diff --git a/docs/start/onboarding-overview.md b/docs/start/onboarding-overview.md
index f398ae4738bf..19d8336efd9a 100644
--- a/docs/start/onboarding-overview.md
+++ b/docs/start/onboarding-overview.md
@@ -139,6 +139,11 @@ setup verifies a real model reply before saving the provider and activating its
 model. A failed or cancelled check preserves the previous configuration. The
 classic wizard also retains its custom-provider setup.
 
+If the endpoint refuses the connection or its hostname cannot be found, setup
+reports the failed connection immediately instead of waiting through normal
+chat retries. Start the server or correct the URL and network settings on the
+Gateway host, then retry. Ordinary agent sessions keep their connection retries.
+
 ## Related
 
 - [Getting started](/start/getting-started)
diff --git a/src/agents/embedded-agent-runner/run/assistant-failure.ts b/src/agents/embedded-agent-runner/run/assistant-failure.ts
index fafa6858a3fc..bd756aff9e78 100644
--- a/src/agents/embedded-agent-runner/run/assistant-failure.ts
+++ b/src/agents/embedded-agent-runner/run/assistant-failure.ts
@@ -489,6 +489,7 @@ export async function handleEmbeddedAssistantFailure(input: {
         profileId: input.authProfileId,
         authMode,
         status,
+        code: failedAssistant?.errorCode,
         rawError: failedAssistant?.errorMessage?.trim(),
         // Retry reason "timeout" also includes 5xx; only the terminal owner records a deadline.
         timeout:
diff --git a/src/agents/embedded-agent-runner/run/attempt-recovery.test-support.ts b/src/agents/embedded-agent-runner/run/attempt-recovery.test-support.ts
index 7aef8ca51571..95bc5f192122 100644
--- a/src/agents/embedded-agent-runner/run/attempt-recovery.test-support.ts
+++ b/src/agents/embedded-agent-runner/run/attempt-recovery.test-support.ts
@@ -36,6 +36,7 @@ export type TransportDropScenario = {
   lastToolError?: Parameters<typeof makeEmbeddedRunnerAttempt>[0]["lastToolError"];
   pluginHarnessOwnsTransport?: boolean;
   retryAvailable?: boolean;
+  retryConnectionErrors?: boolean;
   replaySafe?: boolean;
   fallbackConfigured?: boolean;
   providerRetryMaxDelayMs?: number;
@@ -136,9 +137,11 @@ export async function recoverAfterTransportDrop(scenario: TransportDropScenario
   const continueFromCurrentTranscript = vi.fn();
   const contextRecoveryState = createEmbeddedRunContextRecoveryState();
   const failoverRetryController = createEmbeddedRunFailoverRetryController({
-    runParams: { runId: "run:transport-drop", config: scenario.config } as Parameters<
-      typeof createEmbeddedRunFailoverRetryController
-    >[0]["runParams"],
+    runParams: {
+      runId: "run:transport-drop",
+      config: scenario.config,
+      retryConnectionErrors: scenario.retryConnectionErrors,
+    } as Parameters<typeof createEmbeddedRunFailoverRetryController>[0]["runParams"],
     provider,
     modelId,
     globalLane: "test",
diff --git a/src/agents/embedded-agent-runner/run/attempt-recovery.test.ts b/src/agents/embedded-agent-runner/run/attempt-recovery.test.ts
index efe4f9df5d70..5aab5579184f 100644
--- a/src/agents/embedded-agent-runner/run/attempt-recovery.test.ts
+++ b/src/agents/embedded-agent-runner/run/attempt-recovery.test.ts
@@ -35,6 +35,30 @@ const tempDirs = createTempDirTracker();
 const requireRecord = createRequireRecord("record", "expected-label-object-capitalized");
 afterEach(() => tempDirs.cleanup());
 
+it.each(["ECONNREFUSED", "ENOTFOUND", "EHOSTUNREACH", "ENETUNREACH"])(
+  "fails fast on %s for setup while normal runs retain connection retries",
+  async (errorCode) => {
+    const scenario = {
+      errorMessage: "Connection error.",
+      errorCode,
+      noTools: true,
+      replaySafe: true,
+      content: [],
+    } satisfies TransportDropScenario;
