diff --git a/docs/tools/subagents.md b/docs/tools/subagents.md
index 2d4095cb7c14..5b3114fcd285 100644
--- a/docs/tools/subagents.md
+++ b/docs/tools/subagents.md
@@ -573,14 +573,16 @@ Sub-agents use the same profile and tool-policy pipeline as the parent or
 target agent first. After that, OpenClaw applies the sub-agent restriction
 layer.
 
-Sub-agents always lose `gateway`, `agents_list`, `session_status`, and
-`cron` regardless of depth or role (system-level/interactive tools, or
-tools the main agent should coordinate). Leaf sub-agents (default depth-1
-behavior, and always at depth 2) additionally lose `subagents`,
-`sessions_list`, `sessions_history`, and `sessions_spawn`. Sub-agents never
-get the `message` tool — it is disabled at spawn time, not filtered by
-this deny list — and `sessions_send` stays denied so sub-agents
-communicate only through the announce chain.
+Sub-agents always lose `gateway`, `agents_list`, `session_status`, `cron`,
+`message`, `sessions_send`, and the `conversations_*` tools regardless of
+depth or role (system-level/interactive tools, direct delivery surfaces, or
+tools the main agent should coordinate). This hard-deny layer is derived from
+the persisted sub-agent session envelope on every turn, including resumed and
+visible dashboard sessions; ordinary `allow`/`alsoAllow` entries cannot override
+it. Hidden launches also disable `message` before tool construction as defense in
+depth. Leaf sub-agents (default depth-1 behavior, and always at depth 2)
+additionally lose `subagents`, `sessions_list`, `sessions_history`, and
+`sessions_spawn`, so sub-agent communication stays on the announce chain.
 
 `sessions_history` remains a bounded, sanitized recall view here too — it
 is not a raw transcript dump.
diff --git a/src/agents/agent-tools.create-openclaw-coding-tools.test.ts b/src/agents/agent-tools.create-openclaw-coding-tools.test.ts
index bec2bc9cb121..b0c6df57804e 100644
--- a/src/agents/agent-tools.create-openclaw-coding-tools.test.ts
+++ b/src/agents/agent-tools.create-openclaw-coding-tools.test.ts
@@ -1497,6 +1497,41 @@ describe("createOpenClawCodingTools", () => {
     expectListIncludes(latestCreateOpenClawToolsOptions().pluginToolDenylist, ["pdf"]);
   });
 
+  it("removes message from persisted visible child sessions on every turn", async () => {
+    const storeDir = await fs.mkdtemp(path.join(os.tmpdir(), "openclaw-visible-subagent-message-"));
+    const storeTemplate = path.join(storeDir, "{agentId}", "sessions.json");
+    const agentId = "visible-subagent-message";
+    const childSessionKey = `agent:${agentId}:dashboard:child`;
+    const rootSessionKey = `agent:${agentId}:dashboard:root`;
+    try {
+      await writeSessionStore(storeTemplate, agentId, {
+        [childSessionKey]: {
+          sessionId: "visible-child",
+          updatedAt: Date.now(),
+          spawnDepth: 1,
+          spawnedBy: `agent:${agentId}:main`,
+          subagentRole: "leaf",
+          subagentControlScope: "none",
+        },
+        [rootSessionKey]: {
+          sessionId: "root-dashboard",
+          updatedAt: Date.now(),
+          spawnDepth: 0,
+        },
+      });
+
+      const firstChildTurn = createToolsForStoredSession(storeTemplate, childSessionKey);
+      const resumedChildTurn = createToolsForStoredSession(storeTemplate, childSessionKey);
+      const rootTurn = createToolsForStoredSession(storeTemplate, rootSessionKey);
+
+      expect(toolNameList(firstChildTurn)).not.toContain("message");
+      expect(toolNameList(resumedChildTurn)).not.toContain("message");
+      expect(toolNameList(rootTurn)).toContain("message");
+    } finally {
+      await fs.rm(storeDir, { recursive: true, force: true });
+    }
+  });
+
   it("passes inherited allowlist entries to OpenClaw plugin discovery", async () => {
     const createOpenClawToolsMock = vi.mocked(createOpenClawTools);
     createOpenClawToolsMock.mockClear();
diff --git a/src/agents/agent-tools.policy.test.ts b/src/agents/agent-tools.policy.test.ts
index 1a4be9157614..176d02dc881f 100644
--- a/src/agents/agent-tools.policy.test.ts
+++ b/src/agents/agent-tools.policy.test.ts
@@ -313,6 +313,64 @@ describe("resolveSubagentToolPolicyForSession", () => {
     expect(isToolAllowedByPolicyName("memory_get", policy)).toBe(true);
   });
 
