != "." {
+			if relName == rootName {
+				continue
+			}
+			prefix := rootName + "/"
+			relName = strings.TrimPrefix(relName, prefix)
+		}
+		relName = strings.TrimPrefix(relName, "./")
+		if relName == "" || relName == "." {
+			continue
+		}
+		if strings.Contains(relName, "/") {
+			// CopyFromContainer is recursive; keep only immediate children.
+			continue
+		}
+		if _, ok := seen[relName]; ok {
+			continue
+		}
+		seen[relName] = struct{}{}
+
+		entries = append(entries, &containerDirEntry{
+			name: relName,
+			info: hdr.FileInfo(),
+		})
+	}
+
+	return entries, nil
+}
+
+type containerDirEntry struct {
+	name string
+	info os.FileInfo
+}
+
+func (d *containerDirEntry) Name() string               { return d.name }
+func (d *containerDirEntry) IsDir() bool                { return d.info.IsDir() }
+func (d *containerDirEntry) Type() os.FileMode          { return d.info.Mode().Type() }
+func (d *containerDirEntry) Info() (os.FileInfo, error) { return d.info, nil }
+
 func (c *ContainerSandbox) hostDirForContainerPath(containerDir string) (string, bool) {
-	if c.cfg.WorkspaceAccess == "ro" {
+	if c.cfg.WorkspaceAccess == string(config.WorkspaceAccessRO) {
 		return "", false
 	}
 	workspace := strings.TrimSpace(c.cfg.Workspace)

pkg/agent/sandbox/container_integration_test.go

@@ -103,6 +103,22 @@ func TestContainerSandbox_Integration_ExecReadWrite(t *testing.T) {
 	if strings.TrimSpace(pwdRes.Stdout) != "/workspace/it" {
 		t.Fatalf("pwd mismatch: got %q want %q", strings.TrimSpace(pwdRes.Stdout), "/workspace/it")
 	}
+
+	// Test ReadDir
+	entries, err := sb.Fs().ReadDir(ctx, "it")
+	if err != nil {
+		t.Fatalf("ReadDir failed: %v", err)
+	}
+	found := false
+	for _, e := range entries {
+		if e.Name() == "write.txt" {
+			found = true
+			break
+		}
+	}
+	if !found {
+		t.Errorf("ReadDir result missing 'write.txt'")
+	}
 }
 
 func TestContainerSandbox_Integration_WriteFileMkdirInContainerTmp(t *testing.T) {
@@ -364,3 +380,48 @@ func TestContainerSandbox_Integration_ExecTimeoutRespectsRequest(t *testing.T) {
 		t.Fatalf("expected timeout to trigger early, took %v", time.Since(start))
 	}
 }
+
+func TestContainerSandbox_Integration_ExecTimeoutBreaksStdCopyBlock(t *testing.T) {
+	if os.Getenv("PICOCLAW_RUN_DOCKER_TESTS") != "1" {
+		t.Skip("set PICOCLAW_RUN_DOCKER_TESTS=1 to run docker integration tests")
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	containerName := fmt.Sprintf("picoclaw-test-timeout-block-%d", time.Now().UnixNano())
+	image := strings.TrimSpace(os.Getenv("PICOCLAW_DOCKER_TEST_IMAGE"))
+	if image == "" {
+		image = "debian:bookworm-slim"
+	}
+
+	sb := NewContainerSandbox(ContainerSandboxConfig{
+		Image:         image,
+		ContainerName: containerName,
+		Workspace:     t.TempDir(),
+	})
+	if err := sb.Start(ctx); err != nil {
+		t.Fatalf("sandbox start failed: %v", err)
+	}
+	defer func() {
+		_ = sb.Prune(context.Background())
+		if sb.cli != nil {
+			_ = sb.cli.ContainerRemove(context.Background(), containerName, container.RemoveOptions{Force: true})
+		}
+	}()
+
+	start := time.Now()
+	// Run a command that sleeps for a very long time holding the stream open.
+	// We set a 500ms timeout. If StdCopy isn't broken asynchronously, the Exec call will hang.
+	_, err := sb.Exec(ctx, ExecRequest{
+		Command:   "sh -c 'sleep 1000'",
+		TimeoutMs: 500,
+	})
+	if err == nil {
+		t.Fatal("expected timeout error for hanging command")
+	}
+	elapsed := time.Since(start)
+	if elapsed > 2*time.Second {
+		t.Fatalf("expected timeout to trigger within 2s, but it blocked for %v (StdCopy might be hanging)", elapsed)
+	}
+}

pkg/agent/sandbox/container_test.go

@@ -1,6 +1,8 @@
 package sandbox
 
 import (
+	"archive/tar"
+	"bytes"
 	"context"
 	"errors"
 	"os"
@@ -53,6 +55,48 @@ func TestResolveContainerPath_RejectsAbsoluteOutsideWorkspace(t *testing.T) {
 	}
 }
 
+func TestParseTopLevelDirEntriesFromTar_KeepImmediateChildren(t *testing.T) {
+	var buf bytes.Buffer
+	tw := tar.NewWriter(&buf)
+	writeHeader := func(hdr *tar.Header, content []byte) {
+		t.Helper()
+		if err := tw.WriteHeader(hdr); err != nil {
+			t.Fatalf("write header failed: %v", err)
+		}
+		if len(content) > 0 {
+			if _, err := tw.Write(content); err != nil {
+				t.Fatalf("write content failed: %v", err)
+			}
+		}
+	}
+
+	writeHeader(&tar.Header{Name: "it", Typeflag: tar.TypeDir, Mode: 0o755}, nil)
+	writeHeader(&tar.Header{Name: "it/write.txt", Typeflag: tar.TypeReg, Mode: 0o644, Size: 1}, []byte("x"))
+	writeHeader(&tar.Header{Name: "it/sub", Typeflag: tar.TypeDir, Mode: 0o755}, nil)
+	writeHeader(&tar.Header{Name: "it/sub/nested.txt", Typeflag: tar.TypeReg, Mode: 0o644, Size: 1}, []byte("y"))
+	if err := tw.Close(); err != nil {
+		t.Fatalf("tar close failed: %v", err)
+	}
+
+	entries, err := parseTopLevelDirEntriesFromTar(tar.NewReader(&buf))
+	if err != nil {
+		t.Fatalf("parseTopLevelDirEntriesFromTar failed: %v", err)
+	}
+
+	got := make(map[string]struct{}, len(entries))
+	for _, e := range entries {
+		got[e.Name()] = struct{}{}
+	}
+	for _, want := range []string{"write.txt", "sub"} {
+		if _, ok := got[want]; !ok {
+			t.Fatalf("missing top-level entry %q in %+v", want, got)
+		}
+	}
+	if _, ok := got["nested.txt"]; ok {
+		t.Fatalf("nested file should not appear in top-level entries: %+v", got)
+	}
+}
+
 func TestBuildExecCommand_DefaultShell(t *testing.T) {
 	sb := NewContainerSandbox(ContainerSandboxConfig{})
 	cmd, wd, err := sb.buildExecCommand(ExecRequest{

pkg/agent/sandbox/host.go

@@ -5,6 +5,7 @@ import (
 	"context"
 	"fmt"
 	"io"
+	"io/fs"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -90,7 +91,7 @@ func (h *HostSandbox) ExecStream(
 	}
 
 	if req.WorkingDir != "" {
-		dir, err := validatePath(req.WorkingDir, h.workspace, h.restrict)
+		dir, err := ValidatePath(req.WorkingDir, h.workspace, h.restrict)
 		if err != nil {
 			return nil, err
 		}
@@ -105,6 +106,9 @@ func (h *HostSandbox) ExecStream(
 	if err != nil {
 		return nil, fmt.Errorf("stderr pipe setup failed: %w", err)
 	}
+
+	prepareCommandForTermination(cmd)
+
 	if err := cmd.Start(); err != nil {
 		return nil, err
 	}
@@ -158,9 +162,11 @@ func (h *HostSandbox) ExecStream(
 
 	waitErr := cmd.Wait()
 	if streamErr != nil {
+		_ = terminateProcessTree(cmd)
 		return nil, streamErr
 	}
 	if cmdCtx.Err() != nil {
+		_ = terminateProcessTree(cmd)
 		return nil, cmdCtx.Err()
 	}
 
@@ -205,7 +211,7 @@ func (h *hostFS) getSafeRelPath(path string) (string, error) {
 func (h *hostFS) ReadFile(ctx context.Context, path string) ([]byte, error) {
 	if !h.restrict || h.workspace == "" || h.root == nil {
 		// Unrestricted mode continues to use traditional resolution
-		resolved, err := validatePath(path, h.workspace, h.restrict)
+		resolved, err := ValidatePath(path, h.workspace, h.restrict)
 		if err != nil {
 			return nil, err
 		}
@@ -225,7 +231,7 @@ func (h *hostFS) ReadFile(ctx context.Context, path string) ([]byte, error) {
 func (h *hostFS) WriteFile(ctx context.Context, path string, data []byte, mkdir bool) error {
 	if !h.restrict || h.workspace == "" || h.root == nil {
 		// Unrestricted mode continues to use traditional resolution
-		resolved, err := validatePath(path, h.workspace, h.restrict)
+		resolved, err := ValidatePath(path, h.workspace, h.restrict)
 		if err != nil {
 			return err
 		}
@@ -234,7 +240,17 @@ func (h *hostFS) WriteFile(ctx context.Context, path string, data []byte, mkdir
 				return err
 			}
 		}
-		return os.WriteFile(resolved, data, 0o644)
+		// Atomic write: write to temp file then rename to prevent partial writes.
+		tmpPath := fmt.Sprintf("%s.%d.tmp", resolved, time.Now().UnixNano())
+		if err := os.WriteFile(tmpPath, data, 0o644); err != nil {
+			os.Remove(tmpPath)
+			return fmt.Errorf("failed to write temp file: %w", err)
+		}
+		if err := os.Rename(tmpPath, resolved); err != nil {
+			os.Remove(tmpPath)
+			return fmt.Errorf("failed to replace original file: %w", err)
+		}
+		return nil
 	}
 
 	relPath, err := h.getSafeRelPath(path)
@@ -248,80 +264,32 @@ func (h *hostFS) WriteFile(ctx context.Context, path string, data []byte, mkdir
 			return err
 		}
 	}
-	// Uses OS-level guarantees to restrict the file writing within the root descriptor.
+	// Atomic write within os.Root: write to temp file then rename.
-	return h.root.WriteFile(relPath, data, 0o644)
+	tmpRelPath := fmt.Sprintf("%s.%d.tmp", relPath, time.Now().UnixNano())
+	if err := h.root.WriteFile(tmpRelPath, data, 0o644); err != nil {
+		h.root.Remove(tmpRelPath)
+		return fmt.Errorf("failed to write temp file: %w", err)
+	}
+	if err := h.root.Rename(tmpRelPath, relPath); err != nil {
+		h.root.Remove(tmpRelPath)
+		return fmt.Errorf("failed to rename temp file over target: %w", err)
+	}
+	return nil
 }
 
-// validatePath ensures the given path is within the workspace if restrict is true but does not ensure atomic TOCTOU protection.
+func (h *hostFS) ReadDir(ctx context.Context, path string) ([]os.DirEntry, error) {
-// It is kept for setting string-based fields like cmd.Dir where os.Root cannot be directly mapped.
+	if !h.restrict || h.workspace == "" || h.root == nil {
-// The secure file operations boundary relies on os.Root implemented in FsBridge.
+		resolved, err := ValidatePath(path, h.workspace, h.restrict)
-// validatePath ensures the given path is within the workspace if restrict is true.
-func validatePath(path, workspace string, restrict bool) (string, error) {
-	if workspace == "" {
-		return path, nil
-	}
-
-	absWorkspace, err := filepath.Abs(workspace)
 		if err != nil {
-		return "", fmt.Errorf("failed to resolve workspace path: %w", err)
+			return nil, err
+		}
+		return os.ReadDir(resolved)
 	}
 
