diff --git a/CHANGELOG.md b/CHANGELOG.md
index 1cca33c9f06..640c05f903a 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,8 @@ All notable changes to NanoClaw will be documented in this file.
 
 ## [Unreleased]
 
+- **The guard seam.** Every privileged action crossing the container or channel boundary now passes one decision function — `guard()` in `src/guard/` — before it executes: `allow`, `hold` (the existing approval flows), or `deny`. Today's checks are preserved verbatim as each action's decision; ncl commands and delivery actions cannot register without a guard, and approved replays re-enter carrying the approval row as a **grant** (the forgeable `approved: true` boolean is deleted) with the checks re-run live. Two deliberate outcome changes: **(a)** approving a held a2a message after its destination was revoked no longer delivers it — the requester is told "approved, but not delivered" and the host logs a warning; **(b)** a forged, already-consumed, or mismatched grant refuses the replay instead of executing.
+- **Delivery-registry hardening.** Re-registering a guard-wrapped delivery action *without* a guard spec now throws instead of silently disarming the guard.
 - [BREAKING] **`whatsapp-formatting` and `slack-formatting` container skills moved from trunk to the `channels` branch.** They now install with their channel — `/add-whatsapp` / `/add-slack` and the setup installers copy them in — so installs without those channels stop carrying channel-specific formatting instructions in every agent's context. **Migration — only if this install has the channel wired** (check `src/channels/whatsapp.ts` / `src/channels/slack.ts`): updating removes both skills from the working tree — WhatsApp agents lose the formatting fragment from their composed CLAUDE.md on next spawn, Slack agents lose the mrkdwn skill from `~/.claude/skills`. Re-run the matching skill, `/add-whatsapp` or `/add-slack` (idempotent), to restore. Installs without the channel need nothing — do NOT run the add-skill just in case; it installs the full channel adapter.
 - **Pre-task script failures back their series off instead of spinning.** A `--script` that errors lands the occurrence as a failed run (`script-skip:error` ack → `failed` status); recurrence reads the series' trailing failed streak and re-arms at `max(cron next, now + 2·2^(n−1) min, cap 60)`; after 8 consecutive failures the series is auto-paused with a host-written note in its run log (`ncl tasks resume` revives it). A deliberate `wakeAgent:false` gate is a normal run and never backs off. Also fixed: an explicitly-addressed `<message to>` in a task fire's final text now delivers as a deliberate send (previously suppressed as a turn-reply echo → zero delivery when the agent skipped the MCP tool); identical echoes of an MCP send are dropped in the runner, where the duplication originates.
 - [BREAKING] **Scheduled tasks moved from MCP tools to `ncl tasks`.** The six scheduling MCP tools are no longer exposed to agent containers; agents and operators manage tasks with `ncl tasks list/get/create/update/cancel/pause/resume/delete`. New tasks run from a per-agent-group system session rather than waking the chat session that created them, and task writes are not approval-gated inside the owning group. **Migration:** [docs/ncl-tasks-migration.md](docs/ncl-tasks-migration.md).
diff --git a/CLAUDE.md b/CLAUDE.md
index 4416e14c817..0772957a1eb 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -62,10 +62,12 @@ For ad-hoc queries from skills or scripts, use the in-tree wrapper rather than t
 | `src/index.ts` | Entry point: init DB, migrations, channel adapters, delivery polls, sweep, shutdown |
 | `src/router.ts` | Inbound routing: messaging group → agent group → session → `inbound.db` → wake |
 | `src/delivery.ts` | Polls `outbound.db`, delivers via adapter, handles system actions (schedule, approvals, etc.) |
+| `src/delivery-guard.ts` | `DeliveryGuardSpec` + `runGuarded` — the guard-consult pipeline for privileged delivery actions (registry stays in `delivery.ts`) |
 | `src/host-sweep.ts` | 60s sweep: `processing_ack` sync, stale detection, due-message wake, recurrence |
 | `src/session-manager.ts` | Resolves sessions; opens `inbound.db` / `outbound.db`; manages heartbeat path |
 | `src/container-runner.ts` | Spawns per-agent-group Docker containers with session DB + outbox mounts, OneCLI `ensureAgent` |
 | `src/container-runtime.ts` | Docker CLI wrapper (runtime binary, host-gateway args, mount args), orphan cleanup |
+| `src/guard/` | Privileged-action decision seam: `guard(action, input)` → allow \| hold \| deny. Module-edge `guard.ts` adapters (cli, agent-to-agent, self-mod, permissions) define each action's decision; ncl commands + delivery actions demand a guard at registration; approved replays carry the approval row as a grant and re-run the checks. Conformance test: `src/guard/conformance.test.ts` |
 | `src/modules/permissions/access.ts` | `canAccessAgentGroup` — owner / global admin / scoped admin / member resolution against `user_roles` + `agent_group_members` |
 | `src/modules/approvals/primitive.ts` | `pickApprover`, `pickApprovalDelivery`, `requestApproval`, approval-handler registry |
 | `src/command-gate.ts` | Router-side admin command gate — queries `user_roles` directly (no env var, no container-side check) |
diff --git a/src/cli/crud.ts b/src/cli/crud.ts
index fd786232943..587f5be1cd8 100644
--- a/src/cli/crud.ts
+++ b/src/cli/crud.ts
@@ -402,6 +402,7 @@ export function registerResource(def: ResourceDef): void {
   if (def.operations.list) {
     register({
       name: `${def.plural}-list`,
+      action: `${def.plural}.list`,
       description: `List all ${def.plural}.`,
       access: def.operations.list,
       resource: def.plural,
@@ -414,6 +415,7 @@ export function registerResource(def: ResourceDef): void {
   if (def.operations.get) {
     register({
       name: `${def.plural}-get`,
+      action: `${def.plural}.get`,
       description: `Get a ${def.name} by ID.`,
       access: def.operations.get,
       resource: def.plural,
@@ -426,6 +428,7 @@ export function registerResource(def: ResourceDef): void {
   if (def.operations.create) {
     register({
       name: `${def.plural}-create`,
+      action: `${def.plural}.create`,
       description: `Create a new ${def.name}.`,
       access: def.operations.create,
       resource: def.plural,
@@ -437,6 +440,7 @@ export function registerResource(def: ResourceDef): void {
   if (def.operations.update) {
     register({
       name: `${def.plural}-update`,
+      action: `${def.plural}.update`,
       description: `Update a ${def.name}.`,
       access: def.operations.update,
       resource: def.plural,
@@ -448,6 +452,7 @@ export function registerResource(def: ResourceDef): void {
   if (def.operations.delete) {
     register({
       name: `${def.plural}-delete`,
+      action: `${def.plural}.delete`,
       description: `Delete a ${def.name}.`,
       access: def.operations.delete,
       resource: def.plural,
@@ -464,6 +469,7 @@ export function registerResource(def: ResourceDef): void {
       const declared = op.args;
       register({
         name: `${def.plural}-${verb.replace(/ /g, '-')}`,
+        action: `${def.plural}.${verb.replace(/ /g, '.')}`,
         description: op.description,
         access: op.access,
         resource: def.plural,
diff --git a/src/cli/delivery-action.ts b/src/cli/delivery-action.ts
index 5c693be6f1c..5a759cb21de 100644
--- a/src/cli/delivery-action.ts
+++ b/src/cli/delivery-action.ts
@@ -8,52 +8,57 @@
 import type Database from 'better-sqlite3';
 
 import { registerDeliveryAction } from '../delivery.js';
+import { unguarded } from '../guard/index.js';
 import { insertMessage } from '../db/session-db.js';
 import { log } from '../log.js';
 import { dispatch } from './dispatch.js';
 import type { RequestFrame } from './frame.js';
 import type { Session } from '../types.js';
 
-registerDeliveryAction('cli_request', async (content, session, inDb) => {
-  const requestId = content.requestId as string;
-  const command = content.command as string;
-  const args = (content.args as Record<string, unknown>) ?? {};
-
-  if (!requestId || !command) {
-    log.warn('cli_request missing requestId or command', { sessionId: session.id });
-    return;
-  }
-
-  const req: RequestFrame = { id: requestId, command, args };
-  const ctx = {
-    caller: 'agent' as const,
-    sessionId: session.id,
-    agentGroupId: session.agent_group_id,
-    messagingGroupId: session.messaging_group_id ?? '',
-  };
-
-  log.info('CLI request from agent', { requestId, command, sessionId: session.id });
-
-  const response = await dispatch(req, ctx);
-
-  // Write response to inbound.db so the container can read it.
-  // trigger=0: don't wake the agent — this is an inline response to a tool call.
-  insertMessage(inDb, {
-    id: `cli-resp-${requestId}`,
-    kind: 'system',
-    timestamp: new Date().toISOString(),
-    platformId: null,
-    channelType: null,
-    threadId: null,
-    content: JSON.stringify({
-      type: 'cli_response',
-      requestId,
-      frame: response,
-    }),
-    processAfter: null,
-    recurrence: null,
-    trigger: 0,
-  });
-
-  log.info('CLI response written', { requestId, ok: response.ok, sessionId: session.id });
-});
+registerDeliveryAction(
+  'cli_request',
+  async (content, session, inDb) => {
+    const requestId = content.requestId as string;
+    const command = content.command as string;
+    const args = (content.args as Record<string, unknown>) ?? {};
+
+    if (!requestId || !command) {
+      log.warn('cli_request missing requestId or command', { sessionId: session.id });
+      return;
+    }
+
+    const req: RequestFrame = { id: requestId, command, args };
+    const ctx = {
+      caller: 'agent' as const,
+      sessionId: session.id,
+      agentGroupId: session.agent_group_id,
+      messagingGroupId: session.messaging_group_id ?? '',
+    };
+
+    log.info('CLI request from agent', { requestId, command, sessionId: session.id });
+
+    const response = await dispatch(req, ctx);
+
+    // Write response to inbound.db so the container can read it.
+    // trigger=0: don't wake the agent — this is an inline response to a tool call.
+    insertMessage(inDb, {
+      id: `cli-resp-${requestId}`,
+      kind: 'system',
+      timestamp: new Date().toISOString(),
+      platformId: null,
+      channelType: null,
+      threadId: null,
+      content: JSON.stringify({
+        type: 'cli_response',
+        requestId,
+        frame: response,
+      }),
+      processAfter: null,
+      recurrence: null,
+      trigger: 0,
+    });
+
+    log.info('CLI response written', { requestId, ok: response.ok, sessionId: session.id });
+  },
+  unguarded('transport envelope — every inner command is guarded at dispatch'),
+);
diff --git a/src/cli/dispatch.test.ts b/src/cli/dispatch.test.ts
index 4da3c1dca1a..558154d3770 100644
--- a/src/cli/dispatch.test.ts
+++ b/src/cli/dispatch.test.ts
@@ -9,6 +9,7 @@ const approvalState = vi.hoisted(() => ({
     | ((args: {
         session: unknown;
         payload: Record<string, unknown>;
+        approval: Record<string, unknown>;
         userId: string;
         notify: (text: string) => void;
       }) => Promise<void>),
@@ -18,6 +19,7 @@ const approvalState = vi.hoisted(() => ({
       handler: (args: {
         session: unknown;
         payload: Record<string, unknown>;
+        approval: Record<string, unknown>;
         userId: string;
         notify: (text: string) => void;
       }) => Promise<void>,
@@ -43,8 +45,11 @@ vi.mock('../db/agent-groups.js', () => ({
 }));
 
 const mockGetSession = vi.fn();
+// The guard's grant check re-fetches the approval row to prove it's live.
+const mockGetPendingApproval = vi.fn();
 vi.mock('../db/sessions.js', () => ({
   getSession: (...args: unknown[]) => mockGetSession(...args),
+  getPendingApproval: (...args: unknown[]) => mockGetPendingApproval(...args),
 }));
 
 // dispatch's post-handler looks up the resource's `scopeField` via getResource.
@@ -495,10 +500,20 @@ describe('CLI scope enforcement', () => {
       callerContext: ctx,
     });
 
+    // The approve path hands the handler the live approval row — the grant
+    // the replay carries back into dispatch.
+    const grantRow = {
+      approval_id: 'appr-t1',
+      action: 'cli_command',
+      payload: JSON.stringify(approval.payload),
+    };
+    mockGetPendingApproval.mockReturnValue(grantRow);
+
     expect(approvalState.approvalHandler).toBeTypeOf('function');
     await approvalState.approvalHandler!({
       session: { id: 's1', agent_group_id: 'g1', messaging_group_id: 'mg1' },
       payload: approval.payload,
+      approval: grantRow,
       userId: 'telegram:admin',
       notify: vi.fn(),
     });
@@ -507,6 +522,73 @@ describe('CLI scope enforcement', () => {
     expect(approvalState.requestApproval).toHaveBeenCalledTimes(1);
   });
 
+  // --- Grant-carrying replay (the `approved: true` boolean no longer exists) ---
+
+  it('replay with a dead grant (row deleted) refuses instead of re-holding', async () => {
+    mockGetContainerConfig.mockReturnValue({ cli_scope: 'group' });
+    mockGetSession.mockReturnValue({ id: 's1', agent_group_id: 'g1', messaging_group_id: 'mg1' });
+    mockGetAgentGroup.mockReturnValue({ id: 'g1', name: 'Group One' });
+
+    const ctx = agentCtx();
+    await dispatch({ id: '1', command: 'approval-context-command', args: {} }, ctx);
+    const approval = approvalState.requestApproval.mock.calls[0][0] as { payload: Record<string, unknown> };
+
+    mockGetPendingApproval.mockReturnValue(undefined); // resolution already deleted the row
+    const notify = vi.fn();
+    await approvalState.approvalHandler!({
+      session: { id: 's1', agent_group_id: 'g1', messaging_group_id: 'mg1' },
+      payload: approval.payload,
+      approval: { approval_id: 'appr-dead', action: 'cli_command', payload: JSON.stringify(approval.payload) },
+      userId: 'telegram:admin',
+      notify,
+    });
+
+    expect(approvalState.observedContexts).toHaveLength(0); // handler never ran
+    expect(approvalState.requestApproval).toHaveBeenCalledTimes(1); // no second card
+    expect(notify.mock.calls[0][0]).toContain('failed');
+  });
+
+  it("a grant approved for one command doesn't transfer to another", async () => {
+    mockGetContainerConfig.mockReturnValue({ cli_scope: 'group' });
+
+    // A live cli_command row, but held for a DIFFERENT command.
+    const grantRow = {
+      approval_id: 'appr-other',
+      action: 'cli_command',
+      payload: JSON.stringify({ frame: { command: 'members-add' } }),
+    };
+    mockGetPendingApproval.mockReturnValue(grantRow);
+
+    const resp = await dispatch({ id: '1', command: 'approval-context-command', args: {} }, agentCtx(), {
+      grant: grantRow as never,
+    });
+
+    expect(resp.ok).toBe(false);
+    if (!resp.ok) {
+      expect(resp.error.code).toBe('forbidden');
+      expect(resp.error.message).toContain('grant');
+    }
+    expect(approvalState.observedContexts).toHaveLength(0);
+  });
+
+  it('a fabricated grant object without a live row is refused', async () => {
+    mockGetContainerConfig.mockReturnValue({ cli_scope: 'group' });
+    mockGetPendingApproval.mockReturnValue(undefined);
+
+    const forged = {
+      approval_id: 'appr-forged',
+      action: 'cli_command',
+      payload: JSON.stringify({ frame: { command: 'approval-context-command' } }),
+    };
+    const resp = await dispatch({ id: '1', command: 'approval-context-command', args: {} }, agentCtx(), {
+      grant: forged as never,
+    });
+
+    expect(resp.ok).toBe(false);
+    if (!resp.ok) expect(resp.error.code).toBe('forbidden');
+    expect(approvalState.requestApproval).not.toHaveBeenCalled(); // refusal, not a fresh hold
+  });
+
   // --- Post-handler filtering ---
 
   it('group: groups list filters out other groups', async () => {
diff --git a/src/cli/dispatch.ts b/src/cli/dispatch.ts
index 3885e931ea2..290756c81bc 100644
--- a/src/cli/dispatch.ts
+++ b/src/cli/dispatch.ts
@@ -3,24 +3,38 @@
  * the per-session DB poller (container caller) call dispatch() with the
  * same frame and a transport-supplied CallerContext.
  *
- * Approval gating for risky calls from the container is the only branch
- * that differs by caller. Host callers and `open` commands run inline.
+ * Every command passes the guard before its handler runs — the decision
+ * (allow / hold / deny) comes from the command's catalog entry, derived at
+ * registration (see cli/guard.ts). Dispatch keeps the mechanics: arg
+ * auto-fill, the sessions-get existence oracle, `--help` interception,
+ * parseArgs, and post-handler row filtering. An approved replay re-enters
+ * here carrying the verified approval row as its grant — the guard re-checks
+ * the structural checks live, and the `approved: true` boolean no longer
+ * exists.
  */
 import { getContainerConfig } from '../db/container-configs.js';
 import { getAgentGroup } from '../db/agent-groups.js';
 import { getSession } from '../db/sessions.js';
+import { guard, type GuardActor } from '../guard/index.js';
 import { registerApprovalHandler, requestApproval } from '../modules/approvals/index.js';
+import type { PendingApproval } from '../types.js';
 import type { CallerContext, ErrorCode, RequestFrame, ResponseFrame } from './frame.js';
 import { localizeIsoTimestamps } from './format.js';
 import { getResource } from './crud.js';
 import { listVerbs, renderVerbHelp } from './help-render.js';
-import { GROUP_SCOPE_RESOURCES, listCommands, lookup } from './registry.js';
+import { commandGuard, listCommands, lookup } from './registry.js';
 
 type DispatchOptions = {
-  /** True when a command is being replayed after approval. */
-  approved?: boolean;
+  /** Verified approval row when a command is replayed after approval. */
+  grant?: PendingApproval;
 };
 
