diff --git a/CHANGELOG.md b/CHANGELOG.md
index 8a0b2c7..fda8504 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -50,3 +50,13 @@ Initial implementation.
   through the cancel endpoint before teardown. Permission and question
   requests cannot be satisfied unattended and fail closed (exit 2) naming
   the tool. Runs fall back to `crush run` when the server cannot start.
+- Full-permission consent: `traust-cli init` asks whether to run skills
+  with all tools auto-approved (the harness's skills require unattended
+  bash and file writes) and records the answer in the workspace state;
+  headless installs pass `--full-permission`. Runs translate the consent
+  into the right flag per harness — claude
+  `--dangerously-skip-permissions`, crush `--yolo` / the server's
+  permissions-skip endpoint, codex
+  `--dangerously-bypass-approvals-and-sandbox`, pi's full toolset — and
+  surface it in the notes, the run journal and `doctor`. Without consent,
+  permission requests still fail closed (exit 2) with guidance.
diff --git a/cmd/init.go b/cmd/init.go
index 3b84b99..4544d89 100644
--- a/cmd/init.go
+++ b/cmd/init.go
@@ -14,14 +14,15 @@ import (
 )
 
 var initOpts struct {
-	workspace    string
-	forge        string
-	ref          string
-	configHome   string
-	toolchain    string
-	harness      string
-	writeProfile bool
-	yes          bool
+	workspace      string
+	forge          string
+	ref            string
+	configHome     string
+	toolchain      string
+	harness        string
+	fullPermission bool
+	writeProfile   bool
+	yes            bool
 }
 
 var initCmd = &cobra.Command{
@@ -49,6 +50,7 @@ func init() {
 	initCmd.Flags().StringVar(&initOpts.configHome, "config-home", "", "TRAUST_CONFIG_HOME (default <workspace>/config)")
 	initCmd.Flags().StringVar(&initOpts.toolchain, "toolchain", "none", "scanner toolchain profile (none, secure-code-audit, vuln-scan, triage, …)")
 	initCmd.Flags().StringVar(&initOpts.harness, "harness", "", "default agent harness (default: first detected)")
+	initCmd.Flags().BoolVar(&initOpts.fullPermission, "full-permission", false, "consent to full-permission runs: harness executes with all tools auto-approved (required by the harness's skills)")
 	initCmd.Flags().BoolVar(&initOpts.writeProfile, "write-profile", false, "also record TRAUST_CONFIG_HOME in your shell profile (install_traust --write-profile)")
 	initCmd.Flags().BoolVarP(&initOpts.yes, "yes", "y", false, "non-interactive: accept all defaults")
 }
@@ -85,6 +87,7 @@ func runInit(cmd *cobra.Command, args []string) {
 		initOpts.ref = answers.Ref
 		initOpts.toolchain = answers.ToolchainProfile
 		initOpts.harness = answers.Harness
+		initOpts.fullPermission = answers.FullPermission
 	}
 
 	log := func(format string, a ...any) {
@@ -99,6 +102,7 @@ func runInit(cmd *cobra.Command, args []string) {
 		ToolchainProfile: initOpts.toolchain,
 		WriteProfile:     initOpts.writeProfile,
 		Harness:          initOpts.harness,
+		FullPermission:   initOpts.fullPermission,
 	}, log)
 	if err != nil {
 		fatalf(1, "init: %v", err)
diff --git a/cmd/run.go b/cmd/run.go
index b49a2c7..2bb0f48 100644
--- a/cmd/run.go
+++ b/cmd/run.go
@@ -50,7 +50,7 @@ func init() {
 	runCmd.Flags().StringVar(&runFlags.harnessID, "harness", "", "agent harness (claude, codex, pi, opencode, crush, goose; default: state default)")
 	runCmd.Flags().StringVar(&runFlags.model, "model", "", "model override (default: the registry floor for the skill's role)")
 	runCmd.Flags().IntVar(&runFlags.maxTurns, "max-turns", 0, "cap agent turns (0 = harness default)")
-	runCmd.Flags().BoolVar(&runFlags.unsafeFullPermission, "unsafe-full-permission", false, "grant the harness full permission (--dangerously-skip-permissions / --yolo class)")
+	runCmd.Flags().BoolVar(&runFlags.unsafeFullPermission, "unsafe-full-permission", false, "grant the harness full permission for this run (also settable at init: --full-permission)")
 	runCmd.Flags().StringVar(&runFlags.budgetOverride, "budget-override", "", "reason to proceed past the budget refuse band (recorded)")
 	runCmd.Flags().BoolVar(&runFlags.json, "json", false, "machine-readable result summary on stdout")
 	runCmd.Flags().BoolVar(&runFlags.strictValidation, "strict", false, "strict schema validation of artifacts")
@@ -128,6 +128,13 @@ func executeRun(state *bootstrap.State, skill *catalog.Skill, args []string) {
 