-	var absPath string
+	relPath, err := h.getSafeRelPath(path)
-	if filepath.IsAbs(path) {
-		absPath = filepath.Clean(path)
-	} else {
-		absPath, err = filepath.Abs(filepath.Join(absWorkspace, path))
 	if err != nil {
-			return "", fmt.Errorf("failed to resolve file path: %w", err)
+		return nil, err
-		}
 	}
 
-	if restrict {
+	return fs.ReadDir(h.root.FS(), relPath)
-		if !isWithinWorkspace(absPath, absWorkspace) {
-			return "", ErrOutsideWorkspace
-		}
-
-		var resolved string
-		workspaceReal := absWorkspace
-		if resolved, err = filepath.EvalSymlinks(absWorkspace); err == nil {
-			workspaceReal = resolved
-		}
-
-		if resolved, err = filepath.EvalSymlinks(absPath); err == nil {
-			if !isWithinWorkspace(resolved, workspaceReal) {
-				return "", ErrOutsideWorkspace
-			}
-		} else if os.IsNotExist(err) {
-			var parentResolved string
-			if parentResolved, err = resolveExistingAncestor(filepath.Dir(absPath)); err == nil {
-				if !isWithinWorkspace(parentResolved, workspaceReal) {
-					return "", fmt.Errorf("access denied: symlink resolves outside workspace")
-				}
-			} else if !os.IsNotExist(err) {
-				return "", fmt.Errorf("failed to resolve path: %w", err)
-			}
-		} else {
-			return "", fmt.Errorf("failed to resolve path: %w", err)
-		}
-	}
-
-	return absPath, nil
-}
-
-func resolveExistingAncestor(path string) (string, error) {
-	for current := filepath.Clean(path); ; current = filepath.Dir(current) {
-		if resolved, err := filepath.EvalSymlinks(current); err == nil {
-			return resolved, nil
-		} else if !os.IsNotExist(err) {
-			return "", err
-		}
-		if filepath.Dir(current) == current {
-			return "", os.ErrNotExist
-		}
-	}
-}
-
-func isWithinWorkspace(candidate, workspace string) bool {
-	rel, err := filepath.Rel(filepath.Clean(workspace), filepath.Clean(candidate))
-	return err == nil && filepath.IsLocal(rel)
 }

pkg/agent/sandbox/host_process_unix.go

@@ -1,6 +1,6 @@
 //go:build !windows
 
-package tools
+package sandbox
 
 import (
 	"os/exec"

pkg/agent/sandbox/host_process_windows.go

@@ -1,6 +1,6 @@
 //go:build windows
 
-package tools
+package sandbox
 
 import (
 	"os/exec"

pkg/agent/sandbox/host_test.go

@@ -77,7 +77,7 @@ func TestHostSandbox_ExecAndFs(t *testing.T) {
 func TestHostSandbox_ResolvePathRestrictions(t *testing.T) {
 	root := t.TempDir()
 
-	got, err := validatePath("a/b.txt", root, true)
+	got, err := ValidatePath("a/b.txt", root, true)
 	if err != nil {
 		t.Fatalf("resolvePath relative error: %v", err)
 	}
@@ -86,7 +86,7 @@ func TestHostSandbox_ResolvePathRestrictions(t *testing.T) {
 		t.Fatalf("resolvePath relative got %q, want %q", got, want)
 	}
 
-	_, err = validatePath(filepath.Join(root, "..", "outside.txt"), root, true)
+	_, err = ValidatePath(filepath.Join(root, "..", "outside.txt"), root, true)
 	if err == nil || !errors.Is(err, ErrOutsideWorkspace) {
 		t.Fatalf("expected outside workspace error, got: %v", err)
 	}
@@ -97,7 +97,7 @@ func TestHostSandbox_ResolvePathRestrictions(t *testing.T) {
 	}
 	link := filepath.Join(root, "link.txt")
 	if err := os.Symlink(target, link); err == nil {
-		_, err = validatePath("link.txt", root, true)
+		_, err = ValidatePath("link.txt", root, true)
 		if err == nil || !errors.Is(err, ErrOutsideWorkspace) {
 			t.Fatalf("expected symlink outside error, got: %v", err)
 		}
@@ -158,6 +158,56 @@ func TestHostFS_ReadFileWriteFile_Restricted(t *testing.T) {
 	}
 }
 
+func TestHostFS_ReadDir(t *testing.T) {
+	root := t.TempDir()
+	os.MkdirAll(filepath.Join(root, "a/b"), 0o755)
+	os.WriteFile(filepath.Join(root, "a/f1.txt"), []byte("1"), 0o644)
+	os.WriteFile(filepath.Join(root, "a/b/f2.txt"), []byte("2"), 0o644)
+
+	sb := NewHostSandbox(root, true)
+	if err := sb.Start(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+	defer sb.Prune(context.Background())
+
+	// Test restricted ReadDir on "a"
+	entries, err := sb.Fs().ReadDir(context.Background(), "a")
+	if err != nil {
+		t.Fatalf("ReadDir failed: %v", err)
+	}
+
+	foundF1 := false
+	foundB := false
+	for _, e := range entries {
+		if e.Name() == "f1.txt" && !e.IsDir() {
+			foundF1 = true
+		}
+		if e.Name() == "b" && e.IsDir() {
+			foundB = true
+		}
+	}
+	if !foundF1 || !foundB {
+		t.Errorf("ReadDir result missing expected entries: foundF1=%v, foundB=%v", foundF1, foundB)
+	}
+
+	// Test unrestricted ReadDir
+	sb2 := NewHostSandbox(root, false)
+	entries2, err := sb2.Fs().ReadDir(context.Background(), root)
+	if err != nil {
+		t.Fatalf("ReadDir unrestricted failed: %v", err)
+	}
+	foundA := false
+	for _, e := range entries2 {
+		if e.Name() == "a" {
+			foundA = true
+			break
+		}
+	}
+	if !foundA {
+		t.Errorf("ReadDir unrestricted missing 'a'")
+	}
+}
+
 func TestHostFS_ReadFileWriteFile_Unrestricted(t *testing.T) {
 	root := t.TempDir()
 	sb := NewHostSandbox(root, false)
@@ -264,16 +314,13 @@ func TestHostFS_ReadFileWriteFile_WithoutWorkspaceOrRoot(t *testing.T) {
 
 	content := []byte("hello empty workspace")
 	target := filepath.Join(root, "empty.txt")
-	if err := sb.Fs().WriteFile(context.Background(), target, content, true); err != nil {
+	if err := sb.Fs().WriteFile(context.Background(), target, content, true); err == nil {
-		t.Fatalf("WriteFile failed: %v", err)
+		t.Fatalf("expected WriteFile to fail due to empty workspace with restrict=true")
 	}
 
-	readContent, err := sb.Fs().ReadFile(context.Background(), target)
+	_, err := sb.Fs().ReadFile(context.Background(), target)
-	if err != nil {
+	if err == nil {
-		t.Fatalf("ReadFile failed: %v", err)
+		t.Fatalf("expected ReadFile to fail due to empty workspace with restrict=true")
-	}
-	if string(readContent) != string(content) {
-		t.Fatalf("content mismatch")
 	}
 
 	// Test case where root is nil explicitly
@@ -284,7 +331,7 @@ func TestHostFS_ReadFileWriteFile_WithoutWorkspaceOrRoot(t *testing.T) {
 		t.Fatalf("WriteFile failed: %v", err)
 	}
 
-	readContent, err = sb2.Fs().ReadFile(context.Background(), "nil_root_test.txt")
+	readContent, err := sb2.Fs().ReadFile(context.Background(), "nil_root_test.txt")
 	if err != nil {
 		t.Fatalf("ReadFile failed: %v", err)
 	}
@@ -294,9 +341,9 @@ func TestHostFS_ReadFileWriteFile_WithoutWorkspaceOrRoot(t *testing.T) {
 }
 
 func TestValidatePathErrors(t *testing.T) {
-	_, err := validatePath("/a/b/c", "", true)
+	_, err := ValidatePath("/a/b/c", "", true)
-	if err != nil {
+	if err == nil {
-		t.Fatalf("expected no err for empty workspace with abs path")
+		t.Fatalf("expected err for empty workspace with restrict=true")
 	}
 
 	root := t.TempDir()
@@ -306,7 +353,7 @@ func TestValidatePathErrors(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	_, err = validatePath("a.txt/b.txt", root, true)
+	_, err = ValidatePath("a.txt/b.txt", root, true)
 	if err == nil {
 		t.Fatalf("expected error when ancestor is file")
 	}

pkg/agent/sandbox/manager.go

@@ -27,11 +27,11 @@ func NewFromConfig(workspace string, restrict bool, cfg *config.Config) Sandbox
 }
 
 // NewFromConfigWithAgent builds the sandbox Manager for an agent.
-// Returns nil when sandboxing is disabled (mode=off), so callers can check manager != nil.
+// It always returns a non-nil Manager (falling back to a host manager or error manager if needed).
 func NewFromConfigWithAgent(workspace string, restrict bool, cfg *config.Config, agentID string) Manager {
-	mode := "all"
+	mode := config.SandboxModeAll
-	scope := "agent"
+	scope := config.SandboxScopeAgent
-	workspaceAccess := "none"
+	workspaceAccess := config.WorkspaceAccessNone
 	workspaceRoot := "~/.picoclaw/sandboxes"
 	image := "picoclaw-sandbox:bookworm-slim"
 	containerPrefix := "picoclaw-sandbox-"
@@ -41,14 +41,14 @@ func NewFromConfigWithAgent(workspace string, restrict bool, cfg *config.Config,
 
 	if cfg != nil {
 		sb := cfg.Agents.Defaults.Sandbox
-		if strings.TrimSpace(sb.Mode) != "" {
+		if sb.Mode != "" {
-			mode = strings.TrimSpace(sb.Mode)
+			mode = sb.Mode
 		}
-		if strings.TrimSpace(sb.Scope) != "" {
+		if sb.Scope != "" {
-			scope = strings.TrimSpace(sb.Scope)
+			scope = sb.Scope
 		}
-		if strings.TrimSpace(sb.WorkspaceAccess) != "" {
+		if sb.WorkspaceAccess != "" {
-			workspaceAccess = strings.TrimSpace(sb.WorkspaceAccess)
+			workspaceAccess = sb.WorkspaceAccess
 		}
 		if strings.TrimSpace(sb.WorkspaceRoot) != "" {
 			workspaceRoot = strings.TrimSpace(sb.WorkspaceRoot)
@@ -71,13 +71,14 @@ func NewFromConfigWithAgent(workspace string, restrict bool, cfg *config.Config,
 	agentID = routing.NormalizeAgentID(agentID)
 
 	resolvedMode := normalizeSandboxMode(mode)
-	if resolvedMode == "off" {
-		return nil // sandbox disabled; host-level access is handled directly by tools
-	}
-
 	host := NewHostSandbox(workspace, restrict)
 	_ = host.Start(context.Background())
 
