fix(sse): convert Gemini body to OpenAI format in antigravity MITM handler

CHANGELOG.md

@@ -22,6 +22,7 @@ _In development — bullets added per PR; finalized at release._
 
 ### 🔧 Bug Fixes
 
+- **fix(sse): convert the native Gemini request body to OpenAI format in the Antigravity MITM handler** — `contents` / `systemInstruction` / `generationConfig` / `thinkingConfig` are now translated to OpenAI chat-completions format before forwarding to `/v1/chat/completions`, so thinking-capable models (e.g. `ag/claude-opus-4-6-thinking`) no longer fail with provider-side 400 "invalid argument" errors. ([#4845](https://github.com/diegosouzapw/OmniRoute/pull/4845) — thanks @anuragg-saxenaa)
 - **fix(db): translate the two pt-BR SQLite driver-fallback log lines to English** — `[DB] Pré-inicializando sql.js WASM…` and `[DB] Drivers síncronos indisponíveis…` were the only non-English server log strings, mixing languages in the logs. Now `[DB] Pre-initializing sql.js WASM (synchronous drivers unavailable)…` / `[DB] Synchronous drivers unavailable — falling back to sql.js (WASM)`, guarded by a test that scans the driver path for accented log strings. ([#5103](https://github.com/diegosouzapw/OmniRoute/issues/5103))
 - **fix(diagnostics): non-streaming Claude responses no longer false-502 as `empty_choices`** — the v3.8.37 malformed-200 detector (#4942) only understood OpenAI `choices` and Responses-API `output` shapes, so a `/v1/messages` response that stays in Claude shape (`{type:"message", content:[…]}`) fell through to `empty_choices` → 502 (cascading to "All models failed" in a combo). Most visibly, an extended-thinking turn whose buffered body is a single **empty thinking block with a valid `signature`** (Claude Code's non-streaming Bash classifier) 502'd on every call. `detectMalformedNonStream` now understands the Claude shape: text/tool_use blocks and thinking blocks carrying a signature count as valid output, while a genuinely empty `content:[]` is still flagged. ([#5108](https://github.com/diegosouzapw/OmniRoute/issues/5108), thanks @insoln)
 - **fix(combo): empty-content 502 now fails over within the same request instead of exhausting the provider** — a leg that answers HTTP 200 with no usable completion is rewritten to `502 "Provider returned empty content"`, but the combo exhaustion classifier treated that synthetic 502 as a connection-level failure (`#1731v2`) and marked the whole provider/connection exhausted, skipping every remaining **same-provider** leg in that request. The connection is actually healthy (it just returned an empty body), so empty-content 502s are now classified as model-level transient failures: the request advances to the next leg and the rest of that provider's legs stay eligible. Genuine gateway 502s still trip connection exhaustion. ([#5085](https://github.com/diegosouzapw/OmniRoute/issues/5085), thanks @andrea-kingautomation)

src/mitm/handlers/antigravity.ts

@@ -1,11 +1,20 @@
 /**
  * Antigravity IDE handler.
  *
- * Preserves the historical behavior of `src/mitm/server.cjs::intercept()`:
- *   - parses the incoming JSON body,
- *   - replaces `body.model` with the mapped model,
- *   - forwards to `/v1/chat/completions` on the OmniRoute router,
- *   - pipes the SSE response back to the IDE.
+ * Antigravity (the Gemini-based IDE) sends requests in native Gemini
+ * GenerateContent format (`contents`, `systemInstruction`, `generationConfig`,
+ * `thinkingConfig`, …). The OmniRoute router endpoint `/v1/chat/completions`
+ * expects OpenAI Chat Completions format, so the raw Gemini body must be
+ * converted before forwarding — otherwise the unknown fields are either
+ * ignored or cause upstream providers to return a 400 "invalid argument"
+ * error (especially with thinking-capable models such as
+ * `ag/claude-opus-4-6-thinking`).
+ *
+ * Pipeline:
+ *   - parse the incoming Gemini JSON body,
+ *   - convert it to an OpenAI chat.completions body (model = mapped model),
+ *   - forward to `/v1/chat/completions` on the OmniRoute router,
+ *   - pipe the SSE response back to the IDE.
  *
  * Non-regressive: any change here must keep the Antigravity flow working as
  * before (see `tests/unit/mitm-handler-antigravity.test.ts`).
@@ -14,6 +23,93 @@ import type { IncomingMessage, ServerResponse } from "node:http";
 import type { AgentId } from "../types";
 import { MitmHandlerBase } from "./base";
 
