fix(sse): move Antigravity client system content to first user message to avoid upstream 429 (#9030)

Closes #9030

changelog.d/fixes/9030-antigravity-system-429s.md

@@ -0,0 +1 @@
+- fix(sse): move Antigravity client system content to first user message to avoid upstream 429 RESOURCE_EXHAUSTED on oversized systemInstruction (#9030)

open-sse/translator/request/openai-to-gemini.ts

@@ -734,11 +734,24 @@ function wrapInCloudCodeEnvelope(model, cloudCodeRequest, credentials = null) {
     envelope._toolNameMap = cloudCodeRequest._toolNameMap;
   }
 
+  // #9030 — Client system content must NOT be combined with default in systemInstruction
+  //
+  // The upstream Antigravity / Cloud Code endpoint rejects oversized systemInstruction
+  // with 429 RESOURCE_EXHAUSTED. Keep only the lightweight ANTIGRAVITY_DEFAULT_SYSTEM
+  // in systemInstruction and relocate any client system content (which can be very
+  // large — Hermes ~125k tokens) to the first user message.
   const defaultPart: GeminiPart = { text: ANTIGRAVITY_DEFAULT_SYSTEM };
-  if (envelope.request.systemInstruction?.parts) {
-    envelope.request.systemInstruction.parts.unshift(defaultPart);
-  } else {
-    envelope.request.systemInstruction = { role: "system", parts: [defaultPart] };
+  const clientParts = envelope.request.systemInstruction?.parts?.slice() ?? [];
+  envelope.request.systemInstruction = { role: "system", parts: [defaultPart] };
+
+  if (clientParts.length > 0) {
+    // Prepend client system parts to the first user message so they still guide
+    // the model's behavior early in the conversation.
+    if (envelope.request.contents && envelope.request.contents.length > 0) {
+      envelope.request.contents[0].parts.unshift(...clientParts);
+    } else {
+      envelope.request.contents = [{ role: "user", parts: [...clientParts] }];
+    }
   }
 