+	// When sandbox is disabled, skip building container infrastructure entirely.
+	if resolvedMode == config.SandboxModeOff {
+		return &hostOnlyManager{host: host}
+	}
+
 	resolvedScope := normalizeSandboxScope(scope)
 	normalizedAccess := normalizeWorkspaceAccess(workspaceAccess)
 	workspaceRootAbs := resolveAbsPath(expandHomePath(workspaceRoot))
@@ -105,31 +106,33 @@ func NewFromConfigWithAgent(workspace string, restrict bool, cfg *config.Config,
 	return manager
 }
 
-func normalizeWorkspaceAccess(access string) string {
+func normalizeWorkspaceAccess(access config.WorkspaceAccess) config.WorkspaceAccess {
-	v := strings.ToLower(strings.TrimSpace(access))
+	v := config.WorkspaceAccess(strings.ToLower(strings.TrimSpace(string(access))))
 	switch v {
-	case "ro", "rw":
+	case config.WorkspaceAccessRO, config.WorkspaceAccessRW:
 		return v
 	default:
-		return "none"
+		return config.WorkspaceAccessNone
 	}
 }
 
-func normalizeSandboxMode(mode string) string {
+func normalizeSandboxMode(mode config.SandboxMode) config.SandboxMode {
-	switch strings.ToLower(strings.TrimSpace(mode)) {
+	v := config.SandboxMode(strings.ToLower(strings.TrimSpace(string(mode))))
-	case "all", "non-main":
+	switch v {
-		return strings.ToLower(strings.TrimSpace(mode))
+	case config.SandboxModeAll, config.SandboxModeNonMain:
+		return v
 	default:
-		return "off"
+		return config.SandboxModeOff
 	}
 }
 
-func normalizeSandboxScope(scope string) string {
+func normalizeSandboxScope(scope config.SandboxScope) config.SandboxScope {
-	switch strings.ToLower(strings.TrimSpace(scope)) {
+	v := config.SandboxScope(strings.ToLower(strings.TrimSpace(string(scope))))
-	case "session", "shared":
+	switch v {
-		return strings.ToLower(strings.TrimSpace(scope))
+	case config.SandboxScopeSession, config.SandboxScopeShared:
+		return v
 	default:
-		return "agent"
+		return config.SandboxScopeAgent
 	}
 }
 
@@ -165,13 +168,13 @@ func resolveAbsPath(p string) string {
 }
 
 type scopedSandboxManager struct {
-	mode            string
+	mode            config.SandboxMode
-	scope           string
+	scope           config.SandboxScope
 	agentID         string
 	host            Sandbox
 	image           string
 	containerPrefix string
-	workspaceAccess string
+	workspaceAccess config.WorkspaceAccess
 	workspaceRoot   string
 	agentWorkspace  string
 	pruneIdleHours  int
@@ -188,7 +191,7 @@ type scopedSandboxManager struct {
 }
 
 func (m *scopedSandboxManager) Start(ctx context.Context) error {
-	if m.mode == "off" {
+	if m.mode == config.SandboxModeOff {
 		return nil
 	}
 	if _, err := m.getOrCreateSandbox(ctx, m.defaultScopeKey()); err != nil {
@@ -361,9 +364,9 @@ func (m *scopedSandboxManager) Resolve(ctx context.Context) (Sandbox, error) {
 
 func (m *scopedSandboxManager) shouldSandbox(ctx context.Context) bool {
 	switch m.mode {
-	case "all":
+	case config.SandboxModeAll:
 		return true
-	case "non-main":
+	case config.SandboxModeNonMain:
 		// Sandbox all sessions except the agent's main session.
 		// Normalize before comparing to handle aliases like "main" or bare agent keys
 		return m.normalizeSessionKey(SessionKeyFromContext(ctx)) != m.mainSessionKey()
@@ -397,9 +400,9 @@ func (m *scopedSandboxManager) normalizeSessionKey(raw string) string {
 func (m *scopedSandboxManager) scopeKeyFromContext(ctx context.Context) string {
 	sessionKey := m.normalizeSessionKey(SessionKeyFromContext(ctx))
 	switch m.scope {
-	case "shared":
+	case config.SandboxScopeShared:
 		return "shared"
-	case "session":
+	case config.SandboxScopeSession:
 		return sessionKey
 	default:
 		if parsed := routing.ParseAgentSessionKey(sessionKey); parsed != nil {
@@ -434,7 +437,7 @@ func (m *scopedSandboxManager) getOrCreateSandbox(ctx context.Context, scopeKey
 
 func (m *scopedSandboxManager) buildScopedContainerSandbox(scopeKey string) Sandbox {
 	workspace := m.agentWorkspace
-	if m.workspaceAccess == "none" || strings.TrimSpace(workspace) == "" {
+	if m.workspaceAccess == config.WorkspaceAccessNone || strings.TrimSpace(workspace) == "" {
 		workspace = filepath.Join(m.workspaceRoot, slugScopeKey(scopeKey), "workspace")
 	}
 	return NewContainerSandbox(ContainerSandboxConfig{
@@ -443,7 +446,7 @@ func (m *scopedSandboxManager) buildScopedContainerSandbox(scopeKey string) Sand
 		ContainerPrefix: m.containerPrefix,
 		Workspace:       workspace,
 		AgentWorkspace:  m.agentWorkspace,
-		WorkspaceAccess: m.workspaceAccess,
+		WorkspaceAccess: string(m.workspaceAccess),
 		WorkspaceRoot:   m.workspaceRoot,
 		PruneIdleHours:  m.pruneIdleHours,
 		PruneMaxAgeDays: m.pruneMaxAgeDays,
@@ -494,6 +497,17 @@ func (f *managerFS) WriteFile(ctx context.Context, path string, data []byte, mkd
 	return sb.Fs().WriteFile(ctx, path, data, mkdir)
 }
 
+func (f *managerFS) ReadDir(ctx context.Context, path string) ([]os.DirEntry, error) {
+	if !f.m.shouldSandbox(ctx) {
+		return f.m.host.Fs().ReadDir(ctx, path)
+	}
+	sb, err := f.m.getOrCreateSandbox(ctx, f.m.scopeKeyFromContext(ctx))
+	if err != nil {
+		return nil, err
+	}
+	return sb.Fs().ReadDir(ctx, path)
+}
+
 var nonAlnum = regexp.MustCompile(`[^a-z0-9._-]+`)
 
 func slugScopeKey(scopeKey string) string {
@@ -513,6 +527,29 @@ func slugScopeKey(scopeKey string) string {
 	return safe + "-" + hex.EncodeToString(sum[:4])
 }
 
+// hostOnlyManager is a lightweight Manager used when sandbox mode is "off".
+// It delegates all operations directly to the HostSandbox, avoiding
+// unnecessary container infrastructure setup.
+type hostOnlyManager struct {
+	host Sandbox
+}
+
+func (h *hostOnlyManager) Start(ctx context.Context) error              { return nil }
+func (h *hostOnlyManager) Prune(ctx context.Context) error              { return h.host.Prune(ctx) }
+func (h *hostOnlyManager) Resolve(ctx context.Context) (Sandbox, error) { return h.host, nil }
+func (h *hostOnlyManager) Fs() FsBridge                                 { return h.host.Fs() }
+func (h *hostOnlyManager) Exec(ctx context.Context, req ExecRequest) (*ExecResult, error) {
+	return h.host.Exec(ctx, req)
+}
+
+func (h *hostOnlyManager) ExecStream(
+	ctx context.Context,
+	req ExecRequest,
+	onEvent func(ExecEvent) error,
+) (*ExecResult, error) {
+	return h.host.ExecStream(ctx, req, onEvent)
+}
+
 type unavailableSandboxManager struct {
 	err error
 	fs  FsBridge
@@ -562,3 +599,7 @@ func (e *errorFS) ReadFile(ctx context.Context, path string) ([]byte, error) {
 func (e *errorFS) WriteFile(ctx context.Context, path string, data []byte, mkdir bool) error {
 	return fmt.Errorf("sandbox unavailable: %w", e.err)
 }
+
+func (e *errorFS) ReadDir(ctx context.Context, path string) ([]os.DirEntry, error) {
+	return nil, fmt.Errorf("sandbox unavailable: %w", e.err)
+}

pkg/agent/sandbox/path.go

@@ -0,0 +1,78 @@
+package sandbox
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+)
+
+// ValidatePath ensures the given path is within the workspace if restrict is true.
+func ValidatePath(path, workspace string, restrict bool) (string, error) {
+	if workspace == "" {
+		return "", fmt.Errorf("workspace is not defined")
+	}
+
+	absWorkspace, err := filepath.Abs(workspace)
+	if err != nil {
+		return "", fmt.Errorf("failed to resolve workspace path: %w", err)
+	}
+
+	var absPath string
+	if filepath.IsAbs(path) {
+		absPath = filepath.Clean(path)
+	} else {
+		absPath, err = filepath.Abs(filepath.Join(absWorkspace, path))
+		if err != nil {
+			return "", fmt.Errorf("failed to resolve file path: %w", err)
+		}
+	}
+
+	if restrict {
+		if !isWithinWorkspace(absPath, absWorkspace) {
+			return "", ErrOutsideWorkspace
+		}
+
+		var resolved string
+		workspaceReal := absWorkspace
+		if resolved, err = filepath.EvalSymlinks(absWorkspace); err == nil {
+			workspaceReal = resolved
+		}
+
+		if resolved, err = filepath.EvalSymlinks(absPath); err == nil {
+			if !isWithinWorkspace(resolved, workspaceReal) {
+				return "", ErrOutsideWorkspace
+			}
+		} else if os.IsNotExist(err) {
+			var parentResolved string
+			if parentResolved, err = resolveExistingAncestor(filepath.Dir(absPath)); err == nil {
+				if !isWithinWorkspace(parentResolved, workspaceReal) {
+					return "", fmt.Errorf("access denied: symlink resolves outside workspace")
+				}
+			} else if !os.IsNotExist(err) {
+				return "", fmt.Errorf("failed to resolve path: %w", err)
+			}
+		} else {
+			return "", fmt.Errorf("failed to resolve path: %w", err)
+		}
+	}
+
+	return absPath, nil
+}
+
+func resolveExistingAncestor(path string) (string, error) {
+	for current := filepath.Clean(path); ; current = filepath.Dir(current) {
+		if resolved, err := filepath.EvalSymlinks(current); err == nil {
+			return resolved, nil
+		} else if !os.IsNotExist(err) {
+			return "", err
+		}
+		if filepath.Dir(current) == current {
+			return "", os.ErrNotExist
+		}
+	}
+}
+
+func isWithinWorkspace(candidate, workspace string) bool {
+	rel, err := filepath.Rel(filepath.Clean(workspace), filepath.Clean(candidate))
+	return err == nil && filepath.IsLocal(rel)
+}

pkg/agent/sandbox/sandbox.go

@@ -2,7 +2,10 @@ package sandbox
 
 import (
 	"context"
+	"os"
 	"strings"
+
+	"github.com/sipeed/picoclaw/pkg/logger"
 )
 
 // Sandbox abstracts command execution and filesystem access.
@@ -97,12 +100,41 @@ func WithSandbox(ctx context.Context, sb Sandbox) context.Context {
 	return context.WithValue(ctx, sandboxContextKey{}, sb)
 }
 
