diff --git a/packages/runtime/src/__tests__/provider-conformance.test.ts b/packages/runtime/src/__tests__/provider-conformance.test.ts
index 06c3cad86a..b53a8dbf98 100644
--- a/packages/runtime/src/__tests__/provider-conformance.test.ts
+++ b/packages/runtime/src/__tests__/provider-conformance.test.ts
@@ -956,6 +956,68 @@ describe('models.dev provider conformance', () => {
     assert.equal(probedPath, '/v1/responses');
   });
 
+  test('a Responses relay configured with its host root discovers and probes one /v1 base', async () => {
+    const requests: Array<{ method: string; url: string }> = [];
+    const server = await startJsonServer((request, response) => {
+      requests.push({ method: request.method ?? '', url: request.url ?? '' });
+      if (request.method === 'GET' && request.url === '/v1/models') {
+        respondJson(response, 200, { data: [{ id: 'relay-reasoner' }] });
+        return;
+      }
+      if (request.method === 'POST' && request.url === '/v1/responses') {
+        respondJson(response, 200, {});
+        return;
+      }
+      // A mainstream relay mounts the OpenAI surface under `/v1` only. That is
+      // what let the host-root form pass the probe and 404 on discovery: both
+      // addresses have to come from one normalized API base (#3320).
+      respondJson(response, 404, { error: { message: 'not found' } });
+    });
+    const connection: LlmConnection = {
+      slug: 'responses-relay',
+      name: 'Responses Relay',
+      providerType: 'openai-responses-compatible',
+      baseUrl: server.url,
+      defaultModel: 'relay-reasoner',
+      enabled: true,
+      createdAt: 1,
+      updatedAt: 1,
+    };
+
+    assert.deepEqual(
+      (await fetchProviderModels(connection, 'relay-key')).map(({ id }) => id),
+      ['relay-reasoner'],
+    );
+    assert.equal((await testConnection(connection, 'relay-key')).ok, true);
+    assert.deepEqual(requests, [
+      { method: 'GET', url: '/v1/models' },
+      { method: 'POST', url: '/v1/responses' },
+    ]);
+  });
+
+  test('a DeepSeek root base keeps serving Responses at the root it publishes', async () => {
+    let probedPath: string | undefined;
+    const server = await startJsonServer((request, response) => {
+      probedPath = request.url;
+      respondJson(response, 200, {});
+    });
+    const connection: LlmConnection = {
+      slug: 'deepseek',
+      name: 'DeepSeek',
+      providerType: 'deepseek',
+      baseUrl: server.url,
+      defaultModel: 'deepseek-v4-pro',
+      enabled: true,
+      createdAt: 1,
+      updatedAt: 1,
+    };
+
+    // The relay self-heal is provider-scoped: a built-in that publishes an
+    // unversioned root must not be rewritten to `/v1` behind the user's back.
+    assert.equal((await testConnection(connection, 'deepseek-token')).ok, true);
+    assert.equal(probedPath, '/responses');
+  });
+
   test('Ollama Cloud requests usage in streamed chat completions', async () => {
     let requestBody: Record<string, unknown> | undefined;
     const server = await startJsonServer(async (request, response) => {
diff --git a/packages/runtime/src/__tests__/provider-contract-matrix.test.ts b/packages/runtime/src/__tests__/provider-contract-matrix.test.ts
index 4f3b95ffe5..d87486a1e0 100644
--- a/packages/runtime/src/__tests__/provider-contract-matrix.test.ts
+++ b/packages/runtime/src/__tests__/provider-contract-matrix.test.ts
@@ -202,7 +202,7 @@ function assertDiscoveryRequest(
   // Path: the declared path, or the protocol's default models path.
   assert.equal(
     url.pathname,
-    expectedDiscoveryPathname(discovery),
+    expectedDiscoveryPathname(row, discovery),
     `${where} must request the declared models path`,
   );
 