+function actorFor(ctx: CallerContext): GuardActor {
+  return ctx.caller === 'host'
+    ? { kind: 'host' }
+    : { kind: 'agent', agentGroupId: ctx.agentGroupId, sessionId: ctx.sessionId };
+}
+
 export async function dispatch(
   req: RequestFrame,
   ctx: CallerContext,
@@ -55,43 +69,13 @@ export async function dispatch(
     return err(req.id, 'unknown-command', unknownCommandMessage(req.command));
   }
 
-  // CLI scope enforcement for agent callers
+  // Group-scope mechanics for agent callers (visibility, not policy — the
+  // allow/hold/deny decisions live in the guard decision, cli/guard.ts).
   if (ctx.caller === 'agent') {
     const configRow = getContainerConfig(ctx.agentGroupId);
     const cliScope = configRow?.cli_scope ?? 'group';
 
-    if (cliScope === 'disabled') {
-      return err(req.id, 'forbidden', 'CLI access is disabled for this agent group.');
-    }
-
     if (cliScope === 'group') {
-      // Only allow whitelisted resources and general commands (no resource, like help)
-      if (cmd.resource && !GROUP_SCOPE_RESOURCES.has(cmd.resource)) {
-        return err(req.id, 'forbidden', `CLI access is scoped to this agent group. Cannot access "${cmd.resource}".`);
-      }
-
-      // Enforce group scope on all agent-group-related args.
-      // Different resources use different arg names for the agent group ID.
-      // Only check --id for resources where it IS the agent group ID.
-      const groupArgs = ['agent_group_id', 'group'] as const;
-      for (const key of groupArgs) {
-        if (req.args[key] && req.args[key] !== ctx.agentGroupId) {
-          return err(req.id, 'forbidden', 'CLI access is scoped to this agent group.');
-        }
-      }
-      if (
-        (cmd.resource === 'groups' || cmd.resource === 'destinations') &&
-        req.args.id &&
-        req.args.id !== ctx.agentGroupId
-      ) {
-        return err(req.id, 'forbidden', 'CLI access is scoped to this agent group.');
-      }
-
-      // Block cli_scope changes from group-scoped agents (privilege escalation)
-      if (req.args.cli_scope !== undefined || req.args['cli-scope'] !== undefined) {
-        return err(req.id, 'forbidden', 'Cannot change cli_scope from a group-scoped agent.');
-      }
-
       // Auto-fill agent-group-related args so the agent doesn't need
       // to pass its own group ID explicitly.
       const fill: Record<string, unknown> = {
@@ -117,9 +101,19 @@ export async function dispatch(
     }
   }
 
+  const decision = guard(commandGuard(cmd.name), {
+    actor: actorFor(ctx),
+    payload: req.args,
+    grant: opts.grant ?? null,
+  });
+
+  if (decision.effect === 'deny') {
+    return err(req.id, 'forbidden', decision.reason);
+  }
+
   // `--help` interception: answer with the command's generated help instead of
-  // executing. Placed after scope enforcement (a group-scoped agent can't probe
-  // forbidden resources) and BEFORE approval gating — asking for help on an
+  // executing. Placed after the guard's deny (a group-scoped agent can't probe
+  // forbidden resources) and BEFORE hold execution — asking for help on an
   // approval-gated verb must never mint an approval card.
   if (req.args.help === true) {
     // Carry the help text in `human` too, so both clients print it verbatim
@@ -128,7 +122,12 @@ export async function dispatch(
     return { id: req.id, ok: true, data: helpText, human: helpText };
   }
 
-  if (ctx.caller !== 'host' && cmd.access === 'approval' && !opts.approved) {
+  if (decision.effect === 'hold') {
+    if (ctx.caller !== 'agent') {
+      // Holds only arise for agent callers; anything else is a guard bug —
+      // fail closed rather than card a ghost.
+      return err(req.id, 'forbidden', decision.reason);
+    }
     const session = getSession(ctx.sessionId);
     if (!session) {
       return err(req.id, 'handler-error', 'Session not found.');
@@ -216,10 +215,10 @@ export async function dispatch(
   }
 }
 
-registerApprovalHandler('cli_command', async ({ payload, notify }) => {
+registerApprovalHandler('cli_command', async ({ payload, approval, notify }) => {
   const frame = payload.frame as RequestFrame;
   const callerContext = parseCallerContext(payload.callerContext) ?? { caller: 'host' };
-  const response = await dispatch(frame, callerContext, { approved: true });
+  const response = await dispatch(frame, callerContext, { grant: approval });
 
   if (response.ok) {
     const localized = localizeIsoTimestamps(response.data);
diff --git a/src/cli/guard.ts b/src/cli/guard.ts
new file mode 100644
index 00000000000..49c8602b328
--- /dev/null
+++ b/src/cli/guard.ts
@@ -0,0 +1,86 @@
+/**
+ * CLI guard adapter — the command registry's catalog derivation and
+ * structural decision, moved verbatim out of dispatch.ts.
+ * Declaration is registration: registry.register() derives one
+ * catalog entry per command from the CommandDef itself; no second file is
+ * edited when a command is added.
+ *
+ * The decide fn carries today's decisions exactly:
+ *   host caller → allow (the 0600 socket is the auth story — in code,
+ *   unremovable by data);
+ *   cli_scope 'disabled' → deny; 'group' → resource allowlist, cross-group
+ *   arg denial, cli_scope-change denial;
+ *   access 'approval' for agent callers → hold for the group's admin chain.
+ *
+ * Arg auto-fill, the sessions-get existence oracle, and post-handler row
+ * filtering stay in dispatch.ts — mechanics, not policy.
+ */
+import { getContainerConfig } from '../db/container-configs.js';
+import { ALLOW, DENY, HOLD, type GuardedActionSpec, type GuardInput } from '../guard/index.js';
+import { GROUP_SCOPE_RESOURCES, type CommandDef } from './registry.js';
+
+/** Dotted catalog action name for a command. */
+export function commandGuardAction(cmd: Pick<CommandDef, 'name' | 'action'>): string {
+  return cmd.action ?? `cli.${cmd.name}`;
+}
+
+/** Catalog entry derived from a CommandDef at registration time. */
+export function commandGuardSpec(cmd: CommandDef): GuardedActionSpec {
+  return {
+    action: commandGuardAction(cmd),
+    grantActionName: cmd.access === 'approval' ? 'cli_command' : undefined,
+    // Bind a cli_command grant to the exact command it was approved for.
+    grantCoversRequest: (grant) => {
+      try {
+        const payload = JSON.parse(grant.payload) as { frame?: { command?: string } };
+        return payload.frame?.command === cmd.name;
+      } catch {
+        return false;
+      }
+    },
+    decide: (input) => commandDecide(cmd, input),
+  };
+}
+
+function commandDecide(cmd: CommandDef, input: GuardInput) {
+  const { actor } = input;
+  if (actor.kind === 'host') return ALLOW('host caller (trusted socket)');
+  if (actor.kind !== 'agent') return DENY('CLI commands accept host or agent callers only.');
+
+  const args = input.payload;
+  const cliScope = getContainerConfig(actor.agentGroupId)?.cli_scope ?? 'group';
+
+  if (cliScope === 'disabled') {
+    return DENY('CLI access is disabled for this agent group.');
+  }
+
+  if (cliScope === 'group') {
+    // Only allow whitelisted resources and general commands (no resource, like help)
+    if (cmd.resource && !GROUP_SCOPE_RESOURCES.has(cmd.resource)) {
+      return DENY(`CLI access is scoped to this agent group. Cannot access "${cmd.resource}".`);
+    }
+
+    // Enforce group scope on all agent-group-related args.
+    // Different resources use different arg names for the agent group ID.
+    // Only check --id for resources where it IS the agent group ID.
+    for (const key of ['agent_group_id', 'group'] as const) {
+      if (args[key] && args[key] !== actor.agentGroupId) {
+        return DENY('CLI access is scoped to this agent group.');
+      }
+    }
+    if ((cmd.resource === 'groups' || cmd.resource === 'destinations') && args.id && args.id !== actor.agentGroupId) {
+      return DENY('CLI access is scoped to this agent group.');
+    }
+
+    // Block cli_scope changes from group-scoped agents (privilege escalation)
+    if (args.cli_scope !== undefined || args['cli-scope'] !== undefined) {
+      return DENY('Cannot change cli_scope from a group-scoped agent.');
+    }
+  }
+
+  if (cmd.access === 'approval') {
+    return HOLD(`agent-initiated "${cmd.name}" requires admin approval`);
+  }
+
+  return ALLOW('open command');
+}
diff --git a/src/cli/registry.ts b/src/cli/registry.ts
index e108e114321..8a436ec3257 100644
--- a/src/cli/registry.ts
+++ b/src/cli/registry.ts
@@ -8,6 +8,8 @@
  * registers the help commands, so the registry is populated before the host's
  * CLI server accepts connections.
  */
+import { defineGuardedAction, type GuardedAction } from '../guard/index.js';
+import { commandGuardSpec } from './guard.js';
 import type { CallerContext } from './frame.js';
 
 /**
@@ -23,6 +25,13 @@ export type CommandDef<TArgs = unknown, TData = unknown> = {
   name: string;
   description: string;
   access: Access;
+  /**
+   * Dotted guard-catalog action name (e.g. `roles.grant`,
+   * `groups.config.add-mcp-server`). Set by registerResource from the
+   * resource + verb; commands registered directly (help) fall back to
+   * `cli.<name>`.
+   */
+  action?: string;
   /**
    * The group-scope whitelist key. Under `cli_scope: 'group'` the dispatcher
    * only lets an agent run commands whose `resource` is on the whitelist
@@ -52,12 +61,26 @@ export type CommandDef<TArgs = unknown, TData = unknown> = {
 };
 
 const registry = new Map<string, CommandDef>();
+const commandGuards = new Map<string, GuardedAction>();
 
 export function register<TArgs, TData>(def: CommandDef<TArgs, TData>): void {
   if (registry.has(def.name)) {
     throw new Error(`CLI command "${def.name}" already registered`);
   }
   registry.set(def.name, def as CommandDef);
+  // Declaration is registration: every command gets a guard-catalog entry
+  // derived from its own definition, in the same call that registers it — a
+  // command cannot exist without a guard, and dispatch consults it by value.
+  commandGuards.set(def.name, defineGuardedAction(commandGuardSpec(def as CommandDef)));
+}
+
+/** The guard defined for a registered command — total for anything register() accepted. */
+export function commandGuard(name: string): GuardedAction {
+  const g = commandGuards.get(name);
+  if (!g) {
+    throw new Error(`CLI command "${name}" has no guard — was it registered through register()?`);
+  }
+  return g;
 }
 
 export function lookup(name: string): CommandDef | undefined {
diff --git a/src/delivery-actions.test.ts b/src/delivery-actions.test.ts
index 6a89506cd5c..6bd06fc9067 100644
--- a/src/delivery-actions.test.ts
+++ b/src/delivery-actions.test.ts
@@ -4,7 +4,9 @@
  * `registerDeliveryAction` is the hook modules use to handle system-kind
  * outbound messages; `getDeliveryAction` is the read side that makes those
  * registrations behavior-testable. Goes red if either half of the registry
- * is removed or the two stop sharing the same map.
+ * is removed or the two stop sharing the same map. Every registration now
+ * carries a guard spec or an explicit unguarded(<reason>) declaration —
+ * omission is a type error.
  */
 import { describe, it, expect, vi } from 'vitest';
 
@@ -16,11 +18,14 @@ vi.mock('./container-runner.js', () => ({
 }));
 
 import { registerDeliveryAction, getDeliveryAction, type DeliveryActionHandler } from './delivery.js';
+import { defineGuardedAction, HOLD, unguarded } from './guard/index.js';
+
+const testUnguarded = unguarded('test — registry mechanics only');
 
 describe('delivery action registry', () => {
   it('getDeliveryAction returns the handler registerDeliveryAction registered', () => {
     const handler: DeliveryActionHandler = async () => {};
-    registerDeliveryAction('test_registry_action', handler);
+    registerDeliveryAction('test_registry_action', handler, testUnguarded);
     expect(getDeliveryAction('test_registry_action')).toBe(handler);
   });
 
@@ -31,8 +36,34 @@ describe('delivery action registry', () => {
   it('re-registering an action overwrites the previous handler', () => {
     const first: DeliveryActionHandler = async () => {};
     const second: DeliveryActionHandler = async () => {};
-    registerDeliveryAction('test_overwrite_action', first);
-    registerDeliveryAction('test_overwrite_action', second);
+    registerDeliveryAction('test_overwrite_action', first, testUnguarded);
+    registerDeliveryAction('test_overwrite_action', second, testUnguarded);
     expect(getDeliveryAction('test_overwrite_action')).toBe(second);
   });
+
+  it('refuses to replace a guard-wrapped action with an unguarded handler', () => {
+    const guardAction = defineGuardedAction({
+      action: 'test.guarded-overwrite',
+      decide: () => HOLD('t'),
+    });
+    registerDeliveryAction('test_guarded_overwrite', async () => {}, {
+      guardAction,
+      requestHold: async () => {},
+    });
+
+    // Disarming the guard by re-registering unguarded must throw — otherwise
+    // the action's catalog entry would still exist while the live path runs
+    // unguarded.
+    expect(() => registerDeliveryAction('test_guarded_overwrite', async () => {}, testUnguarded)).toThrow(
+      /disarm the guard/,
+    );
+
+    // Re-registering WITH a spec stays allowed (a legitimate replacement
+    // keeps the action guarded).
+    registerDeliveryAction('test_guarded_overwrite', async () => {}, {
+      guardAction,
+      requestHold: async () => {},
+    });
+    expect(getDeliveryAction('test_guarded_overwrite')).toBeDefined();
+  });
 });
diff --git a/src/delivery-guard.ts b/src/delivery-guard.ts
new file mode 100644
index 00000000000..e4a99d72554
--- /dev/null
+++ b/src/delivery-guard.ts
@@ -0,0 +1,63 @@
+/**
+ * The guard-consult path for privileged delivery actions.
+ *
+ * The registry itself — registration, lookup, approved-replay re-entry —
+ * stays in delivery.ts, close to main's shape. This file holds the new
+ * guard logic: the spec a privileged registration carries, and runGuarded,
+ * the precheck → guard → deny/hold/allow pipeline every consult runs.
+ */
+import { guard, type GuardedAction } from './guard/index.js';
+import { log } from './log.js';
+import type { PendingApproval, Session } from './types.js';
+
+/** Handler shape for guard-wrapped actions — must not touch inDb (replays run without one). */
+export type GuardedDeliveryHandler = (content: Record<string, unknown>, session: Session) => Promise<void>;
+
+export interface DeliveryGuardSpec {
+  /** Guard action consulted before the handler runs — the defined value, not a name. */
+  guardAction: GuardedAction;
+  /**
+   * Domain validation that runs before the guard — malformed requests are
+   * answered (notify) without ever creating a hold. Return false to stop.
+   */
+  precheck?: (content: Record<string, unknown>, session: Session) => boolean | Promise<boolean>;
+  /** Create the hold (the domain's requestApproval call — card text lives with the domain). */
+  requestHold: (content: Record<string, unknown>, session: Session) => Promise<void>;
+  /** Tell the requester about a deny. */
+  onDeny?: (content: Record<string, unknown>, session: Session, reason: string) => void;
+}
+
+/**
+ * Run a guarded delivery action: precheck, consult the guard, then route the
+ * decision — deny → onDeny, hold → requestHold, allow → handler. A fresh
+ * dispatch passes grant=null; an approved replay passes the approval row,
+ * which satisfies a hold but never a deny (the structural checks re-run
+ * live, so approve-then-revoke does not execute).
+ */
+export async function runGuarded(
+  action: string,
+  spec: DeliveryGuardSpec,
+  handler: GuardedDeliveryHandler,
+  content: Record<string, unknown>,
+  session: Session,
+  grant: PendingApproval | null,
+): Promise<void> {
+  if (spec.precheck && !(await spec.precheck(content, session))) return;
+
+  const decision = guard(spec.guardAction, {
+    actor: { kind: 'agent', agentGroupId: session.agent_group_id, sessionId: session.id },
+    payload: content,
+    grant,
+  });
+
+  if (decision.effect === 'deny') {
+    log.warn('Delivery action denied by guard', { action, reason: decision.reason });
+    spec.onDeny?.(content, session, decision.reason);
+    return;
+  }
+  if (decision.effect === 'hold') {
+    await spec.requestHold(content, session);
+    return;
+  }
+  await handler(content, session);
+}
diff --git a/src/delivery.ts b/src/delivery.ts
index d9a0d15f270..4ff8a935448 100644
--- a/src/delivery.ts
+++ b/src/delivery.ts
@@ -20,12 +20,14 @@ import {
   markDeliveryFailed,
   migrateDeliveredTable,
 } from './db/session-db.js';
+import { runGuarded, type DeliveryGuardSpec, type GuardedDeliveryHandler } from './delivery-guard.js';
+import { isUnguarded, type Unguarded } from './guard/index.js';
 import { log } from './log.js';
 import { normalizeOptions } from './channels/ask-question.js';
 import { clearOutbox, openInboundDb, openOutboundDb, readOutboxFiles } from './session-manager.js';
 import { pauseTypingRefreshAfterDelivery, setTypingAdapter } from './modules/typing/index.js';
 import type { OutboundFile } from './channels/adapter.js';
-import type { Session } from './types.js';
+import type { PendingApproval, Session } from './types.js';
 
