diff --git a/AGENTS.md b/AGENTS.md
index 0f01b459c..33b932c2c 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -235,7 +235,7 @@ The `error-context.md` file contains a YAML accessibility tree of the entire pag
 - Has a different `name`/`data-testid` than expected
 - Is behind a lazy-load boundary that hasn't resolved
 
-**Stale state from earlier serial tests** — Mock-LLM tests run serially (`workers: 1`) against a real agent-server. Earlier tests (conversation, automation) persist settings on the server. If your test depends on "clean" state but a prior test configured `llm_base_url`, `llm_model`, etc., the form may render in a different view mode. Use Playwright `page.route()` to intercept and normalize the settings response. Example: `routeOnboardingLlmCatalog` in `tests/e2e/support/onboarding-helpers.ts` intercepts `GET /api/settings` to clear `llm_base_url` so the LLM form always opens in "Basic" view.
+**Stale state from earlier serial tests** — Mock-LLM tests run serially (`workers: 1`) against a real agent-server. Earlier tests (conversation, automation) persist settings on the server. If your test needs "first-run" state (e.g. onboarding), use `interceptSettingsForFirstRun(page)` from `mock-llm-helpers.ts` in the `beforeGoto` callback — it intercepts `GET /api/settings` to strip stale `agent_settings.llm` values so the frontend sees schema defaults. This is necessary because (a) the agent-server PATCH API rejects null/empty values for validated fields like `model` and `api_key` (422), and (b) the `inferInitialView` function checks ALL schema fields against their defaults to decide between Basic/Advanced view. The interception must be registered AFTER `routeSessionApiKey` (Playwright LIFO order).
 
 **View mode mismatch (Basic vs Advanced)** — `LlmSettingsScreen` switches between "Basic" (renders `ModelSelector` with provider/model dropdowns) and "Advanced" (renders plain text inputs). The view is determined by `getInitialView()` which checks `currentSettings.llm_base_url` — a non-default base URL triggers "Advanced" view. If your test expects `input[name="llm-provider-input"]` but sees text inputs instead, the settings have a stale `base_url`.
 
diff --git a/tests/e2e/mock-llm/mock-llm-onboarding-regressions.spec.ts b/tests/e2e/mock-llm/mock-llm-onboarding-regressions.spec.ts
index fcfe90c69..17382b3ff 100644
--- a/tests/e2e/mock-llm/mock-llm-onboarding-regressions.spec.ts
+++ b/tests/e2e/mock-llm/mock-llm-onboarding-regressions.spec.ts
@@ -5,7 +5,11 @@ import {
   showOnboarding,
   waitForOnboardingStep,
 } from "../support/onboarding-helpers";
-import { routeSessionApiKey, SESSION_API_KEY } from "./utils/mock-llm-helpers";
+import {
+  interceptSettingsForFirstRun,
+  routeSessionApiKey,
+  SESSION_API_KEY,
+} from "./utils/mock-llm-helpers";
 
 test.describe.configure({ mode: "serial" });
 
@@ -70,22 +74,8 @@ test.describe("onboarding recent regressions", () => {
       apiKey: SESSION_API_KEY,
       beforeGoto: async () => {
         await routeSessionApiKey(page);
-        // Intercept GET /api/settings so the LLM form sees a clean
-        // base_url, forcing basic view mode regardless of what earlier
-        // specs configured. Registered AFTER routeSessionApiKey so
-        // Playwright's LIFO matching picks this up first for settings.
-        await page.route("**/api/settings", async (route, req) => {
-          if (req.method() !== "GET") {
-            await route.fallback();
-            return;
-          }
-          const response = await route.fetch();
-          const body = await response.json();
-          if (body?.agent_settings?.llm) {
-            body.agent_settings.llm.base_url = null;
-          }
-          await route.fulfill({ response, json: body });
-        });
+        // Strip stale LLM config so the form sees a first-run state.
+        await interceptSettingsForFirstRun(page);
       },
     });
     await advanceOnboardingToLlmStep(page);
diff --git a/tests/e2e/mock-llm/utils/mock-llm-helpers.ts b/tests/e2e/mock-llm/utils/mock-llm-helpers.ts
index b557a6aff..3d713b841 100644
--- a/tests/e2e/mock-llm/utils/mock-llm-helpers.ts
+++ b/tests/e2e/mock-llm/utils/mock-llm-helpers.ts
@@ -422,6 +422,36 @@ export async function resetMockLLM(request: APIRequestContext) {
   expect(resp.ok(), `Reset mock LLM: ${resp.status()}`).toBe(true);
 }
 
+/**
+ * Intercept `GET /api/settings` to strip all LLM configuration, simulating
+ * a first-run state where no model/api_key/base_url has been configured.
+ *
+ * Earlier specs (conversation, automation, etc.) persist custom LLM settings
+ * on the real agent-server. The settings schema's `inferInitialView` checks
+ * ALL non-critical fields against their defaults — if any differ (model,
+ * api_key, base_url), the LLM form renders in "Advanced" view instead of
+ * "Basic". The PATCH API rejects null/empty values for most fields (422),
+ * so we can't reset them server-side. Instead we strip the stale values from
+ * the response so the frontend sees schema defaults.
+ *
+ * Must be registered AFTER `routeSessionApiKey` so Playwright's LIFO order
+ * processes the settings interception first for matching requests.
+ */
+export async function interceptSettingsForFirstRun(page: Page) {
+  await page.route("**/api/settings", async (route, req) => {
+    if (req.method() !== "GET") {
+      await route.fallback();
+      return;
+    }
+    const response = await route.fetch();
+    const body = await response.json();
+    if (body?.agent_settings?.llm) {
+      body.agent_settings.llm = {};
+    }
+    await route.fulfill({ response, json: body });
+  });
+}
+
 /**
  * Fetch all chat-completion request bodies captured by the mock LLM server
  * since the last /admin/reset.