tui: migrate off the deprecated Bubble Tea mouse Type API (SA1019)

The mouse classification helpers checked both the current Button/Action pair and
the deprecated msg.Type / tea.Mouse* constants. Bubble Tea's parser always
populates Button+Action and only DERIVES Type from them, so the Type fallbacks
were pure redundancy (a left-button drag is Action==Motion, already covered by
mouseMotion). Drop them — clears all 11 SA1019 findings in internal/tui with no
behavior change. The left-drag test now asserts the Action==Motion path directly.

Gates: gofmt/vet/build(host+linux+windows)/test -race/staticcheck(SA1019 0)/
deadcode(no new) all pass.

internal/cli/app.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"io"
 	"os"
+	"path/filepath"
 	"strings"
 	"time"
 
@@ -23,6 +24,7 @@ import (
 	"github.com/Gitlawb/zero/internal/sessions"
 	"github.com/Gitlawb/zero/internal/skills"
 	"github.com/Gitlawb/zero/internal/specialist"
+	"github.com/Gitlawb/zero/internal/swarm"
 	"github.com/Gitlawb/zero/internal/tools"
 	"github.com/Gitlawb/zero/internal/tui"
 	"github.com/Gitlawb/zero/internal/update"
@@ -240,6 +242,8 @@ func runWithDeps(args []string, stdout io.Writer, stderr io.Writer, deps appDeps
 	case "exec":
 		// Forward leading --add-dir occurrences so exec's own parser collects them.
 		return runExec(append(addDirFlagArgs(addDirs), args[1:]...), stdout, stderr, deps)
+	case "daemon":
+		return runDaemon(args[1:], stdout, stderr, deps)
 	case "config":
 		return runConfig(args[1:], stdout, stderr, deps)
 	case "models":
@@ -274,6 +278,8 @@ func runWithDeps(args []string, stdout io.Writer, stderr io.Writer, deps appDeps
 		return runHooks(args[1:], stdout, stderr, deps)
 	case "mcp":
 		return runMCP(args[1:], stdout, stderr, deps)
+	case "auth":
+		return runAuth(args[1:], stdout, stderr, deps)
 	case "sandbox":
 		return runSandbox(args[1:], stdout, stderr, deps)
 	case "update":
@@ -415,6 +421,12 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 	if err != nil {
 		return writeAppError(stderr, "failed to resolve user config path: "+err.Error(), 1)
 	}
+	doctorUserConfigPath := ""
+	projectConfigPath := ""
+	if resolveOptions, optErr := config.DefaultResolveOptions(workspaceRoot); optErr == nil {
+		doctorUserConfigPath = resolveOptions.UserConfigPath
+		projectConfigPath = resolveOptions.ProjectConfigPath
+	}
 
 	needsSetup := setupRequired(resolved)
 	setupVisible := forceSetup || needsSetup
@@ -507,20 +519,24 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 	})
 	lastKnownMCPConfig := mcpConfig
 	return deps.runTUI(context.Background(), tui.Options{
-		Cwd:                workspaceRoot,
-		UserConfigPath:     userConfigPath,
-		ProviderName:       resolved.Provider.Name,
-		ModelName:          resolved.Provider.Model,
-		ProviderProfile:    resolved.Provider,
-		FavoriteModels:     resolved.Preferences.FavoriteModels,
-		Provider:           provider,
-		NewProvider:        deps.newProvider,
-		Registry:           registry,
-		SessionStore:       deps.newSessionStore(),
-		SandboxStore:       sandboxStore,
-		MCPConfig:          mcpConfig,
-		MCPPermissionStore: mcpPermissionStore,
-		MCPTokenStore:      mcpTokenStore,
+		Cwd:                  workspaceRoot,
+		UserConfigPath:       userConfigPath,
+		DoctorUserConfigPath: doctorUserConfigPath,
+		ProjectConfigPath:    projectConfigPath,
+		ProviderName:         resolved.Provider.Name,
+		ModelName:            resolved.Provider.Model,
+		ProviderProfile:      resolved.Provider,
+		FavoriteModels:       resolved.Preferences.FavoriteModels,
+		Provider:             provider,
+		NewProvider:          deps.newProvider,
+		ProbeProviderHealth:  deps.probeProviderHealth,
+		UserAgent:            userAgent(),
+		Registry:             registry,
+		SessionStore:         deps.newSessionStore(),
+		SandboxStore:         sandboxStore,
+		MCPConfig:            mcpConfig,
+		MCPPermissionStore:   mcpPermissionStore,
+		MCPTokenStore:        mcpTokenStore,
 		MCPCommand: func(ctx context.Context, args []string) tui.MCPCommandResult {
 			if ctx == nil {
 				ctx = context.Background()
@@ -582,12 +598,37 @@ func newCoreRegistryScoped(workspaceRoot string, scope tools.PathScope) *tools.R
 	return registry
 }
 
-func registerSpecialistTools(registry *tools.Registry, workspaceRoot string) (*specialist.Runtime, error) {
+// agentToolRuntime bundles the specialist runtime with the swarm it backs so
+// their lifetimes (and shutdown) stay paired: registerSpecialistTools brings up
+// both, closeSpecialistRuntime tears down both.
+type agentToolRuntime struct {
+	specialist *specialist.Runtime
+	swarm      *swarm.Swarm
+}
+
+func registerSpecialistTools(registry *tools.Registry, workspaceRoot string) (*agentToolRuntime, error) {
 	paths, err := specialist.DefaultPaths(workspaceRoot)
 	if err != nil {
 		return nil, err
 	}
-	return specialist.RegisterTools(registry, specialist.Executor{Paths: paths})
+	executor := specialist.Executor{Paths: paths}
+	runtime, err := specialist.RegisterTools(registry, executor)
+	if err != nil {
+		return nil, err
+	}
+	// The swarm reuses the same specialist executor to launch each member, so
+	// every member runs under the orchestrator's sandbox + policy. Mailbox state
+	// lives under the workspace so its files fall within the sandbox write rules.
+	sw, err := swarm.New(swarm.Options{
+		BaseDir:  filepath.Join(workspaceRoot, ".zero", "swarm"),
+		Launcher: swarm.NewSpecialistLauncher(executor),
+	})
+	if err != nil {
+		runtime.Close()
+		return nil, err
+	}
+	swarm.RegisterTools(registry, sw)
+	return &agentToolRuntime{specialist: runtime, swarm: sw}, nil
 }
 
 func shouldRegisterExecSpecialistTools(options execOptions) bool {
@@ -609,12 +650,17 @@ func closeMCPRuntime(stderr io.Writer, runtime mcpToolRuntime) {
 	}
 }
 
-func closeSpecialistRuntime(stderr io.Writer, runtime *specialist.Runtime) {
+func closeSpecialistRuntime(stderr io.Writer, runtime *agentToolRuntime) {
 	if runtime == nil {
 		return
 	}
-	if err := runtime.Close(); err != nil {
-		_, _ = fmt.Fprintf(stderr, "[zero] specialist_cleanup_error: %s\n", err)
+	if runtime.swarm != nil {
+		runtime.swarm.Close()
+	}
+	if runtime.specialist != nil {
+		if err := runtime.specialist.Close(); err != nil {
+			_, _ = fmt.Fprintf(stderr, "[zero] specialist_cleanup_error: %s\n", err)
+		}
 	}
 }
 
@@ -637,6 +683,7 @@ Usage:
 
 Commands:
   exec       Run a one-shot prompt through the Go agent runtime
+  daemon     Manage the local background worker daemon (start/stop/status/run/attach)
   setup      Guide first-run provider setup
   config     Inspect resolved Go configuration without leaking secrets
   models     List Zero model registry entries
@@ -655,6 +702,7 @@ Commands:
   tools      Scaffold and list local Zero plugin-tools
   hooks      Inspect Zero hook configuration
   mcp        Manage MCP backend settings
+  auth       Log in to model providers via OAuth
   sandbox    Inspect sandbox policy and persistent grants
   update     Check for Zero CLI updates
   worktrees  Prepare isolated git worktrees

internal/cli/app_test.go

@@ -8,6 +8,7 @@ import (
 	"io"
 	"os"
 	"path/filepath"
+	"runtime"
 	"strings"
 	"testing"
 	"time"
@@ -59,10 +60,38 @@ func TestRunPrintsHelp(t *testing.T) {
 	}
 }
 
+func setCLIUserConfigRoot(t *testing.T) string {
+	t.Helper()
+
+	root := t.TempDir()
+	switch runtime.GOOS {
+	case "windows":
+		t.Setenv("APPDATA", root)
+	case "darwin":
+		t.Setenv("HOME", root)
+	default:
+		t.Setenv("XDG_CONFIG_HOME", root)
+	}
+
+	configRoot, err := os.UserConfigDir()
+	if err != nil {
+		t.Fatalf("UserConfigDir() error = %v", err)
+	}
+	return configRoot
+}
+
 func TestRunNoArgsLaunchesSetupTUIWithNilProviderWhenNoProviderConfigured(t *testing.T) {
 	var stdout bytes.Buffer
 	var stderr bytes.Buffer
 	cwd := t.TempDir()
+	setCLIUserConfigRoot(t)
+	projectConfigPath := filepath.Join(cwd, ".zero", "config.json")
+	if err := os.MkdirAll(filepath.Dir(projectConfigPath), 0o700); err != nil {
+		t.Fatalf("create project config parent: %v", err)
+	}
+	if err := os.WriteFile(projectConfigPath, []byte("{}"), 0o600); err != nil {
+		t.Fatalf("write project config: %v", err)
+	}
 	userConfigPath := filepath.Join(t.TempDir(), "zero", "config.json")
 	var launchedOptions tui.Options
 
@@ -104,6 +133,12 @@ func TestRunNoArgsLaunchesSetupTUIWithNilProviderWhenNoProviderConfigured(t *tes
 	if launchedOptions.UserConfigPath != userConfigPath {
 		t.Fatalf("UserConfigPath = %q, want %q", launchedOptions.UserConfigPath, userConfigPath)
 	}
+	if launchedOptions.DoctorUserConfigPath != "" {
+		t.Fatalf("DoctorUserConfigPath = %q, want empty for missing user config", launchedOptions.DoctorUserConfigPath)
+	}
+	if launchedOptions.ProjectConfigPath != projectConfigPath {
+		t.Fatalf("ProjectConfigPath = %q, want %q", launchedOptions.ProjectConfigPath, projectConfigPath)
+	}
 	if launchedOptions.Provider != nil {
 		t.Fatalf("Provider = %#v, want nil", launchedOptions.Provider)
 	}

internal/cli/auth.go

@@ -0,0 +1,344 @@
+package cli
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/oauth"
+	"github.com/Gitlawb/zero/internal/redaction"
+)
+
+// runAuth dispatches `zero auth <command>` for provider OAuth login. It is
+// additive and independent of `zero mcp oauth` (MCP server auth), which is
+// unchanged.
+func runAuth(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
+	if len(args) == 0 {
+		return writeExecUsageError(stderr, "auth subcommand required. Use `zero auth status`.")
+	}
+	if args[0] == "-h" || args[0] == "--help" || args[0] == "help" {
+		if err := writeAuthHelp(stdout); err != nil {
+			return exitCrash
+		}
+		return exitSuccess
+	}
+	switch args[0] {
+	case "login":
+		return runAuthLogin(args[1:], stdout, stderr, deps)
+	case "logout":
+		return runAuthLogout(args[1:], stdout, stderr, deps)
+	case "status":
+		return runAuthStatus(args[1:], stdout, stderr, deps)
+	case "refresh":
+		return runAuthRefresh(args[1:], stdout, stderr, deps)
+	default:
+		return writeExecUsageError(stderr, fmt.Sprintf("unknown auth subcommand %q", args[0]))
+	}
+}
+
+// authArgs is the parsed form of an auth subcommand's arguments.
+type authArgs struct {
+	positional []string
+	json       bool
+	device     bool
+	watch      bool
+	scopes     []string
+	help       bool
+}
+
+func parseAuthArgs(sub string, args []string) (authArgs, error) {
+	var parsed authArgs
+	addScope := func(scope string) error {
+		scope = strings.TrimSpace(scope)
+		if scope == "" {
+			return fmt.Errorf("--scope requires a non-empty value")
+		}
+		parsed.scopes = append(parsed.scopes, scope)
+		return nil
+	}
+	for i := 0; i < len(args); i++ {
+		arg := args[i]
+		switch {
+		case arg == "-h" || arg == "--help" || arg == "help":
+			parsed.help = true
+		case arg == "--json":
+			parsed.json = true
+		case arg == "--device":
+			parsed.device = true
+		case arg == "--watch":
+			parsed.watch = true
+		case arg == "--scope":
+			if i+1 >= len(args) {
+				return authArgs{}, fmt.Errorf("--scope requires a value")
+			}
+			i++
+			if err := addScope(args[i]); err != nil {
+				return authArgs{}, err
+			}
+		case strings.HasPrefix(arg, "--scope="):
+			if err := addScope(strings.TrimPrefix(arg, "--scope=")); err != nil {
+				return authArgs{}, err
+			}
+		case strings.HasPrefix(arg, "-"):
+			return authArgs{}, fmt.Errorf("unknown flag %q", arg)
+		default:
+			parsed.positional = append(parsed.positional, arg)
+		}
+	}
+	if parsed.help {
+		return parsed, nil // help short-circuits flag validation
+	}
+	if err := validateAuthFlags(sub, parsed); err != nil {
+		return authArgs{}, err
+	}
+	return parsed, nil
+}
+
+// validateAuthFlags rejects flags a subcommand does not accept, so an ambiguous
+// invocation fails fast instead of silently ignoring a flag.
+func validateAuthFlags(sub string, a authArgs) error {
+	allowed := map[string]map[string]bool{
+		"login":   {"device": true, "scope": true},
+		"logout":  {"json": true},
+		"status":  {"json": true},
+		"refresh": {"watch": true},
+	}[sub]
+	bad := func(name string) error { return fmt.Errorf("zero auth %s does not accept %s", sub, name) }
+	if a.json && !allowed["json"] {
+		return bad("--json")
+	}
+	if a.device && !allowed["device"] {
+		return bad("--device")
+	}
+	if a.watch && !allowed["watch"] {
+		return bad("--watch")
+	}
+	if len(a.scopes) > 0 && !allowed["scope"] {
+		return bad("--scope")
+	}
+	return nil
+}
+
+// newAuthManager builds an oauth.Manager backed by the file store, printing the
+// authorization URL / device code to stdout. The store path honors
+// ZERO_OAUTH_TOKENS_PATH (env), so callers/tests can redirect it.
+func newAuthManager(deps appDeps, out io.Writer) (*oauth.Manager, error) {
+	store, err := oauth.NewStore(oauth.StoreOptions{Now: deps.now})
+	if err != nil {
+		return nil, err
+	}
+	return oauth.NewManager(oauth.ManagerOptions{
+		Store:      store,
+		HTTPClient: &http.Client{Timeout: 30 * time.Second},
+		Now:        deps.now,
+		Out:        out,
+		// The opener prints the URL so headless shells can copy it; the URL
+		// carries no token material. A real browser launch is intentionally not
+		// performed (the sandbox/headless contexts make printing the safer default).
+		OpenBrowser: func(string) error { return nil },
+	})
+}
+
+func runAuthLogin(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
+	parsed, err := parseAuthArgs("login", args)
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+	if parsed.help {
+		_ = writeAuthHelp(stdout)
+		return exitSuccess
+	}
+	if len(parsed.positional) != 1 {
+		return writeExecUsageError(stderr, "usage: zero auth login <provider> [--device] [--scope <scope>]")
+	}
+	manager, err := newAuthManager(deps, stdout)
+	if err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	status, err := manager.Login(context.Background(), oauth.LoginOptions{
+		Provider:    parsed.positional[0],
+		Device:      parsed.device,
+		ExtraScopes: parsed.scopes,
+	})
+	if err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	if _, err := fmt.Fprintf(stdout, "Logged in to %s.\n%s\n", parsed.positional[0], oauth.FormatStatuses([]oauth.Status{status})); err != nil {
+		return exitCrash
+	}
+	return exitSuccess
+}
+
+func runAuthLogout(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
+	parsed, err := parseAuthArgs("logout", args)
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+	if parsed.help {
+		_ = writeAuthHelp(stdout)
+		return exitSuccess
+	}
+	if len(parsed.positional) != 1 {
+		return writeExecUsageError(stderr, "usage: zero auth logout <provider>")
+	}
+	provider := parsed.positional[0]
+	manager, err := newAuthManager(deps, stdout)
+	if err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	removed, err := manager.Logout(provider)
+	if err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	if parsed.json {
+		payload := struct {
+			Provider string `json:"provider"`
+			Removed  bool   `json:"removed"`
+		}{Provider: provider, Removed: removed}
+		if err := writePrettyJSON(stdout, payload); err != nil {
+			return exitCrash
+		}
+		return exitSuccess
+	}
+	msg := fmt.Sprintf("No OAuth login stored for %s.\n", provider)
+	if removed {
+		msg = fmt.Sprintf("Logged out of %s.\n", provider)
+	}
+	if _, err := fmt.Fprint(stdout, msg); err != nil {
+		return exitCrash
+	}
+	return exitSuccess
+}
+
+func runAuthStatus(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
+	parsed, err := parseAuthArgs("status", args)
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+	if parsed.help {
+		_ = writeAuthHelp(stdout)
+		return exitSuccess
+	}
+	if len(parsed.positional) > 1 {
+		return writeExecUsageError(stderr, "usage: zero auth status [provider]")
+	}
+	manager, err := newAuthManager(deps, stdout)
+	if err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	statuses, err := manager.StatusAll()
+	if err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	if len(parsed.positional) == 1 {
+		statuses = filterAuthStatuses(statuses, parsed.positional[0])
+	}
+	if parsed.json {
+		payload := struct {
+			Logins []oauth.Status `json:"logins"`
+		}{Logins: statuses}
+		if err := writePrettyJSON(stdout, payload); err != nil {
+			return exitCrash
+		}
+		return exitSuccess
+	}
+	if _, err := fmt.Fprintln(stdout, oauth.FormatStatuses(statuses)); err != nil {
+		return exitCrash
+	}
+	return exitSuccess
+}
+
+func runAuthRefresh(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
+	parsed, err := parseAuthArgs("refresh", args)
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+	if parsed.help {
+		_ = writeAuthHelp(stdout)
+		return exitSuccess
+	}
+	if len(parsed.positional) != 1 {
+		return writeExecUsageError(stderr, "usage: zero auth refresh <provider>")
+	}
+	provider := parsed.positional[0]
+	manager, err := newAuthManager(deps, stdout)
+	if err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	key := oauth.ProviderKey(provider)
+	if parsed.watch {
+		return runAuthRefreshWatch(manager, key, provider, stdout, stderr)
+	}
+	if _, err := manager.Handle401(context.Background(), key); err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	if _, err := fmt.Fprintf(stdout, "Refreshed OAuth token for %s.\n", provider); err != nil {
+		return exitCrash
+	}
+	return exitSuccess
+}
+
+// runAuthRefreshWatch keeps a provider's token fresh in the foreground until
+// interrupted. This is the opt-in proactive-refresh scheduler surface (for a
+// long-running external process that reads the token file). It validates a
+// refreshable token exists first, then schedules refreshes before each expiry.
+func runAuthRefreshWatch(manager *oauth.Manager, key, provider string, stdout io.Writer, stderr io.Writer) int {
+	ctx, stop := signalContext()
+	defer stop()
+	// Validate (and refresh-if-needed) once up front so a missing/non-refreshable
+	// token fails fast instead of silently watching nothing.
+	if _, err := manager.GetFresh(ctx, key); err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	scheduler := oauth.NewRefreshScheduler()
+	scheduler.Start(ctx, manager, key)
+	defer scheduler.Stop()
+	if _, err := fmt.Fprintf(stdout, "Watching %s — refreshing before expiry. Press Ctrl+C to stop.\n", provider); err != nil {
+		return exitCrash
+	}
+	<-ctx.Done()
+	return exitSuccess
+}
+
+func filterAuthStatuses(statuses []oauth.Status, provider string) []oauth.Status {
+	want := oauth.ProviderKey(provider)
+	filtered := make([]oauth.Status, 0, 1)
+	for _, st := range statuses {
+		if st.Key == want {
+			filtered = append(filtered, st)
+		}
+	}
+	return filtered
+}
+
+func writeAuthHelp(w io.Writer) error {
+	_, err := fmt.Fprint(w, `Usage:
+  zero auth <command>
+
+Commands:
+  login <provider> [--device] [--scope <scope>]   Log in to a provider via OAuth
+  logout <provider>                               Delete a provider's stored login
+  status [provider]                               Show login presence/expiry (never the token)
+  refresh <provider> [--watch]                    Force a token refresh (--watch keeps it fresh)
+
+A provider is any OAuth 2.0 / OIDC server you configure via env (no providers are
+built in). For a provider named <name>, set:
+  ZERO_OAUTH_<NAME>_CLIENT_ID       (required)
+  ZERO_OAUTH_<NAME>_CLIENT_SECRET   (optional)
+  ZERO_OAUTH_<NAME>_AUTHORIZE_URL   ZERO_OAUTH_<NAME>_TOKEN_URL
+  ZERO_OAUTH_<NAME>_DEVICE_URL      ZERO_OAUTH_<NAME>_ISSUER_URL (for discovery)
+  ZERO_OAUTH_<NAME>_SCOPES          ZERO_OAUTH_<NAME>_FLOW (loopback|device)
+Endpoint URLs must be https (loopback exempt).
+
+Flags:
+      --device   Use the device-code flow (headless/SSH; no browser)
+      --scope    Add an OAuth scope (repeatable)
+      --watch    Keep the token fresh in the foreground (refresh only)
+      --json     Print result as JSON (status/logout)
+  -h, --help     Show this help
+`)
+	return err
+}

internal/cli/auth_test.go

@@ -0,0 +1,129 @@
+package cli
+
+import (
+	"bytes"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/oauth"
+)
+
+// withAuthStore points the provider OAuth store at a temp file for the test.
+func withAuthStore(t *testing.T) string {
+	t.Helper()
+	path := filepath.Join(t.TempDir(), "oauth-tokens.json")
+	t.Setenv("ZERO_OAUTH_TOKENS_PATH", path)
+	return path
+}
+
+func TestRunAuthStatusEmpty(t *testing.T) {
+	withAuthStore(t)
+	var stdout, stderr bytes.Buffer
+	if code := runWithDeps([]string{"auth", "status"}, &stdout, &stderr, appDeps{}); code != exitSuccess {
+		t.Fatalf("exit = %d stderr=%s", code, stderr.String())
+	}
+	if !strings.Contains(stdout.String(), "No OAuth provider logins are stored.") {
+		t.Fatalf("status output = %q", stdout.String())
+	}
+}
+
+func TestRunAuthStatusReportsLoginWithoutSecret(t *testing.T) {
+	path := withAuthStore(t)
+	store, err := oauth.NewStore(oauth.StoreOptions{FilePath: path})
+	if err != nil {
+		t.Fatalf("NewStore: %v", err)
+	}
+	if err := store.Save(oauth.ProviderKey("demo"), oauth.Token{
+		AccessToken: "super-secret", RefreshToken: "super-secret-rt", Account: "[email protected]",
+	}); err != nil {
+		t.Fatalf("Save: %v", err)
+	}
+	var stdout, stderr bytes.Buffer
+	if code := runWithDeps([]string{"auth", "status"}, &stdout, &stderr, appDeps{}); code != exitSuccess {
+		t.Fatalf("exit = %d stderr=%s", code, stderr.String())
+	}
+	out := stdout.String()
+	if !strings.Contains(out, "demo") || !strings.Contains(out, "[email protected]") {
+		t.Fatalf("status should show provider + account: %q", out)
+	}
+	if strings.Contains(out, "super-secret") {
+		t.Fatalf("status leaked token material: %q", out)
+	}
+}
+
+func TestRunAuthLogoutNothing(t *testing.T) {
+	withAuthStore(t)
+	var stdout, stderr bytes.Buffer
+	if code := runWithDeps([]string{"auth", "logout", "demo"}, &stdout, &stderr, appDeps{}); code != exitSuccess {
+		t.Fatalf("exit = %d stderr=%s", code, stderr.String())
+	}
+	if !strings.Contains(stdout.String(), "No OAuth login stored for demo") {
+		t.Fatalf("logout output = %q", stdout.String())
+	}
+}
+
+func TestRunAuthLoginValidation(t *testing.T) {
+	withAuthStore(t)
+	var stdout, stderr bytes.Buffer
+	// Missing provider.
+	if code := runWithDeps([]string{"auth", "login"}, &stdout, &stderr, appDeps{}); code == exitSuccess {
+		t.Fatal("login with no provider should fail")
+	}
+	// --json is rejected for the interactive login.
+	stdout.Reset()
+	stderr.Reset()
+	if code := runWithDeps([]string{"auth", "login", "demo", "--json"}, &stdout, &stderr, appDeps{}); code == exitSuccess {
+		t.Fatal("login --json should be rejected")
+	}
+}
+
+func TestRunAuthLoginUnknownProvider(t *testing.T) {
+	withAuthStore(t)
+	var stdout, stderr bytes.Buffer
+	if code := runWithDeps([]string{"auth", "login", "does-not-exist"}, &stdout, &stderr, appDeps{}); code == exitSuccess {
+		t.Fatal("unknown provider login should fail")
+	}
+	if !strings.Contains(stderr.String(), "not configured") {
+		t.Fatalf("stderr = %q, want not-configured error", stderr.String())
+	}
+}
+
+func TestRunAuthRefreshNoToken(t *testing.T) {
+	withAuthStore(t)
+	t.Setenv("ZERO_OAUTH_DEMO_CLIENT_ID", "client") // so config resolves; refresh still fails (no token)
+	var stdout, stderr bytes.Buffer
+	if code := runWithDeps([]string{"auth", "refresh", "demo"}, &stdout, &stderr, appDeps{}); code == exitSuccess {
+		t.Fatal("refresh with no stored token should fail")
+	}
+}
+
+func TestRunAuthRejectsWrongFlags(t *testing.T) {
+	withAuthStore(t)
+	cases := [][]string{
+		{"auth", "login", "demo", "--watch"},       // watch is refresh-only
+		{"auth", "login", "demo", "--json"},        // json not for interactive login
+		{"auth", "status", "demo", "--device"},     // device is login-only
+		{"auth", "logout", "demo", "--scope", "x"}, // scope is login-only
+		{"auth", "refresh", "demo", "--json"},      // json not for refresh
+		{"auth", "login", "demo", "--scope", ""},   // empty scope rejected
+	}
+	for _, args := range cases {
+		var stdout, stderr bytes.Buffer
+		if code := runWithDeps(args, &stdout, &stderr, appDeps{}); code == exitSuccess {
+			t.Errorf("args %v should be rejected, got success", args)
+		}
+	}
+}
+
+func TestRunAuthHelp(t *testing.T) {
+	var stdout, stderr bytes.Buffer
+	if code := runWithDeps([]string{"auth", "--help"}, &stdout, &stderr, appDeps{}); code != exitSuccess {
+		t.Fatalf("exit = %d", code)
+	}
+	for _, want := range []string{"zero auth", "login", "logout", "status", "refresh", "--device"} {
+		if !strings.Contains(stdout.String(), want) {
+			t.Fatalf("help missing %q:\n%s", want, stdout.String())
+		}
+	}
+}

internal/cli/backends.go

@@ -17,6 +17,13 @@ type backendStatusOptions struct {
 }
 
 func runBackends(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
+	if len(args) > 0 {
+		switch args[0] {
+		case "doctor":
+			return runBackendsDoctor(args[1:], stdout, stderr, deps)
+		}
+	}
+
 	options, help, err := parseBackendsArgs(args)
 	if err != nil {
 		return writeExecUsageError(stderr, err.Error())
@@ -44,6 +51,44 @@ func runBackends(args []string, stdout io.Writer, stderr io.Writer, deps appDeps
 	return exitSuccess
 }
 
+type backendDoctorOptions struct {
+	json bool
+}
+
+func runBackendsDoctor(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
+	options, help, err := parseBackendsDoctorArgs(args)
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+	if help {
+		if err := writeBackendsDoctorHelp(stdout); err != nil {
+			return exitCrash
+		}
+		return exitSuccess
+	}
+
+	report, err := backendDoctorReport(deps)
+	if err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	if options.json {
+		if err := writePrettyJSON(stdout, report); err != nil {
+			return exitCrash
+		}
+		if !report.OK {
+			return exitProvider
+		}
+		return exitSuccess
+	}
+	if _, err := fmt.Fprintln(stdout, redaction.RedactString(formatBackendDoctorReport(report), redaction.Options{})); err != nil {
+		return exitCrash
+	}
+	if !report.OK {
+		return exitProvider
+	}
+	return exitSuccess
+}
+
 func parseBackendsArgs(args []string) (backendStatusOptions, bool, error) {
 	options := backendStatusOptions{}
 	for _, arg := range args {
@@ -59,6 +104,21 @@ func parseBackendsArgs(args []string) (backendStatusOptions, bool, error) {
 	return options, false, nil
 }
 
+func parseBackendsDoctorArgs(args []string) (backendDoctorOptions, bool, error) {
+	options := backendDoctorOptions{}
+	for _, arg := range args {
+		switch arg {
+		case "-h", "--help", "help":
+			return options, true, nil
+		case "--json":
+			options.json = true
+		default:
+			return options, false, execUsageError{fmt.Sprintf("unknown backends doctor flag %q", arg)}
+		}
+	}
+	return options, false, nil
+}
+
 func backendLifecycleSnapshot(deps appDeps) (zerocommands.BackendLifecycleSnapshot, error) {
 	cwd, err := deps.getwd()
 	if err != nil {
@@ -86,6 +146,32 @@ func backendLifecycleSnapshot(deps appDeps) (zerocommands.BackendLifecycleSnapsh
 	return zerocommands.NewBackendLifecycleSnapshot(servers, hookResult.Config.Hooks, pluginResult.Plugins), nil
 }
 
+func backendDoctorReport(deps appDeps) (zerocommands.BackendDoctorReport, error) {
+	cwd, err := deps.getwd()
+	if err != nil {
+		return zerocommands.BackendDoctorReport{}, fmt.Errorf("failed to resolve workspace: %w", err)
+	}
+
+	cfg, err := deps.resolveMCPConfig(cwd)
+	if err != nil {
+		return zerocommands.BackendDoctorReport{}, err
+	}
+	hookResult, err := deps.loadHooks(hooks.LoadOptions{Cwd: cwd})
+	if err != nil {
+		return zerocommands.BackendDoctorReport{}, err
+	}
+	pluginResult, err := deps.loadPlugins(plugins.LoadOptions{Cwd: cwd})
+	if err != nil {
+		return zerocommands.BackendDoctorReport{}, err
+	}
+
+	return zerocommands.NewBackendDoctorReport(zerocommands.BackendDoctorInput{
+		MCP:     cfg,
+		Hooks:   hookResult,
+		Plugins: pluginResult,
+	}), nil
+}
+
 func formatBackendLifecycleSnapshot(snapshot zerocommands.BackendLifecycleSnapshot) string {
 	lines := []string{"Zero Backends:"}
 	lines = append(lines, fmt.Sprintf("  MCP servers: %d", len(snapshot.MCPServers)))
@@ -135,16 +221,52 @@ func formatBackendLifecycleSnapshot(snapshot zerocommands.BackendLifecycleSnapsh
 	return strings.Join(lines, "\n")
 }
 
+func formatBackendDoctorReport(report zerocommands.BackendDoctorReport) string {
+	lines := []string{
+		"Zero backend doctor",
+		"Overall: " + string(report.Status),
+	}
+	for _, check := range report.Checks {
+		target := ""
+		if check.Target != "" {
+			target = " " + check.Target
+		}
+		lines = append(lines, fmt.Sprintf("[%s] %s%s - %s", check.Status, check.ID, target, check.Message))
+		if check.Action != "" {
+			lines = append(lines, "  action: "+check.Action)
+		}
+	}
+	return strings.Join(lines, "\n")
+}
+
 func writeBackendsHelp(w io.Writer) error {
 	_, err := fmt.Fprint(w, `Usage:
   zero backends [flags]
+  zero backends doctor [flags]
 
 Inspect MCP, hook, and plugin backend lifecycle state without connecting to
 external MCP servers.
 
+Commands:
+  doctor    Diagnose MCP, hook, and plugin backend setup
+
 Flags:
       --json    Print backend lifecycle data as JSON
   -h, --help    Show this help
 `)
 	return err
 }
+
+func writeBackendsDoctorHelp(w io.Writer) error {
+	_, err := fmt.Fprint(w, `Usage:
+  zero backends doctor [flags]
+
+Diagnose MCP, hook, and plugin backend setup without connecting to external MCP
+servers or executing hooks/plugins.
+
+Flags:
+      --json    Print backend diagnostics as JSON
+  -h, --help    Show this help
+`)
+	return err
+}

internal/cli/backends_test.go

@@ -5,7 +5,12 @@ import (
 	"context"
 	"encoding/json"
 	"errors"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
 	"strings"
+	"sync/atomic"
 	"testing"
 
 	"github.com/Gitlawb/zero/internal/config"
@@ -153,3 +158,220 @@ func TestRunBackendsTextAndHelp(t *testing.T) {
 		}
 	}
 }
+
+func TestRunBackendsDoctorJSONAndTextWithoutConnectingMCP(t *testing.T) {
+	cwd := t.TempDir()
+	secret := "sk-proj-" + strings.Repeat("b", 24)
+	deps := appDeps{
+		getwd: func() (string, error) { return cwd, nil },
+		resolveMCPConfig: func(workspaceRoot string) (config.MCPConfig, error) {
+			if workspaceRoot != cwd {
+				t.Fatalf("workspaceRoot = %q, want %q", workspaceRoot, cwd)
+			}
+			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+				"remote": {
+					Type: "http",
+					URL:  "https://api.example.com/mcp?token=" + secret,
+				},
+				"broken": {
+					Type: "http",
+				},
+			}}, nil
+		},
+		loadHooks: func(options hooks.LoadOptions) (hooks.LoadResult, error) {
+			if options.Cwd != cwd {
+				t.Fatalf("hook Cwd = %q, want %q", options.Cwd, cwd)
+			}
+			return hooks.LoadResult{Diagnostics: []hooks.Diagnostic{{
+				Kind:    hooks.DiagnosticSchema,
+				Message: "schema failed " + secret,
+				HookID:  "zero.bad",
+			}}}, nil
+		},
+		loadPlugins: func(options plugins.LoadOptions) (plugins.LoadResult, error) {
+			if options.Cwd != cwd {
+				t.Fatalf("plugin Cwd = %q, want %q", options.Cwd, cwd)
+			}
+			return plugins.LoadResult{Plugins: []plugins.LoadedPlugin{{
+				ID:      "zero.docs",
+				Name:    "Docs",
+				Enabled: true,
+				Source:  plugins.SourceProject,
+			}}}, nil
+		},
+		registerMCPTools: func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			return nil, errors.New("zero backends doctor must not connect to MCP servers")
+		},
+	}
+
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	exitCode := runWithDeps([]string{"backends", "doctor", "--json"}, &stdout, &stderr, deps)
+	if exitCode != exitProvider {
+		t.Fatalf("exitCode = %d, want %d stderr=%s stdout=%s", exitCode, exitProvider, stderr.String(), stdout.String())
+	}
+	if strings.Contains(stdout.String(), secret) || strings.Contains(stdout.String(), "sk-proj-") {
+		t.Fatalf("backend doctor JSON leaked secret material:\n%s", stdout.String())
+	}
+	var payload zerocommands.BackendDoctorReport
+	if err := json.Unmarshal(stdout.Bytes(), &payload); err != nil {
+		t.Fatalf("backend doctor JSON failed to decode: %v\n%s", err, stdout.String())
+	}
+	if payload.OK {
+		t.Fatalf("payload.OK = true, want false for invalid MCP/hook diagnostic: %#v", payload.Checks)
+	}
+	if payload.Status != zerocommands.BackendDoctorStatusFail {
+		t.Fatalf("payload.Status = %q, want %q", payload.Status, zerocommands.BackendDoctorStatusFail)
+	}
+	assertBackendDoctorPayloadCheck(t, payload, "backend.mcp.invalid", "broken")
+	assertBackendDoctorPayloadCheck(t, payload, "backend.hooks.diagnostic", "zero.bad")
+
+	stdout.Reset()
+	stderr.Reset()
+	exitCode = runWithDeps([]string{"backends", "doctor"}, &stdout, &stderr, deps)
+	if exitCode != exitProvider {
+		t.Fatalf("text exitCode = %d, want %d stderr=%s", exitCode, exitProvider, stderr.String())
+	}
+	for _, want := range []string{"Zero backend doctor", "[fail] backend.mcp.invalid", "zero mcp add broken", "[fail] backend.hooks.diagnostic"} {
+		if !strings.Contains(stdout.String(), want) {
+			t.Fatalf("backend doctor text missing %q:\n%s", want, stdout.String())
+		}
+	}
+}
+
+func TestRunBackendsDoctorHelp(t *testing.T) {
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	exitCode := runWithDeps([]string{"backends", "doctor", "--help"}, &stdout, &stderr, appDeps{})
+	if exitCode != exitSuccess {
+		t.Fatalf("exitCode = %d stderr=%s", exitCode, stderr.String())
+	}
+	for _, want := range []string{"Usage:", "zero backends doctor", "--json"} {
+		if !strings.Contains(stdout.String(), want) {
+			t.Fatalf("backend doctor help missing %q:\n%s", want, stdout.String())
+		}
+	}
+}
+
+func TestRunBackendsDoctorDoesNotConnectOrExecuteConfiguredBackends(t *testing.T) {
+	cwd := t.TempDir()
+	hookSentinel := filepath.Join(cwd, "hook-ran")
+	pluginToolSentinel := filepath.Join(cwd, "plugin-tool-ran")
+	pluginHookSentinel := filepath.Join(cwd, "plugin-hook-ran")
+	helperCommand := os.Args[0]
+	helperArgs := func(path string) []string {
+		return []string{"-test.run=TestBackendDoctorHelperProcess", "--", "--zero-backend-doctor-sentinel", path}
+	}
+
+	var hits int32
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		atomic.AddInt32(&hits, 1)
+		w.WriteHeader(http.StatusOK)
+	}))
+	defer server.Close()
+
+	hookConfigPath := filepath.Join(cwd, ".zero", "hooks.json")
+	writeBackendDoctorJSON(t, hookConfigPath, map[string]any{
+		"enabled": true,
+		"hooks": []any{map[string]any{
+			"id":      "zero.sentinel",
+			"event":   "sessionStart",
+			"command": helperCommand,
+			"args":    helperArgs(hookSentinel),
+		}},
+	})
+	pluginDir := filepath.Join(cwd, ".zero", "plugins", "sentinel")
+	writeBackendDoctorJSON(t, filepath.Join(pluginDir, "plugin.json"), map[string]any{
+		"schemaVersion": 1,
+		"id":            "zero.sentinel",
+		"name":          "Sentinel",
+		"version":       "1.0.0",
+		"tools": []any{map[string]any{
+			"name":    "sentinel_tool",
+			"command": helperCommand,
+			"args":    helperArgs(pluginToolSentinel),
+		}},
+		"hooks": []any{map[string]any{
+			"name":    "sentinel_hook",
+			"event":   "sessionStart",
+			"command": helperCommand,
+			"args":    helperArgs(pluginHookSentinel),
+		}},
+	})
+
+	deps := appDeps{
+		getwd: func() (string, error) { return cwd, nil },
+		resolveMCPConfig: func(string) (config.MCPConfig, error) {
+			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+				"remote": {Type: "http", URL: server.URL + "/mcp"},
+			}}, nil
+		},
+		loadHooks: func(options hooks.LoadOptions) (hooks.LoadResult, error) {
+			return hooks.LoadConfig(hooks.LoadOptions{
+				Cwd:               options.Cwd,
+				UserConfigPath:    filepath.Join(cwd, "missing-user-hooks.json"),
+				ProjectConfigPath: hookConfigPath,
+			})
+		},
+		loadPlugins: func(options plugins.LoadOptions) (plugins.LoadResult, error) {
+			return plugins.Load(plugins.LoadOptions{
+				Roots: []plugins.Root{{Source: plugins.SourceProject, Path: filepath.Join(cwd, ".zero", "plugins")}},
+				Cwd:   options.Cwd,
+			})
+		},
+		registerMCPTools: func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			return nil, errors.New("zero backends doctor must not register MCP tools")
+		},
+	}
+
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	exitCode := runWithDeps([]string{"backends", "doctor", "--json"}, &stdout, &stderr, deps)
+	if exitCode != exitSuccess {
+		t.Fatalf("exitCode = %d stderr=%s stdout=%s", exitCode, stderr.String(), stdout.String())
+	}
+	if got := atomic.LoadInt32(&hits); got != 0 {
+		t.Fatalf("backend doctor connected to MCP HTTP server %d time(s)", got)
+	}
+	for _, sentinel := range []string{hookSentinel, pluginToolSentinel, pluginHookSentinel} {
+		if _, err := os.Stat(sentinel); !errors.Is(err, os.ErrNotExist) {
+			t.Fatalf("backend doctor executed configured backend command, sentinel=%s err=%v", sentinel, err)
+		}
+	}
+}
+
+func assertBackendDoctorPayloadCheck(t *testing.T, report zerocommands.BackendDoctorReport, id string, target string) {
+	t.Helper()
+	for _, check := range report.Checks {
+		if check.ID == id && check.Target == target {
+			return
+		}
+	}
+	t.Fatalf("check %s/%s not found in %#v", id, target, report.Checks)
+}
+
+func writeBackendDoctorJSON(t *testing.T, path string, value any) {
+	t.Helper()
+	if err := os.MkdirAll(filepath.Dir(path), 0o700); err != nil {
+		t.Fatalf("create dir for %s: %v", path, err)
+	}
+	data, err := json.MarshalIndent(value, "", "  ")
+	if err != nil {
+		t.Fatalf("marshal %s: %v", path, err)
+	}
+	if err := os.WriteFile(path, data, 0o600); err != nil {
+		t.Fatalf("write %s: %v", path, err)
+	}
+}
+
+func TestBackendDoctorHelperProcess(t *testing.T) {
+	for index, arg := range os.Args {
+		if arg != "--zero-backend-doctor-sentinel" || index+1 >= len(os.Args) {
+			continue
+		}
+		if err := os.WriteFile(os.Args[index+1], []byte("executed"), 0o600); err != nil {
+			os.Exit(2)
+		}
+		os.Exit(0)
+	}
+}

internal/cli/daemon.go

@@ -0,0 +1,587 @@
+package cli
+
+import (
+	"fmt"
+	"io"
+	"os"
+	"os/exec"
+	"os/signal"
+	"path/filepath"
+	"strings"
+	"syscall"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/background"
+	"github.com/Gitlawb/zero/internal/daemon"
+	"github.com/Gitlawb/zero/internal/daemon/remote"
+)
+
+// runDaemon dispatches the `zero daemon ...` subcommands. The daemon supervises a
+// pool of headless `zero exec` workers and routes sessions to them over an
+// owner-only local control socket. It is an ADDITIVE surface — interactive and
+// one-shot exec are unchanged.
+func runDaemon(args []string, stdout io.Writer, stderr io.Writer, _ appDeps) int {
+	if len(args) == 0 {
+		return writeDaemonUsage(stderr, exitUsage)
+	}
+	sub, rest := args[0], args[1:]
+	switch sub {
+	case "start":
+		return runDaemonStart(rest, stdout, stderr)
+	case "stop":
+		return runDaemonStop(rest, stdout, stderr)
+	case "status":
+		return runDaemonStatus(rest, stdout, stderr)
+	case "run":
+		return runDaemonRun(rest, stdout, stderr)
+	case "attach":
+		return runDaemonAttach(rest, stdout, stderr)
+	case "serve-remote":
+		return runDaemonServeRemote(rest, stdout, stderr)
+	case "-h", "--help", "help":
+		return writeDaemonUsage(stdout, exitSuccess)
+	default:
+		fmt.Fprintf(stderr, "unknown daemon subcommand %q\n", sub)
+		return writeDaemonUsage(stderr, exitUsage)
+	}
+}
+
+func writeDaemonUsage(w io.Writer, code int) int {
+	fmt.Fprint(w, `Usage: zero daemon <command>
+
+Commands:
+  start [--foreground]      Start the daemon (background by default).
+  stop                      Gracefully stop the running daemon.
+  status                    Show daemon / pool / session status.
+  run --session <id> [--cwd <dir>] [--prompt <text>] [exec flags...]
+                            Create/route a session and stream its output.
+  attach <session>          Attach to a running session's stream.
+  serve-remote --addr <host:port> --tls-cert <f> --tls-key <f>
+                            Serve an opt-in, TLS-only network bridge to this
+                            daemon. Requires a bearer token in $ZERO_DAEMON_REMOTE_TOKEN
+                            (or $ZERO_DAEMON_REMOTE_TOKEN_FILE).
+
+run and attach accept --remote <host:port> [--token <t>] [--ca-cert <f>]
+[--server-name <name>] to drive a remote daemon over the bridge instead of the
+local socket.
+`)
+	return code
+}
+
+// runDaemonStart starts the daemon. Without --foreground it spawns a detached
+// background process running the foreground daemon and returns once it is up.
+func runDaemonStart(args []string, stdout io.Writer, stderr io.Writer) int {
+	foreground := false
+	for _, a := range args {
+		switch a {
+		case "--foreground", "-f":
+			foreground = true
+		case "-h", "--help":
+			return writeDaemonUsage(stdout, exitSuccess)
+		default:
+			return writeExecUsageError(stderr, fmt.Sprintf("unknown flag %q for daemon start", a))
+		}
+	}
+	paths, err := daemon.DefaultPaths()
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	if foreground {
+		return runDaemonForeground(paths, stdout, stderr)
+	}
+	return runDaemonStartDetached(paths, stdout, stderr)
+}
+
+// runDaemonForeground runs the daemon in this process until SIGINT/SIGTERM. Each
+// worker is a headless `zero exec` child with the sandbox re-entrancy markers
+// scrubbed (NewExecLauncher) so it establishes its own sandbox.
+func runDaemonForeground(paths daemon.Paths, stdout io.Writer, stderr io.Writer) int {
+	launcher, err := daemon.NewExecLauncher(daemon.ExecLauncherConfig{})
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	logf := func(line string) { fmt.Fprintln(stderr, "[daemon] "+line) }
+	pool, err := daemon.NewPool(daemon.PoolOptions{Launcher: launcher, Log: logf})
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	mgr, err := daemon.NewSessionManager(daemon.SessionManagerOptions{Pool: pool})
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	srv, err := daemon.NewServer(daemon.ServerOptions{Paths: paths, Manager: mgr, Pool: pool, Log: logf})
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+
+	sigCh := make(chan os.Signal, 1)
+	signal.Notify(sigCh, os.Interrupt, syscall.SIGTERM)
+	go func() {
+		<-sigCh
+		srv.Shutdown()
+	}()
+
+	if err := srv.Serve(); err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	return exitSuccess
+}
+
+// runDaemonStartDetached spawns the foreground daemon as a detached background
+// process (its own process group, output to a log file) and waits for it to bind.
+func runDaemonStartDetached(paths daemon.Paths, stdout io.Writer, stderr io.Writer) int {
+	if daemonReachable(paths) {
+		fmt.Fprintln(stdout, "zero daemon is already running")
+		return exitSuccess
+	}
+	exe, err := os.Executable()
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	if err := os.MkdirAll(filepath.Dir(paths.Socket), 0o700); err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	logPath := filepath.Join(filepath.Dir(paths.Socket), "daemon.log")
+	logFile, err := os.OpenFile(logPath, os.O_CREATE|os.O_WRONLY|os.O_APPEND, 0o600)
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	defer logFile.Close()
+
+	cmd := exec.Command(exe, "daemon", "start", "--foreground")
+	cmd.Stdout = logFile
+	cmd.Stderr = logFile
+	background.ConfigureChildProcessGroup(cmd) // own process group: outlives this shell
+	if err := cmd.Start(); err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	_ = cmd.Process.Release()
+
+	deadline := time.Now().Add(5 * time.Second)
+	for time.Now().Before(deadline) {
+		if daemonReachable(paths) {
+			fmt.Fprintf(stdout, "zero daemon started (socket %s)\n", paths.Socket)
+			return exitSuccess
+		}
+		time.Sleep(25 * time.Millisecond)
+	}
+	return writeAppError(stderr, "daemon did not come up within timeout; see "+logPath, exitCrash)
+}
+
+func runDaemonStop(args []string, stdout io.Writer, stderr io.Writer) int {
+	if helpRequested(args) {
+		return writeDaemonUsage(stdout, exitSuccess)
+	}
+	if len(args) > 0 {
+		return writeExecUsageError(stderr, "daemon stop takes no arguments")
+	}
+	paths, err := daemon.DefaultPaths()
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	client, err := daemon.Dial(paths.Socket)
+	if err != nil {
+		fmt.Fprintln(stdout, "zero daemon is not running")
+		return exitSuccess
+	}
+	defer client.Close()
+	if err := client.Shutdown(); err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	fmt.Fprintln(stdout, "zero daemon stopped")
+	return exitSuccess
+}
+
+func runDaemonStatus(args []string, stdout io.Writer, stderr io.Writer) int {
+	if helpRequested(args) {
+		return writeDaemonUsage(stdout, exitSuccess)
+	}
+	if len(args) > 0 {
+		return writeExecUsageError(stderr, "daemon status takes no arguments")
+	}
+	paths, err := daemon.DefaultPaths()
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	client, err := daemon.Dial(paths.Socket)
+	if err != nil {
+		fmt.Fprintln(stdout, "zero daemon is not running")
+		return exitSuccess
+	}
+	defer client.Close()
+	report, err := client.Status()
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	fmt.Fprintf(stdout, "daemon pid=%d version=%d socket=%s\n", report.PID, report.Version, report.Socket)
+	fmt.Fprintf(stdout, "pool size=%d busy=%d queue=%d\n", report.PoolSize, len(report.Workers), report.QueueDepth)
+	for _, s := range report.Sessions {
+		fmt.Fprintf(stdout, "  session %s state=%s lines=%d\n", s.ID, s.State, s.Lines)
+	}
+	return exitSuccess
+}
+
+func runDaemonRun(args []string, stdout io.Writer, stderr io.Writer) int {
+	session, cwd, prompt := "", "", ""
+	var remoteFlags remoteDialFlags
+	var forward []string
+	for i := 0; i < len(args); i++ {
+		a := args[i]
+		value := func() (string, bool) {
+			if i+1 >= len(args) {
+				return "", false
+			}
+			i++
+			return args[i], true
+		}
+		switch {
+		case a == "-h" || a == "--help":
+			return writeDaemonUsage(stdout, exitSuccess)
+		case a == "--remote":
+			v, ok := value()
+			if !ok {
+				return writeExecUsageError(stderr, "--remote requires a value")
+			}
+			remoteFlags.Addr = v
+		case strings.HasPrefix(a, "--remote="):
+			remoteFlags.Addr = strings.TrimPrefix(a, "--remote=")
+		case a == "--token":
+			v, ok := value()
+			if !ok {
+				return writeExecUsageError(stderr, "--token requires a value")
+			}
+			remoteFlags.Token = v
+		case strings.HasPrefix(a, "--token="):
+			remoteFlags.Token = strings.TrimPrefix(a, "--token=")
+		case a == "--ca-cert":
+			v, ok := value()
+			if !ok {
+				return writeExecUsageError(stderr, "--ca-cert requires a value")
+			}
+			remoteFlags.CACert = v
+		case strings.HasPrefix(a, "--ca-cert="):
+			remoteFlags.CACert = strings.TrimPrefix(a, "--ca-cert=")
+		case a == "--server-name":
+			v, ok := value()
+			if !ok {
+				return writeExecUsageError(stderr, "--server-name requires a value")
+			}
+			remoteFlags.ServerName = v
+		case strings.HasPrefix(a, "--server-name="):
+			remoteFlags.ServerName = strings.TrimPrefix(a, "--server-name=")
+		case a == "--session":
+			v, ok := value()
+			if !ok {
+				return writeExecUsageError(stderr, "--session requires a value")
+			}
+			session = v
+		case strings.HasPrefix(a, "--session="):
+			session = strings.TrimPrefix(a, "--session=")
+		case a == "--cwd":
+			v, ok := value()
+			if !ok {
+				return writeExecUsageError(stderr, "--cwd requires a value")
+			}
+			cwd = v
+		case strings.HasPrefix(a, "--cwd="):
+			cwd = strings.TrimPrefix(a, "--cwd=")
+		case a == "--prompt" || a == "-p":
+			v, ok := value()
+			if !ok {
+				return writeExecUsageError(stderr, "--prompt requires a value")
+			}
+			prompt = v
+		case strings.HasPrefix(a, "--prompt="):
+			prompt = strings.TrimPrefix(a, "--prompt=")
+		default:
+			// Forwarded verbatim to the worker `zero exec` (reuses exec's own flag
+			// parsing for per-session run options).
+			forward = append(forward, a)
+		}
+	}
+	if strings.TrimSpace(session) == "" {
+		return writeExecUsageError(stderr, "daemon run requires --session <id>")
+	}
+	if prompt == "" && len(forward) == 0 {
+		return writeExecUsageError(stderr, "daemon run requires --prompt <text> or exec args")
+	}
+	client, err := dialForCLI(remoteFlags)
+	if err != nil {
+		return writeAppError(stderr, daemonDialError(remoteFlags, err), exitCrash)
+	}
+	defer client.Close()
+	if err := client.Run(session, cwd, prompt, forward, func(line string) { fmt.Fprintln(stdout, line) }); err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	return exitSuccess
+}
+
+func runDaemonAttach(args []string, stdout io.Writer, stderr io.Writer) int {
+	session := ""
+	var remoteFlags remoteDialFlags
+	extra := 0
+	for i := 0; i < len(args); i++ {
+		a := args[i]
+		value := func() (string, bool) {
+			if i+1 >= len(args) {
+				return "", false
+			}
+			i++
+			return args[i], true
+		}
+		switch {
+		case a == "-h" || a == "--help":
+			return writeDaemonUsage(stdout, exitSuccess)
+		case a == "--remote":
+			v, ok := value()
+			if !ok {
+				return writeExecUsageError(stderr, "--remote requires a value")
+			}
+			remoteFlags.Addr = v
+		case strings.HasPrefix(a, "--remote="):
+			remoteFlags.Addr = strings.TrimPrefix(a, "--remote=")
+		case a == "--token":
+			v, ok := value()
+			if !ok {
+				return writeExecUsageError(stderr, "--token requires a value")
+			}
+			remoteFlags.Token = v
+		case strings.HasPrefix(a, "--token="):
+			remoteFlags.Token = strings.TrimPrefix(a, "--token=")
+		case a == "--ca-cert":
+			v, ok := value()
+			if !ok {
+				return writeExecUsageError(stderr, "--ca-cert requires a value")
+			}
+			remoteFlags.CACert = v
+		case strings.HasPrefix(a, "--ca-cert="):
+			remoteFlags.CACert = strings.TrimPrefix(a, "--ca-cert=")
+		case a == "--server-name":
+			v, ok := value()
+			if !ok {
+				return writeExecUsageError(stderr, "--server-name requires a value")
+			}
+			remoteFlags.ServerName = v
+		case strings.HasPrefix(a, "--server-name="):
+			remoteFlags.ServerName = strings.TrimPrefix(a, "--server-name=")
+		case session == "":
+			session = a
+		default:
+			extra++
+		}
+	}
+	if strings.TrimSpace(session) == "" {
+		return writeExecUsageError(stderr, "daemon attach requires a <session> id")
+	}
+	if extra > 0 {
+		return writeExecUsageError(stderr, "daemon attach accepts exactly one <session> id")
+	}
+	client, err := dialForCLI(remoteFlags)
+	if err != nil {
+		return writeAppError(stderr, daemonDialError(remoteFlags, err), exitCrash)
+	}
+	defer client.Close()
+	if err := client.Attach(session, func(line string) { fmt.Fprintln(stdout, line) }); err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	return exitSuccess
+}
+
+// daemonDialError tailors the connect-failure message for local vs remote.
+func daemonDialError(flags remoteDialFlags, err error) string {
+	if strings.TrimSpace(flags.Addr) != "" {
+		return "failed to reach remote daemon at " + flags.Addr + ": " + err.Error()
+	}
+	return "zero daemon is not running (start it with `zero daemon start`)"
+}
+
+// runDaemonServeRemote starts the local daemon plus an opt-in, TLS-only network
+// bridge. TLS and a bearer token are mandatory (fail closed): it refuses to
+// start without a cert/key pair and a token from the environment.
+func runDaemonServeRemote(args []string, stdout io.Writer, stderr io.Writer) int {
+	addr, certFile, keyFile := "", "", ""
+	minVersion, maxConns := 0, 0
+	for i := 0; i < len(args); i++ {
+		a := args[i]
+		value := func() (string, bool) {
+			if i+1 >= len(args) {
+				return "", false
+			}
+			i++
+			return args[i], true
+		}
+		switch {
+		case a == "-h" || a == "--help":
+			return writeDaemonUsage(stdout, exitSuccess)
+		case a == "--addr":
+			v, ok := value()
+			if !ok {
+				return writeExecUsageError(stderr, "--addr requires a value")
+			}
+			addr = v
+		case strings.HasPrefix(a, "--addr="):
+			addr = strings.TrimPrefix(a, "--addr=")
+		case a == "--tls-cert":
+			v, ok := value()
+			if !ok {
+				return writeExecUsageError(stderr, "--tls-cert requires a value")
+			}
+			certFile = v
+		case strings.HasPrefix(a, "--tls-cert="):
+			certFile = strings.TrimPrefix(a, "--tls-cert=")
+		case a == "--tls-key":
+			v, ok := value()
+			if !ok {
+				return writeExecUsageError(stderr, "--tls-key requires a value")
+			}
+			keyFile = v
+		case strings.HasPrefix(a, "--tls-key="):
+			keyFile = strings.TrimPrefix(a, "--tls-key=")
+		case a == "--min-version":
+			v, ok := value()
+			if !ok {
+				return writeExecUsageError(stderr, "--min-version requires a value")
+			}
+			minVersion = atoiOrZero(v)
+		case a == "--max-conns":
+			v, ok := value()
+			if !ok {
+				return writeExecUsageError(stderr, "--max-conns requires a value")
+			}
+			maxConns = atoiOrZero(v)
+		default:
+			return writeExecUsageError(stderr, fmt.Sprintf("unknown flag %q for daemon serve-remote", a))
+		}
+	}
+	if strings.TrimSpace(addr) == "" {
+		return writeExecUsageError(stderr, "daemon serve-remote requires --addr <host:port>")
+	}
+	// Fail closed: TLS cert/key + a bearer token are mandatory.
+	tlsConfig, err := remote.ServerTLSConfig(certFile, keyFile)
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	token, err := remote.TokenFromEnv()
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	auth, err := remote.NewTokenAuthenticator(token)
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	paths, err := daemon.DefaultPaths()
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	launcher, err := daemon.NewExecLauncher(daemon.ExecLauncherConfig{})
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	logf := func(line string) { fmt.Fprintln(stderr, "[daemon] "+line) }
+	pool, err := daemon.NewPool(daemon.PoolOptions{Launcher: launcher, Log: logf})
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	mgr, err := daemon.NewSessionManager(daemon.SessionManagerOptions{Pool: pool})
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	srv, err := daemon.NewServer(daemon.ServerOptions{Paths: paths, Manager: mgr, Pool: pool, Log: logf})
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	bridge, err := remote.NewBridge(remote.BridgeOptions{
+		Server: srv, Authenticator: auth, MinVersion: minVersion, MaxConnections: maxConns, Log: logf,
+	})
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+
+	// Serve the local control socket too, so local clients keep working.
+	go func() {
+		if serveErr := srv.Serve(); serveErr != nil {
+			logf("local serve error: " + serveErr.Error())
+		}
+	}()
+
+	serveErr := make(chan error, 1)
+	go func() { serveErr <- bridge.ListenAndServeTLS(addr, tlsConfig) }()
+
+	sigCh := make(chan os.Signal, 1)
+	signal.Notify(sigCh, os.Interrupt, syscall.SIGTERM)
+
+	fmt.Fprintf(stdout, "zero daemon remote bridge listening on %s (TLS)\n", addr)
+	select {
+	case <-sigCh:
+		srv.Shutdown()
+		_ = bridge.Close()
+		<-serveErr // wait for the accept loop to unwind
+		return exitSuccess
+	case err := <-serveErr:
+		// Bind/serve failed before any signal (e.g. address in use).
+		srv.Shutdown()
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+}
+
+// remoteDialFlags holds the optional flags that redirect run/attach to a remote
+// daemon. When Addr is empty the local control socket is used.
+type remoteDialFlags struct {
+	Addr       string
+	Token      string
+	CACert     string
+	ServerName string
+}
+
+// dialForCLI returns a daemon.Client for either the local socket (Addr empty) or
+// a remote bridge (Addr set). For remote, the token falls back to the env.
+func dialForCLI(flags remoteDialFlags) (*daemon.Client, error) {
+	if strings.TrimSpace(flags.Addr) == "" {
+		paths, err := daemon.DefaultPaths()
+		if err != nil {
+			return nil, err
+		}
+		return daemon.Dial(paths.Socket)
+	}
+	token := strings.TrimSpace(flags.Token)
+	if token == "" {
+		token, _ = remote.TokenFromEnv() // best effort; DialRemote rejects an empty token
+	}
+	return remote.DialRemote(remote.RemoteConfig{
+		Address:    flags.Addr,
+		Token:      token,
+		CACertFile: flags.CACert,
+		ServerName: flags.ServerName,
+	})
+}
+
+func atoiOrZero(s string) int {
+	n := 0
+	for _, r := range strings.TrimSpace(s) {
+		if r < '0' || r > '9' {
+			return 0
+		}
+		n = n*10 + int(r-'0')
+	}
+	return n
+}
+
+// daemonReachable reports whether a daemon is accepting connections (a successful
+// handshake). Used for the single-instance check and start-up wait.
+func daemonReachable(paths daemon.Paths) bool {
+	client, err := daemon.Dial(paths.Socket)
+	if err != nil {
+		return false
+	}
+	_ = client.Close()
+	return true
+}
+
+func helpRequested(args []string) bool {
+	for _, a := range args {
+		if a == "-h" || a == "--help" {
+			return true
+		}
+	}
+	return false
+}

internal/cli/daemon_test.go

@@ -0,0 +1,123 @@
+package cli
+
+import (
+	"bytes"
+	"strings"
+	"testing"
+)
+
+// isolateDaemonPaths points DefaultPaths at a temp dir so the test never touches
+// a real daemon on the dev machine.
+func isolateDaemonPaths(t *testing.T) {
+	t.Helper()
+	t.Setenv("XDG_RUNTIME_DIR", t.TempDir())
+}
+
+func runDaemonCLI(t *testing.T, args ...string) (int, string, string) {
+	t.Helper()
+	var out, errb bytes.Buffer
+	code := runDaemon(args, &out, &errb, appDeps{})
+	return code, out.String(), errb.String()
+}
+
+func TestDaemonUsage(t *testing.T) {
+	code, _, _ := runDaemonCLI(t)
+	if code != exitUsage {
+		t.Fatalf("no-args exit = %d, want exitUsage", code)
+	}
+	code, out, _ := runDaemonCLI(t, "--help")
+	if code != exitSuccess || !strings.Contains(out, "Usage: zero daemon") {
+		t.Fatalf("--help exit=%d out=%q", code, out)
+	}
+}
+
+func TestDaemonUnknownSubcommand(t *testing.T) {
+	code, _, errb := runDaemonCLI(t, "frobnicate")
+	if code != exitUsage {
+		t.Fatalf("unknown subcommand exit = %d, want exitUsage", code)
+	}
+	if !strings.Contains(errb, "unknown daemon subcommand") {
+		t.Fatalf("stderr = %q, want unknown-subcommand message", errb)
+	}
+}
+
+func TestDaemonRunRequiresSession(t *testing.T) {
+	isolateDaemonPaths(t)
+	code, _, errb := runDaemonCLI(t, "run", "--prompt", "hi")
+	if code != exitUsage {
+		t.Fatalf("run without --session exit = %d, want exitUsage", code)
+	}
+	if !strings.Contains(errb, "--session") {
+		t.Fatalf("stderr = %q, want a --session hint", errb)
+	}
+}
+
+func TestDaemonRunRequiresPromptOrArgs(t *testing.T) {
+	isolateDaemonPaths(t)
+	code, _, errb := runDaemonCLI(t, "run", "--session", "s1")
+	if code != exitUsage {
+		t.Fatalf("run without prompt/args exit = %d, want exitUsage", code)
+	}
+	if !strings.Contains(errb, "--prompt") {
+		t.Fatalf("stderr = %q, want a --prompt hint", errb)
+	}
+}
+
+func TestDaemonStopWhenNotRunning(t *testing.T) {
+	isolateDaemonPaths(t)
+	code, out, _ := runDaemonCLI(t, "stop")
+	if code != exitSuccess {
+		t.Fatalf("stop (not running) exit = %d, want exitSuccess", code)
+	}
+	if !strings.Contains(out, "not running") {
+		t.Fatalf("stop output = %q, want 'not running'", out)
+	}
+}
+
+func TestDaemonStatusWhenNotRunning(t *testing.T) {
+	isolateDaemonPaths(t)
+	code, out, _ := runDaemonCLI(t, "status")
+	if code != exitSuccess {
+		t.Fatalf("status (not running) exit = %d, want exitSuccess", code)
+	}
+	if !strings.Contains(out, "not running") {
+		t.Fatalf("status output = %q, want 'not running'", out)
+	}
+}
+
+func TestDaemonAttachRequiresSession(t *testing.T) {
+	isolateDaemonPaths(t)
+	code, _, errb := runDaemonCLI(t, "attach")
+	if code != exitUsage {
+		t.Fatalf("attach without session exit = %d, want exitUsage", code)
+	}
+	if !strings.Contains(errb, "session") {
+		t.Fatalf("stderr = %q, want a session hint", errb)
+	}
+}
+
+func TestDaemonRunWhenNotRunning(t *testing.T) {
+	isolateDaemonPaths(t)
+	code, _, errb := runDaemonCLI(t, "run", "--session", "s1", "--prompt", "hello")
+	if code != exitCrash {
+		t.Fatalf("run (no daemon) exit = %d, want exitCrash", code)
+	}
+	if !strings.Contains(errb, "not running") {
+		t.Fatalf("stderr = %q, want 'not running'", errb)
+	}
+}
+
+func TestDaemonSubcommandsRejectExtraArgs(t *testing.T) {
+	isolateDaemonPaths(t)
+	cases := [][]string{
+		{"stop", "oops"},
+		{"status", "oops"},
+		{"attach", "s1", "extra"},
+	}
+	for _, args := range cases {
+		code, _, errb := runDaemonCLI(t, args...)
+		if code != exitUsage {
+			t.Fatalf("%v exit = %d, want exitUsage (reject extra args); stderr=%q", args, code, errb)
+		}
+	}
+}

internal/cli/exec.go

@@ -19,7 +19,6 @@ import (
 	"github.com/Gitlawb/zero/internal/providers"
 	"github.com/Gitlawb/zero/internal/sandbox"
 	"github.com/Gitlawb/zero/internal/sessions"
-	"github.com/Gitlawb/zero/internal/specialist"
 	"github.com/Gitlawb/zero/internal/specmode"
 	"github.com/Gitlawb/zero/internal/streamjson"
 	"github.com/Gitlawb/zero/internal/tools"
@@ -163,7 +162,7 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 	if options.allowEscalation {
 		registry.Register(tools.NewEscalateModelTool())
 	}
-	var specialistRuntime *specialist.Runtime
+	var specialistRuntime *agentToolRuntime
 	if shouldRegisterExecSpecialistTools(options) {
 		var err error
 		specialistRuntime, err = registerSpecialistTools(registry, workspaceRoot)
@@ -704,7 +703,8 @@ func buildExecSandboxEngine(workspaceRoot string, resolved config.ResolvedConfig
 // depth — it guarantees that even a direct/programmatic caller passing a typo
 // can never WIDEN the ceiling back to the High default.
 // applyConfiguredSandboxPolicy overlays every config-sourced sandbox knob onto
-// the default policy: the autonomy ceiling and the network mode.
+// the default policy (the autonomy ceiling and the network mode), plus the
+// ZERO_SANDBOX_AUTO_ALLOW_BASH environment opt-in.
 func applyConfiguredSandboxPolicy(policy sandbox.Policy, cfg config.SandboxConfig) sandbox.Policy {
 	policy = applyConfiguredAutonomyCeiling(policy, cfg.MaxAutonomy)
 	if network := strings.TrimSpace(cfg.Network); network != "" {
@@ -721,6 +721,12 @@ func applyConfiguredSandboxPolicy(policy sandbox.Policy, cfg config.SandboxConfi
 	if cfg.MonitorDenials {
 		policy.MonitorDenials = true
 	}
+	// ZERO_SANDBOX_AUTO_ALLOW_BASH opts into auto-allowing a *sandboxed* shell
+	// command without a prompt — the active sandbox is the safety boundary, and
+	// the engine only honors this when the shell sandbox is actually active. Like
+	// the flags above it only ever turns the opt-in ON; unset/blank stays off
+	// (prompt), and an explicit config opt-in is preserved.
+	policy = sandbox.ApplyAutoAllowBashEnv(policy)
 	return policy
 }
 

internal/cli/sandbox_test.go

@@ -662,6 +662,19 @@ func TestApplyConfiguredSandboxPolicyHardeningFlags(t *testing.T) {
 	}
 }
 