 const ACTIVE_POLL_MS = 1000;
 const SWEEP_POLL_MS = 60_000;
@@ -393,14 +395,19 @@ async function deliverMessage(
  * Delivery action registry.
  *
  * Modules register handlers for system-kind outbound message actions via
- * `registerDeliveryAction`. Core checks the registry first in
- * `handleSystemAction` and falls through to the inline switch when no
- * handler is registered. The switch will shrink as modules are extracted
- * (scheduling, approvals, agent-to-agent) and eventually only its default
- * branch remains.
+ * `registerDeliveryAction`. Unknown actions log "Unknown system action".
  *
- * Default when no handler registered and the switch doesn't match: log
- * "Unknown system action" and return.
+ * Privileged delivery actions (create_agent, install_packages,
+ * add_mcp_server) register with a guard spec: every path to the handler body
+ * — dispatch, approved replay, test lookup — goes through the guard consult
+ * (allow / hold / deny), so there is no unguarded route to it. On approve,
+ * the continuation re-enters the same entry carrying the approval row as its
+ * grant (`reenterGuardedDeliveryAction`), so the structural checks are
+ * re-run live. Plain actions (the cli_request bridge — its inner
+ * commands are guarded at dispatch) register with an
+ * explicit `unguarded(<reason>)` declaration instead of a spec — omission is
+ * not representable, so the decision to run unguarded is visible, and
+ * justified, at the registration site.
  */
 export type DeliveryActionHandler = (
   content: Record<string, unknown>,
@@ -408,18 +415,70 @@ export type DeliveryActionHandler = (
   inDb: Database.Database,
 ) => Promise<void>;
 
-const actionHandlers = new Map<string, DeliveryActionHandler>();
+type DeliveryEntry =
+  | { guard: Unguarded; handler: DeliveryActionHandler }
+  | { guard: DeliveryGuardSpec; handler: GuardedDeliveryHandler };
 
-export function registerDeliveryAction(action: string, handler: DeliveryActionHandler): void {
-  if (actionHandlers.has(action)) {
+const deliveryActions = new Map<string, DeliveryEntry>();
+
+function isUnguardedEntry(entry: DeliveryEntry): entry is Extract<DeliveryEntry, { guard: Unguarded }> {
+  return isUnguarded(entry.guard);
+}
+
+export function registerDeliveryAction(action: string, handler: DeliveryActionHandler, unguardedDecl: Unguarded): void;
+export function registerDeliveryAction(action: string, handler: GuardedDeliveryHandler, spec: DeliveryGuardSpec): void;
+export function registerDeliveryAction(
+  action: string,
+  handler: DeliveryActionHandler | GuardedDeliveryHandler,
+  guardDecl: DeliveryGuardSpec | Unguarded,
+): void {
+  const existing = deliveryActions.get(action);
+  if (existing) {
+    // Replacing a guard-wrapped action with an unguarded handler would
+    // disarm the guard while its catalog entry still exists — refuse. A
+    // skill that wants to extend a guarded action must compose at the
+    // module's exported functions instead, or re-register with a guard spec
+    // of its own.
+    if (isUnguarded(guardDecl) && !isUnguardedEntry(existing)) {
+      throw new Error(
+        `delivery action "${action}" is guard-wrapped; re-registering it without a guard spec would disarm the guard`,
+      );
+    }
     log.warn('Delivery action handler overwritten', { action });
   }
-  actionHandlers.set(action, handler);
+  // The overloads pair each handler shape with its declaration; the merged
+  // implementation signature erases that pairing, hence the one cast.
+  deliveryActions.set(action, { guard: guardDecl, handler } as DeliveryEntry);
 }
 
-/** Look up a registered delivery-action handler. Lets module registrations be behavior-tested. */
+/**
+ * Approve continuation for a guard-wrapped delivery action: re-enter the
+ * entry with the approval row as the grant. The guard treats the grant as
+ * hold-satisfied but re-runs the structural checks, so approve-then-revoke
+ * does not execute. Domains register this as their approval handler in the
+ * same line that registers the action.
+ */
+export function reenterGuardedDeliveryAction(action: string) {
+  return async (ctx: { session: Session; payload: Record<string, unknown>; approval: PendingApproval }) => {
+    const entry = deliveryActions.get(action);
+    if (!entry || isUnguardedEntry(entry)) {
+      log.warn('Approved replay for an action that is not guard-wrapped — dropping', { action });
+      return;
+    }
+    await runGuarded(action, entry.guard, entry.handler, ctx.payload, ctx.session, ctx.approval);
+  };
+}
+
+/**
+ * The invocable for a registered action — the raw handler for unguarded
+ * entries, the guard-consulting path for guarded ones. Dispatch and tests
+ * both come through here; there is no route around the guard.
+ */
 export function getDeliveryAction(action: string): DeliveryActionHandler | undefined {
-  return actionHandlers.get(action);
+  const entry = deliveryActions.get(action);
+  if (!entry) return undefined;
+  if (isUnguardedEntry(entry)) return entry.handler;
+  return (content, session) => runGuarded(action, entry.guard, entry.handler, content, session, null);
 }
 
 /**
@@ -435,7 +494,7 @@ async function handleSystemAction(
   const action = content.action as string;
   log.info('System action from agent', { sessionId: session.id, action });
 
