fix(dashboard): test Responses nodes on /v1/responses, not chat completions (#13070) (#13087)

Boarded with 13 sibling PRs into one worktree off release/v3.8.51 and validated as a set: 132 focused tests pass across all 15 test files in the batch, typecheck:core is clean, check-changelog-integrity reports no lost base bullets, and check-file-size is green. Your PR merged without conflict against its siblings.

Thank you — the write-up made this reviewable: measuring the behaviour on the release tip and showing the before/after table meant the defect could be confirmed rather than taken on faith.

changelog.d/fixes/0000-responses-node-model-test.md

@@ -0,0 +1 @@
+- **fix(dashboard):** model health tests for a provider node set to the Responses API now call `/v1/responses` with a Responses-shaped body instead of `/v1/chat/completions` — those models were reported as `Provider returned HTTP 200 but no text content` even though the same model answered normally through `/v1/responses` ([#13070](https://github.com/diegosouzapw/OmniRoute/issues/13070))

src/lib/api/modelTestRunner.ts

@@ -3,7 +3,9 @@ import { POST as postChatCompletion } from "@/app/api/v1/chat/completions/route"
 import { POST as postAudioTranscription } from "@/app/api/v1/audio/transcriptions/route";
 import { handleValidatedEmbeddingRequestBody } from "@/app/api/v1/embeddings/route";
 import { POST as postRerank } from "@/app/api/v1/rerank/route";
+import { POST as postResponses } from "@/app/api/v1/responses/route";
 import {
+  buildComboTestPrompt,
   buildComboTestRequestBody,
   extractComboTestResponseText,
   extractComboTestStreamResult,
@@ -29,6 +31,10 @@ const ZAI_WEB_PROVIDER_ID = "zai-web";
 const ZAI_WEB_TEST_TIMEOUT_MS = 60_000;
 const SLOW_WEB_TEST_MODELS = new Set(["dola-pro"]);
 const STREAMING_CHAT_TEST_MAX_TOKENS = 64;
+// Responses calls the same budget `max_output_tokens`; `max_tokens` is silently
+// ignored on that endpoint, which would let a reasoning model spend the whole
+// default budget before emitting any visible text.
+const RESPONSES_TEST_MAX_OUTPUT_TOKENS = 256;
 
 function asRecord(value: unknown): Record<string, unknown> {
   return value && typeof value === "object" && !Array.isArray(value)
@@ -175,6 +181,26 @@ export function buildInternalChatRequest(
   });
 }
 
+export function buildInternalResponsesRequest(
+  testBody: Record<string, unknown>,
+  signal: AbortSignal,
+  connectionId?: string
+) {
+  return new Request(`${INTERNAL_ORIGIN}/v1/responses`, {
+    method: "POST",
+    headers: {
+      "Content-Type": "application/json",
+      "X-Internal-Test": "combo-health-check",
+      "X-OmniRoute-No-Cache": "true",
+      "X-OmniRoute-Compression": "off",
+      "X-Request-Id": `model-test-${randomUUID()}`,
+      ...(connectionId ? { "X-OmniRoute-Connection": connectionId } : {}),
+    },
+    body: JSON.stringify(testBody),
+    signal,
+  });
+}
+
 export function buildInternalRerankRequest(
   testBody: Record<string, unknown>,
   signal: AbortSignal,
@@ -265,7 +291,22 @@ export function detectTestKind(modelStr: string, customModel: any, nodeApiType?:
       lowerModel.includes("text-embed") ||
       lowerModel.includes("jina-clip") ||
       lowerModel.includes("colbert"));
-  return { isRerank, isEmbedding, isAudioTranscription };
+  // A Responses node answers on /v1/responses only. Without this the model fell
+  // through to the chat branch below, which posts a Chat Completions body to
+  // /v1/chat/completions: the route can still answer 200 while carrying nothing a
+  // Chat Completions reader recognises, so the model was marked unhealthy with
+  // "Provider returned HTTP 200 but no text content" (#13070).
+  //
+  // Last in the chain deliberately: a Responses-typed node can still host an
+  // embedding or rerank model, and those endpoints stay right for it.
+  const isResponses =
+    !isAudioTranscription &&
+    !isRerank &&
+    !isEmbedding &&
+    (apiFormat === "responses" ||
+      nodeType === "responses" ||
+      supportedEndpoints.includes("responses"));
+  return { isRerank, isEmbedding, isAudioTranscription, isResponses };
 }
 