+    vi.mocked(sleepWithAbort).mockClear();
+    const setup = await recoverAfterTransportDrop({ ...scenario, retryConnectionErrors: false });
+    expect(setup.recovery.action).toBe("proceed");
+    expect(sleepWithAbort).not.toHaveBeenCalled();
+    await expect(handleAssistantFailureAfterRecovery(setup)).rejects.toMatchObject({
+      reason: "timeout",
+      code: errorCode,
+    });
+    const ordinary = await recoverAfterTransportDrop(scenario);
+    expect(ordinary.recovery.action).toBe("retry");
+    expect(sleepWithAbort).toHaveBeenCalledOnce();
+  },
+);
+
 function handleAssistantFailureAfterRecovery(
   fixture: Awaited<ReturnType<typeof recoverAfterTransportDrop>>,
   previousRetryFailoverReason: Parameters<
diff --git a/src/agents/embedded-agent-runner/run/attempt-recovery.ts b/src/agents/embedded-agent-runner/run/attempt-recovery.ts
index 6315fd372386..1374d24da810 100644
--- a/src/agents/embedded-agent-runner/run/attempt-recovery.ts
+++ b/src/agents/embedded-agent-runner/run/attempt-recovery.ts
@@ -13,6 +13,7 @@ import {
 } from "../../failover-error.js";
 import { failoverReasonFromClassification } from "../../failover/classification-rules.js";
 import { classifyFailoverSignal } from "../../failover/classify.js";
+import { getFailoverErrorCode } from "../../failover/error.js";
 import { resolveRetryAfterMs } from "../../failover/retry-evidence.js";
 import { LiveSessionModelSwitchError } from "../../live-model-switch-error.js";
 import { shouldSwitchToLiveModel, clearLiveModelSwitchPending } from "../../live-model-switch.js";
@@ -394,6 +395,7 @@ export async function recoverEmbeddedRunAttempt(input: {
     recoveryReason &&
     (await failoverRetryController.maybeRetryTransient({
       reason: recoveryReason,
+      code: promptError ? getFailoverErrorCode(promptError) : assistantSignal?.code,
       message: promptError ? formatErrorMessage(promptError) : assistantSignal?.message,
       retryAfterMs: promptError
         ? resolveRetryAfterMs(formatErrorMessage(promptError), Date.now(), promptError)
diff --git a/src/agents/embedded-agent-runner/run/failover-retry-controller.ts b/src/agents/embedded-agent-runner/run/failover-retry-controller.ts
index f841272d23de..e51f2192e5db 100644
--- a/src/agents/embedded-agent-runner/run/failover-retry-controller.ts
+++ b/src/agents/embedded-agent-runner/run/failover-retry-controller.ts
@@ -252,6 +252,7 @@ export function createEmbeddedRunFailoverRetryController(input: {
     maybeRetryTransient: async (retry: {
       reason: TransientRetryReason;
       message?: string;
+      code?: string;
       retryAfterMs?: number;
       /** Saved retry.provider.maxRetryDelayMs; undefined or 0 disables the cap. */
       maxRetryDelayMs?: number;
@@ -268,6 +269,7 @@ export function createEmbeddedRunFailoverRetryController(input: {
         decision: "accepted" | "rejected",
         reason:
           | "non_transient"
+          | "connection_retry_disabled"
           | "long_window_rate_limit"
           | "retry_budget_exhausted"
           | "retry_delay_unavailable"
@@ -284,6 +286,14 @@ export function createEmbeddedRunFailoverRetryController(input: {
           },
           { config: params.config },
         );
+      if (
+        params.retryConnectionErrors === false &&
+        retry.code !== undefined &&
+        ["ECONNREFUSED", "ENOTFOUND", "EHOSTUNREACH", "ENETUNREACH"].includes(retry.code)
+      ) {
+        recordDecision("rejected", "connection_retry_disabled");
+        return false;
+      }
       if (
         retry.reason !== "rate_limit" &&
         retry.reason !== "overloaded" &&
diff --git a/src/agents/embedded-agent-runner/run/params.ts b/src/agents/embedded-agent-runner/run/params.ts
index a85f31b0423a..3fdec09e6950 100644
--- a/src/agents/embedded-agent-runner/run/params.ts