-// SandboxFromContext returns the sandbox instance attached by WithSandbox.
+// FromContext returns the sandbox instance attached by WithSandbox.
-func SandboxFromContext(ctx context.Context) Sandbox {
+// If no pre-resolved sandbox exists, it attempts to resolve one via the Manager in context.
+func FromContext(ctx context.Context) Sandbox {
 	if ctx == nil {
 		return nil
 	}
-	v, _ := ctx.Value(sandboxContextKey{}).(Sandbox)
+	// 1. Try pre-resolved sandbox
+	if v, ok := ctx.Value(sandboxContextKey{}).(Sandbox); ok && v != nil {
+		return v
+	}
+	// 2. Try on-demand resolution via Manager
+	if m := managerFromContext(ctx); m != nil {
+		sb, err := m.Resolve(ctx)
+		if err != nil {
+			logger.WarnCF("sandbox", "FromContext: manager.Resolve failed", map[string]any{"error": err.Error()})
+		} else if sb != nil {
+			return sb
+		}
+	}
+	return nil
+}
+
+type managerContextKey struct{}
+
+// WithManager returns a derived context carrying the sandbox manager.
+func WithManager(ctx context.Context, m Manager) context.Context {
+	return context.WithValue(ctx, managerContextKey{}, m)
+}
+
+// managerFromContext returns the manager attached by WithManager.
+func managerFromContext(ctx context.Context) Manager {
+	if ctx == nil {
+		return nil
+	}
+	v, _ := ctx.Value(managerContextKey{}).(Manager)
 	return v
 }
 
@@ -113,6 +145,8 @@ type FsBridge interface {
 	// WriteFile writes data to a sandbox-visible path.
 	// When mkdir is true, missing parent directories should be created.
 	WriteFile(ctx context.Context, path string, data []byte, mkdir bool) error
+	// ReadDir reads the named directory and returns a list of directory entries.
+	ReadDir(ctx context.Context, path string) ([]os.DirEntry, error)
 }
 
 func aggregateExecStream(execFn func(onEvent func(ExecEvent) error) (*ExecResult, error)) (*ExecResult, error) {

pkg/config/config.go

@@ -539,10 +539,37 @@ type AgentSandboxDockerConfig struct {
 	Binds           []string                                 `json:"binds"            env:"PICOCLAW_AGENTS_DEFAULTS_SANDBOX_DOCKER_BINDS"`
 }
 
+// SandboxMode defines the operational mode of the agent sandbox.
+type SandboxMode string
+
+const (
+	SandboxModeOff     SandboxMode = "off"      // Sandbox disabled (host execution)
+	SandboxModeNonMain SandboxMode = "non-main" // Sandbox all sessions except main
+	SandboxModeAll     SandboxMode = "all"      // Sandbox all sessions
+)
+
+// SandboxScope defines the isolation scope of the sandbox container.
+type SandboxScope string
+
+const (
+	SandboxScopeSession SandboxScope = "session" // One container per session
+	SandboxScopeAgent   SandboxScope = "agent"   // One container per agent (shared across sessions)
+	SandboxScopeShared  SandboxScope = "shared"  // One container shared by all agents
+)
+
+// WorkspaceAccess defines how the agent workspace is exposed to the sandbox.
+type WorkspaceAccess string
+
+const (
+	WorkspaceAccessNone WorkspaceAccess = "none" // No workspace access
+	WorkspaceAccessRO   WorkspaceAccess = "ro"   // Read-only access
+	WorkspaceAccessRW   WorkspaceAccess = "rw"   // Read-write access
+)
+
 type AgentSandboxConfig struct {
-	Mode            string                   `json:"mode"             env:"PICOCLAW_AGENTS_DEFAULTS_SANDBOX_MODE"`
+	Mode            SandboxMode              `json:"mode"             env:"PICOCLAW_AGENTS_DEFAULTS_SANDBOX_MODE"`
-	Scope           string                   `json:"scope"            env:"PICOCLAW_AGENTS_DEFAULTS_SANDBOX_SCOPE"`
+	Scope           SandboxScope             `json:"scope"            env:"PICOCLAW_AGENTS_DEFAULTS_SANDBOX_SCOPE"`
-	WorkspaceAccess string                   `json:"workspace_access" env:"PICOCLAW_AGENTS_DEFAULTS_SANDBOX_WORKSPACE_ACCESS"`
+	WorkspaceAccess WorkspaceAccess          `json:"workspace_access" env:"PICOCLAW_AGENTS_DEFAULTS_SANDBOX_WORKSPACE_ACCESS"`
 	WorkspaceRoot   string                   `json:"workspace_root"   env:"PICOCLAW_AGENTS_DEFAULTS_SANDBOX_WORKSPACE_ROOT"`
 	Docker          AgentSandboxDockerConfig `json:"docker"           env:"PICOCLAW_AGENTS_DEFAULTS_SANDBOX_DOCKER"`
 	Prune           AgentSandboxPruneConfig  `json:"prune"            env:"PICOCLAW_AGENTS_DEFAULTS_SANDBOX_PRUNE"`

pkg/config/defaults.go

@@ -18,9 +18,9 @@ func DefaultConfig() *Config {
 				Temperature:         nil, // nil means use provider default
 				MaxToolIterations:   20,
 				Sandbox: AgentSandboxConfig{
-					Mode:            "off",
+					Mode:            SandboxModeOff,
-					Scope:           "agent",
+					Scope:           SandboxScopeAgent,
-					WorkspaceAccess: "none",
+					WorkspaceAccess: WorkspaceAccessNone,
 					WorkspaceRoot:   "~/.picoclaw/sandboxes",
 					Docker: AgentSandboxDockerConfig{
 						Image:           "picoclaw-sandbox:bookworm-slim",

pkg/tools/edit.go

@@ -2,9 +2,8 @@ package tools
 
 import (
 	"context"
-	"errors"
 	"fmt"
-	"io/fs"
+	"os"
 	"strings"
 
 	"github.com/sipeed/picoclaw/pkg/agent/sandbox"
@@ -12,19 +11,13 @@ import (
 
 // EditFileTool edits a file by replacing old_text with new_text.
 // The old_text must exist exactly in the file.
-type EditFileTool struct {
+type EditFileTool struct{}
-	fs fileSystem
-}
 
-// NewEditFileTool creates a new EditFileTool with optional directory restriction.
+// NewEditFileTool creates a new EditFileTool.
+// The workspace and restrict parameters are kept for compatibility with tool registry signatures
+// but are no longer used since filesystem access is entirely delegated to the Sandbox context.
 func NewEditFileTool(workspace string, restrict bool) *EditFileTool {
-	var fs fileSystem
+	return &EditFileTool{}
-	if restrict {
-		fs = &sandboxFs{workspace: workspace}
-	} else {
-		fs = &hostFs{}
-	}
-	return &EditFileTool{fs: fs}
 }
 
 func (t *EditFileTool) Name() string {
@@ -72,51 +65,35 @@ func (t *EditFileTool) Execute(ctx context.Context, args map[string]any) *ToolRe
 		return ErrorResult("new_text is required")
 	}
 
-	sb := sandbox.SandboxFromContext(ctx)
+	sb := sandbox.FromContext(ctx)
-	if sb != nil {
+	if sb == nil {
+		return ErrorResult("sandbox environment unavailable")
+	}
+
 	content, err := sb.Fs().ReadFile(ctx, path)
 	if err != nil {
 		return ErrorResult(fmt.Sprintf("failed to read file: %v", err))
 	}
 
-		contentStr := string(content)
+	newContent, err := replaceEditContent(content, oldText, newText)
-		if !strings.Contains(contentStr, oldText) {
+	if err != nil {
-			return ErrorResult("old_text not found in file. Make sure it matches exactly")
+		return ErrorResult(err.Error())
 	}
 
-		count := strings.Count(contentStr, oldText)
+	err = sb.Fs().WriteFile(ctx, path, newContent, true)
-		if count > 1 {
-			return ErrorResult(
-				fmt.Sprintf("old_text appears %d times. Please provide more context to make it unique", count),
-			)
-		}
-
-		newContent := strings.Replace(contentStr, oldText, newText, 1)
-		err = sb.Fs().WriteFile(ctx, path, []byte(newContent), true)
 	if err != nil {
 		return ErrorResult(fmt.Sprintf("failed to write file: %v", err))
 	}
 	return SilentResult(fmt.Sprintf("File edited: %s", path))
-	}
-
-	if err := editFile(t.fs, path, oldText, newText); err != nil {
-		return ErrorResult(err.Error())
-	}
-	return SilentResult(fmt.Sprintf("File edited: %s", path))
 }
 
-type AppendFileTool struct {
+type AppendFileTool struct{}
-	fs fileSystem
-}
 
+// NewAppendFileTool creates a new AppendFileTool.
+// The workspace and restrict parameters are kept for compatibility with tool registry signatures
+// but are no longer used since filesystem access is entirely delegated to the Sandbox context.
 func NewAppendFileTool(workspace string, restrict bool) *AppendFileTool {
-	var fs fileSystem
+	return &AppendFileTool{}
-	if restrict {
-		fs = &sandboxFs{workspace: workspace}
-	} else {
-		fs = &hostFs{}
-	}
-	return &AppendFileTool{fs: fs}
 }
 
 func (t *AppendFileTool) Name() string {
@@ -155,52 +132,22 @@ func (t *AppendFileTool) Execute(ctx context.Context, args map[string]any) *Tool
 		return ErrorResult("content is required")
 	}
 
-	sb := sandbox.SandboxFromContext(ctx)
+	sb := sandbox.FromContext(ctx)
-	if sb != nil {
+	if sb == nil {
-		// Implement Append using Read + Write if no Append in FsBridge
+		return ErrorResult("sandbox environment unavailable")
+	}
+
+	// Implement Append using Read + Write
 	oldContent, err := sb.Fs().ReadFile(ctx, path)
-		if err != nil && !strings.Contains(err.Error(), "no such file") {
+	if err != nil && !os.IsNotExist(err) && !strings.Contains(err.Error(), "not found") {
 		return ErrorResult(fmt.Sprintf("failed to read file for append: %v", err))
 	}
-		newContent := string(oldContent) + content
+	newContent := append(oldContent, []byte(content)...)
-		err = sb.Fs().WriteFile(ctx, path, []byte(newContent), true)
+	err = sb.Fs().WriteFile(ctx, path, newContent, true)
 	if err != nil {
 		return ErrorResult(fmt.Sprintf("failed to append (write) to file: %v", err))
 	}
 	return SilentResult(fmt.Sprintf("Appended to %s", path))
-	}
-
-	if err := appendFile(t.fs, path, content); err != nil {
-		return ErrorResult(err.Error())
-	}
-	return SilentResult(fmt.Sprintf("Appended to %s", path))
-}
-
-// editFile reads the file via sysFs, performs the replacement, and writes back.
-// It uses a fileSystem interface, allowing the same logic for both restricted and unrestricted modes.
-func editFile(sysFs fileSystem, path, oldText, newText string) error {
-	content, err := sysFs.ReadFile(path)
-	if err != nil {
-		return err
-	}
-
-	newContent, err := replaceEditContent(content, oldText, newText)
-	if err != nil {
-		return err
-	}
-
-	return sysFs.WriteFile(path, newContent)
-}
-
-// appendFile reads the existing content (if any) via sysFs, appends new content, and writes back.
-func appendFile(sysFs fileSystem, path, appendContent string) error {
-	content, err := sysFs.ReadFile(path)
-	if err != nil && !errors.Is(err, fs.ErrNotExist) {
-		return err
-	}
-
-	newContent := append(content, []byte(appendContent)...)
-	return sysFs.WriteFile(path, newContent)
 }
 
 // replaceEditContent handles the core logic of finding and replacing a single occurrence of oldText.

pkg/tools/edit_test.go

@@ -8,6 +8,8 @@ import (
 	"testing"
 
 	"github.com/stretchr/testify/assert"
+
+	"github.com/sipeed/picoclaw/pkg/agent/sandbox"
 )
 