@@ -249,7 +249,10 @@ function assertDiscoveryRequest(
   }
 }
 
-function expectedDiscoveryPathname(discovery: ProviderContractDiscoveryPlan): string {
+function expectedDiscoveryPathname(
+  row: ProviderContractRow,
+  discovery: ProviderContractDiscoveryPlan,
+): string {
   switch (discovery.protocol) {
     case 'anthropic':
       return '/v1/models';
@@ -257,7 +260,11 @@ function expectedDiscoveryPathname(discovery: ProviderContractDiscoveryPlan): st
       return '/v1beta/models';
     default: {
       const path = discovery.path ?? '/models';
-      return path.startsWith('/') ? path : `/${path}`;
+      const pathname = path.startsWith('/') ? path : `/${path}`;
+      // A Responses relay normalizes the host root `baseConnection` hands it to
+      // `/v1`, so discovery and the Responses request resolve from one API base
+      // instead of splitting across `/models` and `/v1/models` (#3320).
+      return row.providerType === 'openai-responses-compatible' ? `/v1${pathname}` : pathname;
     }
   }
 }
diff --git a/packages/runtime/src/__tests__/responses-wire-contract.test.ts b/packages/runtime/src/__tests__/responses-wire-contract.test.ts
index 99b1bc93cf..7348ffd715 100644
--- a/packages/runtime/src/__tests__/responses-wire-contract.test.ts
+++ b/packages/runtime/src/__tests__/responses-wire-contract.test.ts
@@ -28,7 +28,11 @@ import { buildProviderOptions, getAIModel } from '../model-factory.js';
 import { resolveModelRuntime } from '../model-runtime.js';
 import { lowerModelTools } from '../model-adapter.js';
 import { openAiCodexCompactionMessages } from '../openai-codex-history-compactor.js';
-import { openAiResponsesBaseUrl, openResponsesUrl } from '../provider-urls.js';
+import {
+  openAiResponsesBaseUrl,
+  openResponsesUrl,
+  responsesRelayApiBaseUrl,
+} from '../provider-urls.js';
 
 function conn(providerType: LlmConnection['providerType'], slug = 'test'): LlmConnection {
   return {
@@ -141,6 +145,68 @@ describe('responses wire contract', () => {
     assert.deepEqual(urls, ['https://relay.example/v1/responses']);
   });
 
+  test('resolves a Responses relay host root to one versioned API base', async () => {
+    assert.equal(
+      responsesRelayApiBaseUrl('http://relay.example:3000'),
+      'http://relay.example:3000/v1',
+    );
+    assert.equal(
+      responsesRelayApiBaseUrl('http://relay.example:3000/'),
+      'http://relay.example:3000/v1',
+    );
+    assert.equal(responsesRelayApiBaseUrl('https://relay.example/v1'), 'https://relay.example/v1');
+    assert.equal(responsesRelayApiBaseUrl('https://relay.example/v1/'), 'https://relay.example/v1');
+    assert.equal(
+      responsesRelayApiBaseUrl('https://relay.example/v1/responses'),
+      'https://relay.example/v1',
+    );
+    assert.equal(
+      responsesRelayApiBaseUrl('https://relay.example/relay/v1'),
+      'https://relay.example/relay/v1',
+    );
+    // No base to normalize: the caller fails where it already failed instead of
+    // on a URL constructor.
+    assert.equal(responsesRelayApiBaseUrl(''), '');
+
+    // The relay resolves send, probe, and discovery from that one base, while a
+    // built-in that publishes an unversioned root keeps it (#3320).
+    assert.equal(
+      resolveModelRuntime(
+        { providerType: 'openai-responses-compatible', baseUrl: 'http://relay.example:3000' },
+        'relay-model',
+      ).baseUrl,
+      'http://relay.example:3000/v1',
+    );
+    assert.equal(
+      resolveModelRuntime({ providerType: 'deepseek' }, 'deepseek-v4-flash').baseUrl,
+      'https://api.deepseek.com',
+    );
+
+    const urls: string[] = [];
+    const fetch = (async (url: string | URL | Request) => {
+      urls.push(String(url));
+      return Response.json({
+        id: 'r',
+        object: 'response',
+        status: 'completed',
+        output: [],
+        usage: { input_tokens: 1, output_tokens: 1 },
+      });
+    }) as typeof globalThis.fetch;
+    const model = getAIModel({
+      connection: { ...conn('openai-responses-compatible'), baseUrl: 'http://relay.example:3000' },
+      apiKey: '[redacted]',
+      modelId: 'relay-model',
+      fetch,
+    });
+
+    await model.doGenerate({
+      prompt: [{ role: 'user', content: [{ type: 'text', text: 'ping' }] }],
+    });
+
+    assert.deepEqual(urls, ['http://relay.example:3000/v1/responses']);
+  });
+
   test('resolves only supported Responses adapter and replay pairings', () => {
     const deepseek = resolveModelRuntime({ providerType: 'deepseek' }, 'deepseek-v4-flash');
     assert.deepEqual(deepseek.reasoningReplay, {
diff --git a/packages/runtime/src/model-fetcher.ts b/packages/runtime/src/model-fetcher.ts
index 3b4e478509..e07aa7ab96 100644
--- a/packages/runtime/src/model-fetcher.ts
+++ b/packages/runtime/src/model-fetcher.ts
@@ -30,7 +30,7 @@ import {
   CONNECTION_MODEL_ID_MAX_LENGTH,
   normalizeConnectionModelDiscoveryResult,
 } from '@maka/core/runtime-policy';
-import { anthropicV1Url, googleApiUrl } from './provider-urls.js';
+import { anthropicV1Url, googleApiUrl, responsesRelayApiBaseUrl } from './provider-urls.js';
 import { openAiCodexHeaders } from './subscription-auth.js';
 import {
   GITHUB_COPILOT_API_VERSION,
@@ -179,7 +179,15 @@ async function fetchProviderModelsStrict(
   apiKey: string,
   fetchFn: ConnectionEffectFetch | undefined,
 ): Promise<ModelInfo[]> {
-  const baseUrl = effectiveBaseUrl(connection);
+  // A Responses relay serves `/models` next to `/responses`, so discovery has
+  // to resolve from the same API base `resolveModelRuntime` hands the send and
+  // probe paths — otherwise a configured host root passes the probe and 404s
+  // here (#3320).
+  const configuredBaseUrl = effectiveBaseUrl(connection);
+  const baseUrl =
+    connection.providerType === 'openai-responses-compatible'
+      ? responsesRelayApiBaseUrl(configuredBaseUrl)
+      : configuredBaseUrl;
   const definition = PROVIDER_DEFAULTS[connection.providerType];
   // Unknown providerType → no discovery path. Throw a clear error (caught and
   // generalized by the caller) rather than crashing on `.modelDiscovery`.
diff --git a/packages/runtime/src/model-runtime.ts b/packages/runtime/src/model-runtime.ts
index 874c3e206b..d2db4f1a39 100644
--- a/packages/runtime/src/model-runtime.ts
+++ b/packages/runtime/src/model-runtime.ts
@@ -31,6 +31,7 @@ import {
   openAiAdapterApiProtocol,
 } from '@maka/core/model-metadata';
 import { isRetiredProvider } from '@maka/core/provider-registry';
+import { responsesRelayApiBaseUrl } from './provider-urls.js';
 import { resolveApplyPatchProfile, type ApplyPatchProfile } from './apply-patch-profile.js';
 