-  const registered = actionHandlers.get(action);
+  const registered = getDeliveryAction(action);
   if (registered) {
     await registered(content, session, inDb);
     return;
diff --git a/src/guard/conformance.test.ts b/src/guard/conformance.test.ts
new file mode 100644
index 00000000000..086f0a40db3
--- /dev/null
+++ b/src/guard/conformance.test.ts
@@ -0,0 +1,64 @@
+/**
+ * Guard conformance — checked with the real registries.
+ *
+ * The old registry walk is gone: an unmapped consult or an undeclared
+ * unguarded registration is now unconstructible — guard() takes the defined
+ * GuardedAction value (a dropped module-edge import or typo'd name is a
+ * compile error), and the keyed registries require a guard spec or an
+ * explicit unguarded(<reason>) declaration. What's left to verify is the
+ * cross-registry pairing the compiler can't see: every holding action has a
+ * registered approve continuation. (At runtime a missing continuation is
+ * handled loudly at click time — the requester is told no handler is
+ * installed; this test keeps the tree from shipping that state.)
+ */
+import { describe, expect, it } from 'vitest';
+
+// Production barrels — side-effect imports populate the real registries.
+import '../cli/commands/index.js';
+import '../modules/index.js';
+import '../cli/delivery-action.js';
+import '../cli/dispatch.js'; // registers the cli_command approval handler
+
+import { commandGuard, listCommands } from '../cli/registry.js';
+import { getApprovalHandler } from '../modules/approvals/primitive.js';
+import { defineGuardedAction, listGuardedActions } from './guard-actions.js';
+import { HOLD } from './types.js';
+
+describe('guard conformance', () => {
+  it('every holding action pairs with a registered approval handler', () => {
+    const holding = listGuardedActions().filter((spec) => spec.grantActionName);
+    expect(holding.length).toBeGreaterThan(0);
+
+    const dangling = holding.filter((spec) => !getApprovalHandler(spec.grantActionName as string));
+    expect(dangling.map((s) => s.action)).toEqual([]);
+  });
+
+  it('every mutating ncl command derives a guard that holds via cli_command', () => {
+    const mutating = listCommands().filter((cmd) => cmd.access === 'approval');
+    expect(mutating.length).toBeGreaterThan(0);
+
+    const wrong = mutating.filter((cmd) => commandGuard(cmd.name).grantActionName !== 'cli_command');
+    expect(wrong.map((c) => c.name)).toEqual([]);
+  });
+
+  it('the domain catalog entries are defined once the module barrels load', () => {
+    const actions = new Set(listGuardedActions().map((s) => s.action));
+    for (const expected of [
+      'agents.create',
+      'a2a.send',
+      'self_mod.install_packages',
+      'self_mod.add_mcp_server',
+      'senders.admit',
+      'channels.register',
+    ]) {
+      expect(actions.has(expected), `catalog is missing "${expected}"`).toBe(true);
+    }
+  });
+
+  it('defining the same action twice throws — names are the catalog key', () => {
+    defineGuardedAction({ action: 'test.dup-define', decide: () => HOLD('x') });
+    expect(() => defineGuardedAction({ action: 'test.dup-define', decide: () => HOLD('x') })).toThrow(
+      /already defined/,
+    );
+  });
+});
diff --git a/src/guard/guard-actions.ts b/src/guard/guard-actions.ts
new file mode 100644
index 00000000000..09f139a3ed0
--- /dev/null
+++ b/src/guard/guard-actions.ts
@@ -0,0 +1,73 @@
+/**
+ * The action catalog — the enforcement boundary.
+ *
+ * An action either is defined here (and every consult passes its decision)
+ * or cannot be consulted at all: guard() takes the GuardedAction VALUE
+ * returned by defineGuardedAction, so the wiring between a consult site and
+ * its decide fn is a symbol reference the compiler checks. A dropped
+ * module-edge import or a typo'd action name is a build error, not a
+ * runtime fail-open — there is no lookup that can miss.
+ *
+ * Definitions are still recorded by name so the catalog can be enumerated:
+ * the conformance test pairs every holding action with its registered
+ * approval handler, and duplicate names are refused at definition time
+ * (grants match on the name).
+ */
+import type { GuardDecision, GuardInput } from './types.js';
+import type { PendingApproval } from '../types.js';
+
+export interface GuardedActionSpec {
+  /** Dotted action name, e.g. 'roles.grant', 'agents.create', 'a2a.send'. */
+  action: string;
+  /**
+   * Today's structural checks for this action, verbatim — the only source of
+   * allow. Runs on every consult, including approved replays (a grant
+   * satisfies a hold, never a deny).
+   */
+  decide: (input: GuardInput) => GuardDecision;
+  /**
+   * The pending_approvals.action its holds resolve through — a grant is only
+   * accepted when its row carries this action. Omit for actions that can
+   * never be held (deny/allow-only decisions).
+   */
+  grantActionName?: string;
+  /**
+   * Extra domain binding between a grant and the replayed input (e.g. the
+   * a2a target must match the held message). Runs in addition to the
+   * grantActionName + live-row checks.
+   */
+  grantCoversRequest?: (grant: PendingApproval, input: GuardInput) => boolean;
+}
+
+declare const guardedActionBrand: unique symbol;
+/**
+ * A defined guarded action — only defineGuardedAction can mint one. The
+ * brand makes the type nominal: a hand-rolled { action, decide } object
+ * does not typecheck at a consult site, and fails the runtime check too.
+ */
+export type GuardedAction = Readonly<GuardedActionSpec> & { readonly [guardedActionBrand]: true };
+
+const defined = new Map<string, GuardedAction>();
+const minted = new WeakSet<object>();
+
+export function defineGuardedAction(spec: GuardedActionSpec): GuardedAction {
+  if (defined.has(spec.action)) {
+    throw new Error(`guarded action "${spec.action}" is already defined — action names are the catalog key`);
+  }
+  const def = Object.freeze({ ...spec }) as GuardedAction;
+  minted.add(def);
+  defined.set(spec.action, def);
+  return def;
+}
+
+/**
+ * Runtime backstop for callers outside the type system (plain JS, casts):
+ * only values minted by defineGuardedAction pass — guard() denies the rest.
+ */
+export function isGuardedAction(value: unknown): value is GuardedAction {
+  return typeof value === 'object' && value !== null && minted.has(value);
+}
+
+export function listGuardedActions(): GuardedAction[] {
+  return [...defined.values()].sort((a, b) => a.action.localeCompare(b.action));
+}
diff --git a/src/guard/guard.test.ts b/src/guard/guard.test.ts
new file mode 100644
index 00000000000..11b3f909057
--- /dev/null
+++ b/src/guard/guard.test.ts
@@ -0,0 +1,160 @@
+/**
+ * Guard decision-function unit tests: decide is the decision (allow /
+ * hold / deny returned as-is), grant semantics (satisfies holds, never
+ * denies; invalid → refuse), the runtime backstop against forged action
+ * values, and the fail-closed posture on a throwing decide.
+ *
+ * Uses synthetic actions defined per test — the catalog is per-worker module
+ * state with no reset, so action names are unique.
+ */
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import { guard } from './guard.js';
+import { defineGuardedAction, type GuardedAction } from './guard-actions.js';
+import { ALLOW, DENY, HOLD, type GuardInput } from './types.js';
+
+const mockGetPendingApproval = vi.fn();
+vi.mock('../db/sessions.js', () => ({
+  getPendingApproval: (...args: unknown[]) => mockGetPendingApproval(...args),
+}));
+vi.mock('../log.js', () => ({
+  log: { info: vi.fn(), warn: vi.fn(), error: vi.fn(), debug: vi.fn() },
+}));
+
+const AGENT = { kind: 'agent', agentGroupId: 'ag-1', sessionId: 'sess-1' } as const;
+
+function input(extra: Partial<GuardInput> = {}): GuardInput {
+  return { actor: AGENT, payload: {}, ...extra };
+}
+
+beforeEach(() => {
+  mockGetPendingApproval.mockReset();
+});
+
+afterEach(() => {
+  vi.clearAllMocks();
+});
+
+describe('decide is the decision', () => {
+  it('decide allow → allow', () => {
+    const action = defineGuardedAction({ action: 't.allow1', decide: () => ALLOW('ok') });
+    expect(guard(action, input()).effect).toBe('allow');
+  });
+
+  it('decide hold → hold, default approver chain', () => {
+    const action = defineGuardedAction({ action: 't.hold1', decide: () => HOLD('needs approval') });
+    const d = guard(action, input());
+    expect(d.effect).toBe('hold');
+    if (d.effect === 'hold') {
+      expect(d.reason).toBe('needs approval');
+      expect(d.approverUserId).toBeUndefined();
+    }
+  });
+
+  it('decide hold → hold, carrying a named approver', () => {
+    const action = defineGuardedAction({ action: 't.hold2', decide: () => HOLD('policy row', 'telegram:dana') });
+    const d = guard(action, input());
+    expect(d.effect).toBe('hold');
+    if (d.effect === 'hold') expect(d.approverUserId).toBe('telegram:dana');
+  });
+
+  it('decide deny → deny, carrying the reason', () => {
+    const action = defineGuardedAction({ action: 't.deny1', decide: () => DENY('structurally unauthorized') });
+    const d = guard(action, input());
+    expect(d.effect).toBe('deny');
+    if (d.effect === 'deny') expect(d.reason).toBe('structurally unauthorized');
+  });
+
+  it('a forged action value (not from defineGuardedAction) is denied', () => {
+    const forged = { action: 't.forged', decide: () => ALLOW('never vetted') } as unknown as GuardedAction;
+    const d = guard(forged, input());
+    expect(d.effect).toBe('deny');
+    if (d.effect === 'deny') expect(d.reason).toContain('undefined action');
+  });
+});
+
+describe('grants', () => {
+  const grantRow = (action: string) =>
+    ({ approval_id: 'appr-1', action, payload: '{}' }) as unknown as NonNullable<GuardInput['grant']>;
+
+  it('a valid live grant satisfies a hold', () => {
+    const action = defineGuardedAction({
+      action: 't.g1',
+      grantActionName: 'g1_approved',
+      decide: () => HOLD('b'),
+    });
+    const grant = grantRow('g1_approved');
+    mockGetPendingApproval.mockReturnValue(grant);
+    expect(guard(action, input({ grant })).effect).toBe('allow');
+  });
+
+  it('a grant never satisfies a deny — the checks re-run live', () => {
+    const action = defineGuardedAction({
+      action: 't.g2',
+      grantActionName: 'g2_approved',
+      decide: () => DENY('revoked since'),
+    });
+    const grant = grantRow('g2_approved');
+    mockGetPendingApproval.mockReturnValue(grant);
+    const d = guard(action, input({ grant }));
+    expect(d.effect).toBe('deny');
+    if (d.effect === 'deny') expect(d.reason).toBe('revoked since');
+  });
+
+  it('a dead grant (row deleted) refuses instead of re-holding', () => {
+    const action = defineGuardedAction({
+      action: 't.g3',
+      grantActionName: 'g3_approved',
+      decide: () => HOLD('b'),
+    });
+    mockGetPendingApproval.mockReturnValue(undefined);
+    const d = guard(action, input({ grant: grantRow('g3_approved') }));
+    expect(d.effect).toBe('deny');
+  });
+
+  it("a grant for a different action doesn't transfer", () => {
+    const action = defineGuardedAction({
+      action: 't.g4',
+      grantActionName: 'g4_approved',
+      decide: () => HOLD('b'),
+    });
+    const grant = grantRow('other_action');
+    mockGetPendingApproval.mockReturnValue(grant);
+    expect(guard(action, input({ grant })).effect).toBe('deny');
+  });
+
+  it('a domain grantCoversRequest binding can refuse a payload mismatch', () => {
+    const action = defineGuardedAction({
+      action: 't.g5',
+      grantActionName: 'g5_approved',
+      grantCoversRequest: () => false,
+      decide: () => HOLD('b'),
+    });
+    const grant = grantRow('g5_approved');
+    mockGetPendingApproval.mockReturnValue(grant);
+    expect(guard(action, input({ grant })).effect).toBe('deny');
+  });
+
+  it('a grant on an already-allowed action is a no-op', () => {
+    const action = defineGuardedAction({
+      action: 't.g6',
+      grantActionName: 'g6_approved',
+      decide: () => ALLOW('ok'),
+    });
+    const grant = grantRow('g6_approved');
+    mockGetPendingApproval.mockReturnValue(grant);
+    expect(guard(action, input({ grant })).effect).toBe('allow');
+  });
+});
+
+describe('fail-closed posture', () => {
+  it('a throwing decide denies', () => {
+    const action = defineGuardedAction({
+      action: 't.f1',
+      decide: () => {
+        throw new Error('boom');
+      },
+    });
+    expect(guard(action, input()).effect).toBe('deny');
+  });
+});
diff --git a/src/guard/guard.ts b/src/guard/guard.ts
new file mode 100644
index 00000000000..327c698b1ea
--- /dev/null
+++ b/src/guard/guard.ts
@@ -0,0 +1,72 @@
+/**
+ * guard() — the one decision function every privileged action consults.
+ *
+ * The decision is the action's decide fn — today's code checks,
+ * defined per action at the module edges. The consult site holds the
+ * GuardedAction value itself (defineGuardedAction), so there is no name
+ * lookup and no fail-open path for an unknown action: an unwired consult is
+ * a compile error, and a value that didn't come from defineGuardedAction is
+ * denied at runtime. Policy-as-data (tighten-only rule sources composing
+ * with the decision) is deliberately deferred — a generalized rules table
+ * can arrive later, with its first operator-visible consumer; until then
+ * the one policy table (agent_message_policies) is consulted inside
+ * a2a.send's decide.
+ *
+ * Grants: an approved replay carries the verified approval row. A valid
+ * grant (live pending row whose action matches the entry's approval action,
+ * plus any domain binding) satisfies a hold — the human already decided —
+ * but NEVER a deny: the checks re-run live, so approve-then-revoke
+ * no longer executes. A grant that is present but invalid fails closed to
+ * deny (no second card).
+ *
+ * The guard itself fails closed: a throwing decide denies.
+ */
+import { getPendingApproval } from '../db/sessions.js';
+import { log } from '../log.js';
+import { isGuardedAction, type GuardedAction } from './guard-actions.js';
+import { ALLOW, DENY, type GuardDecision, type GuardInput } from './types.js';
+
+export function guard(action: GuardedAction, input: GuardInput): GuardDecision {
+  if (!isGuardedAction(action)) {
+    // JS-level backstop — the branded type already forbids this. A
+    // hand-rolled object must not carry a decide fn never vetted at
+    // definition time.
+    log.error('Guard consulted with an undefined action — failing closed', {
+      action: (action as { action?: unknown } | null)?.action,
+    });
+    return DENY('guard consulted with an undefined action (failing closed)');
+  }
+
+  let decision: GuardDecision;
+  try {
+    decision = action.decide(input);
+  } catch (err) {
+    log.error('Guard evaluation threw — failing closed', { action: action.action, err });
+    return DENY('guard failure (failing closed)');
+  }
+
+  if (!input.grant || decision.effect !== 'hold') {
+    // A grant never loosens a deny (the checks re-run live), and a
+    // grant on an already-allowed action is a no-op.
+    return decision;
+  }
+
+  // An invalid grant on a replay is a refusal, not a fresh hold — approved
+  // replays must execute exactly once.
+  if (grantSatisfies(action, input)) {
+    return ALLOW(`hold satisfied by approval ${input.grant.approval_id}`);
+  }
+  return DENY('replay carried an invalid or mismatched grant');
+}
+
+function grantSatisfies(action: GuardedAction, input: GuardInput): boolean {
+  const grant = input.grant;
+  if (!grant || !action.grantActionName) return false;
+  if (grant.action !== action.grantActionName) return false;
+  // The row must still be live — resolution deletes it, so a grant can only
+  // execute once and a fabricated row object doesn't pass.
+  const live = getPendingApproval(grant.approval_id);
+  if (!live || live.action !== action.grantActionName) return false;
+  if (action.grantCoversRequest && !action.grantCoversRequest(grant, input)) return false;
+  return true;
+}
diff --git a/src/guard/index.ts b/src/guard/index.ts
new file mode 100644
index 00000000000..909ce5724ac
--- /dev/null
+++ b/src/guard/index.ts
@@ -0,0 +1,29 @@
+/**
+ * Guard — the privileged-action decision seam.
+ *
+ * One decision function (guard.ts) and a definition-derived action
+ * catalog (guard-actions.ts). Consults carry the GuardedAction value returned by
+ * defineGuardedAction — never a name to look up — so mis-wiring is a build
+ * error, not a runtime fail-open.
+ * Domain-free leaf: domain decisions are defined at the domain modules' edges.
+ */
+export { guard } from './guard.js';
+export {
+  defineGuardedAction,
+  isGuardedAction,
+  listGuardedActions,
+  type GuardedAction,
+  type GuardedActionSpec,
+} from './guard-actions.js';
+export {
+  ALLOW,
+  DENY,
+  GuardDenyError,
+  HOLD,
+  isUnguarded,
+  unguarded,
+  type GuardActor,
+  type GuardDecision,
+  type GuardInput,
+  type Unguarded,
+} from './types.js';
diff --git a/src/guard/types.ts b/src/guard/types.ts
new file mode 100644
index 00000000000..121ed39edea
--- /dev/null
+++ b/src/guard/types.ts
@@ -0,0 +1,88 @@
+/**
+ * Guard vocabulary — the decision seam every privileged action passes.
+ *
+ * The guard is a domain-free leaf: this module may import the DB read layer,
+ * config, log, and shared types — never src/cli/* or src/modules/*. Domain
+ * knowledge (what an action's decide fn checks) arrives via
+ * definition: domain modules call defineGuardedAction (guard-actions.ts) at
+ * their module edges and pass the returned value to every consult and
+ * registration site — the wiring is a symbol reference the compiler checks.
+ */
+import type { PendingApproval } from '../types.js';
+
+/** Who is attempting the action. Mirrors the CLI CallerContext + click identities. */
+export type GuardActor =
+  | { kind: 'host' }
+  | { kind: 'agent'; agentGroupId: string; sessionId?: string }
+  | { kind: 'human'; userId: string }
+  | { kind: 'system' };
+
+export interface GuardInput {
+  actor: GuardActor;
+  /** Domain resource reference, e.g. { from, to } for a2a.send. */
+  resource?: Record<string, string>;
+  /** Action arguments — what the card summarizes and rules may later match on. */
+  payload: Record<string, unknown>;
+  /**
+   * Verified approval row carried by an approved replay. A valid grant
+   * satisfies a hold (the human already decided) but never a deny — the
+   * structural checks re-run live on every replay.
+   */
+  grant?: PendingApproval | null;
+}
+
+const unguardedBrand = Symbol('unguarded');
+/**
+ * A registration that deliberately carries no guard. Where a registry takes
+ * a declaration (delivery actions), omission is not representable —
+ * registration requires either a guard spec or this marker, so the decision
+ * to run unguarded is visible, and justified, in the diff that registers
+ * the handler. The reason travels with the registration;
+ * `grep "unguarded("` is the complete inventory.
+ */
+export type Unguarded = { readonly reason: string; readonly [unguardedBrand]: true };
+
+export function unguarded(reason: string): Unguarded {
+  return Object.freeze({ reason, [unguardedBrand]: true as const });
+}
+
+/**
+ * The one runtime discriminator for guard declarations. The brand symbol is
+ * module-private, so `unguarded()` is the only mint — a look-alike
+ * `{ reason }` object (or a guard spec that someday grows a `reason` field)
+ * doesn't pass.
+ */
+export function isUnguarded(decl: object): decl is Unguarded {
+  return unguardedBrand in decl;
+}
+
+export type GuardDecision =
+  | { effect: 'allow'; reason: string }
+  | { effect: 'hold'; reason: string; approverUserId?: string }
+  | { effect: 'deny'; reason: string };
+
+export const ALLOW = (reason: string): GuardDecision => ({ effect: 'allow', reason });
+export const DENY = (reason: string): GuardDecision => ({ effect: 'deny', reason });
+/**
+ * approverUserId names an exclusive approver for the hold (the a2a policy
+ * row's named approver). Absent, the hold goes to the approvals primitive's
+ * default chain (scoped admins → global admins → owners).
+ */
+export const HOLD = (reason: string, approverUserId?: string): GuardDecision => ({
+  effect: 'hold',
+  reason,
+  approverUserId,
+});
+
+/**
+ * A guard deny travelling as an exception — for flows whose entry point
+ * signals refusal by throwing (the a2a route). Catching it lets a caller
+ * distinguish "the guard refused, as designed" (report to the requester,
+ * log a warning) from a real runtime failure (crash path, stack trace).
+ */
+export class GuardDenyError extends Error {
+  constructor(reason: string) {
+    super(reason);
+    this.name = 'GuardDenyError';
+  }
+}
diff --git a/src/modules/agent-to-agent/agent-route.ts b/src/modules/agent-to-agent/agent-route.ts
index 5ad169a1444..a2b45dad4b2 100644
--- a/src/modules/agent-to-agent/agent-route.ts
+++ b/src/modules/agent-to-agent/agent-route.ts
@@ -27,14 +27,15 @@ import { getAgentGroup } from '../../db/agent-groups.js';
 import { getInboundSourceSessionId, getMostRecentPeerSourceSessionId } from '../../db/session-db.js';
 import { getSession } from '../../db/sessions.js';
 import { wakeContainer } from '../../container-runner.js';
+import { GuardDenyError, guard } from '../../guard/index.js';
 import { log } from '../../log.js';
 import { openInboundDb, resolveSession, sessionDir, writeSessionMessage } from '../../session-manager.js';
-import type { Session } from '../../types.js';
+import type { PendingApproval, Session } from '../../types.js';
 import { requestApproval } from '../approvals/index.js';
-import { hasDestination } from './db/agent-destinations.js';
-import { getMessagePolicy } from './db/agent-message-policies.js';
+import { A2A_MESSAGE_GATE_ACTION, a2aSend } from './guard.js';
 
 export { isSafeAttachmentName };
+export { A2A_MESSAGE_GATE_ACTION } from './guard.js';
 
 export interface ForwardedAttachment {
   name: string;
@@ -230,56 +231,64 @@ function resolveTargetSession(msg: RoutableAgentMessage, sourceSession: Session,
   return resolveSession(targetAgentGroupId, null, null, 'agent-shared').session;
 }
 
-export async function routeAgentMessage(msg: RoutableAgentMessage, session: Session): Promise<void> {
+export async function routeAgentMessage(
+  msg: RoutableAgentMessage,
+  session: Session,
+  opts: { grant?: PendingApproval } = {},
+): Promise<void> {
   const sourceAgentGroupId = session.agent_group_id;
   const targetAgentGroupId = msg.platform_id;
   if (!targetAgentGroupId) {
     throw new Error(`agent-to-agent message ${msg.id} is missing a target agent group id`);
   }
-  const isSelf = targetAgentGroupId === sourceAgentGroupId;
-  if (!isSelf && !hasDestination(sourceAgentGroupId, 'agent', targetAgentGroupId)) {
-    throw new Error(`unauthorized agent-to-agent: ${sourceAgentGroupId} has no destination for ${targetAgentGroupId}`);
-  }
-  if (!getAgentGroup(targetAgentGroupId)) {
-    throw new Error(`target agent group ${targetAgentGroupId} not found for message ${msg.id}`);
+
+  // The a2a.send decision (guard.ts) carries the checks verbatim in their
+  // original order: destination ACL deny, target-exists deny, self-send
+  // allow, agent_message_policies hold. An approved replay carries the
+  // grant — the hold is satisfied but the structure is re-checked live, so
+  // revoking a destination between hold and approve blocks delivery.
+  const decision = guard(a2aSend, {
+    actor: { kind: 'agent', agentGroupId: sourceAgentGroupId, sessionId: session.id },
+    resource: { from: sourceAgentGroupId, to: targetAgentGroupId },
+    payload: { id: msg.id, platform_id: targetAgentGroupId, content: msg.content, in_reply_to: msg.in_reply_to },
+    grant: opts.grant ?? null,
+  });
+
+  if (decision.effect === 'deny') {
+    throw new GuardDenyError(decision.reason);
   }
 