 // TestEditTool_EditFile_Success verifies successful file editing
@@ -17,7 +19,8 @@ func TestEditTool_EditFile_Success(t *testing.T) {
 	os.WriteFile(testFile, []byte("Hello World\nThis is a test"), 0o644)
 
 	tool := NewEditFileTool(tmpDir, true)
-	ctx := context.Background()
+	sb := sandbox.NewHostSandbox(tmpDir, true)
+	ctx := sandbox.WithSandbox(context.Background(), sb)
 	args := map[string]any{
 		"path":     testFile,
 		"old_text": "World",
@@ -61,7 +64,8 @@ func TestEditTool_EditFile_NotFound(t *testing.T) {
 	testFile := filepath.Join(tmpDir, "nonexistent.txt")
 
 	tool := NewEditFileTool(tmpDir, true)
-	ctx := context.Background()
+	sb := sandbox.NewHostSandbox(tmpDir, true)
+	ctx := sandbox.WithSandbox(context.Background(), sb)
 	args := map[string]any{
 		"path":     testFile,
 		"old_text": "old",
@@ -75,10 +79,9 @@ func TestEditTool_EditFile_NotFound(t *testing.T) {
 		t.Errorf("Expected error for non-existent file")
 	}
 
-	// Should mention file not found
+	// Should mention file not found or no such file
-	if !strings.Contains(result.ForLLM, "not found") && !strings.Contains(result.ForUser, "not found") {
+	assert.True(t, strings.Contains(result.ForLLM, "not found") || strings.Contains(result.ForLLM, "no such file"),
-		t.Errorf("Expected 'file not found' message, got ForLLM: %s", result.ForLLM)
+		"Expected 'not found' or 'no such file' message, got ForLLM: %s", result.ForLLM)
-	}
 }
 
 // TestEditTool_EditFile_OldTextNotFound verifies error when old_text doesn't exist
@@ -88,7 +91,8 @@ func TestEditTool_EditFile_OldTextNotFound(t *testing.T) {
 	os.WriteFile(testFile, []byte("Hello World"), 0o644)
 
 	tool := NewEditFileTool(tmpDir, true)
-	ctx := context.Background()
+	sb := sandbox.NewHostSandbox(tmpDir, true)
+	ctx := sandbox.WithSandbox(context.Background(), sb)
 	args := map[string]any{
 		"path":     testFile,
 		"old_text": "Goodbye",
@@ -115,7 +119,8 @@ func TestEditTool_EditFile_MultipleMatches(t *testing.T) {
 	os.WriteFile(testFile, []byte("test test test"), 0o644)
 
 	tool := NewEditFileTool(tmpDir, true)
-	ctx := context.Background()
+	sb := sandbox.NewHostSandbox(tmpDir, true)
+	ctx := sandbox.WithSandbox(context.Background(), sb)
 	args := map[string]any{
 		"path":     testFile,
 		"old_text": "test",
@@ -143,7 +148,8 @@ func TestEditTool_EditFile_OutsideAllowedDir(t *testing.T) {
 	os.WriteFile(testFile, []byte("content"), 0o644)
 
 	tool := NewEditFileTool(tmpDir, true) // Restrict to tmpDir
-	ctx := context.Background()
+	sb := sandbox.NewHostSandbox(tmpDir, true)
+	ctx := sandbox.WithSandbox(context.Background(), sb)
 	args := map[string]any{
 		"path":     testFile,
 		"old_text": "content",
@@ -169,8 +175,9 @@ func TestEditTool_EditFile_OutsideAllowedDir(t *testing.T) {
 
 // TestEditTool_EditFile_MissingPath verifies error handling for missing path
 func TestEditTool_EditFile_MissingPath(t *testing.T) {
-	tool := NewEditFileTool("", false)
+	tool := NewEditFileTool("/", false)
-	ctx := context.Background()
+	sb := sandbox.NewHostSandbox("/", false)
+	ctx := sandbox.WithSandbox(context.Background(), sb)
 	args := map[string]any{
 		"old_text": "old",
 		"new_text": "new",
@@ -186,8 +193,9 @@ func TestEditTool_EditFile_MissingPath(t *testing.T) {
 
 // TestEditTool_EditFile_MissingOldText verifies error handling for missing old_text
 func TestEditTool_EditFile_MissingOldText(t *testing.T) {
-	tool := NewEditFileTool("", false)
+	tool := NewEditFileTool("/", false)
-	ctx := context.Background()
+	sb := sandbox.NewHostSandbox("/", false)
+	ctx := sandbox.WithSandbox(context.Background(), sb)
 	args := map[string]any{
 		"path":     "/tmp/test.txt",
 		"new_text": "new",
@@ -203,8 +211,9 @@ func TestEditTool_EditFile_MissingOldText(t *testing.T) {
 
 // TestEditTool_EditFile_MissingNewText verifies error handling for missing new_text
 func TestEditTool_EditFile_MissingNewText(t *testing.T) {
-	tool := NewEditFileTool("", false)
+	tool := NewEditFileTool("/", false)
-	ctx := context.Background()
+	sb := sandbox.NewHostSandbox("/", false)
+	ctx := sandbox.WithSandbox(context.Background(), sb)
 	args := map[string]any{
 		"path":     "/tmp/test.txt",
 		"old_text": "old",
@@ -224,8 +233,9 @@ func TestEditTool_AppendFile_Success(t *testing.T) {
 	testFile := filepath.Join(tmpDir, "test.txt")
 	os.WriteFile(testFile, []byte("Initial content"), 0o644)
 
-	tool := NewAppendFileTool("", false)
+	tool := NewAppendFileTool("/", false)
-	ctx := context.Background()
+	sb := sandbox.NewHostSandbox("/", false)
+	ctx := sandbox.WithSandbox(context.Background(), sb)
 	args := map[string]any{
 		"path":    testFile,
 		"content": "\nAppended content",
@@ -264,7 +274,7 @@ func TestEditTool_AppendFile_Success(t *testing.T) {
 
 // TestEditTool_AppendFile_MissingPath verifies error handling for missing path
 func TestEditTool_AppendFile_MissingPath(t *testing.T) {
-	tool := NewAppendFileTool("", false)
+	tool := NewAppendFileTool("/", false)
 	ctx := context.Background()
 	args := map[string]any{
 		"content": "test",
@@ -280,7 +290,7 @@ func TestEditTool_AppendFile_MissingPath(t *testing.T) {
 
 // TestEditTool_AppendFile_MissingContent verifies error handling for missing content
 func TestEditTool_AppendFile_MissingContent(t *testing.T) {
-	tool := NewAppendFileTool("", false)
+	tool := NewAppendFileTool("/", false)
 	ctx := context.Background()
 	args := map[string]any{
 		"path": "/tmp/test.txt",
@@ -349,7 +359,8 @@ func TestReplaceEditContent(t *testing.T) {
 func TestAppendFileTool_AppendToNonExistent_Restricted(t *testing.T) {
 	workspace := t.TempDir()
 	tool := NewAppendFileTool(workspace, true)
-	ctx := context.Background()
+	sb := sandbox.NewHostSandbox(workspace, true)
+	ctx := sandbox.WithSandbox(context.Background(), sb)
 
 	args := map[string]any{
 		"path":    "brand_new_file.txt",
@@ -379,7 +390,8 @@ func TestAppendFileTool_Restricted_Success(t *testing.T) {
 	assert.NoError(t, err)
 
 	tool := NewAppendFileTool(workspace, true)
-	ctx := context.Background()
+	sb := sandbox.NewHostSandbox(workspace, true)
+	ctx := sandbox.WithSandbox(context.Background(), sb)
 	args := map[string]any{
 		"path":    testFile,
 		"content": " appended",
@@ -403,7 +415,8 @@ func TestEditFileTool_Restricted_InPlaceEdit(t *testing.T) {
 	assert.NoError(t, err)
 
 	tool := NewEditFileTool(workspace, true)
-	ctx := context.Background()
+	sb := sandbox.NewHostSandbox(workspace, true)
+	ctx := sandbox.WithSandbox(context.Background(), sb)
 	args := map[string]any{
 		"path":     testFile,
 		"old_text": "World",
@@ -424,7 +437,8 @@ func TestEditFileTool_Restricted_InPlaceEdit(t *testing.T) {
 func TestEditFileTool_Restricted_FileNotFound(t *testing.T) {
 	workspace := t.TempDir()
 	tool := NewEditFileTool(workspace, true)
-	ctx := context.Background()
+	sb := sandbox.NewHostSandbox(workspace, true)
+	ctx := sandbox.WithSandbox(context.Background(), sb)
 	args := map[string]any{
 		"path":     "no_such_file.txt",
 		"old_text": "old",
@@ -433,5 +447,6 @@ func TestEditFileTool_Restricted_FileNotFound(t *testing.T) {
 
 	result := tool.Execute(ctx, args)
 	assert.True(t, result.IsError)
-	assert.Contains(t, result.ForLLM, "not found")
+	assert.True(t, strings.Contains(result.ForLLM, "not found") || strings.Contains(result.ForLLM, "no such file"),
+		"Expected 'not found' or 'no such file' message, got ForLLM: %s", result.ForLLM)
 }

pkg/tools/filesystem.go

@@ -3,98 +3,19 @@ package tools
 import (
 	"context"
 	"fmt"
-	"io/fs"
 	"os"
-	"path/filepath"
 	"strings"
-	"time"
 
 	"github.com/sipeed/picoclaw/pkg/agent/sandbox"
 )
 
-// validatePath ensures the given path is within the workspace if restrict is true.
+type ReadFileTool struct{}
-func validatePath(path, workspace string, restrict bool) (string, error) {
-	if workspace == "" {
-		return path, fmt.Errorf("workspace is not defined")
-	}
-
-	absWorkspace, err := filepath.Abs(workspace)
-	if err != nil {
-		return "", fmt.Errorf("failed to resolve workspace path: %w", err)
-	}
-
-	var absPath string
-	if filepath.IsAbs(path) {
-		absPath = filepath.Clean(path)
-	} else {
-		absPath, err = filepath.Abs(filepath.Join(absWorkspace, path))
-		if err != nil {
-			return "", fmt.Errorf("failed to resolve file path: %w", err)
-		}
-	}
-
-	if restrict {
-		if !isWithinWorkspace(absPath, absWorkspace) {
-			return "", fmt.Errorf("access denied: path is outside the workspace")
-		}
-
-		var resolved string
-		workspaceReal := absWorkspace
-		if resolved, err = filepath.EvalSymlinks(absWorkspace); err == nil {
-			workspaceReal = resolved
-		}
-
-		if resolved, err = filepath.EvalSymlinks(absPath); err == nil {
-			if !isWithinWorkspace(resolved, workspaceReal) {
-				return "", fmt.Errorf("access denied: symlink resolves outside workspace")
-			}
-		} else if os.IsNotExist(err) {
-			var parentResolved string
-			if parentResolved, err = resolveExistingAncestor(filepath.Dir(absPath)); err == nil {
-				if !isWithinWorkspace(parentResolved, workspaceReal) {
-					return "", fmt.Errorf("access denied: symlink resolves outside workspace")
-				}
-			} else if !os.IsNotExist(err) {
-				return "", fmt.Errorf("failed to resolve path: %w", err)
-			}
-		} else {
-			return "", fmt.Errorf("failed to resolve path: %w", err)
-		}
-	}
-
-	return absPath, nil
-}
-
-func resolveExistingAncestor(path string) (string, error) {
-	for current := filepath.Clean(path); ; current = filepath.Dir(current) {
-		if resolved, err := filepath.EvalSymlinks(current); err == nil {
-			return resolved, nil
-		} else if !os.IsNotExist(err) {
-			return "", err
-		}
-		if filepath.Dir(current) == current {
-			return "", os.ErrNotExist
-		}
-	}
-}
-
-func isWithinWorkspace(candidate, workspace string) bool {
-	rel, err := filepath.Rel(filepath.Clean(workspace), filepath.Clean(candidate))
-	return err == nil && filepath.IsLocal(rel)
-}
-
-type ReadFileTool struct {
-	fs fileSystem
-}
 