+func TestApplyConfiguredSandboxPolicyHonorsAutoAllowBashEnv(t *testing.T) {
+	// Unset/blank: the opt-in stays off so a sandboxed bash command still prompts.
+	t.Setenv(sandbox.EnvAutoAllowBash, "")
+	if got := applyConfiguredSandboxPolicy(sandbox.DefaultPolicy(), config.SandboxConfig{}); got.AutoAllowBashWhenSandboxed {
+		t.Fatal("auto-allow-bash must be off when ZERO_SANDBOX_AUTO_ALLOW_BASH is unset")
+	}
+	// Truthy: the env opt-in is wired through the central policy builder.
+	t.Setenv(sandbox.EnvAutoAllowBash, "true")
+	if got := applyConfiguredSandboxPolicy(sandbox.DefaultPolicy(), config.SandboxConfig{}); !got.AutoAllowBashWhenSandboxed {
+		t.Fatal("ZERO_SANDBOX_AUTO_ALLOW_BASH=true must enable AutoAllowBashWhenSandboxed")
+	}
+}
+
 func newSandboxTestStore(t *testing.T) *sandbox.GrantStore {
 	t.Helper()
 	store, err := sandbox.NewGrantStore(sandbox.StoreOptions{

internal/config/resolver_test.go

@@ -1094,7 +1094,8 @@ func TestResolveProviderProfileExtendedJSONAliases(t *testing.T) {
 			"auth_scheme": "raw",
 			"auth_header_value": "header-secret",
 			"custom_headers": {"X-Zero": "1"},
-			"model_id": "custom-model"
+			"model_id": "custom-model",
+			"parse_think_tags": true
 		}]
 	}`)
 
@@ -1119,6 +1120,30 @@ func TestResolveProviderProfileExtendedJSONAliases(t *testing.T) {
 	if profile.CustomHeaders["X-Zero"] != "1" {
 		t.Fatalf("CustomHeaders = %#v, want X-Zero header", profile.CustomHeaders)
 	}
+	if profile.ParseThinkTags == nil || !*profile.ParseThinkTags {
+		t.Fatalf("ParseThinkTags = %#v, want true", profile.ParseThinkTags)
+	}
+}
+
+func TestResolveProviderProfileParseThinkTagsFalseAlias(t *testing.T) {
+	path := writeConfig(t, `{
+		"activeProvider": "custom",
+		"providers": [{
+			"name": "custom",
+			"provider_kind": "openai-compatible",
+			"base_url": "https://custom.example/v1",
+			"model_id": "custom-model",
+			"parse_think_tags": false
+		}]
+	}`)
+
+	resolved, err := Resolve(ResolveOptions{UserConfigPath: path, Env: map[string]string{}})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+	if resolved.Provider.ParseThinkTags == nil || *resolved.Provider.ParseThinkTags {
+		t.Fatalf("ParseThinkTags = %#v, want explicit false", resolved.Provider.ParseThinkTags)
+	}
 }
 
 func TestResolveRejectsUnknownProviderCatalogID(t *testing.T) {

internal/config/types.go

@@ -34,6 +34,7 @@ type ProviderProfile struct {
 	AuthHeaderValue string            `json:"authHeaderValue,omitempty"`
 	CustomHeaders   map[string]string `json:"customHeaders,omitempty"`
 	Model           string            `json:"model,omitempty"`
+	ParseThinkTags  *bool             `json:"parseThinkTags,omitempty"`
 	Description     string            `json:"description,omitempty"`
 }
 
@@ -51,6 +52,7 @@ func HasProviderProfile(profile ProviderProfile) bool {
 		strings.TrimSpace(profile.AuthHeaderValue) != "" ||
 		profile.CustomHeaders != nil ||
 		strings.TrimSpace(profile.Model) != "" ||
+		profile.ParseThinkTags != nil ||
 		strings.TrimSpace(profile.Description) != ""
 }
 
@@ -304,6 +306,8 @@ func (profile *ProviderProfile) UnmarshalJSON(data []byte) error {
 		CustomHeadersSnake   map[string]string `json:"custom_headers"`
 		Model                string            `json:"model"`
 		ModelID              string            `json:"model_id"`
+		ParseThinkTags       *bool             `json:"parseThinkTags"`
+		ParseThinkTagsSnake  *bool             `json:"parse_think_tags"`
 		Description          string            `json:"description"`
 	}
 
@@ -325,6 +329,7 @@ func (profile *ProviderProfile) UnmarshalJSON(data []byte) error {
 	profile.AuthHeaderValue = firstNonEmpty(raw.AuthHeaderValue, raw.AuthHeaderValueSnake)
 	profile.CustomHeaders = firstNonNilMap(raw.CustomHeaders, raw.CustomHeadersSnake)
 	profile.Model = strings.TrimSpace(firstNonEmpty(raw.Model, raw.ModelID))
+	profile.ParseThinkTags = firstNonNilBool(raw.ParseThinkTags, raw.ParseThinkTagsSnake)
 	profile.Description = strings.TrimSpace(raw.Description)
 	return nil
 }
@@ -348,6 +353,15 @@ func firstNonNilMap(values ...map[string]string) map[string]string {
 	return nil
 }
 
+func firstNonNilBool(values ...*bool) *bool {
+	for _, value := range values {
+		if value != nil {
+			return value
+		}
+	}
+	return nil
+}
+
 func copyStringMap(values map[string]string) map[string]string {
 	if values == nil {
 		return nil

internal/daemon/backoff.go

@@ -0,0 +1,41 @@
+package daemon
+
+import (
+	"math/rand"
+	"time"
+)
+
+// Backoff bounds. Mirrors reference-daemon-code-agent-js/worker-manager.js:
+// backoffMs(consecutive) = jitter(min(1000 * 2**consecutive, 300000)).
+const (
+	backoffUnit = time.Second
+	backoffCap  = 5 * time.Minute
+)
+
+// backoffBase returns the un-jittered exponential delay for a 1-based restart
+// count: min(1s * 2^attempt, 5m). attempt<=0 yields the unit delay.
+func backoffBase(attempt int) time.Duration {
+	if attempt <= 0 {
+		return backoffUnit
+	}
+	// Shift carefully: cap before overflow. 2^attempt * 1s exceeds the cap well
+	// before any int64 overflow (cap is 5m), so clamp once it would exceed it.
+	d := backoffUnit
+	for i := 0; i < attempt; i++ {
+		d *= 2
+		if d >= backoffCap {
+			return backoffCap
+		}
+	}
+	return d
+}
+
+// BackoffWithJitter returns backoffBase(attempt) scaled by a uniform random
+// factor in [0.5, 1.5), matching worker-manager.js jitter(base) =
+// base*(0.5 + random()). The jitter de-synchronizes simultaneous restarts so a
+// fleet of workers does not thunder back at the same instant.
+func BackoffWithJitter(attempt int) time.Duration {
+	base := backoffBase(attempt)
+	factor := 0.5 + rand.Float64() //nolint:gosec // jitter, not security-sensitive
+	return time.Duration(float64(base) * factor)
+}

internal/daemon/backoff_test.go

@@ -0,0 +1,38 @@
+package daemon
+
+import (
+	"testing"
+	"time"
+)
+
+func TestBackoffBaseExponentialAndCap(t *testing.T) {
+	cases := []struct {
+		attempt int
+		want    time.Duration
+	}{
+		{0, 1 * time.Second},
+		{1, 2 * time.Second},
+		{2, 4 * time.Second},
+		{3, 8 * time.Second},
+		{20, backoffCap}, // far past the cap
+	}
+	for _, tc := range cases {
+		if got := backoffBase(tc.attempt); got != tc.want {
+			t.Fatalf("backoffBase(%d) = %s, want %s", tc.attempt, got, tc.want)
+		}
+	}
+}
+
+func TestBackoffWithJitterBounds(t *testing.T) {
+	for attempt := 1; attempt <= 6; attempt++ {
+		base := backoffBase(attempt)
+		lo := time.Duration(float64(base) * 0.5)
+		hi := time.Duration(float64(base) * 1.5)
+		for i := 0; i < 200; i++ {
+			got := BackoffWithJitter(attempt)
+			if got < lo || got >= hi {
+				t.Fatalf("BackoffWithJitter(%d) = %s, want in [%s,%s)", attempt, got, lo, hi)
+			}
+		}
+	}
+}

internal/daemon/client.go

@@ -0,0 +1,130 @@
+package daemon
+
+import (
+	"fmt"
+	"net"
+	"time"
+)
+
+// Client is a control-socket client used by the `zero daemon run|attach|status|
+// stop` subcommands. It performs the version handshake on Dial.
+type Client struct {
+	conn net.Conn
+}
+
+// Dial connects to the daemon control socket and completes the handshake.
+func Dial(socketPath string) (*Client, error) {
+	conn, err := net.DialTimeout("unix", socketPath, 5*time.Second)
+	if err != nil {
+		return nil, err
+	}
+	c := &Client{conn: conn}
+	if err := c.handshake(); err != nil {
+		_ = conn.Close()
+		return nil, err
+	}
+	return c, nil
+}
+
+// NewClientConn completes the control handshake on an already-connected
+// transport (e.g. a TLS connection the remote bridge has authenticated) and
+// returns a Client ready for Run/Attach/Status. It takes ownership of conn and
+// closes it on handshake failure.
+func NewClientConn(conn net.Conn) (*Client, error) {
+	c := &Client{conn: conn}
+	if err := c.handshake(); err != nil {
+		_ = conn.Close()
+		return nil, err
+	}
+	return c, nil
+}
+
+func (c *Client) handshake() error {
+	if err := WriteControl(c.conn, Ctrl{Type: CtrlHello, Version: ProtoVersion}); err != nil {
+		return err
+	}
+	reply, err := ReadControl(c.conn)
+	if err != nil {
+		return err
+	}
+	if reply.Type != CtrlHelloOK {
+		return fmt.Errorf("daemon: handshake rejected: %s", reply.Message)
+	}
+	return nil
+}
+
+// Close closes the connection.
+func (c *Client) Close() error { return c.conn.Close() }
+
+// Run starts/routes a session and streams its stream-json output lines to onLine
+// until the session ends. It returns the session's terminal error, if any.
+func (c *Client) Run(session, cwd, prompt string, args []string, onLine func(string)) error {
+	if err := WriteControl(c.conn, Ctrl{Type: CtrlRun, Session: session, Cwd: cwd, Prompt: prompt, Args: args}); err != nil {
+		return err
+	}
+	return c.streamLoop(onLine)
+}
+
+// Attach streams a running session's output (buffered history + live) to onLine.
+func (c *Client) Attach(session string, onLine func(string)) error {
+	if err := WriteControl(c.conn, Ctrl{Type: CtrlAttach, Session: session}); err != nil {
+		return err
+	}
+	return c.streamLoop(onLine)
+}
+
+func (c *Client) streamLoop(onLine func(string)) error {
+	for {
+		msg, err := ReadControl(c.conn)
+		if err != nil {
+			return err
+		}
+		switch msg.Type {
+		case CtrlAck:
+			// dispatch acknowledged; keep reading
+		case CtrlData:
+			if onLine != nil {
+				onLine(msg.Line)
+			}
+		case CtrlEnd:
+			return nil
+		case CtrlError:
+			return fmt.Errorf("daemon: %s", msg.Message)
+		default:
+			return fmt.Errorf("daemon: unexpected message %q", msg.Type)
+		}
+	}
+}
+
+// Status requests the daemon/worker/session status report.
+func (c *Client) Status() (*StatusReport, error) {
+	if err := WriteControl(c.conn, Ctrl{Type: CtrlStatus}); err != nil {
+		return nil, err
+	}
+	msg, err := ReadControl(c.conn)
+	if err != nil {
+		return nil, err
+	}
+	if msg.Type == CtrlError {
+		return nil, fmt.Errorf("daemon: %s", msg.Message)
+	}
+	if msg.Type != CtrlStatusResult || msg.Status == nil {
+		return nil, fmt.Errorf("daemon: unexpected status reply %q", msg.Type)
+	}
+	return msg.Status, nil
+}
+
+// Shutdown asks the daemon to drain and stop.
+func (c *Client) Shutdown() error {
+	if err := WriteControl(c.conn, Ctrl{Type: CtrlShutdown}); err != nil {
+		return err
+	}
+	msg, err := ReadControl(c.conn)
+	if err != nil {
+		return err
+	}
+	if msg.Type == CtrlError {
+		return fmt.Errorf("daemon: %s", msg.Message)
+	}
+	return nil
+}

internal/daemon/launcher.go

@@ -0,0 +1,165 @@
+package daemon
+
+import (
+	"bufio"
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"os/exec"
+	"strings"
+
+	"github.com/Gitlawb/zero/internal/background"
+	"github.com/Gitlawb/zero/internal/sandbox"
+)
+
+// reentrancyMarkers are the env vars zero sets on a command it has ALREADY
+// wrapped in a sandbox. The daemon MUST strip them from every worker's
+// environment: a worker that inherited them would trip zero's re-entrancy guard
+// (sandbox.IsAlreadySandboxed needs BOTH) and run with a pass-through plan —
+// i.e. UNSANDBOXED — silently defeating the sandbox. Stripping them forces each
+// worker to (re)establish its own sandbox. This is the daemon's #1 security
+// invariant: it must NOT bypass the sandbox.
+var reentrancyMarkers = []string{sandbox.EnvSandboxed, sandbox.EnvSandboxBackend}
+
+// scrubWorkerEnv returns env with the sandbox re-entrancy markers removed. Every
+// other variable is preserved — including the ZERO_SANDBOX_* policy config the
+// worker reads to rebuild its own policy — so the sandbox policy is PROPAGATED,
+// not bypassed. The caller's slice is never mutated.
+func scrubWorkerEnv(env []string) []string {
+	out := make([]string, 0, len(env))
+	for _, kv := range env {
+		drop := false
+		for _, marker := range reentrancyMarkers {
+			if strings.HasPrefix(kv, marker+"=") {
+				drop = true
+				break
+			}
+		}
+		if !drop {
+			out = append(out, kv)
+		}
+	}
+	return out
+}
+
+// execWorker is a WorkerHandle backed by a `zero exec` child process speaking
+// stream-json on stdout.
+type execWorker struct {
+	cmd    *exec.Cmd
+	stdout io.ReadCloser
+	lines  *scannerLines
+	pid    int
+}
+
+func (w *execWorker) Stdout() Lines { return w.lines }
+func (w *execWorker) Pid() int      { return w.pid }
+
+func (w *execWorker) Wait() (int, error) {
+	err := w.cmd.Wait()
+	if err == nil {
+		return 0, nil
+	}
+	var exitErr *exec.ExitError
+	if errors.As(err, &exitErr) {
+		return exitErr.ExitCode(), nil
+	}
+	return -1, err
+}
+
+func (w *execWorker) Kill() error {
+	if w.cmd.Process == nil {
+		return nil
+	}
+	// background.TerminateProcess is the cross-platform terminate (kills the
+	// process group on POSIX, taskkill /T on Windows).
+	return background.TerminateProcess(w.cmd.Process.Pid)
+}
+
+// scannerLines adapts a bufio.Scanner to the Lines interface.
+type scannerLines struct {
+	sc *bufio.Scanner
+}
+
+func (s *scannerLines) Next() (string, bool, error) {
+	if s.sc.Scan() {
+		return s.sc.Text(), true, nil
+	}
+	if err := s.sc.Err(); err != nil {
+		return "", false, err
+	}
+	return "", false, nil
+}
+
+// ExecLauncherConfig configures the production worker launcher.
+type ExecLauncherConfig struct {
+	// Executable is the zero binary to spawn (defaults to os.Executable()).
+	Executable string
+	// BaseArgs are prepended before the per-session args (defaults to the headless
+	// stream-json exec invocation `exec --output-format stream-json`).
+	BaseArgs []string
+	// Env is the base environment for workers (defaults to os.Environ()). The
+	// re-entrancy markers are always scrubbed from it.
+	Env []string
+}
+
+// NewExecLauncher builds a Launcher that spawns `zero exec` workers which speak
+// stream-json on stdout. Each worker:
+//   - runs with the re-entrancy markers SCRUBBED from its env (so it establishes
+//     its own sandbox; the daemon never bypasses the sandbox), while the rest of
+//     the policy config/env is propagated;
+//   - is placed in its own process group (background.ConfigureChildProcessGroup)
+//     so Kill / drain can terminate it and any children cleanly;
+//   - has its per-session working directory (spec.Cwd) and flags (spec.Args)
+//     applied. The pool does the os/exec itself and uses internal/background only
+//     for process-group setup + cross-platform terminate.
+func NewExecLauncher(cfg ExecLauncherConfig) (Launcher, error) {
+	exe := cfg.Executable
+	if exe == "" {
+		resolved, err := os.Executable()
+		if err != nil {
+			return nil, fmt.Errorf("daemon: locate zero executable: %w", err)
+		}
+		exe = resolved
+	}
+	baseArgs := cfg.BaseArgs
+	if baseArgs == nil {
+		baseArgs = []string{"exec", "--output-format", "stream-json"}
+	}
+	baseEnv := cfg.Env
+	if baseEnv == nil {
+		baseEnv = os.Environ()
+	}
+	env := scrubWorkerEnv(baseEnv)
+
+	return func(ctx context.Context, spec WorkerSpec) (WorkerHandle, error) {
+		args := make([]string, 0, len(baseArgs)+len(spec.Args))
+		args = append(args, baseArgs...)
+		args = append(args, spec.Args...)
+
+		cmd := exec.CommandContext(ctx, exe, args...)
+		cmd.Dir = spec.Cwd
+		cmd.Env = env
+		cmd.Stdin = nil // the prompt is passed via flags; no streamed stdin input
+		background.ConfigureChildProcessGroup(cmd)
+
+		stdout, err := cmd.StdoutPipe()
+		if err != nil {
+			return nil, fmt.Errorf("daemon: worker stdout pipe: %w", err)
+		}
+		if err := cmd.Start(); err != nil {
+			return nil, fmt.Errorf("daemon: start worker: %w", err)
+		}
+
+		scanner := bufio.NewScanner(stdout)
+		// Allow long stream-json lines up to the control-frame cap.
+		scanner.Buffer(make([]byte, 0, 64*1024), MaxFrameSize)
+		return &execWorker{
+			cmd:    cmd,
+			stdout: stdout,
+			lines:  &scannerLines{sc: scanner},
+			pid:    cmd.Process.Pid,
+		}, nil
+	}, nil
+}

internal/daemon/launcher_test.go

@@ -0,0 +1,102 @@
+package daemon
+
+import (
+	"context"
+	"runtime"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/sandbox"
+)
+
+func TestScrubWorkerEnvRemovesReentrancyMarkers(t *testing.T) {
+	in := []string{
+		"PATH=/bin",
+		sandbox.EnvSandboxed + "=1",
+		"ZERO_SANDBOX_AUTO_ALLOW_BASH=1", // policy config — must be PRESERVED
+		sandbox.EnvSandboxBackend + "=bubblewrap",
+		"HOME=/home/u",
+	}
+	out := scrubWorkerEnv(in)
+
+	for _, kv := range out {
+		if strings.HasPrefix(kv, sandbox.EnvSandboxed+"=") || strings.HasPrefix(kv, sandbox.EnvSandboxBackend+"=") {
+			t.Fatalf("re-entrancy marker not scrubbed: %q", kv)
+		}
+	}
+	want := map[string]bool{"PATH=/bin": true, "ZERO_SANDBOX_AUTO_ALLOW_BASH=1": true, "HOME=/home/u": true}
+	got := map[string]bool{}
+	for _, kv := range out {
+		got[kv] = true
+	}
+	for w := range want {
+		if !got[w] {
+			t.Fatalf("scrub dropped a non-marker var: %q", w)
+		}
+	}
+	if len(out) != len(want) {
+		t.Fatalf("scrub result = %v, want exactly %d preserved vars", out, len(want))
+	}
+	if len(in) != 5 {
+		t.Fatal("scrubWorkerEnv must not mutate the caller's slice")
+	}
+}
+
+// TestExecLauncherScrubsEnvEndToEnd proves the spawned WORKER PROCESS actually
+// receives an environment without the re-entrancy markers — the security
+// invariant. It uses /bin/sh to dump the worker env, so it is POSIX-only.
+func TestExecLauncherScrubsEnvEndToEnd(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("uses the POSIX env dumper /bin/sh -c env")
+	}
+	launcher, err := NewExecLauncher(ExecLauncherConfig{
+		Executable: "/bin/sh",
+		BaseArgs:   []string{"-c", "env"},
+		Env: append([]string{"KEEP=yes"},
+			sandbox.EnvSandboxed+"=1",
+			sandbox.EnvSandboxBackend+"=bubblewrap"),
+	})
+	if err != nil {
+		t.Fatalf("NewExecLauncher: %v", err)
+	}
+	h, err := launcher(context.Background(), WorkerSpec{})
+	if err != nil {
+		t.Fatalf("launch: %v", err)
+	}
+	lines := h.Stdout()
+	var got []string
+	for {
+		line, ok, lerr := lines.Next()
+		if lerr != nil {
+			t.Fatalf("read worker stdout: %v", lerr)
+		}
+		if !ok {
+			break
+		}
+		got = append(got, line)
+	}
+	if _, err := h.Wait(); err != nil {
+		t.Fatalf("worker wait: %v", err)
+	}
+
+	keepSeen := false
+	for _, l := range got {
+		if strings.HasPrefix(l, sandbox.EnvSandboxed+"=") || strings.HasPrefix(l, sandbox.EnvSandboxBackend+"=") {
+			t.Fatalf("worker process still has a re-entrancy marker in its env: %q", l)
+		}
+		if l == "KEEP=yes" {
+			keepSeen = true
+		}
+	}
+	if !keepSeen {
+		t.Fatalf("worker env lost a non-marker var (KEEP=yes); got %v", got)
+	}
+}
+
+func TestNewExecLauncherDefaults(t *testing.T) {
+	// With no Executable it resolves the test binary's own path (os.Executable),
+	// and constructs without error.
+	if _, err := NewExecLauncher(ExecLauncherConfig{}); err != nil {
+		t.Fatalf("NewExecLauncher with defaults: %v", err)
+	}
+}

internal/daemon/lock.go

@@ -0,0 +1,87 @@
+package daemon
+
+import (
+	"errors"
+	"fmt"
+	"io/fs"
+	"os"
+	"strconv"
+	"strings"
+)
+
+// Single-instance lock. Mirrors reference-daemon-code-agent-js/supervisor.js's
+// lock file: a PID file created with O_EXCL. A second start fails; a STALE lock
+// left by a dead daemon (the recorded PID is no longer alive) is reclaimed so the
+// daemon recovers from an unclean shutdown without manual cleanup.
+
+// ErrAlreadyRunning is returned when a live daemon already holds the lock.
+var ErrAlreadyRunning = errors.New("daemon: another instance is already running")
+
+// fileLock is an acquired single-instance lock.
+type fileLock struct {
+	path string
+}
+
+// processAlive reports whether pid is a live process. Implemented per-platform
+// (lock_posix.go / lock_windows.go). It is a package var so tests can stub it.
+var processAlive = osProcessAlive
+
+// acquireLock takes the single-instance lock at path, reclaiming a stale lock
+// whose recorded PID is dead. isAlive overrides the liveness check (tests pass a
+// stub); nil uses the real processAlive.
+func acquireLock(path string, isAlive func(pid int) bool) (*fileLock, error) {
+	if isAlive == nil {
+		isAlive = processAlive
+	}
+	// At most two passes: create, or detect-stale-then-retry once.
+	for attempt := 0; attempt < 2; attempt++ {
+		f, err := os.OpenFile(path, os.O_CREATE|os.O_EXCL|os.O_WRONLY, 0o600)
+		if err == nil {
+			// A failed PID write would leave a malformed lock file that another
+			// process reads as stale (unparsable PID) and wrongly reclaims, breaking
+			// the single-instance guarantee — so on write failure, remove it and fail.
+			if _, werr := fmt.Fprintf(f, "%d\n", os.Getpid()); werr != nil {
+				_ = f.Close()
+				_ = os.Remove(path)
+				return nil, werr
+			}
+			if cerr := f.Close(); cerr != nil {
+				_ = os.Remove(path)
+				return nil, cerr
+			}
+			return &fileLock{path: path}, nil
+		}
+		if !errors.Is(err, fs.ErrExist) {
+			return nil, err
+		}
+		pid, perr := readPidFile(path)
+		if perr == nil && pid > 0 && isAlive(pid) {
+			return nil, fmt.Errorf("%w (pid %d)", ErrAlreadyRunning, pid)
+		}
+		// Stale lock (dead PID or unreadable) — remove and retry once.
+		if rmErr := os.Remove(path); rmErr != nil && !errors.Is(rmErr, fs.ErrNotExist) {
+			return nil, rmErr
+		}
+	}
+	return nil, ErrAlreadyRunning
+}
+
+// release removes the lock file. Safe to call once.
+func (l *fileLock) release() error {
+	if l == nil || l.path == "" {
+		return nil
+	}
+	if err := os.Remove(l.path); err != nil && !errors.Is(err, fs.ErrNotExist) {
+		return err
+	}
+	return nil
+}
+
+// readPidFile reads and parses the PID recorded in a lock file.
+func readPidFile(path string) (int, error) {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return 0, err
+	}
+	return strconv.Atoi(strings.TrimSpace(string(data)))
+}

internal/daemon/lock_posix.go

@@ -0,0 +1,22 @@
+//go:build !windows
+
+package daemon
+
+import (
+	"errors"
+	"syscall"
+)
+
+// osProcessAlive reports whether pid is a live process on POSIX. Signal 0 does
+// not deliver a signal; it only checks existence/permission. ESRCH means no such
+// process (dead); EPERM means it exists but we may not signal it (alive).
+func osProcessAlive(pid int) bool {
+	if pid <= 0 {
+		return false
+	}
+	err := syscall.Kill(pid, 0)
+	if err == nil {
+		return true
+	}
+	return errors.Is(err, syscall.EPERM)
+}

internal/daemon/lock_test.go

@@ -0,0 +1,77 @@
+package daemon
+
+import (
+	"errors"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"testing"
+)
+
+func TestLockSingleInstance(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "daemon.lock")
+	alive := func(int) bool { return true }
+
+	l1, err := acquireLock(path, alive)
+	if err != nil {
+		t.Fatalf("first acquireLock: %v", err)
+	}
+	// Second acquire while the holder is "alive" must be refused.
+	if _, err := acquireLock(path, alive); !errors.Is(err, ErrAlreadyRunning) {
+		t.Fatalf("second acquireLock err = %v, want ErrAlreadyRunning", err)
+	}
+	// After release, a new acquire succeeds.
+	if err := l1.release(); err != nil {
+		t.Fatalf("release: %v", err)
+	}
+	l2, err := acquireLock(path, alive)
+	if err != nil {
+		t.Fatalf("acquire after release: %v", err)
+	}
+	_ = l2.release()
+}
+
+func TestLockStaleRecovery(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "daemon.lock")
+	// Simulate a stale lock from a crashed daemon: a PID file whose process is
+	// dead.
+	if err := os.WriteFile(path, []byte("4242\n"), 0o600); err != nil {
+		t.Fatalf("write stale lock: %v", err)
+	}
+	dead := func(int) bool { return false }
+	l, err := acquireLock(path, dead)
+	if err != nil {
+		t.Fatalf("stale-lock recovery failed: %v", err)
+	}
+	// The lock now records OUR pid, not the stale one.
+	data, _ := os.ReadFile(path)
+	if strings.TrimSpace(string(data)) != strconv.Itoa(os.Getpid()) {
+		t.Fatalf("reclaimed lock pid = %q, want %d", strings.TrimSpace(string(data)), os.Getpid())
+	}
+	_ = l.release()
+}
+
+func TestLockReleaseRemovesFile(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "daemon.lock")
+	l, err := acquireLock(path, func(int) bool { return true })
+	if err != nil {
+		t.Fatalf("acquireLock: %v", err)
+	}
+	if err := l.release(); err != nil {
+		t.Fatalf("release: %v", err)
+	}
+	if _, err := os.Stat(path); !errors.Is(err, os.ErrNotExist) {
+		t.Fatalf("lock file still present after release: %v", err)
+	}
+}
+
+func TestProcessAliveSelfAndDead(t *testing.T) {
+	if !osProcessAlive(os.Getpid()) {
+		t.Fatal("osProcessAlive(self) = false, want true")
+	}
+	// PID 0 / negative are never live.
+	if osProcessAlive(0) || osProcessAlive(-1) {
+		t.Fatal("osProcessAlive must reject non-positive pids")
+	}
+}

internal/daemon/lock_windows.go

@@ -0,0 +1,27 @@
+//go:build windows
+
+package daemon
+
+import "golang.org/x/sys/windows"
+
+// osProcessAlive reports whether pid is a live process on Windows. It opens the
+// process for limited query access; a successful open with a non-exited status
+// means the PID is live. x/sys is already a module dependency.
+func osProcessAlive(pid int) bool {
+	if pid <= 0 {
+		return false
+	}
+	handle, err := windows.OpenProcess(windows.PROCESS_QUERY_LIMITED_INFORMATION, false, uint32(pid))
+	if err != nil {
+		return false
+	}
+	defer windows.CloseHandle(handle)
+	var code uint32
+	if err := windows.GetExitCodeProcess(handle, &code); err != nil {
+		// Could not query; treat the open as proof of life (conservative: do not
+		// reclaim a lock we are unsure about).
+		return true
+	}
+	const stillActive = 259 // STILL_ACTIVE
+	return code == stillActive
+}

internal/daemon/paths.go

@@ -0,0 +1,42 @@
+package daemon
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+)
+
+// DefaultDir returns the per-user directory for the daemon's runtime files:
+// $XDG_RUNTIME_DIR/zero when set (a tmpfs owned by the user on Linux), otherwise
+// ~/.zero. Mirrors supervisor.js / config.js choosing a per-user runtime dir.
+func DefaultDir() (string, error) {
+	if rt := strings.TrimSpace(os.Getenv("XDG_RUNTIME_DIR")); rt != "" {
+		return filepath.Join(rt, "zero"), nil
+	}
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return "", err
+	}
+	return filepath.Join(home, ".zero"), nil
+}
+
+// Paths bundles the daemon's socket, lock, and status file paths.
+type Paths struct {
+	Socket string
+	Lock   string
+	Status string
+}
+
+// DefaultPaths returns the daemon control-socket, lock, and status file paths
+// under DefaultDir.
+func DefaultPaths() (Paths, error) {
+	dir, err := DefaultDir()
+	if err != nil {
+		return Paths{}, err
+	}
+	return Paths{
+		Socket: filepath.Join(dir, "daemon.sock"),
+		Lock:   filepath.Join(dir, "daemon.lock"),
+		Status: filepath.Join(dir, "daemon.status"),
+	}, nil
+}

internal/daemon/pool.go

@@ -0,0 +1,355 @@
+package daemon
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"sync"
+	"time"
+)
+
+// Exit codes a worker may use to signal restart policy. Mirrors
+// reference-daemon-code-agent-js/worker-manager.js EXIT_TEMPFAIL/EXIT_PERMANENT.
+const (
+	// ExitTempfail asks for a retry after a fixed cool-off (transient failure).
+	ExitTempfail = 75
+	// ExitPermanent asks the daemon to STOP retrying this work (fatal).
+	ExitPermanent = 76
+)
+
+// defaultPoolSize, defaultMaxAttempts and defaultKillTimeout are used when a
+// PoolOptions field is left zero.
+const (
+	defaultPoolSize    = 4
+	defaultMaxAttempts = 5
+	defaultKillTimeout = 5 * time.Second
+)
+
+// ErrPoolDraining is returned by Run once the pool is shutting down.
+var ErrPoolDraining = errors.New("daemon: pool is draining")
+
+// ErrPermanent is returned when a worker exits ExitPermanent or the attempt cap
+// is exhausted; the request must not be retried.
+var ErrPermanent = errors.New("daemon: worker failed permanently")
+
+// WorkerSpec describes the worker to launch for one session attempt.
+type WorkerSpec struct {
+	Session string
+	Cwd     string
+	// Args are the per-session exec flags (e.g. --prompt ...), parsed by the CLI
+	// and appended after `exec` by the production launcher.
+	Args []string
+}
+
+// WorkerHandle is a launched worker process the pool supervises. The production
+// launcher wraps an exec.Cmd running `zero exec -i/-o stream-json`; tests inject
+// a fake. Stdout yields the worker's stream-json event lines.
+type WorkerHandle interface {
+	Stdout() Lines
+	// Wait blocks until the worker exits and returns its process exit code.
+	Wait() (int, error)
+	// Kill force-terminates the worker (used on drain timeout).
+	Kill() error
+	Pid() int
+}
+
+// Lines is a stream of stream-json output lines from a worker. Next returns the
+// next line and io-EOF-style ok=false when the stream ends.
+type Lines interface {
+	Next() (line string, ok bool, err error)
+}
+
+// Launcher spawns a worker for a single attempt. It must apply the sandbox/env
+// policy (the production launcher scrubs the re-entrancy markers so the worker
+// re-establishes its own sandbox — see newExecLauncher).
+type Launcher func(ctx context.Context, spec WorkerSpec) (WorkerHandle, error)
+
+// PoolOptions configures a Pool. Zero fields take documented defaults.
+type PoolOptions struct {
+	// Size is the max number of concurrent workers (lease slots). Sessions beyond
+	// this queue until a slot frees.
+	Size int
+	// Launcher spawns a worker. Required.
+	Launcher Launcher
+	// MaxAttempts caps the total attempts per request (first + retries). When
+	// exhausted, Run returns ErrPermanent (mirrors capping consecutive restarts).
+	MaxAttempts int
+	// KillTimeout bounds graceful termination before SIGKILL-equivalent on drain.
+	KillTimeout time.Duration
+	// Backoff returns the delay before retry attempt n (1-based crash count).
+	// Defaults to BackoffWithJitter; tests inject a deterministic function.
+	Backoff func(attempt int) time.Duration
+	// TempfailDelay is the cool-off after an ExitTempfail exit. Defaults to 30s.
+	TempfailDelay time.Duration
+	// Log, when set, receives one line per lifecycle event.
+	Log func(string)
+}
+
+// Pool is a bounded, self-healing worker pool. A crashed worker never takes down
+// the pool: Run retries on a fresh worker with backoff up to MaxAttempts, and an
+// ExitPermanent worker stops retries. Drain stops new work and kills stragglers.
+type Pool struct {
+	opts  PoolOptions
+	slots chan struct{}
+
+	mu       sync.Mutex
+	draining bool
+	active   map[int]WorkerHandle // pid -> handle, for drain/kill + status
+	nextID   int
+
+	drainOnce sync.Once
+	drained   chan struct{}
+}
+
+// workerStat tracks one in-flight request's restart count (local to Run).
+type workerStat struct {
+	id       int
+	restarts int
+}
+
+// NewPool builds a pool from opts, filling defaults.
+func NewPool(opts PoolOptions) (*Pool, error) {
+	if opts.Launcher == nil {
+		return nil, errors.New("daemon: pool requires a Launcher")
+	}
+	if opts.Size <= 0 {
+		opts.Size = defaultPoolSize
+	}
+	if opts.MaxAttempts <= 0 {
+		opts.MaxAttempts = defaultMaxAttempts
+	}
+	if opts.KillTimeout <= 0 {
+		opts.KillTimeout = defaultKillTimeout
+	}
+	if opts.TempfailDelay <= 0 {
+		opts.TempfailDelay = 30 * time.Second
+	}
+	if opts.Backoff == nil {
+		opts.Backoff = BackoffWithJitter
+	}
+	return &Pool{
+		opts:    opts,
+		slots:   make(chan struct{}, opts.Size),
+		active:  map[int]WorkerHandle{},
+		drained: make(chan struct{}),
+	}, nil
+}
+
+// Size returns the configured max concurrency.
+func (p *Pool) Size() int { return p.opts.Size }
+
+func (p *Pool) logf(format string, args ...any) {
+	if p.opts.Log != nil {
+		p.opts.Log(fmt.Sprintf(format, args...))
+	}
+}
+
+// acquire takes a lease slot, queuing until one frees or ctx is cancelled / the
+// pool drains.
+func (p *Pool) acquire(ctx context.Context) error {
+	p.mu.Lock()
+	draining := p.draining
+	p.mu.Unlock()
+	if draining {
+		return ErrPoolDraining
+	}
+	select {
+	case p.slots <- struct{}{}:
+		return nil
+	case <-ctx.Done():
+		return ctx.Err()
+	case <-p.drained:
+		return ErrPoolDraining
+	}
+}
+
+func (p *Pool) release() { <-p.slots }
+
+// QueueDepth reports how many lease slots are currently in use.
+func (p *Pool) QueueDepth() int { return len(p.slots) }
+
+// Sink receives a worker's stream-json output lines for one request.
+type Sink interface {
+	Line(line string)
+}
+
+// Run leases a worker slot and dispatches spec to a worker, streaming its
+// stream-json lines to sink. It is at-least-once with bounded retries: a worker
+// that crashes (non-zero, non-permanent) is retried on a fresh worker after a
+// backoff; ExitPermanent or exhausting MaxAttempts returns ErrPermanent. Run
+// queues when all slots are busy. The returned int is the final worker exit code.
+func (p *Pool) Run(ctx context.Context, spec WorkerSpec, sink Sink) (int, error) {
+	if err := p.acquire(ctx); err != nil {
+		return 0, err
+	}
+	defer p.release()
+
+	// Lease hook: a sink that tracks queue->running state is notified the instant
+	// a worker slot is leased (mirrors lease.js acquiring a worker for a session).
+	if s, ok := sink.(interface{ Started() }); ok {
+		s.Started()
+	}
+
+	stat := p.newStat()
+	var lastErr error
+	for attempt := 1; attempt <= p.opts.MaxAttempts; attempt++ {
+		if p.isDraining() {
+			return 0, ErrPoolDraining
+		}
+		code, err := p.runOnce(ctx, stat.id, spec, sink)
+		switch {
+		case err != nil:
+			lastErr = err
+			if ctx.Err() != nil {
+				return 0, ctx.Err()
+			}
+			p.logf("worker %d launch/run error: %v", stat.id, err)
+		case code == 0:
+			return 0, nil // clean success
+		case code == ExitPermanent:
+			p.logf("worker %d exited permanently (code=%d) — not retrying", stat.id, code)
+			return code, ErrPermanent
+		case code == ExitTempfail:
+			lastErr = fmt.Errorf("worker %d tempfail (code=%d)", stat.id, code)
+			p.logf("worker %d tempfail — retry after %s", stat.id, p.opts.TempfailDelay)
+			if !p.sleep(ctx, p.opts.TempfailDelay) {
+				return 0, ctx.Err()
+			}
+			continue // tempfail retries do not count against the crash backoff
+		default:
+			lastErr = fmt.Errorf("worker %d exited code=%d", stat.id, code)
+		}
+
+		if attempt == p.opts.MaxAttempts {
+			break
+		}
+		stat.restarts++
+		delay := p.opts.Backoff(stat.restarts)
+		p.logf("worker %d restart %d after backoff %s", stat.id, stat.restarts, delay)
+		if !p.sleep(ctx, delay) {
+			return 0, ctx.Err()
+		}
+	}
+	if lastErr == nil {
+		lastErr = ErrPermanent
+	}
+	return ExitPermanent, fmt.Errorf("%w: %v", ErrPermanent, lastErr)
+}
+
+// runOnce launches a single worker, pumps its output to sink, and returns its
+// exit code. The worker handle is tracked so Drain can kill it.
+func (p *Pool) runOnce(ctx context.Context, id int, spec WorkerSpec, sink Sink) (int, error) {
+	handle, err := p.opts.Launcher(ctx, spec)
+	if err != nil {
+		return 0, err
+	}
+	p.track(handle)
+	defer p.untrack(handle)
+
+	// Pump stdout lines until the stream ends.
+	lines := handle.Stdout()
+	for {
+		line, ok, lerr := lines.Next()
+		if lerr != nil {
+			break
+		}
+		if !ok {
+			break
+		}
+		if sink != nil {
+			sink.Line(line)
+		}
+	}
+	return handle.Wait()
+}
+
+func (p *Pool) newStat() *workerStat {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	p.nextID++
+	return &workerStat{id: p.nextID}
+}
+
+func (p *Pool) track(h WorkerHandle) {
+	p.mu.Lock()
+	p.active[h.Pid()] = h
+	p.mu.Unlock()
+}
+
+func (p *Pool) untrack(h WorkerHandle) {
+	p.mu.Lock()
+	delete(p.active, h.Pid())
+	p.mu.Unlock()
+}
+
+func (p *Pool) isDraining() bool {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return p.draining
+}
+
+// sleep waits d, returning false if ctx is cancelled first. A non-positive d
+// returns immediately.
+func (p *Pool) sleep(ctx context.Context, d time.Duration) bool {
+	if d <= 0 {
+		return ctx.Err() == nil
+	}
+	timer := time.NewTimer(d)
+	defer timer.Stop()
+	select {
+	case <-timer.C:
+		return true
+	case <-ctx.Done():
+		return false
+	}
+}
+
+// Drain stops accepting new work, gives in-flight workers a grace window
+// (KillTimeout) to finish on their own, then force-kills any straggler. It
+// returns as soon as the pool is idle (graceful) or the window elapses. Safe to
+// call once; subsequent calls are no-ops.
+func (p *Pool) Drain() {
+	p.drainOnce.Do(func() {
+		p.mu.Lock()
+		p.draining = true
+		p.mu.Unlock()
+		close(p.drained)
+
+		// Grace window: poll until idle or the deadline.
+		deadline := time.Now().Add(p.opts.KillTimeout)
+		for time.Now().Before(deadline) {
+			p.mu.Lock()
+			n := len(p.active)
+			p.mu.Unlock()
+			if n == 0 {
+				return // all workers drained gracefully
+			}
+			time.Sleep(5 * time.Millisecond)
+		}
+
+		// Timed out: force-kill remaining workers.
+		p.mu.Lock()
+		handles := make([]WorkerHandle, 0, len(p.active))
+		for _, h := range p.active {
+			handles = append(handles, h)
+		}
+		p.mu.Unlock()
+		for _, h := range handles {
+			p.logf("drain: killing straggler worker pid=%d", h.Pid())
+			_ = h.Kill()
+		}
+	})
+}
+
+// WorkerStats returns a snapshot of the currently-active (busy) workers, bounded
+// by the pool size. The on-demand pool keeps no idle workers, so an empty result
+// means the pool is idle.
+func (p *Pool) WorkerStats() []WorkerStatus {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	out := make([]WorkerStatus, 0, len(p.active))
+	for pid := range p.active {
+		out = append(out, WorkerStatus{PID: pid, State: "busy"})
+	}
+	return out
+}

internal/daemon/pool_test.go

@@ -0,0 +1,248 @@
+package daemon
+
+import (
+	"context"
+	"errors"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+// --- test doubles ---------------------------------------------------------
+
+type fakeLines struct {
+	lines []string
+	i     int
+}
+
+func (f *fakeLines) Next() (string, bool, error) {
+	if f.i >= len(f.lines) {
+		return "", false, nil
+	}
+	l := f.lines[f.i]
+	f.i++
+	return l, true, nil
+}
+
+type fakeWorker struct {
+	pid      int
+	out      []string
+	exitCode int
+	killed   int32
+	waitCh   chan struct{} // when non-nil, Wait blocks until closed (drain tests)
+}
+
+func (w *fakeWorker) Stdout() Lines { return &fakeLines{lines: w.out} }
+func (w *fakeWorker) Wait() (int, error) {
+	if w.waitCh != nil {
+		<-w.waitCh
+	}
+	return w.exitCode, nil
+}
+func (w *fakeWorker) Kill() error {
+	atomic.StoreInt32(&w.killed, 1)
+	if w.waitCh != nil {
+		select {
+		case <-w.waitCh:
+		default:
+			close(w.waitCh)
+		}
+	}
+	return nil
+}
+func (w *fakeWorker) Pid() int { return w.pid }
+
+// seqLauncher hands out the given workers in order, then errors.
+func seqLauncher(workers ...*fakeWorker) (Launcher, *int32) {
+	var calls int32
+	l := func(_ context.Context, _ WorkerSpec) (WorkerHandle, error) {
+		n := atomic.AddInt32(&calls, 1)
+		idx := int(n) - 1
+		if idx >= len(workers) {
+			return nil, errors.New("no more fake workers")
+		}
+		return workers[idx], nil
+	}
+	return l, &calls
+}
+
+type collectSink struct {
+	mu    sync.Mutex
+	lines []string
+}
+
+func (c *collectSink) Line(line string) {
+	c.mu.Lock()
+	c.lines = append(c.lines, line)
+	c.mu.Unlock()
+}
+
+// --- tests ----------------------------------------------------------------
+
+func TestPoolRunStreamsAndSucceeds(t *testing.T) {
+	launcher, calls := seqLauncher(&fakeWorker{pid: 1, out: []string{"a", "b"}, exitCode: 0})
+	pool, err := NewPool(PoolOptions{Size: 2, Launcher: launcher})
+	if err != nil {
+		t.Fatalf("NewPool: %v", err)
+	}
+	sink := &collectSink{}
+	code, err := pool.Run(context.Background(), WorkerSpec{Session: "s"}, sink)
+	if err != nil || code != 0 {
+		t.Fatalf("Run = (%d,%v), want (0,nil)", code, err)
+	}
+	if len(sink.lines) != 2 || sink.lines[0] != "a" || sink.lines[1] != "b" {
+		t.Fatalf("sink lines = %v, want [a b]", sink.lines)
+	}
+	if *calls != 1 {
+		t.Fatalf("launcher calls = %d, want 1", *calls)
+	}
+}
+
+func TestPoolRunRetriesWithBackoff(t *testing.T) {
+	launcher, calls := seqLauncher(
+		&fakeWorker{pid: 1, exitCode: 1},                      // crash
+		&fakeWorker{pid: 2, exitCode: 1},                      // crash
+		&fakeWorker{pid: 3, exitCode: 0, out: []string{"ok"}}, // success
+	)
+	var backoffArgs []int
+	pool, _ := NewPool(PoolOptions{
+		Size: 1, Launcher: launcher, MaxAttempts: 5,
+		Backoff: func(attempt int) time.Duration { backoffArgs = append(backoffArgs, attempt); return 0 },
+	})
+	sink := &collectSink{}
+	code, err := pool.Run(context.Background(), WorkerSpec{Session: "s"}, sink)
+	if err != nil || code != 0 {
+		t.Fatalf("Run = (%d,%v), want (0,nil)", code, err)
+	}
+	if *calls != 3 {
+		t.Fatalf("launcher calls = %d, want 3 (2 crashes + success)", *calls)
+	}
+	// Backoff invoked between attempts with an increasing 1-based restart count.
+	if len(backoffArgs) != 2 || backoffArgs[0] != 1 || backoffArgs[1] != 2 {
+		t.Fatalf("backoff attempts = %v, want [1 2]", backoffArgs)
+	}
+}
+
+func TestPoolRunPermanentStopsImmediately(t *testing.T) {
+	launcher, calls := seqLauncher(&fakeWorker{pid: 1, exitCode: ExitPermanent})
+	pool, _ := NewPool(PoolOptions{Size: 1, Launcher: launcher, Backoff: func(int) time.Duration { return 0 }})
+	code, err := pool.Run(context.Background(), WorkerSpec{Session: "s"}, &collectSink{})
+	if !errors.Is(err, ErrPermanent) {
+		t.Fatalf("Run err = %v, want ErrPermanent", err)
+	}
+	if code != ExitPermanent {
+		t.Fatalf("Run code = %d, want %d", code, ExitPermanent)
+	}
+	if *calls != 1 {
+		t.Fatalf("permanent exit must not retry; launcher calls = %d, want 1", *calls)
+	}
+}
+
+func TestPoolRunCapExhausted(t *testing.T) {
+	launcher, calls := seqLauncher(
+		&fakeWorker{pid: 1, exitCode: 1},
+		&fakeWorker{pid: 2, exitCode: 1},
+		&fakeWorker{pid: 3, exitCode: 1},
+	)
+	pool, _ := NewPool(PoolOptions{Size: 1, Launcher: launcher, MaxAttempts: 3, Backoff: func(int) time.Duration { return 0 }})
+	_, err := pool.Run(context.Background(), WorkerSpec{Session: "s"}, &collectSink{})
+	if !errors.Is(err, ErrPermanent) {
+		t.Fatalf("Run err = %v, want ErrPermanent after cap", err)
+	}
+	if *calls != 3 {
+		t.Fatalf("launcher calls = %d, want 3 (== MaxAttempts)", *calls)
+	}
+}
+
+func TestPoolRunTempfailRetries(t *testing.T) {
+	launcher, calls := seqLauncher(
+		&fakeWorker{pid: 1, exitCode: ExitTempfail},
+		&fakeWorker{pid: 2, exitCode: 0, out: []string{"ok"}},
+	)
+	pool, _ := NewPool(PoolOptions{Size: 1, Launcher: launcher, TempfailDelay: time.Millisecond})
+	code, err := pool.Run(context.Background(), WorkerSpec{Session: "s"}, &collectSink{})
+	if err != nil || code != 0 {
+		t.Fatalf("Run = (%d,%v), want (0,nil)", code, err)
+	}
+	if *calls != 2 {
+		t.Fatalf("launcher calls = %d, want 2 (tempfail + success)", *calls)
+	}
+}
+
+func TestPoolQueuesWhenFull(t *testing.T) {
+	block := make(chan struct{})
+	first := &fakeWorker{pid: 1, waitCh: block}
+	second := &fakeWorker{pid: 2, exitCode: 0}
+	launcher, _ := seqLauncher(first, second)
+	pool, _ := NewPool(PoolOptions{Size: 1, Launcher: launcher})
+
+	started := make(chan struct{})
+	go func() {
+		close(started)
+		_, _ = pool.Run(context.Background(), WorkerSpec{Session: "a"}, &collectSink{})
+	}()
+	<-started
+	// Wait until the first run holds the only slot.
+	waitFor(t, func() bool { return pool.QueueDepth() == 1 })
+
+	secondDone := make(chan struct{})
+	go func() {
+		_, _ = pool.Run(context.Background(), WorkerSpec{Session: "b"}, &collectSink{})
+		close(secondDone)
+	}()
+	// The second run must NOT complete while the slot is held.
+	select {
+	case <-secondDone:
+		t.Fatal("second run completed while the only slot was busy (no queueing)")
+	case <-time.After(50 * time.Millisecond):
+	}
+	// Free the slot; the queued run now proceeds.
+	close(block)
+	select {
+	case <-secondDone:
+	case <-time.After(2 * time.Second):
+		t.Fatal("queued run did not proceed after slot freed")
+	}
+}
+
+func TestPoolDrainKillsStraggler(t *testing.T) {
+	block := make(chan struct{})
+	straggler := &fakeWorker{pid: 1, waitCh: block}
+	launcher, _ := seqLauncher(straggler)
+	pool, _ := NewPool(PoolOptions{Size: 1, Launcher: launcher, MaxAttempts: 1, KillTimeout: 10 * time.Millisecond})
+
+	runDone := make(chan struct{})
+	go func() {
+		_, _ = pool.Run(context.Background(), WorkerSpec{Session: "a"}, &collectSink{})
+		close(runDone)
+	}()
+	waitFor(t, func() bool { return pool.QueueDepth() == 1 })
+
+	pool.Drain() // KillTimeout elapses, straggler is force-killed
+	if atomic.LoadInt32(&straggler.killed) != 1 {
+		t.Fatal("Drain must kill a straggler still running after KillTimeout")
+	}
+	select {
+	case <-runDone:
+	case <-time.After(2 * time.Second):
+		t.Fatal("run did not finish after drain kill")
+	}
+
+	// A new Run after drain is refused.
+	if _, err := pool.Run(context.Background(), WorkerSpec{Session: "b"}, &collectSink{}); !errors.Is(err, ErrPoolDraining) {
+		t.Fatalf("Run after drain err = %v, want ErrPoolDraining", err)
+	}
+}
+
+func waitFor(t *testing.T, cond func() bool) {
+	t.Helper()
+	deadline := time.Now().Add(2 * time.Second)
+	for time.Now().Before(deadline) {
+		if cond() {
+			return
+		}
+		time.Sleep(time.Millisecond)
+	}
+	t.Fatal("condition not met within timeout")
+}

internal/daemon/protocol.go

@@ -0,0 +1,184 @@
+// Package daemon implements zero's long-running daemon/server mode: a local
+// control server that supervises a pool of headless `zero exec` worker processes
+// and routes multiple agent sessions to them over an owner-only local socket.
+//
+// It reuses zero's existing building blocks rather than inventing new ones:
+//   - internal/background : child-process group setup + cross-platform terminate.
+//   - internal/streamjson : the line-based agent event protocol on worker stdio.
+//   - internal/cli (exec) : a worker is a `zero exec -i/-o stream-json` process.
+//
+// The control plane (this file) is a small framed codec mirrored from the
+// reference daemon's protocol.js: a 4-byte big-endian length prefix, a 1-byte
+// frame kind, then a JSON control payload, with a 1 MiB frame cap and a
+// version-negotiation handshake. The agent event stream itself stays stream-json.
+package daemon
+
+import (
+	"encoding/binary"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+)
+
+// Wire constants. Mirrors reference-daemon-code-agent-js/protocol.js
+// (FRAME_DATA/FRAME_CTRL, HEADER_SIZE, MAX_FRAME_SIZE, PROTO_VERSION).
+const (
+	// ProtoVersion is the control-protocol version advertised in the handshake.
+	ProtoVersion = 1
+	// MaxFrameSize caps a single control frame payload at 1 MiB. A larger
+	// advertised length is rejected before any allocation (fail closed).
+	MaxFrameSize = 1 << 20
+	// frameHeaderSize is 4 bytes big-endian length + 1 byte kind.
+	frameHeaderSize = 5
+)
+
+// FrameKind tags a frame's payload. Mirrors protocol.js FRAME_DATA/FRAME_CTRL.
+type FrameKind uint8
+
+const (
+	// KindData carries an opaque payload (e.g. a stream-json line). Reserved for
+	// future use; the control plane uses KindCtrl.
+	KindData FrameKind = 0
+	// KindCtrl carries a JSON-encoded control message (see Ctrl).
+	KindCtrl FrameKind = 1
+)
+
+// ErrFrameTooLarge is returned when a frame's advertised payload length exceeds
+// MaxFrameSize. The reader rejects it without allocating the buffer.
+var ErrFrameTooLarge = errors.New("daemon: control frame exceeds size cap")
+
+// ErrUnknownFrameKind is returned for a frame whose kind byte is not recognized.
+var ErrUnknownFrameKind = errors.New("daemon: unknown control frame kind")
+
+// WriteFrame writes a single framed message: [uint32BE len][uint8 kind][payload].
+// It rejects an over-cap payload before writing anything so a buggy/hostile caller
+// cannot emit a frame a conforming reader would refuse.
+func WriteFrame(w io.Writer, kind FrameKind, payload []byte) error {
+	if len(payload) > MaxFrameSize {
+		return fmt.Errorf("%w (%d > %d)", ErrFrameTooLarge, len(payload), MaxFrameSize)
+	}
+	header := make([]byte, frameHeaderSize)
+	binary.BigEndian.PutUint32(header[:4], uint32(len(payload)))
+	header[4] = byte(kind)
+	if _, err := w.Write(header); err != nil {
+		return err
+	}
+	if len(payload) == 0 {
+		return nil
+	}
+	_, err := w.Write(payload)
+	return err
+}
+
+// ReadFrame reads a single framed message. It validates the advertised length
+// against MaxFrameSize BEFORE allocating, so an oversize or hostile frame is
+// rejected (ErrFrameTooLarge) rather than triggering a huge allocation. An
+// unrecognized kind byte yields ErrUnknownFrameKind after the payload is drained
+// so the stream stays in sync.
+func ReadFrame(r io.Reader) (FrameKind, []byte, error) {
+	header := make([]byte, frameHeaderSize)
+	if _, err := io.ReadFull(r, header); err != nil {
+		return 0, nil, err
+	}
+	length := binary.BigEndian.Uint32(header[:4])
+	if length > MaxFrameSize {
+		return 0, nil, fmt.Errorf("%w (%d > %d)", ErrFrameTooLarge, length, MaxFrameSize)
+	}
+	kind := FrameKind(header[4])
+	payload := make([]byte, length)
+	if length > 0 {
+		if _, err := io.ReadFull(r, payload); err != nil {
+			return 0, nil, err
+		}
+	}
+	if kind != KindData && kind != KindCtrl {
+		return kind, nil, ErrUnknownFrameKind
+	}
+	return kind, payload, nil
+}
+
+// CtrlType is the discriminator for a control message.
+type CtrlType string
+
+const (
+	// CtrlHello is the client's opening handshake; Version is the highest control
+	// version it speaks.
+	CtrlHello CtrlType = "hello"
+	// CtrlHelloOK is the daemon's handshake reply with the negotiated Version.
+	CtrlHelloOK CtrlType = "hello_ok"
+	// CtrlRun asks the daemon to create/route a session and stream its output.
+	CtrlRun CtrlType = "run"
+	// CtrlAttach asks the daemon to attach the client to a running session's stream.
+	CtrlAttach CtrlType = "attach"
+	// CtrlStatus requests daemon/worker/session status.
+	CtrlStatus CtrlType = "status"
+	// CtrlStatusResult carries the status payload (see StatusReport in Status).
+	CtrlStatusResult CtrlType = "status_result"
+	// CtrlData carries one stream-json line of agent output back to the client.
+	CtrlData CtrlType = "data"
+	// CtrlAck acknowledges receipt/dispatch of a request (at-least-once dispatch).
+	CtrlAck CtrlType = "ack"
+	// CtrlEnd marks the end of a session's stream.
+	CtrlEnd CtrlType = "end"
+	// CtrlShutdown asks the daemon to drain and stop.
+	CtrlShutdown CtrlType = "shutdown"
+	// CtrlError carries a human-readable error for the client.
+	CtrlError CtrlType = "error"
+)
+
+// Ctrl is a control message carried in a KindCtrl frame as JSON. Unused fields
+// are omitted so frames stay small.
+type Ctrl struct {
+	Type    CtrlType `json:"type"`
+	Version int      `json:"version,omitempty"`
+	Session string   `json:"session,omitempty"`
+	Cwd     string   `json:"cwd,omitempty"`
+	Args    []string `json:"args,omitempty"`
+	Prompt  string   `json:"prompt,omitempty"`
+	// Line is one stream-json event line (for CtrlData).
+	Line string `json:"line,omitempty"`
+	// Message is a human-readable detail (for CtrlError / CtrlAck).
+	Message string `json:"message,omitempty"`
+	// Status is populated on CtrlStatusResult.
+	Status *StatusReport `json:"status,omitempty"`
+}
+
+// WriteControl marshals a control message into a KindCtrl frame.
+func WriteControl(w io.Writer, msg Ctrl) error {
+	payload, err := json.Marshal(msg)
+	if err != nil {
+		return err
+	}
+	return WriteFrame(w, KindCtrl, payload)
+}
+
+// ReadControl reads the next frame and decodes it as a control message. A
+// non-control frame yields ErrUnknownFrameKind so the caller fails closed.
+func ReadControl(r io.Reader) (Ctrl, error) {
+	kind, payload, err := ReadFrame(r)
+	if err != nil {
+		return Ctrl{}, err
+	}
+	if kind != KindCtrl {
+		return Ctrl{}, ErrUnknownFrameKind
+	}
+	var msg Ctrl
+	if err := json.Unmarshal(payload, &msg); err != nil {
+		return Ctrl{}, fmt.Errorf("daemon: decode control message: %w", err)
+	}
+	return msg, nil
+}
+
+// NegotiateVersion picks the protocol version both ends can speak: the lower of
+// the client's advertised version and ProtoVersion. A non-positive client
+// version is invalid and yields ok=false so the daemon can reject the connection.
+func NegotiateVersion(clientVersion int) (int, bool) {
+	if clientVersion <= 0 {
+		return 0, false
+	}
+	if clientVersion < ProtoVersion {
+		return clientVersion, true
+	}
+	return ProtoVersion, true
+}

internal/daemon/protocol_test.go

@@ -0,0 +1,120 @@
+package daemon
+
+import (
+	"bytes"
+	"encoding/binary"
+	"errors"
+	"io"
+	"strings"
+	"testing"
+)
+
+func TestFrameRoundTrip(t *testing.T) {
+	var buf bytes.Buffer
+	payloads := [][]byte{[]byte("hello"), {}, []byte(strings.Repeat("x", 4096))}
+	for _, p := range payloads {
+		if err := WriteFrame(&buf, KindCtrl, p); err != nil {
+			t.Fatalf("WriteFrame: %v", err)
+		}
+	}
+	for i, want := range payloads {
+		kind, got, err := ReadFrame(&buf)
+		if err != nil {
+			t.Fatalf("ReadFrame[%d]: %v", i, err)
+		}
+		if kind != KindCtrl {
+			t.Fatalf("frame[%d] kind = %d, want KindCtrl", i, kind)
+		}
+		if !bytes.Equal(got, want) {
+			t.Fatalf("frame[%d] payload = %q, want %q", i, got, want)
+		}
+	}
+}
+
+func TestControlRoundTrip(t *testing.T) {
+	var buf bytes.Buffer
+	msg := Ctrl{Type: CtrlRun, Session: "s1", Cwd: "/tmp", Args: []string{"--prompt", "hi"}}
+	if err := WriteControl(&buf, msg); err != nil {
+		t.Fatalf("WriteControl: %v", err)
+	}
+	got, err := ReadControl(&buf)
+	if err != nil {
+		t.Fatalf("ReadControl: %v", err)
+	}
+	if got.Type != CtrlRun || got.Session != "s1" || got.Cwd != "/tmp" || len(got.Args) != 2 {
+		t.Fatalf("round-trip mismatch: %+v", got)
+	}
+}
+
+func TestWriteFrameRejectsOversize(t *testing.T) {
+	var buf bytes.Buffer
+	err := WriteFrame(&buf, KindCtrl, make([]byte, MaxFrameSize+1))
+	if !errors.Is(err, ErrFrameTooLarge) {
+		t.Fatalf("WriteFrame oversize err = %v, want ErrFrameTooLarge", err)
+	}
+	if buf.Len() != 0 {
+		t.Fatalf("oversize WriteFrame must not emit any bytes, wrote %d", buf.Len())
+	}
+}
+
+func TestReadFrameRejectsOversizeBeforeAlloc(t *testing.T) {
+	// A header advertising a 4 GiB payload must be rejected without reading or
+	// allocating that payload.
+	header := make([]byte, frameHeaderSize)
+	binary.BigEndian.PutUint32(header[:4], MaxFrameSize+1)
+	header[4] = byte(KindCtrl)
+	_, _, err := ReadFrame(bytes.NewReader(header))
+	if !errors.Is(err, ErrFrameTooLarge) {
+		t.Fatalf("ReadFrame oversize err = %v, want ErrFrameTooLarge", err)
+	}
+}
+
+func TestReadFrameUnknownKind(t *testing.T) {
+	var buf bytes.Buffer
+	// Hand-write a frame with an unknown kind byte (7) and a small payload.
+	header := make([]byte, frameHeaderSize)
+	binary.BigEndian.PutUint32(header[:4], 2)
+	header[4] = 7
+	buf.Write(header)
+	buf.Write([]byte("ab"))
+	_, _, err := ReadFrame(&buf)
+	if !errors.Is(err, ErrUnknownFrameKind) {
+		t.Fatalf("unknown kind err = %v, want ErrUnknownFrameKind", err)
+	}
+}
+
+func TestReadFrameTruncated(t *testing.T) {
+	// Header promises 10 bytes but only 3 follow → unexpected EOF.
+	var buf bytes.Buffer
+	header := make([]byte, frameHeaderSize)
+	binary.BigEndian.PutUint32(header[:4], 10)
+	header[4] = byte(KindCtrl)
+	buf.Write(header)
+	buf.Write([]byte("abc"))
+	_, _, err := ReadFrame(&buf)
+	if err == nil || errors.Is(err, ErrFrameTooLarge) {
+		t.Fatalf("truncated frame err = %v, want a read error", err)
+	}
+	if !errors.Is(err, io.ErrUnexpectedEOF) {
+		t.Fatalf("truncated frame err = %v, want io.ErrUnexpectedEOF", err)
+	}
+}
+
+func TestNegotiateVersion(t *testing.T) {
+	cases := []struct {
+		client int
+		want   int
+		wantOK bool
+	}{
+		{client: 1, want: 1, wantOK: true},
+		{client: 99, want: ProtoVersion, wantOK: true}, // clamp down to ours
+		{client: 0, want: 0, wantOK: false},
+		{client: -3, want: 0, wantOK: false},
+	}
+	for _, tc := range cases {
+		got, ok := NegotiateVersion(tc.client)
+		if got != tc.want || ok != tc.wantOK {
+			t.Fatalf("NegotiateVersion(%d) = (%d,%v), want (%d,%v)", tc.client, got, ok, tc.want, tc.wantOK)
+		}
+	}
+}

internal/daemon/remote/auth.go

@@ -0,0 +1,158 @@
+// Package remote adds an OPT-IN, TLS-only network bridge on top of zero's local
+// daemon (internal/daemon). It lets a remote client drive the SAME daemon
+// SessionManager/Pool over the network, behind bearer-token authentication and a
+// protocol-version floor. The local Unix-socket daemon is unchanged and remains
+// the default; nothing here activates unless `zero daemon serve-remote` is run.
+//
+// Security (fail closed):
+//   - TLS is mandatory; the bridge refuses to serve without a cert/key.
+//   - A connection is authenticated BEFORE any control frame is dispatched; a
+//     failed/absent token closes the connection (after a small backoff) and no
+//     session frame is ever processed.
+//   - Tokens are compared in constant time and are never logged.
+//   - The control-frame size cap and version negotiation from internal/daemon are
+//     reused unchanged; oversize/old-version handshakes are rejected.
+//   - A remote-driven session runs through the same daemon dispatch, so it stays
+//     under the same sandbox + risk model — remote never bypasses local controls.
+package remote
+
+import (
+	"crypto/subtle"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"strings"
+
+	"github.com/Gitlawb/zero/internal/daemon"
+)
+
+// Env vars the bridge reads for its bearer token.
+const (
+	EnvToken     = "ZERO_DAEMON_REMOTE_TOKEN"
+	EnvTokenFile = "ZERO_DAEMON_REMOTE_TOKEN_FILE"
+)
+
+// ErrUnauthorized is returned when a token does not match.
+var ErrUnauthorized = errors.New("remote: unauthorized")
+
+// Authenticator verifies a bearer token presented by a remote client.
+type Authenticator interface {
+	Authenticate(token string) error
+}
+
+// TokenAuthenticator compares a presented token against a fixed secret in
+// constant time.
+type TokenAuthenticator struct {
+	token string
+}
+
+// NewTokenAuthenticator builds a token authenticator, refusing an empty secret
+// (fail closed — a bridge must never accept everyone).
+func NewTokenAuthenticator(token string) (*TokenAuthenticator, error) {
+	token = strings.TrimSpace(token)
+	if token == "" {
+		return nil, errors.New("remote: a non-empty auth token is required")
+	}
+	return &TokenAuthenticator{token: token}, nil
+}
+
+// Authenticate reports nil when token matches the configured secret.
+func (a *TokenAuthenticator) Authenticate(token string) error {
+	if subtle.ConstantTimeCompare([]byte(token), []byte(a.token)) == 1 {
+		return nil
+	}
+	return ErrUnauthorized
+}
+
+// TokenFromEnv resolves the bridge token from EnvToken, or a file named by
+// EnvTokenFile. It never logs the token.
+func TokenFromEnv() (string, error) {
+	if t := strings.TrimSpace(os.Getenv(EnvToken)); t != "" {
+		return t, nil
+	}
+	if file := strings.TrimSpace(os.Getenv(EnvTokenFile)); file != "" {
+		data, err := os.ReadFile(file)
+		if err != nil {
+			return "", fmt.Errorf("remote: read token file: %w", err)
+		}
+		t := strings.TrimSpace(string(data))
+		if t == "" {
+			return "", errors.New("remote: token file is empty")
+		}
+		return t, nil
+	}
+	return "", fmt.Errorf("remote: set %s or %s", EnvToken, EnvTokenFile)
+}
+
+// Attestation is an optional post-token hook (e.g. workload attestation). The
+// default is a no-op; a deployment can supply a stricter implementation.
+type Attestation interface {
+	Verify(meta map[string]string) error
+}
+
+type noopAttestation struct{}
+
+func (noopAttestation) Verify(map[string]string) error { return nil }
+
+// authRequest is the first frame a remote client sends (before the daemon
+// hello). Token is never logged.
+type authRequest struct {
+	Token   string            `json:"token"`
+	Version int               `json:"version"`
+	Meta    map[string]string `json:"meta,omitempty"`
+}
+
+// authResponse is the bridge's reply to the auth handshake.
+type authResponse struct {
+	OK      bool   `json:"ok"`
+	Version int    `json:"version,omitempty"`
+	Message string `json:"message,omitempty"`
+}
+
+func writeAuthRequest(w io.Writer, req authRequest) error {
+	payload, err := json.Marshal(req)
+	if err != nil {
+		return err
+	}
+	return daemon.WriteFrame(w, daemon.KindCtrl, payload)
+}
+
+func readAuthRequest(r io.Reader) (authRequest, error) {
+	kind, payload, err := daemon.ReadFrame(r)
+	if err != nil {
+		return authRequest{}, err
+	}
+	if kind != daemon.KindCtrl {
+		return authRequest{}, errors.New("remote: expected auth frame")
+	}
+	var req authRequest
+	if err := json.Unmarshal(payload, &req); err != nil {
+		return authRequest{}, fmt.Errorf("remote: decode auth request: %w", err)
+	}
+	return req, nil
+}
+
+func writeAuthResponse(w io.Writer, resp authResponse) error {
+	payload, err := json.Marshal(resp)
+	if err != nil {
+		return err
+	}
+	return daemon.WriteFrame(w, daemon.KindCtrl, payload)
+}
+
+func readAuthResponse(r io.Reader) (authResponse, error) {
+	kind, payload, err := daemon.ReadFrame(r)
+	if err != nil {
+		return authResponse{}, err
+	}
+	if kind != daemon.KindCtrl {
+		return authResponse{}, errors.New("remote: expected auth response frame")
+	}
+	var resp authResponse
+	if err := json.Unmarshal(payload, &resp); err != nil {
+		return authResponse{}, fmt.Errorf("remote: decode auth response: %w", err)
+	}
+	return resp, nil
+}