   // Strip Gemini built-in tool *names* out of functionDeclarations: Antigravity's

tests/unit/repro-9030-antigravity-system-429s.test.ts

@@ -0,0 +1,131 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+const { openaiToAntigravityRequest } =
+  await import("../../open-sse/translator/request/openai-to-gemini.ts");
+const { ANTIGRAVITY_DEFAULT_SYSTEM } = await import("../../open-sse/config/constants.ts");
+
+type PartsArray = Array<{ text?: string; functionCall?: unknown; functionResponse?: unknown }>;
+
+/**
+ * #9030 — Hermes + Antigravity system messages cause 429s
+ *
+ * Root cause: openai-to-gemini.ts's wrapInCloudCodeEnvelope prepends
+ * ANTIGRAVITY_DEFAULT_SYSTEM to systemInstruction.parts that already contain
+ * the client's system message. This doubles the systemInstruction size well
+ * beyond what the upstream Antigravity/Google Cloud Code endpoint accepts,
+ * causing a 429 RESOURCE_EXHAUSTED.
+ *
+ * Fix: move client system content to the first user message and keep only the
+ * default system in systemInstruction.
+ *
+ * This repro test asserts that the fix is in place — after translation, the
+ * systemInstruction contains ONLY ANTIGRAVITY_DEFAULT_SYSTEM, and the client's
+ * system text is in contents[0].
+ */
+test("OpenAI -> Antigravity: client system message must NOT be in systemInstruction with default (#9030)", () => {
+  const result = openaiToAntigravityRequest(
+    "gemini-2.5-flash",
+    {
+      messages: [
+        { role: "system", content: "You are a helpful assistant specialized in coding tasks." },
+        { role: "user", content: "Write a function to sort an array." },
+      ],
+    },
+    false
+  );
+
+  // systemInstruction must contain ONLY the ANTIGRAVITY_DEFAULT_SYSTEM
+  assert.ok(result.request.systemInstruction, "systemInstruction must exist");
+  assert.equal(
+    result.request.systemInstruction.parts.length,
+    1,
+    "systemInstruction must have exactly 1 part (only ANTIGRAVITY_DEFAULT_SYSTEM)"
+  );
+  assert.equal(
+    result.request.systemInstruction.parts[0].text,
+    ANTIGRAVITY_DEFAULT_SYSTEM,
+    "systemInstruction must contain only ANTIGRAVITY_DEFAULT_SYSTEM"
+  );
+
+  // The client system message must be moved to the first user content
+  const parts = result.request.contents[0]?.parts as PartsArray | undefined;
+  assert.ok(parts, "contents[0] must exist");
+  const firstUserText = parts.map((p) => p.text ?? "").join("");
+  assert.ok(
+    firstUserText.includes("You are a helpful assistant specialized in coding tasks."),
+    "client system text must be in first user message contents[0]"
+  );
+  assert.ok(
+    firstUserText.includes("Write a function to sort an array."),
+    "user message must also be in contents[0]"
+  );
+});
+
+test("OpenAI -> Antigravity: systemInstruction with only ANTIGRAVITY_DEFAULT_SYSTEM when no client system (#9030)", () => {
+  const result = openaiToAntigravityRequest(
+    "gemini-2.5-flash",
+    {
+      messages: [
+        { role: "user", content: "Hello, how are you?" },
+      ],
+    },
+    false
+  );
+
+  // systemInstruction must still have only the ANTIGRAVITY_DEFAULT_SYSTEM
+  assert.ok(result.request.systemInstruction, "systemInstruction must exist");
+  assert.equal(
+    result.request.systemInstruction.parts.length,
+    1,
+    "systemInstruction must have exactly 1 part"
+  );
+  assert.equal(
+    result.request.systemInstruction.parts[0].text,
+    ANTIGRAVITY_DEFAULT_SYSTEM,
+    "systemInstruction must contain ANTIGRAVITY_DEFAULT_SYSTEM"
+  );
+});
+
+test("OpenAI -> Antigravity: multiple system messages all moved to first user content (#9030)", () => {
+  const result = openaiToAntigravityRequest(
+    "gemini-2.5-flash",
+    {
+      messages: [
+        { role: "system", content: "Rule one: be concise." },
+        { role: "system", content: "Rule two: use TypeScript." },
+        { role: "user", content: "Write code." },
+      ],
+    },
+    false
+  );
+
+  // systemInstruction must contain ONLY ANTIGRAVITY_DEFAULT_SYSTEM
+  assert.equal(
+    result.request.systemInstruction.parts.length,
+    1,
+    "systemInstruction must have exactly 1 part"
+  );
+  assert.equal(
+    result.request.systemInstruction.parts[0].text,
+    ANTIGRAVITY_DEFAULT_SYSTEM,
+    "systemInstruction must contain only ANTIGRAVITY_DEFAULT_SYSTEM"
+  );
+
+  // The combined system messages + user message must be in contents[0]
+  const parts = result.request.contents[0]?.parts as PartsArray | undefined;
+  assert.ok(parts, "contents[0] must exist");
+  const firstUserText = parts.map((p) => p.text ?? "").join("");
+  assert.ok(
+    firstUserText.includes("Rule one: be concise."),
+    "first system text must be in contents[0]"
+  );
+  assert.ok(
+    firstUserText.includes("Rule two: use TypeScript."),
+    "second system text must be in contents[0]"
+  );
+  assert.ok(
+    firstUserText.includes("Write code."),
+    "user message must be in contents[0]"
+  );
+});

tests/unit/translator-openai-to-gemini.test.ts

@@ -863,7 +863,10 @@ test("OpenAI -> Antigravity maps Claude-family models to Gemini-compatible schem
   assert.match(result.requestId, /^agent\/\d+\/[0-9a-f]{8}$/);
   assert.equal(result.enabledCreditTypes, undefined);
   assert.equal(result.request.systemInstruction.parts[0].text, ANTIGRAVITY_DEFAULT_SYSTEM);
-  assert.equal(result.request.systemInstruction.parts[1].text, "Project rules");
+  assert.equal(result.request.systemInstruction.parts.length, 1, "systemInstruction must contain only ANTIGRAVITY_DEFAULT_SYSTEM (#9030)");
+  // #9030 — Client system content moved to first user message to avoid upstream 429s
+  assert.equal(result.request.contents[0].parts[0].text, "Project rules");
+  assert.equal(result.request.contents[0].parts[1].text, "Read a file");
   assert.equal((result as any).request?.generationConfig.maxOutputTokens, undefined);
   assert.equal((result as any).request?.messages, undefined);
   assert.equal((result as any).request?.system, undefined);