+// NewReadFileTool creates a new ReadFileTool.
+// The workspace and restrict parameters are kept for compatibility with tool registry signatures
+// but are no longer used since filesystem access is entirely delegated to the Sandbox context.
 func NewReadFileTool(workspace string, restrict bool) *ReadFileTool {
-	var fs fileSystem
+	return &ReadFileTool{}
-	if restrict {
-		fs = &sandboxFs{workspace: workspace}
-	} else {
-		fs = &hostFs{}
-	}
-	return &ReadFileTool{fs: fs}
 }
 
 func (t *ReadFileTool) Name() string {
@@ -124,34 +45,22 @@ func (t *ReadFileTool) Execute(ctx context.Context, args map[string]any) *ToolRe
 		return ErrorResult("path is required")
 	}
 
-	sb := sandbox.SandboxFromContext(ctx)
+	sb := sandbox.FromContext(ctx)
-	if sb != nil {
+	if sb == nil {
+		return ErrorResult("sandbox environment unavailable")
+	}
+
 	content, err := sb.Fs().ReadFile(ctx, path)
 	if err != nil {
-			return ErrorResult(fmt.Sprintf("failed to read file from sandbox: %v", err))
+		return ErrorResult(fmt.Sprintf("failed to read file: %v", err))
-		}
-		return NewToolResult(string(content))
-	}
-
-	content, err := t.fs.ReadFile(path)
-	if err != nil {
-		return ErrorResult(err.Error())
 	}
 	return NewToolResult(string(content))
 }
 
-type WriteFileTool struct {
+type WriteFileTool struct{}
-	fs fileSystem
-}
 
 func NewWriteFileTool(workspace string, restrict bool) *WriteFileTool {
-	var fs fileSystem
+	return &WriteFileTool{}
-	if restrict {
-		fs = &sandboxFs{workspace: workspace}
-	} else {
-		fs = &hostFs{}
-	}
-	return &WriteFileTool{fs: fs}
 }
 
 func (t *WriteFileTool) Name() string {
@@ -190,35 +99,21 @@ func (t *WriteFileTool) Execute(ctx context.Context, args map[string]any) *ToolR
 		return ErrorResult("content is required")
 	}
 
-	sb := sandbox.SandboxFromContext(ctx)
+	sb := sandbox.FromContext(ctx)
-	if sb != nil {
+	if sb == nil {
+		return ErrorResult("sandbox environment unavailable")
+	}
+
 	if err := sb.Fs().WriteFile(ctx, path, []byte(content), true); err != nil {
-			return ErrorResult(fmt.Sprintf("failed to write file to sandbox: %v", err))
+		return ErrorResult(fmt.Sprintf("failed to write file: %v", err))
 	}
-		return SilentResult(fmt.Sprintf("File written to sandbox: %s", path))
-	}
-
-	if err := t.fs.WriteFile(path, []byte(content)); err != nil {
-		return ErrorResult(err.Error())
-	}
-
 	return SilentResult(fmt.Sprintf("File written: %s", path))
 }
 
-// ListDirTool lists files and directories at a given path.
+type ListDirTool struct{}
-// Phase 1: host-only execution; sandbox routing is deferred to Phase 2 (see sandbox.md §6.2).
-type ListDirTool struct {
-	fs fileSystem
-}
 
 func NewListDirTool(workspace string, restrict bool) *ListDirTool {
-	var fs fileSystem
+	return &ListDirTool{}
-	if restrict {
-		fs = &sandboxFs{workspace: workspace}
-	} else {
-		fs = &hostFs{}
-	}
-	return &ListDirTool{fs: fs}
 }
 
 func (t *ListDirTool) Name() string {
@@ -248,7 +143,12 @@ func (t *ListDirTool) Execute(ctx context.Context, args map[string]any) *ToolRes
 		path = "."
 	}
 
-	entries, err := t.fs.ReadDir(path)
+	sb := sandbox.FromContext(ctx)
+	if sb == nil {
+		return ErrorResult("sandbox environment unavailable")
+	}
+
+	entries, err := sb.Fs().ReadDir(ctx, path)
 	if err != nil {
 		return ErrorResult(fmt.Sprintf("failed to read directory: %v", err))
 	}
@@ -266,161 +166,3 @@ func formatDirEntries(entries []os.DirEntry) *ToolResult {
 	}
 	return NewToolResult(result.String())
 }
-
-// fileSystem abstracts reading, writing, and listing files, allowing both
-// unrestricted (host filesystem) and sandbox (os.Root) implementations to share the same polymorphic interface.
-type fileSystem interface {
-	ReadFile(path string) ([]byte, error)
-	WriteFile(path string, data []byte) error
-	ReadDir(path string) ([]os.DirEntry, error)
-}
-
-// hostFs is an unrestricted fileReadWriter that operates directly on the host filesystem.
-type hostFs struct{}
-
-func (h *hostFs) ReadFile(path string) ([]byte, error) {
-	content, err := os.ReadFile(path)
-	if err != nil {
-		if os.IsNotExist(err) {
-			return nil, fmt.Errorf("failed to read file: file not found: %w", err)
-		}
-		if os.IsPermission(err) {
-			return nil, fmt.Errorf("failed to read file: access denied: %w", err)
-		}
-		return nil, fmt.Errorf("failed to read file: %w", err)
-	}
-	return content, nil
-}
-
-func (h *hostFs) ReadDir(path string) ([]os.DirEntry, error) {
-	return os.ReadDir(path)
-}
-
-func (h *hostFs) WriteFile(path string, data []byte) error {
-	dir := filepath.Dir(path)
-	if err := os.MkdirAll(dir, 0o755); err != nil {
-		return fmt.Errorf("failed to create parent directories: %w", err)
-	}
-
-	// We use a "write-then-rename" pattern here to ensure an atomic write.
-	// This prevents the target file from being left in a truncated or partial state
-	// if the operation is interrupted, as the rename operation is atomic on Linux.
-	tmpPath := fmt.Sprintf("%s.%d.tmp", path, time.Now().UnixNano())
-	if err := os.WriteFile(tmpPath, data, 0o644); err != nil {
-		os.Remove(tmpPath) // Ensure cleanup of partial/empty temp file
-		return fmt.Errorf("failed to write temp file: %w", err)
-	}
-
-	if err := os.Rename(tmpPath, path); err != nil {
-		os.Remove(tmpPath)
-		return fmt.Errorf("failed to replace original file: %w", err)
-	}
-	return nil
-}
-
-// sandboxFs is a sandboxed fileSystem that operates within a strictly defined workspace using os.Root.
-type sandboxFs struct {
-	workspace string
-}
-
-func (r *sandboxFs) execute(path string, fn func(root *os.Root, relPath string) error) error {
-	if r.workspace == "" {
-		return fmt.Errorf("workspace is not defined")
-	}
-
-	root, err := os.OpenRoot(r.workspace)
-	if err != nil {
-		return fmt.Errorf("failed to open workspace: %w", err)
-	}
-	defer root.Close()
-
-	relPath, err := getSafeRelPath(r.workspace, path)
-	if err != nil {
-		return err
-	}
-
-	return fn(root, relPath)
-}
-
-func (r *sandboxFs) ReadFile(path string) ([]byte, error) {
-	var content []byte
-	err := r.execute(path, func(root *os.Root, relPath string) error {
-		fileContent, err := root.ReadFile(relPath)
-		if err != nil {
-			if os.IsNotExist(err) {
-				return fmt.Errorf("failed to read file: file not found: %w", err)
-			}
-			// os.Root returns "escapes from parent" for paths outside the root
-			if os.IsPermission(err) || strings.Contains(err.Error(), "escapes from parent") ||
-				strings.Contains(err.Error(), "permission denied") {
-				return fmt.Errorf("failed to read file: access denied: %w", err)
-			}
-			return fmt.Errorf("failed to read file: %w", err)
-		}
-		content = fileContent
-		return nil
-	})
-	return content, err
-}
-
-func (r *sandboxFs) WriteFile(path string, data []byte) error {
-	return r.execute(path, func(root *os.Root, relPath string) error {
-		dir := filepath.Dir(relPath)
-		if dir != "." && dir != "/" {
-			if err := root.MkdirAll(dir, 0o755); err != nil {
-				return fmt.Errorf("failed to create parent directories: %w", err)
-			}
-		}
-
-		// We use a "write-then-rename" pattern here to ensure an atomic write.
-		// This prevents the target file from being left in a truncated or partial state
-		// if the operation is interrupted, as the rename operation is atomic on Linux.
-		tmpRelPath := fmt.Sprintf("%s.%d.tmp", relPath, time.Now().UnixNano())
-
-		if err := root.WriteFile(tmpRelPath, data, 0o644); err != nil {
-			root.Remove(tmpRelPath) // Ensure cleanup of partial/empty temp file
-			return fmt.Errorf("failed to write to temp file: %w", err)
-		}
-
-		if err := root.Rename(tmpRelPath, relPath); err != nil {
-			root.Remove(tmpRelPath)
-			return fmt.Errorf("failed to rename temp file over target: %w", err)
-		}
-		return nil
-	})
-}
-
-func (r *sandboxFs) ReadDir(path string) ([]os.DirEntry, error) {
-	var entries []os.DirEntry
-	err := r.execute(path, func(root *os.Root, relPath string) error {
-		dirEntries, err := fs.ReadDir(root.FS(), relPath)
-		if err != nil {
-			return err
-		}
-		entries = dirEntries
-		return nil
-	})
-	return entries, err
-}
-
-// Helper to get a safe relative path for os.Root usage
-func getSafeRelPath(workspace, path string) (string, error) {
-	if workspace == "" {
-		return "", fmt.Errorf("workspace is not defined")
-	}
-
-	rel := filepath.Clean(path)
-	if filepath.IsAbs(rel) {
-		var err error
-		rel, err = filepath.Rel(workspace, rel)
-		if err != nil {
-			return "", fmt.Errorf("failed to calculate relative path: %w", err)
-		}
-	}
-
-	if !filepath.IsLocal(rel) {
-		return "", fmt.Errorf("path escapes workspace: %s", path)
-	}
-
-	return rel, nil
-}

pkg/tools/filesystem_test.go

@@ -2,13 +2,15 @@ package tools
 
 import (
 	"context"
-	"io"
+	"fmt"
 	"os"
 	"path/filepath"
 	"strings"
 	"testing"
 
 	"github.com/stretchr/testify/assert"
+
+	"github.com/sipeed/picoclaw/pkg/agent/sandbox"
 )
 
