diff --git a/src/agents/embedded-agent-runner.run-embedded-agent.auth-profile-rotation.e2e.test.ts b/src/agents/embedded-agent-runner.run-embedded-agent.auth-profile-rotation.e2e.test.ts
index c3ef88c40e29..15a8259d05a3 100644
--- a/src/agents/embedded-agent-runner.run-embedded-agent.auth-profile-rotation.e2e.test.ts
+++ b/src/agents/embedded-agent-runner.run-embedded-agent.auth-profile-rotation.e2e.test.ts
@@ -90,7 +90,6 @@ const installRunEmbeddedMocks = () => {
 };
 
 let runEmbeddedAgent: typeof import("./embedded-agent-runner/run.js").runEmbeddedAgent;
-let authProfileUsageTesting: typeof import("./auth-profiles/usage.test-support.js").testing;
 let createDiagnosticLogRecordCaptureFn: typeof import("../logging/test-helpers/diagnostic-log-capture.js").createDiagnosticLogRecordCapture;
 let cleanupLogCapture: (() => void) | undefined;
 let resetLoggerFn: typeof import("../logging/logger.js").resetLogger;
@@ -101,7 +100,6 @@ beforeAll(async () => {
   vi.resetModules();
   installRunEmbeddedMocks();
   ({ runEmbeddedAgent } = await import("./embedded-agent-runner/run.js"));
-  ({ testing: authProfileUsageTesting } = await import("./auth-profiles/usage.test-support.js"));
   ({ createDiagnosticLogRecordCapture: createDiagnosticLogRecordCaptureFn } =
     await import("../logging/test-helpers/diagnostic-log-capture.js"));
   ({ resetLogger: resetLoggerFn, setLoggerOverride: setLoggerOverrideFn } =
@@ -141,7 +139,6 @@ beforeEach(() => {
 
 afterEach(() => {
   globalThis.fetch = originalFetch;
-  authProfileUsageTesting.setDepsForTest(null);
   cleanupLogCapture?.();
   cleanupLogCapture = undefined;
   setLoggerOverrideFn(null);
@@ -937,46 +934,6 @@ describe("runEmbeddedAgent auth profile rotation", () => {
     expect(sleepWithAbortMock).not.toHaveBeenCalled();
   });
 