 /**
@@ -424,7 +465,7 @@ export async function runSingleModelTest(
     findCustomModelMetadata(providerId, fullModelStr),
     findProviderNodeApiType(providerId),
   ]);
-  const { isRerank, isEmbedding, isAudioTranscription } = detectTestKind(
+  const { isRerank, isEmbedding, isAudioTranscription, isResponses } = detectTestKind(
     fullModelStr,
     customModel,
     nodeApiType
@@ -443,10 +484,22 @@ export async function runSingleModelTest(
       }
     : isAudioTranscription
       ? { model: fullModelStr }
-      : buildComboTestRequestBody(fullModelStr, isEmbedding, {
-          stream: !isEmbedding && streamChat,
-          maxTokens: !isEmbedding && streamChat ? STREAMING_CHAT_TEST_MAX_TOKENS : undefined,
-        });
+      : isResponses
+        ? {
+            model: fullModelStr,
+            // Responses takes `input`, not `messages`.
+            input: buildComboTestPrompt(),
+            max_output_tokens: RESPONSES_TEST_MAX_OUTPUT_TOKENS,
+            // Non-streaming on purpose: the SSE reader below understands Chat
+            // Completions deltas and the `output_text`/`output[]` shapes, but not
+            // Responses stream events (`response.output_text.delta`), so a
+            // streamed answer would read as empty — the very failure being fixed.
+            stream: false,
+          }
+        : buildComboTestRequestBody(fullModelStr, isEmbedding, {
+            stream: !isEmbedding && streamChat,
+            maxTokens: !isEmbedding && streamChat ? STREAMING_CHAT_TEST_MAX_TOKENS : undefined,
+          });
 
   // Per-model AbortController. We track whether the timeout fired so we can
   // distinguish "rate-limit queue aborted" (withRateLimit threw AbortError
@@ -473,6 +526,9 @@ export async function runSingleModelTest(
         buildInternalAudioTranscriptionRequest(fullModelStr, signal, connectionId)
       );
     }
+    if (isResponses) {
+      return postResponses(buildInternalResponsesRequest(testBody, signal, connectionId));
+    }
     return postChatCompletion(buildInternalChatRequest(testBody, signal, connectionId));
   };
 
@@ -577,7 +633,7 @@ export async function runSingleModelTest(
       // deactivated") would run outside runAsProbe and could still reach
       // markAccountUnavailable (#9817).
       const parsedResponse = await runAsProbe(() =>
-        extractModelTestResponseText(res, !isEmbedding && !isRerank && streamChat)
+        extractModelTestResponseText(res, !isEmbedding && !isRerank && !isResponses && streamChat)
       );
       responseText = parsedResponse.text;
       streamError = parsedResponse.error;

src/lib/combos/testHealth.ts

@@ -112,7 +112,7 @@ function getRandomFiveDigitNumber() {
   return COMBO_TEST_OPERAND_MIN + Math.floor(Math.random() * COMBO_TEST_OPERAND_RANGE);
 }
 
-function buildComboTestPrompt() {
+export function buildComboTestPrompt() {
   const left = getRandomFiveDigitNumber();
   const right = getRandomFiveDigitNumber();
 

tests/unit/model-test-runner.test.ts

@@ -74,6 +74,7 @@ test("detectTestKind defaults to a plain chat test for ordinary models", () => {
     isRerank: false,
     isEmbedding: false,
     isAudioTranscription: false,
+    isResponses: false,
   });
 });
 
@@ -95,6 +96,7 @@ test("detectTestKind detects rerank by id and by metadata, and rerank wins over
     isRerank: true,
     isEmbedding: false,
     isAudioTranscription: false,
+    isResponses: false,
   });
   // apiFormat metadata drives detection even when the id is opaque
   assert.equal(detectTestKind("vendor/opaque-model", { apiFormat: "rerank" }).isRerank, true);
@@ -116,6 +118,7 @@ test("detectTestKind detects audio transcription from metadata, and it wins over
     isRerank: false,
     isEmbedding: false,
     isAudioTranscription: true,
+    isResponses: false,
   });
   assert.equal(
     detectTestKind("vendor/opaque-model", { supportedEndpoints: ["audio-transcriptions"] })
@@ -152,6 +155,7 @@ test("detectTestKind falls back to the provider node's configured apiType", () =
     isRerank: false,
     isEmbedding: false,
     isAudioTranscription: false,
+    isResponses: false,
   });
 