internal/daemon/remote/auth_test.go

@@ -0,0 +1,67 @@
+package remote
+
+import (
+	"errors"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestTokenAuthenticator(t *testing.T) {
+	if _, err := NewTokenAuthenticator("  "); err == nil {
+		t.Fatal("empty token must be rejected (fail closed)")
+	}
+	a, err := NewTokenAuthenticator("s3cret")
+	if err != nil {
+		t.Fatalf("NewTokenAuthenticator: %v", err)
+	}
+	if err := a.Authenticate("s3cret"); err != nil {
+		t.Fatalf("matching token should authenticate: %v", err)
+	}
+	if err := a.Authenticate("wrong"); !errors.Is(err, ErrUnauthorized) {
+		t.Fatalf("mismatch err = %v, want ErrUnauthorized", err)
+	}
+	if err := a.Authenticate(""); !errors.Is(err, ErrUnauthorized) {
+		t.Fatalf("empty presented token err = %v, want ErrUnauthorized", err)
+	}
+}
+
+func TestTokenFromEnv(t *testing.T) {
+	// Clear both so the no-config path errors.
+	t.Setenv(EnvToken, "")
+	t.Setenv(EnvTokenFile, "")
+	if _, err := TokenFromEnv(); err == nil {
+		t.Fatal("TokenFromEnv with neither var set must error")
+	}
+	// Direct env.
+	t.Setenv(EnvToken, "from-env")
+	if tok, err := TokenFromEnv(); err != nil || tok != "from-env" {
+		t.Fatalf("TokenFromEnv(env) = %q, %v", tok, err)
+	}
+	// File (env takes precedence, so clear env first).
+	t.Setenv(EnvToken, "")
+	file := filepath.Join(t.TempDir(), "tok")
+	if err := os.WriteFile(file, []byte("  from-file\n"), 0o600); err != nil {
+		t.Fatalf("write token file: %v", err)
+	}
+	t.Setenv(EnvTokenFile, file)
+	if tok, err := TokenFromEnv(); err != nil || tok != "from-file" {
+		t.Fatalf("TokenFromEnv(file) = %q, %v", tok, err)
+	}
+	// Empty file fails closed.
+	if err := os.WriteFile(file, []byte("   \n"), 0o600); err != nil {
+		t.Fatalf("rewrite token file: %v", err)
+	}
+	if _, err := TokenFromEnv(); err == nil {
+		t.Fatal("empty token file must error")
+	}
+}
+
+func TestServerTLSConfigRequiresCertKey(t *testing.T) {
+	if _, err := ServerTLSConfig("", ""); err == nil {
+		t.Fatal("ServerTLSConfig must require a cert and key (TLS mandatory)")
+	}
+	if _, err := ServerTLSConfig("/nope/cert.pem", "/nope/key.pem"); err == nil {
+		t.Fatal("ServerTLSConfig must error on unreadable cert/key")
+	}
+}

internal/daemon/remote/bridge.go

@@ -0,0 +1,214 @@
+package remote
+
+import (
+	"crypto/tls"
+	"errors"
+	"fmt"
+	"net"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/daemon"
+)
+
+const (
+	defaultMaxConnections   = 32
+	defaultHandshakeTimeout = 10 * time.Second
+	defaultAuthFailDelay    = 250 * time.Millisecond
+)
+
+// Bridge serves authenticated remote connections and drives the local daemon's
+// control dispatch for each one.
+type Bridge struct {
+	server           *daemon.Server
+	auth             Authenticator
+	attest           Attestation
+	minVersion       int
+	handshakeTimeout time.Duration
+	authFailDelay    time.Duration
+	log              func(string)
+	sem              chan struct{}
+
+	mu       sync.Mutex
+	listener net.Listener
+}
+
+// BridgeOptions configures a Bridge.
+type BridgeOptions struct {
+	// Server is the local daemon to drive (required).
+	Server *daemon.Server
+	// Authenticator verifies bearer tokens (required).
+	Authenticator Authenticator
+	// Attestation is an optional post-token hook; nil => no-op.
+	Attestation Attestation
+	// MinVersion is the lowest control-protocol version accepted; 0 =>
+	// daemon.ProtoVersion.
+	MinVersion int
+	// MaxConnections bounds concurrent remote connections; 0 => default.
+	MaxConnections int
+	// HandshakeTimeout bounds the auth handshake; 0 => default.
+	HandshakeTimeout time.Duration
+	// AuthFailDelay slows brute-force attempts; <0 => none, 0 => default.
+	AuthFailDelay time.Duration
+	Log           func(string)
+}
+
+// NewBridge validates options and builds a Bridge.
+func NewBridge(opts BridgeOptions) (*Bridge, error) {
+	if opts.Server == nil {
+		return nil, errors.New("remote: bridge requires a daemon server")
+	}
+	if opts.Authenticator == nil {
+		return nil, errors.New("remote: bridge requires an authenticator")
+	}
+	minVersion := opts.MinVersion
+	if minVersion <= 0 {
+		minVersion = daemon.ProtoVersion
+	}
+	maxConns := opts.MaxConnections
+	if maxConns <= 0 {
+		maxConns = defaultMaxConnections
+	}
+	handshakeTimeout := opts.HandshakeTimeout
+	if handshakeTimeout <= 0 {
+		handshakeTimeout = defaultHandshakeTimeout
+	}
+	authFailDelay := opts.AuthFailDelay
+	switch {
+	case authFailDelay < 0:
+		authFailDelay = 0
+	case authFailDelay == 0:
+		authFailDelay = defaultAuthFailDelay
+	}
+	attest := opts.Attestation
+	if attest == nil {
+		attest = noopAttestation{}
+	}
+	return &Bridge{
+		server:           opts.Server,
+		auth:             opts.Authenticator,
+		attest:           attest,
+		minVersion:       minVersion,
+		handshakeTimeout: handshakeTimeout,
+		authFailDelay:    authFailDelay,
+		log:              opts.Log,
+		sem:              make(chan struct{}, maxConns),
+	}, nil
+}
+
+func (b *Bridge) logf(format string, args ...any) {
+	if b.log != nil {
+		b.log(fmt.Sprintf(format, args...))
+	}
+}
+
+// ListenAndServeTLS binds a TLS listener on addr and serves until the listener
+// is closed (via Close). It refuses to serve without a TLS config (no plaintext
+// remote bridge).
+func (b *Bridge) ListenAndServeTLS(addr string, tlsConfig *tls.Config) error {
+	if tlsConfig == nil {
+		return errors.New("remote: TLS config is required; refusing to serve a plaintext remote bridge")
+	}
+	listener, err := tls.Listen("tcp", addr, tlsConfig)
+	if err != nil {
+		return fmt.Errorf("remote: listen %s: %w", addr, err)
+	}
+	b.mu.Lock()
+	b.listener = listener
+	b.mu.Unlock()
+	return b.Serve(listener)
+}
+
+// Close stops the listener started by ListenAndServeTLS, causing Serve to
+// return. Safe to call before serving or more than once.
+func (b *Bridge) Close() error {
+	b.mu.Lock()
+	listener := b.listener
+	b.listener = nil // clear so a repeat Close is a no-op, not a closed-listener error
+	b.mu.Unlock()
+	if listener != nil {
+		return listener.Close()
+	}
+	return nil
+}
+
+// Serve accepts connections from listener until it returns an error (e.g. it is
+// closed). Each connection is authenticated, then handed to the daemon's control
+// dispatch. Connections beyond MaxConnections are refused immediately.
+func (b *Bridge) Serve(listener net.Listener) error {
+	for {
+		conn, err := listener.Accept()
+		if err != nil {
+			return err
+		}
+		select {
+		case b.sem <- struct{}{}:
+		default:
+			// At capacity: refuse fast (fail closed) rather than queueing.
+			b.logf("remote: refused connection (at capacity)")
+			_ = conn.Close()
+			continue
+		}
+		go func() {
+			defer func() { <-b.sem }()
+			b.handle(conn)
+		}()
+	}
+}
+
+// handle authenticates conn then drives the daemon dispatch. It always closes
+// conn (directly on a failed handshake, or via ServeConn on success).
+func (b *Bridge) handle(conn net.Conn) {
+	// Bound the auth handshake so a stalled peer cannot hold a connection slot.
+	_ = conn.SetDeadline(time.Now().Add(b.handshakeTimeout))
+	req, err := readAuthRequest(conn)
+	if err != nil {
+		_ = conn.Close()
+		return
+	}
+	if req.Version < b.minVersion {
+		b.deny(conn, "unsupported protocol version")
+		return
+	}
+	if err := b.auth.Authenticate(req.Token); err != nil {
+		b.deny(conn, "unauthorized")
+		return
+	}
+	if err := b.attest.Verify(req.Meta); err != nil {
+		b.deny(conn, "attestation failed")
+		return
+	}
+	if err := writeAuthResponse(conn, authResponse{OK: true, Version: daemon.ProtoVersion}); err != nil {
+		_ = conn.Close()
+		return
+	}
+	// Clear the handshake deadline: a session may stream for a long time.
+	_ = conn.SetDeadline(time.Time{})
+	b.server.ServeConn(conn) // performs the daemon handshake + one command, then closes conn
+}
+
+// deny rejects an unauthenticated connection after a small backoff (to slow
+// brute force). The reason is generic and never includes any token material.
+func (b *Bridge) deny(conn net.Conn, message string) {
+	if b.authFailDelay > 0 {
+		time.Sleep(b.authFailDelay)
+	}
+	_ = writeAuthResponse(conn, authResponse{OK: false, Message: message})
+	_ = conn.Close()
+}
+
+// ServerTLSConfig builds a server TLS config from a cert/key pair, refusing if
+// either is missing (TLS is mandatory for the remote bridge).
+func ServerTLSConfig(certFile, keyFile string) (*tls.Config, error) {
+	certFile = strings.TrimSpace(certFile)
+	keyFile = strings.TrimSpace(keyFile)
+	if certFile == "" || keyFile == "" {
+		return nil, errors.New("remote: a TLS cert and key are required for serve-remote")
+	}
+	cert, err := tls.LoadX509KeyPair(certFile, keyFile)
+	if err != nil {
+		return nil, fmt.Errorf("remote: load TLS keypair: %w", err)
+	}
+	return &tls.Config{Certificates: []tls.Certificate{cert}, MinVersion: tls.VersionTLS12}, nil
+}

internal/daemon/remote/bridge_test.go

@@ -0,0 +1,342 @@
+package remote
+
+import (
+	"context"
+	"crypto/ecdsa"
+	"crypto/elliptic"
+	"crypto/rand"
+	"crypto/tls"
+	"crypto/x509"
+	"crypto/x509/pkix"
+	"encoding/pem"
+	"errors"
+	"math/big"
+	"net"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/daemon"
+)
+
+// ---- fake daemon worker (the daemon's test fakes are unexported) -------------
+
+type fakeLines struct{ ch <-chan string }
+
+func (l fakeLines) Next() (string, bool, error) {
+	line, ok := <-l.ch
+	return line, ok, nil
+}
+
+type fakeWorker struct {
+	ch  chan string
+	pid int
+}
+
+func (w *fakeWorker) Stdout() daemon.Lines { return fakeLines{ch: w.ch} }
+func (w *fakeWorker) Wait() (int, error)   { return 0, nil }
+func (w *fakeWorker) Kill() error          { return nil }
+func (w *fakeWorker) Pid() int             { return w.pid }
+
+// staticLauncher emits the given lines then ends the stream.
+func staticLauncher(lines ...string) daemon.Launcher {
+	return func(_ context.Context, _ daemon.WorkerSpec) (daemon.WorkerHandle, error) {
+		ch := make(chan string, len(lines))
+		for _, l := range lines {
+			ch <- l
+		}
+		close(ch)
+		return &fakeWorker{ch: ch, pid: 1}, nil
+	}
+}
+
+// blockingLauncher returns a worker whose stream stays open until release is
+// closed, so a session (and its bridge connection slot) is held open.
+func blockingLauncher(release <-chan struct{}) daemon.Launcher {
+	return func(_ context.Context, _ daemon.WorkerSpec) (daemon.WorkerHandle, error) {
+		ch := make(chan string)
+		go func() {
+			<-release
+			close(ch)
+		}()
+		return &fakeWorker{ch: ch, pid: 1}, nil
+	}
+}
+
+func newBridgeServer(t *testing.T, launcher daemon.Launcher) *daemon.Server {
+	t.Helper()
+	dir := t.TempDir()
+	pool, err := daemon.NewPool(daemon.PoolOptions{Size: 4, Launcher: launcher, KillTimeout: 200 * time.Millisecond})
+	if err != nil {
+		t.Fatalf("NewPool: %v", err)
+	}
+	mgr, err := daemon.NewSessionManager(daemon.SessionManagerOptions{Pool: pool})
+	if err != nil {
+		t.Fatalf("NewSessionManager: %v", err)
+	}
+	srv, err := daemon.NewServer(daemon.ServerOptions{
+		Paths:   daemon.Paths{Socket: filepath.Join(dir, "s.sock"), Lock: filepath.Join(dir, "s.lock"), Status: filepath.Join(dir, "s.status")},
+		Manager: mgr,
+		Pool:    pool,
+	})
+	if err != nil {
+		t.Fatalf("NewServer: %v", err)
+	}
+	t.Cleanup(srv.Shutdown)
+	return srv
+}
+
+// genTestCert writes a self-signed cert/key (valid for 127.0.0.1) to temp files.
+func genTestCert(t *testing.T) (certFile, keyFile string) {
+	t.Helper()
+	key, err := ecdsa.GenerateKey(elliptic.P256(), rand.Reader)
+	if err != nil {
+		t.Fatalf("genkey: %v", err)
+	}
+	tmpl := &x509.Certificate{
+		SerialNumber: big.NewInt(1),
+		Subject:      pkix.Name{CommonName: "zero-remote-test"},
+		NotBefore:    time.Now().Add(-time.Hour),
+		NotAfter:     time.Now().Add(time.Hour),
+		KeyUsage:     x509.KeyUsageDigitalSignature | x509.KeyUsageCertSign,
+		ExtKeyUsage:  []x509.ExtKeyUsage{x509.ExtKeyUsageServerAuth},
+		IsCA:         true,
+		IPAddresses:  []net.IP{net.ParseIP("127.0.0.1"), net.ParseIP("::1")},
+		DNSNames:     []string{"localhost"},
+	}
+	der, err := x509.CreateCertificate(rand.Reader, tmpl, tmpl, &key.PublicKey, key)
+	if err != nil {
+		t.Fatalf("create cert: %v", err)
+	}
+	dir := t.TempDir()
+	certFile = filepath.Join(dir, "cert.pem")
+	keyFile = filepath.Join(dir, "key.pem")
+	certOut, _ := os.Create(certFile)
+	_ = pem.Encode(certOut, &pem.Block{Type: "CERTIFICATE", Bytes: der})
+	_ = certOut.Close()
+	keyDER, _ := x509.MarshalECPrivateKey(key)
+	keyOut, _ := os.Create(keyFile)
+	_ = pem.Encode(keyOut, &pem.Block{Type: "EC PRIVATE KEY", Bytes: keyDER})
+	_ = keyOut.Close()
+	return certFile, keyFile
+}
+
+// startBridge starts a TLS bridge on 127.0.0.1:0 and returns its address + cert.
+func startBridge(t *testing.T, srv *daemon.Server, opts BridgeOptions) (addr, caCert string) {
+	t.Helper()
+	certFile, keyFile := genTestCert(t)
+	tlsConfig, err := ServerTLSConfig(certFile, keyFile)
+	if err != nil {
+		t.Fatalf("ServerTLSConfig: %v", err)
+	}
+	opts.Server = srv
+	bridge, err := NewBridge(opts)
+	if err != nil {
+		t.Fatalf("NewBridge: %v", err)
+	}
+	ln, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		t.Fatalf("listen: %v", err)
+	}
+	tlsLn := tls.NewListener(ln, tlsConfig)
+	go func() { _ = bridge.Serve(tlsLn) }()
+	t.Cleanup(func() { _ = tlsLn.Close() })
+	return ln.Addr().String(), certFile
+}
+
+func TestBridgeRunStatusRoundTrip(t *testing.T) {
+	srv := newBridgeServer(t, staticLauncher(`{"type":"event","seq":1}`, `{"type":"event","seq":2}`))
+	auth, _ := NewTokenAuthenticator("tok")
+	addr, ca := startBridge(t, srv, BridgeOptions{Authenticator: auth})
+
+	client, err := DialRemote(RemoteConfig{Address: addr, Token: "tok", CACertFile: ca})
+	if err != nil {
+		t.Fatalf("DialRemote: %v", err)
+	}
+	var got []string
+	if err := client.Run("sess-1", "", "hello", nil, func(l string) { got = append(got, l) }); err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	_ = client.Close()
+	if len(got) != 2 || got[0] != `{"type":"event","seq":1}` {
+		t.Fatalf("run output = %v", got)
+	}
+
+	// Status over a fresh remote connection (same protocol as local).
+	statusClient, err := DialRemote(RemoteConfig{Address: addr, Token: "tok", CACertFile: ca})
+	if err != nil {
+		t.Fatalf("DialRemote(status): %v", err)
+	}
+	report, err := statusClient.Status()
+	if err != nil {
+		t.Fatalf("Status: %v", err)
+	}
+	_ = statusClient.Close()
+	if report == nil || report.PID == 0 {
+		t.Fatalf("status report = %+v", report)
+	}
+}
+
+func TestBridgeRejectsBadToken(t *testing.T) {
+	srv := newBridgeServer(t, staticLauncher())
+	auth, _ := NewTokenAuthenticator("correct")
+	addr, ca := startBridge(t, srv, BridgeOptions{Authenticator: auth, AuthFailDelay: -1})
+
+	_, err := DialRemote(RemoteConfig{Address: addr, Token: "wrong", CACertFile: ca})
+	if !errors.Is(err, ErrUnauthorized) {
+		t.Fatalf("DialRemote(bad token) err = %v, want ErrUnauthorized", err)
+	}
+}
+
+func TestBridgeRejectsLowVersion(t *testing.T) {
+	srv := newBridgeServer(t, staticLauncher())
+	auth, _ := NewTokenAuthenticator("tok")
+	// Require a version higher than the client speaks (daemon.ProtoVersion).
+	addr, ca := startBridge(t, srv, BridgeOptions{Authenticator: auth, MinVersion: daemon.ProtoVersion + 1, AuthFailDelay: -1})
+
+	_, err := DialRemote(RemoteConfig{Address: addr, Token: "tok", CACertFile: ca})
+	if err == nil {
+		t.Fatal("DialRemote with a below-min version must be rejected")
+	}
+}
+
+func TestBridgeRequiresVerifiableCert(t *testing.T) {
+	srv := newBridgeServer(t, staticLauncher())
+	auth, _ := NewTokenAuthenticator("tok")
+	addr, _ := startBridge(t, srv, BridgeOptions{Authenticator: auth})
+	// No CA configured + a self-signed server cert => verification must fail
+	// (NEVER InsecureSkipVerify).
+	if _, err := DialRemote(RemoteConfig{Address: addr, Token: "tok"}); err == nil {
+		t.Fatal("DialRemote must reject an untrusted self-signed cert")
+	}
+}
+
+func TestListenAndServeTLSRequiresConfig(t *testing.T) {
+	srv := newBridgeServer(t, staticLauncher())
+	auth, _ := NewTokenAuthenticator("tok")
+	bridge, _ := NewBridge(BridgeOptions{Server: srv, Authenticator: auth})
+	if err := bridge.ListenAndServeTLS("127.0.0.1:0", nil); err == nil {
+		t.Fatal("ListenAndServeTLS(nil) must refuse to serve plaintext")
+	}
+}
+
+func TestListenAndServeTLSServesAndCloses(t *testing.T) {
+	srv := newBridgeServer(t, staticLauncher(`{"type":"event","seq":1}`))
+	auth, _ := NewTokenAuthenticator("tok")
+	certFile, keyFile := genTestCert(t)
+	tlsConfig, err := ServerTLSConfig(certFile, keyFile)
+	if err != nil {
+		t.Fatalf("ServerTLSConfig: %v", err)
+	}
+	bridge, err := NewBridge(BridgeOptions{Server: srv, Authenticator: auth})
+	if err != nil {
+		t.Fatalf("NewBridge: %v", err)
+	}
+	// Bind an ephemeral port, discover it, then serve via ListenAndServeTLS by
+	// passing the same addr (it rebinds) — instead we bind here to learn the port
+	// and serve through the public entry on that exact address.
+	probe, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		t.Fatalf("probe listen: %v", err)
+	}
+	addr := probe.Addr().String()
+	_ = probe.Close()
+
+	serveErr := make(chan error, 1)
+	go func() { serveErr <- bridge.ListenAndServeTLS(addr, tlsConfig) }()
+
+	// Wait until it accepts a connection, then drive one run.
+	var client *daemon.Client
+	deadline := time.Now().Add(2 * time.Second)
+	for {
+		client, err = DialRemote(RemoteConfig{Address: addr, Token: "tok", CACertFile: certFile, Timeout: 300 * time.Millisecond})
+		if err == nil {
+			break
+		}
+		if time.Now().After(deadline) {
+			t.Fatalf("bridge never came up: %v", err)
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+	var got []string
+	if err := client.Run("s", "", "hi", nil, func(l string) { got = append(got, l) }); err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	_ = client.Close()
+	if len(got) != 1 {
+		t.Fatalf("got %d lines, want 1", len(got))
+	}
+	// Close stops the accept loop; a second Close is a no-op (not an error).
+	if err := bridge.Close(); err != nil {
+		t.Fatalf("Close: %v", err)
+	}
+	if err := bridge.Close(); err != nil {
+		t.Fatalf("second Close should be a no-op, got: %v", err)
+	}
+	select {
+	case <-serveErr: // ListenAndServeTLS returned after Close
+	case <-time.After(2 * time.Second):
+		t.Fatal("ListenAndServeTLS did not return after Close")
+	}
+}
+
+func TestBridgeResumeAfterReconnect(t *testing.T) {
+	srv := newBridgeServer(t, staticLauncher(`{"type":"event","seq":1}`, `{"type":"event","seq":2}`))
+	auth, _ := NewTokenAuthenticator("tok")
+	addr, ca := startBridge(t, srv, BridgeOptions{Authenticator: auth})
+
+	// First connection runs the session to completion.
+	c1, err := DialRemote(RemoteConfig{Address: addr, Token: "tok", CACertFile: ca})
+	if err != nil {
+		t.Fatalf("DialRemote 1: %v", err)
+	}
+	if err := c1.Run("sess-resume", "", "hi", nil, func(string) {}); err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	_ = c1.Close()
+
+	// A new connection re-attaches and replays the buffered session output —
+	// the dropped first connection did not lose the session.
+	c2, err := DialRemote(RemoteConfig{Address: addr, Token: "tok", CACertFile: ca})
+	if err != nil {
+		t.Fatalf("DialRemote 2: %v", err)
+	}
+	defer c2.Close()
+	var replayed []string
+	if err := c2.Attach("sess-resume", func(l string) { replayed = append(replayed, l) }); err != nil {
+		t.Fatalf("Attach: %v", err)
+	}
+	if len(replayed) != 2 {
+		t.Fatalf("re-attach replayed %d lines, want 2", len(replayed))
+	}
+}
+
+func TestBridgeConnectionCap(t *testing.T) {
+	release := make(chan struct{})
+	defer close(release)
+	srv := newBridgeServer(t, blockingLauncher(release))
+	auth, _ := NewTokenAuthenticator("tok")
+	addr, ca := startBridge(t, srv, BridgeOptions{Authenticator: auth, MaxConnections: 1})
+
+	// First connection holds the only slot with a blocking run.
+	c1, err := DialRemote(RemoteConfig{Address: addr, Token: "tok", CACertFile: ca})
+	if err != nil {
+		t.Fatalf("DialRemote 1: %v", err)
+	}
+	defer c1.Close()
+	runErr := make(chan error, 1)
+	go func() { runErr <- c1.Run("sess-block", "", "hi", nil, func(string) {}) }()
+
+	// Wait until the slot is in use, then a second dial must be refused.
+	deadline := time.Now().Add(2 * time.Second)
+	for time.Now().Before(deadline) {
+		c2, err := DialRemote(RemoteConfig{Address: addr, Token: "tok", CACertFile: ca, Timeout: 500 * time.Millisecond})
+		if err != nil {
+			return // refused as expected
+		}
+		_ = c2.Close()
+		time.Sleep(20 * time.Millisecond)
+	}
+	t.Fatal("second connection was not refused at the connection cap")
+}

internal/daemon/remote/client.go

@@ -0,0 +1,108 @@
+package remote
+
+import (
+	"context"
+	"crypto/tls"
+	"crypto/x509"
+	"errors"
+	"fmt"
+	"net"
+	"os"
+	"strings"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/daemon"
+)
+
+const defaultDialTimeout = 15 * time.Second
+
+// RemoteConfig configures a remote dial.
+type RemoteConfig struct {
+	// Address is host:port of the remote bridge (required).
+	Address string
+	// Token is the bearer token (required).
+	Token string
+	// CACertFile, when set, is the only CA trusted for the server cert (use for a
+	// self-signed bridge). Empty => system roots. InsecureSkipVerify is never used.
+	CACertFile string
+	// ServerName overrides the TLS/SNI verification name; empty => host of Address.
+	ServerName string
+	// Timeout bounds the dial + auth handshake; 0 => default.
+	Timeout time.Duration
+}
+
+// DialRemote establishes a verified TLS connection to a remote bridge,
+// authenticates with the bearer token, and returns a daemon.Client ready for
+// Run/Attach/Status — the same client the local socket uses, so remote and local
+// share one protocol. The server certificate is always verified (never
+// InsecureSkipVerify).
+func DialRemote(cfg RemoteConfig) (*daemon.Client, error) {
+	address := strings.TrimSpace(cfg.Address)
+	if address == "" {
+		return nil, errors.New("remote: address is required")
+	}
+	token := strings.TrimSpace(cfg.Token)
+	if token == "" {
+		return nil, errors.New("remote: a non-empty token is required")
+	}
+	tlsConfig, err := clientTLSConfig(cfg)
+	if err != nil {
+		return nil, err
+	}
+	timeout := cfg.Timeout
+	if timeout <= 0 {
+		timeout = defaultDialTimeout
+	}
+	// Context-aware TLS dial (tls.DialWithDialer is deprecated and not
+	// cancellation-aware). The context bounds the dial + TLS handshake only; the
+	// returned conn is independent of it for the subsequent auth/stream I/O.
+	dialCtx, cancel := context.WithTimeout(context.Background(), timeout)
+	defer cancel()
+	dialer := &tls.Dialer{NetDialer: &net.Dialer{}, Config: tlsConfig}
+	conn, err := dialer.DialContext(dialCtx, "tcp", address)
+	if err != nil {
+		return nil, fmt.Errorf("remote: dial %s: %w", address, err)
+	}
+	// Bound the auth handshake; the daemon handshake + stream run without a deadline.
+	_ = conn.SetDeadline(time.Now().Add(timeout))
+	if err := writeAuthRequest(conn, authRequest{Token: token, Version: daemon.ProtoVersion}); err != nil {
+		_ = conn.Close()
+		return nil, fmt.Errorf("remote: send auth: %w", err)
+	}
+	resp, err := readAuthResponse(conn)
+	if err != nil {
+		_ = conn.Close()
+		return nil, fmt.Errorf("remote: auth handshake: %w", err)
+	}
+	if !resp.OK {
+		_ = conn.Close()
+		return nil, fmt.Errorf("%w: %s", ErrUnauthorized, resp.Message)
+	}
+	_ = conn.SetDeadline(time.Time{})
+	return daemon.NewClientConn(conn)
+}
+
+// clientTLSConfig builds a verifying client TLS config. It never disables
+// verification; for a self-signed bridge, set CACertFile to the bridge's cert.
+func clientTLSConfig(cfg RemoteConfig) (*tls.Config, error) {
+	tc := &tls.Config{MinVersion: tls.VersionTLS12}
+	serverName := strings.TrimSpace(cfg.ServerName)
+	if serverName == "" {
+		if host, _, err := net.SplitHostPort(cfg.Address); err == nil {
+			serverName = host
+		}
+	}
+	tc.ServerName = serverName
+	if ca := strings.TrimSpace(cfg.CACertFile); ca != "" {
+		pem, err := os.ReadFile(ca)
+		if err != nil {
+			return nil, fmt.Errorf("remote: read CA cert: %w", err)
+		}
+		pool := x509.NewCertPool()
+		if !pool.AppendCertsFromPEM(pem) {
+			return nil, errors.New("remote: CA cert file contained no certificates")
+		}
+		tc.RootCAs = pool
+	}
+	return tc, nil
+}

internal/daemon/server.go

@@ -0,0 +1,301 @@
+package daemon
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"net"
+	"os"
+	"sync"
+	"time"
+)
+
+// Server is the daemon control plane. Mirrors reference-daemon-code-agent-js/
+// supervisor.js (single-instance lock, status file, lifecycle) + the accept loop
+// that routes framed control requests to the SessionManager/Pool. It listens on
+// an owner-only local Unix socket and NEVER binds a TCP port.
+type Server struct {
+	opts      ServerOptions
+	startedAt time.Time
+
+	listener net.Listener
+	lock     *fileLock
+
+	ctx    context.Context
+	cancel context.CancelFunc
+
+	wg           sync.WaitGroup
+	shutdownOnce sync.Once
+	done         chan struct{}
+}
+
+// ServerOptions configures a Server.
+type ServerOptions struct {
+	Paths   Paths
+	Manager *SessionManager
+	Pool    *Pool
+	Version int
+	Now     func() time.Time
+	Log     func(string)
+	isAlive func(int) bool // test hook for the single-instance lock
+}
+
+// NewServer validates options and builds a Server.
+func NewServer(opts ServerOptions) (*Server, error) {
+	if opts.Manager == nil || opts.Pool == nil {
+		return nil, errors.New("daemon: server requires a Pool and SessionManager")
+	}
+	if opts.Paths.Socket == "" || opts.Paths.Lock == "" || opts.Paths.Status == "" {
+		return nil, errors.New("daemon: server requires socket, lock, and status paths")
+	}
+	if opts.Version <= 0 {
+		opts.Version = ProtoVersion
+	}
+	if opts.Now == nil {
+		opts.Now = time.Now
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	return &Server{
+		opts:   opts,
+		ctx:    ctx,
+		cancel: cancel,
+		done:   make(chan struct{}),
+	}, nil
+}
+
+func (s *Server) logf(format string, args ...any) {
+	if s.opts.Log != nil {
+		s.opts.Log(fmt.Sprintf(format, args...))
+	}
+}
+
+// Serve acquires the single-instance lock, binds the owner-only control socket,
+// writes the status file, and serves connections until Shutdown. It blocks. On
+// return it has released the lock and removed the socket/status files.
+func (s *Server) Serve() error {
+	if err := checkSocketPathLength(s.opts.Paths.Socket); err != nil {
+		return err
+	}
+	if err := secureSocketParent(s.opts.Paths.Socket); err != nil {
+		return err
+	}
+	lock, err := acquireLock(s.opts.Paths.Lock, s.opts.isAlive)
+	if err != nil {
+		return err
+	}
+	s.lock = lock
+	defer s.cleanup()
+
+	// A leftover socket file from an unclean exit would make Listen fail with
+	// "address already in use"; we hold the lock, so any socket here is stale.
+	_ = os.Remove(s.opts.Paths.Socket)
+
+	listener, err := net.Listen("unix", s.opts.Paths.Socket)
+	if err != nil {
+		return fmt.Errorf("daemon: bind control socket: %w", err)
+	}
+	s.listener = listener
+	if err := hardenSocketFile(s.opts.Paths.Socket); err != nil {
+		return fmt.Errorf("daemon: harden control socket: %w", err)
+	}
+	s.startedAt = s.opts.Now()
+	if err := s.writeStatusFile(); err != nil {
+		return err
+	}
+	s.logf("daemon listening on %s (pid %d)", s.opts.Paths.Socket, os.Getpid())
+
+	for {
+		conn, err := listener.Accept()
+		if err != nil {
+			select {
+			case <-s.done:
+				s.wg.Wait()
+				return nil // clean shutdown
+			default:
+				// Transient accept error during normal operation.
+				s.logf("accept error: %v", err)
+				return fmt.Errorf("daemon: accept: %w", err)
+			}
+		}
+		s.wg.Add(1)
+		go func() {
+			defer s.wg.Done()
+			s.handleConn(conn)
+		}()
+	}
+}
+
+// Shutdown stops accepting connections, cancels in-flight runs, drains the pool,
+// and removes the socket/lock/status files. Safe to call multiple times.
+func (s *Server) Shutdown() {
+	s.shutdownOnce.Do(func() {
+		close(s.done)
+		s.cancel() // stop in-flight pool runs
+		if s.listener != nil {
+			_ = s.listener.Close()
+		}
+		s.opts.Pool.Drain()
+	})
+}
+
+func (s *Server) cleanup() {
+	if s.listener != nil {
+		_ = s.listener.Close()
+	}
+	_ = os.Remove(s.opts.Paths.Socket)
+	_ = os.Remove(s.opts.Paths.Status)
+	if s.lock != nil {
+		_ = s.lock.release()
+	}
+}
+
+func (s *Server) writeStatusFile() error {
+	status := StatusFile{
+		PID:       os.Getpid(),
+		Socket:    s.opts.Paths.Socket,
+		Version:   s.opts.Version,
+		StartedAt: s.startedAt,
+	}
+	data, err := json.MarshalIndent(status, "", "  ")
+	if err != nil {
+		return err
+	}
+	if err := os.WriteFile(s.opts.Paths.Status, data, 0o600); err != nil {
+		return fmt.Errorf("daemon: write status file: %w", err)
+	}
+	return nil
+}
+
+// ServeConn runs the control protocol (handshake + one command) on an
+// already-established connection, reusing the exact local dispatch path. The
+// remote bridge calls it AFTER authenticating a TLS connection, so a remote
+// session is handled identically to a local one (same SessionManager/Pool, same
+// sandbox/risk model) — remote never bypasses the local controls. It closes conn.
+func (s *Server) ServeConn(conn net.Conn) { s.handleConn(conn) }
+
+// handleConn performs the handshake then dispatches a single control command.
+func (s *Server) handleConn(conn net.Conn) {
+	defer conn.Close()
+
+	hello, err := ReadControl(conn)
+	if err != nil {
+		return
+	}
+	if hello.Type != CtrlHello {
+		_ = WriteControl(conn, Ctrl{Type: CtrlError, Message: "expected hello"})
+		return
+	}
+	version, ok := NegotiateVersion(hello.Version)
+	if !ok {
+		_ = WriteControl(conn, Ctrl{Type: CtrlError, Message: "unsupported protocol version"})
+		return
+	}
+	if err := WriteControl(conn, Ctrl{Type: CtrlHelloOK, Version: version}); err != nil {
+		return
+	}
+
+	cmd, err := ReadControl(conn)
+	if err != nil {
+		return
+	}
+	switch cmd.Type {
+	case CtrlRun:
+		s.handleRun(conn, cmd)
+	case CtrlAttach:
+		s.handleAttach(conn, cmd)
+	case CtrlStatus:
+		s.handleStatus(conn)
+	case CtrlShutdown:
+		_ = WriteControl(conn, Ctrl{Type: CtrlAck, Message: "shutting down"})
+		s.Shutdown()
+	default:
+		_ = WriteControl(conn, Ctrl{Type: CtrlError, Message: fmt.Sprintf("unknown command %q", cmd.Type)})
+	}
+}
+
+func (s *Server) handleRun(conn net.Conn, cmd Ctrl) {
+	if cmd.Session == "" {
+		_ = WriteControl(conn, Ctrl{Type: CtrlError, Message: "run requires a session id"})
+		return
+	}
+	args := cmd.Args
+	if cmd.Prompt != "" {
+		args = append(args, "--prompt", cmd.Prompt)
+	}
+	sess, err := s.opts.Manager.Start(s.ctx, WorkerSpec{Session: cmd.Session, Cwd: cmd.Cwd, Args: args})
+	if err != nil {
+		_ = WriteControl(conn, Ctrl{Type: CtrlError, Message: err.Error()})
+		return
+	}
+	_ = WriteControl(conn, Ctrl{Type: CtrlAck, Session: sess.ID()})
+	buffered, live, cancel := sess.Subscribe()
+	defer cancel()
+	s.streamToClient(conn, sess, buffered, live)
+}
+
+func (s *Server) handleAttach(conn net.Conn, cmd Ctrl) {
+	if cmd.Session == "" {
+		_ = WriteControl(conn, Ctrl{Type: CtrlError, Message: "attach requires a session id"})
+		return
+	}
+	buffered, live, cancel, err := s.opts.Manager.Attach(cmd.Session)
+	if err != nil {
+		_ = WriteControl(conn, Ctrl{Type: CtrlError, Message: err.Error()})
+		return
+	}
+	defer cancel()
+	sess, _ := s.opts.Manager.Get(cmd.Session)
+	_ = WriteControl(conn, Ctrl{Type: CtrlAck, Session: cmd.Session})
+	s.streamToClient(conn, sess, buffered, live)
+}
+
+// streamToClient writes the buffered history then live lines as CtrlData frames,
+// finishing with CtrlEnd (or CtrlError if the session failed). A write error
+// (client disconnected) ends the stream without affecting the session.
+func (s *Server) streamToClient(conn net.Conn, sess *Session, buffered []string, live <-chan string) {
+	for _, line := range buffered {
+		if err := WriteControl(conn, Ctrl{Type: CtrlData, Line: line}); err != nil {
+			return
+		}
+	}
+	for {
+		select {
+		case line, ok := <-live:
+			if !ok {
+				s.finishStream(conn, sess)
+				return
+			}
+			if err := WriteControl(conn, Ctrl{Type: CtrlData, Line: line}); err != nil {
+				return
+			}
+		case <-s.done:
+			_ = WriteControl(conn, Ctrl{Type: CtrlEnd, Message: "daemon shutting down"})
+			return
+		}
+	}
+}
+
+func (s *Server) finishStream(conn net.Conn, sess *Session) {
+	if sess != nil {
+		if err := sess.Err(); err != nil {
+			_ = WriteControl(conn, Ctrl{Type: CtrlError, Session: sess.ID(), Message: err.Error()})
+			return
+		}
+	}
+	_ = WriteControl(conn, Ctrl{Type: CtrlEnd})
+}
+
+func (s *Server) handleStatus(conn net.Conn) {
+	report := StatusReport{
+		PID:        os.Getpid(),
+		Version:    s.opts.Version,
+		Socket:     s.opts.Paths.Socket,
+		StartedAt:  s.startedAt,
+		PoolSize:   s.opts.Pool.Size(),
+		Workers:    s.opts.Pool.WorkerStats(),
+		Sessions:   s.opts.Manager.Statuses(),
+		QueueDepth: s.opts.Pool.QueueDepth(),
+	}
+	_ = WriteControl(conn, Ctrl{Type: CtrlStatusResult, Status: &report})
+}

internal/daemon/server_test.go

@@ -0,0 +1,176 @@
+package daemon
+
+import (
+	"errors"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+)
+
+func newTestServer(t *testing.T, launcher Launcher) (*Server, Paths) {
+	t.Helper()
+	dir := t.TempDir()
+	paths := Paths{
+		Socket: filepath.Join(dir, "d.sock"),
+		Lock:   filepath.Join(dir, "d.lock"),
+		Status: filepath.Join(dir, "d.status"),
+	}
+	pool, err := NewPool(PoolOptions{Size: 2, Launcher: launcher, KillTimeout: 200 * time.Millisecond})
+	if err != nil {
+		t.Fatalf("NewPool: %v", err)
+	}
+	mgr, err := NewSessionManager(SessionManagerOptions{Pool: pool})
+	if err != nil {
+		t.Fatalf("NewSessionManager: %v", err)
+	}
+	srv, err := NewServer(ServerOptions{Paths: paths, Manager: mgr, Pool: pool})
+	if err != nil {
+		t.Fatalf("NewServer: %v", err)
+	}
+	return srv, paths
+}
+
+func waitForFile(t *testing.T, path string) {
+	t.Helper()
+	deadline := time.Now().Add(2 * time.Second)
+	for time.Now().Before(deadline) {
+		if _, err := os.Stat(path); err == nil {
+			return
+		}
+		time.Sleep(2 * time.Millisecond)
+	}
+	t.Fatalf("file %s did not appear within timeout", path)
+}
+
+func TestServerEndToEnd(t *testing.T) {
+	out := []string{`{"type":"event","seq":1}`, `{"type":"event","seq":2}`}
+	launcher, _ := seqLauncher(&fakeWorker{pid: 1, out: out, exitCode: 0})
+	srv, paths := newTestServer(t, launcher)
+
+	serveErr := make(chan error, 1)
+	go func() { serveErr <- srv.Serve() }()
+	waitForFile(t, paths.Status)
+
+	// --- run a session and collect its stream-json output ---
+	runClient, err := Dial(paths.Socket)
+	if err != nil {
+		t.Fatalf("Dial(run): %v", err)
+	}
+	var got []string
+	if err := runClient.Run("s1", "", "hello", nil, func(line string) { got = append(got, line) }); err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	runClient.Close()
+	if len(got) != 2 || got[0] != out[0] || got[1] != out[1] {
+		t.Fatalf("streamed lines = %v, want %v", got, out)
+	}
+
+	// --- status reflects the session ---
+	statusClient, err := Dial(paths.Socket)
+	if err != nil {
+		t.Fatalf("Dial(status): %v", err)
+	}
+	report, err := statusClient.Status()
+	statusClient.Close()
+	if err != nil {
+		t.Fatalf("Status: %v", err)
+	}
+	if report.PoolSize != 2 {
+		t.Fatalf("status PoolSize = %d, want 2", report.PoolSize)
+	}
+	foundSession := false
+	for _, s := range report.Sessions {
+		if s.ID == "s1" {
+			foundSession = true
+			if s.State != string(SessionDone) {
+				t.Fatalf("session s1 state = %s, want done", s.State)
+			}
+		}
+	}
+	if !foundSession {
+		t.Fatalf("status missing session s1: %+v", report.Sessions)
+	}
+
+	// --- attach to the finished session sees the buffered history ---
+	attachClient, err := Dial(paths.Socket)
+	if err != nil {
+		t.Fatalf("Dial(attach): %v", err)
+	}
+	var attached []string
+	if err := attachClient.Attach("s1", func(line string) { attached = append(attached, line) }); err != nil {
+		t.Fatalf("Attach: %v", err)
+	}
+	attachClient.Close()
+	if len(attached) != 2 {
+		t.Fatalf("attach lines = %v, want 2 buffered", attached)
+	}
+
+	// --- graceful stop, daemon cleans up ---
+	stopClient, err := Dial(paths.Socket)
+	if err != nil {
+		t.Fatalf("Dial(stop): %v", err)
+	}
+	if err := stopClient.Shutdown(); err != nil {
+		t.Fatalf("Shutdown: %v", err)
+	}
+	stopClient.Close()
+
+	select {
+	case err := <-serveErr:
+		if err != nil {
+			t.Fatalf("Serve returned error: %v", err)
+		}
+	case <-time.After(3 * time.Second):
+		t.Fatal("Serve did not return after shutdown")
+	}
+	// Socket, status, and lock files are removed on exit.
+	for _, p := range []string{paths.Socket, paths.Status, paths.Lock} {
+		if _, err := os.Stat(p); !errors.Is(err, os.ErrNotExist) {
+			t.Fatalf("file %s not cleaned up after shutdown: %v", p, err)
+		}
+	}
+}
+
+func TestServerSecondInstanceFails(t *testing.T) {
+	block := make(chan struct{})
+	defer close(block)
+	launcher, _ := seqLauncher(&fakeWorker{pid: 1, waitCh: block})
+	srv1, paths := newTestServer(t, launcher)
+
+	serveErr := make(chan error, 1)
+	go func() { serveErr <- srv1.Serve() }()
+	waitForFile(t, paths.Status)
+
+	// A second server on the same paths must refuse to start (single instance).
+	pool2, _ := NewPool(PoolOptions{Size: 1, Launcher: launcher})
+	mgr2, _ := NewSessionManager(SessionManagerOptions{Pool: pool2})
+	srv2, _ := NewServer(ServerOptions{Paths: paths, Manager: mgr2, Pool: pool2})
+	if err := srv2.Serve(); !errors.Is(err, ErrAlreadyRunning) {
+		t.Fatalf("second Serve err = %v, want ErrAlreadyRunning", err)
+	}
+
+	srv1.Shutdown()
+	<-serveErr
+}
+
+func TestServerRejectsUnknownCommand(t *testing.T) {
+	block := make(chan struct{})
+	defer close(block)
+	launcher, _ := seqLauncher(&fakeWorker{pid: 1, waitCh: block})
+	srv, paths := newTestServer(t, launcher)
+	go func() { _ = srv.Serve() }()
+	defer srv.Shutdown()
+	waitForFile(t, paths.Status)
+
+	client, err := Dial(paths.Socket)
+	if err != nil {
+		t.Fatalf("Dial: %v", err)
+	}
+	defer client.Close()
+	// Run with an empty session id is rejected with an error frame.
+	err = client.Run("", "", "hi", nil, nil)
+	if err == nil {
+		t.Fatal("run with empty session id must return an error")
+	}
+}

internal/daemon/session.go

@@ -0,0 +1,306 @@
+package daemon
+
+import (
+	"context"
+	"errors"
+	"sort"
+	"sync"
+)
+
+// Session/lease/roster model. Mirrors reference-daemon-code-agent-js/
+// session-manager.js (sessionID -> worker + metrics), lease.js (one active
+// request per worker) and roster.js (the live session registry). Because zero's
+// `exec` worker is one-shot, each session is served by exactly one worker via
+// Pool.Run, so "one active request per worker" holds inherently and the pool's
+// bounded slots provide the queue-when-full behavior.
+
+// SessionState is a session's lifecycle phase.
+type SessionState string
+
+const (
+	// SessionQueued: created, waiting for a free pool slot.
+	SessionQueued SessionState = "queued"
+	// SessionRunning: a worker is leased and producing output.
+	SessionRunning SessionState = "running"
+	// SessionDone: the worker finished cleanly.
+	SessionDone SessionState = "done"
+	// SessionFailed: the worker failed permanently / the run errored.
+	SessionFailed SessionState = "failed"
+)
+
+// ErrSessionExists is returned when starting a session ID already in use.
+var ErrSessionExists = errors.New("daemon: session already exists")
+
+// ErrSessionNotFound is returned by Get/Attach for an unknown session ID.
+var ErrSessionNotFound = errors.New("daemon: session not found")
+
+const defaultSessionBuffer = 1024
+
+// Session is one routed agent run. It buffers recent output (so a late `attach`
+// sees history) and fans live lines out to subscribers. It implements the pool's
+// Sink (Line) and the lease "started" hook (Started).
+type Session struct {
+	id  string
+	cwd string
+
+	mu          sync.Mutex
+	state       SessionState
+	buffer      []string
+	maxBuffer   int
+	lines       int
+	finished    bool
+	err         error
+	exitCode    int
+	subscribers map[int]chan string
+	nextSub     int
+	done        chan struct{}
+}
+
+func newSession(id, cwd string, maxBuffer int) *Session {
+	if maxBuffer <= 0 {
+		maxBuffer = defaultSessionBuffer
+	}
+	return &Session{
+		id:          id,
+		cwd:         cwd,
+		state:       SessionQueued,
+		maxBuffer:   maxBuffer,
+		subscribers: map[int]chan string{},
+		done:        make(chan struct{}),
+	}
+}
+
+// ID returns the session ID.
+func (s *Session) ID() string { return s.id }
+
+// Done is closed when the session finishes (success or failure).
+func (s *Session) Done() <-chan struct{} { return s.done }
+
+// Err returns the terminal error (nil on clean completion); valid after Done.
+func (s *Session) Err() error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return s.err
+}
+
+// State returns the current lifecycle phase.
+func (s *Session) State() SessionState {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return s.state
+}
+
+// Started marks the session running once the pool leases a worker (lease hook).
+func (s *Session) Started() {
+	s.mu.Lock()
+	if s.state == SessionQueued {
+		s.state = SessionRunning
+	}
+	s.mu.Unlock()
+}
+
+// Line records and broadcasts one stream-json output line.
+func (s *Session) Line(line string) {
+	s.mu.Lock()
+	s.lines++
+	s.buffer = append(s.buffer, line)
+	if len(s.buffer) > s.maxBuffer {
+		// Drop the oldest line; the ring keeps only the most recent maxBuffer.
+		s.buffer = s.buffer[len(s.buffer)-s.maxBuffer:]
+	}
+	subs := make([]chan string, 0, len(s.subscribers))
+	for _, ch := range s.subscribers {
+		subs = append(subs, ch)
+	}
+	s.mu.Unlock()
+	for _, ch := range subs {
+		// Non-blocking: a slow/stalled subscriber must not stall the worker pump.
+		// Dropped live lines remain available via the buffer for a fresh attach.
+		select {
+		case ch <- line:
+		default:
+		}
+	}
+}
+
+func (s *Session) finish(exitCode int, err error) {
+	s.mu.Lock()
+	s.finished = true
+	s.exitCode = exitCode
+	s.err = err
+	if err != nil {
+		s.state = SessionFailed
+	} else {
+		s.state = SessionDone
+	}
+	subs := s.subscribers
+	s.subscribers = map[int]chan string{}
+	s.mu.Unlock()
+	for _, ch := range subs {
+		close(ch)
+	}
+	close(s.done)
+}
+
+// Subscribe returns the buffered history plus a channel of subsequent live lines
+// and a cancel func. If the session has already finished, the live channel is
+// returned already closed. Mirrors attach.js.
+func (s *Session) Subscribe() (buffered []string, live <-chan string, cancel func()) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	buffered = append([]string(nil), s.buffer...)
+	if s.finished {
+		ch := make(chan string)
+		close(ch)
+		return buffered, ch, func() {}
+	}
+	id := s.nextSub
+	s.nextSub++
+	ch := make(chan string, 256)
+	s.subscribers[id] = ch
+	cancel = func() {
+		s.mu.Lock()
+		if c, ok := s.subscribers[id]; ok {
+			delete(s.subscribers, id)
+			close(c)
+		}
+		s.mu.Unlock()
+	}
+	return buffered, ch, cancel
+}
+
+func (s *Session) status() SessionStatus {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return SessionStatus{ID: s.id, State: string(s.state), Lines: s.lines}
+}
+
+func (s *Session) isFinished() bool {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return s.finished
+}
+
+// defaultMaxSessions bounds the retained session registry so a long-running
+// daemon does not accumulate finished sessions without limit.
+const defaultMaxSessions = 256
+
+// SessionManager owns the live session registry and routes each session to a
+// pool worker. Mirrors session-manager.js + roster.js. Finished sessions are
+// retained (so a late `attach` sees history) up to MaxSessions, past which the
+// oldest FINISHED ones are evicted; running/queued sessions are never evicted.
+type SessionManager struct {
+	pool        *Pool
+	maxBuffer   int
+	maxSessions int
+
+	mu       sync.Mutex
+	sessions map[string]*Session
+	order    []string // creation order, for FIFO eviction of finished sessions
+}
+
+// SessionManagerOptions configures a SessionManager.
+type SessionManagerOptions struct {
+	Pool      *Pool
+	MaxBuffer int // per-session output ring size; 0 => default
+	// MaxSessions caps the retained session registry; 0 => default. The oldest
+	// FINISHED sessions are evicted once the cap is exceeded.
+	MaxSessions int
+}
+
+// NewSessionManager builds a manager over pool.
+func NewSessionManager(opts SessionManagerOptions) (*SessionManager, error) {
+	if opts.Pool == nil {
+		return nil, errors.New("daemon: session manager requires a Pool")
+	}
+	maxSessions := opts.MaxSessions
+	if maxSessions <= 0 {
+		maxSessions = defaultMaxSessions
+	}
+	return &SessionManager{
+		pool:        opts.Pool,
+		maxBuffer:   opts.MaxBuffer,
+		maxSessions: maxSessions,
+		sessions:    map[string]*Session{},
+	}, nil
+}
+
+// Start creates a session for spec and dispatches it to the pool. It returns the
+// Session immediately (non-blocking); the run proceeds in the background and the
+// session transitions queued -> running -> done/failed. A duplicate ID is
+// rejected with ErrSessionExists.
+func (m *SessionManager) Start(ctx context.Context, spec WorkerSpec) (*Session, error) {
+	m.mu.Lock()
+	if _, exists := m.sessions[spec.Session]; exists {
+		m.mu.Unlock()
+		return nil, ErrSessionExists
+	}
+	sess := newSession(spec.Session, spec.Cwd, m.maxBuffer)
+	m.sessions[spec.Session] = sess
+	m.order = append(m.order, spec.Session)
+	m.pruneLocked()
+	m.mu.Unlock()
+
+	go func() {
+		code, err := m.pool.Run(ctx, spec, sess)
+		sess.finish(code, err)
+	}()
+	return sess, nil
+}
+
+// pruneLocked evicts the oldest FINISHED sessions while the registry exceeds
+// maxSessions, so a long-running daemon does not retain finished sessions without
+// bound. Running/queued sessions are never evicted (a transient burst can exceed
+// the cap until those finish). Caller must hold m.mu.
+func (m *SessionManager) pruneLocked() {
+	if len(m.sessions) <= m.maxSessions {
+		return
+	}
+	kept := m.order[:0]
+	for _, id := range m.order {
+		s, ok := m.sessions[id]
+		if !ok {
+			continue // already removed
+		}
+		if len(m.sessions) > m.maxSessions && s.isFinished() {
+			delete(m.sessions, id)
+			continue
+		}
+		kept = append(kept, id)
+	}
+	m.order = kept
+}
+
+// Get returns a session by ID.
+func (m *SessionManager) Get(id string) (*Session, bool) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	s, ok := m.sessions[id]
+	return s, ok
+}
+
+// Attach returns the buffered history + live stream for an existing session.
+func (m *SessionManager) Attach(id string) (buffered []string, live <-chan string, cancel func(), err error) {
+	s, ok := m.Get(id)
+	if !ok {
+		return nil, nil, nil, ErrSessionNotFound
+	}
+	b, l, c := s.Subscribe()
+	return b, l, c, nil
+}
+
+// Statuses returns a snapshot of all sessions, sorted by ID for stable output.
+func (m *SessionManager) Statuses() []SessionStatus {
+	m.mu.Lock()
+	sessions := make([]*Session, 0, len(m.sessions))
+	for _, s := range m.sessions {
+		sessions = append(sessions, s)
+	}
+	m.mu.Unlock()
+	out := make([]SessionStatus, 0, len(sessions))
+	for _, s := range sessions {
+		out = append(out, s.status())
+	}
+	sort.Slice(out, func(i, j int) bool { return out[i].ID < out[j].ID })
+	return out
+}

internal/daemon/session_test.go

@@ -0,0 +1,179 @@
+package daemon
+
+import (
+	"context"
+	"errors"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+func drain(t *testing.T, buffered []string, live <-chan string) []string {
+	t.Helper()
+	got := append([]string(nil), buffered...)
+	timeout := time.After(2 * time.Second)
+	for {
+		select {
+		case line, ok := <-live:
+			if !ok {
+				return got
+			}
+			got = append(got, line)
+		case <-timeout:
+			t.Fatal("timed out draining session stream")
+		}
+	}
+}
+
+func TestSessionStartRoutesAndStreams(t *testing.T) {
+	launcher, _ := seqLauncher(&fakeWorker{pid: 1, out: []string{"e1", "e2", "e3"}, exitCode: 0})
+	pool, _ := NewPool(PoolOptions{Size: 2, Launcher: launcher})
+	mgr, err := NewSessionManager(SessionManagerOptions{Pool: pool})
+	if err != nil {
+		t.Fatalf("NewSessionManager: %v", err)
+	}
+	sess, err := mgr.Start(context.Background(), WorkerSpec{Session: "s1"})
+	if err != nil {
+		t.Fatalf("Start: %v", err)
+	}
+	buffered, live, cancel := sess.Subscribe()
+	defer cancel()
+	got := drain(t, buffered, live)
+	if len(got) != 3 || got[0] != "e1" || got[2] != "e3" {
+		t.Fatalf("streamed lines = %v, want [e1 e2 e3]", got)
+	}
+	<-sess.Done()
+	if sess.State() != SessionDone {
+		t.Fatalf("state = %s, want done", sess.State())
+	}
+}
+
+func TestSessionDuplicateIDRejected(t *testing.T) {
+	block := make(chan struct{})
+	defer close(block)
+	launcher, _ := seqLauncher(&fakeWorker{pid: 1, waitCh: block})
+	pool, _ := NewPool(PoolOptions{Size: 2, Launcher: launcher})
+	mgr, _ := NewSessionManager(SessionManagerOptions{Pool: pool})
+	if _, err := mgr.Start(context.Background(), WorkerSpec{Session: "dup"}); err != nil {
+		t.Fatalf("first Start: %v", err)
+	}
+	if _, err := mgr.Start(context.Background(), WorkerSpec{Session: "dup"}); !errors.Is(err, ErrSessionExists) {
+		t.Fatalf("second Start err = %v, want ErrSessionExists", err)
+	}
+}
+
+func TestSessionLeaseQueuesWhenPoolFull(t *testing.T) {
+	block := make(chan struct{})
+	a := &fakeWorker{pid: 1, waitCh: block}
+	b := &fakeWorker{pid: 2, out: []string{"b1"}, exitCode: 0}
+	launcher, calls := seqLauncher(a, b)
+	pool, _ := NewPool(PoolOptions{Size: 1, Launcher: launcher})
+	mgr, _ := NewSessionManager(SessionManagerOptions{Pool: pool})
+
+	sa, _ := mgr.Start(context.Background(), WorkerSpec{Session: "a"})
+	waitFor(t, func() bool { return sa.State() == SessionRunning })
+
+	sb, _ := mgr.Start(context.Background(), WorkerSpec{Session: "b"})
+	// b must stay queued (its worker not yet launched) while a holds the slot.
+	time.Sleep(50 * time.Millisecond)
+	if sb.State() != SessionQueued {
+		t.Fatalf("session b state = %s, want queued while pool full", sb.State())
+	}
+	if n := atomic.LoadInt32(calls); n != 1 {
+		t.Fatalf("launcher calls = %d, want 1 (b not launched yet)", n)
+	}
+
+	close(block) // a finishes, slot frees, b proceeds
+	<-sb.Done()
+	if sb.State() != SessionDone {
+		t.Fatalf("session b state = %s, want done", sb.State())
+	}
+	if n := atomic.LoadInt32(calls); n != 2 {
+		t.Fatalf("launcher calls = %d, want 2 after slot freed", n)
+	}
+}
+
+func TestSessionAttachAfterFinishSeesBuffer(t *testing.T) {
+	launcher, _ := seqLauncher(&fakeWorker{pid: 1, out: []string{"x", "y"}, exitCode: 0})
+	pool, _ := NewPool(PoolOptions{Size: 1, Launcher: launcher})
+	mgr, _ := NewSessionManager(SessionManagerOptions{Pool: pool})
+	sess, _ := mgr.Start(context.Background(), WorkerSpec{Session: "s"})
+	<-sess.Done()
+
+	buffered, live, cancel, err := mgr.Attach("s")
+	if err != nil {
+		t.Fatalf("Attach: %v", err)
+	}
+	defer cancel()
+	got := drain(t, buffered, live)
+	if len(got) != 2 || got[0] != "x" || got[1] != "y" {
+		t.Fatalf("attach buffered = %v, want [x y]", got)
+	}
+	if _, _, _, err := mgr.Attach("missing"); !errors.Is(err, ErrSessionNotFound) {
+		t.Fatalf("Attach(missing) err = %v, want ErrSessionNotFound", err)
+	}
+}
+
+func TestSessionManagerEvictsFinishedOverCap(t *testing.T) {
+	launcher, _ := seqLauncher(
+		&fakeWorker{pid: 1, out: []string{"a"}, exitCode: 0},
+		&fakeWorker{pid: 2, out: []string{"b"}, exitCode: 0},
+		&fakeWorker{pid: 3, out: []string{"c"}, exitCode: 0},
+	)
+	pool, _ := NewPool(PoolOptions{Size: 2, Launcher: launcher})
+	mgr, _ := NewSessionManager(SessionManagerOptions{Pool: pool, MaxSessions: 2})
+
+	for _, id := range []string{"s1", "s2", "s3"} {
+		s, err := mgr.Start(context.Background(), WorkerSpec{Session: id})
+		if err != nil {
+			t.Fatalf("Start(%s): %v", id, err)
+		}
+		<-s.Done() // finish before starting the next so prune can evict it
+	}
+
+	// Past the cap of 2, the oldest FINISHED session (s1) is evicted; the two
+	// newest are retained.
+	if _, ok := mgr.Get("s1"); ok {
+		t.Fatal("oldest finished session must be evicted past the cap")
+	}
+	if _, ok := mgr.Get("s2"); !ok {
+		t.Fatal("s2 must be retained")
+	}
+	if _, ok := mgr.Get("s3"); !ok {
+		t.Fatal("s3 (newest) must be retained")
+	}
+}
+
+func TestSessionManagerKeepsRunningOverCap(t *testing.T) {
+	// All sessions are running (workers block), so none are evicted even past cap.
+	block := make(chan struct{})
+	defer close(block)
+	launcher, _ := seqLauncher(
+		&fakeWorker{pid: 1, waitCh: block},
+		&fakeWorker{pid: 2, waitCh: block},
+	)
+	pool, _ := NewPool(PoolOptions{Size: 2, Launcher: launcher})
+	mgr, _ := NewSessionManager(SessionManagerOptions{Pool: pool, MaxSessions: 1})
+
+	s1, _ := mgr.Start(context.Background(), WorkerSpec{Session: "r1"})
+	s2, _ := mgr.Start(context.Background(), WorkerSpec{Session: "r2"})
+	waitFor(t, func() bool { return s1.State() == SessionRunning && s2.State() == SessionRunning })
+	if _, ok := mgr.Get("r1"); !ok {
+		t.Fatal("a running session must never be evicted, even past the cap")
+	}
+	if _, ok := mgr.Get("r2"); !ok {
+		t.Fatal("a running session must never be evicted, even past the cap")
+	}
+}
+
+func TestSessionStatuses(t *testing.T) {
+	launcher, _ := seqLauncher(&fakeWorker{pid: 1, out: []string{"l"}, exitCode: 0})
+	pool, _ := NewPool(PoolOptions{Size: 1, Launcher: launcher})
+	mgr, _ := NewSessionManager(SessionManagerOptions{Pool: pool})
+	sess, _ := mgr.Start(context.Background(), WorkerSpec{Session: "only"})
+	<-sess.Done()
+	st := mgr.Statuses()
+	if len(st) != 1 || st[0].ID != "only" || st[0].State != string(SessionDone) || st[0].Lines != 1 {
+		t.Fatalf("statuses = %+v, want one done session with 1 line", st)
+	}
+}

internal/daemon/socket.go

@@ -0,0 +1,27 @@
+package daemon
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+)
+
+// maxUnixSocketPath bounds the socket path to the smallest platform sun_path
+// limit so we surface a clear error instead of a cryptic bind failure. macOS
+// allows 104 bytes incl. the NUL (so 103 chars); Linux allows 108. 103 is the
+// safe cross-platform ceiling.
+const maxUnixSocketPath = 103
+
+// secureSocketParent creates the socket's parent directory owner-only (0700 on
+// POSIX; on Windows the per-user profile directory is already ACL-restricted).
+func secureSocketParent(socketPath string) error {
+	return os.MkdirAll(filepath.Dir(socketPath), 0o700)
+}
+
+// checkSocketPathLength rejects an over-long unix socket path before bind.
+func checkSocketPathLength(socketPath string) error {
+	if len(socketPath) > maxUnixSocketPath {
+		return fmt.Errorf("daemon: socket path too long (%d > %d): %s", len(socketPath), maxUnixSocketPath, socketPath)
+	}
+	return nil
+}

internal/daemon/socket_posix.go

@@ -0,0 +1,12 @@
+//go:build !windows
+
+package daemon
+
+import "os"
+
+// hardenSocketFile restricts the control socket to its owner (0600) so no other
+// local user can connect. Combined with the 0700 parent directory this makes the
+// control plane owner-only, as required by the security model.
+func hardenSocketFile(path string) error {
+	return os.Chmod(path, 0o600)
+}

internal/daemon/socket_windows.go

@@ -0,0 +1,9 @@
+//go:build windows
+
+package daemon
+
+// hardenSocketFile is a no-op on Windows: AF_UNIX socket files do not honor POSIX
+// mode bits, and the daemon places the socket under the per-user profile
+// directory, which is already ACL-restricted to the owner. A dedicated restricted
+// ACL (or a named pipe with an explicit DACL) is a documented follow-up.
+func hardenSocketFile(string) error { return nil }

internal/daemon/status.go

@@ -0,0 +1,43 @@
+package daemon
+
+import "time"
+
+// StatusReport is the daemon/worker/session snapshot returned by `zero daemon
+// status` (CtrlStatusResult). Mirrors reference-daemon-code-agent-js/status.js.
+type StatusReport struct {
+	PID        int             `json:"pid"`
+	Version    int             `json:"version"`
+	Socket     string          `json:"socket"`
+	StartedAt  time.Time       `json:"startedAt"`
+	PoolSize   int             `json:"poolSize"`
+	Workers    []WorkerStatus  `json:"workers"`
+	Sessions   []SessionStatus `json:"sessions"`
+	QueueDepth int             `json:"queueDepth"`
+}
+
+// WorkerStatus is one worker's lifecycle snapshot.
+type WorkerStatus struct {
+	ID                 int    `json:"id"`
+	PID                int    `json:"pid"`
+	State              string `json:"state"`
+	ConsecutiveCrashes int    `json:"consecutiveCrashes"`
+	Restarts           int    `json:"restarts"`
+	Session            string `json:"session,omitempty"`
+}
+
+// SessionStatus is one session's routing/metrics snapshot.
+type SessionStatus struct {
+	ID    string `json:"id"`
+	State string `json:"state"`
+	Lines int    `json:"lines"`
+}
+
+// StatusFile is the on-disk daemon status document (pid, socket, version,
+// started-at), written next to the lock so `status`/`stop` work without a live
+// connection. Mirrors supervisor.js's status file.
+type StatusFile struct {
+	PID       int       `json:"pid"`
+	Socket    string    `json:"socket"`
+	Version   int       `json:"version"`
+	StartedAt time.Time `json:"startedAt"`
+}

internal/doctor/doctor.go

@@ -158,6 +158,9 @@ func providerModelCheck(profile config.ProviderProfile) Check {
 	if emptyProviderProfile(profile) {
 		return check("provider.model", "Provider model", StatusWarn, "Model validity was skipped because provider config is unavailable.", nil)
 	}
+	if strings.TrimSpace(profile.Model) == "" {
+		return check("provider.model", "Provider model", StatusFail, "Provider model is required.", map[string]any{"provider": providerName(profile)})
+	}
 	registry, err := modelregistry.DefaultRegistry()
 	if err != nil {
 		return check("provider.model", "Provider model", StatusFail, "Model registry could not be loaded: "+err.Error(), nil)
@@ -165,7 +168,7 @@ func providerModelCheck(profile config.ProviderProfile) Check {
 	model, err := registry.Require(profile.Model)
 	if err != nil {
 		if profile.ProviderKind == config.ProviderKindOpenAICompatible || profile.ProviderKind == config.ProviderKindAnthropicCompat {
-			return check("provider.model", "Provider model", StatusWarn, fmt.Sprintf("Custom %s model was not found in the Zero registry; runtime will pass it through to the configured provider.", profile.ProviderKind), map[string]any{"model": profile.Model, "provider": providerName(profile)})
+			return check("provider.model", "Provider model", StatusWarn, fmt.Sprintf("Custom %s model was not found in the Zero registry; runtime will pass it through to the configured provider. Run `zero doctor --connectivity` to validate the endpoint and auth.", profile.ProviderKind), map[string]any{"model": profile.Model, "provider": providerName(profile)})
 		}
 		return check("provider.model", "Provider model", StatusFail, "Provider model is invalid: "+err.Error(), map[string]any{"model": profile.Model})
 	}

internal/doctor/doctor_test.go

@@ -86,6 +86,96 @@ func TestRunReportWarnsForUnknownOpenAICompatibleModel(t *testing.T) {
 	}
 }
 
+func TestProviderModelUnknownOpenAICompatibleModelWarnsPassThrough(t *testing.T) {
+	report := Run(Options{
+		Now:     fixedDoctorClock("2026-06-04T15:45:00Z"),
+		Runtime: "go",
+		Provider: config.ProviderProfile{
+			Name:         "openrouter",
+			ProviderKind: config.ProviderKindOpenAICompatible,
+			BaseURL:      "https://openrouter.ai/api/v1",
+			Model:        "vendor/new-dynamic-model",
+		},
+	})
+
+	if !report.OK {
+		t.Fatalf("unknown openai-compatible model should warn, not fail: %#v", report)
+	}
+	check := report.Check("provider.model")
+	if check == nil {
+		t.Fatalf("expected provider.model check: %#v", report.Checks)
+	}
+	if check.Status != StatusWarn {
+		t.Fatalf("provider.model status = %s, want warn: %#v", check.Status, check)
+	}
+	if !strings.Contains(check.Message, "runtime will pass it through") || !strings.Contains(check.Message, "doctor --connectivity") {
+		t.Fatalf("provider.model message should explain pass-through connectivity validation, got %q", check.Message)
+	}
+}
+
+func TestProviderModelMissingModelFails(t *testing.T) {
+	report := Run(Options{
+		Now:     fixedDoctorClock("2026-06-04T15:45:00Z"),
+		Runtime: "go",
+		Provider: config.ProviderProfile{
+			Name:         "ollama-cloud",
+			ProviderKind: config.ProviderKindOpenAICompatible,
+			BaseURL:      "https://ollama.com/v1",
+			Model:        "   ",
+		},
+	})
+
+	if report.OK {
+		t.Fatalf("missing model should fail: %#v", report)
+	}
+	check := report.Check("provider.model")
+	if check == nil || check.Status != StatusFail || !strings.Contains(check.Message, "required") {
+		t.Fatalf("expected missing model failure: %#v", report.Checks)
+	}
+}
+
+func TestProviderModelBuiltInUnknownModelFails(t *testing.T) {
+	report := Run(Options{
+		Now:     fixedDoctorClock("2026-06-04T15:45:00Z"),
+		Runtime: "go",
+		Provider: config.ProviderProfile{
+			Name:         "openai",
+			ProviderKind: config.ProviderKindOpenAI,
+			BaseURL:      config.OpenAIBaseURL,
+			Model:        "vendor/new-dynamic-model",
+		},
+	})
+
+	if report.OK {
+		t.Fatalf("built-in provider unknown model should fail: %#v", report)
+	}
+	check := report.Check("provider.model")
+	if check == nil || check.Status != StatusFail || !strings.Contains(check.Message, "unknown Zero model") {
+		t.Fatalf("expected built-in unknown model failure: %#v", report.Checks)
+	}
+}
+
+func TestProviderModelRegisteredProviderMismatchFails(t *testing.T) {
+	report := Run(Options{
+		Now:     fixedDoctorClock("2026-06-04T15:45:00Z"),
+		Runtime: "go",
+		Provider: config.ProviderProfile{
+			Name:         "openai",
+			ProviderKind: config.ProviderKindOpenAI,
+			BaseURL:      config.OpenAIBaseURL,
+			Model:        "gemini-2.5-pro",
+		},
+	})
+
+	if report.OK {
+		t.Fatalf("registered provider mismatch should fail: %#v", report)
+	}
+	check := report.Check("provider.model")
+	if check == nil || check.Status != StatusFail || !strings.Contains(check.Message, "not available for provider") {
+		t.Fatalf("expected registered provider mismatch failure: %#v", report.Checks)
+	}
+}
+
 func writeDoctorConfig(t *testing.T, body string) string {
 	t.Helper()
 	dir := t.TempDir()

internal/mcp/oauth.go

@@ -2,9 +2,6 @@ package mcp
 
 import (
 	"context"
-	"crypto/rand"
-	"crypto/sha256"
-	"encoding/base64"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -14,20 +11,56 @@ import (
 	"net/url"
 	"strings"
 	"time"
+
+	"github.com/Gitlawb/zero/internal/oauth"
 )
 