 export type ModelRuntimeWire =
@@ -120,10 +121,16 @@ export function resolveModelRuntime(
   const parallelToolCalls = resolveParallelToolCalls(connection, modelId, adapter);
   return {
     adapter,
+    // A Responses relay resolves its send and probe addresses from the same
+    // API base that model discovery uses in `fetchProviderModelsStrict`, so a
+    // configured host root cannot answer `/responses` while `/models` 404s
+    // (#3320).
     baseUrl:
       connection.providerType === 'kimi-coding-plan' && apiProtocol === 'openai-chat'
         ? kimiOpenAiBaseUrl(resolvedBaseUrl)
-        : resolvedBaseUrl,
+        : connection.providerType === 'openai-responses-compatible'
+          ? responsesRelayApiBaseUrl(resolvedBaseUrl)
+          : resolvedBaseUrl,
     ...(apiProtocol ? { apiProtocol } : {}),
     wire,
     reasoningReplay: reasoningReplayContract(adapter, wire),
diff --git a/packages/runtime/src/provider-urls.ts b/packages/runtime/src/provider-urls.ts
index f7ddbffc6a..6e08928d1e 100644
--- a/packages/runtime/src/provider-urls.ts
+++ b/packages/runtime/src/provider-urls.ts
@@ -75,6 +75,35 @@ export function openAiResponsesBaseUrl(baseUrl: string): string {
   return url.toString();
 }
 
+/**
+ * A custom OpenAI Responses relay is configured with a single URL that has to
+ * serve both `…/responses` (send and probe) and `…/models` (discovery). Users
+ * paste the host root — `http://relay.example:3000` — which mainstream relays
+ * mount under `/v1`, so `/responses` answers at the root while `/models` 404s
+ * and the model catalog stops refreshing behind a passing probe (#3320).
+ *
+ * Normalize the configured value into the one API base every relay call site
+ * derives from: a path-less root self-heals to `/v1`, as `anthropicV1BaseUrl`
+ * and `googleV1BetaBaseUrl` already do for their versions; an authored path
+ * (`/v1`, `/relay/v1`) is preserved verbatim; the endpoint form reduces to its
+ * base. Callers scope this to the relay provider, because built-ins such as
+ * DeepSeek legitimately serve `/responses` and `/models` at their own root.
+ *
+ * A value that is not a parsable absolute URL is returned untouched: there is
+ * no base to normalize, and the request it feeds fails where it fails today.
+ */
+export function responsesRelayApiBaseUrl(baseUrl: string): string {
+  let url: URL;
+  try {
+    url = new URL(baseUrl);
+  } catch {
+    return baseUrl;
+  }
+  const basePath = stripTrailing(url.pathname).replace(/\/responses$/i, '');
+  url.pathname = basePath === '' ? '/v1' : basePath;
+  return url.toString();
+}
+
 function stripTrailing(u: string): string {
   return u.replace(/\/+$/, '');
 }