   // Per-model metadata still wins when present.

tests/unit/responses-node-model-test-13070.test.ts

@@ -0,0 +1,184 @@
+/**
+ * #13070 -- the dashboard's per-model health test ignored a provider node's
+ * `apiType: "responses"`.
+ *
+ * `detectTestKind` mapped a node's apiType to audio, rerank and embeddings only,
+ * so every text model on a Responses node fell through to the chat branch and
+ * `buildInternalChatRequest` posted a Chat Completions body to
+ * /v1/chat/completions. A Responses-native upstream can answer 200 to that and
+ * still carry nothing a Chat Completions reader recognises, so the model went
+ * red with "Provider returned HTTP 200 but no text content" while the same
+ * model answered normally through /v1/responses.
+ *
+ * The classification tests below are cheap, but on their own they prove
+ * nothing: reverting the dispatch in runSingleModelTest and leaving
+ * detectTestKind alone keeps them all green. The last test is the one that
+ * fails in that case -- it reads the body that actually leaves for the
+ * upstream and asserts it is Responses-shaped.
+ */
+import test from "node:test";
+import assert from "node:assert/strict";
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+
+const TEST_DATA_DIR = fs.mkdtempSync(path.join(os.tmpdir(), "omniroute-13070-"));
+process.env.DATA_DIR = TEST_DATA_DIR;
+
+const core = await import("../../src/lib/db/core.ts");
+const nodesDb = await import("../../src/lib/db/providers/nodes.ts");
+const providersDb = await import("../../src/lib/db/providers.ts");
+const runner = await import("../../src/lib/api/modelTestRunner.ts");
+const callLogs = await import("../../src/lib/usage/callLogs.ts");
+
+const NODE_ID = "openai-compatible-responses-13070-0000-4000-8000-000000000000";
+const MODEL_ID = "opaque-text-model";
+
+test.after(() => {
+  core.resetDbInstance();
+  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true, maxRetries: 5, retryDelay: 100 });
+});
+
+// ---------------------------------------------------------------------------
+// detectTestKind — a Responses node must be recognised, and must not steal the
+// endpoints that were already right for it.
+// ---------------------------------------------------------------------------
+
+test("detectTestKind reports a Responses node, whichever field carries the signal", () => {
+  // An imported model has no per-model metadata at all; the node's apiType is
+  // the only signal available, which is exactly the reported case.
+  assert.equal(runner.detectTestKind("vendor/opaque-guid", null, "responses").isResponses, true);
+  assert.equal(
+    runner.detectTestKind("vendor/opaque-guid", { apiFormat: "responses" }).isResponses,
+    true
+  );
+  assert.equal(
+    runner.detectTestKind("vendor/opaque-guid", { supportedEndpoints: ["responses"] }).isResponses,
+    true
+  );
+});
+
+test("detectTestKind leaves an ordinary chat model alone", () => {
+  const kind = runner.detectTestKind("openai/gpt-4o", null);
+  assert.equal(kind.isResponses, false);
+  assert.equal(kind.isRerank, false);
+  assert.equal(kind.isEmbedding, false);
+  assert.equal(kind.isAudioTranscription, false);
+});
+
+test("embeddings, rerank and audio still win over a Responses node type", () => {
+  // A Responses-typed node can host these too, and /v1/responses is the wrong
+  // endpoint for all three. Losing this ordering would break working setups
+  // rather than fix a broken one.
+  assert.equal(
+    runner.detectTestKind("baai/bge-m3", null, "responses").isEmbedding,
+    true,
+    "embedding id must still route to embeddings"
+  );
+  assert.equal(runner.detectTestKind("baai/bge-m3", null, "responses").isResponses, false);
+
+  assert.equal(runner.detectTestKind("jina/jina-reranker-v2", null, "responses").isRerank, true);
+  assert.equal(
+    runner.detectTestKind("jina/jina-reranker-v2", null, "responses").isResponses,
+    false
+  );
+
+  const audio = runner.detectTestKind(