 // ServerAuthOAuth is the value of an MCP server's auth field that selects the
 // OAuth 2.0 + PKCE authorization-code flow.
 const ServerAuthOAuth = "oauth"
 
-const (
-	wellKnownAuthServerPath = "/.well-known/oauth-authorization-server"
-	defaultLoginTimeout     = 3 * time.Minute
-	// pkceVerifierBytes yields a 43-character base64url verifier, the minimum
-	// length permitted by the PKCE spec while remaining high-entropy.
-	pkceVerifierBytes = 32
-	stateBytes        = 32
-)
+const defaultLoginTimeout = 3 * time.Minute
+
+// MCP OAuth delegates its transport/identity-agnostic engine to internal/oauth
+// (PKCE, RFC 8414 discovery, authorize-URL build, token exchange/refresh) so the
+// two share one implementation. MCP keeps its own LoginOptions/Login
+// orchestration, OAuthConfig/StoredToken types, loopback handling, CLI, and
+// on-disk token format — all behavior-preserving. The shared engine also adds an
+// https-only token-endpoint guard (loopback exempt), a hardening MCP inherits.
+
+// pkceToParams converts the shared PKCE pair to MCP's local type.
+func pkceToParams(p oauth.PKCE) pkceParams {
+	return pkceParams{Verifier: p.Verifier, Challenge: p.Challenge, Method: p.Method}
+}
+
+// pkceToOAuth converts MCP's local PKCE type back to the shared type.
+func pkceToOAuth(p pkceParams) oauth.PKCE {
+	return oauth.PKCE{Verifier: p.Verifier, Challenge: p.Challenge, Method: p.Method}
+}
+
+// configFor builds the shared oauth.Config from MCP's OAuthConfig + resolved
+// endpoints for a flow.
+func configFor(cfg OAuthConfig, metadata authServerMetadata) oauth.Config {
+	return oauth.Config{
+		ClientID:              cfg.ClientID,
+		ClientSecret:          cfg.ClientSecret,
+		Scopes:                cfg.Scopes,
+		AuthorizationEndpoint: metadata.AuthorizationEndpoint,
+		TokenEndpoint:         metadata.TokenEndpoint,
+	}
+}
+
+// tokenToStored converts a shared oauth.Token to MCP's StoredToken (MCP does not
+// use the Account field).
+func tokenToStored(t oauth.Token) StoredToken {
+	return StoredToken{
+		AccessToken:  t.AccessToken,
+		RefreshToken: t.RefreshToken,
+		TokenType:    t.TokenType,
+		Scopes:       t.Scopes,
+		ExpiresAt:    t.ExpiresAt,
+	}
+}
 
 // OAuthConfig describes how to authenticate to a remote MCP server using OAuth.
 // Endpoints may be discovered from the server's metadata document; explicit
@@ -84,34 +117,20 @@ type authorizationFlow struct {
 	now        func() time.Time
 }
 
-// discoverAuthorizationServer fetches the OAuth 2.0 authorization server
-// metadata document published at the well-known path under the given base URL.
+// discoverAuthorizationServer fetches the RFC 8414 authorization server metadata
+// at the well-known path under baseURL, via the shared engine.
 func discoverAuthorizationServer(ctx context.Context, client *http.Client, baseURL string) (authServerMetadata, error) {
-	if client == nil {
-		client = http.DefaultClient
-	}
-	metadataURL, err := joinWellKnown(baseURL)
+	meta, err := oauth.DiscoverAuthorizationServer(ctx, client, baseURL)
 	if err != nil {
 		return authServerMetadata{}, err
 	}
-	request, err := http.NewRequestWithContext(ctx, http.MethodGet, metadataURL, nil)
-	if err != nil {
-		return authServerMetadata{}, err
-	}
-	request.Header.Set("Accept", "application/json")
-	response, err := client.Do(request)
-	if err != nil {
-		return authServerMetadata{}, fmt.Errorf("fetch authorization server metadata: %w", err)
-	}
-	defer response.Body.Close()
-	if response.StatusCode < http.StatusOK || response.StatusCode >= http.StatusMultipleChoices {
-		return authServerMetadata{}, fmt.Errorf("authorization server metadata returned HTTP %d", response.StatusCode)
-	}
-	var metadata authServerMetadata
-	if err := json.NewDecoder(io.LimitReader(response.Body, 1<<20)).Decode(&metadata); err != nil {
-		return authServerMetadata{}, fmt.Errorf("decode authorization server metadata: %w", err)
-	}
-	return metadata, nil
+	return authServerMetadata{
+		Issuer:                meta.Issuer,
+		AuthorizationEndpoint: meta.AuthorizationEndpoint,
+		TokenEndpoint:         meta.TokenEndpoint,
+		RegistrationEndpoint:  meta.RegistrationEndpoint,
+		ScopesSupported:       meta.ScopesSupported,
+	}, nil
 }
 
 // resolveAuthorizationServer discovers metadata and applies explicit config
@@ -149,44 +168,23 @@ func resolveAuthorizationServer(ctx context.Context, client *http.Client, baseUR
 	return metadata, nil
 }
 
-// newPKCE generates a high-entropy code verifier and its S256 challenge.
+// newPKCE generates a high-entropy code verifier and its S256 challenge via the
+// shared engine.
 func newPKCE() (pkceParams, error) {
-	verifierRaw := make([]byte, pkceVerifierBytes)
-	if _, err := rand.Read(verifierRaw); err != nil {
-		return pkceParams{}, fmt.Errorf("generate PKCE verifier: %w", err)
+	p, err := oauth.NewPKCE()
+	if err != nil {
+		return pkceParams{}, err
 	}
-	verifier := base64.RawURLEncoding.EncodeToString(verifierRaw)
-	sum := sha256.Sum256([]byte(verifier))
-	challenge := base64.RawURLEncoding.EncodeToString(sum[:])
-	return pkceParams{Verifier: verifier, Challenge: challenge, Method: "S256"}, nil
+	return pkceToParams(p), nil
 }
 
 func newState() (string, error) {
-	raw := make([]byte, stateBytes)
-	if _, err := rand.Read(raw); err != nil {
-		return "", fmt.Errorf("generate state: %w", err)
-	}
-	return base64.RawURLEncoding.EncodeToString(raw), nil
+	return oauth.NewState()
 }
 
-// authorizationURL builds the authorization request URL.
+// authorizationURL builds the authorization request URL via the shared engine.
 func (flow *authorizationFlow) authorizationURL(redirectURI string) (string, error) {
-	parsed, err := url.Parse(flow.metadata.AuthorizationEndpoint)
-	if err != nil {
-		return "", fmt.Errorf("parse authorization endpoint: %w", err)
-	}
-	query := parsed.Query()
-	query.Set("response_type", "code")
-	query.Set("client_id", flow.config.ClientID)
-	query.Set("redirect_uri", redirectURI)
-	query.Set("state", flow.state)
-	query.Set("code_challenge", flow.pkce.Challenge)
-	query.Set("code_challenge_method", flow.pkce.Method)
-	if len(flow.config.Scopes) > 0 {
-		query.Set("scope", strings.Join(flow.config.Scopes, " "))
-	}
-	parsed.RawQuery = query.Encode()
-	return parsed.String(), nil
+	return oauth.BuildAuthorizationURL(configFor(flow.config, flow.metadata), pkceToOAuth(flow.pkce), flow.state, redirectURI, nil)
 }
 
 // parseCallback validates the redirect query and returns the authorization
@@ -209,117 +207,30 @@ func (flow *authorizationFlow) parseCallback(values url.Values) (string, error)
 	return code, nil
 }
 
-// exchangeCode swaps an authorization code + PKCE verifier for tokens.
+// exchangeCode swaps an authorization code + PKCE verifier for tokens via the
+// shared engine.
 func (flow *authorizationFlow) exchangeCode(ctx context.Context, code string, redirectURI string) (StoredToken, error) {
-	form := url.Values{}
-	form.Set("grant_type", "authorization_code")
-	form.Set("code", code)
-	form.Set("redirect_uri", redirectURI)
-	form.Set("client_id", flow.config.ClientID)
-	form.Set("code_verifier", flow.pkce.Verifier)
-	if secret := strings.TrimSpace(flow.config.ClientSecret); secret != "" {
-		form.Set("client_secret", secret)
+	token, err := oauth.ExchangeCode(ctx, flow.httpClient, configFor(flow.config, flow.metadata), code, flow.pkce.Verifier, redirectURI, flow.now)
+	if err != nil {
+		return StoredToken{}, err
 	}
-	return postTokenRequest(ctx, flow.httpClient, flow.metadata.TokenEndpoint, form, StoredToken{Scopes: flow.config.Scopes}, flow.now)
+	return tokenToStored(token), nil
 }
 
-// refreshAccessToken exchanges a refresh token for a fresh access token. A
-// response that omits a new refresh token preserves the previous one.
+// refreshAccessToken exchanges a refresh token for a fresh access token via the
+// shared engine. A response that omits a new refresh token preserves the
+// previous one.
 func refreshAccessToken(ctx context.Context, client *http.Client, cfg OAuthConfig, current StoredToken, now func() time.Time) (StoredToken, error) {
-	refresh := strings.TrimSpace(current.RefreshToken)
-	if refresh == "" {
-		return StoredToken{}, errors.New("no refresh token available")
-	}
-	tokenEndpoint := strings.TrimSpace(cfg.TokenEndpoint)
-	if tokenEndpoint == "" {
-		return StoredToken{}, errors.New("no token endpoint configured for refresh")
-	}
-	form := url.Values{}
-	form.Set("grant_type", "refresh_token")
-	form.Set("refresh_token", refresh)
-	form.Set("client_id", cfg.ClientID)
-	if secret := strings.TrimSpace(cfg.ClientSecret); secret != "" {
-		form.Set("client_secret", secret)
-	}
-	if len(cfg.Scopes) > 0 {
-		form.Set("scope", strings.Join(cfg.Scopes, " "))
-	}
-	refreshed, err := postTokenRequest(ctx, client, tokenEndpoint, form, StoredToken{Scopes: current.Scopes, RefreshToken: refresh}, now)
+	token, err := oauth.Refresh(ctx, client, oauth.Config{
+		ClientID:      cfg.ClientID,
+		ClientSecret:  cfg.ClientSecret,
+		Scopes:        cfg.Scopes,
+		TokenEndpoint: cfg.TokenEndpoint,
+	}, oauth.Token{AccessToken: current.AccessToken, RefreshToken: current.RefreshToken, TokenType: current.TokenType, Scopes: current.Scopes, ExpiresAt: current.ExpiresAt}, now)
 	if err != nil {
 		return StoredToken{}, err
 	}
-	return refreshed, nil
-}
-
-// tokenResponse is the JSON returned by the token endpoint.
-type tokenResponse struct {
-	AccessToken      string `json:"access_token"`
-	RefreshToken     string `json:"refresh_token"`
-	TokenType        string `json:"token_type"`
-	ExpiresIn        int64  `json:"expires_in"`
-	Scope            string `json:"scope"`
-	Error            string `json:"error"`
-	ErrorDescription string `json:"error_description"`
-}
-
-// postTokenRequest performs a token endpoint POST and maps the response onto a
-// StoredToken. The base token supplies values to preserve (e.g. an existing
-// refresh token or scope set) when the response omits them.
-func postTokenRequest(ctx context.Context, client *http.Client, tokenEndpoint string, form url.Values, base StoredToken, now func() time.Time) (StoredToken, error) {
-	if client == nil {
-		client = http.DefaultClient
-	}
-	if now == nil {
-		now = time.Now
-	}
-	request, err := http.NewRequestWithContext(ctx, http.MethodPost, tokenEndpoint, strings.NewReader(form.Encode()))
-	if err != nil {
-		return StoredToken{}, err
-	}
-	request.Header.Set("Content-Type", "application/x-www-form-urlencoded")
-	request.Header.Set("Accept", "application/json")
-
-	response, err := client.Do(request)
-	if err != nil {
-		return StoredToken{}, fmt.Errorf("token request failed: %w", err)
-	}
-	defer response.Body.Close()
-
-	body, _ := io.ReadAll(io.LimitReader(response.Body, 1<<20))
-	var parsed tokenResponse
-	if len(body) > 0 {
-		// A malformed body on an error status is reported as a status error below.
-		_ = json.Unmarshal(body, &parsed)
-	}
-
-	if response.StatusCode < http.StatusOK || response.StatusCode >= http.StatusMultipleChoices {
-		if parsed.Error != "" {
-			if parsed.ErrorDescription != "" {
-				return StoredToken{}, fmt.Errorf("token endpoint error %q: %s", parsed.Error, parsed.ErrorDescription)
-			}
-			return StoredToken{}, fmt.Errorf("token endpoint error %q", parsed.Error)
-		}
-		return StoredToken{}, fmt.Errorf("token endpoint returned HTTP %d", response.StatusCode)
-	}
-	if strings.TrimSpace(parsed.AccessToken) == "" {
-		return StoredToken{}, errors.New("token endpoint returned no access token")
-	}
-
-	token := base
-	token.AccessToken = parsed.AccessToken
-	if strings.TrimSpace(parsed.RefreshToken) != "" {
-		token.RefreshToken = parsed.RefreshToken
-	}
-	if strings.TrimSpace(parsed.TokenType) != "" {
-		token.TokenType = parsed.TokenType
-	}
-	if parsed.ExpiresIn > 0 {
-		token.ExpiresAt = now().Add(time.Duration(parsed.ExpiresIn) * time.Second).UTC()
-	}
-	if scope := strings.TrimSpace(parsed.Scope); scope != "" {
-		token.Scopes = strings.Fields(scope)
-	}
-	return token, nil
+	return tokenToStored(token), nil
 }
 
 // registerClient performs dynamic client registration against the registration
@@ -505,22 +416,3 @@ func Login(ctx context.Context, options LoginOptions) (StoredToken, error) {
 		return StoredToken{}, fmt.Errorf("timed out waiting for OAuth authorization callback: %w", loginCtx.Err())
 	}
 }
-
-func joinWellKnown(baseURL string) (string, error) {
-	parsed, err := url.Parse(strings.TrimSpace(baseURL))
-	if err != nil || parsed.Host == "" {
-		return "", fmt.Errorf("invalid base URL for metadata discovery: %q", baseURL)
-	}
-	// RFC 8414: the well-known segment is inserted between the host and any issuer
-	// path component, so a path-based issuer (https://host/tenant-a) is probed at
-	// https://host/.well-known/oauth-authorization-server/tenant-a — not at the
-	// host root (which would break tenant-scoped discovery).
-	issuerPath := strings.Trim(parsed.Path, "/")
-	parsed.Path = wellKnownAuthServerPath
-	if issuerPath != "" {
-		parsed.Path = strings.TrimRight(wellKnownAuthServerPath, "/") + "/" + issuerPath
-	}
-	parsed.RawQuery = ""
-	parsed.Fragment = ""
-	return parsed.String(), nil
-}

internal/oauth/device.go

@@ -0,0 +1,213 @@
+package oauth
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"net/url"
+	"strings"
+	"time"
+)
+
+const deviceGrantType = "urn:ietf:params:oauth:grant-type:device_code"
+
+// defaultDeviceCodeLifetime bounds a device authorization that omits expires_in,
+// so the expiry gate in PollDeviceToken is always active (fail closed).
+const defaultDeviceCodeLifetime = 10 * time.Minute
+
+// DeviceAuth is the result of an RFC 8628 device-authorization request.
+type DeviceAuth struct {
+	DeviceCode              string
+	UserCode                string
+	VerificationURI         string
+	VerificationURIComplete string
+	Interval                time.Duration
+	ExpiresAt               time.Time
+}
+
+type deviceAuthResponse struct {
+	DeviceCode              string `json:"device_code"`
+	UserCode                string `json:"user_code"`
+	VerificationURI         string `json:"verification_uri"`
+	VerificationURIComplete string `json:"verification_uri_complete"`
+	ExpiresIn               int64  `json:"expires_in"`
+	Interval                int64  `json:"interval"`
+	Error                   string `json:"error"`
+	ErrorDescription        string `json:"error_description"`
+}
+
+// RequestDeviceCode performs the RFC 8628 device-authorization request. This is
+// the headless/SSH login path (no browser, no loopback).
+func RequestDeviceCode(ctx context.Context, client *http.Client, cfg Config, now func() time.Time) (DeviceAuth, error) {
+	endpoint := trimmed(cfg.DeviceAuthorizationEndpoint)
+	if endpoint == "" {
+		return DeviceAuth{}, errors.New("oauth: provider has no device authorization endpoint")
+	}
+	if err := validateTokenEndpoint(endpoint); err != nil {
+		return DeviceAuth{}, err
+	}
+	if client == nil {
+		client = http.DefaultClient
+	}
+	if now == nil {
+		now = time.Now
+	}
+	form := url.Values{}
+	form.Set("client_id", cfg.ClientID)
+	if secret := trimmed(cfg.ClientSecret); secret != "" {
+		// Confidential clients must authenticate on the device endpoint too,
+		// consistent with the token poll (pollDeviceOnce).
+		form.Set("client_secret", secret)
+	}
+	if len(cfg.Scopes) > 0 {
+		form.Set("scope", strings.Join(cfg.Scopes, " "))
+	}
+	request, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, strings.NewReader(form.Encode()))
+	if err != nil {
+		return DeviceAuth{}, err
+	}
+	request.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+	request.Header.Set("Accept", "application/json")
+	response, err := client.Do(request)
+	if err != nil {
+		return DeviceAuth{}, fmt.Errorf("oauth: device authorization request failed: %w", err)
+	}
+	defer response.Body.Close()
+	body, _ := io.ReadAll(io.LimitReader(response.Body, tokenResponseLimit))
+	var parsed deviceAuthResponse
+	if len(body) > 0 {
+		_ = json.Unmarshal(body, &parsed)
+	}
+	if response.StatusCode < http.StatusOK || response.StatusCode >= http.StatusMultipleChoices {
+		if parsed.Error != "" {
+			return DeviceAuth{}, fmt.Errorf("oauth: device authorization error %q", parsed.Error)
+		}
+		return DeviceAuth{}, fmt.Errorf("oauth: device authorization returned HTTP %d", response.StatusCode)
+	}
+	if trimmed(parsed.DeviceCode) == "" || trimmed(parsed.UserCode) == "" {
+		return DeviceAuth{}, errors.New("oauth: device authorization response missing device_code/user_code")
+	}
+	interval := time.Duration(parsed.Interval) * time.Second
+	if interval <= 0 {
+		interval = 5 * time.Second // RFC 8628 default
+	}
+	auth := DeviceAuth{
+		DeviceCode:              parsed.DeviceCode,
+		UserCode:                parsed.UserCode,
+		VerificationURI:         parsed.VerificationURI,
+		VerificationURIComplete: parsed.VerificationURIComplete,
+		Interval:                interval,
+	}
+	lifetime := time.Duration(parsed.ExpiresIn) * time.Second
+	if lifetime <= 0 {
+		// Fail closed: a response without expires_in must still get a bounded
+		// lifetime so the poll loop's expiry gate stays active.
+		lifetime = defaultDeviceCodeLifetime
+	}
+	auth.ExpiresAt = now().Add(lifetime)
+	return auth, nil
+}
+
+// PollDeviceToken polls the token endpoint for the device grant until the user
+// approves, the device code expires, ctx is done, or a terminal error occurs. It
+// honors authorization_pending (keep waiting), slow_down (increase the
+// interval), and expired_token.
+func PollDeviceToken(ctx context.Context, client *http.Client, cfg Config, auth DeviceAuth, now func() time.Time) (Token, error) {
+	if now == nil {
+		now = time.Now
+	}
+	interval := auth.Interval
+	if interval <= 0 {
+		interval = 5 * time.Second
+	}
+	for {
+		if !auth.ExpiresAt.IsZero() && !auth.ExpiresAt.After(now()) {
+			return Token{}, errors.New("oauth: device code expired before authorization")
+		}
+		select {
+		case <-ctx.Done():
+			return Token{}, fmt.Errorf("oauth: device authorization canceled: %w", ctx.Err())
+		case <-time.After(interval):
+		}
+		// Re-check expiry after sleeping: slow_down (or a long interval) can push
+		// the next poll past ExpiresAt, and we must not poll after the deadline.
+		if !auth.ExpiresAt.IsZero() && !auth.ExpiresAt.After(now()) {
+			return Token{}, errors.New("oauth: device code expired before authorization")
+		}
+		token, err := pollDeviceOnce(ctx, client, cfg, auth.DeviceCode, now)
+		switch {
+		case err == nil:
+			return token, nil
+		case errors.Is(err, ErrAuthorizationPending):
+			// keep waiting at the current interval
+		case errors.Is(err, ErrSlowDown):
+			interval += 5 * time.Second // RFC 8628: add 5s on slow_down
+		default:
+			return Token{}, err
+		}
+	}
+}
+
+// pollDeviceOnce makes a single device-token poll, translating the RFC 8628
+// error codes into ErrAuthorizationPending / ErrSlowDown / terminal errors.
+func pollDeviceOnce(ctx context.Context, client *http.Client, cfg Config, deviceCode string, now func() time.Time) (Token, error) {
+	if err := validateTokenEndpoint(cfg.TokenEndpoint); err != nil {
+		return Token{}, err
+	}
+	if client == nil {
+		client = http.DefaultClient
+	}
+	form := url.Values{}
+	form.Set("grant_type", deviceGrantType)
+	form.Set("device_code", deviceCode)
+	form.Set("client_id", cfg.ClientID)
+	if secret := trimmed(cfg.ClientSecret); secret != "" {
+		form.Set("client_secret", secret)
+	}
+	request, err := http.NewRequestWithContext(ctx, http.MethodPost, trimmed(cfg.TokenEndpoint), strings.NewReader(form.Encode()))
+	if err != nil {
+		return Token{}, err
+	}
+	request.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+	request.Header.Set("Accept", "application/json")
+	response, err := client.Do(request)
+	if err != nil {
+		return Token{}, fmt.Errorf("oauth: device token poll failed: %w", err)
+	}
+	defer response.Body.Close()
+	body, _ := io.ReadAll(io.LimitReader(response.Body, tokenResponseLimit))
+	var parsed tokenResponse
+	if len(body) > 0 {
+		_ = json.Unmarshal(body, &parsed)
+	}
+	if response.StatusCode >= http.StatusOK && response.StatusCode < http.StatusMultipleChoices && trimmed(parsed.AccessToken) != "" {
+		token := Token{Scopes: cfg.Scopes}
+		token.AccessToken = parsed.AccessToken
+		token.RefreshToken = parsed.RefreshToken
+		token.TokenType = parsed.TokenType
+		if parsed.ExpiresIn > 0 {
+			token.ExpiresAt = now().Add(time.Duration(parsed.ExpiresIn) * time.Second).UTC()
+		}
+		if scope := trimmed(parsed.Scope); scope != "" {
+			token.Scopes = strings.Fields(scope)
+		}
+		return token, nil
+	}
+	switch parsed.Error {
+	case "authorization_pending":
+		return Token{}, ErrAuthorizationPending
+	case "slow_down":
+		return Token{}, ErrSlowDown
+	case "expired_token":
+		return Token{}, errors.New("oauth: device code expired before authorization")
+	case "access_denied":
+		return Token{}, errors.New("oauth: authorization denied by the user")
+	case "":
+		return Token{}, fmt.Errorf("oauth: device token poll returned HTTP %d", response.StatusCode)
+	default:
+		return Token{}, fmt.Errorf("oauth: device token error %q", parsed.Error)
+	}
+}

internal/oauth/device_test.go

@@ -0,0 +1,143 @@
+package oauth
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+func TestRequestDeviceCodeSendsClientSecret(t *testing.T) {
+	var gotSecret string
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_ = r.ParseForm()
+		gotSecret = r.Form.Get("client_secret")
+		_, _ = w.Write([]byte(`{"device_code":"dc","user_code":"U","verification_uri":"https://x","expires_in":600,"interval":5}`))
+	}))
+	defer server.Close()
+	// A confidential client must authenticate on the device endpoint too.
+	_, err := RequestDeviceCode(context.Background(), server.Client(), Config{ClientID: "c", ClientSecret: "shh", DeviceAuthorizationEndpoint: server.URL}, nil)
+	if err != nil {
+		t.Fatalf("RequestDeviceCode: %v", err)
+	}
+	if gotSecret != "shh" {
+		t.Fatalf("client_secret = %q, want shh", gotSecret)
+	}
+}
+
+func TestRequestDeviceCode(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		_, _ = w.Write([]byte(`{"device_code":"dc","user_code":"WXYZ-1234","verification_uri":"https://x/dev","verification_uri_complete":"https://x/dev?c=WXYZ","expires_in":600,"interval":3}`))
+	}))
+	defer server.Close()
+	cfg := Config{ClientID: "c", DeviceAuthorizationEndpoint: server.URL, Scopes: []string{"a"}}
+	auth, err := RequestDeviceCode(context.Background(), server.Client(), cfg, nil)
+	if err != nil {
+		t.Fatalf("RequestDeviceCode: %v", err)
+	}
+	if auth.DeviceCode != "dc" || auth.UserCode != "WXYZ-1234" {
+		t.Fatalf("device auth = %+v", auth)
+	}
+	if auth.Interval != 3*time.Second {
+		t.Fatalf("interval = %v, want 3s", auth.Interval)
+	}
+	if auth.VerificationURIComplete == "" {
+		t.Fatal("missing verification_uri_complete")
+	}
+}
+
+func TestRequestDeviceCodeDefaultsInterval(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		_, _ = w.Write([]byte(`{"device_code":"dc","user_code":"U","verification_uri":"https://x"}`))
+	}))
+	defer server.Close()
+	auth, err := RequestDeviceCode(context.Background(), server.Client(), Config{ClientID: "c", DeviceAuthorizationEndpoint: server.URL}, nil)
+	if err != nil {
+		t.Fatalf("RequestDeviceCode: %v", err)
+	}
+	if auth.Interval != 5*time.Second {
+		t.Fatalf("default interval = %v, want 5s", auth.Interval)
+	}
+	// A response without expires_in must still get a bounded expiry (fail closed),
+	// so the poll loop's expiry gate stays active.
+	if auth.ExpiresAt.IsZero() {
+		t.Fatal("missing expires_in must default to a bounded ExpiresAt, got zero")
+	}
+}
+
+func TestPollDeviceTokenHonorsPendingThenSucceeds(t *testing.T) {
+	var calls atomic.Int32
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		n := calls.Add(1)
+		if n < 3 {
+			w.WriteHeader(http.StatusBadRequest)
+			_, _ = w.Write([]byte(`{"error":"authorization_pending"}`))
+			return
+		}
+		_, _ = w.Write([]byte(`{"access_token":"at","token_type":"Bearer","expires_in":3600}`))
+	}))
+	defer server.Close()
+	cfg := Config{ClientID: "c", TokenEndpoint: server.URL}
+	// Small interval + future expiry so the loop is fast.
+	auth := DeviceAuth{DeviceCode: "dc", Interval: 5 * time.Millisecond, ExpiresAt: time.Now().Add(5 * time.Second)}
+	tok, err := PollDeviceToken(context.Background(), server.Client(), cfg, auth, nil)
+	if err != nil {
+		t.Fatalf("PollDeviceToken: %v", err)
+	}
+	if tok.AccessToken != "at" {
+		t.Fatalf("token = %+v", tok)
+	}
+	if calls.Load() != 3 {
+		t.Fatalf("polled %d times, want 3 (2 pending + 1 success)", calls.Load())
+	}
+}
+
+func TestPollDeviceTokenSlowDown(t *testing.T) {
+	var calls atomic.Int32
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		if calls.Add(1) == 1 {
+			w.WriteHeader(http.StatusBadRequest)
+			_, _ = w.Write([]byte(`{"error":"slow_down"}`))
+			return
+		}
+		_, _ = w.Write([]byte(`{"access_token":"at"}`))
+	}))
+	defer server.Close()
+	cfg := Config{ClientID: "c", TokenEndpoint: server.URL}
+	// Expiry well beyond the slow_down back-off (+5s) so the success path never
+	// depends on polling after the device code has expired.
+	auth := DeviceAuth{DeviceCode: "dc", Interval: 5 * time.Millisecond, ExpiresAt: time.Now().Add(30 * time.Second)}
+	tok, err := PollDeviceToken(context.Background(), server.Client(), cfg, auth, nil)
+	if err != nil {
+		t.Fatalf("PollDeviceToken (slow_down): %v", err)
+	}
+	if tok.AccessToken != "at" {
+		t.Fatalf("token = %+v", tok)
+	}
+}
+
+func TestPollDeviceTokenExpired(t *testing.T) {
+	cfg := Config{ClientID: "c", TokenEndpoint: "https://unused/token"}
+	auth := DeviceAuth{DeviceCode: "dc", Interval: time.Millisecond, ExpiresAt: time.Now().Add(-time.Second)}
+	_, err := PollDeviceToken(context.Background(), http.DefaultClient, cfg, auth, nil)
+	if err == nil {
+		t.Fatal("expected expired error")
+	}
+}
+
+func TestPollDeviceTokenAccessDenied(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(http.StatusBadRequest)
+		_, _ = w.Write([]byte(`{"error":"access_denied"}`))
+	}))
+	defer server.Close()
+	cfg := Config{ClientID: "c", TokenEndpoint: server.URL}
+	auth := DeviceAuth{DeviceCode: "dc", Interval: 5 * time.Millisecond, ExpiresAt: time.Now().Add(5 * time.Second)}
+	_, err := PollDeviceToken(context.Background(), server.Client(), cfg, auth, nil)
+	if err == nil || !strings.Contains(err.Error(), "denied") {
+		t.Fatalf("err = %v, want access denied", err)
+	}
+}

internal/oauth/discovery.go

@@ -0,0 +1,79 @@
+package oauth
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"net/url"
+	"strings"
+)
+
+const wellKnownAuthServerPath = "/.well-known/oauth-authorization-server"
+
+// ServerMetadata is the subset of an RFC 8414 authorization-server metadata
+// document the engine consumes.
+type ServerMetadata struct {
+	Issuer                      string   `json:"issuer"`
+	AuthorizationEndpoint       string   `json:"authorization_endpoint"`
+	TokenEndpoint               string   `json:"token_endpoint"`
+	DeviceAuthorizationEndpoint string   `json:"device_authorization_endpoint"`
+	RegistrationEndpoint        string   `json:"registration_endpoint"`
+	ScopesSupported             []string `json:"scopes_supported"`
+}
+
+// DiscoverAuthorizationServer fetches the RFC 8414 metadata document at the
+// well-known path under baseURL.
+func DiscoverAuthorizationServer(ctx context.Context, client *http.Client, baseURL string) (ServerMetadata, error) {
+	metadataURL, err := joinWellKnown(baseURL)
+	if err != nil {
+		return ServerMetadata{}, err
+	}
+	return fetchMetadata(ctx, client, metadataURL)
+}
+
+// fetchMetadata GETs and decodes an authorization-server metadata document at an
+// already-resolved URL (no well-known recomputation), so callers can also use it
+// for the OIDC openid-configuration path.
+func fetchMetadata(ctx context.Context, client *http.Client, metadataURL string) (ServerMetadata, error) {
+	if client == nil {
+		client = http.DefaultClient
+	}
+	request, err := http.NewRequestWithContext(ctx, http.MethodGet, metadataURL, nil)
+	if err != nil {
+		return ServerMetadata{}, err
+	}
+	request.Header.Set("Accept", "application/json")
+	response, err := client.Do(request)
+	if err != nil {
+		return ServerMetadata{}, fmt.Errorf("oauth: fetch authorization server metadata: %w", err)
+	}
+	defer response.Body.Close()
+	if response.StatusCode < http.StatusOK || response.StatusCode >= http.StatusMultipleChoices {
+		return ServerMetadata{}, fmt.Errorf("oauth: authorization server metadata returned HTTP %d", response.StatusCode)
+	}
+	var metadata ServerMetadata
+	if err := json.NewDecoder(io.LimitReader(response.Body, tokenResponseLimit)).Decode(&metadata); err != nil {
+		return ServerMetadata{}, fmt.Errorf("oauth: decode authorization server metadata: %w", err)
+	}
+	return metadata, nil
+}
+
+// joinWellKnown builds the RFC 8414 metadata URL, inserting the well-known
+// segment between host and any issuer path component (so a path-based issuer is
+// probed tenant-scoped, not at the host root).
+func joinWellKnown(baseURL string) (string, error) {
+	parsed, err := url.Parse(trimmed(baseURL))
+	if err != nil || parsed.Host == "" {
+		return "", fmt.Errorf("oauth: invalid base URL for metadata discovery: %q", baseURL)
+	}
+	issuerPath := strings.Trim(parsed.Path, "/")
+	parsed.Path = wellKnownAuthServerPath
+	if issuerPath != "" {
+		parsed.Path = strings.TrimRight(wellKnownAuthServerPath, "/") + "/" + issuerPath
+	}
+	parsed.RawQuery = ""
+	parsed.Fragment = ""
+	return parsed.String(), nil
+}

internal/oauth/flow.go

@@ -0,0 +1,212 @@
+package oauth
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"net"
+	"net/http"
+	"net/url"
+	"strings"
+	"time"
+)
+
+const tokenResponseLimit = 1 << 20 // 1 MiB cap on token-endpoint bodies
+
+// tokenResponse is the JSON returned by a token endpoint.
+type tokenResponse struct {
+	AccessToken      string `json:"access_token"`
+	RefreshToken     string `json:"refresh_token"`
+	TokenType        string `json:"token_type"`
+	ExpiresIn        int64  `json:"expires_in"`
+	Scope            string `json:"scope"`
+	Error            string `json:"error"`
+	ErrorDescription string `json:"error_description"`
+}
+
+// BuildAuthorizationURL constructs the authorization request URL for an
+// authorization-code + PKCE flow. PKCE S256 is always included; a non-S256
+// method is refused (ErrPKCEDowngrade). extraParams (and cfg.ExtraAuthParams)
+// are appended last.
+func BuildAuthorizationURL(cfg Config, pkce PKCE, state, redirectURI string, extraParams map[string]string) (string, error) {
+	if pkce.Method != MethodS256 {
+		return "", ErrPKCEDowngrade
+	}
+	endpoint := trimmed(cfg.AuthorizationEndpoint)
+	if endpoint == "" {
+		return "", errors.New("oauth: no authorization endpoint configured")
+	}
+	parsed, err := url.Parse(endpoint)
+	if err != nil {
+		return "", fmt.Errorf("oauth: parse authorization endpoint: %w", err)
+	}
+	query := parsed.Query()
+	query.Set("response_type", "code")
+	query.Set("client_id", cfg.ClientID)
+	query.Set("redirect_uri", redirectURI)
+	query.Set("state", state)
+	query.Set("code_challenge", pkce.Challenge)
+	query.Set("code_challenge_method", pkce.Method)
+	if len(cfg.Scopes) > 0 {
+		query.Set("scope", strings.Join(cfg.Scopes, " "))
+	}
+	// Extra params must never override the reserved OAuth/PKCE fields above
+	// (e.g. forcing code_challenge_method=plain or rewriting state/redirect_uri),
+	// which would break the flow's security guarantees.
+	for k, v := range cfg.ExtraAuthParams {
+		if isReservedAuthParam(k) {
+			continue
+		}
+		query.Set(k, v)
+	}
+	for k, v := range extraParams {
+		if isReservedAuthParam(k) {
+			continue
+		}
+		query.Set(k, v)
+	}
+	parsed.RawQuery = query.Encode()
+	return parsed.String(), nil
+}
+
+// isReservedAuthParam reports whether a query key is one the engine controls and
+// must not be overridable by caller-supplied extra params.
+func isReservedAuthParam(key string) bool {
+	switch key {
+	case "response_type", "client_id", "redirect_uri", "state", "code_challenge", "code_challenge_method":
+		return true
+	default:
+		return false
+	}
+}
+
+// validateTokenEndpoint refuses to send a credential to a token endpoint that is
+// not https, unless it is a loopback host (mirrors oauth2-client.js
+// validateTokenEndpointURL). This prevents leaking codes/tokens over cleartext.
+func validateTokenEndpoint(endpoint string) error {
+	parsed, err := url.Parse(trimmed(endpoint))
+	if err != nil || parsed.Host == "" {
+		return fmt.Errorf("oauth: invalid token endpoint %q", endpoint)
+	}
+	if parsed.Scheme == "https" {
+		return nil
+	}
+	if parsed.Scheme == "http" && isLoopbackHost(parsed.Hostname()) {
+		return nil
+	}
+	return fmt.Errorf("%w: %s", ErrInsecureTokenEndpoint, parsed.Scheme+"://"+parsed.Host)
+}
+
+func isLoopbackHost(host string) bool {
+	if host == "localhost" {
+		return true
+	}
+	if ip := net.ParseIP(host); ip != nil {
+		return ip.IsLoopback()
+	}
+	return false
+}
+
+// ExchangeCode swaps an authorization code + PKCE verifier for tokens.
+func ExchangeCode(ctx context.Context, client *http.Client, cfg Config, code, verifier, redirectURI string, now func() time.Time) (Token, error) {
+	form := url.Values{}
+	form.Set("grant_type", "authorization_code")
+	form.Set("code", code)
+	form.Set("redirect_uri", redirectURI)
+	form.Set("client_id", cfg.ClientID)
+	form.Set("code_verifier", verifier)
+	if secret := trimmed(cfg.ClientSecret); secret != "" {
+		form.Set("client_secret", secret)
+	}
+	return PostToken(ctx, client, cfg.TokenEndpoint, form, Token{Scopes: cfg.Scopes}, now)
+}
+
+// Refresh exchanges a refresh token for a fresh access token. A response that
+// omits a new refresh token preserves the current one.
+func Refresh(ctx context.Context, client *http.Client, cfg Config, current Token, now func() time.Time) (Token, error) {
+	refresh := trimmed(current.RefreshToken)
+	if refresh == "" {
+		return Token{}, ErrNoRefreshToken
+	}
+	if trimmed(cfg.TokenEndpoint) == "" {
+		return Token{}, errors.New("oauth: no token endpoint configured for refresh")
+	}
+	form := url.Values{}
+	form.Set("grant_type", "refresh_token")
+	form.Set("refresh_token", refresh)
+	form.Set("client_id", cfg.ClientID)
+	if secret := trimmed(cfg.ClientSecret); secret != "" {
+		form.Set("client_secret", secret)
+	}
+	if len(cfg.Scopes) > 0 {
+		form.Set("scope", strings.Join(cfg.Scopes, " "))
+	}
+	base := Token{Scopes: current.Scopes, RefreshToken: refresh, Account: current.Account}
+	return PostToken(ctx, client, cfg.TokenEndpoint, form, base, now)
+}
+
+// PostToken performs a token-endpoint POST and maps the response onto a Token.
+// The https guard is applied first. The base token supplies values to preserve
+// (e.g. an existing refresh token or scopes) when the response omits them.
+// Error messages carry only the server's error/error_description — never the raw
+// body — so token material in an unexpected payload is not leaked.
+func PostToken(ctx context.Context, client *http.Client, tokenEndpoint string, form url.Values, base Token, now func() time.Time) (Token, error) {
+	if err := validateTokenEndpoint(tokenEndpoint); err != nil {
+		return Token{}, err
+	}
+	if client == nil {
+		client = http.DefaultClient
+	}
+	if now == nil {
+		now = time.Now
+	}
+	request, err := http.NewRequestWithContext(ctx, http.MethodPost, trimmed(tokenEndpoint), strings.NewReader(form.Encode()))
+	if err != nil {
+		return Token{}, err
+	}
+	request.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+	request.Header.Set("Accept", "application/json")
+
+	response, err := client.Do(request)
+	if err != nil {
+		return Token{}, fmt.Errorf("oauth: token request failed: %w", err)
+	}
+	defer response.Body.Close()
+
+	body, _ := io.ReadAll(io.LimitReader(response.Body, tokenResponseLimit))
+	var parsed tokenResponse
+	if len(body) > 0 {
+		_ = json.Unmarshal(body, &parsed)
+	}
+
+	if response.StatusCode < http.StatusOK || response.StatusCode >= http.StatusMultipleChoices {
+		if parsed.Error != "" {
+			if parsed.ErrorDescription != "" {
+				return Token{}, fmt.Errorf("oauth: token endpoint error %q: %s", parsed.Error, parsed.ErrorDescription)
+			}
+			return Token{}, fmt.Errorf("oauth: token endpoint error %q", parsed.Error)
+		}
+		return Token{}, fmt.Errorf("oauth: token endpoint returned HTTP %d", response.StatusCode)
+	}
+	if trimmed(parsed.AccessToken) == "" {
+		return Token{}, errors.New("oauth: token endpoint returned no access token")
+	}
+
+	token := base
+	token.AccessToken = parsed.AccessToken
+	if trimmed(parsed.RefreshToken) != "" {
+		token.RefreshToken = parsed.RefreshToken
+	}
+	if trimmed(parsed.TokenType) != "" {
+		token.TokenType = parsed.TokenType
+	}
+	if parsed.ExpiresIn > 0 {
+		token.ExpiresAt = now().Add(time.Duration(parsed.ExpiresIn) * time.Second).UTC()
+	}
+	if scope := trimmed(parsed.Scope); scope != "" {
+		token.Scopes = strings.Fields(scope)
+	}
+	return token, nil
+}

internal/oauth/flow_test.go

@@ -0,0 +1,212 @@
+package oauth
+
+import (
+	"context"
+	"crypto/sha256"
+	"encoding/base64"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"strings"
+	"testing"
+	"time"
+)
+
+func TestNewPKCE(t *testing.T) {
+	p, err := NewPKCE()
+	if err != nil {
+		t.Fatalf("NewPKCE: %v", err)
+	}
+	if p.Method != MethodS256 {
+		t.Fatalf("method = %q, want S256", p.Method)
+	}
+	if len(p.Verifier) != 43 { // base64url(32 bytes) = 43 chars
+		t.Fatalf("verifier length = %d, want 43", len(p.Verifier))
+	}
+	sum := sha256.Sum256([]byte(p.Verifier))
+	want := base64.RawURLEncoding.EncodeToString(sum[:])
+	if p.Challenge != want {
+		t.Fatalf("challenge = %q, want base64url(sha256(verifier))", p.Challenge)
+	}
+}
+
+func TestNewStateUnique(t *testing.T) {
+	seen := map[string]bool{}
+	for i := 0; i < 100; i++ {
+		s, err := NewState()
+		if err != nil {
+			t.Fatalf("NewState: %v", err)
+		}
+		if s == "" || seen[s] {
+			t.Fatalf("state not unique/non-empty: %q", s)
+		}
+		seen[s] = true
+	}
+}
+
+func TestBuildAuthorizationURL(t *testing.T) {
+	cfg := Config{
+		ClientID:              "client-123",
+		Scopes:                []string{"read", "write"},
+		AuthorizationEndpoint: "https://auth.example.com/authorize",
+		ExtraAuthParams:       map[string]string{"login_hint": "[email protected]"},
+	}
+	pkce := PKCE{Verifier: "v", Challenge: "chal", Method: MethodS256}
+	raw, err := BuildAuthorizationURL(cfg, pkce, "state-xyz", "http://127.0.0.1:9/callback", map[string]string{"prompt": "login"})
+	if err != nil {
+		t.Fatalf("BuildAuthorizationURL: %v", err)
+	}
+	u, _ := url.Parse(raw)
+	q := u.Query()
+	checks := map[string]string{
+		"response_type":         "code",
+		"client_id":             "client-123",
+		"redirect_uri":          "http://127.0.0.1:9/callback",
+		"state":                 "state-xyz",
+		"code_challenge":        "chal",
+		"code_challenge_method": "S256",
+		"scope":                 "read write",
+		"login_hint":            "[email protected]",
+		"prompt":                "login",
+	}
+	for k, want := range checks {
+		if got := q.Get(k); got != want {
+			t.Errorf("query %q = %q, want %q", k, got, want)
+		}
+	}
+}
+
+func TestBuildAuthorizationURLIgnoresReservedExtras(t *testing.T) {
+	cfg := Config{
+		ClientID:              "c",
+		AuthorizationEndpoint: "https://auth.example.com/authorize",
+		// A hostile config tries to weaken the flow via extra params.
+		ExtraAuthParams: map[string]string{"state": "attacker", "code_challenge_method": "plain", "login_hint": "keep-me"},
+	}
+	pkce := PKCE{Challenge: "chal", Method: MethodS256}
+	raw, err := BuildAuthorizationURL(cfg, pkce, "real-state", "http://127.0.0.1:9/cb", map[string]string{"redirect_uri": "http://evil/cb"})
+	if err != nil {
+		t.Fatalf("BuildAuthorizationURL: %v", err)
+	}
+	parsed, _ := url.Parse(raw)
+	q := parsed.Query()
+	if q.Get("state") != "real-state" {
+		t.Fatalf("state overridden: %q", q.Get("state"))
+	}
+	if q.Get("code_challenge_method") != "S256" {
+		t.Fatalf("PKCE method downgraded: %q", q.Get("code_challenge_method"))
+	}
+	if q.Get("redirect_uri") != "http://127.0.0.1:9/cb" {
+		t.Fatalf("redirect_uri overridden: %q", q.Get("redirect_uri"))
+	}
+	if q.Get("login_hint") != "keep-me" {
+		t.Fatalf("non-reserved extra param should still apply: %q", q.Get("login_hint"))
+	}
+}
+
+func TestBuildAuthorizationURLRejectsPlain(t *testing.T) {
+	_, err := BuildAuthorizationURL(Config{AuthorizationEndpoint: "https://a/x"}, PKCE{Method: "plain"}, "s", "r", nil)
+	if !errors.Is(err, ErrPKCEDowngrade) {
+		t.Fatalf("err = %v, want ErrPKCEDowngrade", err)
+	}
+}
+
+func TestValidateTokenEndpoint(t *testing.T) {
+	cases := map[string]bool{
+		"https://auth.example.com/token": true,
+		"http://127.0.0.1:1455/token":    true, // loopback exempt
+		"http://localhost:1455/token":    true,
+		"http://[::1]:1455/token":        true,
+		"http://auth.example.com/token":  false, // non-https, non-loopback
+		"ftp://auth.example.com/token":   false,
+	}
+	for endpoint, ok := range cases {
+		err := validateTokenEndpoint(endpoint)
+		if ok && err != nil {
+			t.Errorf("validateTokenEndpoint(%q) = %v, want nil", endpoint, err)
+		}
+		if !ok && err == nil {
+			t.Errorf("validateTokenEndpoint(%q) = nil, want error", endpoint)
+		}
+	}
+}
+
+func TestExchangeCodeSuccess(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_ = r.ParseForm()
+		if r.FormValue("grant_type") != "authorization_code" || r.FormValue("code_verifier") != "verifier" {
+			w.WriteHeader(http.StatusBadRequest)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"access_token":"at","refresh_token":"rt","token_type":"Bearer","expires_in":3600,"scope":"read write"}`))
+	}))
+	defer server.Close()
+
+	now := time.Date(2020, 1, 1, 0, 0, 0, 0, time.UTC)
+	cfg := Config{ClientID: "c", TokenEndpoint: server.URL}
+	tok, err := ExchangeCode(context.Background(), server.Client(), cfg, "code", "verifier", "http://127.0.0.1/cb", func() time.Time { return now })
+	if err != nil {
+		t.Fatalf("ExchangeCode: %v", err)
+	}
+	if tok.AccessToken != "at" || tok.RefreshToken != "rt" || tok.TokenType != "Bearer" {
+		t.Fatalf("token = %+v", tok)
+	}
+	if !tok.ExpiresAt.Equal(now.Add(time.Hour)) {
+		t.Fatalf("expiresAt = %v, want now+1h", tok.ExpiresAt)
+	}
+	if strings.Join(tok.Scopes, " ") != "read write" {
+		t.Fatalf("scopes = %v", tok.Scopes)
+	}
+}
+
+func TestExchangeCodeErrorRedactsBody(t *testing.T) {
+	// The error body contains a stray access_token-looking field; it must NOT
+	// surface in the returned error (only error/error_description do).
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(http.StatusBadRequest)
+		_, _ = w.Write([]byte(`{"error":"invalid_grant","error_description":"bad code","access_token":"SECRET-LEAK"}`))
+	}))
+	defer server.Close()
+	cfg := Config{ClientID: "c", TokenEndpoint: server.URL}
+	_, err := ExchangeCode(context.Background(), server.Client(), cfg, "code", "v", "http://127.0.0.1/cb", nil)
+	if err == nil {
+		t.Fatal("expected error")
+	}
+	if !strings.Contains(err.Error(), "invalid_grant") || !strings.Contains(err.Error(), "bad code") {
+		t.Fatalf("error should carry error/description, got %q", err.Error())
+	}
+	if strings.Contains(err.Error(), "SECRET-LEAK") {
+		t.Fatalf("error leaked raw body token material: %q", err.Error())
+	}
+}
+
+func TestPostTokenRefusesInsecureEndpoint(t *testing.T) {
+	_, err := PostToken(context.Background(), http.DefaultClient, "http://auth.example.com/token", url.Values{}, Token{}, nil)
+	if !errors.Is(err, ErrInsecureTokenEndpoint) {
+		t.Fatalf("err = %v, want ErrInsecureTokenEndpoint", err)
+	}
+}
+
+func TestRefreshNoToken(t *testing.T) {
+	_, err := Refresh(context.Background(), http.DefaultClient, Config{TokenEndpoint: "https://a/token"}, Token{}, nil)
+	if !errors.Is(err, ErrNoRefreshToken) {
+		t.Fatalf("err = %v, want ErrNoRefreshToken", err)
+	}
+}
+
+func TestRefreshPreservesRefreshTokenWhenOmitted(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		_, _ = w.Write([]byte(`{"access_token":"new-at","expires_in":3600}`)) // no refresh_token in response
+	}))
+	defer server.Close()
+	cfg := Config{ClientID: "c", TokenEndpoint: server.URL}
+	tok, err := Refresh(context.Background(), server.Client(), cfg, Token{RefreshToken: "keep-me"}, nil)
+	if err != nil {
+		t.Fatalf("Refresh: %v", err)
+	}
+	if tok.AccessToken != "new-at" || tok.RefreshToken != "keep-me" {
+		t.Fatalf("refresh should preserve old refresh token: %+v", tok)
+	}
+}

internal/oauth/lock.go

@@ -0,0 +1,74 @@
+package oauth
+
+import (
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"sync/atomic"
+	"time"
+)
+
+const (
+	fileLockTimeout    = 5 * time.Second
+	fileLockStaleAfter = 30 * time.Second
+)
+
+var lockSeq atomic.Uint64
+
+// acquireFileLock takes a cross-process exclusive lock by creating lockPath with
+// O_EXCL. It retries with a short backoff until a timeout, breaking a lock whose
+// file is older than fileLockStaleAfter (so a crashed holder cannot deadlock the
+// store). Release is ownership-aware: it removes the lock only if it still holds
+// our token, so a stale-broken holder cannot delete a newer holder's lock.
+func acquireFileLock(lockPath string, now func() time.Time) (func(), error) {
+	if now == nil {
+		now = time.Now
+	}
+	if err := os.MkdirAll(filepath.Dir(lockPath), 0o700); err != nil {
+		return nil, err
+	}
+	token := fmt.Sprintf("%d-%d-%d", os.Getpid(), now().UnixNano(), lockSeq.Add(1))
+	deadline := now().Add(fileLockTimeout)
+	for {
+		f, err := os.OpenFile(lockPath, os.O_CREATE|os.O_EXCL|os.O_WRONLY, 0o600)
+		if err == nil {
+			// A partial/failed write would leave a lock file without our token, so
+			// the releaser below could never delete it — stranding the lock for
+			// other processes. Fail closed: remove the file and surface the error.
+			if _, werr := f.WriteString(token); werr != nil {
+				_ = f.Close()
+				_ = os.Remove(lockPath)
+				return nil, fmt.Errorf("oauth: write token lock: %w", werr)
+			}
+			if cerr := f.Close(); cerr != nil {
+				_ = os.Remove(lockPath)
+				return nil, fmt.Errorf("oauth: close token lock: %w", cerr)
+			}
+			var released bool
+			return func() {
+				if released {
+					return
+				}
+				released = true
+				if data, rerr := os.ReadFile(lockPath); rerr == nil && string(data) == token {
+					_ = os.Remove(lockPath)
+				}
+			}, nil
+		}
+		if !errors.Is(err, os.ErrExist) {
+			return nil, fmt.Errorf("oauth: acquire token lock: %w", err)
+		}
+		if info, statErr := os.Stat(lockPath); statErr == nil && time.Since(info.ModTime()) > fileLockStaleAfter {
+			stale, _ := os.ReadFile(lockPath)
+			if data, rerr := os.ReadFile(lockPath); rerr == nil && string(data) == string(stale) {
+				_ = os.Remove(lockPath)
+			}
+			continue
+		}
+		if now().After(deadline) {
+			return nil, fmt.Errorf("oauth: timed out acquiring token lock %s", filepath.Base(lockPath))
+		}
+		time.Sleep(10 * time.Millisecond)
+	}
+}

internal/oauth/loopback.go

@@ -0,0 +1,112 @@
+package oauth
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"net"
+	"net/http"
+	"net/url"
+	"strings"
+	"time"
+)
+
+// LoopbackListener is a single-use loopback HTTP server that captures an OAuth
+// redirect (?code=&state=) on 127.0.0.1 only: bind an OS-assigned loopback port,
+// verify the CSRF state, hand back the code, then close.
+type LoopbackListener struct {
+	listener net.Listener
+	state    string
+	result   chan callbackResult
+	server   *http.Server
+}
+
+type callbackResult struct {
+	code string
+	err  error
+}
+
+// NewLoopbackListener binds 127.0.0.1:0 (loopback only, OS-assigned port) and
+// begins serving. state is the CSRF value the callback must echo back. Call
+// RedirectURI to build the redirect_uri, then Wait for the code. Always Close.
+func NewLoopbackListener(state string) (*LoopbackListener, error) {
+	// Refuse an empty state: the callback check compares the query's state to this
+	// value, so an empty state would match a callback carrying no state at all,
+	// defeating CSRF protection (fail closed).
+	if strings.TrimSpace(state) == "" {
+		return nil, errors.New("oauth: loopback listener requires a non-empty CSRF state")
+	}
+	ln, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		return nil, fmt.Errorf("oauth: start loopback redirect listener: %w", err)
+	}
+	l := &LoopbackListener{
+		listener: ln,
+		state:    state,
+		result:   make(chan callbackResult, 1),
+	}
+	l.server = &http.Server{Handler: http.HandlerFunc(l.handle)}
+	go func() { _ = l.server.Serve(ln) }()
+	return l, nil
+}
+
+// RedirectURI returns the http://127.0.0.1:<port>/callback redirect URI.
+func (l *LoopbackListener) RedirectURI() string {
+	return fmt.Sprintf("http://%s/callback", l.listener.Addr().String())
+}
+
+func (l *LoopbackListener) handle(w http.ResponseWriter, r *http.Request) {
+	if r.URL.Path != "/callback" {
+		http.NotFound(w, r)
+		return
+	}
+	code, err := parseCallback(r.URL.Query(), l.state)
+	if err != nil {
+		w.WriteHeader(http.StatusBadRequest)
+		_, _ = io.WriteString(w, "Authorization failed. You may close this window.")
+	} else {
+		_, _ = io.WriteString(w, "Authorization complete. You may close this window.")
+	}
+	select {
+	case l.result <- callbackResult{code: code, err: err}:
+	default:
+	}
+}
+
+// Wait blocks until the callback arrives or ctx is done, returning the
+// authorization code.
+func (l *LoopbackListener) Wait(ctx context.Context) (string, error) {
+	select {
+	case res := <-l.result:
+		return res.code, res.err
+	case <-ctx.Done():
+		return "", fmt.Errorf("oauth: timed out waiting for authorization callback: %w", ctx.Err())
+	}
+}
+
+// Close shuts the listener down (bounded), idempotent.
+func (l *LoopbackListener) Close() {
+	shutdownCtx, cancel := context.WithTimeout(context.Background(), time.Second)
+	defer cancel()
+	_ = l.server.Shutdown(shutdownCtx)
+}
+
+// parseCallback validates the redirect query and returns the authorization code,
+// rejecting a mismatched state (CSRF) and surfacing provider errors.
+func parseCallback(values url.Values, wantState string) (string, error) {
+	if got := values.Get("state"); got != wantState {
+		return "", ErrStateMismatch
+	}
+	if providerErr := strings.TrimSpace(values.Get("error")); providerErr != "" {
+		if desc := strings.TrimSpace(values.Get("error_description")); desc != "" {
+			return "", fmt.Errorf("oauth: authorization server returned error %q: %s", providerErr, desc)
+		}
+		return "", fmt.Errorf("oauth: authorization server returned error %q", providerErr)
+	}
+	code := strings.TrimSpace(values.Get("code"))
+	if code == "" {
+		return "", errors.New("oauth: callback missing authorization code")
+	}
+	return code, nil
+}

internal/oauth/loopback_test.go

@@ -0,0 +1,86 @@
+package oauth
+
+import (
+	"context"
+	"errors"
+	"net/http"
+	"strings"
+	"testing"
+	"time"
+)
+
+func TestLoopbackCapturesCode(t *testing.T) {
+	l, err := NewLoopbackListener("the-state")
+	if err != nil {
+		t.Fatalf("NewLoopbackListener: %v", err)
+	}
+	defer l.Close()
+	if !strings.HasPrefix(l.RedirectURI(), "http://127.0.0.1:") {
+		t.Fatalf("redirect URI not loopback: %q", l.RedirectURI())
+	}
+	go func() {
+		_, _ = http.Get(l.RedirectURI() + "?code=auth-code&state=the-state")
+	}()
+	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Second)
+	defer cancel()
+	code, err := l.Wait(ctx)
+	if err != nil {
+		t.Fatalf("Wait: %v", err)
+	}
+	if code != "auth-code" {
+		t.Fatalf("code = %q, want auth-code", code)
+	}
+}
+
+func TestNewLoopbackListenerRejectsEmptyState(t *testing.T) {
+	if _, err := NewLoopbackListener("   "); err == nil {
+		t.Fatal("an empty CSRF state must be rejected (fail closed)")
+	}
+}
+
+func TestLoopbackRejectsStateMismatch(t *testing.T) {
+	l, err := NewLoopbackListener("expected-state")
+	if err != nil {
+		t.Fatalf("NewLoopbackListener: %v", err)
+	}
+	defer l.Close()
+	go func() {
+		_, _ = http.Get(l.RedirectURI() + "?code=x&state=WRONG")
+	}()
+	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Second)
+	defer cancel()
+	_, err = l.Wait(ctx)
+	if !errors.Is(err, ErrStateMismatch) {
+		t.Fatalf("Wait err = %v, want ErrStateMismatch", err)
+	}
+}
+
+func TestLoopbackTimesOut(t *testing.T) {
+	l, err := NewLoopbackListener("s")
+	if err != nil {
+		t.Fatalf("NewLoopbackListener: %v", err)
+	}
+	defer l.Close()
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Millisecond)
+	defer cancel()
+	if _, err := l.Wait(ctx); err == nil {
+		t.Fatal("Wait should time out cleanly")
+	}
+}
+
+func TestLoopbackProviderError(t *testing.T) {
+	l, err := NewLoopbackListener("s")
+	if err != nil {
+		t.Fatalf("NewLoopbackListener: %v", err)
+	}
+	defer l.Close()
+	go func() {
+		_, _ = http.Get(l.RedirectURI() + "?error=access_denied&error_description=nope&state=s")
+	}()
+	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Second)
+	defer cancel()
+	_, err = l.Wait(ctx)
+	if err == nil || !strings.Contains(err.Error(), "access_denied") {
+		t.Fatalf("Wait err = %v, want provider error", err)
+	}
+}

internal/oauth/manager.go

@@ -0,0 +1,292 @@
+package oauth
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"time"
+)
+
+// defaultRefreshBuffer refreshes a token this long before its hard expiry.
+const defaultRefreshBuffer = 60 * time.Second
+
+const oidcWellKnownPath = "/.well-known/openid-configuration"
+
+// Manager ties the token store, provider registry, and HTTP client together to
+// run logins and serve fresh access tokens. It is the high-level entrypoint the
+// CLI and request paths use.
+type Manager struct {
+	store    *Store
+	registry *Registry
+	client   *http.Client
+	env      map[string]string
+	now      func() time.Time
+	buffer   time.Duration
+	out      io.Writer
+	// openBrowser is invoked with the authorization URL for loopback logins.
+	// Tests inject a function that drives the loopback redirect.
+	openBrowser func(authURL string) error
+}
+
+// ManagerOptions configures a Manager.
+type ManagerOptions struct {
+	Store         *Store
+	Registry      *Registry
+	HTTPClient    *http.Client
+	Env           map[string]string
+	Now           func() time.Time
+	RefreshBuffer time.Duration
+	Out           io.Writer
+	OpenBrowser   func(authURL string) error
+}
+
+// NewManager builds a Manager, filling defaults.
+func NewManager(opts ManagerOptions) (*Manager, error) {
+	if opts.Store == nil {
+		return nil, fmt.Errorf("oauth: manager requires a store")
+	}
+	registry := opts.Registry
+	if registry == nil {
+		registry = NewRegistry()
+	}
+	client := opts.HTTPClient
+	if client == nil {
+		client = http.DefaultClient
+	}
+	now := opts.Now
+	if now == nil {
+		now = time.Now
+	}
+	buffer := opts.RefreshBuffer
+	if buffer <= 0 {
+		buffer = defaultRefreshBuffer
+	}
+	out := opts.Out
+	if out == nil {
+		out = io.Discard
+	}
+	open := opts.OpenBrowser
+	if open == nil {
+		open = func(string) error { return nil }
+	}
+	return &Manager{
+		store: opts.Store, registry: registry, client: client,
+		env: opts.Env, now: now, buffer: buffer, out: out, openBrowser: open,
+	}, nil
+}
+
+// LoginOptions configures a single provider login.
+type LoginOptions struct {
+	Provider    string
+	Device      bool          // force device-code flow
+	ExtraScopes []string      // appended to the provider's scopes
+	Timeout     time.Duration // bounds the whole interactive login
+}
+
+// Login runs the provider login (loopback by default, device-code when
+// requested or when the provider only supports device), stores the token under
+// "provider:<name>", and returns a redaction-safe status.
+func (m *Manager) Login(ctx context.Context, opts LoginOptions) (Status, error) {
+	cfg, flow, err := m.registry.ResolveConfig(opts.Provider, m.env)
+	if err != nil {
+		return Status{}, err
+	}
+	if len(opts.ExtraScopes) > 0 {
+		cfg.Scopes = append(append([]string{}, cfg.Scopes...), opts.ExtraScopes...)
+	}
+	timeout := opts.Timeout
+	if timeout <= 0 {
+		timeout = 5 * time.Minute
+	}
+	loginCtx, cancel := context.WithTimeout(ctx, timeout)
+	defer cancel()
+
+	cfg, err = m.resolveEndpoints(loginCtx, cfg)
+	if err != nil {
+		return Status{}, err
+	}
+
+	useDevice := opts.Device || flow == FlowDevice
+	var token Token
+	if useDevice {
+		token, err = m.loginDevice(loginCtx, cfg)
+	} else {
+		token, err = m.loginLoopback(loginCtx, cfg)
+	}
+	if err != nil {
+		return Status{}, err
+	}
+
+	key := ProviderKey(opts.Provider)
+	if err := m.store.Save(key, token); err != nil {
+		return Status{}, err
+	}
+	return m.statusFor(key)
+}
+
+// resolveEndpoints fills missing authorize/token/device endpoints from issuer
+// discovery (RFC 8414 then OIDC), leaving any explicitly-pinned endpoint intact.
+func (m *Manager) resolveEndpoints(ctx context.Context, cfg Config) (Config, error) {
+	if trimmed(cfg.IssuerURL) == "" {
+		return cfg, nil
+	}
+	if cfg.AuthorizationEndpoint != "" && cfg.TokenEndpoint != "" && cfg.DeviceAuthorizationEndpoint != "" {
+		return cfg, nil
+	}
+	// Discovery is best-effort: a failure is non-fatal because pinned endpoints
+	// may already be sufficient for the chosen flow. Only merge on success (and
+	// never return from the error branch, which would trip the nilerr linter).
+	if meta, err := m.discover(ctx, cfg.IssuerURL); err == nil {
+		if cfg.AuthorizationEndpoint == "" {
+			cfg.AuthorizationEndpoint = meta.AuthorizationEndpoint
+		}
+		if cfg.TokenEndpoint == "" {
+			cfg.TokenEndpoint = meta.TokenEndpoint
+		}
+		if cfg.DeviceAuthorizationEndpoint == "" {
+			cfg.DeviceAuthorizationEndpoint = meta.DeviceAuthorizationEndpoint
+		}
+	}
+	return cfg, nil
+}
+
+// discover tries the OAuth (RFC 8414) well-known path, then the OIDC
+// openid-configuration path (some issuers publish only the latter).
+func (m *Manager) discover(ctx context.Context, issuer string) (ServerMetadata, error) {
+	meta, err := DiscoverAuthorizationServer(ctx, m.client, issuer)
+	if err == nil && (meta.AuthorizationEndpoint != "" || meta.TokenEndpoint != "") {
+		return meta, nil
+	}
+	oidcURL := strings.TrimRight(strings.TrimSpace(issuer), "/") + oidcWellKnownPath
+	return fetchMetadata(ctx, m.client, oidcURL)
+}
+
+func (m *Manager) loginLoopback(ctx context.Context, cfg Config) (Token, error) {
+	if trimmed(cfg.AuthorizationEndpoint) == "" {
+		return Token{}, fmt.Errorf("oauth: no authorization endpoint (set the authorize URL or a discoverable issuer)")
+	}
+	state, err := NewState()
+	if err != nil {
+		return Token{}, err
+	}
+	pkce, err := NewPKCE()
+	if err != nil {
+		return Token{}, err
+	}
+	listener, err := NewLoopbackListener(state)
+	if err != nil {
+		return Token{}, err
+	}
+	defer listener.Close()
+	redirectURI := listener.RedirectURI()
+	authURL, err := BuildAuthorizationURL(cfg, pkce, state, redirectURI, nil)
+	if err != nil {
+		return Token{}, err
+	}
+	fmt.Fprintf(m.out, "Open this URL to authorize:\n  %s\n", authURL)
+	if err := m.openBrowser(authURL); err != nil {
+		return Token{}, fmt.Errorf("oauth: open authorization URL: %w", err)
+	}
+	code, err := listener.Wait(ctx)
+	if err != nil {
+		return Token{}, err
+	}
+	return ExchangeCode(ctx, m.client, cfg, code, pkce.Verifier, redirectURI, m.now)
+}
+
+func (m *Manager) loginDevice(ctx context.Context, cfg Config) (Token, error) {
+	auth, err := RequestDeviceCode(ctx, m.client, cfg, m.now)
+	if err != nil {
+		return Token{}, err
+	}
+	target := auth.VerificationURIComplete
+	if target == "" {
+		target = auth.VerificationURI
+	}
+	// user_code is meant to be displayed to the user; it is not a secret.
+	fmt.Fprintf(m.out, "To authorize, visit:\n  %s\nand enter code: %s\n", target, auth.UserCode)
+	return PollDeviceToken(ctx, m.client, cfg, auth, m.now)
+}
+
+// GetFresh returns a valid access token for key, refreshing on-demand if the
+// stored token is expired or within the refresh buffer. Mirrors
+// checkAndRefreshOAuthTokenIfNeeded.
+func (m *Manager) GetFresh(ctx context.Context, key string) (string, error) {
+	token, cfg, err := m.loadForKey(key)
+	if err != nil {
+		return "", err
+	}
+	if !token.NeedsRefresh(m.now(), m.buffer) {
+		return token.AccessToken, nil
+	}
+	return m.refreshAndSave(ctx, key, cfg, token)
+}
+
+// Handle401 forces a refresh after an upstream 401, returning the new access
+// token. Mirrors handleOAuth401Error.
+func (m *Manager) Handle401(ctx context.Context, key string) (string, error) {
+	token, cfg, err := m.loadForKey(key)
+	if err != nil {
+		return "", err
+	}
+	return m.refreshAndSave(ctx, key, cfg, token)
+}
+
+func (m *Manager) refreshAndSave(ctx context.Context, key string, cfg Config, current Token) (string, error) {
+	refreshed, err := Refresh(ctx, m.client, cfg, current, m.now)
+	if err != nil {
+		return "", err
+	}
+	if err := m.store.Save(key, refreshed); err != nil {
+		return "", err
+	}
+	return refreshed.AccessToken, nil
+}
+
+// loadForKey loads a stored token and resolves the provider config for its key.
+func (m *Manager) loadForKey(key string) (Token, Config, error) {
+	if err := ValidateKey(key); err != nil {
+		return Token{}, Config{}, err
+	}
+	token, ok, err := m.store.Load(key)
+	if err != nil {
+		return Token{}, Config{}, err
+	}
+	if !ok {
+		return Token{}, Config{}, fmt.Errorf("oauth: no stored token for %q", key)
+	}
+	name := strings.TrimPrefix(key, KeyPrefixProvider)
+	if name == key {
+		return Token{}, Config{}, fmt.Errorf("oauth: refresh is only supported for provider tokens (got %q)", key)
+	}
+	cfg, _, err := m.registry.ResolveConfig(name, m.env)
+	if err != nil {
+		return Token{}, Config{}, err
+	}
+	return token, cfg, nil
+}
+
+// Logout removes a provider's stored token, reporting whether one was present.
+func (m *Manager) Logout(name string) (bool, error) {
+	return m.store.Delete(ProviderKey(name))
+}
+
+// StatusAll returns the status of every provider login.
+func (m *Manager) StatusAll() ([]Status, error) {
+	return m.store.Status(KeyPrefixProvider)
+}
+
+func (m *Manager) statusFor(key string) (Status, error) {
+	statuses, err := m.store.Status(KeyPrefixProvider)
+	if err != nil {
+		return Status{}, err
+	}
+	for _, st := range statuses {
+		if st.Key == key {
+			return st, nil
+		}
+	}
+	return Status{Key: key}, nil
+}