-  it("does not wait for prompt failure cooldown marking before retrying", async () => {
-    let releaseMark: (() => void) | undefined;
-    const markCanFinish = new Promise<void>((resolve) => {
-      releaseMark = resolve;
-    });
-    let markStarted = false;
-    authProfileUsageTesting.setDepsForTest({
-      updateAuthProfileStoreWithLock: async () => {
-        markStarted = true;
-        await markCanFinish;
-        return null;
-      },
-    });
-
-    try {
-      await withAgentWorkspace(async ({ agentDir, workspaceDir }) => {
-        await writeAuthStore(agentDir);
-        mockPromptErrorThenSuccessfulAttempt("rate limit exceeded");
-
-        const runPromise = runAutoPinnedOpenAiTurn({
-          agentDir,
-          workspaceDir,
-          sessionKey: "agent:test:prompt-deferred-mark",
-          runId: "run:prompt-deferred-mark",
-        });
-
-        await vi.waitFor(() => expect(runEmbeddedAttemptMock).toHaveBeenCalledTimes(2));
-        expect(markStarted).toBe(true);
-        releaseMark?.();
-        releaseMark = undefined;
-        await runPromise;
-
-        const usageStats = await readUsageStats(agentDir);
-        expect(typeof usageStats["openai:p2"]?.lastUsed).toBe("number");
-      });
-    } finally {
-      releaseMark?.();
-    }
-  });
-
   it("rotates on timeout without cooling down the timed-out profile", async () => {
     const { usageStats } = await runAutoPinnedRotationCase({
       errorMessage: "request ended without sending any chunks",
diff --git a/src/agents/embedded-agent-runner/run/prompt-failure.test.ts b/src/agents/embedded-agent-runner/run/prompt-failure.test.ts
new file mode 100644
index 000000000000..d1919f6a8cb9
--- /dev/null
+++ b/src/agents/embedded-agent-runner/run/prompt-failure.test.ts
@@ -0,0 +1,107 @@
+// Prompt failure tests cover retry composition around profile rotation.
+import { describe, expect, it, vi } from "vitest";
+import { handleEmbeddedPromptFailure } from "./prompt-failure.js";
+
+type Params = Parameters<typeof handleEmbeddedPromptFailure>[0];
+
+function makeParams(overrides: Partial<Params> = {}): Params {
+  const provider = "openai";
+  const modelId = "gpt-5";
+  const defaults: Params = {
+    runParams: {
+      config: undefined,
+      runId: "run:prompt-failure-test",
+    } as Params["runParams"],
+    attempt: {
+      replayMetadata: {
+        replaySafe: true,
+      },
+    } as Params["attempt"],
+    promptError: new Error("rate limit exceeded"),
+    promptErrorSource: "prompt",
+    activeErrorContext: { provider, model: modelId },
+    provider,
+    modelId,
+    authProfileId: "openai:p1",
+    authProfileStore: {
+      version: 1,
+      profiles: {},
+    },
+    sessionIdUsed: "session:prompt-failure-test",
+    lane: "test",
+    agentDir: "/tmp/openclaw-prompt-failure-test",
+    suspensionSessionId: "session:prompt-failure-test",
+    runtimeAuthRetry: false,
+    maybeRefreshRuntimeAuthForAuthError: vi.fn(async () => false),
+    suspendForFailure: vi.fn(),
+    resolveReplayInvalid: vi.fn(() => false),
+    setTerminalLifecycleMeta: vi.fn(),
+    buildErrorAgentMeta: vi.fn(),
+    startedAtMs: 0,
+    fallbackConfigured: true,
+    aborted: false,
+    externalAbort: false,
+    pluginHarnessOwnsTransport: false,
+    timedOutByRunBudget: false,
+    resolveAuthProfileFailureReason: vi.fn(() => "rate_limit"),
+    maybeEscalateRateLimitProfileFallback: vi.fn(),
+    advanceAttemptAuthProfile: vi.fn(async () => true),
+    maybeMarkAuthProfileFailure: vi.fn(async () => {}),
+    maybeBackoffBeforeOverloadFailover: vi.fn(async () => {}),
+    attemptedThinking: new Set(),
+    thinkLevel: "low",
+    getThinkLevel: () => "low",
+    traceAttempts: [],
+    previousRetryFailoverReason: null,
+  };
+  return { ...defaults, ...overrides };
+}
+
+describe("handleEmbeddedPromptFailure", () => {
+  it("returns the profile-rotation retry before failure marking finishes", async () => {
+    const events: string[] = [];
+    let releaseMark: (() => void) | undefined;
+    const markCanFinish = new Promise<void>((resolve) => {
+      releaseMark = resolve;
+    });
+    const maybeMarkAuthProfileFailure = vi.fn(async () => {
+      events.push("mark-start");
+      await markCanFinish;
+      events.push("mark-finish");
+    });
+
+    try {
+      const outcome = await handleEmbeddedPromptFailure(
+        makeParams({
+          advanceAttemptAuthProfile: vi.fn(async () => {
+            events.push("advance");
+            return true;
+          }),
+          maybeMarkAuthProfileFailure,
+          maybeBackoffBeforeOverloadFailover: vi.fn(async () => {
+            events.push("backoff");
+          }),
+        }),
+      );
+
+      expect(outcome).toEqual({
+        action: "retry",
+        thinkLevel: "low",
+        authRetryPending: false,
+        lastRetryFailoverReason: "rate_limit",
+      });
+      expect(events).toEqual(["advance", "mark-start", "backoff"]);
+      expect(maybeMarkAuthProfileFailure).toHaveBeenCalledWith({
+        profileId: "openai:p1",
+        reason: "rate_limit",
+        modelId: "gpt-5",
+      });
+    } finally {
+      releaseMark?.();
+    }
+
+    await vi.waitFor(() =>
+      expect(events).toEqual(["advance", "mark-start", "backoff", "mark-finish"]),
+    );
+  });
+});