+    "vendor/whisper",
+    { apiFormat: "audio-transcriptions" },
+    "responses"
+  );
+  assert.equal(audio.isAudioTranscription, true);
+  assert.equal(audio.isResponses, false);
+});
+
+// ---------------------------------------------------------------------------
+// buildInternalResponsesRequest — the endpoint, and the bypass headers the
+// other builders carry. A health check that lost X-Internal-Test would be
+// rejected by strict mode instead of testing anything.
+// ---------------------------------------------------------------------------
+
+test("buildInternalResponsesRequest targets /v1/responses with the health-check headers", async () => {
+  const controller = new AbortController();
+  const req = runner.buildInternalResponsesRequest(
+    { model: "vendor/opaque", input: "hi" },
+    controller.signal,
+    "conn-1"
+  );
+
+  assert.equal(new URL(req.url).pathname, "/v1/responses");
+  assert.equal(req.method, "POST");
+  assert.equal(req.headers.get("X-Internal-Test"), "combo-health-check");
+  assert.equal(req.headers.get("X-OmniRoute-No-Cache"), "true");
+  assert.equal(req.headers.get("X-OmniRoute-Compression"), "off");
+  assert.equal(req.headers.get("X-OmniRoute-Connection"), "conn-1");
+  assert.deepEqual(await req.json(), { model: "vendor/opaque", input: "hi" });
+});
+
+test("buildInternalResponsesRequest omits the connection header when there is no connection", () => {
+  const req = runner.buildInternalResponsesRequest({ model: "m" }, new AbortController().signal);
+  assert.equal(req.headers.get("X-OmniRoute-Connection"), null);
+});
+
+// ---------------------------------------------------------------------------
+// The wiring. Everything above passes against the unfixed runner as long as
+// detectTestKind alone is changed; this one does not.
+// ---------------------------------------------------------------------------
+
+test("a model on a Responses node is probed on the internal /v1/responses route", async () => {
+  await nodesDb.createProviderNode({
+    id: NODE_ID,
+    type: "openai-compatible",
+    name: "Responses Node 13070",
+    prefix: "resp13070",
+    apiType: "responses",
+    baseUrl: "https://example.test/v1",
+  });
+  const connection = await providersDb.createProviderConnection({
+    provider: NODE_ID,
+    authType: "apikey",
+    name: "responses-node-13070",
+    apiKey: "sk-responses-node-13070",
+    isActive: true,
+    testStatus: "active",
+  });
+
+  const originalFetch = globalThis.fetch;
+  globalThis.fetch = (async () =>
+    // A minimal Responses reply. `output_text` is a field the existing
+    // extractor already understands, which is why this fix needs no reader
+    // change -- only the request side was ever wrong.
+    new Response(JSON.stringify({ output_text: "4" }), {
+      status: 200,
+      headers: { "content-type": "application/json" },
+    })) as typeof globalThis.fetch;
+
+  try {
+    await runner.runSingleModelTest({
+      providerId: NODE_ID,
+      modelId: MODEL_ID,
+      connectionId: String(connection.id),
+      timeoutMs: 15_000,
+    });
+  } finally {
+    globalThis.fetch = originalFetch;
+  }
+
+  await callLogs.waitForCallLogSaves(10_000);
+  const logs = await callLogs.getCallLogs({});
+  const probe = logs.find((entry: { model?: string | null }) =>
+    String(entry.model ?? "").includes(MODEL_ID)
+  );
+
+  assert.ok(probe, "the model test should have produced a call log entry");
+  // This is the line from the report: the call log showed
+  // path=/v1/chat/completions for a Responses node. Asserting on the
+  // upstream request instead would prove nothing -- the router translates a
+  // chat body into Responses shape for such a node either way, so that
+  // assertion stays green with the dispatch below reverted.
+  assert.equal(
+    probe.path,
+    "/v1/responses",
+    `a Responses node must be probed on /v1/responses (call log says ${probe.path})`
+  );
+});