+interface GeminiPart {
+  text?: string;
+}
+
+interface GeminiContent {
+  role?: string;
+  parts?: GeminiPart[];
+}
+
+interface GeminiGenerationConfig {
+  maxOutputTokens?: number;
+  temperature?: number;
+  topP?: number;
+  stopSequences?: string[];
+}
+
+interface GeminiRequestBody {
+  systemInstruction?: GeminiContent;
+  contents?: GeminiContent[];
+  generationConfig?: GeminiGenerationConfig;
+  [key: string]: unknown;
+}
+
+interface OpenAIChatMessage {
+  role: "system" | "user" | "assistant";
+  content: string;
+}
+
+interface OpenAIChatBody {
+  model: string;
+  messages: OpenAIChatMessage[];
+  stream: boolean;
+  max_tokens?: number;
+  temperature?: number;
+  top_p?: number;
+  stop?: string[];
+}
+
+function joinPartsText(parts: GeminiPart[] | undefined): string {
+  return (parts || [])
+    .map((p) => p.text)
+    .filter((t): t is string => Boolean(t))
+    .join("\n");
+}
+
+/**
+ * Convert a Gemini GenerateContent request body to an OpenAI
+ * chat.completions body.
+ *
+ * @param geminiBody parsed Gemini request
+ * @param model      resolved OmniRoute model string
+ * @param stream     whether the original request was streaming
+ */
+export function convertGeminiToOpenAI(
+  geminiBody: GeminiRequestBody,
+  model: string,
+  stream: boolean,
+): OpenAIChatBody {
+  const messages: OpenAIChatMessage[] = [];
+
+  // System instruction
+  if (geminiBody.systemInstruction) {
+    const systemText = joinPartsText(geminiBody.systemInstruction.parts);
+    if (systemText) messages.push({ role: "system", content: systemText });
+  }
+
+  // Chat turns
+  for (const content of geminiBody.contents || []) {
+    const role: OpenAIChatMessage["role"] = content.role === "model" ? "assistant" : "user";
+    messages.push({ role, content: joinPartsText(content.parts) });
+  }
+
+  const openaiBody: OpenAIChatBody = {
+    model,
+    messages,
+    stream: !!stream,
+  };
+
+  const cfg = geminiBody.generationConfig || {};
+  if (cfg.maxOutputTokens != null) openaiBody.max_tokens = cfg.maxOutputTokens;
+  if (cfg.temperature != null) openaiBody.temperature = cfg.temperature;
+  if (cfg.topP != null) openaiBody.top_p = cfg.topP;
+  if (cfg.stopSequences?.length) openaiBody.stop = cfg.stopSequences;
+
+  return openaiBody;
+}
+
 export class AntigravityHandler extends MitmHandlerBase {
   readonly agentId: AgentId = "antigravity";
 
@@ -27,8 +123,12 @@ export class AntigravityHandler extends MitmHandlerBase {
     const intercepted = await this.hookBufferStart(req, body, mappedModel);
 
     try {
-      const payload = JSON.parse(body.toString());
-      payload.model = mappedModel;
+      const geminiBody = JSON.parse(body.toString()) as GeminiRequestBody;
+
+      // Streaming intent: Antigravity uses :streamGenerateContent for streaming.
+      const isStream = (req.url || "").includes(":streamGenerateContent");
+
+      const payload = convertGeminiToOpenAI(geminiBody, mappedModel, isStream);
 
       const upstreamStart = this.now();
       const upstream = await this.fetchRouter(payload, "/v1/chat/completions", req.headers);

tests/unit/_mitmHandlerHarness.ts

@@ -19,10 +19,13 @@ export interface HarnessResult {
   responseChunks: string[];
 }
 