+++ b/src/agents/embedded-agent-runner/run/params.ts
@@ -116,6 +116,8 @@ export type RunEmbeddedAgentParams = {
   codeModeOverride?: boolean | "auto";
   /** Internal one-shot model probe mode: no tools, no workspace/chat prompt policy. */
   modelRun?: boolean;
+  /** Setup can reject unavailable endpoints without spending the session retry budget. */
+  retryConnectionErrors?: boolean;
   /** Disable trajectory persistence for auxiliary runs with no durable session owner. */
   disableTrajectory?: boolean;
   /** Restrict Skill Workshop to a bounded pending-proposal budget for an internal review run. */
diff --git a/src/system-agent/setup-inference-core.ts b/src/system-agent/setup-inference-core.ts
index ef0a6c115783..3863b4cf8368 100644
--- a/src/system-agent/setup-inference-core.ts
+++ b/src/system-agent/setup-inference-core.ts
@@ -8,6 +8,7 @@ import type { AgentRunResultView } from "../agents/agent-run-result.js";
 import type { loadAuthProfileStoreForRuntime } from "../agents/auth-profiles/store-runtime.js";
 import type { readCodexCliActiveApiKey } from "../agents/cli-credentials.js";
 import type { AgentExecutionAuthBinding } from "../agents/execution-auth-binding.js";
+import { describeFailoverError } from "../agents/failover-error.js";
 import type { FailoverReason } from "../agents/failover/signal.js";
 import { DEFAULT_AGENT_WORKSPACE_DIR } from "../agents/workspace-default.js";
 import type {
@@ -402,12 +403,34 @@ const SETUP_STATUS_BY_FAILOVER_REASON = {
   unknown: "unknown",
 } satisfies Record<FailoverReason, SetupInferenceFailureStatus>;
 
-export function mapFailoverReasonToSetupStatus(
+function mapFailoverReasonToSetupStatus(
   reason?: FailoverReason | null,
 ): SetupInferenceFailureStatus {
   return reason ? SETUP_STATUS_BY_FAILOVER_REASON[reason] : "unknown";
 }
 
+export function describeSetupInferenceError(
+  error: unknown,
+  route: SystemAgentConfiguredRoute,
+): { status: SetupInferenceFailureStatus; error: string } {
+  const described = describeFailoverError(error);
+  const origin = URL.parse(
+    route.runConfig.models?.providers?.[route.provider]?.baseUrl ?? "",
+  )?.origin;
+  const connectionError = !origin
+    ? undefined
+    : described.code === "ECONNREFUSED"
+      ? `Nothing is listening at ${origin}. Start the server or check the URL, then retry setup.`
+      : described.code === "ENOTFOUND"
+        ? `The server name in ${origin} could not be found. Check the URL and DNS settings, then retry setup.`
+        : described.code === "EHOSTUNREACH" || described.code === "ENETUNREACH"
+          ? `Cannot reach ${origin}. Check the URL and network connection from the Gateway host, then retry setup.`
+          : undefined;
+  return connectionError
+    ? { status: "unavailable", error: `${connectionError} No default model was changed.` }
+    : { status: mapFailoverReasonToSetupStatus(described.reason), error: described.message };
+}
+
 export function validateSetupInferenceOwnerEvidence(params: {
   runner: "cli" | "embedded";
   configuredHarnessId?: string;
diff --git a/src/system-agent/setup-inference-turn.test.ts b/src/system-agent/setup-inference-turn.test.ts
index ba77518df8b6..d2a3da205072 100644
--- a/src/system-agent/setup-inference-turn.test.ts
+++ b/src/system-agent/setup-inference-turn.test.ts
@@ -1,5 +1,6 @@
 import { describe, expect, it, vi } from "vitest";
 import { createDeferred } from "../../test/helpers/promise.js";
+import { FailoverError } from "../agents/failover-error.js";
 import {
   createPluginMetadataSnapshot,
   makeRegistry,
@@ -67,6 +68,49 @@ function embeddedRoute(runtime: "codex" | "openclaw" = "codex"): SystemAgentConf
 }
 
 describe("setup inference plugin ownership", () => {
+  it.each([
+    ["ECONNREFUSED", "Nothing is listening at http://127.0.0.1:43210", "Start the server"],