+  it.each(["allow", "alsoAllow"] as const)(
+    "does not let configured %s entries re-enable hard-denied tools",
+    async (allowField) => {
+      const storePath = createSessionStorePath(`openclaw-subagent-hard-deny-${allowField}`);
+      const sessionKeys = {
+        leaf: "agent:main:subagent:hard-deny-leaf",
+        orchestrator: "agent:main:subagent:hard-deny-orchestrator",
+      } as const;
+      await writeSessionEntries(storePath, {
+        [sessionKeys.leaf]: {
+          sessionId: "hard-deny-leaf",
+          updatedAt: Date.now(),
+          spawnDepth: 2,
+          subagentRole: "leaf",
+          subagentControlScope: "none",
+        },
+        [sessionKeys.orchestrator]: {
+          sessionId: "hard-deny-orchestrator",
+          updatedAt: Date.now(),
+          spawnDepth: 1,
+          subagentRole: "orchestrator",
+          subagentControlScope: "children",
+        },
+      });
+      const hardDeniedTools = [
+        "gateway",
+        "agents_list",
+        "session_status",
+        "automations",
+        "cron",
+        "message",
+        "sessions_send",
+        "conversations_list",
+        "conversations_send",
+        "conversations_turn",
+      ];
+      const cfg = {
+        ...baseCfg,
+        session: { store: storePath },
+        tools: {
+          subagents: {
+            tools: {
+              [allowField]: [...hardDeniedTools, "memory_search"],
+            },
+          },
+        },
+      } as unknown as OpenClawConfig;
+
+      for (const sessionKey of Object.values(sessionKeys)) {
+        const policy = resolveSubagentToolPolicyForSession(cfg, sessionKey);
+        for (const toolName of hardDeniedTools) {
+          expect(isToolAllowedByPolicyName(toolName, policy), toolName).toBe(false);
+        }
+        expect(isToolAllowedByPolicyName("memory_search", policy)).toBe(true);
+      }
+    },
+  );
+
   it("resolves inherited tool denies from stored subagent sessions", async () => {
     const storePath = createSessionStorePath("openclaw-subagent-inherited-deny");
     await writeSessionEntries(storePath, {
diff --git a/src/agents/agent-tools.policy.ts b/src/agents/agent-tools.policy.ts
index d29f9f5ecbbb..169a34e5b409 100644
--- a/src/agents/agent-tools.policy.ts
+++ b/src/agents/agent-tools.policy.ts
@@ -36,11 +36,7 @@ import {
   type SubagentSessionRole,
 } from "./subagent-capabilities.js";
 import { isToolAllowedByPolicyName } from "./tool-policy-match.js";
-import {
-  mergeAlsoAllowPolicy,
-  normalizeToolName,
-  resolveToolProfilePolicy,
-} from "./tool-policy.js";
+import { mergeAlsoAllowPolicy, resolveToolProfilePolicy } from "./tool-policy.js";
 import { AUTOMATIONS_TOOL_NAME } from "./tools/automations-tool-name.js";
 
 export { resolveProviderToolPolicy };
@@ -56,7 +52,8 @@ const SUBAGENT_TOOL_DENY_ALWAYS = [
   // Status/scheduling - main agent coordinates
   "session_status",
   AUTOMATIONS_TOOL_NAME,
-  // Direct session sends - subagents communicate through announce chain
+  // Direct user/session sends - subagents communicate through announce chain
+  "message",
   "sessions_send",
   "conversations_list",
   "conversations_send",
@@ -105,13 +102,8 @@ export function resolveSubagentToolPolicyForSession(
   });
   const allow = Array.isArray(configured?.allow) ? configured.allow : undefined;
   const alsoAllow = Array.isArray(configured?.alsoAllow) ? configured.alsoAllow : undefined;
-  const explicitAllow = new Set(
-    [...(allow ?? []), ...(alsoAllow ?? [])].map((toolName) => normalizeToolName(toolName)),
-  );
   const deny = [
-    ...resolveSubagentDenyListForRole(capabilities.role).filter(
-      (toolName) => !explicitAllow.has(normalizeToolName(toolName)),
-    ),
+    ...resolveSubagentDenyListForRole(capabilities.role),
     ...(Array.isArray(configured?.deny) ? configured.deny : []),
   ];
   const mergedAllow = mergeConfiguredSubagentAllow(allow, alsoAllow);
diff --git a/src/agents/sandbox/workspace-authority.test.ts b/src/agents/sandbox/workspace-authority.test.ts
index 0b40aeb99b37..da1fd6c7b782 100644
--- a/src/agents/sandbox/workspace-authority.test.ts
+++ b/src/agents/sandbox/workspace-authority.test.ts
@@ -102,6 +102,9 @@ describe("resolveSandboxWorkspaceAuthority", () => {
     });
     expect(elevated.confinementError).toContain("elevated execution");
 
+    // Config-driven delegation cannot happen anymore: the subagent hard-deny
+    // list is non-overridable, so alsoAllow cannot re-enable sessions_spawn.
+    // The worker stays confined (no error) instead of being rejected.
     const delegatingConfig = configWithSandbox({ mode: "all", workspaceAccess: "rw" });
     delegatingConfig.tools!.sandbox!.tools!.allow = [...SAFE_WORKBOARD_TOOLS, "sessions_spawn"];
     delegatingConfig.tools!.subagents = { tools: { alsoAllow: ["sessions_spawn"] } };
@@ -110,7 +113,8 @@ describe("resolveSandboxWorkspaceAuthority", () => {
       agentId: "main",
       sessionKey: "agent:main:subagent:workboard-card",
     });
-    expect(delegating.confinementError).toContain("sessions_spawn");
+    expect(delegating.confinementError).toBeUndefined();
+    expect(delegating.sandboxed).toBe(true);
   });
 
   it("uses the runtime session visibility clamp", () => {