internal/oauth/manager_test.go

@@ -0,0 +1,182 @@
+package oauth
+
+import (
+	"context"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"path/filepath"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+// fakeProvider is an httptest server that plays a token + device endpoint.
+type fakeProvider struct {
+	server    *httptest.Server
+	tokenHits atomic.Int32
+}
+
+func newFakeProvider(t *testing.T, tokenJSON string) *fakeProvider {
+	t.Helper()
+	fp := &fakeProvider{}
+	mux := http.NewServeMux()
+	mux.HandleFunc("/token", func(w http.ResponseWriter, _ *http.Request) {
+		fp.tokenHits.Add(1)
+		_, _ = io.WriteString(w, tokenJSON)
+	})
+	mux.HandleFunc("/device", func(w http.ResponseWriter, _ *http.Request) {
+		// Approve immediately on poll; short interval so the test is fast.
+		_, _ = io.WriteString(w, `{"device_code":"dc","user_code":"U-1","verification_uri":"https://example/dev","expires_in":600,"interval":1}`)
+	})
+	fp.server = httptest.NewServer(mux)
+	t.Cleanup(fp.server.Close)
+	return fp
+}
+
+func managerFor(t *testing.T, env map[string]string, openBrowser func(string) error) *Manager {
+	t.Helper()
+	store, err := NewStore(StoreOptions{FilePath: filepath.Join(t.TempDir(), "tok.json")})
+	if err != nil {
+		t.Fatalf("NewStore: %v", err)
+	}
+	m, err := NewManager(ManagerOptions{
+		Store:       store,
+		Env:         env,
+		OpenBrowser: openBrowser,
+	})
+	if err != nil {
+		t.Fatalf("NewManager: %v", err)
+	}
+	return m
+}
+
+func TestManagerLoginLoopback(t *testing.T) {
+	fp := newFakeProvider(t, `{"access_token":"at","refresh_token":"rt","token_type":"Bearer","expires_in":3600}`)
+	env := map[string]string{
+		"ZERO_OAUTH_DEMO_CLIENT_ID":     "client",
+		"ZERO_OAUTH_DEMO_AUTHORIZE_URL": "https://auth.example.com/authorize",
+		"ZERO_OAUTH_DEMO_TOKEN_URL":     fp.server.URL + "/token",
+	}
+	// The fake browser drives the loopback redirect with the captured state.
+	openBrowser := func(authURL string) error {
+		u, err := url.Parse(authURL)
+		if err != nil {
+			return err
+		}
+		redirect := u.Query().Get("redirect_uri")
+		state := u.Query().Get("state")
+		_, err = http.Get(redirect + "?code=the-code&state=" + url.QueryEscape(state))
+		return err
+	}
+	m := managerFor(t, env, openBrowser)
+
+	status, err := m.Login(context.Background(), LoginOptions{Provider: "demo"})
+	if err != nil {
+		t.Fatalf("Login: %v", err)
+	}
+	if !status.HasToken || status.Key != ProviderKey("demo") {
+		t.Fatalf("status = %+v", status)
+	}
+	// The token is persisted and GetFresh returns it without a needless refresh.
+	access, err := m.GetFresh(context.Background(), ProviderKey("demo"))
+	if err != nil {
+		t.Fatalf("GetFresh: %v", err)
+	}
+	if access != "at" {
+		t.Fatalf("access = %q, want at", access)
+	}
+}
+
+func TestManagerLoginDevice(t *testing.T) {
+	fp := newFakeProvider(t, `{"access_token":"dev-at","token_type":"Bearer","expires_in":3600}`)
+	env := map[string]string{
+		"ZERO_OAUTH_DEMODEV_CLIENT_ID":  "client",
+		"ZERO_OAUTH_DEMODEV_TOKEN_URL":  fp.server.URL + "/token",
+		"ZERO_OAUTH_DEMODEV_DEVICE_URL": fp.server.URL + "/device",
+	}
+	m := managerFor(t, env, nil)
+	status, err := m.Login(context.Background(), LoginOptions{Provider: "demodev", Device: true})
+	if err != nil {
+		t.Fatalf("device Login: %v", err)
+	}
+	if !status.HasToken {
+		t.Fatalf("device status = %+v", status)
+	}
+}
+
+func TestManagerGetFreshRefreshesExpired(t *testing.T) {
+	fp := newFakeProvider(t, `{"access_token":"fresh-at","expires_in":3600}`)
+	env := map[string]string{
+		"ZERO_OAUTH_DEMO_CLIENT_ID": "client",
+		"ZERO_OAUTH_DEMO_TOKEN_URL": fp.server.URL + "/token",
+	}
+	m := managerFor(t, env, nil)
+	// Seed an expired token with a refresh token.
+	if err := m.store.Save(ProviderKey("demo"), Token{AccessToken: "stale", RefreshToken: "rt", ExpiresAt: time.Now().Add(-time.Hour)}); err != nil {
+		t.Fatalf("seed: %v", err)
+	}
+	access, err := m.GetFresh(context.Background(), ProviderKey("demo"))
+	if err != nil {
+		t.Fatalf("GetFresh: %v", err)
+	}
+	if access != "fresh-at" {
+		t.Fatalf("access = %q, want fresh-at (refreshed)", access)
+	}
+	if fp.tokenHits.Load() != 1 {
+		t.Fatalf("token endpoint hit %d times, want 1 refresh", fp.tokenHits.Load())
+	}
+	// The refreshed token is persisted.
+	stored, _, _ := m.store.Load(ProviderKey("demo"))
+	if stored.AccessToken != "fresh-at" {
+		t.Fatalf("stored token not updated: %+v", stored)
+	}
+}
+
+func TestManagerGetFreshSkipsValidToken(t *testing.T) {
+	fp := newFakeProvider(t, `{"access_token":"should-not-be-used"}`)
+	env := map[string]string{
+		"ZERO_OAUTH_DEMO_CLIENT_ID": "client",
+		"ZERO_OAUTH_DEMO_TOKEN_URL": fp.server.URL + "/token",
+	}
+	m := managerFor(t, env, nil)
+	_ = m.store.Save(ProviderKey("demo"), Token{AccessToken: "valid", RefreshToken: "rt", ExpiresAt: time.Now().Add(time.Hour)})
+	access, err := m.GetFresh(context.Background(), ProviderKey("demo"))
+	if err != nil {
+		t.Fatalf("GetFresh: %v", err)
+	}
+	if access != "valid" || fp.tokenHits.Load() != 0 {
+		t.Fatalf("a valid token must not be refreshed (access=%q hits=%d)", access, fp.tokenHits.Load())
+	}
+}
+
+func TestManagerHandle401ForcesRefresh(t *testing.T) {
+	fp := newFakeProvider(t, `{"access_token":"after-401"}`)
+	env := map[string]string{
+		"ZERO_OAUTH_DEMO_CLIENT_ID": "client",
+		"ZERO_OAUTH_DEMO_TOKEN_URL": fp.server.URL + "/token",
+	}
+	m := managerFor(t, env, nil)
+	// Token is still valid by clock, but Handle401 forces a refresh anyway.
+	_ = m.store.Save(ProviderKey("demo"), Token{AccessToken: "valid", RefreshToken: "rt", ExpiresAt: time.Now().Add(time.Hour)})
+	access, err := m.Handle401(context.Background(), ProviderKey("demo"))
+	if err != nil {
+		t.Fatalf("Handle401: %v", err)
+	}
+	if access != "after-401" || fp.tokenHits.Load() != 1 {
+		t.Fatalf("Handle401 must force a refresh (access=%q hits=%d)", access, fp.tokenHits.Load())
+	}
+}
+
+func TestManagerLogout(t *testing.T) {
+	m := managerFor(t, map[string]string{"ZERO_OAUTH_DEMO_CLIENT_ID": "c"}, nil)
+	_ = m.store.Save(ProviderKey("demo"), Token{AccessToken: "a"})
+	removed, err := m.Logout("demo")
+	if err != nil || !removed {
+		t.Fatalf("Logout = %v %v", removed, err)
+	}
+	if removed2, _ := m.Logout("demo"); removed2 {
+		t.Fatal("second logout should report nothing removed")
+	}
+}

internal/oauth/oauth.go

@@ -0,0 +1,95 @@
+// Package oauth is zero's reusable OAuth 2.0 engine. It generalizes the
+// authorization-code + PKCE flow that internal/mcp already uses for MCP servers
+// into a transport/identity-agnostic core, and adds the pieces provider login
+// needs: a device-authorization (RFC 8628) grant, a provider registry, a
+// namespaced + refreshing token store, and a proactive refresh scheduler.
+//
+// Security invariants (enforced throughout, never relaxed):
+//   - PKCE S256 is mandatory on every authorization-code flow; "plain" is refused.
+//   - A per-flow CSRF state is generated and verified on the callback.
+//   - The loopback callback server binds 127.0.0.1 only, on an OS-assigned port,
+//     serves a single request, then closes.
+//   - The token endpoint must be https (loopback exempt) before any credential is
+//     sent — see validateTokenEndpoint.
+//   - Tokens, codes, and verifiers are never logged; error bodies are redacted.
+//   - Token files are 0600, base-confined, and file-locked.
+//
+// This package holds NO vendor secrets: provider client IDs and any overridable
+// endpoints come from config/env, so no third-party OAuth client identity is
+// baked into the binary.
+package oauth
+
+import (
+	"errors"
+	"strings"
+	"time"
+)
+
+// Token holds the credentials issued by an authorization server. The token
+// fields are sensitive: callers must never log them, and the store persists them
+// 0600. It mirrors the on-disk shape internal/mcp uses so the two stay
+// format-compatible.
+type Token struct {
+	AccessToken  string    `json:"access_token"`
+	RefreshToken string    `json:"refresh_token,omitempty"`
+	TokenType    string    `json:"token_type,omitempty"`
+	Scopes       []string  `json:"scopes,omitempty"`
+	ExpiresAt    time.Time `json:"expires_at,omitempty"`
+	// Account is an optional non-secret identifier (email / account id) shown in
+	// status output; never a credential.
+	Account string `json:"account,omitempty"`
+}
+
+// Expired reports whether the token has an expiry that is at or before now.
+// A zero ExpiresAt means "no known expiry" and is treated as not expired.
+func (t Token) Expired(now time.Time) bool {
+	return !t.ExpiresAt.IsZero() && !t.ExpiresAt.After(now)
+}
+
+// NeedsRefresh reports whether the token is expired or falls within buffer of
+// expiry (so a proactive/on-demand refresh should run). A token with no expiry
+// never needs a refresh on a timer.
+func (t Token) NeedsRefresh(now time.Time, buffer time.Duration) bool {
+	if t.ExpiresAt.IsZero() {
+		return false
+	}
+	return !t.ExpiresAt.After(now.Add(buffer))
+}
+
+// Config describes how to talk to one authorization server for a flow.
+// Endpoints may be discovered (RFC 8414) and overridden here.
+type Config struct {
+	ClientID     string
+	ClientSecret string
+	Scopes       []string
+
+	AuthorizationEndpoint       string
+	TokenEndpoint               string
+	DeviceAuthorizationEndpoint string
+	RegistrationEndpoint        string
+	// IssuerURL is the base for metadata discovery when endpoints are not set.
+	IssuerURL string
+
+	// ExtraAuthParams are appended to the authorization URL (e.g. login_hint).
+	ExtraAuthParams map[string]string
+}
+
+// Errors returned by the engine. Callers can match these with errors.Is.
+var (
+	// ErrPKCEDowngrade is returned if a flow is asked to use anything but S256.
+	ErrPKCEDowngrade = errors.New("oauth: PKCE S256 is mandatory; plain is refused")
+	// ErrStateMismatch is returned when the callback state does not match (CSRF).
+	ErrStateMismatch = errors.New("oauth: callback state mismatch; possible CSRF, login aborted")
+	// ErrInsecureTokenEndpoint is returned when a credential would be sent over a
+	// non-https, non-loopback endpoint.
+	ErrInsecureTokenEndpoint = errors.New("oauth: refusing to send credential to a non-https token endpoint")
+	// ErrNoRefreshToken is returned when a refresh is attempted without one.
+	ErrNoRefreshToken = errors.New("oauth: no refresh token available")
+	// ErrAuthorizationPending is the RFC 8628 "keep polling" signal.
+	ErrAuthorizationPending = errors.New("oauth: authorization pending")
+	// ErrSlowDown is the RFC 8628 "increase the poll interval" signal.
+	ErrSlowDown = errors.New("oauth: slow down")
+)
+
+// trimmed is a tiny helper used across the package.
+func trimmed(s string) string { return strings.TrimSpace(s) }

internal/oauth/pkce.go

@@ -0,0 +1,49 @@
+package oauth
+
+import (
+	"crypto/rand"
+	"crypto/sha256"
+	"encoding/base64"
+	"fmt"
+)
+
+const (
+	// pkceVerifierBytes yields a 43-character base64url verifier — the minimum
+	// length the PKCE spec permits while remaining high-entropy.
+	pkceVerifierBytes = 32
+	stateBytes        = 32
+	// MethodS256 is the only code-challenge method this engine accepts.
+	MethodS256 = "S256"
+)
+
+// PKCE is a verifier/challenge pair for an authorization-code flow. The method
+// is always S256 (the engine refuses "plain").
+type PKCE struct {
+	Verifier  string
+	Challenge string
+	Method    string
+}
+
+// NewPKCE generates a high-entropy code verifier and its S256 challenge. It
+// matches internal/mcp's newPKCE byte-for-byte (RawURLEncoding of 32 random
+// bytes; challenge = base64url(SHA-256(verifier))) so the two engines are
+// interchangeable.
+func NewPKCE() (PKCE, error) {
+	raw := make([]byte, pkceVerifierBytes)
+	if _, err := rand.Read(raw); err != nil {
+		return PKCE{}, fmt.Errorf("oauth: generate PKCE verifier: %w", err)
+	}
+	verifier := base64.RawURLEncoding.EncodeToString(raw)
+	sum := sha256.Sum256([]byte(verifier))
+	challenge := base64.RawURLEncoding.EncodeToString(sum[:])
+	return PKCE{Verifier: verifier, Challenge: challenge, Method: MethodS256}, nil
+}
+
+// NewState generates a high-entropy CSRF state value.
+func NewState() (string, error) {
+	raw := make([]byte, stateBytes)
+	if _, err := rand.Read(raw); err != nil {
+		return "", fmt.Errorf("oauth: generate state: %w", err)
+	}
+	return base64.RawURLEncoding.EncodeToString(raw), nil
+}

internal/oauth/providers.go

@@ -0,0 +1,100 @@
+package oauth
+
+import (
+	"fmt"
+	"regexp"
+	"strings"
+)
+
+// Flow selects how a provider delivers the authorization result.
+type Flow string
+
+const (
+	// FlowLoopback uses a 127.0.0.1 callback server (browser required).
+	FlowLoopback Flow = "loopback"
+	// FlowDevice uses the RFC 8628 device-code flow (headless/SSH).
+	FlowDevice Flow = "device"
+)
+
+// providerNamePattern bounds a provider name to a safe identifier that is also a
+// valid store-key segment.
+var providerNamePattern = regexp.MustCompile(`^[A-Za-z0-9][A-Za-z0-9._-]{0,63}$`)
+
+// ValidateProviderName reports whether name is a safe provider identifier.
+func ValidateProviderName(name string) error {
+	if !providerNamePattern.MatchString(name) {
+		return fmt.Errorf("oauth: invalid provider name %q", name)
+	}
+	return nil
+}
+
+// Registry resolves a provider's Config from env/config. It ships NO built-in
+// providers, endpoints, or client identities: every provider is defined entirely
+// by the operator via ZERO_OAUTH_<NAME>_* variables, so no third-party brand or
+// OAuth client identity is baked into the binary.
+type Registry struct{}
+
+// NewRegistry returns the (stateless) env-driven registry.
+func NewRegistry() *Registry { return &Registry{} }
+
+// envKey builds the ZERO_OAUTH_<NAME>_<suffix> variable name for a provider.
+func envKey(name, suffix string) string {
+	up := strings.ToUpper(name)
+	up = strings.NewReplacer("-", "_", ".", "_").Replace(up)
+	return "ZERO_OAUTH_" + up + "_" + suffix
+}
+
+// ResolveConfig builds the oauth.Config and default Flow for a provider from its
+// env/config:
+//
+//	ZERO_OAUTH_<NAME>_CLIENT_ID       (required)
+//	ZERO_OAUTH_<NAME>_CLIENT_SECRET   (optional)
+//	ZERO_OAUTH_<NAME>_AUTHORIZE_URL   (loopback flow; or discovered via issuer)
+//	ZERO_OAUTH_<NAME>_TOKEN_URL       (or discovered via issuer)
+//	ZERO_OAUTH_<NAME>_DEVICE_URL      (device flow; or discovered via issuer)
+//	ZERO_OAUTH_<NAME>_ISSUER_URL      (RFC 8414 / OIDC discovery base)
+//	ZERO_OAUTH_<NAME>_SCOPES          (space-separated)
+//	ZERO_OAUTH_<NAME>_FLOW            ("loopback" [default] | "device")
+//
+// Pinned credential-bearing endpoints must be https (loopback exempt).
+func (r *Registry) ResolveConfig(name string, env map[string]string) (Config, Flow, error) {
+	if err := ValidateProviderName(name); err != nil {
+		return Config{}, "", err
+	}
+	cfg := Config{
+		ClientID:                    strings.TrimSpace(envValue(env, envKey(name, "CLIENT_ID"))),
+		ClientSecret:                strings.TrimSpace(envValue(env, envKey(name, "CLIENT_SECRET"))),
+		AuthorizationEndpoint:       strings.TrimSpace(envValue(env, envKey(name, "AUTHORIZE_URL"))),
+		TokenEndpoint:               strings.TrimSpace(envValue(env, envKey(name, "TOKEN_URL"))),
+		DeviceAuthorizationEndpoint: strings.TrimSpace(envValue(env, envKey(name, "DEVICE_URL"))),
+		IssuerURL:                   strings.TrimSpace(envValue(env, envKey(name, "ISSUER_URL"))),
+		Scopes:                      strings.Fields(envValue(env, envKey(name, "SCOPES"))),
+	}
+	if cfg.ClientID == "" {
+		return Config{}, "", fmt.Errorf("oauth: provider %q is not configured; set %s (and its endpoints or an issuer)", name, envKey(name, "CLIENT_ID"))
+	}
+	var flow Flow
+	switch strings.ToLower(strings.TrimSpace(envValue(env, envKey(name, "FLOW")))) {
+	case "", string(FlowLoopback):
+		flow = FlowLoopback
+	case string(FlowDevice):
+		flow = FlowDevice
+	default:
+		return Config{}, "", fmt.Errorf("oauth: provider %q has invalid %s (want loopback or device)", name, envKey(name, "FLOW"))
+	}
+	// A token endpoint (for exchange/refresh) must be reachable directly or via
+	// discovery. The per-flow authorize/device endpoints are checked at login
+	// time (after discovery), so a refresh-only config needs only a token URL.
+	if cfg.IssuerURL == "" && cfg.TokenEndpoint == "" {
+		return Config{}, "", fmt.Errorf("oauth: provider %q needs %s or %s", name, envKey(name, "TOKEN_URL"), envKey(name, "ISSUER_URL"))
+	}
+	for _, ep := range []string{cfg.TokenEndpoint, cfg.AuthorizationEndpoint, cfg.DeviceAuthorizationEndpoint, cfg.IssuerURL} {
+		if ep == "" {
+			continue
+		}
+		if err := validateTokenEndpoint(ep); err != nil {
+			return Config{}, "", err
+		}
+	}
+	return cfg, flow, nil
+}

internal/oauth/providers_test.go

@@ -0,0 +1,97 @@
+package oauth
+
+import (
+	"errors"
+	"testing"
+)
+
+func TestResolveConfigFromEnv(t *testing.T) {
+	r := NewRegistry()
+	env := map[string]string{
+		"ZERO_OAUTH_DEMO_CLIENT_ID":     "my-client",
+		"ZERO_OAUTH_DEMO_CLIENT_SECRET": "shh",
+		"ZERO_OAUTH_DEMO_SCOPES":        "read write",
+		"ZERO_OAUTH_DEMO_AUTHORIZE_URL": "https://auth.example.com/authorize",
+		"ZERO_OAUTH_DEMO_TOKEN_URL":     "https://auth.example.com/token",
+	}
+	cfg, flow, err := r.ResolveConfig("demo", env)
+	if err != nil {
+		t.Fatalf("ResolveConfig: %v", err)
+	}
+	if cfg.ClientID != "my-client" || cfg.ClientSecret != "shh" {
+		t.Fatalf("client creds not applied: %+v", cfg)
+	}
+	if len(cfg.Scopes) != 2 || cfg.Scopes[0] != "read" {
+		t.Fatalf("scopes = %v", cfg.Scopes)
+	}
+	if cfg.TokenEndpoint != "https://auth.example.com/token" {
+		t.Fatalf("token endpoint = %q", cfg.TokenEndpoint)
+	}
+	if flow != FlowLoopback {
+		t.Fatalf("flow = %q, want loopback default", flow)
+	}
+}
+
+func TestResolveConfigDeviceFlow(t *testing.T) {
+	r := NewRegistry()
+	env := map[string]string{
+		"ZERO_OAUTH_DEMO_CLIENT_ID":  "c",
+		"ZERO_OAUTH_DEMO_TOKEN_URL":  "https://auth.example.com/token",
+		"ZERO_OAUTH_DEMO_DEVICE_URL": "https://auth.example.com/device",
+		"ZERO_OAUTH_DEMO_FLOW":       "device",
+	}
+	_, flow, err := r.ResolveConfig("demo", env)
+	if err != nil {
+		t.Fatalf("ResolveConfig: %v", err)
+	}
+	if flow != FlowDevice {
+		t.Fatalf("flow = %q, want device", flow)
+	}
+}
+
+func TestResolveConfigRequiresClientID(t *testing.T) {
+	r := NewRegistry()
+	if _, _, err := r.ResolveConfig("demo", map[string]string{}); err == nil {
+		t.Fatal("missing client id must error")
+	}
+}
+
+func TestResolveConfigRequiresEndpointsOrIssuer(t *testing.T) {
+	r := NewRegistry()
+	// client id but no token endpoint and no issuer => error.
+	_, _, err := r.ResolveConfig("demo", map[string]string{"ZERO_OAUTH_DEMO_CLIENT_ID": "c"})
+	if err == nil {
+		t.Fatal("missing endpoints/issuer must error")
+	}
+}
+
+func TestResolveConfigRejectsInsecureEndpoint(t *testing.T) {
+	r := NewRegistry()
+	env := map[string]string{
+		"ZERO_OAUTH_DEMO_CLIENT_ID":     "c",
+		"ZERO_OAUTH_DEMO_AUTHORIZE_URL": "https://auth.example.com/authorize",
+		"ZERO_OAUTH_DEMO_TOKEN_URL":     "http://insecure.example/token", // non-https, non-loopback
+	}
+	_, _, err := r.ResolveConfig("demo", env)
+	if !errors.Is(err, ErrInsecureTokenEndpoint) {
+		t.Fatalf("err = %v, want ErrInsecureTokenEndpoint", err)
+	}
+}
+
+func TestResolveConfigInvalidName(t *testing.T) {
+	r := NewRegistry()
+	for _, bad := range []string{"", "has space", "../escape", "a/b"} {
+		if _, _, err := r.ResolveConfig(bad, nil); err == nil {
+			t.Errorf("ResolveConfig(%q) should reject invalid name", bad)
+		}
+	}
+}
+
+func TestEnvKey(t *testing.T) {
+	if got := envKey("my-svc", "CLIENT_ID"); got != "ZERO_OAUTH_MY_SVC_CLIENT_ID" {
+		t.Fatalf("envKey = %q", got)
+	}
+	if got := envKey("two.part", "SCOPES"); got != "ZERO_OAUTH_TWO_PART_SCOPES" {
+		t.Fatalf("envKey = %q", got)
+	}
+}

internal/oauth/scheduler.go

@@ -0,0 +1,125 @@
+package oauth
+
+import (
+	"context"
+	"sync"
+	"time"
+)
+
+// maxSchedulerLoadErrors bounds consecutive token-load failures before the
+// scheduler gives up, so a transient error retries but a permanent one (e.g. a
+// deleted token) does not loop forever.
+const maxSchedulerLoadErrors = 5
+
+// RefreshScheduler proactively refreshes a stored provider token shortly before
+// it expires. It is an OPTIMIZATION over the on-demand GetFresh path, never the
+// source of truth: a failed scheduled refresh is non-fatal and simply retried at
+// the next expiry. It is a no-op for a token with no refresh token or no expiry.
+type RefreshScheduler struct {
+	mu      sync.Mutex
+	cancel  context.CancelFunc
+	done    chan struct{}
+	started bool
+}
+
+// NewRefreshScheduler returns an idle scheduler.
+func NewRefreshScheduler() *RefreshScheduler {
+	return &RefreshScheduler{}
+}
+
+// Start begins proactively refreshing key via the manager until the context is
+// canceled or Stop is called. It returns immediately; refresh happens in a
+// goroutine. Calling Start twice is a no-op after the first.
+func (s *RefreshScheduler) Start(ctx context.Context, m *Manager, key string) {
+	s.mu.Lock()
+	if s.started {
+		s.mu.Unlock()
+		return
+	}
+	runCtx, cancel := context.WithCancel(ctx)
+	s.cancel = cancel
+	s.done = make(chan struct{})
+	s.started = true
+	done := s.done
+	s.mu.Unlock()
+
+	go s.loop(runCtx, m, key, done)
+}
+
+func (s *RefreshScheduler) loop(ctx context.Context, m *Manager, key string, done chan struct{}) {
+	defer close(done)
+	loadErrors := 0
+	for {
+		token, _, err := m.loadForKey(key)
+		if err != nil {
+			// A load error may be transient (a concurrent store write, a brief I/O
+			// hiccup): back off and retry rather than permanently stopping proactive
+			// refresh. Give up only after repeated consecutive failures.
+			loadErrors++
+			if loadErrors >= maxSchedulerLoadErrors {
+				return
+			}
+			select {
+			case <-ctx.Done():
+				return
+			case <-time.After(m.buffer):
+			}
+			continue
+		}
+		loadErrors = 0
+		// No refresh token or no expiry => nothing to schedule on a timer.
+		if token.RefreshToken == "" || token.ExpiresAt.IsZero() {
+			return
+		}
+		delay := s.delayUntilRefresh(token.ExpiresAt, m.buffer, m.now())
+		select {
+		case <-ctx.Done():
+			return
+		case <-time.After(delay):
+		}
+		// Best-effort refresh; failure is non-fatal — reload and reschedule (the
+		// on-demand GetFresh remains the safety net for callers).
+		if _, err := m.GetFresh(ctx, key); err != nil {
+			// Wait a bounded backoff before retrying so a persistently failing
+			// refresh does not hot-loop.
+			select {
+			case <-ctx.Done():
+				return
+			case <-time.After(m.buffer):
+			}
+		}
+	}
+}
+
+// delayUntilRefresh computes how long to wait before refreshing: aim for buffer
+// before expiry, clamped to >= 0, plus a small deterministic jitter so many
+// sessions do not refresh in lockstep.
+func (s *RefreshScheduler) delayUntilRefresh(expiresAt time.Time, buffer time.Duration, now time.Time) time.Duration {
+	target := expiresAt.Add(-buffer)
+	delay := target.Sub(now)
+	if delay < 0 {
+		delay = 0
+	}
+	// Jitter up to ~10% of the buffer, derived from the expiry nanos (no RNG dep).
+	if buffer > 0 {
+		jitter := time.Duration(expiresAt.UnixNano()%int64(buffer/10+1)) % (buffer/10 + 1)
+		delay += jitter
+	}
+	return delay
+}
+
+// Stop cancels the scheduler and waits for its goroutine to exit. Safe to call
+// more than once and on a never-started scheduler.
+func (s *RefreshScheduler) Stop() {
+	s.mu.Lock()
+	cancel := s.cancel
+	done := s.done
+	s.cancel = nil
+	s.mu.Unlock()
+	if cancel != nil {
+		cancel()
+	}
+	if done != nil {
+		<-done
+	}
+}

internal/oauth/scheduler_test.go

@@ -0,0 +1,72 @@
+package oauth
+
+import (
+	"context"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"path/filepath"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+func TestRefreshSchedulerNoRefreshTokenIsNoop(t *testing.T) {
+	m := managerFor(t, map[string]string{"ZERO_OAUTH_DEMO_CLIENT_ID": "c"}, nil)
+	// Token without a refresh token => scheduler must exit promptly (no-op).
+	if err := m.store.Save(ProviderKey("demo"), Token{AccessToken: "a", ExpiresAt: time.Now().Add(time.Hour)}); err != nil {
+		t.Fatalf("seed Save: %v", err)
+	}
+	s := NewRefreshScheduler()
+	s.Start(context.Background(), m, ProviderKey("demo"))
+	// Stop waits for the goroutine; if it didn't exit on its own this still returns
+	// because Stop cancels — but the loop should already be done.
+	s.Stop()
+}
+
+func TestRefreshSchedulerRefreshesBeforeExpiry(t *testing.T) {
+	var hits atomic.Int32
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		hits.Add(1)
+		_, _ = io.WriteString(w, `{"access_token":"refreshed","expires_in":3600}`)
+	}))
+	defer server.Close()
+
+	store, err := NewStore(StoreOptions{FilePath: filepath.Join(t.TempDir(), "tok.json")})
+	if err != nil {
+		t.Fatalf("NewStore: %v", err)
+	}
+	m, err := NewManager(ManagerOptions{
+		Store: store,
+		Env: map[string]string{
+			"ZERO_OAUTH_DEMO_CLIENT_ID": "c",
+			"ZERO_OAUTH_DEMO_TOKEN_URL": server.URL,
+		},
+		RefreshBuffer: 10 * time.Millisecond,
+	})
+	if err != nil {
+		t.Fatalf("NewManager: %v", err)
+	}
+	// Expiry just past the buffer so the scheduled delay is ~0 and a refresh fires
+	// almost immediately.
+	if err := store.Save(ProviderKey("demo"), Token{AccessToken: "old", RefreshToken: "rt", ExpiresAt: time.Now().Add(15 * time.Millisecond)}); err != nil {
+		t.Fatalf("seed Save: %v", err)
+	}
+
+	s := NewRefreshScheduler()
+	s.Start(context.Background(), m, ProviderKey("demo"))
+	defer s.Stop()
+
+	deadline := time.Now().Add(2 * time.Second)
+	for time.Now().Before(deadline) {
+		stored, _, err := store.Load(ProviderKey("demo"))
+		if err != nil {
+			t.Fatalf("Load: %v", err)
+		}
+		if stored.AccessToken == "refreshed" {
+			return // success
+		}
+		time.Sleep(5 * time.Millisecond)
+	}
+	t.Fatalf("scheduler did not refresh before expiry (token endpoint hits=%d)", hits.Load())
+}

internal/oauth/store.go

@@ -0,0 +1,324 @@
+package oauth
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"regexp"
+	"sort"
+	"strings"
+	"sync"
+	"time"
+)
+
+const (
+	storeSchemaVersion = 1
+	// KeyPrefixProvider namespaces provider-login tokens; MCP server tokens live
+	// under KeyPrefixMCP in the same format (so a future MCP migration is a key
+	// rename, not a format change).
+	KeyPrefixProvider = "provider:"
+	KeyPrefixMCP      = "mcp:"
+)
+
+// keyPattern bounds a token key to a safe, single-segment namespaced identifier
+// so a key can never traverse or collide with store internals.
+var keyPattern = regexp.MustCompile(`^(provider|mcp):[A-Za-z0-9][A-Za-z0-9._-]{0,127}$`)
+
+// ValidateKey reports whether key is a well-formed namespaced token key.
+func ValidateKey(key string) error {
+	if !keyPattern.MatchString(key) {
+		return fmt.Errorf("oauth: invalid token key %q (want \"provider:<name>\" or \"mcp:<name>\")", key)
+	}
+	return nil
+}
+
+// ProviderKey builds the store key for a provider login.
+func ProviderKey(name string) string { return KeyPrefixProvider + name }
+
+// Status is a redaction-safe summary of a stored token (no secret material).
+type Status struct {
+	Key             string    `json:"key"`
+	HasToken        bool      `json:"hasToken"`
+	HasRefreshToken bool      `json:"hasRefreshToken"`
+	TokenType       string    `json:"tokenType,omitempty"`
+	Account         string    `json:"account,omitempty"`
+	Scopes          []string  `json:"scopes,omitempty"`
+	ExpiresAt       time.Time `json:"expiresAt,omitempty"`
+	Expired         bool      `json:"expired"`
+}
+
+// StoreOptions configures where provider OAuth tokens are persisted.
+type StoreOptions struct {
+	FilePath string
+	Env      map[string]string
+	Now      func() time.Time
+}
+
+// Store persists OAuth tokens (provider + MCP namespaces) in a 0600 JSON file,
+// guarded by a cross-process lock and written atomically.
+type Store struct {
+	filePath string
+	now      func() time.Time
+	mu       sync.Mutex
+}
+
+type storeFile struct {
+	SchemaVersion int              `json:"schemaVersion"`
+	Tokens        map[string]Token `json:"tokens"`
+}
+
+// ResolveStorePath determines the on-disk location for provider OAuth tokens,
+// honoring ZERO_OAUTH_TOKENS_PATH, then XDG_CONFIG_HOME, then the home dir.
+func ResolveStorePath(env map[string]string) (string, error) {
+	if override := strings.TrimSpace(envValue(env, "ZERO_OAUTH_TOKENS_PATH")); override != "" {
+		if filepath.IsAbs(override) {
+			return filepath.Clean(override), nil
+		}
+		return filepath.Abs(override)
+	}
+	configHome := strings.TrimSpace(envValue(env, "XDG_CONFIG_HOME"))
+	if configHome == "" {
+		home := strings.TrimSpace(firstNonEmpty(envValue(env, "HOME"), envValue(env, "USERPROFILE")))
+		if home == "" {
+			var err error
+			home, err = os.UserHomeDir()
+			if err != nil {
+				return "", fmt.Errorf("oauth: resolve user home: %w", err)
+			}
+		}
+		configHome = filepath.Join(home, ".config")
+	} else if !filepath.IsAbs(configHome) {
+		resolved, err := filepath.Abs(configHome)
+		if err != nil {
+			return "", err
+		}
+		configHome = resolved
+	}
+	return filepath.Join(configHome, "zero", "oauth-tokens.json"), nil
+}
+
+// NewStore builds a file-backed token store.
+func NewStore(options StoreOptions) (*Store, error) {
+	filePath := options.FilePath
+	var err error
+	if strings.TrimSpace(filePath) == "" {
+		filePath, err = ResolveStorePath(options.Env)
+		if err != nil {
+			return nil, err
+		}
+	}
+	if !filepath.IsAbs(filePath) {
+		filePath, err = filepath.Abs(filePath)
+		if err != nil {
+			return nil, err
+		}
+	}
+	now := options.Now
+	if now == nil {
+		now = time.Now
+	}
+	return &Store{filePath: filepath.Clean(filePath), now: now}, nil
+}
+
+// Save persists a token under key, replacing any existing entry.
+func (s *Store) Save(key string, token Token) error {
+	if err := ValidateKey(key); err != nil {
+		return err
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	unlock, err := acquireFileLock(s.filePath+".lockfile", s.now)
+	if err != nil {
+		return err
+	}
+	defer unlock()
+	state, err := s.readState()
+	if err != nil {
+		return err
+	}
+	state.Tokens[key] = token
+	return s.writeState(state)
+}
+
+// Load returns the token for key; the bool is false when none is stored.
+func (s *Store) Load(key string) (Token, bool, error) {
+	if err := ValidateKey(key); err != nil {
+		return Token{}, false, err
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	state, err := s.readState()
+	if err != nil {
+		return Token{}, false, err
+	}
+	token, ok := state.Tokens[key]
+	return token, ok, nil
+}
+
+// Delete removes the token for key, reporting whether one was present.
+func (s *Store) Delete(key string) (bool, error) {
+	if err := ValidateKey(key); err != nil {
+		return false, err
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	unlock, err := acquireFileLock(s.filePath+".lockfile", s.now)
+	if err != nil {
+		return false, err
+	}
+	defer unlock()
+	state, err := s.readState()
+	if err != nil {
+		return false, err
+	}
+	if _, ok := state.Tokens[key]; !ok {
+		return false, nil
+	}
+	delete(state.Tokens, key)
+	if err := s.writeState(state); err != nil {
+		return false, err
+	}
+	return true, nil
+}
+
+// Status returns redaction-safe summaries of every stored token, sorted by key.
+// An optional prefix filters to one namespace (e.g. KeyPrefixProvider).
+func (s *Store) Status(prefix string) ([]Status, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	state, err := s.readState()
+	if err != nil {
+		return nil, err
+	}
+	keys := make([]string, 0, len(state.Tokens))
+	for k := range state.Tokens {
+		if prefix == "" || strings.HasPrefix(k, prefix) {
+			keys = append(keys, k)
+		}
+	}
+	sort.Strings(keys)
+	now := s.now()
+	out := make([]Status, 0, len(keys))
+	for _, k := range keys {
+		token := state.Tokens[k]
+		out = append(out, Status{
+			Key:             k,
+			HasToken:        strings.TrimSpace(token.AccessToken) != "",
+			HasRefreshToken: strings.TrimSpace(token.RefreshToken) != "",
+			TokenType:       token.TokenType,
+			Account:         token.Account,
+			Scopes:          token.Scopes,
+			ExpiresAt:       token.ExpiresAt,
+			Expired:         token.Expired(now),
+		})
+	}
+	return out, nil
+}
+
+func (s *Store) readState() (storeFile, error) {
+	data, err := os.ReadFile(s.filePath)
+	if err != nil {
+		if errors.Is(err, os.ErrNotExist) {
+			return emptyStoreFile(), nil
+		}
+		return storeFile{}, err
+	}
+	var state storeFile
+	if err := json.Unmarshal(data, &state); err != nil {
+		return storeFile{}, fmt.Errorf("oauth: invalid token file at %s: %w", s.filePath, err)
+	}
+	if state.SchemaVersion != storeSchemaVersion {
+		return storeFile{}, fmt.Errorf("oauth: invalid token file at %s: unsupported schemaVersion", s.filePath)
+	}
+	if state.Tokens == nil {
+		state.Tokens = map[string]Token{}
+	}
+	for key := range state.Tokens {
+		if err := ValidateKey(key); err != nil {
+			return storeFile{}, fmt.Errorf("oauth: invalid token file at %s: %w", s.filePath, err)
+		}
+	}
+	return state, nil
+}
+
+func (s *Store) writeState(state storeFile) error {
+	if err := os.MkdirAll(filepath.Dir(s.filePath), 0o700); err != nil {
+		return err
+	}
+	data, err := json.MarshalIndent(state, "", "  ")
+	if err != nil {
+		return err
+	}
+	tempPath := fmt.Sprintf("%s.tmp-%d-%d", s.filePath, os.Getpid(), s.now().UnixNano())
+	if err := os.WriteFile(tempPath, append(data, '\n'), 0o600); err != nil {
+		return err
+	}
+	if err := os.Rename(tempPath, s.filePath); err != nil {
+		_ = os.Remove(tempPath)
+		return err
+	}
+	return nil
+}
+
+func emptyStoreFile() storeFile {
+	return storeFile{SchemaVersion: storeSchemaVersion, Tokens: map[string]Token{}}
+}
+
+// FormatStatuses renders a human-readable status table without leaking token
+// material.
+func FormatStatuses(statuses []Status) string {
+	if len(statuses) == 0 {
+		return "No OAuth provider logins are stored."
+	}
+	var b strings.Builder
+	for i, st := range statuses {
+		if i > 0 {
+			b.WriteByte('\n')
+		}
+		name := strings.TrimPrefix(st.Key, KeyPrefixProvider)
+		b.WriteString(name)
+		b.WriteString(": ")
+		if !st.HasToken {
+			b.WriteString("no token")
+			continue
+		}
+		b.WriteString("logged in")
+		if st.Account != "" {
+			b.WriteString(" as " + st.Account)
+		}
+		if st.HasRefreshToken {
+			b.WriteString(" (refreshable)")
+		}
+		if !st.ExpiresAt.IsZero() {
+			if st.Expired {
+				b.WriteString(", expired at ")
+			} else {
+				b.WriteString(", expires ")
+			}
+			b.WriteString(st.ExpiresAt.UTC().Format(time.RFC3339))
+		}
+	}
+	return b.String()
+}
+
+// envValue reads a variable. A non-nil env map is authoritative (hermetic): a
+// missing key returns "" rather than falling back to the process environment, so
+// a caller/test that passes a controlled map can never pick up ambient
+// ZERO_OAUTH_* / HOME / XDG_CONFIG_HOME values. Only a nil map uses os.Getenv.
+func envValue(env map[string]string, key string) string {
+	if env != nil {
+		return env[key]
+	}
+	return os.Getenv(key)
+}
+
+func firstNonEmpty(values ...string) string {
+	for _, v := range values {
+		if strings.TrimSpace(v) != "" {
+			return v
+		}
+	}
+	return ""
+}

internal/oauth/store_test.go

@@ -0,0 +1,128 @@
+package oauth
+
+import (
+	"os"
+	"path/filepath"
+	"runtime"
+	"testing"
+	"time"
+)
+
+func newTestStore(t *testing.T) (*Store, string) {
+	t.Helper()
+	path := filepath.Join(t.TempDir(), "oauth-tokens.json")
+	s, err := NewStore(StoreOptions{FilePath: path})
+	if err != nil {
+		t.Fatalf("NewStore: %v", err)
+	}
+	return s, path
+}
+
+func TestStoreSaveLoadDelete(t *testing.T) {
+	s, _ := newTestStore(t)
+	tok := Token{AccessToken: "at", RefreshToken: "rt", Account: "me@x"}
+	if err := s.Save(ProviderKey("demo"), tok); err != nil {
+		t.Fatalf("Save: %v", err)
+	}
+	if err := s.Save(KeyPrefixMCP+"server1", Token{AccessToken: "mcp-at"}); err != nil {
+		t.Fatalf("Save mcp: %v", err)
+	}
+	got, ok, err := s.Load(ProviderKey("demo"))
+	if err != nil || !ok || got.AccessToken != "at" || got.Account != "me@x" {
+		t.Fatalf("Load = %+v ok=%v err=%v", got, ok, err)
+	}
+	removed, err := s.Delete(ProviderKey("demo"))
+	if err != nil || !removed {
+		t.Fatalf("Delete = %v %v", removed, err)
+	}
+	if _, ok, _ := s.Load(ProviderKey("demo")); ok {
+		t.Fatal("token should be gone after delete")
+	}
+	// The mcp-namespaced token is untouched.
+	if _, ok, _ := s.Load(KeyPrefixMCP + "server1"); !ok {
+		t.Fatal("mcp token should survive provider delete")
+	}
+}
+
+func TestStoreRejectsInvalidKeys(t *testing.T) {
+	s, _ := newTestStore(t)
+	for _, bad := range []string{"demo", "provider:", "provider:../escape", "mcp:bad/key", "other:x", ""} {
+		if err := s.Save(bad, Token{AccessToken: "x"}); err == nil {
+			t.Errorf("Save(%q) should be rejected", bad)
+		}
+	}
+	for _, ok := range []string{"provider:demo", "mcp:server-1", "provider:two-svc"} {
+		if err := ValidateKey(ok); err != nil {
+			t.Errorf("ValidateKey(%q) = %v, want nil", ok, err)
+		}
+	}
+}
+
+func TestStoreStatusFiltersByPrefix(t *testing.T) {
+	s, _ := newTestStore(t)
+	_ = s.Save(ProviderKey("demo"), Token{AccessToken: "a", RefreshToken: "r", ExpiresAt: time.Now().Add(time.Hour)})
+	_ = s.Save(KeyPrefixMCP+"srv", Token{AccessToken: "m"})
+	statuses, err := s.Status(KeyPrefixProvider)
+	if err != nil {
+		t.Fatalf("Status: %v", err)
+	}
+	if len(statuses) != 1 || statuses[0].Key != ProviderKey("demo") {
+		t.Fatalf("provider status = %+v", statuses)
+	}
+	if !statuses[0].HasRefreshToken {
+		t.Fatal("status should report a refresh token")
+	}
+}
+
+func TestStoreFileMode0600(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("unix file modes")
+	}
+	s, path := newTestStore(t)
+	if err := s.Save(ProviderKey("x"), Token{AccessToken: "a"}); err != nil {
+		t.Fatalf("Save: %v", err)
+	}
+	info, err := os.Stat(path)
+	if err != nil {
+		t.Fatalf("stat: %v", err)
+	}
+	if perm := info.Mode().Perm(); perm != 0o600 {
+		t.Fatalf("token file mode = %o, want 600", perm)
+	}
+}
+
+func TestStoreMalformedFailsClosed(t *testing.T) {
+	s, path := newTestStore(t)
+	if err := os.WriteFile(path, []byte("{not json"), 0o600); err != nil {
+		t.Fatalf("corrupt: %v", err)
+	}
+	if _, _, err := s.Load(ProviderKey("x")); err == nil {
+		t.Fatal("malformed store must fail closed")
+	}
+}
+
+func TestEnvValueHermetic(t *testing.T) {
+	t.Setenv("ZERO_OAUTH_DEMO_CLIENT_ID", "ambient")
+	// A non-nil map omitting the key must NOT leak the ambient process value.
+	if got := envValue(map[string]string{}, "ZERO_OAUTH_DEMO_CLIENT_ID"); got != "" {
+		t.Fatalf("non-nil env map must be hermetic, got %q", got)
+	}
+	// A nil map reads the process environment.
+	if got := envValue(nil, "ZERO_OAUTH_DEMO_CLIENT_ID"); got != "ambient" {
+		t.Fatalf("nil env map should read os env, got %q", got)
+	}
+}
+
+func TestResolveStorePathHonorsOverride(t *testing.T) {
+	// Use an OS-appropriate absolute path: a unix-style "/tmp/..." literal is not
+	// absolute on Windows (no drive), so it would be resolved against the current
+	// drive and a verbatim comparison would fail there.
+	override := filepath.Join(t.TempDir(), "custom", "tok.json")
+	path, err := ResolveStorePath(map[string]string{"ZERO_OAUTH_TOKENS_PATH": override})
+	if err != nil {
+		t.Fatalf("ResolveStorePath: %v", err)
+	}
+	if path != override {
+		t.Fatalf("path = %q, want %q", path, override)
+	}
+}

internal/providers/factory.go

@@ -40,6 +40,7 @@ func New(profile config.ProviderProfile, options Options) (zeroruntime.Provider,
 			MaxTokens:       resolved.maxOutputTokens,
 			HTTPClient:      options.HTTPClient,
 			UserAgent:       options.UserAgent,
+			ParseThinkTags:  parseThinkTagsForProfile(profile, resolved),
 		})
 	case config.ProviderKindAnthropic, config.ProviderKindAnthropicCompat:
 		return anthropic.New(anthropic.Options{
@@ -72,6 +73,40 @@ func New(profile config.ProviderProfile, options Options) (zeroruntime.Provider,
 	}
 }
 
+func parseThinkTagsForProfile(profile config.ProviderProfile, resolved resolvedProfile) bool {
+	if profile.ParseThinkTags != nil {
+		return *profile.ParseThinkTags
+	}
+	if resolved.providerKind != config.ProviderKindOpenAICompatible {
+		return false
+	}
+	return modelMayEmitThinkTags(resolved.apiModel)
+}
+
+func modelMayEmitThinkTags(model string) bool {
+	model = strings.ToLower(strings.TrimSpace(model))
+	for _, marker := range []string{
+		"deepseek-r1",
+		"deepseek-reasoner",
+		"gpt-oss",
+		"glm-z1",
+		"kimi-k2-thinking",
+		"magistral",
+		"minimax-m3",
+		"nemotron",
+		"qwen3",
+		"qwq",
+		"reasoner",
+		"reasoning",
+		"thinking",
+	} {
+		if strings.Contains(model, marker) {
+			return true
+		}
+	}
+	return false
+}
+
 type resolvedProfile struct {
 	providerKind    config.ProviderKind
 	apiModel        string

internal/providers/factory_test.go

@@ -108,6 +108,33 @@ func TestNewSupportsOpenAIProviderKind(t *testing.T) {
 	}
 }
 
+func TestParseThinkTagsForProfileUsesConservativeDefaultsAndOverride(t *testing.T) {
+	openAICompatible := resolvedProfile{providerKind: config.ProviderKindOpenAICompatible, apiModel: "qwen3-coder:480b"}
+	if !parseThinkTagsForProfile(config.ProviderProfile{}, openAICompatible) {
+		t.Fatal("qwen3 OpenAI-compatible model should parse inline think tags by default")
+	}
+
+	generic := resolvedProfile{providerKind: config.ProviderKindOpenAICompatible, apiModel: "factory-model"}
+	if parseThinkTagsForProfile(config.ProviderProfile{}, generic) {
+		t.Fatal("generic OpenAI-compatible model should preserve literal think tags by default")
+	}
+
+	official := resolvedProfile{providerKind: config.ProviderKindOpenAI, apiModel: "gpt-4.1"}
+	if parseThinkTagsForProfile(config.ProviderProfile{}, official) {
+		t.Fatal("official OpenAI model should preserve literal think tags by default")
+	}
+
+	enabled := true
+	if !parseThinkTagsForProfile(config.ProviderProfile{ParseThinkTags: &enabled}, generic) {
+		t.Fatal("explicit parseThinkTags=true should enable inline think parsing")
+	}
+
+	disabled := false
+	if parseThinkTagsForProfile(config.ProviderProfile{ParseThinkTags: &disabled}, openAICompatible) {
+		t.Fatal("explicit parseThinkTags=false should disable inline think parsing")
+	}
+}
+
 func TestNewResolvesKnownModelToAPIModelAndProvider(t *testing.T) {
 	transport := &captureTransport{
 		responseBody: "data: {\"type\":\"message_stop\"}\n\n",

internal/providers/openai/provider.go

@@ -40,6 +40,9 @@ type Options struct {
 	// StreamIdleTimeout aborts the stream if no data arrives for this long.
 	// Zero uses defaultStreamIdleTimeout.
 	StreamIdleTimeout time.Duration
+	// ParseThinkTags converts streamed <think>...</think> content into reasoning
+	// events for OpenAI-compatible models known to emit that legacy format.
+	ParseThinkTags bool
 }
 
 // Provider streams completions from an OpenAI-compatible chat completions API.
@@ -55,6 +58,7 @@ type Provider struct {
 	httpClient        *http.Client
 	userAgent         string
 	streamIdleTimeout time.Duration
+	parseThinkTags    bool
 }
 
 // New creates an OpenAI-compatible provider.
@@ -100,6 +104,7 @@ func New(options Options) (*Provider, error) {
 		httpClient:        httpClient,
 		userAgent:         options.UserAgent,
 		streamIdleTimeout: idleTimeout,
+		parseThinkTags:    options.ParseThinkTags,
 	}, nil
 }
 
@@ -162,7 +167,7 @@ func (provider *Provider) stream(ctx context.Context, body []byte, events chan<-
 		return
 	}
 
-	state := newToolState()
+	state := newToolState(provider.parseThinkTags)
 	// Use the shared SSE reader (also used by the Anthropic/Gemini providers) so
 	// multi-line "data:" continuation fields are joined into one payload, and the
 	// idle watchdog / context cancellation are handled uniformly.
@@ -170,7 +175,8 @@ func (provider *Provider) stream(ctx context.Context, body []byte, events chan<-
 		return provider.emitPayload(ctx, data, state, events)
 	})
 	if errors.Is(err, providerio.ErrStreamIdle) {
-		state.closeOpen(ctx, events)
+		state.flushBufferedContent(events)
+		state.closeBufferedOpen(events)
 		sendEvent(ctx, events, zeroruntime.StreamEvent{
 			Type:  zeroruntime.StreamEventError,
 			Error: provider.redact(fmt.Sprintf("provider stream error: idle timeout after %s (upstream stopped sending data)", provider.streamIdleTimeout)),
@@ -178,16 +184,19 @@ func (provider *Provider) stream(ctx context.Context, body []byte, events chan<-
 		return
 	}
 	if err != nil {
-		state.closeOpen(ctx, events)
+		state.flushBufferedContent(events)
+		state.closeBufferedOpen(events)
 		sendEvent(ctx, events, zeroruntime.StreamEvent{Type: zeroruntime.StreamEventError, Error: provider.redact("provider stream error: " + err.Error())})
 		return
 	}
 	if ctxErr := ctx.Err(); ctxErr != nil {
-		state.closeOpen(ctx, events)
+		state.flushBufferedContent(events)
+		state.closeBufferedOpen(events)
 		sendEvent(ctx, events, zeroruntime.StreamEvent{Type: zeroruntime.StreamEventError, Error: provider.redact("provider stream error: " + ctxErr.Error())})
 		return
 	}
 	if !state.done {
+		state.flushContent(ctx, events)
 		state.closeOpen(ctx, events)
 		sendEvent(ctx, events, zeroruntime.StreamEvent{Type: zeroruntime.StreamEventDone, FinishReason: state.finishReason})
 	}
@@ -199,6 +208,7 @@ func (provider *Provider) stream(ctx context.Context, body []byte, events chan<-
 func (provider *Provider) emitPayload(ctx context.Context, data string, state *toolState, events chan<- zeroruntime.StreamEvent) bool {
 	var chunk streamChunk
 	if err := json.Unmarshal([]byte(data), &chunk); err != nil {
+		state.flushContent(ctx, events)
 		state.closeOpen(ctx, events)
 		sendEvent(ctx, events, zeroruntime.StreamEvent{
 			Type:  zeroruntime.StreamEventError,
@@ -208,6 +218,7 @@ func (provider *Provider) emitPayload(ctx context.Context, data string, state *t
 		return false
 	}
 	if chunk.Error != nil {
+		state.flushContent(ctx, events)
 		state.closeOpen(ctx, events)
 		sendEvent(ctx, events, zeroruntime.StreamEvent{
 			Type:  zeroruntime.StreamEventError,
@@ -234,15 +245,13 @@ func (provider *Provider) emitChunk(
 			})
 		}
 		if choice.Delta.Content != "" {
-			sendEvent(ctx, events, zeroruntime.StreamEvent{
-				Type:    zeroruntime.StreamEventText,
-				Content: choice.Delta.Content,
-			})
+			state.emitContent(ctx, events, choice.Delta.Content)
 		}
 		for _, toolCall := range choice.Delta.ToolCalls {
 			state.applyDelta(ctx, toolCall, events)
 		}
 		if choice.FinishReason == "tool_calls" {
+			state.flushContent(ctx, events)
 			state.closeOpen(ctx, events)
 		}
 		if reason := mapFinishReason(choice.FinishReason); reason != "" {
@@ -314,6 +323,13 @@ func sendEvent(ctx context.Context, events chan<- zeroruntime.StreamEvent, event
 	}
 }
 
+func sendBufferedEvent(events chan<- zeroruntime.StreamEvent, event zeroruntime.StreamEvent) {
+	select {
+	case events <- event:
+	default:
+	}
+}
+
 func (provider *Provider) openAIRequest(request zeroruntime.CompletionRequest) chatCompletionRequest {
 	messages := make([]chatMessage, 0, len(request.Messages))
 	for _, message := range request.Messages {

internal/providers/openai/provider_test.go

@@ -244,6 +244,53 @@ func TestStreamCompletionEmitsReasoningBeforeRegularContent(t *testing.T) {
 	}
 }
 
+func TestStreamCompletionPreservesLiteralThinkTagsByDefault(t *testing.T) {
+	provider := newTestProvider(t, func(w http.ResponseWriter, r *http.Request) {
+		writeSSE(w, `{"choices":[{"delta":{"content":"show <think>literal</think> markup"}}]}`)
+		writeSSE(w, `[DONE]`)
+	})
+
+	events := collectProviderEvents(t, provider)
+	assertEvent(t, events[0], zeroruntime.StreamEventText, "show <think>literal</think> markup")
+	if reasoning := eventsOfType(events, zeroruntime.StreamEventReasoning); len(reasoning) != 0 {
+		t.Fatalf("literal think tags must not emit reasoning by default, got %#v", reasoning)
+	}
+}
+
+func TestStreamCompletionSplitsInlineThinkTagsFromContent(t *testing.T) {
+	provider := newTestProviderWithThinkTags(t, func(w http.ResponseWriter, r *http.Request) {
+		writeSSE(w, `{"choices":[{"delta":{"content":"<think>private reasoning</think>public answer"}}]}`)
+		writeSSE(w, `[DONE]`)
+	})
+
+	events := collectProviderEvents(t, provider)
+	assertEvent(t, events[0], zeroruntime.StreamEventReasoning, "private reasoning")
+	assertEvent(t, events[1], zeroruntime.StreamEventText, "public answer")
+	if events[2].Type != zeroruntime.StreamEventDone {
+		t.Fatalf("third event = %#v, want done", events[2])
+	}
+}
+
+func TestStreamCompletionSplitsInlineThinkTagsAcrossChunks(t *testing.T) {
+	provider := newTestProviderWithThinkTags(t, func(w http.ResponseWriter, r *http.Request) {
+		writeSSE(w, `{"choices":[{"delta":{"content":"<thi"}}]}`)
+		writeSSE(w, `{"choices":[{"delta":{"content":"nk>reason"}}]}`)
+		writeSSE(w, `{"choices":[{"delta":{"content":"ing</"}}]}`)
+		writeSSE(w, `{"choices":[{"delta":{"content":"think> answer"}}]}`)
+		writeSSE(w, `[DONE]`)
+	})
+
+	events := collectProviderEvents(t, provider)
+	reasoning := eventsOfType(events, zeroruntime.StreamEventReasoning)
+	if len(reasoning) != 2 || reasoning[0].Content != "reason" || reasoning[1].Content != "ing" {
+		t.Fatalf("reasoning events = %#v, want split reasoning content", reasoning)
+	}
+	text := eventsOfType(events, zeroruntime.StreamEventText)
+	if len(text) != 1 || text[0].Content != " answer" {
+		t.Fatalf("text events = %#v, want answer-only content", text)
+	}
+}
+
 func TestStreamCompletionBuffersToolArgsUntilIDAndNameArrive(t *testing.T) {
 	provider := newTestProvider(t, func(w http.ResponseWriter, r *http.Request) {
 		writeSSE(w, `{"choices":[{"delta":{"tool_calls":[{"index":0,"function":{"arguments":"{\"path\":"}}]}}]}`)
@@ -385,19 +432,83 @@ func TestStreamCompletionEmitsErrorWhenContextCancels(t *testing.T) {
 	}
 }
 
+func TestStreamCompletionFlushesBufferedContentWhenContextCancels(t *testing.T) {
+	release := make(chan struct{})
+	provider := newTestProviderWithThinkTags(t, func(w http.ResponseWriter, r *http.Request) {
+		writeSSE(w, `{"choices":[{"delta":{"content":"visible <thi"}}]}`)
+		select {
+		case <-r.Context().Done():
+		case <-release:
+		}
+	})
+	defer close(release)
+
+	ctx, cancel := context.WithCancel(context.Background())
+	stream, err := provider.StreamCompletion(ctx, zeroruntime.CompletionRequest{})
+	if err != nil {
+		t.Fatalf("StreamCompletion returned setup error: %v", err)
+	}
+
+	events := []zeroruntime.StreamEvent{}
+	select {
+	case event, ok := <-stream:
+		if !ok {
+			t.Fatal("stream closed before first text event")
+		}
+		events = append(events, event)
+	case <-time.After(2 * time.Second):
+		t.Fatal("timed out waiting for first text event")
+	}
+	cancel()
+	events = append(events, readAll(stream)...)
+
+	if len(events) != 3 ||
+		events[0].Type != zeroruntime.StreamEventText || events[0].Content != "visible " ||
+		events[1].Type != zeroruntime.StreamEventText || events[1].Content != "<thi" ||
+		events[2].Type != zeroruntime.StreamEventError {
+		t.Fatalf("events = %#v, want text, buffered text, then context error", events)
+	}
+	if done := eventsOfType(events, zeroruntime.StreamEventDone); len(done) != 0 {
+		t.Fatalf("events = %#v, want no done after context cancel", events)
+	}
+}
+
 func newTestProvider(t *testing.T, handler http.HandlerFunc) *Provider {
 	t.Helper()
 	return newTestProviderWithKey(t, "", handler)
 }
 
 func newTestProviderWithKey(t *testing.T, apiKey string, handler http.HandlerFunc) *Provider {
+	t.Helper()
+	return newTestProviderWithOptions(t, Options{APIKey: apiKey}, handler)
+}
+
+func newTestProviderWithThinkTags(t *testing.T, handler http.HandlerFunc) *Provider {
+	t.Helper()
+	return newTestProviderWithOptions(t, Options{ParseThinkTags: true}, handler)
+}
+
+func newTestProviderWithOptions(t *testing.T, options Options, handler http.HandlerFunc) *Provider {
 	t.Helper()
 	server := httptest.NewServer(handler)
 	t.Cleanup(server.Close)
+	options.BaseURL = server.URL
+	if strings.TrimSpace(options.Model) == "" {
+		options.Model = "gpt-test"
+	}
 	provider, err := New(Options{
-		APIKey:  apiKey,
-		BaseURL: server.URL,
-		Model:   "gpt-test",
+		APIKey:            options.APIKey,
+		BaseURL:           options.BaseURL,
+		Model:             options.Model,
+		AuthHeader:        options.AuthHeader,
+		AuthScheme:        options.AuthScheme,
+		AuthHeaderValue:   options.AuthHeaderValue,
+		CustomHeaders:     options.CustomHeaders,
+		HTTPClient:        options.HTTPClient,
+		UserAgent:         options.UserAgent,
+		MaxTokens:         options.MaxTokens,
+		StreamIdleTimeout: options.StreamIdleTimeout,
+		ParseThinkTags:    options.ParseThinkTags,
 	})
 	if err != nil {
 		t.Fatalf("New returned error: %v", err)

internal/providers/openai/tool_state.go

@@ -3,12 +3,20 @@ package openai
 import (
 	"context"
 	"sort"
+	"strings"
 
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
+const (
+	thinkOpenTag  = "<think>"
+	thinkCloseTag = "</think>"
+)
+
 type toolState struct {
-	calls map[int]*pendingToolCall
+	calls          map[int]*pendingToolCall
+	think          thinkTagSplitter
+	parseThinkTags bool
 	// finishReason holds the normalized terminal stop reason (zeroruntime
 	// FinishReason*) when the response ended abnormally (length/content_filter),
 	// so the provider can attach it to the done event. Empty for a normal finish.
@@ -26,8 +34,128 @@ type pendingToolCall struct {
 	ended     bool
 }
 
-func newToolState() *toolState {
-	return &toolState{calls: make(map[int]*pendingToolCall)}
+func newToolState(parseThinkTags bool) *toolState {
+	return &toolState{calls: make(map[int]*pendingToolCall), parseThinkTags: parseThinkTags}
+}
+
+type streamEventEmitter func(zeroruntime.StreamEvent)
+
+type thinkTagSplitter struct {
+	pending    string
+	inThinking bool
+}
+
+func (state *toolState) emitContent(ctx context.Context, events chan<- zeroruntime.StreamEvent, content string) {
+	state.emitContentWith(content, func(event zeroruntime.StreamEvent) {
+		sendEvent(ctx, events, event)
+	})
+}
+
+func (state *toolState) flushContent(ctx context.Context, events chan<- zeroruntime.StreamEvent) {
+	state.flushContentWith(func(event zeroruntime.StreamEvent) {
+		sendEvent(ctx, events, event)
+	})
+}
+
+func (state *toolState) flushBufferedContent(events chan<- zeroruntime.StreamEvent) {
+	state.flushContentWith(func(event zeroruntime.StreamEvent) {
+		sendBufferedEvent(events, event)
+	})
+}
+
+func (state *toolState) emitContentWith(content string, emit streamEventEmitter) {
+	if !state.parseThinkTags {
+		emit(zeroruntime.StreamEvent{Type: zeroruntime.StreamEventText, Content: content})
+		return
+	}
+	state.think.push(content, func(eventType zeroruntime.StreamEventType, text string) {
+		emit(zeroruntime.StreamEvent{Type: eventType, Content: text})
+	})
+}
+
+func (state *toolState) flushContentWith(emit streamEventEmitter) {
+	if !state.parseThinkTags {
+		return
+	}
+	state.think.flush(func(eventType zeroruntime.StreamEventType, text string) {
+		emit(zeroruntime.StreamEvent{Type: eventType, Content: text})
+	})
+}
+
+func (splitter *thinkTagSplitter) push(content string, emit func(zeroruntime.StreamEventType, string)) {
+	if content == "" {
+		return
+	}
+	splitter.pending += content
+	splitter.drain(false, emit)
+}
+
+func (splitter *thinkTagSplitter) flush(emit func(zeroruntime.StreamEventType, string)) {
+	splitter.drain(true, emit)
+}
+
+func (splitter *thinkTagSplitter) drain(final bool, emit func(zeroruntime.StreamEventType, string)) {
+	for {
+		tag := thinkOpenTag
+		if splitter.inThinking {
+			tag = thinkCloseTag
+		}
+		if index := indexFold(splitter.pending, tag); index >= 0 {
+			splitter.emitCurrent(emit, splitter.pending[:index])
+			splitter.pending = splitter.pending[index+len(tag):]
+			splitter.inThinking = !splitter.inThinking
+			continue
+		}
+		if final {
+			splitter.emitCurrent(emit, splitter.pending)
+			splitter.pending = ""
+			return
+		}
+		keep := trailingTagPrefixLen(splitter.pending, tag)
+		if keep == len(splitter.pending) {
+			return
+		}
+		emitText := splitter.pending[:len(splitter.pending)-keep]
+		splitter.emitCurrent(emit, emitText)
+		splitter.pending = splitter.pending[len(splitter.pending)-keep:]
+		return
+	}
+}
+
+func (splitter *thinkTagSplitter) emitCurrent(emit func(zeroruntime.StreamEventType, string), text string) {
+	if text == "" {
+		return
+	}
+	eventType := zeroruntime.StreamEventText
+	if splitter.inThinking {
+		eventType = zeroruntime.StreamEventReasoning
+	}
+	emit(eventType, text)
+}
+
+func trailingTagPrefixLen(text string, tag string) int {
+	maxLen := min(len(text), len(tag)-1)
+	for length := maxLen; length > 0; length-- {
+		if strings.EqualFold(text[len(text)-length:], tag[:length]) {
+			return length
+		}
+	}
+	return 0
+}
+
+func indexFold(text string, needle string) int {
+	if needle == "" {
+		return 0
+	}
+	if len(text) < len(needle) {
+		return -1
+	}
+	for index := 0; index <= len(text)-len(needle); index++ {
+		if strings.EqualFold(text[index:index+len(needle)], needle) {
+			return index
+		}
+	}
+	return -1
 }
 
 func (state *toolState) applyDelta(
@@ -78,6 +206,18 @@ func (state *toolState) applyDelta(
 }
 
 func (state *toolState) closeOpen(ctx context.Context, events chan<- zeroruntime.StreamEvent) {
+	state.closeOpenWith(func(event zeroruntime.StreamEvent) {
+		sendEvent(ctx, events, event)
+	})
+}
+
+func (state *toolState) closeBufferedOpen(events chan<- zeroruntime.StreamEvent) {
+	state.closeOpenWith(func(event zeroruntime.StreamEvent) {
+		sendBufferedEvent(events, event)
+	})
+}
+
+func (state *toolState) closeOpenWith(emit streamEventEmitter) {
 	indexes := make([]int, 0, len(state.calls))
 	for index := range state.calls {
 		indexes = append(indexes, index)
@@ -95,20 +235,20 @@ func (state *toolState) closeOpen(ctx context.Context, events chan<- zeroruntime
 		if call.id == "" || call.name == "" {
 			if call.id != "" || call.name != "" || call.arguments != "" {
 				call.ended = true
-				sendEvent(ctx, events, zeroruntime.StreamEvent{Type: zeroruntime.StreamEventToolCallDropped})
+				emit(zeroruntime.StreamEvent{Type: zeroruntime.StreamEventToolCallDropped})
 			}
 			continue
 		}
 		if !call.started {
 			call.started = true
-			sendEvent(ctx, events, zeroruntime.StreamEvent{
+			emit(zeroruntime.StreamEvent{
 				Type:       zeroruntime.StreamEventToolCallStart,
 				ToolCallID: call.id,
 				ToolName:   call.name,
 			})
 		}
 		if call.arguments != "" {
-			sendEvent(ctx, events, zeroruntime.StreamEvent{
+			emit(zeroruntime.StreamEvent{
 				Type:              zeroruntime.StreamEventToolCallDelta,
 				ToolCallID:        call.id,
 				ArgumentsFragment: call.arguments,
@@ -116,7 +256,7 @@ func (state *toolState) closeOpen(ctx context.Context, events chan<- zeroruntime
 			call.arguments = ""
 		}
 		call.ended = true
-		sendEvent(ctx, events, zeroruntime.StreamEvent{
+		emit(zeroruntime.StreamEvent{
 			Type:       zeroruntime.StreamEventToolCallEnd,
 			ToolCallID: call.id,
 		})

internal/sandbox/adapters.go

@@ -23,15 +23,26 @@ type Backend struct {
 	// under a seatbelt profile that restricts outbound to the proxy port);
 	// bubblewrap isolates the network namespace with no bridge to the host proxy,
 	// so scoped egress there collapses to deny until a real relay exists.
-	ScopedEgress bool   `json:"scopedEgress,omitempty"`
-	Executable   string `json:"executable,omitempty"`
-	Message      string `json:"message,omitempty"`
+	ScopedEgress bool `json:"scopedEgress,omitempty"`
+	// ProxyEgress reports a backend that routes scoped egress through the local
+	// filtering proxy on a BEST-EFFORT basis (the child must honor the proxy env)
+	// rather than via OS-level network isolation. The WSL policy-only fallback uses
+	// this: there is no netns/seatbelt to enforce egress, but the proxy still
+	// applies the allow/deny gate to well-behaved clients.
+	ProxyEgress bool   `json:"proxyEgress,omitempty"`
+	Executable  string `json:"executable,omitempty"`
+	Message     string `json:"message,omitempty"`
 }
 
