fix(e2e): echo unique launch replies in mock inference

Author: prekshivyas

test/e2e/fixtures/fake-openai-compatible.ts

@@ -40,6 +40,7 @@ export interface FakeOpenAiCompatibleServerOptions {
   readonly apiKey?: string;
   readonly chatContent?: string;
   readonly forbiddenMarkers?: readonly string[];
+  readonly launchReplyFromPrompt?: boolean;
   /** Non-secret marker expected in a request under test. */
   readonly requestCanaryMarker?: string;
   readonly host?: string;
@@ -184,6 +185,7 @@ export async function startFakeOpenAiCompatibleServer(
           NEMOCLAW_FAKE_OPENAI_FORBIDDEN_MARKERS: JSON.stringify(options.forbiddenMarkers ?? []),
           NEMOCLAW_FAKE_OPENAI_HOST: host,
           NEMOCLAW_FAKE_OPENAI_LOG_FILE: logFile,
+          NEMOCLAW_FAKE_OPENAI_LAUNCH_REPLY_FROM_PROMPT: options.launchReplyFromPrompt ? "1" : "0",
           NEMOCLAW_FAKE_OPENAI_MAX_MODEL_LEN:
             options.maxModelLen !== undefined ? String(options.maxModelLen) : "",
           NEMOCLAW_FAKE_OPENAI_MODEL: options.model ?? "test-model",

test/e2e/fixtures/inference-adapter.ts

@@ -390,6 +390,7 @@ export async function createE2EInferenceAdapter(
       // alias, so listen on the bridge-facing interfaces. The workflow uses an
       // ephemeral ubuntu-latest VM, an OS-assigned port, and a per-run credential.
       host: "0.0.0.0",
+      launchReplyFromPrompt: true,
       model,
       publicHost: SANDBOX_HOST_ALIAS,
       progress: options.progress,

test/e2e/lib/fake-openai-compatible-api.mts

@@ -28,6 +28,8 @@ const requireAuth = process.env.NEMOCLAW_FAKE_OPENAI_REQUIRE_AUTH === "1";
 const requireAuthModels = process.env.NEMOCLAW_FAKE_OPENAI_REQUIRE_AUTH_MODELS === "1";
 const chatContent = process.env.NEMOCLAW_FAKE_OPENAI_CHAT_CONTENT || "ok";
 const responseText = process.env.NEMOCLAW_FAKE_OPENAI_RESPONSE_TEXT || chatContent;
+const launchReplyFromPrompt =
+  process.env.NEMOCLAW_FAKE_OPENAI_LAUNCH_REPLY_FROM_PROMPT === "1";
 const requestCanaryMarker = process.env.NEMOCLAW_FAKE_OPENAI_REQUEST_CANARY_MARKER || "";
 const forbiddenMarkers = (() => {
   try {
@@ -145,6 +147,27 @@ function requestCanaryPresent(req: IncomingMessage, raw: Buffer): boolean | unde
   return requestMaterial.includes(requestCanaryMarker);
 }
 
+function latestUserPrompt(payload: JsonObject): string | null {
+  const entries = Array.isArray(payload.messages) ? payload.messages : [];
+  for (let index = entries.length - 1; index >= 0; index -= 1) {
+    const entry = entries[index];
+    if (!entry || typeof entry !== "object") continue;
+    const message = entry as JsonObject;
+    if (message.role !== "user") continue;
+    if (typeof message.content === "string") return message.content;
+  }
+  return null;
+}
+
+function requestedLaunchReply(payload: JsonObject): string | null {
+  if (!launchReplyFromPrompt) return null;
+  const prompt = latestUserPrompt(payload);
+  const match = prompt?.match(
+    /^Join these four fragments with underscores and put only the result on its own line: NEMOCLAW, ([0-9A-F]{12}), (FIRST|SECOND), OK\. Do not use tools\.$/u,
+  );
+  return match ? `NEMOCLAW_${match[1]}_${match[2]}_OK` : null;
+}
+
 const server = createServer(async (req, res) => {
   const path = requestPath(req);
 
@@ -198,8 +221,9 @@ const server = createServer(async (req, res) => {
       sendJson(res, 401, { error: { message: "missing bearer credential" } });
       return;
     }
+    const content = requestedLaunchReply(payload) ?? chatContent;
     if (payload.stream) {
-      sendChatSse(res, chatContent);
+      sendChatSse(res, content);
       return;
     }
     sendJson(res, 200, {
@@ -210,7 +234,7 @@ const server = createServer(async (req, res) => {
       choices: [
         {
           index: 0,
-          message: { role: "assistant", content: chatContent },
+          message: { role: "assistant", content },
           finish_reason: "stop",
         },
       ],

test/e2e/support/inference-adapter.test.ts

@@ -180,6 +180,23 @@ describe("E2E inference adapter", () => {
     );
   });
 
+  it("returns the unique launch reply requested by the mock prompt (#9046)", async () => {
+    const adapter = await createAdapter({ env: {} });
+    const expectedReply = "NEMOCLAW_0123456789AB_FIRST_OK";
+    const prompt =
+      "Join these four fragments with underscores and put only the result on its own line: " +
+      "NEMOCLAW, 0123456789AB, FIRST, OK. Do not use tools.";
+
+    expect(await adapter.directChat(prompt)).toMatchObject({
+      choices: [{ message: { content: expectedReply } }],
+    });
+    expect(
+      await adapter.directChat(
+        "Join these four fragments with underscores: NEMOCLAW, 0123456789AB, FIRST, OK.",
+      ),
+    ).toMatchObject({ choices: [{ message: { content: "PONG" } }] });
+  });
+
   it("keeps unrelated ambient secrets out of adapter and fake-server child environments", async () => {
     const secretName = "UNRELATED_E2E_SENTINEL_SECRET";
     const secretValue = "sentinel-value-that-must-not-propagate";