   // Gated edge: hold the message and return (not throw) so the delivery loop
-  // consumes the outbound row; `applyA2aMessageGate` re-routes it on approve.
-  if (!isSelf) {
-    const policy = getMessagePolicy(sourceAgentGroupId, targetAgentGroupId);
-    if (policy) {
-      const { approver } = policy;
-      const sourceName = getAgentGroup(sourceAgentGroupId)?.name ?? sourceAgentGroupId;
-      const targetName = getAgentGroup(targetAgentGroupId)?.name ?? targetAgentGroupId;
-      await requestApproval({
-        session,
-        agentName: sourceName,
-        action: A2A_MESSAGE_GATE_ACTION,
-        approverUserId: approver,
-        title: 'Message approval',
-        question: buildGateQuestion(sourceName, targetName, msg.content),
-        payload: {
-          id: msg.id,
-          platform_id: targetAgentGroupId,
-          content: msg.content,
-          in_reply_to: msg.in_reply_to,
-        },
-      });
-      log.info('Agent message held for approval', {
-        from: sourceAgentGroupId,
-        to: targetAgentGroupId,
-        msgId: msg.id,
-      });
-      return;
-    }
+  // consumes the outbound row; `applyA2aMessageGate` re-enters here with the
+  // grant on approve.
+  if (decision.effect === 'hold') {
+    const sourceName = getAgentGroup(sourceAgentGroupId)?.name ?? sourceAgentGroupId;
+    const targetName = getAgentGroup(targetAgentGroupId)?.name ?? targetAgentGroupId;
+    await requestApproval({
+      session,
+      agentName: sourceName,
+      action: A2A_MESSAGE_GATE_ACTION,
+      approverUserId: decision.approverUserId,
+      title: 'Message approval',
+      question: buildGateQuestion(sourceName, targetName, msg.content),
+      payload: {
+        id: msg.id,
+        platform_id: targetAgentGroupId,
+        content: msg.content,
+        in_reply_to: msg.in_reply_to,
+      },
+    });
+    log.info('Agent message held for approval', {
+      from: sourceAgentGroupId,
+      to: targetAgentGroupId,
+      msgId: msg.id,
+    });
+    return;
   }
 
   await performAgentRoute(msg, session, targetAgentGroupId);
 }
 
-export const A2A_MESSAGE_GATE_ACTION = 'a2a_message_gate';
-
 const GATE_CARD_BODY_MAX = 1500;
 
 function parseMessageContent(contentStr: string): { text: string; files: string[] } {
@@ -308,9 +317,11 @@ function buildGateQuestion(sourceName: string, targetName: string, contentStr: s
 
 /**
  * Cross-session route: pick the target session, forward files, write to its
- * inbound DB, wake it. Authorization is the caller's responsibility.
+ * inbound DB, wake it. Module-private — the only door is routeAgentMessage's
+ * guard decision (the approve continuation re-enters with a grant rather
+ * than calling this directly).
  */
-export async function performAgentRoute(
+async function performAgentRoute(
   msg: RoutableAgentMessage,
   session: Session,
   targetAgentGroupId: string,
diff --git a/src/modules/agent-to-agent/create-agent.test.ts b/src/modules/agent-to-agent/create-agent.test.ts
index f9f6219c784..96ebb920f6c 100644
--- a/src/modules/agent-to-agent/create-agent.test.ts
+++ b/src/modules/agent-to-agent/create-agent.test.ts
@@ -2,26 +2,48 @@
  * Tests for create_agent host-side authorization.
  *
  * Regression guard for the audit finding: `create_agent` is a privileged
- * central-DB write with no host-side authz. The fix authorizes by CLI scope —
- * trusted owner agent groups ('global') create directly; confined groups
- * ('group', the default and the prompt-injection victim) must get admin
- * approval. These tests pin that branch decision.
+ * central-DB write with no host-side authz. Authorization is the guard's
+ * `agents.create` decision — trusted owner agent groups ('global') create
+ * directly; confined groups ('group', the default and the prompt-injection
+ * victim) hold for admin approval. These tests drive the REAL wrapped
+ * delivery action (the only reachable path) and the approve continuation's
+ * grant-carrying re-entry.
  */
 import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
 
-import type { Session } from '../../types.js';
+import type { PendingApproval, Session } from '../../types.js';
 
 // Mocks for the collaborators the branch decides between / depends on.
-const mockRequestApproval = vi.fn().mockResolvedValue(undefined);
-const mockGetContainerConfig = vi.fn();
-const mockCreateAgentGroup = vi.fn();
-const mockInitGroupFilesystem = vi.fn();
-const mockUpdateScalars = vi.fn();
-const mockWriteDestinations = vi.fn();
-const mockNotifyWrite = vi.fn();
+// vi.hoisted: the module barrel import below runs before this file's const
+// initializers, and the mock factories close over this state.
+const {
+  mockRequestApproval,
+  mockGetContainerConfig,
+  mockCreateAgentGroup,
+  mockInitGroupFilesystem,
+  mockUpdateScalars,
+  mockWriteDestinations,
+  mockNotifyWrite,
+  liveApprovals,
+  approvalHandlers,
+} = vi.hoisted(() => ({
+  mockRequestApproval: vi.fn().mockResolvedValue(undefined),
+  mockGetContainerConfig: vi.fn(),
+  mockCreateAgentGroup: vi.fn(),
+  mockInitGroupFilesystem: vi.fn(),
+  mockUpdateScalars: vi.fn(),
+  mockWriteDestinations: vi.fn(),
+  mockNotifyWrite: vi.fn(),
+  liveApprovals: new Map<string, import('../../types.js').PendingApproval>(),
+  approvalHandlers: new Map<string, (ctx: Record<string, unknown>) => Promise<void>>(),
+}));
 
 vi.mock('../approvals/index.js', () => ({
   requestApproval: (...a: unknown[]) => mockRequestApproval(...a),
+  notifyAgent: vi.fn(),
+  registerApprovalHandler: (action: string, handler: (ctx: Record<string, unknown>) => Promise<void>) => {
+    approvalHandlers.set(action, handler);
+  },
 }));
 vi.mock('../../db/container-configs.js', () => ({
   getContainerConfig: (...a: unknown[]) => mockGetContainerConfig(...a),
@@ -42,36 +64,81 @@ vi.mock('./write-destinations.js', () => ({
 vi.mock('./db/agent-destinations.js', () => ({
   getDestinationByName: () => undefined,
   createDestination: vi.fn(),
+  hasDestination: () => true,
   normalizeName: (s: string) => s.toLowerCase().replace(/[^a-z0-9]+/g, '-'),
 }));
 // notifyAgent writes to the session inbound.db + wakes the container; stub both.
+// delivery.ts and agent-route.ts pull more session-manager exports at import time.
 vi.mock('../../session-manager.js', () => ({
   writeSessionMessage: (...a: unknown[]) => mockNotifyWrite(...a),
+  openInboundDb: vi.fn(),
+  openOutboundDb: vi.fn(),
+  clearOutbox: vi.fn(),
+  readOutboxFiles: vi.fn().mockReturnValue([]),
+  resolveSession: vi.fn(),
+  sessionDir: vi.fn().mockReturnValue('/tmp/nowhere'),
+  inboundDbPath: vi.fn().mockReturnValue('/tmp/nowhere/inbound.db'),
 }));
 vi.mock('../../container-runner.js', () => ({
   wakeContainer: vi.fn().mockResolvedValue(undefined),
 }));
 vi.mock('../../db/sessions.js', () => ({
   getSession: (id: string) => ({ id, agent_group_id: 'ag-1' }),
+  getPendingApproval: (id: string) => liveApprovals.get(id),
+  getRunningSessions: () => [],
+  getActiveSessions: () => [],
+  createPendingQuestion: vi.fn(),
 }));
 
-import { handleCreateAgent } from './create-agent.js';
+// The a2a module barrel registers ./guard.js (catalog entries) and the
+// guard-wrapped create_agent delivery action — the path under test.
+import './index.js';
+import { getDeliveryAction } from '../../delivery.js';
 
 const SESSION = { id: 'sess-1', agent_group_id: 'ag-1' } as Session;
 
+async function runCreateAgent(content: Record<string, unknown>): Promise<void> {
+  const wrapped = getDeliveryAction('create_agent');
+  expect(wrapped).toBeDefined();
+  await wrapped!(content, SESSION, undefined as never);
+}
+
+function liveGrant(approvalId: string, payload: Record<string, unknown>): PendingApproval {
+  const row = {
+    approval_id: approvalId,
+    session_id: SESSION.id,
+    request_id: approvalId,
+    action: 'create_agent',
+    payload: JSON.stringify(payload),
+    created_at: new Date().toISOString(),
+    agent_group_id: 'ag-1',
+    channel_type: null,
+    platform_id: null,
+    platform_message_id: null,
+    expires_at: null,
+    status: 'pending',
+    title: '',
+    options_json: '[]',
+    approver_user_id: null,
+  } as PendingApproval;
+  liveApprovals.set(approvalId, row);
+  return row;
+}
+
 beforeEach(() => {
   vi.clearAllMocks();
+  liveApprovals.clear();
 });
 
 afterEach(() => {
   vi.restoreAllMocks();
 });
 
-describe('handleCreateAgent — scope-based authorization', () => {
+describe('create_agent — guard-based authorization (wrapped delivery action)', () => {
   it('global scope: creates directly, no approval requested', async () => {
     mockGetContainerConfig.mockReturnValue({ cli_scope: 'global' });
 
-    await handleCreateAgent({ name: 'Scout', instructions: 'help' }, SESSION);
+    await runCreateAgent({ name: 'Scout', instructions: 'help' });
 
     expect(mockRequestApproval).not.toHaveBeenCalled();
     expect(mockCreateAgentGroup).toHaveBeenCalledTimes(1);
@@ -84,7 +151,7 @@ describe('handleCreateAgent — scope-based authorization', () => {
     // dropping the inheritance leaves the child provider-less (→ claude).
     mockGetContainerConfig.mockReturnValue({ cli_scope: 'global', provider: 'codex' });
 
-    await handleCreateAgent({ name: 'Scout', instructions: 'help' }, SESSION);
+    await runCreateAgent({ name: 'Scout', instructions: 'help' });
 
     expect(mockInitGroupFilesystem).toHaveBeenCalledWith(
       expect.anything(),
@@ -96,7 +163,7 @@ describe('handleCreateAgent — scope-based authorization', () => {
   it('claude creator leaves the child provider unset (built-in default)', async () => {
     mockGetContainerConfig.mockReturnValue({ cli_scope: 'global' }); // no provider
 
-    await handleCreateAgent({ name: 'Scout', instructions: 'help' }, SESSION);
+    await runCreateAgent({ name: 'Scout', instructions: 'help' });
 
     expect(mockUpdateScalars).not.toHaveBeenCalled();
   });
@@ -104,7 +171,7 @@ describe('handleCreateAgent — scope-based authorization', () => {
   it('group scope (default): requires approval, does NOT create directly', async () => {
     mockGetContainerConfig.mockReturnValue({ cli_scope: 'group' });
 
-    await handleCreateAgent({ name: 'Scout', instructions: 'help' }, SESSION);
+    await runCreateAgent({ name: 'Scout', instructions: 'help' });
 
     expect(mockRequestApproval).toHaveBeenCalledTimes(1);
     expect(mockRequestApproval.mock.calls[0][0]).toMatchObject({ action: 'create_agent' });
@@ -115,7 +182,7 @@ describe('handleCreateAgent — scope-based authorization', () => {
   it('missing config: fails closed to approval (no direct create)', async () => {
     mockGetContainerConfig.mockReturnValue(undefined);
 
-    await handleCreateAgent({ name: 'Scout' }, SESSION);
+    await runCreateAgent({ name: 'Scout' });
 
     expect(mockRequestApproval).toHaveBeenCalledTimes(1);
     expect(mockCreateAgentGroup).not.toHaveBeenCalled();
@@ -124,7 +191,7 @@ describe('handleCreateAgent — scope-based authorization', () => {
   it('disabled/other scope: requires approval', async () => {
     mockGetContainerConfig.mockReturnValue({ cli_scope: 'disabled' });
 
-    await handleCreateAgent({ name: 'Scout' }, SESSION);
+    await runCreateAgent({ name: 'Scout' });
 
     expect(mockRequestApproval).toHaveBeenCalledTimes(1);
     expect(mockCreateAgentGroup).not.toHaveBeenCalled();
@@ -133,9 +200,58 @@ describe('handleCreateAgent — scope-based authorization', () => {
   it('empty name: neither creates nor requests approval', async () => {
     mockGetContainerConfig.mockReturnValue({ cli_scope: 'global' });
 
-    await handleCreateAgent({ name: '' }, SESSION);
+    await runCreateAgent({ name: '' });
 
     expect(mockRequestApproval).not.toHaveBeenCalled();
     expect(mockCreateAgentGroup).not.toHaveBeenCalled();
   });
 });
+
+describe('create_agent — approved replay (grant-carrying re-entry)', () => {
+  it('valid grant executes exactly once — decide hold is satisfied, create runs', async () => {
+    mockGetContainerConfig.mockReturnValue({ cli_scope: 'group' });
+    const payload = { name: 'Scout', instructions: 'help' };
+    const approval = liveGrant('appr-ca-1', payload);
+
+    const continuation = approvalHandlers.get('create_agent');
+    expect(continuation).toBeDefined();
+    await continuation!({ session: SESSION, payload, approval, userId: 'telegram:admin', notify: vi.fn() });
+
+    expect(mockCreateAgentGroup).toHaveBeenCalledTimes(1);
+    expect(mockRequestApproval).not.toHaveBeenCalled(); // no second card
+  });
+
+  it('dead grant (row already resolved) refuses the replay', async () => {
+    mockGetContainerConfig.mockReturnValue({ cli_scope: 'group' });
+    const payload = { name: 'Scout', instructions: 'help' };
+    const approval = liveGrant('appr-ca-2', payload);
+    liveApprovals.delete('appr-ca-2'); // resolution consumed the row
+
+    await approvalHandlers.get('create_agent')!({
+      session: SESSION,
+      payload,
+      approval,
+      userId: 'telegram:admin',
+      notify: vi.fn(),
+    });
+
+    expect(mockCreateAgentGroup).not.toHaveBeenCalled();
+    expect(mockRequestApproval).not.toHaveBeenCalled(); // refused, not re-held
+  });
+
+  it('mismatched grant (approved for a different name) refuses the replay', async () => {
+    mockGetContainerConfig.mockReturnValue({ cli_scope: 'group' });
+    const approval = liveGrant('appr-ca-3', { name: 'OtherAgent' });
+
+    await approvalHandlers.get('create_agent')!({
+      session: SESSION,
+      payload: { name: 'Scout' },
+      approval,
+      userId: 'telegram:admin',
+      notify: vi.fn(),
+    });
+
+    expect(mockCreateAgentGroup).not.toHaveBeenCalled();
+    expect(mockRequestApproval).not.toHaveBeenCalled();
+  });
+});
diff --git a/src/modules/agent-to-agent/create-agent.ts b/src/modules/agent-to-agent/create-agent.ts
index b8044da390d..4f8e557c8c5 100644
--- a/src/modules/agent-to-agent/create-agent.ts
+++ b/src/modules/agent-to-agent/create-agent.ts
@@ -1,16 +1,18 @@
 /**
- * `create_agent` delivery-action handler.
+ * `create_agent` delivery-action bodies.
  *
  * SECURITY: `create_agent` writes to the CENTRAL DB (agent_groups,
  * container_configs, agent_destinations) and scaffolds host filesystem state —
  * a privileged operation a confined container is otherwise architecturally
  * barred from. The container's MCP tool gate is inside the (untrusted)
  * container and is trivially bypassed by writing the outbound system row
- * directly, so authorization MUST be enforced host-side. Trusted owner agent
- * groups (CLI scope 'global') create directly; every other (confined) group
- * requires admin approval via `requestApproval` — matching `ncl groups create`
- * (access: 'approval') and the self-mod actions. `applyCreateAgent` runs the
- * creation on approve; `performCreateAgent` is the shared body.
+ * directly, so authorization MUST be enforced host-side: the delivery
+ * registry wraps this action with the guard, whose `agents.create` decision
+ * (./guard.ts) is the old cli_scope branch verbatim — trusted global-scope
+ * groups allow, everything else (including unknown config, fail-closed)
+ * holds for admin approval. On approve the continuation re-enters the
+ * wrapped action with the approval row as its grant and `createAgent` runs.
+ * `performCreateAgent` is the module-private body.
  */
 import path from 'path';
 
@@ -23,7 +25,7 @@ import { initGroupFilesystem } from '../../group-init.js';
 import { log } from '../../log.js';
 import { writeSessionMessage } from '../../session-manager.js';
 import type { AgentGroup, Session } from '../../types.js';
-import { requestApproval, type ApprovalHandler } from '../approvals/index.js';
+import { requestApproval } from '../approvals/index.js';
 import { createDestination, getDestinationByName, normalizeName } from './db/agent-destinations.js';
 import { writeDestinations } from './write-destinations.js';
 
@@ -43,41 +45,27 @@ function notifyAgent(session: Session, text: string): void {
   }
 }
 
