fix(cli): authorize gateway model probe overrides

CHANGELOG.md

@@ -46,6 +46,7 @@ Docs: https://docs.openclaw.ai
 - Chat commands: include configured model-catalog reasoning metadata when building `/think` argument menus so Ollama Cloud and other provider-owned reasoning models show supported levels instead of only `off`. Fixes #73515; supersedes #73568. Thanks @danielzinhu99 and @neeravmakwana.
 - Channels/Telegram: suppress generic tool-progress chatter when preview streaming is off, so non-streaming Telegram turns only deliver final replies while approvals, media, and errors still route normally. Refs #72363 and #72482. Thanks @neeravmakwana and @SweetSophia.
 - CLI/model probes: add repeatable image `--file` inputs to `infer model run` for local and gateway multimodal model smokes, so vision models such as Ollama Qwen VL and Gemini can be tested through the raw model-probe surface. Fixes #63700. Thanks @cedricjanssens.
+- CLI/model probes: request trusted operator scope for `infer model run --gateway --model <provider/model>` so Gateway raw model smokes can use one-off provider/model overrides instead of being rejected before provider auth resolution. Fixes #73759. Thanks @chrislro.
 - CLI/image describe: pass `--prompt` and `--timeout-ms` through `infer image describe` and `describe-many`, so custom vision instructions and slow local model budgets reach media-understanding providers such as Ollama, OpenAI, Google, and OpenRouter. Refs #63700. Thanks @cedricjanssens.
 - WhatsApp/Web: pass explicit Baileys socket timings into every WhatsApp Web socket and expose `web.whatsapp.*` keepalive, connect, and query timeout settings so unstable networks can avoid repeated 408 disconnect and opening-handshake timeout loops. Fixes #56365. (#73580) Thanks @velvet-shark.
 - Channels/Telegram: persist native command metadata on target sessions so topic, helper, and ACP-bound slash commands keep their session metadata attached to the routed conversation. (#57548) Thanks @GaosCode.

docs/cli/infer.md

@@ -135,6 +135,7 @@ This table maps common inference tasks to the corresponding infer command.
 - `model run --file` accepts image files, detects their MIME type, and sends them with the supplied prompt to the selected model. Repeat `--file` for multiple images.
 - `model run --file` rejects non-image inputs. Use `infer audio transcribe` for audio files and `infer video describe` for video files.
 - `model run --gateway` exercises Gateway routing, saved auth, provider selection, and the embedded runtime, but still runs as a raw model probe: it sends the supplied prompt and any image attachments without prior session transcript, bootstrap/AGENTS context, context-engine assembly, tools, or bundled MCP servers.
+- `model run --gateway --model <provider/model>` requires a trusted operator gateway credential because the request asks the Gateway to run a one-off provider/model override.
 
 ## Model
 

src/cli/capability-cli.test.ts

@@ -589,6 +589,38 @@ describe("capability cli", () => {
     );
   });
 
+  it("requests admin scope for gateway model probes with provider/model overrides", async () => {
+    await runRegisteredCli({
+      register: registerCapabilityCli as (program: Command) => void,
+      argv: [
+        "capability",
+        "model",
+        "run",
+        "--prompt",
+        "hello",
+        "--gateway",
+        "--model",
+        "anthropic/claude-haiku-4-5",
+        "--json",
+      ],
+    });
+
+    expect(mocks.callGateway).toHaveBeenCalledWith(
+      expect.objectContaining({
+        clientName: "gateway-client",
+        method: "agent",
+        mode: "backend",
+        scopes: ["operator.admin"],
+        params: expect.objectContaining({
+          provider: "anthropic",
+          model: "claude-haiku-4-5",
+          modelRun: true,
+          promptMode: "none",
+        }),
+      }),
+    );
+  });
+
   it("rejects empty model run prompts before gateway dispatch", async () => {
     await expect(
       runRegisteredCli({

src/cli/capability-cli.ts

@@ -22,6 +22,7 @@ import type { OpenClawConfig } from "../config/types.openclaw.js";
 import { callGateway, randomIdempotencyKey } from "../gateway/call.js";
 import { buildGatewayConnectionDetailsWithResolvers } from "../gateway/connection-details.js";
 import { isLoopbackHost } from "../gateway/net.js";
+import { ADMIN_SCOPE } from "../gateway/operator-scopes.js";
 import { GATEWAY_CLIENT_MODES, GATEWAY_CLIENT_NAMES } from "../gateway/protocol/client-info.js";
 import { generateImage, listRuntimeImageGenerationProviders } from "../image-generation/runtime.js";
 import type {
@@ -702,6 +703,9 @@ async function runModelRun(params: {
   }
 
   const { provider, model } = resolveModelRefOverride(params.model);
+  // Provider/model overrides require trusted-operator scope. Use the backend
+  // shared-secret lane so local gateway smokes do not depend on paired CLI device scopes.
+  const hasModelOverride = Boolean(provider || model);
   const response: {
     result?: {
       payloads?: Array<{ text?: string; mediaUrl?: string | null; mediaUrls?: string[] }>;
@@ -730,8 +734,9 @@ async function runModelRun(params: {
     },
     expectFinal: true,
     timeoutMs: 120_000,
-    clientName: GATEWAY_CLIENT_NAMES.CLI,
-    mode: GATEWAY_CLIENT_MODES.CLI,
+    clientName: hasModelOverride ? GATEWAY_CLIENT_NAMES.GATEWAY_CLIENT : GATEWAY_CLIENT_NAMES.CLI,
+    mode: hasModelOverride ? GATEWAY_CLIENT_MODES.BACKEND : GATEWAY_CLIENT_MODES.CLI,
+    ...(hasModelOverride ? { scopes: [ADMIN_SCOPE] } : {}),
   });
   return {
     ok: true,