 	client := api.NewPythonCLI(state.VenvPython(), state.ChildEnv(), state.TraustRepo())
 
+	// Full-permission mode: explicit per-run flag, or the init-time consent
+	// recorded in state. The harness's skills require unattended tool use.
+	unsafeFullPermission := runFlags.unsafeFullPermission || state.FullPermission
+	if unsafeFullPermission && !runFlags.unsafeFullPermission {
+		fmt.Fprintln(os.Stderr, theme.Default().Warn.Render("note: full-permission mode active (consented at init; re-run `traust-cli init` to change) — harness runs with all tools auto-approved"))
+	}
+
 	if skill.Standalone == nil {
 		styles := theme.Default()
 		fmt.Fprintln(os.Stderr, styles.Warn.Render("warning: "+fmt.Sprintf("%q is not in the standalone profile — it may expect campaign infrastructure (siblings, ledger, tracker)", skill.Name)))
@@ -141,7 +148,7 @@ func executeRun(state *bootstrap.State, skill *catalog.Skill, args []string) {
 		Args:                 args,
 		Model:                runFlags.model,
 		MaxTurns:             runFlags.maxTurns,
-		UnsafeFullPermission: runFlags.unsafeFullPermission,
+		UnsafeFullPermission: unsafeFullPermission,
 		BudgetOverrideReason: runFlags.budgetOverride,
 		StrictValidation:     runFlags.strictValidation,
 		DeclareSpend:         true,
diff --git a/internal/bootstrap/bootstrap.go b/internal/bootstrap/bootstrap.go
index 8995313..1b893f9 100644
--- a/internal/bootstrap/bootstrap.go
+++ b/internal/bootstrap/bootstrap.go
@@ -33,6 +33,9 @@ type InitOptions struct {
 	WriteProfile bool
 	// Harness is the default harness id to store in state; empty auto-detects.
 	Harness string
+	// FullPermission records the consent to run harnesses with all tools
+	// auto-approved (the traust skills require it).
+	FullPermission bool
 }
 
 // LogFn receives progress lines during init.
@@ -166,6 +169,12 @@ func Init(ctx context.Context, opts InitOptions, log LogFn) (*State, error) {
 		state.Harness = "claude"
 	}
 
+	// 7. Full-permission consent (explicit --full-permission or a prior
+	// consent survives re-inits; the wizard's answer arrives via opts).
+	if opts.FullPermission {
+		state.FullPermission = true
+	}
+
 	if err := state.Save(); err != nil {
 		return nil, err
 	}
diff --git a/internal/bootstrap/doctor.go b/internal/bootstrap/doctor.go
index 0299a29..93e7fdc 100644
--- a/internal/bootstrap/doctor.go
+++ b/internal/bootstrap/doctor.go
@@ -39,6 +39,11 @@ func Doctor(ctx context.Context, state *State, log LogFn) (*DoctorReport, error)
 
 	// State file present and parseable (caller already loaded it).
 	add("state", true, StatePath(state.Workspace))
+	if state.FullPermission {
+		add("permission-mode", true, "full (consented at init): harnesses run with all tools auto-approved")
+	} else {
+		add("permission-mode", true, "scoped (fail-closed): permission requests abort runs with exit 2 — the harness's skills need `traust-cli init --full-permission`")
+	}
 