 // EnforcesScopedEgress reports whether a populated NetworkScoped allowlist can be
-// enforced by this backend. When false, scoped egress must fail closed (collapse
-// to deny) rather than silently run with unrestricted networking.
+// routed through the filtering proxy by this backend. When false, scoped egress
+// must fail closed (collapse to deny) rather than silently run with unrestricted
+// networking. A native backend needs an executable; the WSL fallback uses
+// best-effort proxy egress (ProxyEgress) with no native isolation.
 func (backend Backend) EnforcesScopedEgress() bool {
+	if backend.ProxyEgress {
+		return true
+	}
 	return backend.Available && backend.Executable != "" && backend.ScopedEgress
 }
 
@@ -65,6 +76,12 @@ func SelectBackend(options BackendOptions) Backend {
 		if path, err := lookup("bwrap"); err == nil && path != "" {
 			return nativeBackend(goos, BackendBubblewrap, path, "bubblewrap sandbox available")
 		}
+		// Under WSL (no bubblewrap), do NOT silently degrade to a plain policy-only
+		// runner: select the WSL backend, which still routes egress through the
+		// filtering proxy and is gated on AllowPolicyOnlyRunner in the runner.
+		if info := detectWSL(); info.IsWSL {
+			return wslBackend(goos, info)
+		}
 		return policyOnlyBackend(goos, "policy-only fallback: bubblewrap is not installed")
 	case "darwin":
 		if path, err := lookup("sandbox-exec"); err == nil && path != "" {
@@ -94,6 +111,26 @@ func nativeBackend(goos string, name BackendName, executable string, message str
 	}
 }
 
+// wslBackend is the policy-only WSL fallback: no native isolation, but scoped
+// egress is routed best-effort through the filtering proxy (ProxyEgress). The
+// runner gates it on Policy.AllowPolicyOnlyRunner and records a downgrade note.
+func wslBackend(goos string, info WSLInfo) Backend {
+	msg := "policy-only WSL fallback: bubblewrap unavailable under WSL; egress routed through the filtering proxy"
+	if info.IsWSL2 {
+		msg = "policy-only WSL2 fallback: bubblewrap unavailable/unreliable under WSL2; egress routed through the filtering proxy"
+	}
+	return Backend{
+		Name:            BackendWSL,
+		Available:       false,
+		Platform:        goos,
+		Fallback:        true,
+		CommandWrapping: false,
+		NativeIsolation: false,
+		ProxyEgress:     true,
+		Message:         msg,
+	}
+}
+
 func policyOnlyBackend(goos string, message string) Backend {
 	return Backend{
 		Name:            BackendPolicyOnly,

internal/sandbox/batch.go

@@ -1,33 +0,0 @@
-package sandbox
-
-// RequiredAutonomyForBatch returns the highest autonomy tier required to run a
-// batch of operations without further prompting, given each operation's
-// classified risk. RiskCritical and RiskHigh both map to AutonomyHigh because
-// the autonomy ladder has three tiers and the risk ladder has four. Any risk
-// level that is empty or unrecognised fails closed to AutonomyHigh so an
-// unclassified operation can never silently lower the batch ceiling. An empty
-// batch requires only AutonomyLow.
-func RequiredAutonomyForBatch(risks []Risk) Autonomy {
-	required := AutonomyLow
-	for _, risk := range risks {
-		tier := riskToAutonomy(risk.Level)
-		if autonomyRank[tier] > autonomyRank[required] {
-			required = tier
-		}
-	}
-	return required
-}
-
-func riskToAutonomy(level RiskLevel) Autonomy {
-	switch level {
-	case RiskLow:
-		return AutonomyLow
-	case RiskMedium:
-		return AutonomyMedium
-	case RiskHigh, RiskCritical:
-		return AutonomyHigh
-	default:
-		// Unknown / empty risk is treated as the highest tier (fail-closed).
-		return AutonomyHigh
-	}
-}

internal/sandbox/batch_test.go

@@ -1,26 +0,0 @@
-package sandbox
-
-import "testing"
-
-func TestRequiredAutonomyForBatch(t *testing.T) {
-	cases := []struct {
-		name  string
-		risks []Risk
-		want  Autonomy
-	}{
-		{"empty slice is low", nil, AutonomyLow},
-		{"single low", []Risk{{Level: RiskLow}}, AutonomyLow},
-		{"mixed low and medium is medium", []Risk{{Level: RiskLow}, {Level: RiskMedium}}, AutonomyMedium},
-		{"any high is high", []Risk{{Level: RiskLow}, {Level: RiskHigh}}, AutonomyHigh},
-		{"any critical is high", []Risk{{Level: RiskMedium}, {Level: RiskCritical}}, AutonomyHigh},
-		{"unknown risk fails closed to high", []Risk{{Level: RiskLevel("weird")}}, AutonomyHigh},
-		{"empty-level risk fails closed to high", []Risk{{Level: RiskLevel("")}}, AutonomyHigh},
-	}
-	for _, tc := range cases {
-		t.Run(tc.name, func(t *testing.T) {
-			if got := RequiredAutonomyForBatch(tc.risks); got != tc.want {
-				t.Fatalf("RequiredAutonomyForBatch(%v) = %q, want %q", tc.risks, got, tc.want)
-			}
-		})
-	}
-}

internal/sandbox/egress.go

@@ -2,6 +2,8 @@ package sandbox
 
 import (
 	"context"
+	"crypto/tls"
+	"crypto/x509"
 	"errors"
 	"fmt"
 	"io"
@@ -32,9 +34,21 @@ import (
 type egressProxy struct {
 	listener net.Listener
 	server   *http.Server
-	allowed  []string
-	denied   []string
-	log      func(string)
+	// socksListener serves the SOCKS5 CONNECT front-end on a SEPARATE loopback
+	// port. It runs the same authorize()/domainAllowed() gate as the HTTP proxy, so
+	// a client configured with ALL_PROXY=socks5://<socksAddr> is filtered
+	// identically. Bound to 127.0.0.1:0 like the HTTP listener.
+	socksListener net.Listener
+	allowed       []string
+	denied        []string
+	log           func(string)
+
+	// mitm, when non-nil (Policy.InspectTLS), makes handleConnect TERMINATE TLS
+	// using a locally generated CA so the decrypted Host can be re-checked against
+	// the allow/deny gate. mitmTransport dials upstreams with full validation
+	// (system roots unless upstreamRoots is set for tests) — never InsecureSkipVerify.
+	mitm          *mitmCA
+	mitmTransport *http.Transport
 
 	// domainPrompt, when set, is asked to authorize an UNKNOWN host (one neither
 	// allowed nor explicitly denied) instead of failing closed. It must return
@@ -71,6 +85,14 @@ type egressOptions struct {
 	// fail-closed behavior.
 	DomainPrompt  func(ctx context.Context, host string, port int) (bool, error)
 	PromptTimeout time.Duration
+	// InspectTLS enables TLS termination (MITM) so the proxy can enforce the
+	// allow/deny gate on the decrypted request Host. Off => pure CONNECT
+	// passthrough (the default, byte-for-byte unchanged).
+	InspectTLS bool
+	// upstreamRoots overrides the root CAs used to validate UPSTREAM TLS. nil =
+	// system roots (production). Tests set it to an httptest server's pool so
+	// validation is exercised WITHOUT InsecureSkipVerify.
+	upstreamRoots *x509.CertPool
 }
 
 // errEmptyAllowlist is returned when a scoped-egress proxy is requested with no
@@ -89,9 +111,18 @@ func newEgressProxy(options egressOptions) (*egressProxy, error) {
 	if err != nil {
 		return nil, fmt.Errorf("bind scoped-egress proxy: %w", err)
 	}
+	// Bind the SOCKS5 front-end on its own loopback port. Fail closed: if it
+	// cannot bind, tear down the HTTP listener and return an error so the caller
+	// treats scoped egress as a full deny rather than starting a partial proxy.
+	socksListener, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		_ = listener.Close()
+		return nil, fmt.Errorf("bind scoped-egress socks proxy: %w", err)
+	}
 
 	proxy := &egressProxy{
 		listener:      listener,
+		socksListener: socksListener,
 		allowed:       allowed,
 		denied:        denied,
 		log:           options.Log,
@@ -100,6 +131,31 @@ func newEgressProxy(options egressOptions) (*egressProxy, error) {
 		decisionCache: map[string]bool{},
 		inflight:      map[string]chan struct{}{},
 	}
+	if options.InspectTLS {
+		ca, err := newMITMCA()
+		if err != nil {
+			_ = listener.Close()
+			_ = socksListener.Close()
+			return nil, err
+		}
+		proxy.mitm = ca
+		dialer := &net.Dialer{Timeout: 30 * time.Second, KeepAlive: 30 * time.Second}
+		proxy.mitmTransport = &http.Transport{
+			// Bound the TCP connect phase so a blackholed upstream cannot park a
+			// handler+goroutine for the OS default (minutes) — mirrors the
+			// net.DialTimeout the passthrough CONNECT path uses.
+			DialContext: dialer.DialContext,
+			// Upstream TLS is validated normally; upstreamRoots is nil in production
+			// (system roots) and set only by tests. NEVER InsecureSkipVerify.
+			TLSClientConfig: &tls.Config{MinVersion: tls.VersionTLS12, RootCAs: options.upstreamRoots},
+			// Keep upstream on HTTP/1.1: the decrypted client side is HTTP/1.1, so an
+			// HTTP/2 upstream response would serialize an invalid status line back to
+			// the client. (handleConnectMITM also normalizes the response proto.)
+			ForceAttemptHTTP2:     false,
+			TLSHandshakeTimeout:   30 * time.Second,
+			ResponseHeaderTimeout: 30 * time.Second,
+		}
+	}
 	proxy.server = &http.Server{
 		Handler:           http.HandlerFunc(proxy.handle),
 		ReadHeaderTimeout: 30 * time.Second,
@@ -108,10 +164,11 @@ func newEgressProxy(options egressOptions) (*egressProxy, error) {
 		// Serve returns ErrServerClosed on Close; nothing else to do.
 		_ = proxy.server.Serve(listener)
 	}()
+	go proxy.serveSocks(socksListener)
 	return proxy, nil
 }
 
-// Addr returns the loopback host:port the proxy is listening on.
+// Addr returns the loopback host:port the HTTP proxy is listening on.
 func (proxy *egressProxy) Addr() string {
 	if proxy == nil || proxy.listener == nil {
 		return ""
@@ -119,6 +176,25 @@ func (proxy *egressProxy) Addr() string {
 	return proxy.listener.Addr().String()
 }
 
+// SocksAddr returns the loopback host:port the SOCKS5 front-end is listening on,
+// or "" when no SOCKS listener is active.
+func (proxy *egressProxy) SocksAddr() string {
+	if proxy == nil || proxy.socksListener == nil {
+		return ""
+	}
+	return proxy.socksListener.Addr().String()
+}
+
+// CACertPEM returns the MITM CA's public certificate (PEM) when TLS inspection
+// is active, or nil for a plain passthrough proxy. The runner writes this to a
+// file and surfaces it via ZERO_SANDBOX_CA_CERT.
+func (proxy *egressProxy) CACertPEM() []byte {
+	if proxy == nil || proxy.mitm == nil {
+		return nil
+	}
+	return proxy.mitm.caPEM()
+}
+
 // Close stops the proxy and releases its listener. It is safe to call multiple
 // times.
 func (proxy *egressProxy) Close() error {
@@ -130,12 +206,26 @@ func (proxy *egressProxy) Close() error {
 		if proxy.server != nil {
 			err = proxy.server.Close()
 		}
+		// Closing the SOCKS listener unblocks serveSocks's Accept loop so the
+		// goroutine returns. Report its error only if the HTTP close succeeded.
+		if proxy.socksListener != nil {
+			if cerr := proxy.socksListener.Close(); cerr != nil && err == nil {
+				err = cerr
+			}
+		}
+		if proxy.mitmTransport != nil {
+			proxy.mitmTransport.CloseIdleConnections()
+		}
 	})
 	return err
 }
 
 func (proxy *egressProxy) handle(w http.ResponseWriter, r *http.Request) {
 	if r.Method == http.MethodConnect {
+		if proxy.mitm != nil {
+			proxy.handleConnectMITM(w, r)
+			return
+		}
 		proxy.handleConnect(w, r)
 		return
 	}

internal/sandbox/egress_socks.go

@@ -0,0 +1,156 @@
+package sandbox
+
+import (
+	"encoding/binary"
+	"io"
+	"net"
+	"strconv"
+	"time"
+)
+
+// SOCKS5 wire constants (RFC 1928). The proxy speaks only the no-auth CONNECT
+// subset: enough for a sandboxed process configured with ALL_PROXY=socks5://...
+// to tunnel TCP through the same allow/deny gate the HTTP proxy enforces.
+const (
+	socksVersion5          = 0x05
+	socksNoAuth            = 0x00
+	socksNoAcceptable      = 0xFF
+	socksCmdConnect        = 0x01
+	socksAtypIPv4          = 0x01
+	socksAtypDomain        = 0x03
+	socksAtypIPv6          = 0x04
+	socksRepSuccess        = 0x00
+	socksRepGeneralFailure = 0x01
+	socksRepNotAllowed     = 0x02
+	socksRepCmdUnsupported = 0x07
+)
+
+// socksHandshakeTimeout bounds the greeting+request exchange so a stalled or
+// malformed client cannot pin a goroutine; it is cleared before the relay so an
+// established tunnel is not torn down.
+const socksHandshakeTimeout = 30 * time.Second
+
+// serveSocks accepts SOCKS5 connections until the listener is closed.
+func (proxy *egressProxy) serveSocks(listener net.Listener) {
+	for {
+		conn, err := listener.Accept()
+		if err != nil {
+			return // listener closed
+		}
+		go proxy.handleSocksConn(conn)
+	}
+}
+
+// handleSocksConn runs one SOCKS5 CONNECT through the SAME authorize() gate as
+// the HTTP proxy: an allowed (host, port) is dialed and relayed; a denied target
+// gets a SOCKS "connection not allowed" reply and is never dialed. Fail-closed:
+// any protocol/read error just closes the connection.
+func (proxy *egressProxy) handleSocksConn(conn net.Conn) {
+	defer conn.Close()
+	_ = conn.SetDeadline(time.Now().Add(socksHandshakeTimeout))
+
+	host, port, ok := socksNegotiate(conn)
+	if !ok {
+		return
+	}
+	targetLabel := net.JoinHostPort(host, strconv.Itoa(port))
+	if !proxy.authorize(host, port) {
+		proxy.logDecision("deny", "SOCKS", targetLabel)
+		_ = socksReply(conn, socksRepNotAllowed)
+		return
+	}
+	upstream, err := net.DialTimeout("tcp", targetLabel, 30*time.Second)
+	if err != nil {
+		proxy.logDecision("deny", "SOCKS", targetLabel)
+		_ = socksReply(conn, socksRepGeneralFailure)
+		return
+	}
+	defer upstream.Close()
+	if err := socksReply(conn, socksRepSuccess); err != nil {
+		return
+	}
+	// Clear the handshake deadline; the tunnel is long-lived.
+	_ = conn.SetDeadline(time.Time{})
+	proxy.logDecision("allow", "SOCKS", targetLabel)
+	relay(conn, upstream)
+}
+
+// socksNegotiate performs the no-auth greeting and a CONNECT request, returning
+// the requested host and port. ok=false means negotiation failed (an error reply
+// has already been written where the protocol calls for one) and the caller
+// should close the connection.
+func socksNegotiate(conn net.Conn) (host string, port int, ok bool) {
+	greeting := make([]byte, 2) // VER, NMETHODS
+	if _, err := io.ReadFull(conn, greeting); err != nil || greeting[0] != socksVersion5 {
+		return "", 0, false
+	}
+	methods := make([]byte, int(greeting[1]))
+	if _, err := io.ReadFull(conn, methods); err != nil {
+		return "", 0, false
+	}
+	// RFC 1928: select a method the client actually offered. We only support
+	// no-auth; if the client did not offer it, reply "no acceptable methods"
+	// (0xFF) and abort rather than forcing a method it never advertised.
+	supportsNoAuth := false
+	for _, method := range methods {
+		if method == socksNoAuth {
+			supportsNoAuth = true
+			break
+		}
+	}
+	if !supportsNoAuth {
+		_, _ = conn.Write([]byte{socksVersion5, socksNoAcceptable})
+		return "", 0, false
+	}
+	if _, err := conn.Write([]byte{socksVersion5, socksNoAuth}); err != nil {
+		return "", 0, false
+	}
+
+	header := make([]byte, 4) // VER, CMD, RSV, ATYP
+	if _, err := io.ReadFull(conn, header); err != nil || header[0] != socksVersion5 {
+		return "", 0, false
+	}
+	if header[1] != socksCmdConnect {
+		_ = socksReply(conn, socksRepCmdUnsupported)
+		return "", 0, false
+	}
+	switch header[3] {
+	case socksAtypIPv4:
+		addr := make([]byte, net.IPv4len)
+		if _, err := io.ReadFull(conn, addr); err != nil {
+			return "", 0, false
+		}
+		host = net.IP(addr).String()
+	case socksAtypIPv6:
+		addr := make([]byte, net.IPv6len)
+		if _, err := io.ReadFull(conn, addr); err != nil {
+			return "", 0, false
+		}
+		host = net.IP(addr).String()
+	case socksAtypDomain:
+		length := make([]byte, 1)
+		if _, err := io.ReadFull(conn, length); err != nil {
+			return "", 0, false
+		}
+		name := make([]byte, int(length[0]))
+		if _, err := io.ReadFull(conn, name); err != nil {
+			return "", 0, false
+		}
+		host = string(name)
+	default:
+		_ = socksReply(conn, socksRepGeneralFailure)
+		return "", 0, false
+	}
+	portBytes := make([]byte, 2)
+	if _, err := io.ReadFull(conn, portBytes); err != nil {
+		return "", 0, false
+	}
+	return host, int(binary.BigEndian.Uint16(portBytes)), true
+}
+
+// socksReply writes a minimal SOCKS5 reply with the given status and a zero IPv4
+// bound address (clients ignore BND.ADDR/BND.PORT for CONNECT).
+func socksReply(conn net.Conn, status byte) error {
+	_, err := conn.Write([]byte{socksVersion5, status, 0x00, socksAtypIPv4, 0, 0, 0, 0, 0, 0})
+	return err
+}

internal/sandbox/egress_socks_test.go

@@ -0,0 +1,187 @@
+package sandbox
+
+import (
+	"io"
+	"net"
+	"strconv"
+	"testing"
+	"time"
+)
+
+// socks5Connect performs a SOCKS5 no-auth greeting and a CONNECT request for the
+// given ATYP/address/port through proxyAddr, returning the open connection and
+// the server's REP byte. The caller closes the connection.
+func socks5Connect(t *testing.T, proxyAddr string, atyp byte, addr []byte, port int) (net.Conn, byte) {
+	t.Helper()
+	conn, err := net.DialTimeout("tcp", proxyAddr, 2*time.Second)
+	if err != nil {
+		t.Fatalf("dial socks proxy: %v", err)
+	}
+	_ = conn.SetDeadline(time.Now().Add(2 * time.Second))
+
+	// Greeting: VER=5, NMETHODS=1, METHOD=0 (no auth).
+	if _, err := conn.Write([]byte{socksVersion5, 0x01, socksNoAuth}); err != nil {
+		t.Fatalf("write greeting: %v", err)
+	}
+	method := make([]byte, 2)
+	if _, err := io.ReadFull(conn, method); err != nil {
+		t.Fatalf("read method selection: %v", err)
+	}
+	if method[0] != socksVersion5 || method[1] != socksNoAuth {
+		t.Fatalf("method selection = %v, want [5 0]", method)
+	}
+
+	// Request: VER=5, CMD=CONNECT, RSV=0, ATYP, ADDR, PORT(2, big-endian).
+	request := append([]byte{socksVersion5, socksCmdConnect, 0x00, atyp}, addr...)
+	request = append(request, byte(port>>8), byte(port&0xff))
+	if _, err := conn.Write(request); err != nil {
+		t.Fatalf("write request: %v", err)
+	}
+	reply := make([]byte, 10) // VER, REP, RSV, ATYP(IPv4), BND.ADDR(4), BND.PORT(2)
+	if _, err := io.ReadFull(conn, reply); err != nil {
+		t.Fatalf("read reply: %v", err)
+	}
+	return conn, reply[1]
+}
+
+func TestEgressProxySocksAddrBindsLoopback(t *testing.T) {
+	proxy, err := newEgressProxy(egressOptions{Allowed: []string{"github.com"}})
+	if err != nil {
+		t.Fatalf("newEgressProxy: %v", err)
+	}
+	defer proxy.Close()
+
+	if proxy.SocksAddr() == "" {
+		t.Fatal("SocksAddr must be non-empty while the proxy is running")
+	}
+	if proxy.SocksAddr() == proxy.Addr() {
+		t.Fatalf("SOCKS and HTTP listeners must be distinct ports, both %q", proxy.Addr())
+	}
+	host, _, err := net.SplitHostPort(proxy.SocksAddr())
+	if err != nil {
+		t.Fatalf("SplitHostPort(%q): %v", proxy.SocksAddr(), err)
+	}
+	if host != "127.0.0.1" {
+		t.Fatalf("SOCKS bound to %q, want loopback 127.0.0.1", host)
+	}
+}
+
+// TestEgressProxySocksRejectsUnsupportedAuth verifies the proxy honors SOCKS5
+// method negotiation: a client offering only methods the proxy does not support
+// (here user/pass 0x02, no no-auth) gets the "no acceptable methods" reply (0xFF)
+// rather than the proxy forcing a method the client never advertised.
+func TestEgressProxySocksRejectsUnsupportedAuth(t *testing.T) {
+	proxy, err := newEgressProxy(egressOptions{Allowed: []string{"github.com"}})
+	if err != nil {
+		t.Fatalf("newEgressProxy: %v", err)
+	}
+	defer proxy.Close()
+
+	conn, err := net.DialTimeout("tcp", proxy.SocksAddr(), 2*time.Second)
+	if err != nil {
+		t.Fatalf("dial socks proxy: %v", err)
+	}
+	defer conn.Close()
+	_ = conn.SetDeadline(time.Now().Add(2 * time.Second))
+
+	// Greeting: VER=5, NMETHODS=1, METHOD=0x02 (user/pass) — no-auth NOT offered.
+	if _, err := conn.Write([]byte{socksVersion5, 0x01, 0x02}); err != nil {
+		t.Fatalf("write greeting: %v", err)
+	}
+	resp := make([]byte, 2)
+	if _, err := io.ReadFull(conn, resp); err != nil {
+		t.Fatalf("read method selection: %v", err)
+	}
+	if resp[0] != socksVersion5 || resp[1] != socksNoAcceptable {
+		t.Fatalf("method selection = %v, want [5 0xFF] (no acceptable methods)", resp)
+	}
+}
+
+// TestEgressProxySocksConnectAllowed verifies an allowlisted SOCKS5 CONNECT is
+// authorized through the SAME gate, dialed, and relayed to a fake upstream.
+func TestEgressProxySocksConnectAllowed(t *testing.T) {
+	upstream, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		t.Fatalf("listen upstream: %v", err)
+	}
+	defer upstream.Close()
+	go func() {
+		c, aerr := upstream.Accept()
+		if aerr != nil {
+			return
+		}
+		defer c.Close()
+		buf := make([]byte, 4)
+		if _, rerr := io.ReadFull(c, buf); rerr != nil {
+			return
+		}
+		_, _ = c.Write([]byte("pong"))
+	}()
+
+	_, portStr, _ := net.SplitHostPort(upstream.Addr().String())
+	port, _ := strconv.Atoi(portStr)
+
+	// Allow the loopback IP so the IPv4 CONNECT is authorized.
+	proxy, err := newEgressProxy(egressOptions{Allowed: []string{"127.0.0.1"}})
+	if err != nil {
+		t.Fatalf("newEgressProxy: %v", err)
+	}
+	defer proxy.Close()
+
+	conn, rep := socks5Connect(t, proxy.SocksAddr(), socksAtypIPv4, net.IPv4(127, 0, 0, 1).To4(), port)
+	defer conn.Close()
+	if rep != socksRepSuccess {
+		t.Fatalf("SOCKS CONNECT to allowed host REP = %#x, want success (0)", rep)
+	}
+	if _, err := conn.Write([]byte("ping")); err != nil {
+		t.Fatalf("write through tunnel: %v", err)
+	}
+	got := make([]byte, 4)
+	if _, err := io.ReadFull(conn, got); err != nil {
+		t.Fatalf("read through tunnel: %v", err)
+	}
+	if string(got) != "pong" {
+		t.Fatalf("tunneled response = %q, want pong", string(got))
+	}
+}
+
+// TestEgressProxySocksConnectDenied verifies a non-allowlisted SOCKS5 CONNECT
+// (domain form) is refused with the "connection not allowed" reply.
+func TestEgressProxySocksConnectDenied(t *testing.T) {
+	proxy, err := newEgressProxy(egressOptions{Allowed: []string{"allowed.example"}})
+	if err != nil {
+		t.Fatalf("newEgressProxy: %v", err)
+	}
+	defer proxy.Close()
+
+	domain := "denied.example"
+	addr := append([]byte{byte(len(domain))}, []byte(domain)...)
+	conn, rep := socks5Connect(t, proxy.SocksAddr(), socksAtypDomain, addr, 443)
+	conn.Close()
+	if rep != socksRepNotAllowed {
+		t.Fatalf("SOCKS CONNECT to denied host REP = %#x, want not-allowed (2)", rep)
+	}
+}
+
+// TestEgressProxySocksDenyWinsOverAllow verifies the SOCKS gate honors
+// deny-wins: a denied subdomain of an allowed domain is refused (REP 2), proving
+// the SOCKS path runs the same authorize()/domainAllowed() logic as the HTTP
+// proxy rather than a parallel, weaker check.
+func TestEgressProxySocksDenyWinsOverAllow(t *testing.T) {
+	proxy, err := newEgressProxy(egressOptions{
+		Allowed: []string{"allowed.example"},
+		Denied:  []string{"blocked.allowed.example"},
+	})
+	if err != nil {
+		t.Fatalf("newEgressProxy: %v", err)
+	}
+	defer proxy.Close()
+
+	denied := "blocked.allowed.example"
+	addr := append([]byte{byte(len(denied))}, []byte(denied)...)
+	conn, rep := socks5Connect(t, proxy.SocksAddr(), socksAtypDomain, addr, 443)
+	conn.Close()
+	if rep != socksRepNotAllowed {
+		t.Fatalf("denied subdomain REP = %#x, want not-allowed (2) (deny must win)", rep)
+	}
+}

internal/sandbox/engine.go

@@ -65,6 +65,35 @@ func (engine *Engine) Scope() *Scope {
 	return engine.scope
 }
 
+// ReadExclusions returns the resolved DenyRead/AllowRead exclusion matcher for
+// this engine's policy, resolving each policy entry ONCE. The search tools build
+// it a single time per run and reuse it across the whole walk so the predicates
+// don't re-run Abs/EvalSymlinks per visited path. Returns nil for a nil engine
+// (the matcher's methods treat nil as "exclude nothing").
+func (engine *Engine) ReadExclusions() *ReadExclusions {
+	// A disabled policy enforces nothing, so it must not filter search results
+	// either (Evaluate already allows every request under ModeDisabled).
+	if engine == nil || engine.policy.Mode == ModeDisabled {
+		return nil
+	}
+	return &ReadExclusions{
+		workspaceRoot: engine.workspaceRoot,
+		denyRoots:     resolvePolicyPaths(engine.policy.DenyRead),
+		allowRoots:    resolvePolicyPaths(engine.policy.AllowRead),
+	}
+}
+
+// ReadExclusionGlobs returns the ripgrep-style --glob exclusion args for this
+// engine's policy + scope (see the package-level ReadExclusionGlobs). Empty when
+// DenyRead is unset or the engine has no scope.
+func (engine *Engine) ReadExclusionGlobs() []string {
+	// A disabled policy filters nothing (parity with ReadExclusions / Evaluate).
+	if engine == nil || engine.policy.Mode == ModeDisabled {
+		return nil
+	}
+	return ReadExclusionGlobs(engine.policy, engine.scope)
+}
+
 // effectiveNetworkMode is the single source of truth for the engine's active
 // network mode: it collapses an empty-allowlist scoped policy to deny, and ALSO
 // downgrades scoped to deny when the backend cannot actually route scoped egress
@@ -90,6 +119,13 @@ func (engine *Engine) effectiveNetworkMode(policy Policy) NetworkMode {
 // their pre-sandbox behaviour); deny blocks everything; scoped allows only hosts
 // in AllowedDomains minus DeniedDomains (an empty allowlist, or a backend that
 // can't enforce scoped egress, collapses to deny).
+//
+// By default the first-party, in-process tools that consult this gate are NOT
+// subject to the network policy (EnforceToolNetwork is off): that policy exists
+// to confine the sandboxed SHELL's egress, which these tools don't use, and they
+// retain their own SSRF/port/redirect safeguards. Set EnforceToolNetwork to also
+// hold them to the allow/scoped/deny policy. The sandboxed-shell egress decision
+// lives in Evaluate via effectiveNetworkMode and is unaffected by this flag.
 func (engine *Engine) NetworkHostAllowed(host string) (bool, NetworkMode) {
 	if engine == nil {
 		return true, NetworkAllow
@@ -98,6 +134,10 @@ func (engine *Engine) NetworkHostAllowed(host string) (bool, NetworkMode) {
 	if policy.Mode == ModeDisabled {
 		return true, NetworkAllow
 	}
+	if !policy.EnforceToolNetwork {
+		// First-party in-process tools are exempt unless the operator opts in.
+		return true, NetworkAllow
+	}
 	switch mode := engine.effectiveNetworkMode(policy); mode {
 	case NetworkAllow:
 		return true, mode
@@ -109,6 +149,19 @@ func (engine *Engine) NetworkHostAllowed(host string) (bool, NetworkMode) {
 	}
 }
 
+// toolNetworkExempt reports whether a request is exempt from the engine-level
+// network deny because it is a first-party, in-process network TOOL — one that
+// declares SideEffectNetwork (web_search / web_fetch) — and the operator has not
+// opted into EnforceToolNetwork. Such tools do not use the sandboxed shell's
+// egress; they keep their own SSRF/host safeguards, and NetworkHostAllowed (also
+// gated by EnforceToolNetwork) governs them at run time. A SHELL command merely
+// classified as network (SideEffectShell) is NOT exempt, so shell egress stays
+// blocked under deny. Mirrors the NetworkHostAllowed exemption so the Evaluate
+// gate and the per-tool gate never diverge.
+func (engine *Engine) toolNetworkExempt(policy Policy, request Request) bool {
+	return !policy.EnforceToolNetwork && request.SideEffect == SideEffectNetwork
+}
+
 // scopeFor returns the scope to validate request paths against. The engine's
 // shared scope applies only when the request targets the engine's own
 // workspace root; a per-request override root gets an ad-hoc single-root scope
@@ -221,11 +274,17 @@ func (engine *Engine) Evaluate(ctx context.Context, request Request) Decision {
 	if request.Permission == PermissionDeny {
 		return deny(request, risk, ViolationDeniedPermission, "", permissionReason(request), false)
 	}
-	if policy.EnforceWorkspace && request.WorkspaceRoot != "" {
-		for _, requested := range requestPaths(request) {
-			if violation := scope.validate(requested); violation != nil {
-				return deny(request, risk, violation.Code, violation.Path, violation.Reason, false)
-			}
+	// The fine-grained path lists (DenyRead/DenyWrite/AllowRead/AllowWrite) apply
+	// whenever the sandbox is enforcing, independent of EnforceWorkspace and even
+	// when there is no workspace root (absolute paths are still resolved and
+	// matched), so they are honored consistently with the grep/glob exclusion path
+	// and can't be bypassed by an engine built without a workspace root. The
+	// workspace boundary itself needs a root, so it is gated on having one. Mode is
+	// already known to be enforcing here (ModeDisabled returned above).
+	enforceWorkspace := policy.EnforceWorkspace && request.WorkspaceRoot != ""
+	for _, requested := range requestPaths(request) {
+		if violation := validatePathWithPolicy(scope, policy, request.SideEffect, enforceWorkspace, request.WorkspaceRoot, requested); violation != nil {
+			return deny(request, risk, violation.Code, violation.Path, violation.Reason, false)
 		}
 	}
 	// effectiveNetworkMode collapses an empty-allowlist scoped policy to deny, and
@@ -235,7 +294,7 @@ func (engine *Engine) Evaluate(ctx context.Context, request Request) Decision {
 	// than running with unrestricted networking. Shared with NetworkHostAllowed so
 	// the per-tool gate can't diverge from this decision. Allow is unchanged.
 	netMode := engine.effectiveNetworkMode(policy)
-	if netMode == NetworkDeny && HasRiskCategory(risk, "network") {
+	if netMode == NetworkDeny && HasRiskCategory(risk, "network") && !engine.toolNetworkExempt(policy, request) {
 		return deny(request, risk, ViolationNetwork, "", "network access is blocked by sandbox policy", false)
 	}
 	if policy.DenyDestructiveShell && HasRiskCategory(risk, "destructive") {

internal/sandbox/engine_test.go

@@ -9,6 +9,53 @@ import (
 	"time"
 )
 
+// TestEvaluateExemptsNetworkToolsFromDenyByDefault reproduces the web_search
+// hard-deny: under NetworkDeny, the engine-level network gate must NOT deny a
+// first-party network TOOL (SideEffect=network) by default, so it follows the
+// normal permission flow instead of being blocked outright. EnforceToolNetwork
+// restores the strict deny, and a network SHELL command stays denied either way.
+func TestEvaluateExemptsNetworkToolsFromDenyByDefault(t *testing.T) {
+	base := Policy{Mode: ModeEnforce, Network: NetworkDeny, MaxAutonomy: AutonomyHigh}
+
+	// Default (EnforceToolNetwork off): web_search is NOT network-denied.
+	engine := NewEngine(EngineOptions{Policy: base})
+	d := engine.Evaluate(context.Background(), Request{
+		ToolName: "web_search", SideEffect: SideEffectNetwork, Permission: PermissionPrompt,
+	})
+	if d.Action == ActionDeny {
+		t.Fatalf("web_search must not be network-denied by default, got deny: %s", d.ErrorString())
+	}
+
+	// Granted permission → it actually runs (ActionAllow), not blocked.
+	allowed := engine.Evaluate(context.Background(), Request{
+		ToolName: "web_search", SideEffect: SideEffectNetwork, Permission: PermissionPrompt, PermissionGranted: true,
+	})
+	if allowed.Action != ActionAllow {
+		t.Fatalf("granted web_search must be allowed under deny by default, got %q (%s)", allowed.Action, allowed.ErrorString())
+	}
+
+	// EnforceToolNetwork on → strict: the network tool is denied at the gate.
+	strict := base
+	strict.EnforceToolNetwork = true
+	strictEngine := NewEngine(EngineOptions{Policy: strict})
+	ds := strictEngine.Evaluate(context.Background(), Request{
+		ToolName: "web_search", SideEffect: SideEffectNetwork, Permission: PermissionPrompt, PermissionGranted: true,
+	})
+	if ds.Action != ActionDeny || ds.Violation == nil || ds.Violation.Code != ViolationNetwork {
+		t.Fatalf("with EnforceToolNetwork, web_search must be network-denied, got %q (%+v)", ds.Action, ds.Violation)
+	}
+
+	// A network SHELL command (SideEffect=shell) stays denied under deny — shell
+	// egress is never exempt, even with EnforceToolNetwork off.
+	shell := engine.Evaluate(context.Background(), Request{
+		ToolName: "bash", SideEffect: SideEffectShell, PermissionGranted: true,
+		Args: map[string]any{"command": "curl https://evil.test"},
+	})
+	if shell.Action != ActionDeny || shell.Violation == nil || shell.Violation.Code != ViolationNetwork {
+		t.Fatalf("a network shell command must stay denied under deny, got %q (%+v)", shell.Action, shell.Violation)
+	}
+}
+
 func TestEngineEvaluatesReadPromptAndPersistentDecisions(t *testing.T) {
 	root := t.TempDir()
 	store, err := NewGrantStore(StoreOptions{
@@ -295,7 +342,11 @@ func TestEngineClassifiesNetworkAndDestructiveShellCommands(t *testing.T) {
 }
 
 func TestEngineDeniesNetworkSideEffectWhenPolicyBlocksNetwork(t *testing.T) {
-	engine := NewEngine(EngineOptions{WorkspaceRoot: t.TempDir(), Policy: DefaultPolicy()})
+	// EnforceToolNetwork opts the in-process network tools into the policy; by
+	// default they are exempt (see TestEvaluateExemptsNetworkToolsFromDenyByDefault).
+	policy := DefaultPolicy()
+	policy.EnforceToolNetwork = true
+	engine := NewEngine(EngineOptions{WorkspaceRoot: t.TempDir(), Policy: policy})
 
 	decision := engine.Evaluate(context.Background(), Request{
 		ToolName:       "web_fetch",
@@ -736,7 +787,11 @@ func TestEngineNetworkHostAllowed(t *testing.T) {
 	}
 	for _, tc := range cases {
 		t.Run(tc.name, func(t *testing.T) {
-			engine := NewEngine(EngineOptions{Policy: tc.policy, Backend: tc.backend})
+			// This table exercises the ENFORCED tool-network gate; the default
+			// (exempt) posture is covered by TestEngineNetworkHostAllowedToolNetworkDefaultExempt.
+			policy := tc.policy
+			policy.EnforceToolNetwork = true
+			engine := NewEngine(EngineOptions{Policy: policy, Backend: tc.backend})
 			allowed, mode := engine.NetworkHostAllowed(tc.host)
 			if allowed != tc.allowed || mode != tc.mode {
 				t.Fatalf("NetworkHostAllowed(%q) = (%v, %q), want (%v, %q)", tc.host, allowed, mode, tc.allowed, tc.mode)
@@ -749,3 +804,21 @@ func TestEngineNetworkHostAllowed(t *testing.T) {
 		t.Fatalf("nil engine NetworkHostAllowed = (%v, %q), want (true, allow)", allowed, mode)
 	}
 }
+
+// TestEngineNetworkHostAllowedToolNetworkDefaultExempt verifies the default
+// posture: with EnforceToolNetwork off, the in-process tool gate is exempt from
+// the network policy, so even a deny / scoped-unlisted host is allowed (the
+// sandboxed-shell egress decision is separate and unaffected).
+func TestEngineNetworkHostAllowedToolNetworkDefaultExempt(t *testing.T) {
+	egressBackend := Backend{Name: BackendSandboxExec, Available: true, Executable: "/usr/bin/sandbox-exec", ScopedEgress: true}
+	policies := []Policy{
+		{Mode: ModeEnforce, Network: NetworkDeny},
+		{Mode: ModeEnforce, Network: NetworkScoped, AllowedDomains: []string{"allowed.test"}},
+	}
+	for _, policy := range policies {
+		engine := NewEngine(EngineOptions{Policy: policy, Backend: egressBackend})
+		if allowed, _ := engine.NetworkHostAllowed("evil.test"); !allowed {
+			t.Fatalf("default (EnforceToolNetwork off) must allow the tool host under network=%q", policy.Network)
+		}
+	}
+}

internal/sandbox/mitm.go

@@ -0,0 +1,261 @@
+package sandbox
+
+import (
+	"bufio"
+	"crypto/ecdsa"
+	"crypto/elliptic"
+	"crypto/rand"
+	"crypto/tls"
+	"crypto/x509"
+	"crypto/x509/pkix"
+	"encoding/pem"
+	"fmt"
+	"io"
+	"math/big"
+	"net"
+	"net/http"
+	"os"
+	"strings"
+	"sync"
+	"time"
+)
+
+// MITM TLS interception (opt-in via Policy.InspectTLS). Mirrors
+// reference-sandbox-wsl-mitm-code-agent-js/sandbox-manager.js mitmCA/getMitmCA
+// and network/proxy.js. The proxy terminates TLS toward the SANDBOXED CLIENT
+// using an ephemeral, locally generated CA so it can read the decrypted request
+// and apply the SAME allow/deny gate on the real Host — visibility, not new
+// authority. The UPSTREAM connection is always validated against real roots
+// (never InsecureSkipVerify).
+
+// mitmCA is an ephemeral certificate authority that mints per-host leaf certs.
+// It lives only for the proxy's lifetime; only its PUBLIC cert is ever exposed
+// (so the sandboxed client can trust the minted leaves).
+type mitmCA struct {
+	cert    *x509.Certificate
+	key     *ecdsa.PrivateKey
+	certDER []byte
+
+	mu     sync.Mutex
+	leaves map[string]*tls.Certificate
+}
+
+// maxMITMLeafCacheEntries caps the per-host leaf cache so an attacker cannot grow
+// it without bound. Combined with minting for the AUTHORIZED host (not arbitrary
+// client SNI), this keeps the cache size to roughly the allowlist size.
+const maxMITMLeafCacheEntries = 512
+
+func randomSerial() (*big.Int, error) {
+	serial, err := rand.Int(rand.Reader, new(big.Int).Lsh(big.NewInt(1), 128))
+	if err != nil {
+		return nil, fmt.Errorf("mitm: serial: %w", err)
+	}
+	return serial, nil
+}
+
+// newMITMCA generates a fresh ECDSA CA. The key never leaves the process.
+func newMITMCA() (*mitmCA, error) {
+	key, err := ecdsa.GenerateKey(elliptic.P256(), rand.Reader)
+	if err != nil {
+		return nil, fmt.Errorf("mitm: generate CA key: %w", err)
+	}
+	serial, err := randomSerial()
+	if err != nil {
+		return nil, err
+	}
+	now := time.Now()
+	tmpl := &x509.Certificate{
+		SerialNumber:          serial,
+		Subject:               pkix.Name{CommonName: "zero sandbox MITM CA", Organization: []string{"zero-sandbox"}},
+		NotBefore:             now.Add(-time.Hour),
+		NotAfter:              now.Add(24 * time.Hour),
+		KeyUsage:              x509.KeyUsageCertSign | x509.KeyUsageCRLSign | x509.KeyUsageDigitalSignature,
+		BasicConstraintsValid: true,
+		IsCA:                  true,
+		MaxPathLenZero:        true,
+	}
+	der, err := x509.CreateCertificate(rand.Reader, tmpl, tmpl, &key.PublicKey, key)
+	if err != nil {
+		return nil, fmt.Errorf("mitm: create CA cert: %w", err)
+	}
+	cert, err := x509.ParseCertificate(der)
+	if err != nil {
+		return nil, fmt.Errorf("mitm: parse CA cert: %w", err)
+	}
+	return &mitmCA{cert: cert, key: key, certDER: der, leaves: map[string]*tls.Certificate{}}, nil
+}
+
+// caPEM returns the CA's PUBLIC certificate in PEM (safe to write/expose).
+func (ca *mitmCA) caPEM() []byte {
+	return pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: ca.certDER})
+}
+
+// leafFor mints (and caches) a leaf certificate for host, signed by the CA, with
+// the CA appended so the chain verifies against the CA alone.
+func (ca *mitmCA) leafFor(host string) (*tls.Certificate, error) {
+	host = hostnameOnly(host)
+	if host == "" {
+		return nil, fmt.Errorf("mitm: empty host for leaf cert")
+	}
+	ca.mu.Lock()
+	defer ca.mu.Unlock()
+	if leaf, ok := ca.leaves[host]; ok {
+		return leaf, nil
+	}
+	// Bound the cache: past the cap, mint without caching so memory stays bounded
+	// (a small per-request cost rather than unbounded growth).
+	cacheable := len(ca.leaves) < maxMITMLeafCacheEntries
+	leafKey, err := ecdsa.GenerateKey(elliptic.P256(), rand.Reader)
+	if err != nil {
+		return nil, err
+	}
+	serial, err := randomSerial()
+	if err != nil {
+		return nil, err
+	}
+	now := time.Now()
+	tmpl := &x509.Certificate{
+		SerialNumber: serial,
+		Subject:      pkix.Name{CommonName: host},
+		NotBefore:    now.Add(-time.Hour),
+		NotAfter:     now.Add(24 * time.Hour),
+		KeyUsage:     x509.KeyUsageDigitalSignature,
+		ExtKeyUsage:  []x509.ExtKeyUsage{x509.ExtKeyUsageServerAuth},
+	}
+	if ip := net.ParseIP(host); ip != nil {
+		tmpl.IPAddresses = []net.IP{ip}
+	} else {
+		tmpl.DNSNames = []string{host}
+	}
+	der, err := x509.CreateCertificate(rand.Reader, tmpl, ca.cert, &leafKey.PublicKey, ca.key)
+	if err != nil {
+		return nil, fmt.Errorf("mitm: mint leaf for %s: %w", host, err)
+	}
+	leaf := &tls.Certificate{Certificate: [][]byte{der, ca.certDER}, PrivateKey: leafKey}
+	if cacheable {
+		ca.leaves[host] = leaf
+	}
+	return leaf, nil
+}
+
+// writeMITMCAFile writes the CA's PUBLIC cert PEM to an owner-only temp file and
+// returns its path, so the runner can surface it via ZERO_SANDBOX_CA_CERT.
+func writeMITMCAFile(pemBytes []byte) (string, error) {
+	f, err := os.CreateTemp("", "zero-sandbox-ca-*.pem")
+	if err != nil {
+		return "", fmt.Errorf("mitm: create CA file: %w", err)
+	}
+	if err := f.Chmod(0o600); err != nil {
+		_ = f.Close()
+		_ = os.Remove(f.Name())
+		return "", err
+	}
+	if _, err := f.Write(pemBytes); err != nil {
+		_ = f.Close()
+		_ = os.Remove(f.Name())
+		return "", err
+	}
+	if err := f.Close(); err != nil {
+		_ = os.Remove(f.Name())
+		return "", err
+	}
+	return f.Name(), nil
+}
+
+// handleConnectMITM terminates TLS toward the client (minting a leaf for the
+// requested host), then for each decrypted request re-checks the real Host
+// through the SAME authorize gate and forwards it upstream with full system-root
+// validation. A denied host gets a 403 over the decrypted channel and is never
+// forwarded.
+func (proxy *egressProxy) handleConnectMITM(w http.ResponseWriter, r *http.Request) {
+	target := r.Host
+	if target == "" {
+		target = r.URL.Host
+	}
+	if !proxy.authorizeTarget(target, 443) {
+		proxy.logDecision("deny", "CONNECT", target)
+		http.Error(w, "scoped egress: host not allowed", http.StatusForbidden)
+		return
+	}
+	hijacker, ok := w.(http.Hijacker)
+	if !ok {
+		http.Error(w, "scoped egress: hijack unsupported", http.StatusInternalServerError)
+		return
+	}
+	clientConn, _, err := hijacker.Hijack()
+	if err != nil {
+		return
+	}
+	defer clientConn.Close()
+	if _, err := io.WriteString(clientConn, "HTTP/1.1 200 Connection Established\r\n\r\n"); err != nil {
+		return
+	}
+
+	host := hostnameOnly(target)
+	tlsConn := tls.Server(clientConn, &tls.Config{
+		MinVersion: tls.VersionTLS12,
+		GetCertificate: func(_ *tls.ClientHelloInfo) (*tls.Certificate, error) {
+			// Mint for the AUTHORIZED CONNECT host, never the arbitrary client SNI:
+			// honoring SNI would let an attacker churn distinct names to force
+			// unbounded key generation and cache growth.
+			return proxy.mitm.leafFor(host)
+		},
+	})
+	if err := tlsConn.Handshake(); err != nil {
+		return
+	}
+	defer tlsConn.Close()
+
+	reader := bufio.NewReader(tlsConn)
+	for {
+		req, err := http.ReadRequest(reader)
+		if err != nil {
+			return // client closed or EOF
+		}
+		reqHost := req.Host
+		if reqHost == "" {
+			reqHost = host
+		}
+		// MITM widens VISIBILITY, never authority: the decrypted Host passes the
+		// SAME full authorize() gate as the CONNECT check (deny-wins, empty allowlist
+		// => deny, and an interactive DomainPrompt is honored consistently).
+		if !proxy.authorizeTarget(reqHost, 443) {
+			proxy.logDecision("deny", req.Method, reqHost)
+			writeMITMResponse(tlsConn, http.StatusForbidden, "scoped egress: host not allowed")
+			return
+		}
+		req.URL.Scheme = "https"
+		req.URL.Host = normalizeConnectTarget(target)
+		req.RequestURI = ""
+		resp, err := proxy.mitmTransport.RoundTrip(req)
+		if err != nil {
+			proxy.logDecision("deny", req.Method, reqHost)
+			writeMITMResponse(tlsConn, http.StatusBadGateway, "scoped egress: upstream request failed")
+			return
+		}
+		proxy.logDecision("allow", req.Method, reqHost)
+		// Re-frame the response as HTTP/1.1 for the decrypted (HTTP/1.1) client
+		// connection, regardless of how the upstream replied.
+		resp.Proto, resp.ProtoMajor, resp.ProtoMinor = "HTTP/1.1", 1, 1
+		writeErr := resp.Write(tlsConn)
+		_ = resp.Body.Close()
+		if writeErr != nil || req.Close || resp.Close {
+			return
+		}
+	}
+}
+
+// writeMITMResponse writes a minimal plaintext HTTP/1.1 response over the
+// decrypted client connection.
+func writeMITMResponse(conn net.Conn, code int, msg string) {
+	body := msg + "\n"
+	resp := &http.Response{
+		StatusCode:    code,
+		ProtoMajor:    1,
+		ProtoMinor:    1,
+		Header:        http.Header{"Content-Type": []string{"text/plain; charset=utf-8"}},
+		Body:          io.NopCloser(strings.NewReader(body)),
+		ContentLength: int64(len(body)),
+	}
+	_ = resp.Write(conn)
+}

internal/sandbox/mitm_test.go

@@ -0,0 +1,158 @@
+package sandbox
+
+import (
+	"bufio"
+	"crypto/tls"
+	"crypto/x509"
+	"fmt"
+	"io"
+	"net"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+)
+
+func TestEgressProxyPassthroughNoCAByDefault(t *testing.T) {
+	proxy, err := newEgressProxy(egressOptions{Allowed: []string{"github.com"}})
+	if err != nil {
+		t.Fatalf("newEgressProxy: %v", err)
+	}
+	defer proxy.Close()
+	if proxy.mitm != nil {
+		t.Fatal("InspectTLS off must not create a MITM CA (pure passthrough)")
+	}
+	if proxy.CACertPEM() != nil {
+		t.Fatal("passthrough proxy must expose no CA cert")
+	}
+}
+
+func TestMITMCAMintsChainingLeaf(t *testing.T) {
+	ca, err := newMITMCA()
+	if err != nil {
+		t.Fatalf("newMITMCA: %v", err)
+	}
+	leaf, err := ca.leafFor("api.example.com")
+	if err != nil {
+		t.Fatalf("leafFor: %v", err)
+	}
+	leafCert, err := x509.ParseCertificate(leaf.Certificate[0])
+	if err != nil {
+		t.Fatalf("parse leaf: %v", err)
+	}
+	roots := x509.NewCertPool()
+	if !roots.AppendCertsFromPEM(ca.caPEM()) {
+		t.Fatal("append CA PEM")
+	}
+	if _, err := leafCert.Verify(x509.VerifyOptions{
+		Roots:     roots,
+		DNSName:   "api.example.com",
+		KeyUsages: []x509.ExtKeyUsage{x509.ExtKeyUsageServerAuth},
+	}); err != nil {
+		t.Fatalf("minted leaf must chain to the CA: %v", err)
+	}
+	// Per-host caching.
+	if leaf2, _ := ca.leafFor("api.example.com"); leaf2 != leaf {
+		t.Fatal("leafFor must cache one leaf per host")
+	}
+}
+
+// mitmConnect performs a CONNECT through the proxy then opens a TLS client that
+// trusts ONLY the MITM CA, returning the decrypted client connection.
+func mitmConnect(t *testing.T, proxyAddr, target string, caPEM []byte, serverName string) *tls.Conn {
+	t.Helper()
+	conn, err := net.DialTimeout("tcp", proxyAddr, 2*time.Second)
+	if err != nil {
+		t.Fatalf("dial proxy: %v", err)
+	}
+	// Bound the whole exchange (CONNECT read + TLS handshake + request/response) so
+	// a regression cannot hang the test indefinitely.
+	if err := conn.SetDeadline(time.Now().Add(10 * time.Second)); err != nil {
+		t.Fatalf("set deadline: %v", err)
+	}
+	fmt.Fprintf(conn, "CONNECT %s HTTP/1.1\r\nHost: %s\r\n\r\n", target, target)
+	br := bufio.NewReader(conn)
+	resp, err := http.ReadResponse(br, &http.Request{Method: "CONNECT"})
+	if err != nil {
+		t.Fatalf("read CONNECT response: %v", err)
+	}
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("CONNECT status = %d, want 200", resp.StatusCode)
+	}
+	if br.Buffered() != 0 {
+		t.Fatalf("proxy over-sent %d bytes after CONNECT 200", br.Buffered())
+	}
+	pool := x509.NewCertPool()
+	if !pool.AppendCertsFromPEM(caPEM) {
+		t.Fatal("append MITM CA PEM")
+	}
+	tlsConn := tls.Client(conn, &tls.Config{RootCAs: pool, ServerName: serverName})
+	if err := tlsConn.Handshake(); err != nil {
+		t.Fatalf("client TLS handshake trusting the MITM CA: %v", err)
+	}
+	return tlsConn
+}
+
+func TestEgressProxyMITMForwardsAllowlistedHost(t *testing.T) {
+	upstream := httptest.NewTLSServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		fmt.Fprint(w, "upstream-ok")
+	}))
+	defer upstream.Close()
+	host := mustHost(t, upstream.URL)
+	target := mustHostPort(t, upstream.URL)
+
+	// Trust the upstream's real cert explicitly (NOT InsecureSkipVerify): proves
+	// the MITM still validates the upstream chain.
+	upstreamRoots := x509.NewCertPool()
+	upstreamRoots.AddCert(upstream.Certificate())
+
+	proxy, err := newEgressProxy(egressOptions{Allowed: []string{host}, InspectTLS: true, upstreamRoots: upstreamRoots})
+	if err != nil {
+		t.Fatalf("newEgressProxy: %v", err)
+	}
+	defer proxy.Close()
+	if proxy.CACertPEM() == nil {
+		t.Fatal("InspectTLS must generate a CA")
+	}
+
+	tlsConn := mitmConnect(t, proxy.Addr(), target, proxy.CACertPEM(), host)
+	defer tlsConn.Close()
+	fmt.Fprintf(tlsConn, "GET / HTTP/1.1\r\nHost: %s\r\nConnection: close\r\n\r\n", host)
+	resp, err := http.ReadResponse(bufio.NewReader(tlsConn), &http.Request{Method: http.MethodGet})
+	if err != nil {
+		t.Fatalf("read decrypted response: %v", err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("allowlisted forwarded request status = %d, want 200", resp.StatusCode)
+	}
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		t.Fatalf("read decrypted body: %v", err)
+	}
+	if !strings.Contains(string(body), "upstream-ok") {
+		t.Fatalf("MITM did not forward the allowlisted request, got %q", string(body))
+	}
+}
+
+func TestEgressProxyMITMBlocksDeniedDecryptedHost(t *testing.T) {
+	// CONNECT to an allowed host, but send a Host header that is NOT allowlisted —
+	// the MITM re-checks the DECRYPTED Host and blocks it before forwarding.
+	proxy, err := newEgressProxy(egressOptions{Allowed: []string{"127.0.0.1"}, InspectTLS: true})
+	if err != nil {
+		t.Fatalf("newEgressProxy: %v", err)
+	}
+	defer proxy.Close()
+
+	tlsConn := mitmConnect(t, proxy.Addr(), "127.0.0.1:443", proxy.CACertPEM(), "127.0.0.1")
+	defer tlsConn.Close()
+	fmt.Fprint(tlsConn, "GET / HTTP/1.1\r\nHost: blocked.example\r\nConnection: close\r\n\r\n")
+	resp, err := http.ReadResponse(bufio.NewReader(tlsConn), nil)
+	if err != nil {
+		t.Fatalf("read decrypted response: %v", err)
+	}
+	if resp.StatusCode != http.StatusForbidden {
+		t.Fatalf("denied decrypted Host status = %d, want 403", resp.StatusCode)
+	}
+}

internal/sandbox/pathlists.go

@@ -0,0 +1,374 @@
+package sandbox
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+)
+
+// This file implements the fine-grained AllowRead/DenyRead/AllowWrite/DenyWrite
+// path lists (Policy fields). They layer ON TOP of the workspace + Scope guards
+// and never bypass them: AllowRead only re-includes inside a DenyRead carve-out,
+// AllowWrite is consulted only after the workspace/Scope guard already denied a
+// write, and every match is symlink-resolved so a symlink prefix cannot evade a
+// deny or sneak past an allow. All lists default empty, so an unconfigured policy
+// behaves exactly as before.
+
+// resolvePolicyPath home-expands, makes absolute, and symlink-resolves a single
+// policy path entry. ok is false for a blank entry or one that does not exist
+// (EvalSymlinks requires existence) so a bogus entry is dropped — a non-existent
+// deny protects nothing and a non-existent allow grants nothing.
+func resolvePolicyPath(entry string) (string, bool) {
+	trimmed := strings.TrimSpace(entry)
+	if trimmed == "" {
+		return "", false
+	}
+	if trimmed == "~" || strings.HasPrefix(trimmed, "~/") || strings.HasPrefix(trimmed, "~"+string(filepath.Separator)) {
+		home, err := os.UserHomeDir()
+		if err != nil {
+			return "", false
+		}
+		trimmed = filepath.Join(home, strings.TrimPrefix(strings.TrimPrefix(trimmed[1:], "/"), string(filepath.Separator)))
+	}
+	absolute, err := filepath.Abs(trimmed)
+	if err != nil {
+		return "", false
+	}
+	resolved, err := filepath.EvalSymlinks(absolute)
+	if err != nil {
+		return "", false
+	}
+	return resolved, true
+}
+
+// resolvePolicyPaths resolves and de-duplicates a list of policy path entries,
+// dropping blanks and non-existent entries. Files and directories are both kept
+// (a DenyRead/DenyWrite entry may target a single sensitive file).
+func resolvePolicyPaths(entries []string) []string {
+	if len(entries) == 0 {
+		return nil
+	}
+	seen := make(map[string]struct{}, len(entries))
+	out := make([]string, 0, len(entries))
+	for _, entry := range entries {
+		resolved, ok := resolvePolicyPath(entry)
+		if !ok {
+			continue
+		}
+		if _, dup := seen[resolved]; dup {
+			continue
+		}
+		seen[resolved] = struct{}{}
+		out = append(out, resolved)
+	}
+	return out
+}
+
+// resolveWriteRootPaths is resolvePolicyPaths restricted to existing directories
+// that are not the filesystem root — the only valid targets for an OS write bind
+// and for an AllowWrite grant root.
+func resolveWriteRootPaths(entries []string) []string {
+	resolved := resolvePolicyPaths(entries)
+	if len(resolved) == 0 {
+		return nil
+	}
+	out := make([]string, 0, len(resolved))
+	for _, path := range resolved {
+		info, err := os.Stat(path)
+		if err != nil || !info.IsDir() {
+			continue
+		}
+		if filepath.Dir(path) == path {
+			continue // refuse the filesystem root as a write root
+		}
+		out = append(out, path)
+	}
+	return out
+}
+
+// pathUnderPolicyRoot reports whether requestedPath lies within root. A relative
+// requestedPath is resolved against workspaceRoot; the portion of an absolute
+// path outside root is symlink-resolved (via NormalizePrefixForRoot) so a
+// symlink prefix cannot evade the match. root must be an already-resolved
+// absolute path.
+func pathUnderPolicyRoot(requestedPath, root, workspaceRoot string) bool {
+	if root == "" {
+		return false
+	}
+	abs := requestedPath
+	if !filepath.IsAbs(abs) {
+		if workspaceRoot == "" {
+			return false
+		}
+		abs = filepath.Join(workspaceRoot, abs)
+	}
+	normalized := NormalizePrefixForRoot(abs, root)
+	return pathWithinRoot(root, normalized)
+}
+
+// readDenied reports whether path is excluded by the DenyRead list with no
+// more-specific AllowRead re-inclusion. "More specific" means an AllowRead entry
+// nested inside the matched DenyRead entry — that subtree is read back in while
+// the rest of the denied tree stays blocked. It resolves the policy entries on
+// each call, so it suits a one-shot check (the Evaluate gate). A search walk that
+// checks many paths should use ReadExclusions, which resolves the entries once.
+func readDenied(policy Policy, workspaceRoot, path string) bool {
+	return readDeniedResolved(workspaceRoot, resolvePolicyPaths(policy.DenyRead), resolvePolicyPaths(policy.AllowRead), path)
+}
+
+// readDeniedResolved is readDenied operating on ALREADY-resolved deny/allow roots,
+// so a caller that resolved the policy entries once can reuse them across many
+// path checks instead of re-running Abs/EvalSymlinks per path.
+func readDeniedResolved(workspaceRoot string, denyRoots, allowRoots []string, path string) bool {
+	if len(denyRoots) == 0 {
+		return false
+	}
+	for _, deny := range denyRoots {
+		if !pathUnderPolicyRoot(path, deny, workspaceRoot) {
+			continue
+		}
+		reincluded := false
+		for _, allow := range allowRoots {
+			// The allow entry must sit inside the deny entry to be "more specific",
+			// and the path must fall under that allow entry.
+			if pathWithinRoot(deny, allow) && pathUnderPolicyRoot(path, allow, workspaceRoot) {
+				reincluded = true
+				break
+			}
+		}
+		if !reincluded {
+			return true
+		}
+	}
+	return false
+}
+
+// ReadExclusions holds the resolved DenyRead/AllowRead roots for a policy so a
+// search walk resolves each policy entry ONCE (Abs/EvalSymlinks) and reuses the
+// result across every visited path, rather than re-resolving per path. Build it
+// with Engine.ReadExclusions and reuse it for the whole grep/glob walk.
+type ReadExclusions struct {
+	workspaceRoot string
+	denyRoots     []string
+	allowRoots    []string
+}
+
+// Active reports whether any DenyRead root is configured. When false the
+// exclusions are a no-op and the search behaves exactly as before.
+func (rx *ReadExclusions) Active() bool {
+	return rx != nil && len(rx.denyRoots) > 0
+}
+
+// PathExcluded reports whether reading path is excluded by DenyRead, honoring a
+// more-specific AllowRead re-inclusion. It is the per-file predicate for a walk.
+func (rx *ReadExclusions) PathExcluded(path string) bool {
+	if !rx.Active() {
+		return false
+	}
+	return readDeniedResolved(rx.workspaceRoot, rx.denyRoots, rx.allowRoots, path)
+}
+
+// DirExcluded reports whether a directory subtree can be skipped wholesale during
+// a walk: it is read-denied AND contains no nested AllowRead root (descending is
+// required to reach a re-included subtree). When it returns false on a denied dir
+// because of a nested allow, PathExcluded still filters the denied siblings.
+func (rx *ReadExclusions) DirExcluded(path string) bool {
+	if !rx.Active() {
+		return false
+	}
+	if !readDeniedResolved(rx.workspaceRoot, rx.denyRoots, rx.allowRoots, path) {
+		return false
+	}
+	abs := path
+	if !filepath.IsAbs(abs) && rx.workspaceRoot != "" {
+		abs = filepath.Join(rx.workspaceRoot, path)
+	}
+	// allowRoots are symlink-resolved (resolvePolicyPaths), so resolve abs the
+	// same way before the prefix comparison — otherwise a dir reached THROUGH a
+	// symlink would fail to match a nested AllowRead root and be wrongly skipped,
+	// dropping a re-included subtree. Best-effort: keep abs if it can't resolve
+	// (e.g. a non-existent path), matching the deny check's tolerant behavior.
+	if resolved, err := filepath.EvalSymlinks(abs); err == nil {
+		abs = resolved
+	}
+	return !hasNestedAllowReadResolved(rx.allowRoots, abs)
+}
+
+// allowWriteScope builds an ad-hoc Scope from the resolved AllowWrite roots so a
+// write to an AllowWrite path is validated with the SAME symlink-traversal logic
+// the workspace Scope uses. Returns nil when there are no usable AllowWrite roots.
+func allowWriteScope(policy Policy) *Scope {
+	roots := resolveWriteRootPaths(policy.AllowWrite)
+	if len(roots) == 0 {
+		return nil
+	}
+	return &Scope{workspaceRoot: roots[0], extraRoots: roots[1:]}
+}
+
+// validateWritePath enforces the write precedence: DenyWrite wins, then (when
+// workspace enforcement is on) a workspace/Scope-writable path is allowed, then
+// an absolute path under an AllowWrite root is allowed, otherwise the base
+// workspace/Scope violation stands. The DenyWrite list applies regardless of
+// enforceWorkspace; the workspace boundary itself applies only when
+// enforceWorkspace. It never bypasses the symlink/out-of-workspace guards.
+func validateWritePath(scope *Scope, policy Policy, enforceWorkspace bool, workspaceRoot, path string) *pathViolation {
+	// DenyWrite wins regardless of workspace enforcement.
+	for _, deny := range resolvePolicyPaths(policy.DenyWrite) {
+		if pathUnderPolicyRoot(path, deny, workspaceRoot) {
+			return &pathViolation{
+				Code:   ViolationPolicyDenied,
+				Path:   path,
+				Reason: path + " is excluded by the sandbox DenyWrite policy",
+			}
+		}
+	}
+	if !enforceWorkspace {
+		// Workspace confinement is off: only the explicit DenyWrite list restricts.
+		return nil
+	}
+	base := scope.validate(path)
+	if base == nil {
+		return nil // writable under the workspace / Scope guard
+	}
+	// AllowWrite only extends ABSOLUTE paths: a relative path is inherently
+	// workspace-relative and already resolved by the base guard above.
+	if filepath.IsAbs(path) {
+		if allow := allowWriteScope(policy); allow != nil && allow.validate(path) == nil {
+			return nil
+		}
+	}
+	return base
+}
+
+// validatePathWithPolicy is the single entry point the engine uses to validate a
+// request path. The fine-grained read/write lists (DenyRead/DenyWrite, with
+// AllowRead/AllowWrite) apply whenever the sandbox is enforcing, INDEPENDENT of
+// enforceWorkspace, so they match the grep/glob path that also honors DenyRead
+// directly. The workspace boundary (scope.validate) applies only when
+// enforceWorkspace. Behavior is unchanged when the lists are empty.
+func validatePathWithPolicy(scope *Scope, policy Policy, sideEffect SideEffect, enforceWorkspace bool, workspaceRoot, path string) *pathViolation {
+	// A relative path cannot be anchored without a workspace root, so it cannot be
+	// checked against the (absolute) path lists or workspace boundary. Fail closed
+	// when there is anything to enforce; otherwise it is a no-op (unchanged from the
+	// pre-path-list behavior, where an empty workspace root skipped validation).
+	if workspaceRoot == "" && !filepath.IsAbs(path) {
+		if enforceWorkspace || policyHasPathLists(policy) {
+			return &pathViolation{
+				Code:   ViolationOutsideWorkspace,
+				Path:   path,
+				Reason: path + " cannot be validated without a workspace root",
+			}
+		}
+		return nil
+	}
+	switch sideEffect {
+	case SideEffectRead:
+		if readDenied(policy, workspaceRoot, path) {
+			return &pathViolation{
+				Code:   ViolationPolicyDenied,
+				Path:   path,
+				Reason: path + " is excluded by the sandbox DenyRead policy",
+			}
+		}
+		if enforceWorkspace {
+			return scope.validate(path)
+		}
+		return nil
+	case SideEffectWrite, SideEffectOutOfWorkspace:
+		return validateWritePath(scope, policy, enforceWorkspace, workspaceRoot, path)
+	default:
+		if enforceWorkspace {
+			return scope.validate(path)
+		}
+		return nil
+	}
+}
+
+// hasNestedAllowReadResolved reports whether any already-resolved AllowRead root
+// sits strictly inside dir (an already-resolved absolute path). When true, a
+// read-denied dir must still be descended during a walk so the re-included
+// subtree is reachable.
+func hasNestedAllowReadResolved(allowRoots []string, dir string) bool {
+	for _, allow := range allowRoots {
+		if allow != dir && pathWithinRoot(dir, allow) {
+			return true
+		}
+	}
+	return false
+}
+
+// workspaceRelGlob returns target as a clean, slash-separated path relative to
+// workspaceRoot, or ok=false when target is the root itself or lies outside it
+// (a workspace-rooted search never reaches such a path, so no glob is needed).
+func workspaceRelGlob(workspaceRoot, target string) (string, bool) {
+	rel, err := filepath.Rel(workspaceRoot, target)
+	if err != nil {
+		return "", false
+	}
+	if rel == "." || rel == ".." || strings.HasPrefix(rel, ".."+string(filepath.Separator)) || filepath.IsAbs(rel) {
+		return "", false
+	}
+	return filepath.ToSlash(rel), true
+}
+
+// ReadExclusionGlobs returns ripgrep-style --glob exclusion args for the policy's
+// DenyRead subtrees that fall inside the scope's workspace root, so a
+// ripgrep-based search never descends into a read-denied subtree. For each such
+// entry it emits `--glob`, `!<rel>` and `--glob`, `!<rel>/**`. Mirrors the
+// read-subtree exclusion globs used by comparable executor sandboxes.
+//
+// The projection is exclusions-only: a positive ripgrep glob would switch the
+// search into whitelist mode and restrict it to only matching files, so AllowRead
+// re-inclusion is NOT expressed here. The Go-native grep/glob tools honor
+// AllowRead precisely via the cached predicate (Engine.ReadExclusions); this
+// function is the coarser ripgrep-format export for an external rg-based
+// consumer. Empty when DenyRead is unset (the default), so search behavior is
+// unchanged.
+func ReadExclusionGlobs(policy Policy, scope *Scope) []string {
+	denyRoots := resolvePolicyPaths(policy.DenyRead)
+	if len(denyRoots) == 0 || scope == nil {
+		return nil
+	}
+	workspaceRoot := scope.WorkspaceRoot()
+	if workspaceRoot == "" {
+		return nil
+	}
+	var globs []string
+	for _, deny := range denyRoots {
+		rel, ok := workspaceRelGlob(workspaceRoot, deny)
+		if !ok {
+			continue
+		}
+		globs = append(globs, "--glob", "!"+rel, "--glob", "!"+rel+"/**")
+	}
+	return globs
+}
+
+// policyHasPathLists reports whether any fine-grained path list has an
+// ENFORCEABLE entry. It resolves the lists (matching how the rest of this file
+// normalizes them) rather than counting raw config, so a typo or non-existent
+// entry — which resolution drops — doesn't spuriously fail-close relative
+// requests when there is no workspace root.
+func policyHasPathLists(policy Policy) bool {
+	return len(resolvePolicyPaths(policy.DenyRead)) > 0 ||
+		len(resolvePolicyPaths(policy.AllowRead)) > 0 ||
+		len(resolvePolicyPaths(policy.DenyWrite)) > 0 ||
+		len(resolveWriteRootPaths(policy.AllowWrite)) > 0
+}
+
+// dedupeStrings returns xs with duplicates removed, preserving first-seen order.
+func dedupeStrings(xs []string) []string {
+	if len(xs) <= 1 {
+		return xs
+	}
+	seen := make(map[string]struct{}, len(xs))
+	out := make([]string, 0, len(xs))
+	for _, x := range xs {
+		if _, dup := seen[x]; dup {
+			continue
+		}
+		seen[x] = struct{}{}
+		out = append(out, x)
+	}
+	return out
+}