 // TestFilesystemTool_ReadFile_Success verifies successful file reading
@@ -18,7 +20,11 @@ func TestFilesystemTool_ReadFile_Success(t *testing.T) {
 	os.WriteFile(testFile, []byte("test content"), 0o644)
 
 	tool := NewReadFileTool("", false)
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{
+		fs: sandbox.NewHostSandbox(tmpDir, false).Fs(),
+	})
+	// We must ensure the mock FsBridge can actually read the TempDir.
+	// but stubSandbox uses hostFs internally for Fs(). ReadFile so this just works if injected.
 	args := map[string]any{
 		"path": testFile,
 	}
@@ -45,7 +51,9 @@ func TestFilesystemTool_ReadFile_Success(t *testing.T) {
 // TestFilesystemTool_ReadFile_NotFound verifies error handling for missing file
 func TestFilesystemTool_ReadFile_NotFound(t *testing.T) {
 	tool := NewReadFileTool("", false)
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{
+		err: fmt.Errorf("failed to read file: file not found"),
+	})
 	args := map[string]any{
 		"path": "/nonexistent_file_12345.txt",
 	}
@@ -66,7 +74,7 @@ func TestFilesystemTool_ReadFile_NotFound(t *testing.T) {
 // TestFilesystemTool_ReadFile_MissingPath verifies error handling for missing path
 func TestFilesystemTool_ReadFile_MissingPath(t *testing.T) {
 	tool := &ReadFileTool{}
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{})
 	args := map[string]any{}
 
 	result := tool.Execute(ctx, args)
@@ -88,7 +96,9 @@ func TestFilesystemTool_WriteFile_Success(t *testing.T) {
 	testFile := filepath.Join(tmpDir, "newfile.txt")
 
 	tool := NewWriteFileTool("", false)
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{
+		fs: sandbox.NewHostSandbox(tmpDir, false).Fs(),
+	})
 	args := map[string]any{
 		"path":    testFile,
 		"content": "hello world",
@@ -127,7 +137,9 @@ func TestFilesystemTool_WriteFile_CreateDir(t *testing.T) {
 	testFile := filepath.Join(tmpDir, "subdir", "newfile.txt")
 
 	tool := NewWriteFileTool("", false)
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{
+		fs: sandbox.NewHostSandbox(tmpDir, false).Fs(),
+	})
 	args := map[string]any{
 		"path":    testFile,
 		"content": "test",
@@ -153,7 +165,7 @@ func TestFilesystemTool_WriteFile_CreateDir(t *testing.T) {
 // TestFilesystemTool_WriteFile_MissingPath verifies error handling for missing path
 func TestFilesystemTool_WriteFile_MissingPath(t *testing.T) {
 	tool := NewWriteFileTool("", false)
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{})
 	args := map[string]any{
 		"content": "test",
 	}
@@ -169,7 +181,7 @@ func TestFilesystemTool_WriteFile_MissingPath(t *testing.T) {
 // TestFilesystemTool_WriteFile_MissingContent verifies error handling for missing content
 func TestFilesystemTool_WriteFile_MissingContent(t *testing.T) {
 	tool := NewWriteFileTool("", false)
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{})
 	args := map[string]any{
 		"path": "/tmp/test.txt",
 	}
@@ -195,8 +207,10 @@ func TestFilesystemTool_ListDir_Success(t *testing.T) {
 	os.WriteFile(filepath.Join(tmpDir, "file2.txt"), []byte("content"), 0o644)
 	os.Mkdir(filepath.Join(tmpDir, "subdir"), 0o755)
 
-	tool := NewListDirTool("", false)
+	tool := NewListDirTool(tmpDir, false)
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{
+		fs: sandbox.NewHostSandbox(tmpDir, false).Fs(),
+	})
 	args := map[string]any{
 		"path": tmpDir,
 	}
@@ -219,8 +233,12 @@ func TestFilesystemTool_ListDir_Success(t *testing.T) {
 
 // TestFilesystemTool_ListDir_NotFound verifies error handling for non-existent directory
 func TestFilesystemTool_ListDir_NotFound(t *testing.T) {
-	tool := NewListDirTool("", false)
+	tmpDir := t.TempDir()
-	ctx := context.Background()
+	tool := NewListDirTool(tmpDir, false)
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{
+		fs:  sandbox.NewHostSandbox(tmpDir, false).Fs(),
+		err: fmt.Errorf("failed to read directory: file not found"),
+	})
 	args := map[string]any{
 		"path": "/nonexistent_directory_12345",
 	}
@@ -240,8 +258,10 @@ func TestFilesystemTool_ListDir_NotFound(t *testing.T) {
 
 // TestFilesystemTool_ListDir_DefaultPath verifies default to current directory
 func TestFilesystemTool_ListDir_DefaultPath(t *testing.T) {
-	tool := NewListDirTool("", false)
+	tool := NewListDirTool(".", false)
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{
+		fs: sandbox.NewHostSandbox(".", false).Fs(),
+	})
 	args := map[string]any{}
 
 	result := tool.Execute(ctx, args)
@@ -271,7 +291,9 @@ func TestFilesystemTool_ReadFile_RejectsSymlinkEscape(t *testing.T) {
 	}
 
 	tool := NewReadFileTool(workspace, true)
-	result := tool.Execute(context.Background(), map[string]any{
+	result := tool.Execute(sandbox.WithSandbox(context.Background(), &stubSandbox{
+		fs: sandbox.NewHostSandbox(workspace, true).Fs(),
+	}), map[string]any{
 		"path": link,
 	})
 
@@ -295,7 +317,9 @@ func TestFilesystemTool_EmptyWorkspace_AccessDenied(t *testing.T) {
 	secretFile := filepath.Join(tmpDir, "shadow")
 	os.WriteFile(secretFile, []byte("secret data"), 0o600)
 
-	result := tool.Execute(context.Background(), map[string]any{
+	result := tool.Execute(sandbox.WithSandbox(context.Background(), &stubSandbox{
+		err: fmt.Errorf("workspace is not defined"),
+	}), map[string]any{
 		"path": secretFile,
 	})
 
@@ -342,7 +366,9 @@ func TestRootMkdirAll(t *testing.T) {
 func TestFilesystemTool_WriteFile_Restricted_CreateDir(t *testing.T) {
 	workspace := t.TempDir()
 	tool := NewWriteFileTool(workspace, true)
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{
+		fs: sandbox.NewHostSandbox(workspace, true).Fs(),
+	})
 
 	testFile := "deep/nested/path/to/file.txt"
 	content := "deep content"
@@ -360,129 +386,3 @@ func TestFilesystemTool_WriteFile_Restricted_CreateDir(t *testing.T) {
 	assert.NoError(t, err)
 	assert.Equal(t, content, string(data))
 }
-
-// TestHostRW_Read_PermissionDenied verifies that hostRW.Read surfaces access denied errors.
-func TestHostRW_Read_PermissionDenied(t *testing.T) {
-	if os.Getuid() == 0 {
-		t.Skip("skipping permission test: running as root")
-	}
-	tmpDir := t.TempDir()
-	protected := filepath.Join(tmpDir, "protected.txt")
-	err := os.WriteFile(protected, []byte("secret"), 0o000)
-	assert.NoError(t, err)
-	defer os.Chmod(protected, 0o644) // ensure cleanup
-
-	_, err = (&hostFs{}).ReadFile(protected)
-	assert.Error(t, err)
-	assert.Contains(t, err.Error(), "access denied")
-}
-
-// TestHostRW_Read_Directory verifies that hostRW.Read returns an error when given a directory path.
-func TestHostRW_Read_Directory(t *testing.T) {
-	tmpDir := t.TempDir()
-
-	_, err := (&hostFs{}).ReadFile(tmpDir)
-	assert.Error(t, err, "expected error when reading a directory as a file")
-}
-
-// TestRootRW_Read_Directory verifies that rootRW.Read returns an error when given a directory.
-func TestRootRW_Read_Directory(t *testing.T) {
-	workspace := t.TempDir()
-	root, err := os.OpenRoot(workspace)
-	assert.NoError(t, err)
-	defer root.Close()
-
-	// Create a subdirectory
-	err = root.Mkdir("subdir", 0o755)
-	assert.NoError(t, err)
-
-	_, err = (&sandboxFs{workspace: workspace}).ReadFile("subdir")
-	assert.Error(t, err, "expected error when reading a directory as a file")
-}
-
-// TestHostRW_Write_ParentDirMissing verifies that hostRW.Write creates parent dirs automatically.
-func TestHostRW_Write_ParentDirMissing(t *testing.T) {
-	tmpDir := t.TempDir()
-	target := filepath.Join(tmpDir, "a", "b", "c", "file.txt")
-
-	err := (&hostFs{}).WriteFile(target, []byte("hello"))
-	assert.NoError(t, err)
-
-	data, err := os.ReadFile(target)
-	assert.NoError(t, err)
-	assert.Equal(t, "hello", string(data))
-}
-
-// TestRootRW_Write_ParentDirMissing verifies that rootRW.Write creates
-// nested parent directories automatically within the sandbox.
-func TestRootRW_Write_ParentDirMissing(t *testing.T) {
-	workspace := t.TempDir()
-
-	relPath := "x/y/z/file.txt"
-	err := (&sandboxFs{workspace: workspace}).WriteFile(relPath, []byte("nested"))
-	assert.NoError(t, err)
-
-	data, err := os.ReadFile(filepath.Join(workspace, relPath))
-	assert.NoError(t, err)
-	assert.Equal(t, "nested", string(data))
-}
-
-// TestHostRW_Write verifies the hostRW.Write helper function
-func TestHostRW_Write(t *testing.T) {
-	tmpDir := t.TempDir()
-	testFile := filepath.Join(tmpDir, "atomic_test.txt")
-	testData := []byte("atomic test content")
-
-	err := (&hostFs{}).WriteFile(testFile, testData)
-	assert.NoError(t, err)
-
-	content, err := os.ReadFile(testFile)
-	assert.NoError(t, err)
-	assert.Equal(t, testData, content)
-
-	// Verify it overwrites correctly
-	newData := []byte("new atomic content")
-	err = (&hostFs{}).WriteFile(testFile, newData)
-	assert.NoError(t, err)
-
-	content, err = os.ReadFile(testFile)
-	assert.NoError(t, err)
-	assert.Equal(t, newData, content)
-}
-
-// TestRootRW_Write verifies the rootRW.Write helper function
-func TestRootRW_Write(t *testing.T) {
-	tmpDir := t.TempDir()
-
-	relPath := "atomic_root_test.txt"
-	testData := []byte("atomic root test content")
-
-	erw := &sandboxFs{workspace: tmpDir}
-	err := erw.WriteFile(relPath, testData)
-	assert.NoError(t, err)
-
-	root, err := os.OpenRoot(tmpDir)
-	assert.NoError(t, err)
-	defer root.Close()
-
-	f, err := root.Open(relPath)
-	assert.NoError(t, err)
-	defer f.Close()
-
-	content, err := io.ReadAll(f)
-	assert.NoError(t, err)
-	assert.Equal(t, testData, content)
-
-	// Verify it overwrites correctly
-	newData := []byte("new root atomic content")
-	err = erw.WriteFile(relPath, newData)
-	assert.NoError(t, err)
-
-	f2, err := root.Open(relPath)
-	assert.NoError(t, err)
-	defer f2.Close()
-
-	content, err = io.ReadAll(f2)
-	assert.NoError(t, err)
-	assert.Equal(t, newData, content)
-}

pkg/tools/shell.go

@@ -1,16 +1,13 @@
 package tools
 
 import (
-	"bytes"
 	"context"
 	"errors"
 	"fmt"
 	"os"
-	"os/exec"
 	"path"
 	"path/filepath"
 	"regexp"
-	"runtime"
 	"strings"
 	"time"
 