 	// Traust checkout + venv.
 	repo := state.TraustRepo()
diff --git a/internal/bootstrap/state.go b/internal/bootstrap/state.go
index e54fa0f..80194a5 100644
--- a/internal/bootstrap/state.go
+++ b/internal/bootstrap/state.go
@@ -24,16 +24,22 @@ var ErrNotInitialized = errors.New("traust-cli is not initialized — run `traus
 // <workspace>/cli/state.json. It is deliberately outside TRAUST_CONFIG_HOME,
 // which is schema-gated deployment configuration owned by the harness.
 type State struct {
-	Version           int       `json:"version"`
-	Workspace         string    `json:"workspace"`
-	ConfigHome        string    `json:"config_home"`
-	TraustRef         string    `json:"traust_ref"`
-	TraustSHA         string    `json:"traust_sha"`
-	UvVersion         string    `json:"uv_version,omitempty"`
-	Harness           string    `json:"harness,omitempty"`
-	ToolchainProfiles []string  `json:"toolchain_profiles,omitempty"`
-	CreatedAt         time.Time `json:"created_at"`
-	UpdatedAt         time.Time `json:"updated_at"`
+	Version           int      `json:"version"`
+	Workspace         string   `json:"workspace"`
+	ConfigHome        string   `json:"config_home"`
+	TraustRef         string   `json:"traust_ref"`
+	TraustSHA         string   `json:"traust_sha"`
+	UvVersion         string   `json:"uv_version,omitempty"`
+	Harness           string   `json:"harness,omitempty"`
+	ToolchainProfiles []string `json:"toolchain_profiles,omitempty"`
+	// FullPermission records the init-time consent to run skills with all
+	// tools auto-approved (claude --dangerously-skip-permissions, crush
+	// --yolo / permissions/skip, codex --dangerously-bypass-approvals...).
+	// The traust harness's skills require it; without it, permission
+	// requests fail closed (exit 2).
+	FullPermission bool      `json:"full_permission,omitempty"`
+	CreatedAt      time.Time `json:"created_at"`
+	UpdatedAt      time.Time `json:"updated_at"`
 }
 
 // DefaultWorkspace returns the default managed workspace root (~/.traust).
diff --git a/internal/harness/adapters_test.go b/internal/harness/adapters_test.go
index 24afaf9..4529fc0 100644
--- a/internal/harness/adapters_test.go
+++ b/internal/harness/adapters_test.go
@@ -76,6 +76,12 @@ func TestPiCommand(t *testing.T) {
 	if !contains(args, "--tools") || !contains(args, "read,bash,edit,write,grep,find,ls") {
 		t.Errorf("pi default tools wrong: %v", args)
 	}
+	// Full-permission mode: the complete builtin toolset.
+	unsafe := spec()
+	unsafe.UnsafeFullPermission = true
+	if args := argsOf(t, &Pi{}, unsafe); !contains(args, "read,bash,edit,write,grep,find,ls") {
+		t.Errorf("pi full-permission tools wrong: %v", args)
+	}
 	// Task and AskUserQuestion are unmappable and dropped.
 	mapped := mapTools([]string{"Read", "Task", "AskUserQuestion", "Grep"})
 	if mapped != "read,grep" {
@@ -101,6 +107,11 @@ func TestCrushCommand(t *testing.T) {
 			t.Errorf("crush args missing %q: %v", want, cmd.Args)
 		}
 	}
+	unsafe := spec()
+	unsafe.UnsafeFullPermission = true
+	if args := argsOf(t, &Crush{}, unsafe); !contains(args, "--yolo") {