internal/sandbox/pathlists_test.go

@@ -0,0 +1,426 @@
+package sandbox
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// resolvedTempDir returns a symlink-resolved temp dir so paths built under it
+// compare equal to the EvalSymlinks-normalized forms the path lists produce
+// (macOS /var -> /private/var would otherwise diverge).
+func resolvedTempDir(t *testing.T) string {
+	t.Helper()
+	resolved, err := filepath.EvalSymlinks(t.TempDir())
+	if err != nil {
+		t.Fatalf("EvalSymlinks(tempdir): %v", err)
+	}
+	return resolved
+}
+
+func mkdir(t *testing.T, path string) string {
+	t.Helper()
+	if err := os.MkdirAll(path, 0o755); err != nil {
+		t.Fatalf("MkdirAll(%q): %v", path, err)
+	}
+	return path
+}
+
+func TestReadDenyAllowPrecedence(t *testing.T) {
+	ws := resolvedTempDir(t)
+	mkdir(t, filepath.Join(ws, "src"))
+	secret := mkdir(t, filepath.Join(ws, "secret"))
+	public := mkdir(t, filepath.Join(ws, "secret", "public"))
+
+	cases := []struct {
+		name       string
+		policy     Policy
+		path       string
+		wantDenied bool
+	}{
+		{
+			name:       "no lists: nothing denied",
+			policy:     Policy{},
+			path:       filepath.Join(secret, "data"),
+			wantDenied: false,
+		},
+		{
+			name:       "denyread blocks subtree",
+			policy:     Policy{DenyRead: []string{secret}},
+			path:       filepath.Join(secret, "data"),
+			wantDenied: true,
+		},
+		{
+			name:       "denyread does not touch siblings",
+			policy:     Policy{DenyRead: []string{secret}},
+			path:       filepath.Join(ws, "src", "main.go"),
+			wantDenied: false,
+		},
+		{
+			name:       "more-specific allowread re-includes",
+			policy:     Policy{DenyRead: []string{secret}, AllowRead: []string{public}},
+			path:       filepath.Join(public, "ok.txt"),
+			wantDenied: false,
+		},
+		{
+			name:       "allowread only re-includes its own subtree",
+			policy:     Policy{DenyRead: []string{secret}, AllowRead: []string{public}},
+			path:       filepath.Join(secret, "other.txt"),
+			wantDenied: true,
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := readDenied(tc.policy, ws, tc.path); got != tc.wantDenied {
+				t.Fatalf("readDenied(%q) = %v, want %v", tc.path, got, tc.wantDenied)
+			}
+		})
+	}
+}
+
+func TestWritePrecedenceMatrix(t *testing.T) {
+	ws := resolvedTempDir(t)
+	mkdir(t, filepath.Join(ws, "src"))
+	wsSecret := mkdir(t, filepath.Join(ws, "secret"))
+	ext := resolvedTempDir(t)
+	mkdir(t, filepath.Join(ext, "build"))
+	extProtected := mkdir(t, filepath.Join(ext, "build", "protected"))
+	outside := resolvedTempDir(t) // never allowed
+
+	scope, err := NewScope(ws, nil)
+	if err != nil {
+		t.Fatalf("NewScope: %v", err)
+	}
+
+	cases := []struct {
+		name      string
+		policy    Policy
+		path      string
+		wantAllow bool
+	}{
+		{
+			name:      "workspace path writable by default",
+			policy:    Policy{},
+			path:      filepath.Join(ws, "src", "main.go"),
+			wantAllow: true,
+		},
+		{
+			name:      "outside path denied by default",
+			policy:    Policy{},
+			path:      filepath.Join(outside, "x"),
+			wantAllow: false,
+		},
+		{
+			name:      "allowwrite extends to external root",
+			policy:    Policy{AllowWrite: []string{filepath.Join(ext, "build")}},
+			path:      filepath.Join(ext, "build", "out.o"),
+			wantAllow: true,
+		},
+		{
+			name:      "denywrite wins over workspace",
+			policy:    Policy{DenyWrite: []string{wsSecret}},
+			path:      filepath.Join(wsSecret, "creds"),
+			wantAllow: false,
+		},
+		{
+			name:      "denywrite wins over allowwrite",
+			policy:    Policy{AllowWrite: []string{filepath.Join(ext, "build")}, DenyWrite: []string{extProtected}},
+			path:      filepath.Join(extProtected, "x"),
+			wantAllow: false,
+		},
+		{
+			name:      "allowwrite does not allow unlisted external path",
+			policy:    Policy{AllowWrite: []string{filepath.Join(ext, "build")}},
+			path:      filepath.Join(outside, "x"),
+			wantAllow: false,
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			// This matrix exercises the workspace-enforced write precedence.
+			violation := validateWritePath(scope, tc.policy, true, ws, tc.path)
+			gotAllow := violation == nil
+			if gotAllow != tc.wantAllow {
+				t.Fatalf("validateWritePath(%q) allow=%v (violation=%v), want allow=%v", tc.path, gotAllow, violation, tc.wantAllow)
+			}
+		})
+	}
+}
+
+func TestReadExclusionGlobs(t *testing.T) {
+	ws := resolvedTempDir(t)
+	secret := mkdir(t, filepath.Join(ws, "secret"))
+	scope, err := NewScope(ws, nil)
+	if err != nil {
+		t.Fatalf("NewScope: %v", err)
+	}
+
+	// Empty DenyRead → no globs (default behavior unchanged).
+	if globs := ReadExclusionGlobs(Policy{}, scope); globs != nil {
+		t.Fatalf("empty DenyRead must yield no globs, got %v", globs)
+	}
+
+	// DenyRead inside the workspace → ripgrep exclusion globs for the rel path.
+	got := strings.Join(ReadExclusionGlobs(Policy{DenyRead: []string{secret}}, scope), " ")
+	want := "--glob !secret --glob !secret/**"
+	if got != want {
+		t.Fatalf("ReadExclusionGlobs = %q, want %q", got, want)
+	}
+
+	// DenyRead OUTSIDE the workspace → skipped (a rooted search never reaches it).
+	outside := resolvedTempDir(t)
+	if globs := ReadExclusionGlobs(Policy{DenyRead: []string{outside}}, scope); globs != nil {
+		t.Fatalf("out-of-workspace DenyRead must yield no globs, got %v", globs)
+	}
+}
+
+// TestReadDirExcludedDescendsForNestedAllow verifies a read-denied dir is NOT
+// skipped wholesale when it contains a nested AllowRead root, so the walk can
+// still reach the re-included subtree.
+func TestReadDirExcludedDescendsForNestedAllow(t *testing.T) {
+	ws := resolvedTempDir(t)
+	secret := mkdir(t, filepath.Join(ws, "secret"))
+	public := mkdir(t, filepath.Join(ws, "secret", "public"))
+
+	engine := NewEngine(EngineOptions{
+		WorkspaceRoot: ws,
+		Policy:        Policy{Mode: ModeEnforce, EnforceWorkspace: true, DenyRead: []string{secret}, AllowRead: []string{public}},
+	})
+	// The denied dir must be descended (nested allow), but the denied dir's own
+	// files are still excluded per-file.
+	rx := engine.ReadExclusions()
+	if rx.DirExcluded(secret) {
+		t.Fatalf("a denied dir with a nested AllowRead must not be skipped wholesale")
+	}
+	if !rx.PathExcluded(filepath.Join(secret, "creds")) {
+		t.Fatalf("a denied file outside the re-included subtree must be excluded")
+	}
+	if rx.PathExcluded(filepath.Join(public, "ok")) {
+		t.Fatalf("a re-included AllowRead file must not be excluded")
+	}
+}
+
+// TestDirExcludedResolvesSymlinkPrefix verifies DirExcluded resolves a symlink
+// in the path prefix before checking for a nested AllowRead root, so a denied dir
+// reached through a symlink is NOT skipped wholesale when it contains a
+// re-included subtree (the resolved allowRoots would otherwise never match).
+func TestDirExcludedResolvesSymlinkPrefix(t *testing.T) {
+	ws := resolvedTempDir(t)
+	real := mkdir(t, filepath.Join(ws, "real"))
+	secret := mkdir(t, filepath.Join(real, "secret"))
+	public := mkdir(t, filepath.Join(secret, "public"))
+	link := filepath.Join(ws, "link")
+	if err := os.Symlink(real, link); err != nil {
+		t.Skipf("symlinks unavailable on this platform: %v", err)
+	}
+
+	engine := NewEngine(EngineOptions{
+		WorkspaceRoot: ws,
+		Policy:        Policy{Mode: ModeEnforce, EnforceWorkspace: true, DenyRead: []string{secret}, AllowRead: []string{public}},
+	})
+	rx := engine.ReadExclusions()
+	// secret reached THROUGH the symlink still contains the re-included AllowRead
+	// subtree once the prefix is resolved, so it must not be skipped wholesale.
+	if rx.DirExcluded(filepath.Join(link, "secret")) {
+		t.Fatalf("DirExcluded must resolve the symlink prefix and detect the nested AllowRead (must not skip)")
+	}
+}
+
+// TestPathListsApplyWithoutWorkspaceEnforcement verifies the explicit path lists
+// (DenyWrite/DenyRead) are honored even when EnforceWorkspace is off, while the
+// workspace boundary itself is not enforced in that mode.
+func TestPathListsApplyWithoutWorkspaceEnforcement(t *testing.T) {
+	ws := resolvedTempDir(t)
+	secret := mkdir(t, filepath.Join(ws, "secret"))
+	outside := resolvedTempDir(t)
+	scope, err := NewScope(ws, nil)
+	if err != nil {
+		t.Fatalf("NewScope: %v", err)
+	}
+
+	// DenyWrite still blocks even with workspace enforcement off.
+	denyPolicy := Policy{DenyWrite: []string{secret}}
+	if v := validateWritePath(scope, denyPolicy, false, ws, filepath.Join(secret, "creds")); v == nil {
+		t.Fatal("DenyWrite must block even when EnforceWorkspace is off")
+	}
+	// With enforcement off and no DenyWrite match, an out-of-workspace write is
+	// NOT blocked (the workspace boundary is intentionally not applied).
+	if v := validateWritePath(scope, denyPolicy, false, ws, filepath.Join(outside, "x")); v != nil {
+		t.Fatalf("with EnforceWorkspace off, a non-denied path must be allowed, got %v", v)
+	}
+	// DenyRead still blocks reads even with workspace enforcement off.
+	if v := validatePathWithPolicy(scope, Policy{DenyRead: []string{secret}}, SideEffectRead, false, ws, filepath.Join(secret, "x")); v == nil {
+		t.Fatal("DenyRead must block reads even when EnforceWorkspace is off")
+	}
+}
+
+// TestEvaluatePathListsWithoutWorkspaceRoot verifies the path lists are enforced
+// even on an engine built WITHOUT a workspace root: absolute paths are matched
+// against DenyWrite, and an unanchorable relative path fails closed.
+func TestEvaluatePathListsWithoutWorkspaceRoot(t *testing.T) {
+	secret := resolvedTempDir(t)
+	other := resolvedTempDir(t)
+	engine := NewEngine(EngineOptions{
+		Policy: Policy{Mode: ModeEnforce, MaxAutonomy: AutonomyHigh, DenyWrite: []string{secret}},
+	})
+
+	denied := engine.Evaluate(context.Background(), Request{
+		ToolName: "write_file", SideEffect: SideEffectWrite,
+		Args: map[string]any{"path": filepath.Join(secret, "x")},
+	})
+	if denied.Action != ActionDeny {
+		t.Fatalf("absolute DenyWrite path must be denied without a workspace root, got %q", denied.Action)
+	}
+
+	ok := engine.Evaluate(context.Background(), Request{
+		ToolName: "write_file", SideEffect: SideEffectWrite,
+		Args: map[string]any{"path": filepath.Join(other, "x")},
+	})
+	if ok.Action == ActionDeny {
+		t.Fatalf("a non-denied absolute path must not be denied without a workspace root, got deny: %s", ok.ErrorString())
+	}
+
+	rel := engine.Evaluate(context.Background(), Request{
+		ToolName: "write_file", SideEffect: SideEffectWrite,
+		Args: map[string]any{"path": "x"},
+	})
+	if rel.Action != ActionDeny {
+		t.Fatalf("an unanchorable relative path must fail closed when path lists exist, got %q", rel.Action)
+	}
+}
+
+// TestReadExclusionsInactiveWhenDisabled verifies a disabled policy filters
+// nothing — ReadExclusions/ReadExclusionGlobs are inert under ModeDisabled even
+// when DenyRead is configured.
+func TestReadExclusionsInactiveWhenDisabled(t *testing.T) {
+	ws := resolvedTempDir(t)
+	secret := mkdir(t, filepath.Join(ws, "secret"))
+	engine := NewEngine(EngineOptions{
+		WorkspaceRoot: ws,
+		Policy:        Policy{Mode: ModeDisabled, DenyRead: []string{secret}},
+	})
+	if rx := engine.ReadExclusions(); rx.Active() {
+		t.Fatal("ReadExclusions must be inactive under a disabled policy")
+	}
+	if globs := engine.ReadExclusionGlobs(); globs != nil {
+		t.Fatalf("ReadExclusionGlobs must be empty under a disabled policy, got %v", globs)
+	}
+}
+
+// TestEvaluateAppliesReadDeny verifies the lists are wired into Evaluate: a read
+// under DenyRead is denied, while a sibling read is not.
+func TestEvaluateAppliesReadDeny(t *testing.T) {
+	ws := resolvedTempDir(t)
+	mkdir(t, filepath.Join(ws, "src"))
+	secret := mkdir(t, filepath.Join(ws, "secret"))
+
+	engine := NewEngine(EngineOptions{
+		WorkspaceRoot: ws,
+		Policy: Policy{
+			Mode:             ModeEnforce,
+			EnforceWorkspace: true,
+			MaxAutonomy:      AutonomyHigh,
+			DenyRead:         []string{secret},
+		},
+	})
+
+	denied := engine.Evaluate(context.Background(), Request{
+		ToolName:   "read_file",
+		SideEffect: SideEffectRead,
+		Args:       map[string]any{"path": filepath.Join(secret, "creds.txt")},
+	})
+	if denied.Action != ActionDeny {
+		t.Fatalf("read under DenyRead = %q, want deny", denied.Action)
+	}
+
+	ok := engine.Evaluate(context.Background(), Request{
+		ToolName:   "read_file",
+		SideEffect: SideEffectRead,
+		Args:       map[string]any{"path": filepath.Join(ws, "src", "main.go")},
+	})
+	if ok.Action == ActionDeny {
+		t.Fatalf("read of a non-denied sibling must not be denied, got %q (%s)", ok.Action, ok.ErrorString())
+	}
+}
+
+// TestEvaluateAppliesWriteAllow verifies an external write is denied by default
+// but permitted once the path is on AllowWrite.
+func TestEvaluateAppliesWriteAllow(t *testing.T) {
+	ws := resolvedTempDir(t)
+	ext := resolvedTempDir(t)
+	build := mkdir(t, filepath.Join(ext, "build"))
+
+	base := Policy{Mode: ModeEnforce, EnforceWorkspace: true, MaxAutonomy: AutonomyHigh}
+	target := filepath.Join(build, "out.o")
+
+	denyEngine := NewEngine(EngineOptions{WorkspaceRoot: ws, Policy: base})
+	denied := denyEngine.Evaluate(context.Background(), Request{
+		ToolName:   "write_file",
+		SideEffect: SideEffectWrite,
+		Args:       map[string]any{"path": target},
+	})
+	if denied.Action != ActionDeny {
+		t.Fatalf("external write without AllowWrite = %q, want deny", denied.Action)
+	}
+
+	allowPolicy := base
+	allowPolicy.AllowWrite = []string{build}
+	allowEngine := NewEngine(EngineOptions{WorkspaceRoot: ws, Policy: allowPolicy})
+	allowed := allowEngine.Evaluate(context.Background(), Request{
+		ToolName:   "write_file",
+		SideEffect: SideEffectWrite,
+		Args:       map[string]any{"path": target},
+	})
+	if allowed.Action == ActionDeny {
+		t.Fatalf("external write under AllowWrite must not be denied, got deny: %s", allowed.ErrorString())
+	}
+}
+
+// TestWriteRootsReflectAllowWrite verifies AllowWrite roots flow into the OS
+// backend write binds (so a sandboxed shell can write there).
+func TestWriteRootsReflectAllowWrite(t *testing.T) {
+	ws := resolvedTempDir(t)
+	ext := resolvedTempDir(t)
+	build := mkdir(t, filepath.Join(ext, "build"))
+
+	engine := NewEngine(EngineOptions{
+		WorkspaceRoot: ws,
+		Policy:        Policy{Mode: ModeEnforce, EnforceWorkspace: true, AllowWrite: []string{build}},
+	})
+	roots := engine.writeRoots(ws)
+	found := false
+	for _, r := range roots {
+		if r == build {
+			found = true
+		}
+	}
+	if !found {
+		t.Fatalf("writeRoots must include AllowWrite root %q, got %v", build, roots)
+	}
+}
+
+// TestSandboxExecProfileEmitsDenyWriteRule verifies a DenyWrite directory becomes
+// an explicit seatbelt deny clause AFTER the write-allow (last-match-wins).
+func TestSandboxExecProfileEmitsDenyWriteRule(t *testing.T) {
+	ws := resolvedTempDir(t)
+	secret := mkdir(t, filepath.Join(ws, "secret"))
+
+	policy := Policy{Mode: ModeEnforce, EnforceWorkspace: true, DenyWrite: []string{secret}}
+	profile := sandboxExecProfile([]string{ws}, policy, "", "", "")
+
+	// The path is escaped the same way the profile escapes it (e.g. Windows
+	// backslashes are doubled), so the assertion holds on every platform.
+	wantDeny := `(deny file-write* (subpath "` + sandboxProfileString(secret) + `"))`
+	if !strings.Contains(profile, wantDeny) {
+		t.Fatalf("profile missing DenyWrite rule %q:\n%s", wantDeny, profile)
+	}
+	// The deny clause must appear AFTER the write-allow so seatbelt's
+	// last-match-wins lets it override the bind.
+	allowIdx := strings.Index(profile, "(allow file-write*")
+	denyIdx := strings.Index(profile, wantDeny)
+	if allowIdx < 0 || denyIdx < 0 || denyIdx < allowIdx {
+		t.Fatalf("DenyWrite rule must follow the write-allow (allow@%d deny@%d):\n%s", allowIdx, denyIdx, profile)
+	}
+}

internal/sandbox/reentrancy_test.go

@@ -0,0 +1,90 @@
+package sandbox
+
+import "testing"
+
+func TestIsAlreadySandboxed(t *testing.T) {
+	// A lone ZERO_SANDBOXED=1 (no corroborating backend marker) must NOT count as
+	// sandboxed — stronger provenance than a single ambient flag.
+	t.Setenv(EnvSandboxBackend, "")
+	t.Setenv(EnvSandboxed, "1")
+	if IsAlreadySandboxed() {
+		t.Fatalf("IsAlreadySandboxed must be false when only %s=1 is set without %s", EnvSandboxed, EnvSandboxBackend)
+	}
+	// The backend marker alone (no ZERO_SANDBOXED=1) must not count either.
+	t.Setenv(EnvSandboxBackend, string(BackendBubblewrap))
+	t.Setenv(EnvSandboxed, "")
+	if IsAlreadySandboxed() {
+		t.Fatalf("IsAlreadySandboxed must be false when only %s is set", EnvSandboxBackend)
+	}
+	// Both correlated markers present (as sandboxEnvironment sets them) → sandboxed.
+	t.Setenv(EnvSandboxed, "1")
+	if !IsAlreadySandboxed() {
+		t.Fatalf("IsAlreadySandboxed must be true when both %s=1 and %s are set", EnvSandboxed, EnvSandboxBackend)
+	}
+	t.Setenv(EnvSandboxed, "0")
+	if IsAlreadySandboxed() {
+		t.Fatalf("IsAlreadySandboxed must be false when %s=0", EnvSandboxed)
+	}
+}
+
+func TestSandboxEnvironmentMarksSandboxed(t *testing.T) {
+	env := sandboxEnvironment(DefaultPolicy(), BackendBubblewrap, "/home/agent")
+	// Both correlated markers must be set: IsAlreadySandboxed requires BOTH, so a
+	// regression that drops either one would silently break re-entrancy detection.
+	wantSandboxed := EnvSandboxed + "=1"
+	wantBackend := EnvSandboxBackend + "=" + string(BackendBubblewrap)
+	var gotSandboxed, gotBackend bool
+	for _, entry := range env {
+		switch entry {
+		case wantSandboxed:
+			gotSandboxed = true
+		case wantBackend:
+			gotBackend = true
+		}
+	}
+	if !gotSandboxed || !gotBackend {
+		t.Fatalf("sandboxEnvironment must set %q and %q so a wrapped command is detectable; got %#v", wantSandboxed, wantBackend, env)
+	}
+}
+
+func TestBuildCommandPlanWrapsWhenNotAlreadySandboxed(t *testing.T) {
+	// Ensure neither re-entrancy marker is set so we are NOT seen as sandboxed.
+	t.Setenv(EnvSandboxed, "")
+	t.Setenv(EnvSandboxBackend, "")
+	root := t.TempDir()
+	engine := NewEngine(EngineOptions{
+		WorkspaceRoot: root,
+		Policy:        DefaultPolicy(),
+		Backend:       Backend{Name: BackendBubblewrap, Available: true, Executable: "/usr/bin/bwrap"},
+	})
+	plan, err := engine.BuildCommandPlan(CommandSpec{Name: "/bin/sh", Args: []string{"-c", "pwd"}, Dir: root})
+	if err != nil {
+		t.Fatalf("BuildCommandPlan: %v", err)
+	}
+	if !plan.Wrapped || plan.Name != "/usr/bin/bwrap" {
+		t.Fatalf("expected a wrapped bubblewrap plan, got wrapped=%v name=%q", plan.Wrapped, plan.Name)
+	}
+}
+
+func TestBuildCommandPlanReEntrancyGuardReturnsPassThrough(t *testing.T) {
+	// Simulate running inside an already-wrapped process: BOTH correlated markers
+	// that sandboxEnvironment sets together must be present for the guard to fire.
+	t.Setenv(EnvSandboxed, "1")
+	t.Setenv(EnvSandboxBackend, string(BackendBubblewrap))
+	root := t.TempDir()
+	engine := NewEngine(EngineOptions{
+		WorkspaceRoot: root,
+		Policy:        DefaultPolicy(),
+		Backend:       Backend{Name: BackendBubblewrap, Available: true, Executable: "/usr/bin/bwrap"},
+	})
+	plan, err := engine.BuildCommandPlan(CommandSpec{Name: "/bin/sh", Args: []string{"-c", "pwd"}, Dir: root})
+	if err != nil {
+		t.Fatalf("BuildCommandPlan: %v", err)
+	}
+	if plan.Wrapped {
+		t.Fatalf("re-entrancy guard must return an unwrapped pass-through plan, got wrapped=%v name=%q args=%v", plan.Wrapped, plan.Name, plan.Args)
+	}
+	if plan.Name != "/bin/sh" {
+		t.Fatalf("pass-through plan must run the command directly, got name=%q", plan.Name)
+	}
+}

internal/sandbox/risk.go

@@ -106,6 +106,22 @@ func classifyWithScope(request Request, scope *Scope) Risk {
 		if pipedInstallerPattern.MatchString(command) {
 			add("piped_installer", RiskCritical)
 		}
+		// AST second opinion (analyzer.go): walks the parsed shell tree, so it
+		// catches destructive/network programs the regexes miss — e.g. shred,
+		// fdisk, parted, and commands hidden behind sudo/env wrappers or a
+		// `sh -c <payload>` launcher — and flags an unparseable (obfuscated)
+		// script as elevated risk. It only ADDS categories, so a benign,
+		// parseable command is classified exactly as before.
+		analysis := AnalyzeCommand(command)
+		if analysis.Network {
+			add("network", RiskCritical)
+		}
+		if analysis.Destructive {
+			add("destructive", RiskCritical)
+		}
+		if analysis.TooComplex {
+			add("unparseable_command", RiskHigh)
+		}
 	}
 
 	for _, path := range requestPaths(request) {

internal/sandbox/risk_hardening_test.go

@@ -229,3 +229,69 @@ func TestClassifyNoneSideEffectIsLowRisk(t *testing.T) {
 		t.Fatalf("control-only tool must not classify as out_of_workspace: %#v", risk)
 	}
 }
+
+// The following tests cover the AST analyzer wired into classifyWithScope as a
+// second opinion to the regex detectors.
+
+func TestClassifyASTCatchesDestructiveProgramsRegexMisses(t *testing.T) {
+	// shred/fdisk/parted are irrecoverably destructive but absent from the regex
+	// pattern; the AST analyzer flags them — including behind a sh -c launcher or
+	// a sudo/env wrapper (effectiveProgram resolves the real program). The escalated
+	// level (Critical) is part of the contract, so assert it alongside the category.
+	for _, command := range []string{
+		"shred -u secret.txt",
+		"fdisk /dev/sda",
+		"parted /dev/sda mklabel gpt",
+		"bash -c 'shred /etc/passwd'",
+		"sudo shred -u secret.txt",
+		"env shred -u secret.txt",
+	} {
+		risk := classifyCommand(command)
+		if risk.Level != RiskCritical || !HasRiskCategory(risk, "destructive") {
+			t.Fatalf("Classify(%q) = level %s, categories %v; want critical destructive", command, risk.Level, risk.Categories)
+		}
+	}
+}
+
+func TestClassifyASTCatchesNetworkProgramsRegexMisses(t *testing.T) {
+	for _, command := range []string{
+		"telnet example.com 23",
+		"ftp ftp.example.com",
+		"sftp user@host",
+		"sudo telnet example.com 23",
+	} {
+		risk := classifyCommand(command)
+		if risk.Level != RiskCritical || !HasRiskCategory(risk, "network") {
+			t.Fatalf("Classify(%q) = level %s, categories %v; want critical network", command, risk.Level, risk.Categories)
+		}
+	}
+}
+
+func TestClassifyFlagsUnparseableCommand(t *testing.T) {
+	// An unparseable (e.g. obfuscated) script can't be analyzed statically; the
+	// AST analyzer reports TooComplex so the classifier elevates it to High.
+	risk := classifyCommand(`echo "unterminated`)
+	if risk.Level != RiskHigh || !HasRiskCategory(risk, "unparseable_command") {
+		t.Fatalf("Classify of unparseable command = level %s, categories %v; want high unparseable_command", risk.Level, risk.Categories)
+	}
+}
+
+func TestClassifyASTDoesNotFlagQuotedProgramName(t *testing.T) {
+	// A program name inside a quoted argument is not a command, so the AST
+	// analyzer must not flag it (documenting a destructive command in an echo).
+	risk := classifyCommand(`echo "shred wipes files"`)
+	if HasRiskCategory(risk, "destructive") {
+		t.Fatalf("Classify(%q) wrongly flagged destructive: %v", `echo "shred wipes files"`, risk.Categories)
+	}
+}
+
+func TestClassifyBenignCommandStaysClean(t *testing.T) {
+	for _, command := range []string{"echo hello", "ls -la", "go build ./..."} {
+		risk := classifyCommand(command)
+		for _, category := range []string{"destructive", "network", "unparseable_command"} {
+			if HasRiskCategory(risk, category) {
+				t.Fatalf("Classify(%q) wrongly flagged %s: %v", command, category, risk.Categories)
+			}
+		}
+	}
+}

internal/sandbox/runner.go

@@ -17,6 +17,13 @@ const bubblewrapWorkspace = "/workspace"
 
 var errPolicyOnlyRunnerDisabled = errors.New("policy-only sandbox runner is disabled")
 
+// errWSLPolicyOnlyDisabled is returned when running under WSL without bubblewrap
+// and the policy has NOT opted into the policy-only runner — fail closed rather
+// than run with no native isolation.
+var errWSLPolicyOnlyDisabled = errors.New("bubblewrap is unavailable under WSL and Policy.AllowPolicyOnlyRunner is disabled: " +
+	"enable AllowPolicyOnlyRunner to run under the policy-only WSL fallback (engine policy + proxy egress, no native isolation), " +
+	"or install/enable bubblewrap")
+
 type CommandSpec struct {
 	Name string
 	Args []string
@@ -39,6 +46,10 @@ type CommandPlan struct {
 	// StartDenialMonitor to capture what the sandbox blocked. Empty unless
 	// Policy.MonitorDenials is set on a macOS sandbox-exec plan.
 	MonitorTag string `json:"monitorTag,omitempty"`
+	// Notes records auditable least-privilege downgrade notes — e.g. the WSL
+	// policy-only fallback noting that native isolation was unavailable. Surfaced to
+	// the operator; never affects enforcement.
+	Notes []string `json:"notes,omitempty"`
 	// cleanup releases resources tied to the plan's lifetime — currently the
 	// scoped-egress proxy, which must outlive the command run and be shut down
 	// afterwards. It is never serialized; callers invoke it via Cleanup() once the
@@ -99,6 +110,29 @@ func ProxyEnv(addr string) []string {
 	}
 }
 
+// ProxyEnvWithSocks is the SOCKS-aware form of ProxyEnv built ON TOP of it:
+// HTTP_PROXY/HTTPS_PROXY stay on the HTTP listener (httpAddr) for HTTP(S)-aware
+// clients, while ALL_PROXY/all_proxy are re-pointed at the SOCKS5 front-end
+// (socksAddr) so a client that honors ALL_PROXY tunnels arbitrary TCP through the
+// same allow/deny gate. When socksAddr is empty it returns exactly ProxyEnv's
+// output, so the SOCKS upgrade is purely additive and never weakens the default.
+func ProxyEnvWithSocks(httpAddr, socksAddr string) []string {
+	env := ProxyEnv(httpAddr)
+	if socksAddr == "" {
+		return env
+	}
+	socksURL := "socks5://" + socksAddr
+	for i, kv := range env {
+		switch {
+		case strings.HasPrefix(kv, "ALL_PROXY="):
+			env[i] = "ALL_PROXY=" + socksURL
+		case strings.HasPrefix(kv, "all_proxy="):
+			env[i] = "all_proxy=" + socksURL
+		}
+	}
+	return env
+}
+
 func (engine *Engine) CommandContext(ctx context.Context, spec CommandSpec) (*exec.Cmd, CommandPlan, error) {
 	if ctx == nil {
 		ctx = context.Background()
@@ -118,10 +152,20 @@ func (engine *Engine) CommandContext(ctx context.Context, spec CommandSpec) (*ex
 // only applies to engines built without a workspace root (NewEngine always
 // builds a scope otherwise); it is kept as defense in depth.
 func (engine *Engine) writeRoots(workspaceRoot string) []string {
+	var roots []string
 	if engine.scope != nil {
-		return engine.scope.Roots()
+		roots = engine.scope.Roots()
+	} else {
+		roots = []string{workspaceRoot}
 	}
-	return []string{workspaceRoot}
+	// Reflect the policy's AllowWrite roots in the OS backend write binds so a
+	// sandboxed shell command may write where the policy grants writes. DenyWrite
+	// is enforced at the policy gate, and on sandbox-exec additionally as an
+	// explicit deny rule (see sandboxExecProfile).
+	if extra := resolveWriteRootPaths(engine.policy.AllowWrite); len(extra) > 0 {
+		roots = dedupeStrings(append(roots, extra...))
+	}
+	return roots
 }
 
 func (engine *Engine) BuildCommandPlan(spec CommandSpec) (CommandPlan, error) {
@@ -146,6 +190,15 @@ func (engine *Engine) BuildCommandPlan(spec CommandSpec) (CommandPlan, error) {
 	if backend.Name == "" {
 		backend = Backend{Name: BackendPolicyOnly, Message: "policy-only fallback: sandbox backend was not selected"}
 	}
+	// Re-entrancy guard: a command spawned by a process we already wrapped (both
+	// ZERO_SANDBOXED=1 and ZERO_SANDBOX_BACKEND set in its env — see
+	// IsAlreadySandboxed) must not be wrapped again — nested bwrap / sandbox-exec
+	// fails and a second egress proxy would be redundant. Return a pass-through
+	// plan. Mirrors the already-sandboxed guard used by comparable executor
+	// sandboxes.
+	if IsAlreadySandboxed() {
+		return directCommandPlan(spec, backend, policy, workspaceRoot), nil
+	}
 	if policy.Mode == ModeDisabled {
 		return directCommandPlan(spec, backend, policy, workspaceRoot), nil
 	}
@@ -166,6 +219,18 @@ func (engine *Engine) BuildCommandPlan(spec CommandSpec) (CommandPlan, error) {
 			}
 			return sandboxExecCommandPlan(spec, workspaceRoot, engine.writeRoots(workspaceRoot), policy, backend, egress), nil
 		}
+	case BackendWSL:
+		// WSL fallback: no native isolation available. Fail closed unless the policy
+		// opted into the policy-only runner; otherwise run policy-only with the
+		// command's egress routed through the filtering proxy and an auditable note.
+		if !policy.AllowPolicyOnlyRunner {
+			return CommandPlan{}, errWSLPolicyOnlyDisabled
+		}
+		egress, err := startScopedEgress(policy, backend)
+		if err != nil {
+			return CommandPlan{}, err
+		}
+		return wslCommandPlan(spec, workspaceRoot, policy, backend, egress), nil
 	}
 	if !policy.AllowPolicyOnlyRunner {
 		return CommandPlan{}, errPolicyOnlyRunnerDisabled
@@ -173,12 +238,79 @@ func (engine *Engine) BuildCommandPlan(spec CommandSpec) (CommandPlan, error) {
 	return directCommandPlan(spec, backend, policy, workspaceRoot), nil
 }
 
+// wslCommandPlan builds the WSL policy-only fallback plan: the command runs
+// DIRECTLY (no native wrap, since bubblewrap is unavailable under WSL) but with
+// the sandbox env markers (ZERO_SANDBOXED=1, ZERO_SANDBOX_BACKEND=wsl) and, when
+// a scoped-egress proxy was started, the proxy env so well-behaved clients route
+// through the allow/deny gate. It carries a least-privilege downgrade note.
+func wslCommandPlan(spec CommandSpec, workspaceRoot string, policy Policy, backend Backend, egress *scopedEgress) CommandPlan {
+	// The WSL fallback runs the command DIRECTLY (no native wrap), so it inherits
+	// the caller's environment like the generic policy-only runner; the sandbox
+	// markers are appended last so they cannot be shadowed by the caller's env.
+	var env []string
+	if spec.Env != nil {
+		env = cloneStrings(spec.Env)
+	} else {
+		env = append(env, os.Environ()...)
+	}
+	env = append(env,
+		EnvSandboxBackend+"="+string(BackendWSL),
+		"ZERO_SANDBOX_NETWORK="+string(policy.Network),
+		EnvSandboxed+"=1",
+	)
+	note := "WSL: native process isolation (bubblewrap) is unavailable; downgraded to the policy-only runner " +
+		"(least privilege). The engine still enforces the full policy."
+	if egress != nil {
+		env = append(env, ProxyEnvWithSocks(egress.addr, egress.socksAddr)...)
+		env = appendCACertEnv(env, egress)
+		note = "WSL: native process isolation (bubblewrap) is unavailable; downgraded to the policy-only runner " +
+			"with proxy egress (least privilege). The engine still enforces the full policy."
+	}
+	plan := CommandPlan{
+		Backend:       backend,
+		WorkspaceRoot: workspaceRoot,
+		Policy:        policy,
+		Wrapped:       false,
+		Name:          spec.Name,
+		Args:          cloneStrings(spec.Args),
+		Dir:           spec.Dir,
+		Env:           env,
+		Notes:         []string{note},
+	}
+	if egress != nil {
+		plan.cleanup = egress.cleanup
+	}
+	return plan
+}
+
 // scopedEgress holds the address of a started scoped-egress proxy and the
 // cleanup that shuts it down. A nil *scopedEgress means scoped egress is not in
 // effect for this command (the network mode is allow or deny-equivalent).
 type scopedEgress struct {
-	addr    string
-	cleanup func()
+	addr string
+	// socksAddr is the loopback address of the proxy's SOCKS5 front-end, used to
+	// set ALL_PROXY=socks5://<socksAddr>. Empty when the proxy exposes no SOCKS
+	// listener, in which case ALL_PROXY falls back to the HTTP proxy.
+	socksAddr string
+	// caCertPath is the path to the MITM CA's public cert (PEM), set only when
+	// Policy.InspectTLS started a TLS-terminating proxy. Surfaced to the sandboxed
+	// client via ZERO_SANDBOX_CA_CERT so it can trust the minted leaves.
+	caCertPath string
+	cleanup    func()
+}
+
+// EnvCACert is the env var pointing the sandboxed client at the MITM CA's public
+// cert so it can trust the proxy's minted leaf certificates (only set when
+// Policy.InspectTLS is on).
+const EnvCACert = "ZERO_SANDBOX_CA_CERT"
+
+// appendCACertEnv adds ZERO_SANDBOX_CA_CERT when the egress proxy is terminating
+// TLS (InspectTLS). It is a no-op for a plain passthrough proxy.
+func appendCACertEnv(env []string, egress *scopedEgress) []string {
+	if egress == nil || egress.caCertPath == "" {
+		return env
+	}
+	return append(env, EnvCACert+"="+egress.caCertPath)
 }
 
 // startScopedEgress starts the local filtering proxy when the policy's effective
@@ -196,13 +328,32 @@ func startScopedEgress(policy Policy, backend Backend) (*scopedEgress, error) {
 		return nil, nil
 	}
 	proxy, err := startEgressProxy(egressOptions{
-		Allowed: policy.AllowedDomains,
-		Denied:  policy.DeniedDomains,
+		Allowed:    policy.AllowedDomains,
+		Denied:     policy.DeniedDomains,
+		InspectTLS: policy.InspectTLS,
 	})
 	if err != nil {
 		return nil, fmt.Errorf("scoped egress unavailable, denying network: %w", err)
 	}
-	return &scopedEgress{addr: proxy.Addr(), cleanup: func() { _ = proxy.Close() }}, nil
+	se := &scopedEgress{addr: proxy.Addr(), socksAddr: proxy.SocksAddr(), cleanup: func() { _ = proxy.Close() }}
+	if policy.InspectTLS {
+		// Write the MITM CA's PUBLIC cert so the sandboxed client can trust the
+		// minted leaves; surface it via ZERO_SANDBOX_CA_CERT and clean it up with
+		// the proxy. Reuses no new bind path: sandbox-exec already allows file-read,
+		// and the WSL policy-only runner shares the host filesystem.
+		caPath, werr := writeMITMCAFile(proxy.CACertPEM())
+		if werr != nil {
+			_ = proxy.Close()
+			return nil, werr
+		}
+		se.caCertPath = caPath
+		base := se.cleanup
+		se.cleanup = func() {
+			base()
+			_ = os.Remove(caPath)
+		}
+	}
+	return se, nil
 }
 
 func directCommandPlan(spec CommandSpec, backend Backend, policy Policy, workspaceRoot string) CommandPlan {
@@ -374,21 +525,25 @@ func findSeccompHelper() string {
 }
 
 func sandboxExecCommandPlan(spec CommandSpec, workspaceRoot string, writeRoots []string, policy Policy, backend Backend, egress *scopedEgress) CommandPlan {
-	var proxyPort string
+	var proxyPort, socksPort string
 	if egress != nil {
 		if _, port, err := net.SplitHostPort(egress.addr); err == nil {
 			proxyPort = port
 		}
+		if _, port, err := net.SplitHostPort(egress.socksAddr); err == nil {
+			socksPort = port
+		}
 	}
 	denialTag := ""
 	if policy.MonitorDenials {
 		denialTag = nextSandboxDenialTag()
 	}
-	args := []string{"-p", sandboxExecProfile(writeRoots, policy, proxyPort, denialTag), spec.Name}
+	args := []string{"-p", sandboxExecProfile(writeRoots, policy, proxyPort, socksPort, denialTag), spec.Name}
 	args = append(args, spec.Args...)
 	env := sandboxEnvironment(policy, BackendSandboxExec, workspaceRoot)
 	if egress != nil {
-		env = append(env, ProxyEnv(egress.addr)...)
+		env = append(env, ProxyEnvWithSocks(egress.addr, egress.socksAddr)...)
+		env = appendCACertEnv(env, egress)
 	}
 	plan := CommandPlan{
 		Backend:       backend,
@@ -426,8 +581,9 @@ func sandboxEnvironment(policy Policy, backend BackendName, home string) []strin
 		"HOME=" + home,
 		"PATH=" + firstEnv("PATH", defaultPath()),
 		"TERM=" + firstEnv("TERM", "dumb"),
-		"ZERO_SANDBOX_BACKEND=" + string(backend),
+		EnvSandboxBackend + "=" + string(backend),
 		"ZERO_SANDBOX_NETWORK=" + string(policy.Network),
+		EnvSandboxed + "=1",
 	}
 	if runtime.GOOS == "windows" {
 		env = append(env, "COMSPEC="+firstEnv("COMSPEC", "cmd.exe"))
@@ -540,8 +696,8 @@ func sandboxMachLookupRule() string {
 	return "(allow mach-lookup\n  " + strings.Join(filters, "\n  ") + ")"
 }
 
-func sandboxExecProfile(writeRoots []string, policy Policy, proxyPort string, denialTag string) string {
-	networkRule := networkRuleFor(policy, proxyPort)
+func sandboxExecProfile(writeRoots []string, policy Policy, proxyPort string, socksPort string, denialTag string) string {
+	networkRule := networkRuleFor(policy, proxyPort, socksPort)
 	writeRule := "(allow file-write*)"
 	if policy.EnforceWorkspace {
 		// The granted write roots are the only writable *project* locations. Temp
@@ -566,7 +722,7 @@ func sandboxExecProfile(writeRoots []string, policy Policy, proxyPort string, de
 		// filters `log stream` for this exact (per-plan) tag.
 		denyDefault = `(deny default (with message "` + sandboxProfileString(denialTag) + `"))`
 	}
-	return strings.Join([]string{
+	rules := []string{
 		"(version 1)",
 		denyDefault,
 		"(allow process*)",
@@ -579,27 +735,67 @@ func sandboxExecProfile(writeRoots []string, policy Policy, proxyPort string, de
 		sandboxMachLookupRule(),
 		"(allow file-read*)",
 		writeRule,
-		networkRule,
-	}, "\n")
+	}
+	// DenyWrite entries are emitted AFTER the write allow so seatbelt's
+	// last-match-wins makes them override the workspace/AllowWrite binds — so
+	// "DenyWrite wins" holds for shell commands too, not just the policy gate.
+	rules = append(rules, denyWriteRules(policy)...)
+	rules = append(rules, networkRule)
+	return strings.Join(rules, "\n")
+}
+
+// denyWriteRules returns seatbelt deny clauses for the policy's resolved
+// DenyWrite paths: a (subpath ...) clause for a directory, a (literal ...) clause
+// for a single file. Empty when DenyWrite is unset.
+func denyWriteRules(policy Policy) []string {
+	resolved := resolvePolicyPaths(policy.DenyWrite)
+	if len(resolved) == 0 {
+		return nil
+	}
+	out := make([]string, 0, len(resolved))
+	for _, path := range resolved {
+		filter := `(subpath "` + sandboxProfileString(path) + `")`
+		if info, err := os.Stat(path); err == nil && !info.IsDir() {
+			filter = `(literal "` + sandboxProfileString(path) + `")`
+		}
+		out = append(out, "(deny file-write* "+filter+")")
+	}
+	return out
 }
 
 // networkRuleFor returns the seatbelt network clause for a policy. allow opens
 // all network; deny (and an empty-allowlist scoped policy, which effectiveNetwork
 // collapses to deny) blocks all network; scoped denies general network but
-// permits only outbound to the local proxy port on localhost, so traffic must
-// flow through the filtering proxy. A scoped policy with no resolvable proxy port
-// falls back to a full deny (fail closed).
-func networkRuleFor(policy Policy, proxyPort string) string {
+// permits only outbound to the local proxy ports on localhost, so traffic must
+// flow through the filtering proxy. Both the HTTP proxy port and the SOCKS5
+// front-end port are allowed (a sandboxed process configured with
+// ALL_PROXY=socks5://... must reach the SOCKS listener). A scoped policy with no
+// resolvable proxy port falls back to a full deny (fail closed).
+func networkRuleFor(policy Policy, proxyPort string, socksPort string) string {
 	switch effectiveNetwork(policy) {
 	case NetworkAllow:
 		return "(allow network*)"
 	case NetworkScoped:
-		if strings.TrimSpace(proxyPort) == "" {
+		rules := []string{"(deny network*)"}
+		seen := map[string]struct{}{}
+		// Deny by default, then allow only outbound to each distinct proxy port on
+		// loopback. Duplicate/empty ports are skipped so the clause stays minimal.
+		for _, port := range []string{proxyPort, socksPort} {
+			port = strings.TrimSpace(port)
+			if port == "" {
+				continue
+			}
+			if _, ok := seen[port]; ok {
+				continue
+			}
+			seen[port] = struct{}{}
+			rules = append(rules, `(allow network-outbound (remote ip "localhost:`+sandboxProfileString(port)+`"))`)
+		}
+		if len(rules) == 1 {
+			// No resolvable proxy port: fail closed.
 			return "(deny network*)"
 		}
-		// Deny by default, then allow only outbound to the proxy on loopback.
-		return "(deny network*)\n" +
-			`(allow network-outbound (remote ip "localhost:` + sandboxProfileString(proxyPort) + `"))`
+		return strings.Join(rules, "\n")
 	default:
 		return "(deny network*)"
 	}

internal/sandbox/runner_scoped_test.go

@@ -23,8 +23,17 @@ func TestScopedNetworkGateInEvaluate(t *testing.T) {
 	}
 	sandboxExec := Backend{Name: BackendSandboxExec, Available: true, Executable: "/usr/sbin/sandbox-exec", ScopedEgress: true}
 
+	// This test exercises the network GATE for a network tool, so it opts the
+	// in-process tools into the network policy (EnforceToolNetwork); by default
+	// those tools are exempt and would not be gated here.
+	enforcedScoped := func(allowed, denied []string) Policy {
+		p := scopedPolicy(allowed, denied)
+		p.EnforceToolNetwork = true
+		return p
+	}
+
 	// An enforcing backend lets a populated scoped policy permit a network tool.
-	enforcing := NewEngine(EngineOptions{WorkspaceRoot: root, Policy: scopedPolicy([]string{"github.com"}, nil), Backend: sandboxExec})
+	enforcing := NewEngine(EngineOptions{WorkspaceRoot: root, Policy: enforcedScoped([]string{"github.com"}, nil), Backend: sandboxExec})
 	if decision := enforcing.Evaluate(context.Background(), networkRequest); decision.Action == ActionDeny {
 		t.Fatalf("enforcing backend + scoped policy denied a network tool: %#v", decision)
 	}
@@ -35,7 +44,7 @@ func TestScopedNetworkGateInEvaluate(t *testing.T) {
 		"policy-only": {},
 	}
 	for name, backend := range unenforceable {
-		engine := NewEngine(EngineOptions{WorkspaceRoot: root, Policy: scopedPolicy([]string{"github.com"}, nil), Backend: backend})
+		engine := NewEngine(EngineOptions{WorkspaceRoot: root, Policy: enforcedScoped([]string{"github.com"}, nil), Backend: backend})
 		decision := engine.Evaluate(context.Background(), networkRequest)
 		if decision.Action != ActionDeny || decision.Violation == nil || decision.Violation.Code != ViolationNetwork {
 			t.Fatalf("%s backend must deny scoped network it cannot enforce, got %#v", name, decision)
@@ -43,7 +52,7 @@ func TestScopedNetworkGateInEvaluate(t *testing.T) {
 	}
 
 	// Empty allowlist still fails closed regardless of backend.
-	empty := NewEngine(EngineOptions{WorkspaceRoot: root, Policy: scopedPolicy(nil, nil), Backend: sandboxExec})
+	empty := NewEngine(EngineOptions{WorkspaceRoot: root, Policy: enforcedScoped(nil, nil), Backend: sandboxExec})
 	decision := empty.Evaluate(context.Background(), networkRequest)
 	if decision.Action != ActionDeny || decision.Violation == nil || decision.Violation.Code != ViolationNetwork {
 		t.Fatalf("empty scoped policy must deny network like NetworkDeny, got %#v", decision)

internal/sandbox/runner_test.go

@@ -233,7 +233,7 @@ func resolvedTestPath(t *testing.T, path string) string {
 }
 
 func TestSandboxExecProfileIncludesExtraWriteRoots(t *testing.T) {
-	profile := sandboxExecProfile([]string{"/ws", "/extra root"}, Policy{Mode: ModeEnforce, EnforceWorkspace: true}, "", "")
+	profile := sandboxExecProfile([]string{"/ws", "/extra root"}, Policy{Mode: ModeEnforce, EnforceWorkspace: true}, "", "", "")
 	if !strings.Contains(profile, "(allow file-write*") {
 		t.Fatalf("profile missing file-write rule:\n%s", profile)
 	}
@@ -251,7 +251,7 @@ func TestSandboxExecProfileIncludesExtraWriteRoots(t *testing.T) {
 }
 
 func TestSandboxExecProfileTagsDenialsWhenMonitoring(t *testing.T) {
-	off := sandboxExecProfile([]string{"/ws"}, Policy{Mode: ModeEnforce, EnforceWorkspace: true}, "", "")
+	off := sandboxExecProfile([]string{"/ws"}, Policy{Mode: ModeEnforce, EnforceWorkspace: true}, "", "", "")
 	if strings.Contains(off, "with message") {
 		t.Fatalf("denials must not be tagged when monitoring is off:\n%s", off)
 	}
@@ -259,7 +259,7 @@ func TestSandboxExecProfileTagsDenialsWhenMonitoring(t *testing.T) {
 		t.Fatalf("profile missing the plain default-deny:\n%s", off)
 	}
 
-	on := sandboxExecProfile([]string{"/ws"}, Policy{Mode: ModeEnforce, EnforceWorkspace: true, MonitorDenials: true}, "", "run-tag-123")
+	on := sandboxExecProfile([]string{"/ws"}, Policy{Mode: ModeEnforce, EnforceWorkspace: true, MonitorDenials: true}, "", "", "run-tag-123")
 	if !strings.Contains(on, `(deny default (with message "run-tag-123"))`) {
 		t.Fatalf("denials must be tagged when monitoring is on:\n%s", on)
 	}
@@ -291,7 +291,7 @@ func TestSandboxExecCommandPlanUsesUniquePerPlanDenialTag(t *testing.T) {
 }
 
 func TestSandboxExecProfileGrantsSignalAndMachLookup(t *testing.T) {
-	profile := sandboxExecProfile([]string{"/ws"}, Policy{Mode: ModeEnforce, EnforceWorkspace: true}, "", "")
+	profile := sandboxExecProfile([]string{"/ws"}, Policy{Mode: ModeEnforce, EnforceWorkspace: true}, "", "", "")
 
 	// Signalling own process group lets a sandboxed script kill the children it
 	// spawns; without it seatbelt denies kill() with "Operation not permitted".
@@ -465,3 +465,47 @@ func TestProxyEnv(t *testing.T) {
 		t.Fatalf("ProxyEnv NO_PROXY must exclude loopback, got %q/%q", env["NO_PROXY"], env["no_proxy"])
 	}
 }
+
+func TestProxyEnvWithSocks(t *testing.T) {
+	env := map[string]string{}
+	for _, kv := range ProxyEnvWithSocks("127.0.0.1:8899", "127.0.0.1:1080") {
+		if i := strings.IndexByte(kv, '='); i >= 0 {
+			env[kv[:i]] = kv[i+1:]
+		}
+	}
+	// HTTP_PROXY/HTTPS_PROXY stay on the HTTP listener...
+	const httpURL = "http://127.0.0.1:8899"
+	for _, key := range []string{"HTTP_PROXY", "HTTPS_PROXY", "http_proxy", "https_proxy"} {
+		if env[key] != httpURL {
+			t.Fatalf("ProxyEnvWithSocks[%s] = %q, want %q", key, env[key], httpURL)
+		}
+	}
+	// ...while ALL_PROXY/all_proxy point at the SOCKS5 front-end.
+	const socksURL = "socks5://127.0.0.1:1080"
+	for _, key := range []string{"ALL_PROXY", "all_proxy"} {
+		if env[key] != socksURL {
+			t.Fatalf("ProxyEnvWithSocks[%s] = %q, want %q", key, env[key], socksURL)
+		}
+	}
+	if !strings.Contains(env["NO_PROXY"], "127.0.0.1") {
+		t.Fatalf("ProxyEnvWithSocks NO_PROXY must exclude loopback, got %q", env["NO_PROXY"])
+	}
+}
+
+// TestProxyEnvWithSocksEmptyFallsBackToHTTP verifies the SOCKS upgrade is purely
+// additive: with no SOCKS address, ALL_PROXY falls back to the HTTP proxy exactly
+// as the legacy ProxyEnv produced, never weakening the default.
+func TestProxyEnvWithSocksEmptyFallsBackToHTTP(t *testing.T) {
+	env := map[string]string{}
+	for _, kv := range ProxyEnvWithSocks("127.0.0.1:8899", "") {
+		if i := strings.IndexByte(kv, '='); i >= 0 {
+			env[kv[:i]] = kv[i+1:]
+		}
+	}
+	const httpURL = "http://127.0.0.1:8899"
+	for _, key := range []string{"ALL_PROXY", "all_proxy"} {
+		if env[key] != httpURL {
+			t.Fatalf("ProxyEnvWithSocks(empty socks)[%s] = %q, want HTTP fallback %q", key, env[key], httpURL)
+		}
+	}
+}

internal/sandbox/types.go

@@ -12,6 +12,31 @@ import (
 // only an explicit truthy value enables it.
 const EnvAutoAllowBash = "ZERO_SANDBOX_AUTO_ALLOW_BASH"
 
+// EnvSandboxed marks a process that zero has already wrapped in a sandbox: every
+// wrapped command carries ZERO_SANDBOXED=1 in its environment. When such a
+// process spawns another command through the engine, the re-entrancy guard
+// returns a pass-through plan instead of double-wrapping it — nested bwrap /
+// sandbox-exec fails, and a second egress proxy would be redundant. Mirrors the
+// already-sandboxed guard used by comparable executor sandboxes. Unset by default.
+const EnvSandboxed = "ZERO_SANDBOXED"
+
+// EnvSandboxBackend records which backend wrapped the command. sandboxEnvironment
+// always sets it alongside EnvSandboxed, so it serves as a corroborating marker:
+// the re-entrancy guard requires BOTH, raising the provenance bar above a single
+// ambient flag (a stray or hand-exported ZERO_SANDBOXED=1 with no backend marker
+// no longer forces an unsandboxed pass-through).
+const EnvSandboxBackend = "ZERO_SANDBOX_BACKEND"
+
+// IsAlreadySandboxed reports whether the current process is already running
+// inside a zero-created sandbox. It requires BOTH correlated markers that
+// sandboxEnvironment sets together — EnvSandboxed == "1" AND a non-empty
+// EnvSandboxBackend — so a single user-set/inherited ZERO_SANDBOXED=1 cannot by
+// itself disable wrapping. zero sets both only on genuinely wrapped commands;
+// pass-through (direct) plans set neither.
+func IsAlreadySandboxed() bool {
+	return os.Getenv(EnvSandboxed) == "1" && strings.TrimSpace(os.Getenv(EnvSandboxBackend)) != ""
+}
+
 type SideEffect string
 type Permission string
 type PermissionMode string
@@ -106,6 +131,11 @@ const (
 	BackendBubblewrap  BackendName = "bubblewrap"
 	BackendSandboxExec BackendName = "sandbox-exec"
 	BackendPolicyOnly  BackendName = "policy-only"
+	// BackendWSL is the policy-only fallback used under WSL when bubblewrap is
+	// unavailable/unreliable: there is no native OS isolation, but network egress
+	// is still routed through the local filtering proxy and the command runs under
+	// the full policy engine. It is surfaced via ZERO_SANDBOX_BACKEND=wsl.
+	BackendWSL BackendName = "wsl"
 )
 
 const (
@@ -133,6 +163,14 @@ type Policy struct {
 	// NetworkDeny so existing policies keep their exact behaviour.
 	AllowedDomains []string `json:"allowedDomains,omitempty"`
 	DeniedDomains  []string `json:"deniedDomains,omitempty"`
+	// EnforceToolNetwork, when true, subjects the first-party in-process network
+	// tools (web_search / web_fetch) to the sandbox network policy via
+	// NetworkHostAllowed. It is OFF by default: that policy exists to confine the
+	// sandboxed SHELL's egress, which these tools do not use, so by default they
+	// are allowed to run (keeping their own SSRF/port/redirect/redaction
+	// safeguards). The sandboxed-shell egress decision is independent of this flag.
+	// Turn it on to also hold web_search/web_fetch to the allow/scoped/deny policy.
+	EnforceToolNetwork bool `json:"enforceToolNetwork,omitempty"`
 	// AutoAllowBashWhenSandboxed, when true, auto-allows the bash tool WITHOUT a
 	// permission prompt — but only when the sandbox is actually active (a
 	// native-isolation backend wraps the command). The sandbox is then the safety
@@ -153,6 +191,40 @@ type Policy struct {
 	// (runs without the filter) when the helper binary is not found. Ignored on
 	// non-bubblewrap backends.
 	BlockUnixSockets bool `json:"blockUnixSockets,omitempty"`
+	// AllowRead/DenyRead/AllowWrite/DenyWrite are fine-grained path lists layered
+	// ON TOP of the workspace + Scope guards; they never bypass the symlink /
+	// out-of-workspace protections. Each entry is home-expanded, made absolute, and
+	// symlink-resolved (an entry that does not exist is dropped). All default empty,
+	// so an unconfigured policy behaves exactly as before. Semantics:
+	//
+	//   - Read: a path readable under the base workspace/Scope guard is denied if it
+	//     falls under a DenyRead entry, UNLESS a more-specific AllowRead entry (one
+	//     nested inside that DenyRead entry) re-includes it. AllowRead only
+	//     re-includes within a DenyRead carve-out; it never extends reads beyond the
+	//     workspace.
+	//   - Write: DenyWrite wins over everything; otherwise a path writable under the
+	//     workspace/Scope guard is allowed; otherwise an absolute path under an
+	//     AllowWrite root is allowed; otherwise it is denied. AllowWrite roots are
+	//     also reflected in the OS backend write binds, and on sandbox-exec DenyWrite
+	//     entries are emitted as explicit deny rules so the precedence holds for
+	//     shell commands too.
+	AllowRead  []string `json:"allowRead,omitempty"`
+	DenyRead   []string `json:"denyRead,omitempty"`
+	AllowWrite []string `json:"allowWrite,omitempty"`
+	DenyWrite  []string `json:"denyWrite,omitempty"`
+	// InspectTLS, when true, lets the in-process scoped-egress proxy TERMINATE TLS
+	// (using a locally generated, ephemeral CA) so it can enforce the per-host
+	// allow/deny rules on the DECRYPTED request Host — today a CONNECT tunnel only
+	// reveals the host once. OFF by default: the proxy stays a pure CONNECT
+	// passthrough, byte-for-byte unchanged. SECURITY/TRUST: enabling this means the
+	// sandbox can read the plaintext of the sandboxed process's TLS traffic. The
+	// MITM only re-signs toward the sandboxed client; the upstream connection is
+	// ALWAYS validated against the system roots (never InsecureSkipVerify), and the
+	// decrypted Host still passes the SAME authorize()/domainAllowed() gate — MITM
+	// widens visibility, never authority. The generated CA's public cert is written
+	// to the sandbox runtime dir and surfaced via ZERO_SANDBOX_CA_CERT so the
+	// sandboxed client can trust it.
+	InspectTLS bool `json:"inspectTLS,omitempty"`
 }
 
 type Request struct {

internal/sandbox/wsl.go

@@ -0,0 +1,26 @@
+package sandbox
+
+import "strings"
+
+// WSLInfo describes the Windows Subsystem for Linux environment, if any. It is
+// derived from /proc/version. Mirrors reference wsl/index.js isWSL/isWSL2 and
+// sandbox-manager.js isWsl (reads /proc/version, matches "microsoft").
+type WSLInfo struct {
+	IsWSL  bool
+	IsWSL2 bool
+	Kernel string
+}
+
+// parseWSL classifies a /proc/version string. IsWSL is set when it contains
+// "microsoft" or "wsl"; IsWSL2 additionally requires "microsoft" and the absence
+// of the legacy "wsl1" marker. Factored out (pure) so it is table-testable
+// without a real /proc on any GOOS.
+func parseWSL(procVersion string) WSLInfo {
+	lower := strings.ToLower(procVersion)
+	isWSL := strings.Contains(lower, "microsoft") || strings.Contains(lower, "wsl")
+	info := WSLInfo{IsWSL: isWSL, Kernel: strings.TrimSpace(procVersion)}
+	if isWSL {
+		info.IsWSL2 = strings.Contains(lower, "microsoft") && !strings.Contains(lower, "wsl1")
+	}
+	return info
+}

internal/sandbox/wsl_linux.go

@@ -0,0 +1,15 @@
+//go:build linux
+
+package sandbox
+
+import "os"
+
+// detectWSL reads /proc/version once and classifies the WSL environment. A read
+// error (no /proc) yields the zero WSLInfo (not WSL).
+func detectWSL() WSLInfo {
+	data, err := os.ReadFile("/proc/version")
+	if err != nil {
+		return WSLInfo{}
+	}
+	return parseWSL(string(data))
+}

internal/sandbox/wsl_other.go

@@ -0,0 +1,8 @@
+//go:build !linux
+
+package sandbox
+
+// detectWSL is a no-op on non-Linux platforms: WSL is a Linux-under-Windows
+// environment, so the host GOOS is never WSL. Keeps the package building on all
+// platforms.
+func detectWSL() WSLInfo { return WSLInfo{} }

internal/sandbox/wsl_test.go

@@ -0,0 +1,139 @@
+package sandbox
+
+import (
+	"errors"
+	"strings"
+	"testing"
+)
+
+func TestParseWSL(t *testing.T) {
+	cases := []struct {
+		name     string
+		input    string
+		wantWSL  bool
+		wantWSL2 bool
+	}{
+		{
+			name:     "wsl2 microsoft kernel",
+			input:    "Linux version 5.15.90.1-microsoft-standard-WSL2 (...)",
+			wantWSL:  true,
+			wantWSL2: true,
+		},
+		{
+			name:     "wsl1 legacy marker",
+			input:    "Linux version 4.4.0-19041-Microsoft (wsl1 build) ...",
+			wantWSL:  true,
+			wantWSL2: false,
+		},
+		{
+			name:     "plain linux",
+			input:    "Linux version 6.8.0-generic (gcc ...) #1 SMP",
+			wantWSL:  false,
+			wantWSL2: false,
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			got := parseWSL(tc.input)
+			if got.IsWSL != tc.wantWSL || got.IsWSL2 != tc.wantWSL2 {
+				t.Fatalf("parseWSL(%q) = {IsWSL:%v IsWSL2:%v}, want {%v %v}", tc.input, got.IsWSL, got.IsWSL2, tc.wantWSL, tc.wantWSL2)
+			}
+			if got.IsWSL && got.Kernel == "" {
+				t.Fatalf("parseWSL should record the kernel string")
+			}
+		})
+	}
+}
+
+func wslBackendForTest() Backend {
+	return Backend{Name: BackendWSL, Platform: "linux", Fallback: true, ProxyEgress: true}
+}
+
+func TestWSLPlanFailsClosedWithoutPolicyOnly(t *testing.T) {
+	root := t.TempDir()
+	policy := scopedPolicy([]string{"github.com"}, nil)
+	policy.AllowPolicyOnlyRunner = false // explicitly refuse the degraded runner
+	engine := NewEngine(EngineOptions{WorkspaceRoot: root, Policy: policy, Backend: wslBackendForTest()})
+
+	_, err := engine.BuildCommandPlan(CommandSpec{Name: "/bin/sh", Args: []string{"-c", "pwd"}, Dir: root})
+	if !errors.Is(err, errWSLPolicyOnlyDisabled) {
+		t.Fatalf("WSL without AllowPolicyOnlyRunner err = %v, want errWSLPolicyOnlyDisabled", err)
+	}
+}
+
+func TestWSLPolicyOnlyPlanCarriesProxyAndMarkers(t *testing.T) {
+	root := t.TempDir()
+	// scopedPolicy uses DefaultPolicy (AllowPolicyOnlyRunner=true) + an allowlist,
+	// so the WSL fallback starts the filtering proxy and runs policy-only.
+	engine := NewEngine(EngineOptions{WorkspaceRoot: root, Policy: scopedPolicy([]string{"github.com"}, nil), Backend: wslBackendForTest()})
+
+	plan, err := engine.BuildCommandPlan(CommandSpec{Name: "/bin/sh", Args: []string{"-c", "pwd"}, Dir: root, Env: []string{}})
+	if err != nil {
+		t.Fatalf("BuildCommandPlan: %v", err)
+	}
+	defer plan.Cleanup()
+
+	// Policy-only: the command runs directly, not OS-wrapped.
+	if plan.Wrapped || plan.Name != "/bin/sh" {
+		t.Fatalf("WSL plan must run the command directly, got wrapped=%v name=%q", plan.Wrapped, plan.Name)
+	}
+	env := strings.Join(plan.Env, "\n")
+	if !strings.Contains(env, EnvSandboxed+"=1") {
+		t.Fatalf("WSL plan env missing %s=1:\n%s", EnvSandboxed, env)
+	}
+	if !strings.Contains(env, EnvSandboxBackend+"=wsl") {
+		t.Fatalf("WSL plan env missing %s=wsl:\n%s", EnvSandboxBackend, env)
+	}
+	if !strings.Contains(env, "HTTP_PROXY=") || !strings.Contains(env, "ALL_PROXY=") {
+		t.Fatalf("WSL plan env missing proxy env (ProxyEnvWithSocks):\n%s", env)
+	}
+	if len(plan.Notes) == 0 || !strings.Contains(strings.Join(plan.Notes, " "), "policy-only") {
+		t.Fatalf("WSL plan must record a least-privilege downgrade note, got %v", plan.Notes)
+	}
+}
+
+func TestWSLPlanPreservesCallerEnv(t *testing.T) {
+	root := t.TempDir()
+	engine := NewEngine(EngineOptions{WorkspaceRoot: root, Policy: DefaultPolicy(), Backend: wslBackendForTest()})
+	plan, err := engine.BuildCommandPlan(CommandSpec{
+		Name: "/bin/sh", Args: []string{"-c", "true"}, Dir: root,
+		Env: []string{"FOO=bar", "PATH=/custom/bin"},
+	})
+	if err != nil {
+		t.Fatalf("BuildCommandPlan: %v", err)
+	}
+	defer plan.Cleanup()
+	env := strings.Join(plan.Env, "\n")
+	// The WSL direct-run inherits the caller's env (it is not OS-wrapped)...
+	if !strings.Contains(env, "FOO=bar") || !strings.Contains(env, "PATH=/custom/bin") {
+		t.Fatalf("WSL plan must preserve caller env, got:\n%s", env)
+	}
+	// ...with the sandbox markers appended (so they win over caller env).
+	if !strings.Contains(env, EnvSandboxed+"=1") || !strings.Contains(env, EnvSandboxBackend+"=wsl") {
+		t.Fatalf("WSL plan must append sandbox markers:\n%s", env)
+	}
+}
+
+func TestWSLPolicyOnlyDenyPlanHasMarkersNoProxy(t *testing.T) {
+	root := t.TempDir()
+	// Default policy (NetworkDeny) → no proxy started, but the WSL markers + note
+	// still apply and the plan still runs (AllowPolicyOnlyRunner default true).
+	engine := NewEngine(EngineOptions{WorkspaceRoot: root, Policy: DefaultPolicy(), Backend: wslBackendForTest()})
+	// Explicit empty env so the assertion below is not confused by a dev/CI
+	// HTTP_PROXY inherited via os.Environ().
+	plan, err := engine.BuildCommandPlan(CommandSpec{Name: "/bin/sh", Args: []string{"-c", "pwd"}, Dir: root, Env: []string{}})
+	if err != nil {
+		t.Fatalf("BuildCommandPlan: %v", err)
+	}
+	defer plan.Cleanup()
+	env := strings.Join(plan.Env, "\n")
+	if !strings.Contains(env, EnvSandboxBackend+"=wsl") || !strings.Contains(env, EnvSandboxed+"=1") {
+		t.Fatalf("WSL deny plan must still carry sandbox markers:\n%s", env)
+	}
+	if len(plan.Notes) == 0 || !strings.Contains(strings.Join(plan.Notes, " "), "policy-only") {
+		t.Fatalf("WSL deny plan must record a least-privilege downgrade note, got %v", plan.Notes)
+	}
+	if strings.Contains(env, "HTTP_PROXY=") || strings.Contains(env, "HTTPS_PROXY=") || strings.Contains(env, "ALL_PROXY=") {
+		t.Fatalf("WSL deny plan must NOT export any proxy env vars (network denied):\n%s", env)
+	}
+}

internal/specialist/exec.go

@@ -58,14 +58,20 @@ type BuildArgsInput struct {
 	CurrentDepth          int
 	Description           string
 	Cwd                   string
+	// PermissionMode is the parent's resolved permission mode. Empty preserves
+	// the historical "--auto high" (the Task tool omits it); a caller that
+	// passes a non-unsafe mode (e.g. the swarm propagating a non-unsafe
+	// orchestrator) gets a non-unsafe child so authority is never widened.
+	PermissionMode string
 }
 
 type BuildResumeArgsInput struct {
-	SessionID    string
-	Prompt       string
-	CurrentDepth int
-	Manifest     Manifest
-	Cwd          string
+	SessionID      string
+	Prompt         string
+	CurrentDepth   int
+	Manifest       Manifest
+	Cwd            string
+	PermissionMode string
 }
 
 type BuildArgsResult struct {
@@ -90,8 +96,32 @@ type TaskRunOptions struct {
 	ParentReasoningEffort string
 	CurrentDepth          int
 	Cwd                   string
+	// PermissionMode propagates the parent's resolved permission mode to the
+	// child. Empty keeps the historical "--auto high"; a non-unsafe value runs
+	// the child non-unsafe so it never gains more authority than the parent.
+	PermissionMode string
 }
 