+    ["ENOTFOUND", "The server name in http://127.0.0.1:43210 could not be found", "DNS"],
+    ["EHOSTUNREACH", "Cannot reach http://127.0.0.1:43210", "network"],
+  ])(
+    "explains a failed connection without calling it a timeout (%s)",
+    async (code, message, nextStep) => {
+      const route = embeddedRoute("openclaw");
+      route.runConfig.models = {
+        providers: {
+          openai: {
+            baseUrl: "http://127.0.0.1:43210/v1?key=synthetic-private-value",
+            models: [],
+          },
+        },
+      };
+      let retryConnectionErrors: boolean | undefined;
+      const result = await runSetupInferenceTurn({
+        route,
+        requireExecutionOwner: false,
+        deps: {
+          createTempDir: async () => "/tmp/openclaw-setup-inference-test",
+          removeTempDir: async () => {},
+          runEmbeddedAgent: async (params) => {
+            retryConnectionErrors = params.retryConnectionErrors;
+            throw new FailoverError("Connection error.", { reason: "timeout", code });
+          },
+        },
+      });
+      expect(result).toMatchObject({
+        ok: false,
+        status: "unavailable",
+        error: expect.stringContaining(message),
+      });
+      expect(retryConnectionErrors).toBe(false);
+      if (!result.ok) {
+        expect(result.error).toContain(nextStep);
+        expect(result.error).toContain("No default model was changed.");
+        expect(result.error).not.toContain("synthetic-private-value");
+      }
+    },
+  );
+
   it("waits for the isolated probe runtime to release its plugin work", async () => {
     const route = embeddedRoute();
     const cleanupStarted = createDeferred();
diff --git a/src/system-agent/setup-inference-turn.ts b/src/system-agent/setup-inference-turn.ts
index 08c21aa0e280..3aac5bd177de 100644
--- a/src/system-agent/setup-inference-turn.ts
+++ b/src/system-agent/setup-inference-turn.ts
@@ -11,7 +11,6 @@ import {
 } from "../agents/agent-run-result.js";
 import { resolveAgentWorkspaceDir } from "../agents/agent-scope.js";
 import type { AgentExecutionAuthBinding } from "../agents/execution-auth-binding.js";
-import { describeFailoverError } from "../agents/failover-error.js";
 import type { AgentHarnessPluginSelection } from "../agents/harness/runtime-plugin-load-plan.js";
 import { loadAgentRuntimePluginRegistryHandle } from "../agents/runtime-plugins.js";
 import { SessionManager } from "../agents/sessions/index.js";
@@ -38,8 +37,8 @@ import {
   type ActivateSetupInferenceDeps,
   type BoundVerifySetupInferenceResult,
   type CompleteSetupInferenceResult,
+  describeSetupInferenceError,
   invalidSetupConfigError,
-  mapFailoverReasonToSetupStatus,
   parseInferenceRef,
   redactSetupInferenceError,
   resolveSetupInferenceWinnerError,
@@ -178,6 +177,7 @@ export async function runSetupInferenceTurn(params: {
         ...(route.authProfileId ? { authProfileIdSource: "user" as const } : {}),
         authProfileStateMode: "read-only",
         allowAuthProfileFallback: false,
+        retryConnectionErrors: false,
         preparedModelRuntimeMode: "isolated-read-only",
         ...(harness === "codex" ? { cleanupBundleMcpOnRunEnd: true } : {}),
         ...(harness ? { agentHarnessRuntimeOverride: harness } : {}),
@@ -230,8 +230,8 @@ export async function runSetupInferenceTurn(params: {
       auth: successfulAuth ?? (route.authProfileId ? { authProfileId: route.authProfileId } : {}),
     };
   } catch (error) {
-    const described = describeFailoverError(error);
-    return failed(mapFailoverReasonToSetupStatus(described.reason), described.message);
+    const described = describeSetupInferenceError(error, route);
+    return failed(described.status, described.error);
   } finally {
     preparedRunAdmission.close();
     clearAgentRunContext(runId);