+		t.Errorf("full-permission crush run should pass --yolo: %v", args)
+	}
 	if cmd.Stdin == nil {
 		t.Error("crush should take the prompt on stdin")
 	}
diff --git a/internal/harness/crush.go b/internal/harness/crush.go
index b28d974..14a845c 100644
--- a/internal/harness/crush.go
+++ b/internal/harness/crush.go
@@ -30,6 +30,9 @@ func (c *Crush) Command(ctx context.Context, spec *RunSpec) *exec.Cmd {
 	if spec.Model != "" {
 		args = append(args, "--model", spec.Model)
 	}
+	if spec.UnsafeFullPermission {
+		args = append(args, "--yolo")
+	}
 	cmd := exec.CommandContext(ctx, c.Binary(), args...)
 	cmd.Dir = spec.Cwd
 	cmd.Env = spec.Env
@@ -40,7 +43,7 @@ func (c *Crush) Command(ctx context.Context, spec *RunSpec) *exec.Cmd {
 func (c *Crush) Notes() []string {
 	return []string{
 		"running through crush server mode: thinking, tool calls and token usage stream live into the monitor",
-		"permission and question requests cannot be answered unattended: the run fails closed (exit 2) naming the tool — pre-allow the tools skills need in a .crush/crushrc at the harness root (`permissions allow`)",
+		"without full-permission mode (consent at `traust-cli init`), permission and question requests fail closed (exit 2) — with it, the workspace runs with permissions skipped (--yolo equivalent)",
 		"if the server cannot start, the run falls back to `crush run` (plain text, no live events, no usage)",
 	}
 }
diff --git a/internal/harness/crush_serve.go b/internal/harness/crush_serve.go
index 23ce4b6..2b4a9df 100644
--- a/internal/harness/crush_serve.go
+++ b/internal/harness/crush_serve.go
@@ -45,6 +45,18 @@ func (c *Crush) RunSession(ctx context.Context, spec *RunSpec, emit func(Event))
 	if err != nil {
 		return SessionResult{}, fmt.Errorf("create workspace: %w", err)
 	}
+
+	// Full-permission consent (init): skip the workspace's permission
+	// prompts — the server equivalent of `crush --yolo`. The traust
+	// skills cannot run unattended without it; without consent, permission
+	// requests fail closed (needs-input) instead.
+	if spec.UnsafeFullPermission {
+		code, err := srv.call(ctx, http.MethodPost, "/v1/workspaces/"+wsID+"/permissions/skip", map[string]bool{"skip": true}, nil)
+		if err != nil || code != http.StatusOK {
+			return SessionResult{}, fmt.Errorf("set permissions skip: %d %v", code, err)
+		}
+	}
+
 	sessID, err := srv.createSession(ctx, wsID)
 	if err != nil {
 		return SessionResult{}, fmt.Errorf("create session: %w", err)
diff --git a/internal/harness/crush_serve_test.go b/internal/harness/crush_serve_test.go
index a240f40..de2103e 100644
--- a/internal/harness/crush_serve_test.go
+++ b/internal/harness/crush_serve_test.go
@@ -27,6 +27,7 @@ type fakeCrush struct {
 	agentRunID  string
 	cancelCalls []string
 	modelReqs   []map[string]any
+	skipReqs    []bool
 	script      func(w io.Writer, r *http.Request, sid, runID string)
 	agentPosted chan struct{}
 	postOnce    sync.Once
@@ -88,6 +89,17 @@ func startFakeCrush(t *testing.T, script func(w io.Writer, r *http.Request, sid,
 		w.WriteHeader(http.StatusOK)
 	})
 