-/**
- * Delivery-action entry.
- *
- * Authorization depends on the calling group's CLI scope:
- *   - `global` (set by init-first-agent for trusted owner agent groups):
- *     create immediately. create_agent is the intended primitive for these
- *     privileged agents, and an approval tap on every sub-agent spawn would be
- *     needless friction.
- *   - anything else (the default `group` scope — the realistic
- *     prompt-injection victim): require an admin to approve before any
- *     central-DB write. `applyCreateAgent` runs on approve.
- * Unknown/missing config fails closed to the approval path.
- */
-export async function handleCreateAgent(content: Record<string, unknown>, session: Session): Promise<void> {
+/** Guard precheck: malformed requests are answered without ever creating a hold. */
+export function validateCreateAgent(content: Record<string, unknown>, session: Session): boolean {
   const name = typeof content.name === 'string' ? content.name : '';
-  const instructions = typeof content.instructions === 'string' ? content.instructions : null;
-
   if (!name) {
     notifyAgent(session, 'create_agent failed: name is required.');
-    return;
+    return false;
   }
-
-  const sourceGroup = getAgentGroup(session.agent_group_id);
-  if (!sourceGroup) {
+  if (!getAgentGroup(session.agent_group_id)) {
     notifyAgent(session, 'create_agent failed: source agent group not found.');
     log.warn('create_agent failed: missing source group', { sessionAgentGroup: session.agent_group_id, name });
-    return;
+    return false;
   }
+  return true;
+}
 
-  const cliScope = getContainerConfig(session.agent_group_id)?.cli_scope ?? 'group';
-  if (cliScope === 'global') {
-    // Trusted owner agent group — create directly, then notify (+wake) it.
-    await performCreateAgent(name, instructions, session, sourceGroup, (text) => notifyAgent(session, text));
-    return;
-  }
+/** Guard hold: card the requesting group's admin chain. */
+export async function requestCreateAgentHold(content: Record<string, unknown>, session: Session): Promise<void> {
+  const name = typeof content.name === 'string' ? content.name : '';
+  const instructions = typeof content.instructions === 'string' ? content.instructions : null;
+  const sourceGroup = getAgentGroup(session.agent_group_id);
+  if (!sourceGroup) return;
 
   await requestApproval({
     session,
@@ -89,35 +77,22 @@ export async function handleCreateAgent(content: Record<string, unknown>, sessio
   });
 }
 
-/**
- * Approval handler: performs the creation once an admin approves a request from
- * a confined (non-global) agent group. `session` is the requesting parent.
- */
-export const applyCreateAgent: ApprovalHandler = async ({ session, payload, notify }) => {
-  const name = typeof payload.name === 'string' ? payload.name : '';
-  const instructions = typeof payload.instructions === 'string' ? payload.instructions : null;
-
-  if (!name) {
-    notify('create_agent approved but the request had no name.');
-    return;
-  }
-
+/** Guard allow body: performs the creation (fresh global-scope call or approved replay). */
+export async function createAgent(content: Record<string, unknown>, session: Session): Promise<void> {
+  const name = typeof content.name === 'string' ? content.name : '';
+  const instructions = typeof content.instructions === 'string' ? content.instructions : null;
   const sourceGroup = getAgentGroup(session.agent_group_id);
-  if (!sourceGroup) {
-    notify('create_agent approved but the source agent group no longer exists.');
-    log.warn('create_agent apply failed: missing source group', { sessionAgentGroup: session.agent_group_id, name });
-    return;
-  }
+  if (!name || !sourceGroup) return; // precheck already answered the requester
 
-  await performCreateAgent(name, instructions, session, sourceGroup, notify);
-};
+  await performCreateAgent(name, instructions, session, sourceGroup, (text) => notifyAgent(session, text));
+}
 
 /**
  * Core creation: writes the new agent group + bidirectional destinations and
  * scaffolds its filesystem, then reports via `notify`. Authorization is the
- * CALLER's responsibility (the global-scope shortcut in handleCreateAgent or
- * admin approval via applyCreateAgent) — never call this from an unauthorized
- * path, as it performs privileged central-DB writes a confined container is
+ * CALLER's responsibility (the guard's agents.create decision) — never call
+ * this from an unauthorized path, as it performs privileged central-DB
+ * writes a confined container is
  * otherwise barred from.
  */
 async function performCreateAgent(
diff --git a/src/modules/agent-to-agent/guard.ts b/src/modules/agent-to-agent/guard.ts
new file mode 100644
index 00000000000..c778aa24f26
--- /dev/null
+++ b/src/modules/agent-to-agent/guard.ts
@@ -0,0 +1,88 @@
+/**
+ * Agent-to-agent guard adapter — the module's catalog entries, composed at
+ * the module edge (imported by ./index.ts).
+ *
+ * agents.create — the cli_scope branch moved verbatim out of
+ * create-agent.ts: `global` scope creates directly (create_agent is the
+ * intended primitive for trusted owner agent groups); anything else — the
+ * default `group` scope, and unknown/missing config, fail-closed — holds for
+ * the requesting group's admin chain.
+ *
+ * a2a.send — the decision moved verbatim out of routeAgentMessage, in its
+ * original check order: a missing destination row denies; a missing target
+ * group denies; self-sends allow without a destination row; an
+ * agent_message_policies row for the (from, to) pair holds for the row's
+ * named approver. The ghost-policy edge (policy row with no destination row)
+ * denies — the destination check precedes the policy check, exactly today's
+ * outcome. Policy rows can only tighten (hold), never allow: absence of a
+ * row falls through to the structural checks.
+ */
+import { getAgentGroup } from '../../db/agent-groups.js';
+import { getContainerConfig } from '../../db/container-configs.js';
+import { ALLOW, DENY, HOLD, defineGuardedAction } from '../../guard/index.js';
+import { hasDestination } from './db/agent-destinations.js';
+import { getMessagePolicy } from './db/agent-message-policies.js';
+
+/**
+ * pending_approvals action string for held a2a messages. Lives here (not in
+ * agent-route.ts) so agent-route can import this adapter — loading the
+ * consult site guarantees its catalog entry is registered — without a cycle.
+ */
+export const A2A_MESSAGE_GATE_ACTION = 'a2a_message_gate';
+
+export const agentsCreate = defineGuardedAction({
+  action: 'agents.create',
+  grantActionName: 'create_agent',
+  // Bind a create_agent grant to the name that was approved.
+  grantCoversRequest: (grant, input) => {
+    try {
+      return (JSON.parse(grant.payload) as { name?: string }).name === input.payload.name;
+    } catch {
+      return false;
+    }
+  },
+  decide: (input) => {
+    if (input.actor.kind !== 'agent') return DENY('create_agent is a container-originated action.');
+    const cliScope = getContainerConfig(input.actor.agentGroupId)?.cli_scope ?? 'group';
+    if (cliScope === 'global') {
+      // Trusted owner agent group — an approval tap on every sub-agent spawn
+      // would be needless friction.
+      return ALLOW('trusted global-scope agent group');
+    }
+    // The realistic prompt-injection victim (default `group` scope) — and any
+    // unknown config value, fail-closed — requires an admin before any
+    // central-DB write.
+    return HOLD('agent-initiated create_agent requires admin approval');
+  },
+});
+
+export const a2aSend = defineGuardedAction({
+  action: 'a2a.send',
+  grantActionName: A2A_MESSAGE_GATE_ACTION,
+  // Bind an a2a grant to the exact held message target.
+  grantCoversRequest: (grant, input) => {
+    try {
+      return (JSON.parse(grant.payload) as { platform_id?: string }).platform_id === input.resource?.to;
+    } catch {
+      return false;
+    }
+  },
+  decide: (input) => {
+    if (input.actor.kind !== 'agent') return DENY('agent-to-agent send requires an agent actor');
+    const from = input.actor.agentGroupId;
+    const to = input.resource?.to ?? '';
+    const isSelf = to === from;
+    if (!isSelf && !hasDestination(from, 'agent', to)) {
+      return DENY(`unauthorized agent-to-agent: ${from} has no destination for ${to}`);
+    }
+    if (!getAgentGroup(to)) {
+      return DENY(`target agent group ${to} not found for message ${String(input.payload.id)}`);
+    }
+    if (isSelf) return ALLOW('self-send');
+    const policy = getMessagePolicy(from, to);
+    if (policy) {
+      return HOLD(`a2a message policy ${from}→${to} holds for ${policy.approver}`, policy.approver);
+    }
+    return ALLOW('destination grant exists');
+  },
+});
diff --git a/src/modules/agent-to-agent/index.ts b/src/modules/agent-to-agent/index.ts
index 95dfd36c679..0a79f77dc7f 100644
--- a/src/modules/agent-to-agent/index.ts
+++ b/src/modules/agent-to-agent/index.ts
@@ -1,13 +1,14 @@
 /**
  * Agent-to-agent module — inter-agent messaging and on-demand agent creation.
  *
- * Registers one delivery action (`create_agent`) plus its matching approval
- * handler — `create_agent` writes central-DB state, so confined (non-global)
- * groups require admin approval (the delivery action queues the request;
- * `applyCreateAgent` runs on approve); trusted global-scope groups create
- * directly. The sibling `channel_type === 'agent'` routing path is NOT a system
- * action — core `delivery.ts` dispatches into `./agent-route.js` via a dynamic
- * import when it sees `msg.channel_type === 'agent'`.
+ * Registers its guard-catalog entries (./guard.js) and one guard-wrapped
+ * delivery action (`create_agent`) — `create_agent` writes central-DB state,
+ * so the guard's agents.create decision holds confined (non-global) groups
+ * for admin approval while trusted global-scope groups create directly; the
+ * approval handler re-enters the wrapped action carrying the approval row as
+ * its grant. The sibling `channel_type === 'agent'` routing path is NOT a
+ * system action — core `delivery.ts` dispatches into `./agent-route.js` via
+ * a dynamic import when it sees `msg.channel_type === 'agent'`.
  *
  * Host integration points:
  *   - `src/container-runner.ts::spawnContainer` dynamically imports
@@ -20,13 +21,19 @@
  * system action logs "Unknown system action", `channel_type='agent'` messages
  * throw because the module isn't installed.
  */
-import { registerDeliveryAction } from '../../delivery.js';
-import { registerApprovalHandler } from '../approvals/index.js';
+import { reenterGuardedDeliveryAction, registerDeliveryAction } from '../../delivery.js';
+import { notifyAgent, registerApprovalHandler } from '../approvals/index.js';
 import { A2A_MESSAGE_GATE_ACTION } from './agent-route.js';
-import { applyCreateAgent, handleCreateAgent } from './create-agent.js';
+import { createAgent, requestCreateAgentHold, validateCreateAgent } from './create-agent.js';
+import { agentsCreate } from './guard.js';
 import { applyA2aMessageGate } from './message-gate.js';
 
-registerDeliveryAction('create_agent', handleCreateAgent);
-registerApprovalHandler('create_agent', applyCreateAgent);
+registerDeliveryAction('create_agent', createAgent, {
+  guardAction: agentsCreate,
+  precheck: validateCreateAgent,
+  requestHold: requestCreateAgentHold,
+  onDeny: (_content, session, reason) => notifyAgent(session, `create_agent denied: ${reason}`),
+});
+registerApprovalHandler('create_agent', reenterGuardedDeliveryAction('create_agent'));
 
 registerApprovalHandler(A2A_MESSAGE_GATE_ACTION, applyA2aMessageGate);
diff --git a/src/modules/agent-to-agent/message-gate.test.ts b/src/modules/agent-to-agent/message-gate.test.ts
index fa1a485f686..98bd380c3f9 100644
--- a/src/modules/agent-to-agent/message-gate.test.ts
+++ b/src/modules/agent-to-agent/message-gate.test.ts
@@ -2,16 +2,17 @@ import Database from 'better-sqlite3';
 import fs from 'fs';
 import { describe, expect, it, beforeEach, afterEach, vi } from 'vitest';
 
+import './guard.js'; // register the a2a.send catalog entry (incl. the policy hold)
 import { routeAgentMessage } from './agent-route.js';
 import { createDestination, deleteDestination, deleteAllDestinationsTouching } from './db/agent-destinations.js';
 import { getMessagePolicy, removeMessagePolicy, setMessagePolicy } from './db/agent-message-policies.js';
 import { applyA2aMessageGate } from './message-gate.js';
 import { initTestDb, closeDb, runMigrations, createAgentGroup } from '../../db/index.js';
 import { getDb } from '../../db/connection.js';
-import { createSession } from '../../db/sessions.js';
+import { createPendingApproval, createSession, deletePendingApproval, getPendingApproval } from '../../db/sessions.js';
 import { requestApproval } from '../approvals/index.js';
 import { initSessionFolder, inboundDbPath } from '../../session-manager.js';
-import type { Session } from '../../types.js';
+import type { PendingApproval, Session } from '../../types.js';
 
 vi.mock('../../container-runner.js', () => ({
   wakeContainer: vi.fn().mockResolvedValue(undefined),
@@ -67,6 +68,23 @@ function makeSession(id: string, agentGroupId: string): Session {
   };
 }
 
+/** Seed a live a2a hold row (what requestApproval writes) and return it as the grant. */
+function seedA2aHold(approvalId: string, payload: Record<string, unknown>): PendingApproval {
+  createPendingApproval({
+    approval_id: approvalId,
+    session_id: 'sess-A',
+    request_id: approvalId,
+    action: 'a2a_message_gate',
+    payload: JSON.stringify(payload),
+    created_at: now(),
+    agent_group_id: A,
+    title: 'Message approval',
+    options_json: '[]',
+    approver_user_id: 'telegram:dana',
+  });
+  return getPendingApproval(approvalId)!;
+}
+
 describe('agent message policies', () => {
   let SA: Session;
   let SB: Session;
@@ -129,7 +147,7 @@ describe('agent message policies', () => {
     expect(requestApproval).not.toHaveBeenCalled();
   });
 
-  it('policy present → holds the message and requests approval from the policy approver scoped to the target', async () => {
+  it('policy present → holds the message and requests approval from the policy approver', async () => {
     setMessagePolicy(A, B, 'telegram:dana', now());
 
     await routeAgentMessage(
@@ -139,7 +157,7 @@ describe('agent message policies', () => {
 
     // Held: nothing routed to B.
     expect(readInbound(B, SB.id)).toHaveLength(0);
-    // One approval requested, to the policy's approver, scoped to the target group.
+    // One approval requested, to the policy's approver.
     expect(requestApproval).toHaveBeenCalledTimes(1);
     const opts = vi.mocked(requestApproval).mock.calls[0][0];
     expect(opts.action).toBe('a2a_message_gate');
@@ -158,21 +176,76 @@ describe('agent message policies', () => {
     expect(readInbound(A, SA.id)).toHaveLength(1);
   });
 
-  // ── approve handler re-routes the held message ──
+  it('ghost policy (policy row, no destination row) still denies — deny beats the policy hold', async () => {
+    deleteDestination(A, 'b'); // removes A→B — the destination ACL now denies
+    setMessagePolicy(A, B, 'telegram:dana', now()); // ...but a stale policy row remains
+
+    await expect(
+      routeAgentMessage({ id: 'ghost', platform_id: B, content: JSON.stringify({ text: 'x' }), in_reply_to: null }, SA),
+    ).rejects.toThrow(/unauthorized agent-to-agent/);
+    expect(requestApproval).not.toHaveBeenCalled();
+    expect(readInbound(B, SB.id)).toHaveLength(0);
+  });
+
+  // ── approve handler re-enters the guarded route with the grant ──
+
+  it('applyA2aMessageGate delivers the held message to the target (valid grant)', async () => {
+    setMessagePolicy(A, B, 'telegram:dana', now());
+    const payload = { id: 'held-1', platform_id: B, content: JSON.stringify({ text: 'approved!' }), in_reply_to: null };
+    const approval = seedA2aHold('appr-a2a-1', payload);
 