+// specialistAutonomy maps a parent permission mode to the child's "--auto"
+// level. A headless specialist child cannot answer interactive prompts, so it
+// runs autonomously — but only at "high" (unsafe) when the parent is itself
+// unsafe (or when no mode is provided, preserving the Task tool's behavior).
+// Any other parent mode yields the non-unsafe "low" level, so the child's
+// authority never exceeds the parent's.
+func specialistAutonomy(permissionMode string) string {
+	switch strings.TrimSpace(permissionMode) {
+	case "", string(permissionModeUnsafe):
+		return "high"
+	default:
+		return "low"
+	}
+}
+
+// permissionModeUnsafe mirrors agent.PermissionModeUnsafe without importing the
+// agent package (which would create an import cycle): exec resolves "--auto high"
+// to this mode.
+const permissionModeUnsafe = "unsafe"
+
 type ExecResult struct {
 	Result    tools.Result
 	SessionID string
@@ -147,7 +177,7 @@ func (executor Executor) BuildArgs(input BuildArgsInput) (BuildArgsResult, error
 	args := []string{"exec", "--init-session-id", sessionID}
 	args = append(args, promptArgs...)
 	args = appendModelArgs(args, input.Manifest, input.ParentModel, input.ParentReasoningEffort)
-	args = append(args, "--auto", "high", "--output-format", "stream-json")
+	args = append(args, "--auto", specialistAutonomy(input.PermissionMode), "--output-format", "stream-json")
 	toolAllowlist, err := resolvedToolAllowlist(input.Manifest)
 	if err != nil {
 		return BuildArgsResult{}, err
@@ -199,7 +229,7 @@ func (executor Executor) BuildResumeArgs(input BuildResumeArgsInput) (BuildArgsR
 	}
 	args := []string{"exec", "--resume", sessionID}
 	args = append(args, promptArgs...)
-	args = append(args, "--auto", "high", "--output-format", "stream-json")
+	args = append(args, "--auto", specialistAutonomy(input.PermissionMode), "--output-format", "stream-json")
 	toolAllowlist, err := resolvedToolAllowlist(input.Manifest)
 	if err != nil {
 		return BuildArgsResult{}, err
@@ -230,6 +260,7 @@ func (executor Executor) runFresh(ctx context.Context, params TaskParameters, op
 		CurrentDepth:          options.CurrentDepth,
 		Description:           params.Description,
 		Cwd:                   options.Cwd,
+		PermissionMode:        options.PermissionMode,
 	})
 	if err != nil {
 		return ExecResult{}, err
@@ -257,11 +288,12 @@ func (executor Executor) runResume(ctx context.Context, params TaskParameters, o
 		return ExecResult{}, err
 	}
 	built, err := executor.BuildResumeArgs(BuildResumeArgsInput{
-		SessionID:    params.Resume,
-		Prompt:       params.Prompt,
-		CurrentDepth: options.CurrentDepth,
-		Manifest:     manifest,
-		Cwd:          options.Cwd,
+		SessionID:      params.Resume,
+		Prompt:         params.Prompt,
+		CurrentDepth:   options.CurrentDepth,
+		Manifest:       manifest,
+		Cwd:            options.Cwd,
+		PermissionMode: options.PermissionMode,
 	})
 	if err != nil {
 		return ExecResult{}, err

internal/specialist/exec_test.go

@@ -64,6 +64,65 @@ func TestBuildArgsCreatesFreshSpecialistExecInvocation(t *testing.T) {
 	}
 }
 
+func TestSpecialistAutonomyByPermissionMode(t *testing.T) {
+	cases := map[string]string{
+		"":           "high", // back-compat: the Task tool omits the mode
+		"unsafe":     "high",
+		"auto":       "low",
+		"ask":        "low",
+		"spec-draft": "low",
+		"whatever":   "low",
+	}
+	for mode, want := range cases {
+		if got := specialistAutonomy(mode); got != want {
+			t.Errorf("specialistAutonomy(%q) = %q, want %q", mode, got, want)
+		}
+	}
+}
+
+func TestBuildArgsAutonomyHonorsPermissionMode(t *testing.T) {
+	executor := Executor{NewSessionID: func() (string, error) { return "child", nil }}
+	manifest := Manifest{Metadata: Metadata{Name: "reviewer"}, SystemPrompt: "x", ResolvedTools: []string{"read_file"}}
+
+	// A non-unsafe parent mode yields a non-unsafe child (--auto low), so a
+	// swarm member never gains more authority than a non-unsafe orchestrator.
+	res, err := executor.BuildArgs(BuildArgsInput{Manifest: manifest, Prompt: "p", PermissionMode: "auto"})
+	if err != nil {
+		t.Fatalf("BuildArgs: %v", err)
+	}
+	if !containsSequence(res.Args, []string{"--auto", "low"}) {
+		t.Fatalf("non-unsafe parent must yield --auto low, got %v", res.Args)
+	}
+	if containsSequence(res.Args, []string{"--auto", "high"}) {
+		t.Fatalf("non-unsafe parent must NOT yield --auto high: %v", res.Args)
+	}
+
+	// An unsafe parent (and the empty back-compat case) keep --auto high.
+	for _, mode := range []string{"unsafe", ""} {
+		out, err := executor.BuildArgs(BuildArgsInput{Manifest: manifest, Prompt: "p", PermissionMode: mode})
+		if err != nil {
+			t.Fatalf("BuildArgs(%q): %v", mode, err)
+		}
+		if !containsSequence(out.Args, []string{"--auto", "high"}) {
+			t.Fatalf("mode %q must yield --auto high, got %v", mode, out.Args)
+		}
+	}
+}
+
+func TestBuildResumeArgsAutonomyHonorsPermissionMode(t *testing.T) {
+	executor := Executor{}
+	manifest := Manifest{Metadata: Metadata{Name: "reviewer"}, ResolvedTools: []string{"read_file"}}
+	res, err := executor.BuildResumeArgs(BuildResumeArgsInput{
+		SessionID: "child_session", Prompt: "p", Manifest: manifest, PermissionMode: "auto",
+	})
+	if err != nil {
+		t.Fatalf("BuildResumeArgs: %v", err)
+	}
+	if !containsSequence(res.Args, []string{"--auto", "low"}) || containsSequence(res.Args, []string{"--auto", "high"}) {
+		t.Fatalf("resume non-unsafe parent must yield --auto low, got %v", res.Args)
+	}
+}
+
 func TestBuildArgsSessionTitleFallsBackToNameWithoutDescription(t *testing.T) {
 	executor := Executor{NewSessionID: func() (string, error) { return "child", nil }}
 	manifest := Manifest{

internal/swarm/coordinator.go

@@ -0,0 +1,259 @@
+package swarm
+
+import (
+	"errors"
+	"fmt"
+	"sort"
+	"sync"
+	"time"
+)
+
+// TaskStatus is the lifecycle state of a swarm task.
+type TaskStatus string
+
+const (
+	// StatusPending: registered but the member has not started.
+	StatusPending TaskStatus = "pending"
+	// StatusRunning: the member is executing.
+	StatusRunning TaskStatus = "running"
+	// StatusDone: the member finished successfully.
+	StatusDone TaskStatus = "done"
+	// StatusFailed: the member exited with an error.
+	StatusFailed TaskStatus = "failed"
+	// StatusHandedOff: ownership was transferred to another agent.
+	StatusHandedOff TaskStatus = "handed-off"
+)
+
+// terminal reports whether a status is final (no further transitions expected).
+func (s TaskStatus) terminal() bool {
+	return s == StatusDone || s == StatusFailed || s == StatusHandedOff
+}
+
+// valid reports whether a status is one of the known lifecycle states. Unknown
+// statuses are rejected (fail closed) so a task can never enter an undefined
+// state that Summarize and the lifecycle logic cannot reason about.
+func (s TaskStatus) valid() bool {
+	switch s {
+	case StatusPending, StatusRunning, StatusDone, StatusFailed, StatusHandedOff:
+		return true
+	default:
+		return false
+	}
+}
+
+// Task is one unit of work tracked by the coordinator. It is returned by value
+// from snapshots so callers cannot mutate coordinator state without the lock.
+type Task struct {
+	ID          string
+	AgentID     string
+	Team        string
+	Description string
+	Status      TaskStatus
+	Result      string
+	Err         string
+	CreatedAt   time.Time
+	UpdatedAt   time.Time
+}
+
+// agentColors palette for stable per-agent coloring in the TUI/status. Provider
+// -neutral ANSI-ish names; the renderer maps them to styles.
+var agentColors = []string{"cyan", "magenta", "green", "yellow", "blue", "red"}
+
+// Coordinator is the in-memory task registry + team color assigner shared by an
+// orchestrator and its members. It is safe for concurrent use.
+type Coordinator struct {
+	mu         sync.RWMutex
+	tasks      map[string]*Task
+	colors     map[string]string // agentID -> color
+	colorIndex int
+	now        func() time.Time // injectable clock for tests
+}
+
+// NewCoordinator returns an empty coordinator using the wall clock.
+func NewCoordinator() *Coordinator {
+	return &Coordinator{
+		tasks:  map[string]*Task{},
+		colors: map[string]string{},
+		now:    time.Now,
+	}
+}
+
+// ErrTaskExists is returned when registering a duplicate task ID.
+var ErrTaskExists = errors.New("swarm: task already registered")
+
+// ErrUnknownTask is returned when updating/removing a missing task ID.
+var ErrUnknownTask = errors.New("swarm: unknown task")
+
+// Register adds a new pending task. The ID must be unique and non-empty.
+func (c *Coordinator) Register(id, agentID, team, description string) (Task, error) {
+	if id == "" {
+		return Task{}, errors.New("swarm: task id is required")
+	}
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	if _, ok := c.tasks[id]; ok {
+		return Task{}, fmt.Errorf("%w: %s", ErrTaskExists, id)
+	}
+	now := c.now()
+	t := &Task{
+		ID:          id,
+		AgentID:     agentID,
+		Team:        team,
+		Description: description,
+		Status:      StatusPending,
+		CreatedAt:   now,
+		UpdatedAt:   now,
+	}
+	c.tasks[id] = t
+	c.assignColorLocked(agentID)
+	return *t, nil
+}
+
+// SetStatus transitions a task. Transitions out of a terminal state are
+// rejected (fail closed) so a late member update cannot resurrect a finished
+// task.
+func (c *Coordinator) SetStatus(id string, status TaskStatus) error {
+	if !status.valid() {
+		return fmt.Errorf("swarm: invalid task status %q", status)
+	}
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	t, ok := c.tasks[id]
+	if !ok {
+		return fmt.Errorf("%w: %s", ErrUnknownTask, id)
+	}
+	if t.Status.terminal() && t.Status != status {
+		return fmt.Errorf("swarm: task %s is %s (terminal); cannot move to %s", id, t.Status, status)
+	}
+	t.Status = status
+	t.UpdatedAt = c.now()
+	return nil
+}
+
+// Complete marks a task done with its result.
+func (c *Coordinator) Complete(id, result string) error {
+	return c.finish(id, StatusDone, result, "")
+}
+
+// Fail marks a task failed with an error message.
+func (c *Coordinator) Fail(id, errMsg string) error {
+	return c.finish(id, StatusFailed, "", errMsg)
+}
+
+func (c *Coordinator) finish(id string, status TaskStatus, result, errMsg string) error {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	t, ok := c.tasks[id]
+	if !ok {
+		return fmt.Errorf("%w: %s", ErrUnknownTask, id)
+	}
+	if t.Status.terminal() {
+		return fmt.Errorf("swarm: task %s already %s", id, t.Status)
+	}
+	t.Status = status
+	t.Result = result
+	t.Err = errMsg
+	t.UpdatedAt = c.now()
+	return nil
+}
+
+// Reassign transfers a task to a new agent (handoff/orphan-adoption). The task
+// returns to pending under the new owner unless it is already terminal.
+func (c *Coordinator) Reassign(id, newAgentID string) error {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	t, ok := c.tasks[id]
+	if !ok {
+		return fmt.Errorf("%w: %s", ErrUnknownTask, id)
+	}
+	if t.Status.terminal() {
+		return fmt.Errorf("swarm: task %s already %s; cannot reassign", id, t.Status)
+	}
+	t.AgentID = newAgentID
+	t.Status = StatusPending
+	t.UpdatedAt = c.now()
+	c.assignColorLocked(newAgentID)
+	return nil
+}
+
+// Get returns a snapshot of one task.
+func (c *Coordinator) Get(id string) (Task, bool) {
+	c.mu.RLock()
+	defer c.mu.RUnlock()
+	t, ok := c.tasks[id]
+	if !ok {
+		return Task{}, false
+	}
+	return *t, true
+}
+
+// List returns snapshots of all tasks, ordered by creation time then ID for
+// stable output.
+func (c *Coordinator) List() []Task {
+	c.mu.RLock()
+	out := make([]Task, 0, len(c.tasks))
+	for _, t := range c.tasks {
+		out = append(out, *t)
+	}
+	c.mu.RUnlock()
+	sort.Slice(out, func(i, j int) bool {
+		if out[i].CreatedAt.Equal(out[j].CreatedAt) {
+			return out[i].ID < out[j].ID
+		}
+		return out[i].CreatedAt.Before(out[j].CreatedAt)
+	})
+	return out
+}
+
+// Color returns the stable color assigned to an agent (assigning one on first
+// use), so status output colors each member consistently.
+func (c *Coordinator) Color(agentID string) string {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	return c.assignColorLocked(agentID)
+}
+
+func (c *Coordinator) assignColorLocked(agentID string) string {
+	if agentID == "" {
+		return ""
+	}
+	if color, ok := c.colors[agentID]; ok {
+		return color
+	}
+	color := agentColors[c.colorIndex%len(agentColors)]
+	c.colors[agentID] = color
+	c.colorIndex++
+	return color
+}
+
+// Summary is an aggregate count of tasks by status for the status tool.
+type Summary struct {
+	Total     int
+	Pending   int
+	Running   int
+	Done      int
+	Failed    int
+	HandedOff int
+}
+
+// Summarize aggregates current task counts.
+func (c *Coordinator) Summarize() Summary {
+	c.mu.RLock()
+	defer c.mu.RUnlock()
+	s := Summary{Total: len(c.tasks)}
+	for _, t := range c.tasks {
+		switch t.Status {
+		case StatusPending:
+			s.Pending++
+		case StatusRunning:
+			s.Running++
+		case StatusDone:
+			s.Done++
+		case StatusFailed:
+			s.Failed++
+		case StatusHandedOff:
+			s.HandedOff++
+		}
+	}
+	return s
+}

internal/swarm/coordinator_test.go

@@ -0,0 +1,151 @@
+package swarm
+
+import (
+	"errors"
+	"sync"
+	"testing"
+)
+
+func TestCoordinatorRegisterAndDuplicate(t *testing.T) {
+	c := NewCoordinator()
+	if _, err := c.Register("t1", "a1", "team", "desc"); err != nil {
+		t.Fatalf("Register: %v", err)
+	}
+	task, ok := c.Get("t1")
+	if !ok || task.Status != StatusPending {
+		t.Fatalf("Get after register: ok=%v status=%v", ok, task.Status)
+	}
+	if _, err := c.Register("t1", "a1", "team", "desc"); !errors.Is(err, ErrTaskExists) {
+		t.Fatalf("duplicate Register err = %v, want ErrTaskExists", err)
+	}
+	if _, err := c.Register("", "a", "team", "d"); err == nil {
+		t.Fatal("empty id must error")
+	}
+}
+
+func TestCoordinatorLifecycleTransitions(t *testing.T) {
+	c := NewCoordinator()
+	_, _ = c.Register("t1", "a1", "team", "desc")
+	if err := c.SetStatus("t1", StatusRunning); err != nil {
+		t.Fatalf("SetStatus running: %v", err)
+	}
+	if err := c.Complete("t1", "result-data"); err != nil {
+		t.Fatalf("Complete: %v", err)
+	}
+	task, _ := c.Get("t1")
+	if task.Status != StatusDone || task.Result != "result-data" {
+		t.Fatalf("after Complete: status=%v result=%q", task.Status, task.Result)
+	}
+	// Terminal guard: cannot move done -> running, nor complete twice.
+	if err := c.SetStatus("t1", StatusRunning); err == nil {
+		t.Fatal("moving a done task to running must fail")
+	}
+	if err := c.Complete("t1", "x"); err == nil {
+		t.Fatal("completing a done task must fail")
+	}
+	if err := c.Fail("t1", "boom"); err == nil {
+		t.Fatal("failing a done task must fail")
+	}
+}
+
+func TestCoordinatorRejectsInvalidStatus(t *testing.T) {
+	c := NewCoordinator()
+	_, _ = c.Register("t1", "a1", "team", "desc")
+	if err := c.SetStatus("t1", TaskStatus("bogus")); err == nil {
+		t.Fatal("SetStatus must reject an unknown status")
+	}
+	// The task stays in its prior (valid) state.
+	if task, _ := c.Get("t1"); task.Status != StatusPending {
+		t.Fatalf("rejected status must not mutate the task, got %v", task.Status)
+	}
+}
+
+func TestCoordinatorUnknownTask(t *testing.T) {
+	c := NewCoordinator()
+	if err := c.SetStatus("missing", StatusRunning); !errors.Is(err, ErrUnknownTask) {
+		t.Fatalf("SetStatus(missing) = %v, want ErrUnknownTask", err)
+	}
+	if err := c.Complete("missing", "x"); !errors.Is(err, ErrUnknownTask) {
+		t.Fatalf("Complete(missing) = %v, want ErrUnknownTask", err)
+	}
+}
+
+func TestCoordinatorReassign(t *testing.T) {
+	c := NewCoordinator()
+	_, _ = c.Register("t1", "a1", "team", "desc")
+	_ = c.SetStatus("t1", StatusRunning)
+	if err := c.Reassign("t1", "a2"); err != nil {
+		t.Fatalf("Reassign: %v", err)
+	}
+	task, _ := c.Get("t1")
+	if task.AgentID != "a2" || task.Status != StatusPending {
+		t.Fatalf("after Reassign: agent=%q status=%v", task.AgentID, task.Status)
+	}
+	// Cannot reassign a terminal task.
+	_ = c.Complete("t1", "done")
+	if err := c.Reassign("t1", "a3"); err == nil {
+		t.Fatal("reassigning a terminal task must fail")
+	}
+}
+
+func TestCoordinatorColorStability(t *testing.T) {
+	c := NewCoordinator()
+	first := c.Color("a1")
+	if first == "" {
+		t.Fatal("Color should assign a non-empty color")
+	}
+	if again := c.Color("a1"); again != first {
+		t.Fatalf("Color not stable: %q vs %q", first, again)
+	}
+	if other := c.Color("a2"); other == "" {
+		t.Fatal("second agent should also get a color")
+	}
+	if c.Color("") != "" {
+		t.Fatal("empty agent id should get no color")
+	}
+}
+
+func TestCoordinatorSummarize(t *testing.T) {
+	c := NewCoordinator()
+	_, _ = c.Register("p", "a", "team", "")
+	_, _ = c.Register("r", "b", "team", "")
+	_ = c.SetStatus("r", StatusRunning)
+	_, _ = c.Register("d", "c", "team", "")
+	_ = c.Complete("d", "ok")
+	s := c.Summarize()
+	if s.Total != 3 || s.Pending != 1 || s.Running != 1 || s.Done != 1 {
+		t.Fatalf("Summarize = %+v", s)
+	}
+}
+
+func TestCoordinatorConcurrentRegister(t *testing.T) {
+	c := NewCoordinator()
+	var wg sync.WaitGroup
+	for i := 0; i < 50; i++ {
+		wg.Add(1)
+		go func(n int) {
+			defer wg.Done()
+			id := "task-" + string(rune('a'+n%26)) + "-" + itoa(n)
+			_, _ = c.Register(id, id, "team", "desc")
+			_ = c.SetStatus(id, StatusRunning)
+			_ = c.Complete(id, "done")
+		}(i)
+	}
+	wg.Wait()
+	if got := len(c.List()); got != 50 {
+		t.Fatalf("List len = %d, want 50", got)
+	}
+}
+
+// itoa avoids importing strconv just for the concurrency test id.
+func itoa(n int) string {
+	if n == 0 {
+		return "0"
+	}
+	var b []byte
+	for n > 0 {
+		b = append([]byte{byte('0' + n%10)}, b...)
+		n /= 10
+	}
+	return string(b)
+}

internal/swarm/definition.go

@@ -0,0 +1,133 @@
+// Package swarm adds a multi-agent SWARM on top of zero's single sub-agent
+// mechanism (internal/specialist): an orchestrator can spawn and coordinate
+// MULTIPLE specialist members that run concurrently, communicate via per-agent
+// mailboxes, and hand work off. It composes internal/specialist (to launch each
+// member), internal/daemon (pooled workers when a daemon is running) and
+// internal/background. Everything is additive and opt-in — the existing single
+// "task" tool is unchanged; swarm tools are only active when invoked.
+//
+// Every member runs under the SAME sandbox + risk/autonomy policy as the
+// orchestrator: the swarm never grants a member more authority than its parent.
+package swarm
+
+import (
+	"errors"
+	"sort"
+	"strings"
+	"sync"
+)
+
+// ErrUnknownAgentType is returned when a definition lookup misses.
+var ErrUnknownAgentType = errors.New("swarm: unknown agent type")
+
+// modelInherit is the sentinel meaning "use the orchestrator's model" — members
+// never silently pick a different (e.g. more capable) model.
+const modelInherit = "inherit"
+
+// Definition is one entry in the agent roster: how a member of a given type
+// behaves (agent type, when to use it, model — "inherit" by default —,
+// permission mode, and a system-prompt builder).
+type Definition struct {
+	AgentType      string
+	WhenToUse      string
+	Model          string // "inherit" => use the orchestrator's model
+	PermissionMode string
+	// SystemPrompt returns the member's system prompt for the given task context.
+	// It is a func so a definition can fold the task briefing in.
+	SystemPrompt func(ctx PromptContext) string
+}
+
+// PromptContext is the minimal context handed to a definition's SystemPrompt.
+type PromptContext struct {
+	Team string
+	Task string
+}
+
+// Registry is the agent roster: definitions looked up by agent type. Built-ins
+// are seeded at construction; user-defined definitions extend or override them.
+type Registry struct {
+	mu   sync.RWMutex
+	defs map[string]Definition
+}
+
+// NewRegistry builds a registry seeded with the built-in roster.
+func NewRegistry() *Registry {
+	r := &Registry{defs: map[string]Definition{}}
+	for _, def := range builtinDefinitions() {
+		r.defs[def.AgentType] = def
+	}
+	return r
+}
+
+// Register adds or overrides a definition (user-defined agents extend the
+// built-ins). An empty AgentType is rejected.
+func (r *Registry) Register(def Definition) error {
+	agentType := strings.TrimSpace(def.AgentType)
+	if agentType == "" {
+		return errors.New("swarm: definition requires an agentType")
+	}
+	if def.Model == "" {
+		def.Model = modelInherit
+	}
+	def.AgentType = agentType
+	r.mu.Lock()
+	r.defs[agentType] = def
+	r.mu.Unlock()
+	return nil
+}
+
+// Lookup returns the definition for agentType, or ErrUnknownAgentType.
+func (r *Registry) Lookup(agentType string) (Definition, error) {
+	r.mu.RLock()
+	def, ok := r.defs[strings.TrimSpace(agentType)]
+	r.mu.RUnlock()
+	if !ok {
+		return Definition{}, ErrUnknownAgentType
+	}
+	return def, nil
+}
+
+// AgentTypes returns the registered agent types, sorted, for status/help.
+func (r *Registry) AgentTypes() []string {
+	r.mu.RLock()
+	types := make([]string, 0, len(r.defs))
+	for t := range r.defs {
+		types = append(types, t)
+	}
+	r.mu.RUnlock()
+	sort.Strings(types)
+	return types
+}
+
+// builtinDefinitions returns the seeded roster (teammate, subagent) — kept
+// minimal; user-defined agents extend it via Register.
+func builtinDefinitions() []Definition {
+	return []Definition{
+		{
+			AgentType: "teammate",
+			WhenToUse: "In-process teammate for parallel task execution; delegate work to run alongside the orchestrator.",
+			Model:     modelInherit,
+			// Empty PermissionMode => inherit the orchestrator's mode (never widened).
+			SystemPrompt: func(ctx PromptContext) string {
+				return "You are a teammate agent collaborating with an orchestrator on team " +
+					displayTeam(ctx.Team) + ". Complete your assigned task and report results.\n\nTask: " + ctx.Task
+			},
+		},
+		{
+			AgentType: "subagent",
+			WhenToUse: "General-purpose subagent for an isolated, delegated task; starts with zero prior context.",
+			Model:     modelInherit,
+			SystemPrompt: func(ctx PromptContext) string {
+				return "You are a subagent spawned to complete a specific task. You start with zero context — " +
+					"the briefing below is all you know.\n\nTask: " + ctx.Task
+			},
+		},
+	}
+}
+
+func displayTeam(team string) string {
+	if strings.TrimSpace(team) == "" {
+		return "default"
+	}
+	return team
+}

internal/swarm/definition_test.go

@@ -0,0 +1,74 @@
+package swarm
+
+import (
+	"errors"
+	"testing"
+)
+
+func TestRegistryBuiltins(t *testing.T) {
+	r := NewRegistry()
+	for _, want := range []string{"teammate", "subagent"} {
+		def, err := r.Lookup(want)
+		if err != nil {
+			t.Fatalf("builtin %q missing: %v", want, err)
+		}
+		if def.Model != modelInherit {
+			t.Fatalf("builtin %q model = %q, want inherit", want, def.Model)
+		}
+		if def.SystemPrompt == nil {
+			t.Fatalf("builtin %q must have a SystemPrompt", want)
+		}
+		if got := def.SystemPrompt(PromptContext{Team: "t", Task: "do x"}); got == "" {
+			t.Fatalf("builtin %q SystemPrompt returned empty", want)
+		}
+	}
+}
+
+func TestRegistryLookupUnknown(t *testing.T) {
+	r := NewRegistry()
+	if _, err := r.Lookup("nope"); !errors.Is(err, ErrUnknownAgentType) {
+		t.Fatalf("Lookup(nope) err = %v, want ErrUnknownAgentType", err)
+	}
+}
+
+func TestRegistryRegisterOverrideAndDefaultModel(t *testing.T) {
+	r := NewRegistry()
+	if err := r.Register(Definition{AgentType: "  custom  "}); err != nil {
+		t.Fatalf("Register: %v", err)
+	}
+	def, err := r.Lookup("custom")
+	if err != nil {
+		t.Fatalf("Lookup(custom): %v", err)
+	}
+	if def.AgentType != "custom" {
+		t.Fatalf("agentType not trimmed: %q", def.AgentType)
+	}
+	if def.Model != modelInherit {
+		t.Fatalf("default model = %q, want inherit", def.Model)
+	}
+	// Override a builtin.
+	if err := r.Register(Definition{AgentType: "teammate", Model: "fixed-model"}); err != nil {
+		t.Fatalf("override Register: %v", err)
+	}
+	if got, _ := r.Lookup("teammate"); got.Model != "fixed-model" {
+		t.Fatalf("override model = %q, want fixed-model", got.Model)
+	}
+}
+
+func TestRegistryRegisterRejectsEmpty(t *testing.T) {
+	r := NewRegistry()
+	if err := r.Register(Definition{AgentType: "   "}); err == nil {
+		t.Fatal("Register with blank agentType must error")
+	}
+}
+
+func TestRegistryAgentTypesSorted(t *testing.T) {
+	r := NewRegistry()
+	_ = r.Register(Definition{AgentType: "alpha"})
+	types := r.AgentTypes()
+	for i := 1; i < len(types); i++ {
+		if types[i-1] > types[i] {
+			t.Fatalf("AgentTypes not sorted: %v", types)
+		}
+	}
+}

internal/swarm/launcher_specialist.go

@@ -0,0 +1,47 @@
+package swarm
+
+import (
+	"context"
+
+	"github.com/Gitlawb/zero/internal/specialist"
+)
+
+// NewSpecialistLauncher adapts internal/specialist.Executor into a
+// MemberLauncher: each swarm member runs as a specialist sub-agent. This is the
+// production "direct" launch path (used when no daemon pool is wired). The
+// member inherits the orchestrator's model via spec.Model and runs in spec.Cwd,
+// so it stays under the same sandbox + policy as its parent — the swarm never
+// grants a member more authority.
+//
+// agentType maps to the specialist agent name; the specialist's own definition
+// for that name governs the member's system prompt and tools. RunInBackground is
+// deliberately left false: the swarm provides concurrency itself (one goroutine
+// per member), so each Run executes to completion inside its goroutine.
+//
+// spec.PermissionMode (resolved + clamped by buildSpec, never above the
+// orchestrator's) is propagated so the child only runs unsafe/high autonomy when
+// the orchestrator is itself unsafe; otherwise it runs non-unsafe. An empty mode
+// is substituted with the non-unsafe "auto" default so a swarm member can never
+// fall into the historical empty-means-high path — that is the Task tool's
+// behavior, not the swarm's.
+func NewSpecialistLauncher(executor specialist.Executor) MemberLauncher {
+	return FuncLauncher{Run: func(ctx context.Context, spec MemberSpec) (MemberResult, error) {
+		permissionMode := spec.PermissionMode
+		if permissionMode == "" {
+			permissionMode = permissionModeAuto
+		}
+		res, err := executor.Run(ctx, specialist.TaskParameters{
+			Name:        spec.AgentType,
+			Prompt:      spec.Task,
+			Description: spec.AgentType + " · team " + spec.Team,
+		}, specialist.TaskRunOptions{
+			ParentModel:    spec.Model,
+			Cwd:            spec.Cwd,
+			PermissionMode: permissionMode,
+		})
+		if err != nil {
+			return MemberResult{}, err
+		}
+		return MemberResult{Result: res.Result.Output, SessionID: res.SessionID}, nil
+	}}
+}

internal/swarm/lifecycle.go

@@ -0,0 +1,176 @@
+package swarm
+
+import (
+	"fmt"
+)
+
+// Spawn registers a task and launches a member of agentType to run it under the
+// given team, inheriting the orchestrator's policy. It returns the task id
+// immediately; the member runs concurrently (under the Swarm's base context, so
+// it outlives this call) and reports its result through the coordinator. If the
+// team is at its slot cap the member is queued and launches when a slot frees
+// (it stays pending in the coordinator until then).
+func (s *Swarm) Spawn(pol Policy, teamName, agentType, task, cwd string) (string, error) {
+	def, err := s.registry.Lookup(agentType)
+	if err != nil {
+		return "", err
+	}
+	team := sanitizeName(teamName)
+	id := s.nextID(agentType)
+	if _, err := s.coord.Register(id, id, team, task); err != nil {
+		return "", err
+	}
+	s.rememberCwd(id, cwd)
+	spec := s.buildSpec(pol, id, id, team, def, task, cwd)
+	s.dispatch(spec)
+	return id, nil
+}
+
+// dispatch admits a spec to its team (launching now or queuing for a slot).
+func (s *Swarm) dispatch(spec MemberSpec) {
+	t := s.team(spec.Team)
+	if t.admit(spec) {
+		s.launch(t, spec)
+	}
+	// Otherwise the spec is queued; the coordinator task stays pending until a
+	// slot frees and afterExit launches it.
+}
+
+// launch starts a member for spec and supervises it. A synchronous launch
+// failure fails the task and frees the slot.
+func (s *Swarm) launch(t *Team, spec MemberSpec) {
+	handle, err := s.launcher.Launch(s.baseCtx, spec)
+	if err != nil {
+		_ = s.coord.Fail(spec.TaskID, "launch: "+err.Error())
+		s.afterExit(t)
+		return
+	}
+	m := &Member{ID: spec.ID, AgentType: spec.AgentType, TaskID: spec.TaskID, handle: handle}
+	t.addMember(m)
+	_ = s.coord.SetStatus(spec.TaskID, StatusRunning)
+	go s.watch(t, m, spec)
+}
+
+// watch awaits a member, applies bounded relaunch on temporary failures, records
+// the terminal outcome, then frees the slot and drains the queue.
+//
+// The relaunch counter is per-member (m.restarts) and the same spec is reused on
+// retry. This is sound because a Member is bound 1:1 to its spec for its whole
+// life (Member.ID == MemberSpec.ID); a retry never reuses the struct for a
+// different spec.
+func (s *Swarm) watch(t *Team, m *Member, spec MemberSpec) {
+	res, err := m.handle.Wait()
+	if err != nil {
+		if isRetryable(err) && m.restarts < maxMemberRestarts {
+			if nh, relErr := s.launcher.Launch(s.baseCtx, spec); relErr == nil {
+				m.restarts++
+				m.handle = nh
+				go s.watch(t, m, spec)
+				return
+			}
+			// fall through to record the original error if relaunch fails
+		}
+		_ = s.coord.Fail(m.TaskID, memberError(err))
+	} else {
+		_ = s.coord.Complete(m.TaskID, res.Result)
+	}
+	t.removeMember(m.ID)
+	s.afterExit(t)
+}
+
+// afterExit releases the just-vacated slot and launches the next queued spec, if
+// any. Each exit drains at most one queued member; that member's own exit drains
+// the next, so the queue empties one-per-slot without unbounded recursion.
+func (s *Swarm) afterExit(t *Team) {
+	next, ok := t.onExit()
+	if !ok {
+		return
+	}
+	s.launch(t, next)
+}
+
+// Handoff transfers a task to a fresh member of toAgentType, delivering a note to
+// the new member's inbox and marking the original task handed-off. It returns the
+// new task id. A handoff of an already-terminal task is rejected (fail closed).
+func (s *Swarm) Handoff(pol Policy, teamName, taskID, toAgentType, note string) (string, error) {
+	task, ok := s.coord.Get(taskID)
+	if !ok {
+		return "", fmt.Errorf("%w: %s", ErrUnknownTask, taskID)
+	}
+	if task.Status.terminal() {
+		return "", fmt.Errorf("swarm: task %s already %s; cannot hand off", taskID, task.Status)
+	}
+	def, err := s.registry.Lookup(toAgentType)
+	if err != nil {
+		return "", err
+	}
+	team := sanitizeName(teamName)
+	newID := s.nextID(toAgentType)
+	handoffTask := task.Description
+	if note != "" {
+		handoffTask += "\n\nHandoff note: " + note
+	}
+	if _, err := s.coord.Register(newID, newID, team, handoffTask); err != nil {
+		return "", err
+	}
+	cwd := s.cwdFor(taskID)
+	s.rememberCwd(newID, cwd)
+	// Deliver the handoff note to the new member's inbox. A mailbox failure must
+	// not silently strand the handoff, so it is surfaced to the caller.
+	if note != "" {
+		if mbErr := s.mailbox.Send(team, newID, Message{
+			From: task.AgentID, Subject: "handoff", Body: note, Type: "handoff", Time: nowRFC3339(),
+		}); mbErr != nil {
+			return "", fmt.Errorf("swarm: deliver handoff note: %w", mbErr)
+		}
+	}
+	// Retire the original task (it has been re-delegated).
+	_ = s.coord.SetStatus(taskID, StatusHandedOff)
+	spec := s.buildSpec(pol, newID, newID, team, def, handoffTask, cwd)
+	s.dispatch(spec)
+	return newID, nil
+}
+
+// AdoptOrphans re-parents tasks in a team whose owning member is no longer live
+// (e.g. a crashed worker) onto fresh members of toAgentType, returning the
+// adopted task ids. Terminal tasks and tasks with a live owner are left alone.
+func (s *Swarm) AdoptOrphans(pol Policy, teamName, toAgentType string) ([]string, error) {
+	def, err := s.registry.Lookup(toAgentType)
+	if err != nil {
+		return nil, err
+	}
+	team := sanitizeName(teamName)
+	t := s.team(team)
+	live := t.liveAgents()
+	var adopted []string
+	for _, task := range s.coord.List() {
+		if task.Team != team || task.Status.terminal() {
+			continue
+		}
+		if _, ok := live[task.AgentID]; ok {
+			continue // still has a live member
+		}
+		newAgent := s.nextID(toAgentType)
+		if err := s.coord.Reassign(task.ID, newAgent); err != nil {
+			continue // raced to terminal; skip
+		}
+		cwd := s.cwdFor(task.ID)
+		s.rememberCwd(task.ID, cwd)
+		spec := s.buildSpec(pol, newAgent, task.ID, team, def, task.Description, cwd)
+		s.dispatch(spec)
+		adopted = append(adopted, task.ID)
+	}
+	return adopted, nil
+}
+
+// Collect returns snapshots of every task in a team, for swarm_collect/status.
+func (s *Swarm) Collect(teamName string) []Task {
+	team := sanitizeName(teamName)
+	var out []Task
+	for _, task := range s.coord.List() {
+		if task.Team == team {
+			out = append(out, task)
+		}
+	}
+	return out
+}

internal/swarm/lifecycle_test.go

@@ -0,0 +1,345 @@
+package swarm
+
+import (
+	"context"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+)
+
+// controllableLauncher records every launched spec and lets a test control each
+// member's outcome and (optionally) gate completion until a channel is closed.
+type controllableLauncher struct {
+	mu        sync.Mutex
+	specs     []MemberSpec
+	attempts  map[string]int
+	gate      chan struct{} // nil => members return immediately
+	result    func(spec MemberSpec, attempt int) (MemberResult, error)
+	launchErr error
+}
+
+func newLauncher(result func(MemberSpec, int) (MemberResult, error)) *controllableLauncher {
+	return &controllableLauncher{attempts: map[string]int{}, result: result}
+}
+
+func (l *controllableLauncher) Launch(ctx context.Context, spec MemberSpec) (MemberHandle, error) {
+	l.mu.Lock()
+	if l.launchErr != nil {
+		err := l.launchErr
+		l.mu.Unlock()
+		return nil, err
+	}
+	l.specs = append(l.specs, spec)
+	l.attempts[spec.ID]++
+	attempt := l.attempts[spec.ID]
+	gate := l.gate
+	result := l.result
+	l.mu.Unlock()
+
+	h := &funcHandle{id: spec.ID, done: make(chan struct{})}
+	go func() {
+		defer close(h.done)
+		if gate != nil {
+			select {
+			case <-gate:
+			case <-ctx.Done():
+				h.err = ctx.Err()
+				return
+			}
+		}
+		if result != nil {
+			h.res, h.err = result(spec, attempt)
+		}
+	}()
+	return h, nil
+}
+
+func (l *controllableLauncher) recorded() []MemberSpec {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	out := make([]MemberSpec, len(l.specs))
+	copy(out, l.specs)
+	return out
+}
+
+func (l *controllableLauncher) attemptCount(id string) int {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	return l.attempts[id]
+}
+
+func newSwarmFor(t *testing.T, l MemberLauncher) *Swarm {
+	t.Helper()
+	sw, err := New(Options{BaseDir: t.TempDir(), Launcher: l, MaxTeamSize: 2})
+	if err != nil {
+		t.Fatalf("New: %v", err)
+	}
+	t.Cleanup(sw.Close)
+	return sw
+}
+
+func waitFor(t *testing.T, what string, cond func() bool) {
+	t.Helper()
+	deadline := time.Now().Add(3 * time.Second)
+	for time.Now().Before(deadline) {
+		if cond() {
+			return
+		}
+		time.Sleep(5 * time.Millisecond)
+	}
+	t.Fatalf("timed out waiting for %s", what)
+}
+
+func okFor(spec MemberSpec, _ int) (MemberResult, error) {
+	return MemberResult{Result: "ok:" + spec.Task, SessionID: "sess-" + spec.ID}, nil
+}
+
+func TestSpawnCompletes(t *testing.T) {
+	l := newLauncher(okFor)
+	sw := newSwarmFor(t, l)
+	id, err := sw.Spawn(Policy{Model: "m"}, "team", "teammate", "build widget", "/work")
+	if err != nil {
+		t.Fatalf("Spawn: %v", err)
+	}
+	waitFor(t, "task done", func() bool {
+		task, ok := sw.Coordinator().Get(id)
+		return ok && task.Status == StatusDone
+	})
+	task, _ := sw.Coordinator().Get(id)
+	if task.Result != "ok:build widget" {
+		t.Fatalf("result = %q", task.Result)
+	}
+}
+
+func TestSpawnInheritsPolicy(t *testing.T) {
+	l := newLauncher(okFor)
+	sw := newSwarmFor(t, l)
+	_, err := sw.Spawn(Policy{Model: "orch-model", PermissionMode: permissionModeAuto}, "team", "teammate", "task", "/cwd")
+	if err != nil {
+		t.Fatalf("Spawn: %v", err)
+	}
+	waitFor(t, "spec recorded", func() bool { return len(l.recorded()) == 1 })
+	spec := l.recorded()[0]
+	if spec.Model != "orch-model" {
+		t.Fatalf("member model = %q, want inherited orch-model", spec.Model)
+	}
+	if spec.PermissionMode != permissionModeAuto {
+		t.Fatalf("member permission mode = %q, want inherited auto", spec.PermissionMode)
+	}
+	if spec.Cwd != "/cwd" {
+		t.Fatalf("member cwd = %q, want /cwd", spec.Cwd)
+	}
+	if spec.SystemPrompt == "" {
+		t.Fatal("member should carry a resolved system prompt")
+	}
+}
+
+func TestConcurrencyCapAndQueueDrains(t *testing.T) {
+	gate := make(chan struct{})
+	l := newLauncher(okFor)
+	l.gate = gate
+	sw := newSwarmFor(t, l) // MaxTeamSize 2
+	pol := Policy{Model: "m"}
+	for i := 0; i < 5; i++ {
+		if _, err := sw.Spawn(pol, "team", "teammate", "task", ""); err != nil {
+			t.Fatalf("Spawn %d: %v", i, err)
+		}
+	}
+	team := sw.team("team")
+	if team.Running() != 2 || team.QueueDepth() != 3 {
+		t.Fatalf("cap not enforced: running=%d queue=%d, want 2/3", team.Running(), team.QueueDepth())
+	}
+	// Release everyone; the queue should drain one-per-slot until all are done.
+	close(gate)
+	waitFor(t, "all tasks done", func() bool { return sw.Coordinator().Summarize().Done == 5 })
+	if team.Running() != 0 || team.QueueDepth() != 0 {
+		t.Fatalf("after drain running=%d queue=%d, want 0/0", team.Running(), team.QueueDepth())
+	}
+	if got := len(l.recorded()); got != 5 {
+		t.Fatalf("launched %d members, want 5", got)
+	}
+}
+
+func TestRetryOnTemporaryError(t *testing.T) {
+	l := newLauncher(func(spec MemberSpec, attempt int) (MemberResult, error) {
+		if attempt < 3 {
+			return MemberResult{}, ErrMemberTemporary
+		}
+		return MemberResult{Result: "recovered"}, nil
+	})
+	sw := newSwarmFor(t, l)
+	id, _ := sw.Spawn(Policy{}, "team", "teammate", "task", "")
+	waitFor(t, "task recovered", func() bool {
+		task, ok := sw.Coordinator().Get(id)
+		return ok && task.Status == StatusDone
+	})
+	if got := l.attemptCount(id); got != 3 {
+		t.Fatalf("attempts = %d, want 3 (initial + 2 retries)", got)
+	}
+	task, _ := sw.Coordinator().Get(id)
+	if task.Result != "recovered" {
+		t.Fatalf("result = %q, want recovered", task.Result)
+	}
+}
+
+func TestRetryExhaustionFails(t *testing.T) {
+	l := newLauncher(func(MemberSpec, int) (MemberResult, error) {
+		return MemberResult{}, ErrMemberTemporary
+	})
+	sw := newSwarmFor(t, l)
+	id, _ := sw.Spawn(Policy{}, "team", "teammate", "task", "")
+	waitFor(t, "task failed", func() bool {
+		task, ok := sw.Coordinator().Get(id)
+		return ok && task.Status == StatusFailed
+	})
+	if got := l.attemptCount(id); got != maxMemberRestarts+1 {
+		t.Fatalf("attempts = %d, want %d", got, maxMemberRestarts+1)
+	}
+}
+
+func TestPermanentErrorNoRetry(t *testing.T) {
+	l := newLauncher(func(MemberSpec, int) (MemberResult, error) {
+		return MemberResult{}, errPlain("hard failure")
+	})
+	sw := newSwarmFor(t, l)
+	id, _ := sw.Spawn(Policy{}, "team", "teammate", "task", "")
+	waitFor(t, "task failed", func() bool {
+		task, ok := sw.Coordinator().Get(id)
+		return ok && task.Status == StatusFailed
+	})
+	if got := l.attemptCount(id); got != 1 {
+		t.Fatalf("attempts = %d, want 1 (no retry on permanent error)", got)
+	}
+	task, _ := sw.Coordinator().Get(id)
+	if task.Err == "" {
+		t.Fatal("failed task should record an error message")
+	}
+}
+
+func TestHandoffDeliversNoteAndRetiresOriginal(t *testing.T) {
+	gate := make(chan struct{})
+	l := newLauncher(okFor)
+	l.gate = gate // keep the original member running so it is non-terminal
+	sw := newSwarmFor(t, l)
+	pol := Policy{Model: "m"}
+	origID, _ := sw.Spawn(pol, "team", "teammate", "original task", "/w")
+	waitFor(t, "original running", func() bool {
+		task, ok := sw.Coordinator().Get(origID)
+		return ok && task.Status == StatusRunning
+	})
+
+	newID, err := sw.Handoff(pol, "team", origID, "subagent", "please continue")
+	if err != nil {
+		t.Fatalf("Handoff: %v", err)
+	}
+	// Original retired.
+	orig, _ := sw.Coordinator().Get(origID)
+	if orig.Status != StatusHandedOff {
+		t.Fatalf("original status = %v, want handed-off", orig.Status)
+	}
+	// Note delivered to the new member's inbox.
+	msgs, err := sw.Mailbox().ReadAndConsume("team", newID)
+	if err != nil {
+		t.Fatalf("read new inbox: %v", err)
+	}
+	if len(msgs) != 1 || msgs[0].Body != "please continue" || msgs[0].Type != "handoff" {
+		t.Fatalf("handoff note not delivered: %+v", msgs)
+	}
+	// The new member carries the handoff note in its task and preserves cwd.
+	close(gate)
+	waitFor(t, "spec for new member", func() bool {
+		for _, s := range l.recorded() {
+			if s.ID == newID {
+				return true
+			}
+		}
+		return false
+	})
+	for _, s := range l.recorded() {
+		if s.ID == newID {
+			if s.Cwd != "/w" {
+				t.Fatalf("handoff lost cwd: %q", s.Cwd)
+			}
+		}
+	}
+	// A handoff of an already-terminal task is rejected.
+	waitFor(t, "new task done", func() bool {
+		task, ok := sw.Coordinator().Get(newID)
+		return ok && task.Status == StatusDone
+	})
+	if _, err := sw.Handoff(pol, "team", newID, "teammate", "again"); err == nil {
+		t.Fatal("handoff of a terminal task must fail")
+	}
+}
+
+func TestAdoptOrphans(t *testing.T) {
+	l := newLauncher(okFor)
+	sw := newSwarmFor(t, l)
+	// Simulate a crashed member: a running task in the coordinator whose owning
+	// agent has no live member in the team.
+	if _, err := sw.Coordinator().Register("orphan-1", "ghost", "team", "stranded work"); err != nil {
+		t.Fatalf("Register: %v", err)
+	}
+	_ = sw.Coordinator().SetStatus("orphan-1", StatusRunning)
+
+	adopted, err := sw.AdoptOrphans(Policy{Model: "m"}, "team", "subagent")
+	if err != nil {
+		t.Fatalf("AdoptOrphans: %v", err)
+	}
+	if len(adopted) != 1 || adopted[0] != "orphan-1" {
+		t.Fatalf("adopted = %v, want [orphan-1]", adopted)
+	}
+	// The orphan is relaunched under a fresh agent and completes.
+	waitFor(t, "orphan completed", func() bool {
+		task, ok := sw.Coordinator().Get("orphan-1")
+		return ok && task.Status == StatusDone
+	})
+	task, _ := sw.Coordinator().Get("orphan-1")
+	if task.AgentID == "ghost" {
+		t.Fatal("orphan should be reassigned to a new agent")
+	}
+	// A second adoption finds nothing (the task now has a live/owned outcome).
+	again, _ := sw.AdoptOrphans(Policy{}, "team", "subagent")
+	if len(again) != 0 {
+		t.Fatalf("second adoption = %v, want none", again)
+	}
+}
+
+func TestCollectScopesToTeam(t *testing.T) {
+	l := newLauncher(okFor)
+	sw := newSwarmFor(t, l)
+	a, _ := sw.Spawn(Policy{}, "alpha", "teammate", "ta", "")
+	_, _ = sw.Spawn(Policy{}, "beta", "teammate", "tb", "")
+	waitFor(t, "alpha done", func() bool {
+		task, ok := sw.Coordinator().Get(a)
+		return ok && task.Status == StatusDone
+	})
+	collected := sw.Collect("alpha")
+	if len(collected) != 1 || collected[0].Team != "alpha" {
+		t.Fatalf("Collect(alpha) = %+v, want one alpha task", collected)
+	}
+}
+
+func TestFuncLauncherRecoversPanic(t *testing.T) {
+	// A panic inside a member's Run must surface as that member's error, never
+	// escape the goroutine and crash the orchestrator.
+	l := FuncLauncher{Run: func(context.Context, MemberSpec) (MemberResult, error) {
+		panic("boom in member")
+	}}
+	h, err := l.Launch(context.Background(), MemberSpec{ID: "m1"})
+	if err != nil {
+		t.Fatalf("Launch: %v", err)
+	}
+	_, err = h.Wait()
+	if err == nil || !strings.Contains(err.Error(), "panicked") {
+		t.Fatalf("panic must surface as a member error, got %v", err)
+	}
+}
+
+func TestSpawnUnknownAgentType(t *testing.T) {
+	sw := newSwarmFor(t, newLauncher(okFor))
+	if _, err := sw.Spawn(Policy{}, "team", "does-not-exist", "task", ""); err == nil {
+		t.Fatal("Spawn with unknown agent type must error")
+	}
+}

internal/swarm/lock_other.go

@@ -0,0 +1,15 @@
+//go:build !windows
+
+package swarm
+
+import (
+	"errors"
+	"os"
+)
+
+// isLockContended reports whether an O_EXCL lock-create error means the lock is
+// currently held (so the caller should retry/wait) rather than a hard failure.
+// On POSIX an existing lock surfaces only as os.ErrExist.
+func isLockContended(err error) bool {
+	return errors.Is(err, os.ErrExist)
+}

internal/swarm/lock_windows.go

@@ -0,0 +1,30 @@
+//go:build windows
+
+package swarm
+
+import (
+	"errors"
+	"os"
+	"syscall"
+)
+
+// Windows error codes that an O_EXCL create (or a racing delete) can surface
+// when another writer holds the lock file open, instead of ERROR_FILE_EXISTS.
+// Treating them as "contended" makes the lock retry/wait rather than fail the
+// whole send/read spuriously under contention.
+const (
+	errorAccessDenied     = syscall.Errno(5)  // ERROR_ACCESS_DENIED
+	errorSharingViolation = syscall.Errno(32) // ERROR_SHARING_VIOLATION
+)
+
+// isLockContended reports whether an O_EXCL lock-create error means the lock is
+// currently held (so the caller should retry/wait) rather than a hard failure.
+// On Windows a held lock can appear as ERROR_FILE_EXISTS (mapped to os.ErrExist)
+// but also as ERROR_ACCESS_DENIED / ERROR_SHARING_VIOLATION when the file is
+// concurrently open or pending deletion.
+func isLockContended(err error) bool {
+	if errors.Is(err, os.ErrExist) {
+		return true
+	}
+	return errors.Is(err, errorAccessDenied) || errors.Is(err, errorSharingViolation)
+}

internal/swarm/mailbox.go

@@ -0,0 +1,385 @@
+package swarm
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+	"sync/atomic"
+	"time"
+)
+
+// Mailbox is a per-agent, per-team message inbox persisted as a JSON array on
+// disk at <baseDir>/<team>/inboxes/<agent>.json, written atomically under a
+// lock file. It is hardened for untrusted input:
+//
+//   - inbox files and their parent dirs are owner-only (0600/0700);
+//   - every write takes an exclusive lock file (bounded retry, stale-break);
+//   - message bodies and whole inboxes are size-capped (fail closed on oversize);
+//   - team/agent names are sanitized so a name can never escape the base dir;
+//   - malformed inbox JSON is rejected rather than silently reset (fail closed).
+type Mailbox struct {
+	// BaseDir is the root under which <team>/inboxes/<agent>.json live. It must
+	// be set (NewMailbox enforces this).
+	BaseDir string
+	// MaxMessageBytes caps a single serialized message. 0 => defaultMaxMessageBytes.
+	MaxMessageBytes int
+	// MaxMessages caps the number of messages retained in one inbox. 0 =>
+	// defaultMaxMessages. Sends past the cap fail closed.
+	MaxMessages int
+	// LockTimeout bounds how long Send/MarkRead wait for the inbox lock.
+	LockTimeout time.Duration
+}
+
+const (
+	defaultMaxMessageBytes = 64 * 1024 // 64 KiB per message
+	defaultMaxMessages     = 1000      // per inbox
+	defaultLockTimeout     = 5 * time.Second
+	lockStaleAfter         = 30 * time.Second
+	inboxFileMode          = 0o600
+	inboxDirMode           = 0o700
+)
+
+// ErrMailboxFull is returned when an inbox is at MaxMessages.
+var ErrMailboxFull = errors.New("swarm: mailbox is full")
+
+// ErrMessageTooLarge is returned when a single message exceeds MaxMessageBytes.
+var ErrMessageTooLarge = errors.New("swarm: message exceeds size cap")
+
+// Message is one inbox entry. Time is RFC3339 and set on Send if empty.
+type Message struct {
+	From    string `json:"from"`
+	To      string `json:"to"`
+	Subject string `json:"subject,omitempty"`
+	Body    string `json:"body"`
+	Type    string `json:"type"`
+	Read    bool   `json:"read"`
+	Time    string `json:"time,omitempty"`
+}
+
+// NewMailbox validates baseDir and returns a Mailbox with defaults applied.
+func NewMailbox(baseDir string) (*Mailbox, error) {
+	if strings.TrimSpace(baseDir) == "" {
+		return nil, errors.New("swarm: mailbox base dir is required")
+	}
+	abs, err := filepath.Abs(baseDir)
+	if err != nil {
+		return nil, fmt.Errorf("swarm: resolve mailbox base dir: %w", err)
+	}
+	return &Mailbox{
+		BaseDir:         abs,
+		MaxMessageBytes: defaultMaxMessageBytes,
+		MaxMessages:     defaultMaxMessages,
+		LockTimeout:     defaultLockTimeout,
+	}, nil
+}
+
+var nameUnsafe = regexp.MustCompile(`[^a-zA-Z0-9\-_]`)
+
+// sanitizeName replaces any char outside [A-Za-z0-9-_] with '-', defaulting
+// empty input to "default" and an all-unsafe result to "unknown". This
+// guarantees the name is a single safe path segment (no '/', '\\', '.', '..').
+func sanitizeName(name string) string {
+	trimmed := strings.TrimSpace(name)
+	if trimmed == "" {
+		trimmed = "default"
+	}
+	cleaned := nameUnsafe.ReplaceAllString(trimmed, "-")
+	if cleaned == "" {
+		return "unknown"
+	}
+	return cleaned
+}
+
+func (m *Mailbox) maxMessageBytes() int {
+	if m.MaxMessageBytes > 0 {
+		return m.MaxMessageBytes
+	}
+	return defaultMaxMessageBytes
+}
+
+func (m *Mailbox) maxMessages() int {
+	if m.MaxMessages > 0 {
+		return m.MaxMessages
+	}
+	return defaultMaxMessages
+}
+
+func (m *Mailbox) lockTimeout() time.Duration {
+	if m.LockTimeout > 0 {
+		return m.LockTimeout
+	}
+	return defaultLockTimeout
+}
+
+// inboxPath returns the sanitized, base-confined inbox path for (team, agent)
+// and verifies it cannot escape BaseDir lexically (defense in depth on top of
+// sanitize). Symlink-aware confinement happens in ensureInboxDir.
+func (m *Mailbox) inboxPath(team, agent string) (string, error) {
+	dir := filepath.Join(m.BaseDir, sanitizeName(team), "inboxes")
+	path := filepath.Join(dir, sanitizeName(agent)+".json")
+	if within, err := pathWithin(m.BaseDir, path); err != nil || !within {
+		return "", fmt.Errorf("swarm: inbox path escapes base dir (team=%q agent=%q)", team, agent)
+	}
+	return path, nil
+}
+
+// pathWithin reports whether target is base or lives under it (lexically).
+func pathWithin(base, target string) (bool, error) {
+	rel, err := filepath.Rel(base, target)
+	if err != nil {
+		return false, err
+	}
+	if rel == ".." || strings.HasPrefix(rel, ".."+string(os.PathSeparator)) {
+		return false, nil
+	}
+	return true, nil
+}
+
+// ensureInboxDir creates the inbox's parent directories owner-only, then
+// verifies — after resolving symlinks — that the real directory stays within the
+// real base dir. This defends against a symlink planted under BaseDir (e.g. a
+// pre-existing "inboxes" -> /etc) that the lexical inboxPath check cannot catch.
+// Only after confinement is confirmed are the directories tightened to 0700, so
+// a chmod can never land outside the base via a symlink.
+func (m *Mailbox) ensureInboxDir(inboxPath string) error {
+	dir := filepath.Dir(inboxPath) // <base>/<team>/inboxes
+	if err := os.MkdirAll(dir, inboxDirMode); err != nil {
+		return fmt.Errorf("swarm: create inbox dir: %w", err)
+	}
+	realBase, err := filepath.EvalSymlinks(m.BaseDir)
+	if err != nil {
+		return fmt.Errorf("swarm: resolve base dir: %w", err)
+	}
+	realDir, err := filepath.EvalSymlinks(dir)
+	if err != nil {
+		return fmt.Errorf("swarm: resolve inbox dir: %w", err)
+	}
+	if within, err := pathWithin(realBase, realDir); err != nil || !within {
+		return fmt.Errorf("swarm: inbox dir escapes base dir via symlink")
+	}
+	// MkdirAll leaves an existing dir's permissions untouched; tighten the team
+	// and inboxes dirs to owner-only so mailbox contents are never group/world
+	// readable even if a dir pre-existed with broad perms.
+	_ = os.Chmod(dir, inboxDirMode)
+	_ = os.Chmod(filepath.Dir(dir), inboxDirMode) // <base>/<team>
+	return nil
+}
+
+// Send appends msg to the recipient's inbox under the team, taking the inbox
+// lock. It fails closed on oversize messages or a full inbox.
+//
+// Each send rewrites the whole inbox (read all → append one → write all), so it
+// is O(N) in the inbox size. That is intentional given the small bounds
+// (MaxMessages × MaxMessageBytes); if those caps are raised substantially,
+// switch to an append-only log with periodic compaction.
+func (m *Mailbox) Send(team, recipient string, msg Message) error {
+	path, err := m.inboxPath(team, recipient)
+	if err != nil {
+		return err
+	}
+	if msg.Type == "" {
+		msg.Type = "message"
+	}
+	if msg.To == "" {
+		msg.To = sanitizeName(recipient)
+	}
+	if msg.Time == "" {
+		msg.Time = time.Now().UTC().Format(time.RFC3339)
+	}
+	encoded, err := json.Marshal(msg)
+	if err != nil {
+		return fmt.Errorf("swarm: encode message: %w", err)
+	}
+	if len(encoded) > m.maxMessageBytes() {
+		return fmt.Errorf("%w: %d > %d bytes", ErrMessageTooLarge, len(encoded), m.maxMessageBytes())
+	}
+	if err := m.ensureInboxDir(path); err != nil {
+		return err
+	}
+	release, err := acquireLock(path+".lock", m.lockTimeout())
+	if err != nil {
+		return err
+	}
+	defer release()
+
+	messages, err := m.readLocked(path)
+	if err != nil {
+		return err
+	}
+	if len(messages) >= m.maxMessages() {
+		return fmt.Errorf("%w: %d messages", ErrMailboxFull, len(messages))
+	}
+	messages = append(messages, msg)
+	return atomicWriteJSON(path, messages)
+}
+
+// ReadAndConsume reads the recipient's inbox and marks every previously-unread
+// message read under one lock, returning the messages with their read flags as
+// they were BEFORE this read (so the caller can tell which were new). The inbox
+// is empty (no error) if it does not yet exist; malformed inbox JSON is reported
+// as an error (fail closed) and nothing is written.
+func (m *Mailbox) ReadAndConsume(team, recipient string) ([]Message, error) {
+	path, err := m.inboxPath(team, recipient)
+	if err != nil {
+		return nil, err
+	}
+	// The lock file lives beside the inbox; create the dir so it can be taken
+	// even on first read of a not-yet-existing inbox.
+	if err := m.ensureInboxDir(path); err != nil {
+		return nil, err
+	}
+	release, err := acquireLock(path+".lock", m.lockTimeout())
+	if err != nil {
+		return nil, err
+	}
+	defer release()
+
+	messages, err := m.readLocked(path)
+	if err != nil {
+		return nil, err
+	}
+	if len(messages) == 0 {
+		return messages, nil
+	}
+	// Snapshot pre-consume state to return, then flip unread -> read on disk.
+	snapshot := make([]Message, len(messages))
+	copy(snapshot, messages)
+	changed := false
+	for i := range messages {
+		if !messages[i].Read {
+			messages[i].Read = true
+			changed = true
+		}
+	}
+	if changed {
+		if err := atomicWriteJSON(path, messages); err != nil {
+			return nil, err
+		}
+	}
+	return snapshot, nil
+}
+
+// readLocked reads and decodes an inbox file. Callers that mutate must hold the
+// lock; pure readers may call it directly (a torn read surfaces as a decode
+// error and is retried by the caller, never silently treated as empty).
+func (m *Mailbox) readLocked(path string) ([]Message, error) {
+	raw, err := os.ReadFile(path)
+	if err != nil {
+		if errors.Is(err, os.ErrNotExist) {
+			return nil, nil
+		}
+		return nil, fmt.Errorf("swarm: read inbox: %w", err)
+	}
+	if len(raw) == 0 {
+		return nil, nil
+	}
+	// Reject an oversized inbox file outright rather than decoding untrusted,
+	// unbounded JSON into memory. The bound is the cap on total message bytes
+	// (maxMessages × maxMessageBytes) doubled to allow for pretty-printed JSON
+	// (indentation, field names, brackets) overhead versus the compact bytes the
+	// per-message cap measures.
+	if maxFile := m.maxMessages() * m.maxMessageBytes() * 2; len(raw) > maxFile {
+		return nil, fmt.Errorf("swarm: inbox file too large (%d bytes)", len(raw))
+	}
+	var messages []Message
+	if err := json.Unmarshal(raw, &messages); err != nil {
+		return nil, fmt.Errorf("swarm: malformed inbox %s: %w", filepath.Base(path), err)
+	}
+	for i := range messages {
+		if messages[i].Type == "" {
+			messages[i].Type = "message"
+		}
+	}
+	return messages, nil
+}
+
+// atomicWriteJSON writes data as pretty JSON to a sibling temp file (0600) then
+// renames it over path, so a reader never observes a partial write.
+func atomicWriteJSON(path string, data any) error {
+	encoded, err := json.MarshalIndent(data, "", "  ")
+	if err != nil {
+		return fmt.Errorf("swarm: encode inbox: %w", err)
+	}
+	dir := filepath.Dir(path)
+	tmp, err := os.CreateTemp(dir, ".inbox-*.tmp")
+	if err != nil {
+		return fmt.Errorf("swarm: create temp inbox: %w", err)
+	}
+	tmpName := tmp.Name()
+	defer os.Remove(tmpName) // no-op after a successful rename
+	if err := tmp.Chmod(inboxFileMode); err != nil {
+		tmp.Close()
+		return fmt.Errorf("swarm: chmod temp inbox: %w", err)
+	}
+	if _, err := tmp.Write(encoded); err != nil {
+		tmp.Close()
+		return fmt.Errorf("swarm: write temp inbox: %w", err)
+	}
+	if err := tmp.Close(); err != nil {
+		return fmt.Errorf("swarm: close temp inbox: %w", err)
+	}
+	if err := os.Rename(tmpName, path); err != nil {
+		return fmt.Errorf("swarm: commit inbox: %w", err)
+	}
+	return nil
+}
+
+// lockSeq makes each lock token unique within the process even when two
+// goroutines acquire in the same nanosecond.
+var lockSeq atomic.Uint64
+
+// acquireLock takes an exclusive lock by creating lockPath with O_EXCL. It
+// retries with a short backoff until timeout, breaking a lock whose file is
+// older than lockStaleAfter (so a crashed holder cannot deadlock the swarm).
+//
+// Each holder writes a unique token into the lock file and release is
+// OWNERSHIP-AWARE: it removes the lock only if it still holds our token. After a
+// stale-break another writer may legitimately own the lock; unconditionally
+// removing it would let a third writer in and create split-brain writers that
+// can corrupt or drop mailbox updates.
+func acquireLock(lockPath string, timeout time.Duration) (func(), error) {
+	token := fmt.Sprintf("%d-%d-%d", os.Getpid(), time.Now().UnixNano(), lockSeq.Add(1))
+	deadline := time.Now().Add(timeout)
+	for {
+		f, err := os.OpenFile(lockPath, os.O_CREATE|os.O_EXCL|os.O_WRONLY, inboxFileMode)
+		if err == nil {
+			// Record our ownership token; the exclusive create already holds the
+			// lock, so a write error only weakens diagnostics, not safety.
+			_, _ = f.WriteString(token)
+			f.Close()
+			var released bool
+			return func() {
+				if released {
+					return
+				}
+				released = true
+				if data, rerr := os.ReadFile(lockPath); rerr == nil && string(data) == token {
+					os.Remove(lockPath)
+				}
+			}, nil
+		}
+		if !isLockContended(err) {
+			return nil, fmt.Errorf("swarm: acquire lock: %w", err)
+		}
+		// Lock held: break it if stale, otherwise wait. The break is conditional
+		// on the file's content being unchanged since we observed it, so we do not
+		// delete a lock another writer rotated in between.
+		if info, statErr := os.Stat(lockPath); statErr == nil && time.Since(info.ModTime()) > lockStaleAfter {
+			stale, _ := os.ReadFile(lockPath)
+			if data, rerr := os.ReadFile(lockPath); rerr == nil && string(data) == string(stale) {
+				os.Remove(lockPath)
+			}
+			continue
+		}
+		if time.Now().After(deadline) {
+			return nil, fmt.Errorf("swarm: timed out acquiring lock %s", filepath.Base(lockPath))
+		}
+		// Short retry so a freed lock is re-acquired promptly under heavy
+		// contention (Windows file ops are slow; a coarse sleep starves waiters).
+		time.Sleep(2 * time.Millisecond)
+	}
+}