+	mux.HandleFunc("POST /v1/workspaces/{id}/permissions/skip", func(w http.ResponseWriter, r *http.Request) {
+		var body struct {
+			Skip bool `json:"skip"`
+		}
+		json.NewDecoder(r.Body).Decode(&body)
+		f.mu.Lock()
+		f.skipReqs = append(f.skipReqs, body.Skip)
+		f.mu.Unlock()
+		w.WriteHeader(http.StatusOK)
+	})
+
 	mux.HandleFunc("GET /v1/workspaces/{id}/events", func(w http.ResponseWriter, r *http.Request) {
 		if r.URL.Query().Get("client_id") == "" {
 			w.WriteHeader(http.StatusBadRequest)
@@ -266,6 +278,37 @@ func TestCrushServeIgnoresForeignRunComplete(t *testing.T) {
 	}
 }
 
+func TestCrushServeFullPermissionSkipsPermissions(t *testing.T) {
+	f := startFakeCrush(t, happyScript)
+	_, err := (&Crush{}).RunSession(context.Background(), &RunSpec{
+		Prompt: "do the audit", Cwd: t.TempDir(), Env: os.Environ(),
+		UnsafeFullPermission: true,
+	}, func(Event) {})
+	if err != nil {
+		t.Fatal(err)
+	}
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	if len(f.skipReqs) != 1 || !f.skipReqs[0] {
+		t.Errorf("permissions/skip requests = %v (want one skip=true)", f.skipReqs)
+	}
+}
+
+func TestCrushServeScopedModeDoesNotSkipPermissions(t *testing.T) {
+	f := startFakeCrush(t, happyScript)
+	_, err := (&Crush{}).RunSession(context.Background(), &RunSpec{
+		Prompt: "do the audit", Cwd: t.TempDir(), Env: os.Environ(),
+	}, func(Event) {})
+	if err != nil {
+		t.Fatal(err)
+	}
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	if len(f.skipReqs) != 0 {
+		t.Errorf("scoped mode must not touch permissions/skip: %v", f.skipReqs)
+	}
+}
+
 func TestCrushServeModelOverride(t *testing.T) {
 	f := startFakeCrush(t, happyScript)
 	// Fake `crush models` listing; the runner resolves the floor model here.
diff --git a/internal/harness/pi.go b/internal/harness/pi.go
index d4b3cd3..f3cc423 100644
--- a/internal/harness/pi.go
+++ b/internal/harness/pi.go
@@ -71,7 +71,11 @@ func (p *Pi) Command(ctx context.Context, spec *RunSpec) *exec.Cmd {
 	if spec.Model != "" {
 		args = append(args, "--model", spec.Model)
 	}
-	if !spec.UnsafeFullPermission {
+	if spec.UnsafeFullPermission {
+		// Full-permission mode: pi gates by tool declaration, not per-call
+		// approval, so "all tools auto-approved" is the full builtin set.
+		args = append(args, "--tools", "read,bash,edit,write,grep,find,ls")
+	} else {
 		args = append(args, "--tools", mapTools(spec.AllowedTools))
 	}
 	args = append(args, "--", spec.Prompt)
diff --git a/internal/runner/runner.go b/internal/runner/runner.go
index 808bb59..3c1a0d2 100644
--- a/internal/runner/runner.go
+++ b/internal/runner/runner.go
@@ -118,6 +118,11 @@ func Run(ctx context.Context, opts Options, sink Sink) (*RunResult, error) {
 	runID := time.Now().Format("20060102-150405") + "-" + skill.Name
 	res := &RunResult{RunID: runID, Skill: skill.Name, Harness: h.ID(), ExitCode: ExitOK}
 
+	// Full-permission consent: the per-run flag OR the init-time consent
+	// recorded in state (the SDK Provider path has no cmd layer to resolve
+	// it).
+	unsafeFullPermission := opts.UnsafeFullPermission || state.FullPermission
+
 	// Pre-flight: model floor.
 	pf := ResolveModel(ctx, opts.API, skill, opts.Model)
 	res.Preflight = pf
@@ -147,7 +152,7 @@ func Run(ctx context.Context, opts Options, sink Sink) (*RunResult, error) {
 		Harness: h.ID(), Model: pf.Model,
 		RegistryRole: pf.Role, FloorModel: pf.ResolvedModel, ModelWarning: pf.Warning,
 		BudgetCap: res.BudgetCap, BudgetOverride: opts.BudgetOverrideReason,
-		UnsafeFullPermission: opts.UnsafeFullPermission,
+		UnsafeFullPermission: unsafeFullPermission,
 		TraustRef:            state.TraustRef, TraustSHA: state.TraustSHA,
 	})
 	if err != nil {
@@ -185,6 +190,9 @@ func Run(ctx context.Context, opts Options, sink Sink) (*RunResult, error) {
 	for _, note := range h.Notes() {
 		sink.Progress("note: " + note)
 	}
+	if unsafeFullPermission {
+		sink.Progress("note: full permission mode (consented) — the harness runs with all tools auto-approved; S9 isolation, budget bands and the run journal still apply")
+	}
 
 	// Launch.
 	runCtx, cancel := context.WithCancel(ctx)
@@ -192,7 +200,7 @@ func Run(ctx context.Context, opts Options, sink Sink) (*RunResult, error) {
 	spec := &harness.RunSpec{
 		Prompt: prompt, Model: pf.Model, Cwd: root, Env: state.ChildEnv(),
 		AllowedTools: skill.AllowedTools, MaxTurns: opts.MaxTurns,
-		UnsafeFullPermission: opts.UnsafeFullPermission, OutDir: journal.Dir,
+		UnsafeFullPermission: unsafeFullPermission, OutDir: journal.Dir,
 	}
 