-  it('applyA2aMessageGate delivers the held message to the target', async () => {
     const notify = vi.fn();
-    await applyA2aMessageGate({
-      session: SA,
-      userId: 'slack:dana',
-      notify,
-      payload: { id: 'held-1', platform_id: B, content: JSON.stringify({ text: 'approved!' }), in_reply_to: null },
-    });
+    await applyA2aMessageGate({ session: SA, userId: 'telegram:dana', notify, payload, approval });
 
     const bRows = readInbound(B, SB.id);
     expect(bRows).toHaveLength(1);
     expect(JSON.parse(bRows[0].content).text).toBe('approved!');
     expect(notify).not.toHaveBeenCalled();
+    // The hold is satisfied by the grant — no second card.
+    expect(requestApproval).not.toHaveBeenCalled();
+  });
+
+  it('destination revoked between hold and approve → refused cleanly, requester told, nothing delivered', async () => {
+    setMessagePolicy(A, B, 'telegram:dana', now());
+    const payload = { id: 'held-2', platform_id: B, content: JSON.stringify({ text: 'stale' }), in_reply_to: null };
+    const approval = seedA2aHold('appr-a2a-2', payload);
+
+    deleteDestination(A, 'b'); // revoke A→B while the card is pending
+
+    const notify = vi.fn();
+    // An expected policy refusal — resolves (no throw), so the response
+    // handler never records it as a handler crash.
+    await applyA2aMessageGate({ session: SA, userId: 'telegram:dana', notify, payload, approval });
+
+    expect(readInbound(B, SB.id)).toHaveLength(0);
+    expect(notify).toHaveBeenCalledWith(expect.stringMatching(/not delivered.*no destination for/));
+  });
+
+  it('mismatched grant (held for another target) refuses the replay cleanly', async () => {
+    setMessagePolicy(A, B, 'telegram:dana', now());
+    // Grant was approved for a message to A (different target than the replay).
+    const approval = seedA2aHold('appr-a2a-3', { id: 'other', platform_id: A, content: '{}', in_reply_to: null });
+    const payload = { id: 'held-3', platform_id: B, content: JSON.stringify({ text: 'swap' }), in_reply_to: null };
+
+    const notify = vi.fn();
+    await applyA2aMessageGate({ session: SA, userId: 'telegram:dana', notify, payload, approval });
+
+    expect(readInbound(B, SB.id)).toHaveLength(0);
+    expect(notify).toHaveBeenCalledWith(expect.stringMatching(/not delivered.*invalid or mismatched grant/));
+  });
+
+  it('a grant only works while its row is live (executes once)', async () => {
+    setMessagePolicy(A, B, 'telegram:dana', now());
+    const payload = { id: 'held-4', platform_id: B, content: JSON.stringify({ text: 'once' }), in_reply_to: null };
+    const approval = seedA2aHold('appr-a2a-4', payload);
+
+    deletePendingApproval(approval.approval_id); // resolution already consumed the row
+
+    const notify = vi.fn();
+    await applyA2aMessageGate({ session: SA, userId: 'telegram:dana', notify, payload, approval });
+
+    expect(readInbound(B, SB.id)).toHaveLength(0);
+    expect(notify).toHaveBeenCalledWith(expect.stringMatching(/not delivered.*invalid or mismatched grant/));
   });
 
   // ── ghost-gate cleanup ──
diff --git a/src/modules/agent-to-agent/message-gate.ts b/src/modules/agent-to-agent/message-gate.ts
index 74c20c4cb22..93f7d799174 100644
--- a/src/modules/agent-to-agent/message-gate.ts
+++ b/src/modules/agent-to-agent/message-gate.ts
@@ -1,9 +1,10 @@
 /** Approve handler for a held a2a message. (Reject is handled by the generic response-handler path.) */
+import { GuardDenyError } from '../../guard/index.js';
 import { log } from '../../log.js';
 import type { ApprovalHandler } from '../approvals/index.js';
-import { performAgentRoute, type RoutableAgentMessage } from './agent-route.js';
+import { routeAgentMessage, type RoutableAgentMessage } from './agent-route.js';
 
-export const applyA2aMessageGate: ApprovalHandler = async ({ session, payload, notify }) => {
+export const applyA2aMessageGate: ApprovalHandler = async ({ session, payload, approval, notify }) => {
   const { id, platform_id, content, in_reply_to } = payload;
   if (typeof platform_id !== 'string' || !platform_id) {
     notify('Message approved but the target agent group was missing from the request.');
@@ -18,7 +19,27 @@ export const applyA2aMessageGate: ApprovalHandler = async ({ session, payload, n
     in_reply_to: typeof in_reply_to === 'string' ? in_reply_to : null,
   };
 
-  await performAgentRoute(msg, session, platform_id);
+  // One replay semantics: re-enter the guarded route carrying the approval
+  // row as the grant. The policy hold is satisfied, but the structural
+  // checks run live — a deny here (destination revoked while the card was
+  // pending, dead or mismatched grant) is an EXPECTED policy outcome, not a
+  // crash: tell the requester, log a warning, and let anything else keep the
+  // response handler's failure path.
+  try {
+    await routeAgentMessage(msg, session, { grant: approval });
+  } catch (err) {
+    if (err instanceof GuardDenyError) {
+      log.warn('Approved a2a replay refused by the guard', {
+        from: session.agent_group_id,
+        to: platform_id,
+        msgId: msg.id,
+        reason: err.message,
+      });
+      notify(`Message approved, but not delivered — no longer authorized: ${err.message}`);
+      return;
+    }
+    throw err;
+  }
   log.info('Held agent message delivered after approval', {
     from: session.agent_group_id,
     to: platform_id,
diff --git a/src/modules/approvals/primitive.ts b/src/modules/approvals/primitive.ts
index 2d3028a64dc..48047e676dc 100644
--- a/src/modules/approvals/primitive.ts
+++ b/src/modules/approvals/primitive.ts
@@ -59,6 +59,12 @@ const APPROVAL_OPTIONS: RawOption[] = [
 export interface ApprovalHandlerContext {
   session: Session;
   payload: Record<string, unknown>;
+  /**
+   * The verified approval row — the grant an approved continuation carries
+   * when it re-enters its guarded entry point. Still live here; resolution
+   * deletes it after the handler returns, so a grant executes exactly once.
+   */
+  approval: PendingApproval;
   /** User ID of the admin who approved. Empty string if unknown. */
   userId: string;
   /** Send a system chat message to the requesting agent's session. */
diff --git a/src/modules/approvals/response-handler.ts b/src/modules/approvals/response-handler.ts
index 52fc705efea..79014597ece 100644
--- a/src/modules/approvals/response-handler.ts
+++ b/src/modules/approvals/response-handler.ts
@@ -112,7 +112,7 @@ async function handleRegisteredApproval(
 
   const payload = JSON.parse(approval.payload);
   try {
-    await handler({ session, payload, userId, notify });
+    await handler({ session, payload, approval, userId, notify });
     log.info('Approval handled', { approvalId: approval.approval_id, action: approval.action, userId });
   } catch (err) {
     log.error('Approval handler threw', { approvalId: approval.approval_id, action: approval.action, err });
diff --git a/src/modules/permissions/channel-approval.test.ts b/src/modules/permissions/channel-approval.test.ts
index 103f06d5687..db084787eb3 100644
--- a/src/modules/permissions/channel-approval.test.ts
+++ b/src/modules/permissions/channel-approval.test.ts
@@ -68,7 +68,11 @@ vi.mock('./user-dm.js', () => ({
 
 vi.mock('../../config.js', async () => {
   const actual = await vi.importActual('../../config.js');
-  return { ...actual, DATA_DIR: '/tmp/nanoclaw-test-channel-approval' };
+  return {
+    ...actual,
+    DATA_DIR: '/tmp/nanoclaw-test-channel-approval',
+    GROUPS_DIR: '/tmp/nanoclaw-test-channel-approval/groups',
+  };
 });
 
 const TEST_DIR = '/tmp/nanoclaw-test-channel-approval';
@@ -586,6 +590,108 @@ describe('unknown-channel registration flow', () => {
       .c;
     expect(stillPending).toBe(1);
   });
+
+  it('create new agent: the free-text name reply creates the group and wires the channel', async () => {
+    const { routeInbound } = await import('../../router.js');
+    const { getResponseHandlers } = await import('../../response-registry.js');
+    const { getDb } = await import('../../db/connection.js');
+
+    await routeInbound(groupMention('chat-create-new'));
+    await new Promise((r) => setTimeout(r, 10));
+    const pending = getDb().prepare('SELECT messaging_group_id FROM pending_channel_approvals').get() as {
+      messaging_group_id: string;
+    };
+
+    // Owner clicks "Connect new agent" → name prompt lands in their DM.
+    for (const handler of getResponseHandlers()) {
+      const claimed = await handler({
+        questionId: pending.messaging_group_id,
+        value: 'new_agent',
+        userId: 'owner',
+        channelType: 'telegram',
+        platformId: 'dm-owner',
+        threadId: null,
+      });
+      if (claimed) break;
+    }
+
+    // Owner replies with the agent name in the same DM — the interceptor
+    // captures it and creates.
+    await routeInbound({
+      channelType: 'telegram',
+      platformId: 'dm-owner',
+      threadId: null,
+      message: {
+        id: 'name-reply-1',
+        kind: 'chat' as const,
+        content: JSON.stringify({ senderId: 'owner', senderName: 'Owner', text: 'Newbie' }),
+        timestamp: now(),
+      },
+    });
+
+    const created = getDb().prepare("SELECT id FROM agent_groups WHERE name = 'Newbie'").get() as
+      | { id: string }
+      | undefined;
+    expect(created).toBeDefined();
+    const mgaCount = (
+      getDb()
+        .prepare('SELECT COUNT(*) AS c FROM messaging_group_agents WHERE messaging_group_id = ? AND agent_group_id = ?')
+        .get(pending.messaging_group_id, created!.id) as { c: number }
+    ).c;
+    expect(mgaCount).toBe(1);
+    const stillPending = (getDb().prepare('SELECT COUNT(*) AS c FROM pending_channel_approvals').get() as { c: number })
+      .c;
+    expect(stillPending).toBe(0);
+  });
+
+  it('a name reply after the registration vanished is consumed without creating anything', async () => {
+    const { routeInbound } = await import('../../router.js');
+    const { getResponseHandlers } = await import('../../response-registry.js');
+    const { getDb } = await import('../../db/connection.js');
+
+    await routeInbound(groupMention('chat-vanished'));
+    await new Promise((r) => setTimeout(r, 10));
+    const pending = getDb().prepare('SELECT messaging_group_id FROM pending_channel_approvals').get() as {
+      messaging_group_id: string;
+    };
+
+    for (const handler of getResponseHandlers()) {
+      const claimed = await handler({
+        questionId: pending.messaging_group_id,
+        value: 'new_agent',
+        userId: 'owner',
+        channelType: 'telegram',
+        platformId: 'dm-owner',
+        threadId: null,
+      });
+      if (claimed) break;
+    }
+
+    // The registration disappears between the click and the reply (rejected
+    // from another card, group delete cascade, …) — the interceptor no
+    // longer finds a pending registration, so the reply must not create.
+    getDb()
+      .prepare('DELETE FROM pending_channel_approvals WHERE messaging_group_id = ?')
+      .run(pending.messaging_group_id);
+
+    const agentGroupsBefore = (getDb().prepare('SELECT COUNT(*) AS c FROM agent_groups').get() as { c: number }).c;
+    await routeInbound({
+      channelType: 'telegram',
+      platformId: 'dm-owner',
+      threadId: null,
+      message: {
+        id: 'name-reply-2',
+        kind: 'chat' as const,
+        content: JSON.stringify({ senderId: 'owner', senderName: 'Owner', text: 'Ghost' }),
+        timestamp: now(),
+      },
+    });
+
+    const agentGroupsAfter = (getDb().prepare('SELECT COUNT(*) AS c FROM agent_groups').get() as { c: number }).c;
+    expect(agentGroupsAfter).toBe(agentGroupsBefore);
+    const mgaCount = (getDb().prepare('SELECT COUNT(*) AS c FROM messaging_group_agents').get() as { c: number }).c;
+    expect(mgaCount).toBe(0);
+  });
 });
 
 describe('no-owner / no-agent failure modes', () => {
diff --git a/src/modules/permissions/guard.ts b/src/modules/permissions/guard.ts
new file mode 100644
index 00000000000..810a9cf3f45
--- /dev/null
+++ b/src/modules/permissions/guard.ts
@@ -0,0 +1,55 @@
+/**
+ * Permissions guard adapter — the module's catalog entries, composed at the
+ * module edge (imported by ./index.ts).
+ *
+ * senders.admit — the `unknown_sender_policy` switch moved verbatim out of
+ * handleUnknownSender: `public` allows (short-circuited before the gate
+ * anyway), `request_approval` holds, `strict` denies. The hold is executed by
+ * the caller through the module's own pending_sender_approvals flow (card,
+ * in-flight dedup) — not the approvals primitive — so this entry has no
+ * grantActionName: the approve continuation adds the member and replays
+ * routeInbound, which then passes the gate structurally via membership, no
+ * grant needed.
+ *
+ * channels.register — click authorization for the channel-registration flow,
+ * verbatim from today's response handler: the delivered approver, or an
+ * admin of the pending row's anchor agent group. Consulted inline by the
+ * card-click response handler only. The free-text name reply is deliberately
+ * NOT re-authorized (main's behavior): the click arms the capture and stands
+ * as the auth — a privilege revoked between the click and the reply still
+ * completes the flow.
+ */
+import { ALLOW, DENY, HOLD, defineGuardedAction } from '../../guard/index.js';
+import { getPendingChannelApproval } from './db/pending-channel-approvals.js';
+import { hasAdminPrivilege } from './db/user-roles.js';
+
+export const sendersAdmit = defineGuardedAction({
+  action: 'senders.admit',
+  decide: (input) => {
+    const policy = input.payload.policy;
+    if (policy === 'public') return ALLOW('public messaging group');
+    if (policy === 'request_approval') {
+      return HOLD(
+        `unknown sender requires admin approval on messaging group ${String(input.payload.messagingGroupId)}`,
+      );
+    }
+    return DENY('unknown sender on a strict messaging group');
+  },
+});
+
+export const channelsRegister = defineGuardedAction({
+  action: 'channels.register',
+  decide: (input) => {
+    if (input.actor.kind !== 'human') return DENY('channel registration resolves via human clicks/replies');
+    const questionId = typeof input.payload.questionId === 'string' ? input.payload.questionId : '';
+    const row = getPendingChannelApproval(questionId);
+    if (!row) return DENY(`no pending channel registration for ${questionId || '(missing questionId)'}`);
+    if (
+      input.actor.userId &&
+      (input.actor.userId === row.approver_user_id || hasAdminPrivilege(input.actor.userId, row.agent_group_id))
+    ) {
+      return ALLOW('delivered approver or anchor-group admin');
+    }
+    return DENY('not an eligible channel-registration approver');
+  },
+});
diff --git a/src/modules/permissions/index.ts b/src/modules/permissions/index.ts
index 3cd5159b97d..f8d3dbf5b1c 100644
--- a/src/modules/permissions/index.ts
+++ b/src/modules/permissions/index.ts
@@ -33,6 +33,8 @@ import { registerResponseHandler, type ResponsePayload } from '../../response-re
 import { getDeliveryAdapter } from '../../delivery.js';
 import { log } from '../../log.js';
 import type { MessagingGroup, MessagingGroupAgent } from '../../types.js';
+import { guard } from '../../guard/index.js';
+import { channelsRegister, sendersAdmit } from './guard.js';
 import { canAccessAgentGroup } from './access.js';
 import {
   buildAgentSelectionOptions,
@@ -131,43 +133,49 @@ function handleUnknownSender(
     agent_group_id: agentGroupId,
   };
 
-  if (mg.unknown_sender_policy === 'strict') {
-    log.info('MESSAGE DROPPED — unknown sender (strict policy)', {
+  // The admission decision is the guard's senders.admit decision (./guard.ts)
+  // — unknown_sender_policy verbatim: strict → deny, request_approval → hold,
+  // public → allow (short-circuited before the gate). Drop-recording and the
+  // hold creation stay here.
+  const decision = guard(sendersAdmit, {
+    actor: userId ? { kind: 'human', userId } : { kind: 'system' },
+    payload: {
       messagingGroupId: mg.id,
       agentGroupId,
-      userId,
-      accessReason,
-    });
-    recordDroppedMessage(dropRecord);
-    return;
-  }
+      senderIdentity: userId,
+      policy: mg.unknown_sender_policy,
+    },
+  });
+
+  if (decision.effect === 'allow') return; // 'public' — handled before the gate; fall through silently.
 