@@ -148,11 +145,11 @@ func (t *ExecTool) Execute(ctx context.Context, args map[string]any) *ToolResult
 	cwd := t.workingDir
 	if wd != "" {
 		if t.restrictToWorkspace && t.workingDir != "" {
-			resolvedWD, err := validatePath(wd, t.workingDir, true)
+			resolvedWD, err := sandbox.ValidatePath(wd, t.workingDir, true)
 			if err != nil {
 				// In sandbox mode, allow explicit container workspace paths when
 				// restrict_to_workspace is enabled.
-				sb := sandbox.SandboxFromContext(ctx)
+				sb := sandbox.FromContext(ctx)
 				if sb != nil && filepath.IsAbs(wd) && isSandboxWorkspaceAbsolutePath(wd) {
 					cwd = wd
 				} else {
@@ -176,8 +173,11 @@ func (t *ExecTool) Execute(ctx context.Context, args map[string]any) *ToolResult
 		return ErrorResult(guardError)
 	}
 
-	sb := sandbox.SandboxFromContext(ctx)
+	sb := sandbox.FromContext(ctx)
-	if sb != nil {
+	if sb == nil {
+		return ErrorResult("sandbox environment unavailable")
+	}
+
 	sandboxWD := t.resolveSandboxWorkingDir(cwd)
 	res, err := sb.Exec(ctx, sandbox.ExecRequest{
 		Command:    command,
@@ -185,6 +185,14 @@ func (t *ExecTool) Execute(ctx context.Context, args map[string]any) *ToolResult
 		TimeoutMs:  t.timeout.Milliseconds(),
 	})
 	if err != nil {
+		if errors.Is(err, context.DeadlineExceeded) || strings.Contains(err.Error(), "context deadline exceeded") {
+			msg := fmt.Sprintf("command timed out after %v", t.timeout)
+			return &ToolResult{
+				ForLLM:  msg,
+				ForUser: msg,
+				IsError: true,
+			}
+		}
 		return ErrorResult(fmt.Sprintf("sandbox exec failed: %v", err))
 	}
 	output := res.Stdout
@@ -194,6 +202,12 @@ func (t *ExecTool) Execute(ctx context.Context, args map[string]any) *ToolResult
 	if output == "" {
 		output = "(no output)"
 	}
+
+	maxLen := 10000
+	if len(output) > maxLen {
+		output = output[:maxLen] + fmt.Sprintf("\n... (truncated, %d more chars)", len(output)-maxLen)
+	}
+
 	if res.ExitCode != 0 {
 		output += fmt.Sprintf("\nExit code: %d", res.ExitCode)
 		return &ToolResult{
@@ -207,94 +221,6 @@ func (t *ExecTool) Execute(ctx context.Context, args map[string]any) *ToolResult
 		ForUser: output,
 		IsError: false,
 	}
-	}
-
-	// timeout == 0 means no timeout
-	var cmdCtx context.Context
-	var cancel context.CancelFunc
-	if t.timeout > 0 {
-		cmdCtx, cancel = context.WithTimeout(ctx, t.timeout)
-	} else {
-		cmdCtx, cancel = context.WithCancel(ctx)
-	}
-	defer cancel()
-
-	var cmd *exec.Cmd
-	if runtime.GOOS == "windows" {
-		cmd = exec.CommandContext(cmdCtx, "powershell", "-NoProfile", "-NonInteractive", "-Command", command)
-	} else {
-		cmd = exec.CommandContext(cmdCtx, "sh", "-c", command)
-	}
-	if cwd != "" {
-		cmd.Dir = cwd
-	}
-
-	prepareCommandForTermination(cmd)
-
-	var stdout, stderr bytes.Buffer
-	cmd.Stdout = &stdout
-	cmd.Stderr = &stderr
-
-	if err := cmd.Start(); err != nil {
-		return ErrorResult(fmt.Sprintf("failed to start command: %v", err))
-	}
-
-	done := make(chan error, 1)
-	go func() {
-		done <- cmd.Wait()
-	}()
-
-	var err error
-	select {
-	case err = <-done:
-	case <-cmdCtx.Done():
-		terminateProcessTree(cmd)
-		select {
-		case err = <-done:
-		case <-time.After(2 * time.Second):
-			if cmd.Process != nil {
-				_ = cmd.Process.Kill()
-			}
-			err = <-done
-		}
-	}
-
-	output := stdout.String()
-	if stderr.Len() > 0 {
-		output += "\nSTDERR:\n" + stderr.String()
-	}
-
-	if err != nil {
-		if errors.Is(cmdCtx.Err(), context.DeadlineExceeded) {
-			msg := fmt.Sprintf("command timed out after %v", t.timeout)
-			return &ToolResult{
-				ForLLM:  msg,
-				ForUser: msg,
-				IsError: true,
-			}
-		}
-		var exitErr *exec.ExitError
-		if errors.As(err, &exitErr) {
-			output += fmt.Sprintf("\nExit code: %d", exitErr.ExitCode())
-		} else {
-			output += fmt.Sprintf("\nError: %v", err)
-		}
-	}
-
-	if output == "" {
-		output = "(no output)"
-	}
-
-	maxLen := 10000
-	if len(output) > maxLen {
-		output = output[:maxLen] + fmt.Sprintf("\n... (truncated, %d more chars)", len(output)-maxLen)
-	}
-
-	return &ToolResult{
-		ForLLM:  output,
-		ForUser: output,
-		IsError: err != nil,
-	}
 }
 
 func (t *ExecTool) guardCommand(command, cwd string) string {

pkg/tools/shell_test.go

@@ -16,6 +16,7 @@ type stubSandbox struct {
 	lastReq sandbox.ExecRequest
 	err     error
 	res     *sandbox.ExecResult
+	fs      sandbox.FsBridge
 }
 
 func (s *stubSandbox) Start(ctx context.Context) error { return nil }
@@ -24,7 +25,14 @@ func (s *stubSandbox) Prune(ctx context.Context) error { return nil }
 func (s *stubSandbox) Resolve(ctx context.Context) (sandbox.Sandbox, error) {
 	return s, nil
 }
-func (s *stubSandbox) Fs() sandbox.FsBridge { return nil }
+
+func (s *stubSandbox) Fs() sandbox.FsBridge {
+	if s.fs != nil {
+		return s.fs
+	}
+	return sandbox.NewHostSandbox("", false).Fs()
+}
+
 func (s *stubSandbox) Exec(ctx context.Context, req sandbox.ExecRequest) (*sandbox.ExecResult, error) {
 	return sandboxAggregateFromStub(ctx, req, s.ExecStream)
 }
@@ -103,7 +111,9 @@ func sandboxAggregateFromStub(
 func TestShellTool_Success(t *testing.T) {
 	tool := NewExecTool("", false)
 
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{
+		res: &sandbox.ExecResult{Stdout: "hello world", ExitCode: 0},
+	})
 	args := map[string]any{
 		"command": "echo 'hello world'",
 	}
@@ -130,7 +140,13 @@ func TestShellTool_Success(t *testing.T) {
 func TestShellTool_Failure(t *testing.T) {
 	tool := NewExecTool("", false)
 
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{
+		res: &sandbox.ExecResult{
+			Stdout:   "",
+			Stderr:   "ls: cannot access '/nonexistent_directory_12345': No such file or directory",
+			ExitCode: 2,
+		},
+	})
 	args := map[string]any{
 		"command": "ls /nonexistent_directory_12345",
 	}
@@ -158,7 +174,9 @@ func TestShellTool_Timeout(t *testing.T) {
 	tool := NewExecTool("", false)
 	tool.SetTimeout(100 * time.Millisecond)
 
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{
+		err: context.DeadlineExceeded,
+	})
 	args := map[string]any{
 		"command": "sleep 10",
 	}
@@ -185,7 +203,9 @@ func TestShellTool_WorkingDir(t *testing.T) {
 
 	tool := NewExecTool("", false)
 
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{
+		res: &sandbox.ExecResult{Stdout: "test content\n", ExitCode: 0},
+	})
 	args := map[string]any{
 		"command":     "cat test.txt",
 		"working_dir": tmpDir,
@@ -206,7 +226,9 @@ func TestShellTool_WorkingDir(t *testing.T) {
 func TestShellTool_DangerousCommand(t *testing.T) {
 	tool := NewExecTool("", false)
 
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{
+		res: &sandbox.ExecResult{Stdout: "", Stderr: "", ExitCode: 1},
+	})
 	args := map[string]any{
 		"command": "rm -rf /",
 	}
@@ -227,7 +249,9 @@ func TestShellTool_DangerousCommand(t *testing.T) {
 func TestShellTool_MissingCommand(t *testing.T) {
 	tool := NewExecTool("", false)
 
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{
+		res: &sandbox.ExecResult{Stdout: "", Stderr: "", ExitCode: 1},
+	})
 	args := map[string]any{}
 
 	result := tool.Execute(ctx, args)
@@ -242,7 +266,9 @@ func TestShellTool_MissingCommand(t *testing.T) {
 func TestShellTool_StderrCapture(t *testing.T) {
 	tool := NewExecTool("", false)
 
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{
+		res: &sandbox.ExecResult{Stdout: "stdout", Stderr: "stderr", ExitCode: 0},
+	})
 	args := map[string]any{
 		"command": "sh -c 'echo stdout; echo stderr >&2'",
 	}
@@ -262,7 +288,9 @@ func TestShellTool_StderrCapture(t *testing.T) {
 func TestShellTool_OutputTruncation(t *testing.T) {
 	tool := NewExecTool("", false)
 
-	ctx := context.Background()
+	ctx := sandbox.WithSandbox(context.Background(), &stubSandbox{
+		res: &sandbox.ExecResult{Stdout: strings.Repeat("x", 20000), ExitCode: 0},
+	})
 	// Generate long output (>10000 chars)
 	args := map[string]any{
 		"command": "python3 -c \"print('x' * 20000)\" || echo " + strings.Repeat("x", 20000),
@@ -289,7 +317,9 @@ func TestShellTool_WorkingDir_OutsideWorkspace(t *testing.T) {
 	}
 
 	tool := NewExecTool(workspace, true)
-	result := tool.Execute(context.Background(), map[string]any{
+	result := tool.Execute(sandbox.WithSandbox(context.Background(), &stubSandbox{
+		res: &sandbox.ExecResult{Stdout: "", Stderr: "", ExitCode: 1},
+	}), map[string]any{
 		"command":     "pwd",
 		"working_dir": outsideDir,
 	})

pkg/tools/shell_timeout_unix_test.go

@@ -11,6 +11,8 @@ import (
 	"syscall"
 	"testing"
 	"time"
+
+	"github.com/sipeed/picoclaw/pkg/agent/sandbox"
 )
 
 func processExists(pid int) bool {
@@ -22,15 +24,25 @@ func processExists(pid int) bool {
 }
 
 func TestShellTool_TimeoutKillsChildProcess(t *testing.T) {
-	tool := NewExecTool(t.TempDir(), false)
+	workspace := t.TempDir()
+	tool := NewExecTool(workspace, false)
 	tool.SetTimeout(500 * time.Millisecond)
 
+	sb := sandbox.NewHostSandbox(workspace, false)
+	err := sb.Start(context.Background())
+	if err != nil {
+		t.Fatalf("failed to start sandbox: %v", err)
+	}
+	defer sb.Prune(context.Background())
+
+	ctx := sandbox.WithSandbox(context.Background(), sb)
+
 	args := map[string]any{
 		// Spawn a child process that would outlive the shell unless process-group kill is used.
 		"command": "sleep 60 & echo $! > child.pid; wait",
 	}
 
-	result := tool.Execute(context.Background(), args)
+	result := tool.Execute(ctx, args)
 	if !result.IsError {
 		t.Fatalf("expected timeout error, got success: %s", result.ForLLM)
 	}