 	// handleEvent folds one normalized event into journal, sink, result and
@@ -298,7 +306,7 @@ func Run(ctx context.Context, opts Options, sink Sink) (*RunResult, error) {
 		sink.Progress(fmt.Sprintf("budget refuse band hit (%d output tokens, cap %d) — harness terminated; re-run with --budget-override \"<reason>\" if this is expected", res.Usage.TokensOut, res.BudgetCap))
 	case res.NeedsInput:
 		res.ExitCode = ExitNeedsInput
-		sink.Progress("the skill asked for interactive input or a permission grant, which a headless run cannot provide — run it interactively in your harness and answer its questions, then re-run")
+		sink.Progress("the skill raised a permission request or question that a headless run cannot answer — enable full-permission mode (re-run `traust-cli init`, or --unsafe-full-permission for this run), or pre-allow the tool in your harness's config")
 	case ctx.Err() != nil:
 		res.Aborted = true
 		res.ExitCode = ExitError
diff --git a/internal/runner/session_test.go b/internal/runner/session_test.go
index 967c3ed..5cf4a7d 100644
--- a/internal/runner/session_test.go
+++ b/internal/runner/session_test.go
@@ -35,6 +35,9 @@ func fakeServeCrush(t *testing.T, script string) {
 	mux.HandleFunc("POST /v1/workspaces/{id}/current-session", func(w http.ResponseWriter, r *http.Request) {
 		w.WriteHeader(200)
 	})
+	mux.HandleFunc("POST /v1/workspaces/{id}/permissions/skip", func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(200)
+	})
 	mux.HandleFunc("POST /v1/workspaces/{id}/config/model", func(w http.ResponseWriter, r *http.Request) {
 		w.WriteHeader(200)
 	})
@@ -160,6 +163,46 @@ data: {"type":"run_complete","payload":{"type":"updated","payload":{"session_id"
 	}
 }
 
+func TestRunSessionModeFullPermissionFromState(t *testing.T) {
+	fakeServeCrush(t, happyServeScript)
+	state := testState(t)
+	state.FullPermission = true
+	if err := state.Save(); err != nil {
+		t.Fatal(err)
+	}
+	sink := &recSink{}
+
+	skill := testSkill()
+	skill.Standalone = &catalog.StandaloneInfo{MinModelClass: "opus"}
+
+	res, err := Run(context.Background(), Options{
+		State: state, API: &fakeAPI{}, Harness: &harness.Crush{},
+		Skill: skill, DeclareSpend: true,
+	}, sink)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.ExitCode != ExitOK {
+		t.Fatalf("exit = %d (%+v)", res.ExitCode, res)
+	}
+	noted := false
+	for _, p := range sink.progress {
+		if strings.Contains(p, "full permission mode") {
+			noted = true
+		}
+	}
+	if !noted {
+		t.Error("full-permission note not surfaced")
+	}
+	specData, err := os.ReadFile(filepath.Join(res.Dir, "spec.json"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(specData), `"unsafe_full_permission": true`) {
+		t.Errorf("journal spec does not record full permission: %s", specData)
+	}
+}
+
 func TestRunSessionModeRunError(t *testing.T) {
 	fakeServeCrush(t, `{"type":"run_complete","payload":{"type":"updated","payload":{"session_id":"sess-1","run_id":"{RUN_ID}","error":"provider 401"}}}`)
 	state := testState(t)
diff --git a/internal/tui/init.go b/internal/tui/init.go
index 35f7f8d..aa7847d 100644
--- a/internal/tui/init.go
+++ b/internal/tui/init.go
@@ -11,6 +11,7 @@ type InitAnswers struct {
 	Ref              string
 	ToolchainProfile string
 	Harness          string
+	FullPermission   bool
 }
 
 // RunInitWizard asks the interactive setup questions. detected lists the
@@ -70,6 +71,16 @@ func RunInitWizard(defaultWorkspace, defaultForge string, detected []string, def
 				Options(harnessOptions...).
 				Value(&answers.Harness),
 		),
+		huh.NewGroup(
+			huh.NewConfirm().
+				Title("Run skills with full permissions?").
+				Description("The traust harness's skills REQUIRE unattended tool use (bash, file writes) — "+
+					"without full permission they fail closed on the first permission request. "+
+					"With it, your harness runs with all tools auto-approved "+
+					"(claude --dangerously-skip-permissions, crush --yolo, codex --dangerously-bypass-approvals-and-sandbox). "+
+					"S9 isolation still applies: agents run from the harness root and never load the audit target's configuration.").
+				Value(&answers.FullPermission),
+		),
 	)
 	if err := form.Run(); err != nil {
 		return nil, err