-  if (mg.unknown_sender_policy === 'request_approval') {
-    log.info('MESSAGE DROPPED — unknown sender (approval requested)', {
+  log.info(
+    decision.effect === 'hold'
+      ? 'MESSAGE DROPPED — unknown sender (approval requested)'
+      : 'MESSAGE DROPPED — unknown sender (strict policy)',
+    {
       messagingGroupId: mg.id,
       agentGroupId,
       userId,
       accessReason,
-    });
-    recordDroppedMessage(dropRecord);
-    // Fire-and-forget; pick-approver + delivery + row-insert are all async.
-    // If it fails it logs internally — the user's message still stays dropped
-    // either way. Requires a resolved userId (senderResolver populates users
-    // row before the gate fires); if we got here without one, there's nothing
-    // to identify for approval and we just stay in the "silent strict" branch.
-    if (userId) {
-      requestSenderApproval({
-        messagingGroupId: mg.id,
-        agentGroupId,
-        senderIdentity: userId,
-        senderName,
-        event,
-      }).catch((err) => log.error('Sender-approval flow threw', { err }));
-    }
-    return;
+    },
+  );
+  recordDroppedMessage(dropRecord);
+
+  // Fire-and-forget; pick-approver + delivery + row-insert are all async.
+  // If it fails it logs internally — the user's message still stays dropped
+  // either way. Requires a resolved userId (senderResolver populates users
+  // row before the gate fires); if we got here without one, there's nothing
+  // to identify for approval and we just drop silently.
+  if (decision.effect === 'hold' && userId) {
+    requestSenderApproval({
+      messagingGroupId: mg.id,
+      agentGroupId,
+      senderIdentity: userId,
+      senderName,
+      event,
+    }).catch((err) => log.error('Sender-approval flow threw', { err }));
   }
-
-  // 'public' should have been handled before the gate; fall through silently.
 }
 
 setSenderResolver(extractAndUpsertUser);
@@ -411,18 +419,23 @@ async function handleChannelApprovalResponse(payload: ResponsePayload): Promise<
   const row = getPendingChannelApproval(payload.questionId);
   if (!row) return false;
 
+  // Click authorization is the guard's channels.register decision (./guard.ts):
+  // the delivered approver, or an admin of the pending row's anchor agent group.
   const clickerId = payload.userId
     ? payload.userId.includes(':')
       ? payload.userId
       : `${payload.channelType}:${payload.userId}`
     : null;
-  const isAuthorized =
-    clickerId !== null && (clickerId === row.approver_user_id || hasAdminPrivilege(clickerId, row.agent_group_id));
-  if (!isAuthorized) {
+  const decision = guard(channelsRegister, {
+    actor: { kind: 'human', userId: clickerId ?? '' },
+    payload: { questionId: payload.questionId },
+  });
+  if (!clickerId || decision.effect !== 'allow') {
     log.warn('Channel registration click rejected — unauthorized clicker', {
       messagingGroupId: row.messaging_group_id,
       clickerId,
       expectedApprover: row.approver_user_id,
+      reason: decision.reason,
     });
     return true;
   }
diff --git a/src/modules/self-mod/apply.ts b/src/modules/self-mod/apply.ts
index c5318ffbfb6..6f5fa7a813c 100644
--- a/src/modules/self-mod/apply.ts
+++ b/src/modules/self-mod/apply.ts
@@ -1,11 +1,12 @@
 /**
- * Approval handlers for self-modification actions.
+ * Guarded handler bodies for self-modification actions.
  *
- * The approvals module calls these when an admin clicks Approve on a
- * pending_approvals row whose action matches. Each handler mutates the
- * container config in the DB, rebuilds/kills the container as needed,
- * and writes an on_wake message so the fresh container picks up where
- * the old one left off.
+ * The delivery registry's guard wrapper runs these only on `allow` — which,
+ * for self-mod, means an approved replay carrying a valid grant (the
+ * decision holds unconditionally from the container path; see ./guard.ts).
+ * Each body mutates the container config in the DB, rebuilds/kills the
+ * container as needed, and writes an on_wake message so the fresh container
+ * picks up where the old one left off.
  *
  * install_packages: update DB + rebuild image + kill container + on_wake.
  * add_mcp_server: update DB + kill container + on_wake.
@@ -17,18 +18,19 @@ import { getSession } from '../../db/sessions.js';
 import type { McpServerConfig } from '../../container-config.js';
 import { log } from '../../log.js';
 import { writeSessionMessage } from '../../session-manager.js';
-import type { ApprovalHandler } from '../approvals/index.js';
+import type { Session } from '../../types.js';
+import { notifyAgent } from '../approvals/index.js';
 
-export const applyInstallPackages: ApprovalHandler = async ({ session, payload, userId, notify }) => {
+export async function applyInstallPackages(payload: Record<string, unknown>, session: Session): Promise<void> {
   const agentGroup = getAgentGroup(session.agent_group_id);
   if (!agentGroup) {
-    notify('install_packages approved but agent group missing.');
+    notifyAgent(session, 'install_packages approved but agent group missing.');
     return;
   }
 
   const configRow = getContainerConfig(agentGroup.id);
   if (!configRow) {
-    notify('install_packages approved but container config missing.');
+    notifyAgent(session, 'install_packages approved but container config missing.');
     return;
   }
 
@@ -52,7 +54,7 @@ export const applyInstallPackages: ApprovalHandler = async ({ session, payload,
     ...((payload.apt as string[] | undefined) || []),
     ...((payload.npm as string[] | undefined) || []),
   ].join(', ');
-  log.info('Package install approved', { agentGroupId: session.agent_group_id, userId });
+  log.info('Package install approved', { agentGroupId: session.agent_group_id });
   try {
     await buildAgentGroupImage(session.agent_group_id);
     writeSessionMessage(session.agent_group_id, session.id, {
@@ -75,23 +77,24 @@ export const applyInstallPackages: ApprovalHandler = async ({ session, payload,
     });
     log.info('Container rebuild completed (bundled with install)', { agentGroupId: session.agent_group_id });
   } catch (e) {
-    notify(
+    notifyAgent(
+      session,
       `Packages added to config (${pkgs}) but rebuild failed: ${e instanceof Error ? e.message : String(e)}. Tell the user — an admin will need to retry the install_packages request or inspect the build logs.`,
     );
     log.error('Bundled rebuild failed after install approval', { agentGroupId: session.agent_group_id, err: e });
   }
-};
+}
 
-export const applyAddMcpServer: ApprovalHandler = async ({ session, payload, userId, notify }) => {
+export async function applyAddMcpServer(payload: Record<string, unknown>, session: Session): Promise<void> {
   const agentGroup = getAgentGroup(session.agent_group_id);
   if (!agentGroup) {
-    notify('add_mcp_server approved but agent group missing.');
+    notifyAgent(session, 'add_mcp_server approved but agent group missing.');
     return;
   }
 
   const configRow = getContainerConfig(agentGroup.id);
   if (!configRow) {
-    notify('add_mcp_server approved but container config missing.');
+    notifyAgent(session, 'add_mcp_server approved but container config missing.');
     return;
   }
 
@@ -122,5 +125,5 @@ export const applyAddMcpServer: ApprovalHandler = async ({ session, payload, use
     const s = getSession(session.id);
     if (s) wakeContainer(s);
   });
-  log.info('MCP server add approved', { agentGroupId: session.agent_group_id, userId });
-};
+  log.info('MCP server add approved', { agentGroupId: session.agent_group_id });
+}
diff --git a/src/modules/self-mod/guard.ts b/src/modules/self-mod/guard.ts
new file mode 100644
index 00000000000..bbdb3a11e07
--- /dev/null
+++ b/src/modules/self-mod/guard.ts
@@ -0,0 +1,32 @@
+/**
+ * Self-mod guard adapter — the module's catalog entries, composed at the
+ * module edge (imported by ./index.ts).
+ *
+ * The decision is today's behavior verbatim: from the container
+ * path, self-modification is held unconditionally for the agent group's
+ * admin chain. (The equivalent host-side mutations — `ncl groups config
+ * add-package` etc. — are separate catalog actions derived from the command
+ * registry.)
+ */
+import { DENY, HOLD, defineGuardedAction, type GuardInput } from '../../guard/index.js';
+
+function selfModDecide(label: string) {
+  return (input: GuardInput) => {
+    if (input.actor.kind !== 'agent') {
+      return DENY(`${label} is a container-originated action.`);
+    }
+    return HOLD(`${label} always requires admin approval from the container path`);
+  };
+}
+
+export const selfModInstallPackages = defineGuardedAction({
+  action: 'self_mod.install_packages',
+  grantActionName: 'install_packages',
+  decide: selfModDecide('install_packages'),
+});
+
+export const selfModAddMcpServer = defineGuardedAction({
+  action: 'self_mod.add_mcp_server',
+  grantActionName: 'add_mcp_server',
+  decide: selfModDecide('add_mcp_server'),
+});
diff --git a/src/modules/self-mod/index.ts b/src/modules/self-mod/index.ts
index e1f49e212b3..bfdc08797bc 100644
--- a/src/modules/self-mod/index.ts
+++ b/src/modules/self-mod/index.ts
@@ -2,29 +2,51 @@
  * Self-modification module — admin-approved container mutations.
  *
  * Optional tier. Depends on the approvals default module for the request/
- * handler plumbing. On install the module registers:
- *   - Two delivery actions (install_packages, add_mcp_server) that validate
- *     input and queue an approval via requestApproval().
- *   - Two matching approval handlers that run on approve and perform the
- *     complete follow-up:
- *       install_packages → update container.json, rebuild image, kill
+ * handler plumbing and on the guard for the decision. On install the module
+ * registers:
+ *   - Its guard-catalog entries (./guard.ts): unconditional hold from the
+ *     container path.
+ *   - Two guard-wrapped delivery actions (install_packages, add_mcp_server):
+ *     validation runs as the wrapper's precheck, the hold builders card the
+ *     admin, and the handler bodies (./apply.ts) run only on allow — i.e. on
+ *     an approved replay:
+ *       install_packages → update container_configs, rebuild image, kill
  *         container (next wake respawns on the new image), schedule a
  *         verify-and-report follow-up prompt.
- *       add_mcp_server → update container.json, kill container. No image
+ *       add_mcp_server → update container_configs, kill container. No image
  *         rebuild — bun runs TS directly, so the new MCP server is wired
  *         by the next container start.
+ *   - Two approval handlers that re-enter the wrapped actions with the
+ *     approval row as the grant (one replay semantics — the guard re-checks
+ *     the structural checks live).
  *
  * Without this module: the MCP tools in the container still write outbound
  * system messages with these actions, but delivery logs "Unknown system
  * action" and drops them. Admin never sees a card; nothing changes.
  */
-import { registerDeliveryAction } from '../../delivery.js';
-import { registerApprovalHandler } from '../approvals/index.js';
+import { reenterGuardedDeliveryAction, registerDeliveryAction } from '../../delivery.js';
+import { notifyAgent, registerApprovalHandler } from '../approvals/index.js';
 import { applyAddMcpServer, applyInstallPackages } from './apply.js';
-import { handleAddMcpServer, handleInstallPackages } from './request.js';
+import { selfModAddMcpServer, selfModInstallPackages } from './guard.js';
+import {
+  requestAddMcpServerHold,
+  requestInstallPackagesHold,
+  validateAddMcpServer,
+  validateInstallPackages,
+} from './request.js';
 
-registerDeliveryAction('install_packages', handleInstallPackages);
-registerDeliveryAction('add_mcp_server', handleAddMcpServer);
+registerDeliveryAction('install_packages', applyInstallPackages, {
+  guardAction: selfModInstallPackages,
+  precheck: validateInstallPackages,
+  requestHold: requestInstallPackagesHold,
+  onDeny: (_content, session, reason) => notifyAgent(session, `install_packages denied: ${reason}`),
+});
+registerDeliveryAction('add_mcp_server', applyAddMcpServer, {
+  guardAction: selfModAddMcpServer,
+  precheck: validateAddMcpServer,
+  requestHold: requestAddMcpServerHold,
+  onDeny: (_content, session, reason) => notifyAgent(session, `add_mcp_server denied: ${reason}`),
+});
 
-registerApprovalHandler('install_packages', applyInstallPackages);
-registerApprovalHandler('add_mcp_server', applyAddMcpServer);
+registerApprovalHandler('install_packages', reenterGuardedDeliveryAction('install_packages'));
+registerApprovalHandler('add_mcp_server', reenterGuardedDeliveryAction('add_mcp_server'));
diff --git a/src/modules/self-mod/request.ts b/src/modules/self-mod/request.ts
index 6cd7f05c9c3..21e558c18ff 100644
--- a/src/modules/self-mod/request.ts
+++ b/src/modules/self-mod/request.ts
@@ -1,12 +1,12 @@
 /**
- * Delivery-action handlers for agent-initiated self-modification requests.
+ * Validation + hold-request builders for agent-initiated self-modification.
  *
  * Two actions the container can write into messages_out (via the self-mod
- * MCP tools): install_packages, add_mcp_server. Each one validates input
- * and queues an approval request. The admin's approval triggers the
- * matching approval handler in ./apply.ts, which also performs the
- * required follow-up (rebuild+restart for install_packages, restart-only
- * for add_mcp_server).
+ * MCP tools): install_packages, add_mcp_server. The delivery registry wraps
+ * each one with the guard (see ./guard.ts — unconditional hold from the
+ * container path): validation here runs as the wrapper's precheck, and the
+ * hold builders create the approval card when the guard holds. On approve,
+ * the continuation re-enters the wrapped action and ./apply.ts runs.
  *
  * Host-side sanitization for install_packages is defense-in-depth — the MCP
  * tool validates first. Both layers matter: the DB row carries the payload
@@ -17,40 +17,48 @@ import { log } from '../../log.js';
 import type { Session } from '../../types.js';
 import { notifyAgent, requestApproval } from '../approvals/index.js';
 
-export async function handleInstallPackages(content: Record<string, unknown>, session: Session): Promise<void> {
+export function validateInstallPackages(content: Record<string, unknown>, session: Session): boolean {
   const agentGroup = getAgentGroup(session.agent_group_id);
   if (!agentGroup) {
     notifyAgent(session, 'install_packages failed: agent group not found.');
-    return;
+    return false;
   }
 
   const apt = (content.apt as string[]) || [];
   const npm = (content.npm as string[]) || [];
-  const reason = (content.reason as string) || '';
 
   const APT_RE = /^[a-z0-9][a-z0-9._+-]*$/;
   const NPM_RE = /^(@[a-z0-9][a-z0-9._-]*\/)?[a-z0-9][a-z0-9._-]*$/;
   const MAX_PACKAGES = 20;
   if (apt.length + npm.length === 0) {
     notifyAgent(session, 'install_packages failed: at least one apt or npm package is required.');
-    return;
+    return false;
   }
   if (apt.length + npm.length > MAX_PACKAGES) {
     notifyAgent(session, `install_packages failed: max ${MAX_PACKAGES} packages per request.`);
-    return;
+    return false;
   }
   const invalidApt = apt.find((p) => !APT_RE.test(p));
   if (invalidApt) {
     notifyAgent(session, `install_packages failed: invalid apt package name "${invalidApt}".`);
     log.warn('install_packages: invalid apt package rejected', { pkg: invalidApt });
-    return;
+    return false;
   }
   const invalidNpm = npm.find((p) => !NPM_RE.test(p));
   if (invalidNpm) {
     notifyAgent(session, `install_packages failed: invalid npm package name "${invalidNpm}".`);
     log.warn('install_packages: invalid npm package rejected', { pkg: invalidNpm });
-    return;
+    return false;
   }
+  return true;
+}
+
+export async function requestInstallPackagesHold(content: Record<string, unknown>, session: Session): Promise<void> {
+  const agentGroup = getAgentGroup(session.agent_group_id);
+  if (!agentGroup) return;
+  const apt = (content.apt as string[]) || [];
+  const npm = (content.npm as string[]) || [];
+  const reason = (content.reason as string) || '';
 
   const packageList = [...apt.map((p) => `apt: ${p}`), ...npm.map((p) => `npm: ${p}`)].join(', ');
   await requestApproval({
@@ -63,18 +71,26 @@ export async function handleInstallPackages(content: Record<string, unknown>, se
   });
 }
 
-export async function handleAddMcpServer(content: Record<string, unknown>, session: Session): Promise<void> {
+export function validateAddMcpServer(content: Record<string, unknown>, session: Session): boolean {
   const agentGroup = getAgentGroup(session.agent_group_id);
   if (!agentGroup) {
     notifyAgent(session, 'add_mcp_server failed: agent group not found.');
-    return;
+    return false;
   }
   const serverName = content.name as string;
   const command = content.command as string;
   if (!serverName || !command) {
     notifyAgent(session, 'add_mcp_server failed: name and command are required.');
-    return;
+    return false;
   }
+  return true;
+}
+
+export async function requestAddMcpServerHold(content: Record<string, unknown>, session: Session): Promise<void> {
+  const agentGroup = getAgentGroup(session.agent_group_id);
+  if (!agentGroup) return;
+  const serverName = content.name as string;
+  const command = content.command as string;
   await requestApproval({
     session,
     agentName: agentGroup.name,