-function fakeReq(headers: Record<string, string> = {}): IncomingMessage {
+function fakeReq(
+  headers: Record<string, string> = {},
+  url = "/v1/chat/completions"
+): IncomingMessage {
   return {
     method: "POST",
-    url: "/v1/chat/completions",
+    url,
     headers: {
       host: "api.example.com",
       "user-agent": "ut",
@@ -68,10 +71,11 @@ export async function runHandler(
     upstreamStatus?: number;
     upstreamBody?: string;
     headers?: Record<string, string>;
+    url?: string;
   } = {}
 ): Promise<HarnessResult> {
   const { res, out } = fakeRes();
-  const req = fakeReq(opts.headers);
+  const req = fakeReq(opts.headers, opts.url);
   const buf = Buffer.from(typeof body === "string" ? body : JSON.stringify(body));
 
   const originalFetch = globalThis.fetch;

tests/unit/mitm-handler-antigravity.test.ts

@@ -1,6 +1,9 @@
 import test from "node:test";
 import assert from "node:assert/strict";
-import { AntigravityHandler } from "../../src/mitm/handlers/antigravity.ts";
+import {
+  AntigravityHandler,
+  convertGeminiToOpenAI,
+} from "../../src/mitm/handlers/antigravity.ts";
 import { runHandler } from "./_mitmHandlerHarness.ts";
 
 test("antigravity handler — forwards to OmniRoute and pipes SSE", async () => {
@@ -27,3 +30,80 @@ test("antigravity handler — propagates upstream failure as 500", async () => {
   // Error must NOT include raw stack trace (Hard Rule #12 sanitization).
   assert.ok(!body.includes("at /"));
 });
+
+test("convertGeminiToOpenAI — maps Gemini fields to OpenAI chat body", () => {
+  const out = convertGeminiToOpenAI(
+    {
+      systemInstruction: { parts: [{ text: "be brief" }] },
+      contents: [
+        { role: "user", parts: [{ text: "hello" }] },
+        { role: "model", parts: [{ text: "hi there" }] },
+      ],
+      generationConfig: {
+        maxOutputTokens: 256,
+        temperature: 0.4,
+        topP: 0.9,
+        stopSequences: ["STOP"],
+      },
+      // Gemini-only field that must NOT leak into the OpenAI body.
+      thinkingConfig: { thinkingBudget: 1024 },
+    } as Record<string, unknown>,
+    "claude-opus-4-6-thinking",
+    true
+  );
+
+  assert.equal(out.model, "claude-opus-4-6-thinking");
+  assert.equal(out.stream, true);
+  assert.deepEqual(out.messages, [
+    { role: "system", content: "be brief" },
+    { role: "user", content: "hello" },
+    { role: "assistant", content: "hi there" },
+  ]);
+  assert.equal(out.max_tokens, 256);
+  assert.equal(out.temperature, 0.4);
+  assert.equal(out.top_p, 0.9);
+  assert.deepEqual(out.stop, ["STOP"]);
+  // Gemini-native fields must be stripped, not forwarded.
+  assert.equal((out as Record<string, unknown>).contents, undefined);
+  assert.equal((out as Record<string, unknown>).generationConfig, undefined);
+  assert.equal((out as Record<string, unknown>).thinkingConfig, undefined);
+});
+
+test("antigravity handler — converts raw Gemini body before forwarding", async () => {
+  const r = await runHandler(
+    new AntigravityHandler(),
+    {
+      contents: [{ role: "user", parts: [{ text: "ping" }] }],
+      generationConfig: { maxOutputTokens: 64 },
+      thinkingConfig: { thinkingBudget: 512 },
+    },
+    "ag-claude-opus-4-6-thinking",
+    {
+      upstreamBody: "data: pong\n\n",
+      url: "/v1beta/models/gemini:streamGenerateContent",
+    }
+  );
+
+  assert.ok(r.fetchCalled);
+  const forwarded = JSON.parse(r.fetchBody);
+  // The router must receive OpenAI format, not the raw Gemini body.
+  assert.equal(forwarded.model, "ag-claude-opus-4-6-thinking");
+  assert.equal(forwarded.stream, true);
+  assert.deepEqual(forwarded.messages, [{ role: "user", content: "ping" }]);
+  assert.equal(forwarded.max_tokens, 64);
+  // Gemini-native fields that caused upstream 400s must be gone.
+  assert.equal(forwarded.contents, undefined);
+  assert.equal(forwarded.generationConfig, undefined);
+  assert.equal(forwarded.thinkingConfig, undefined);
+});
+
+test("antigravity handler — non-streaming URL yields stream:false", async () => {
+  const r = await runHandler(
+    new AntigravityHandler(),
+    { contents: [{ role: "user", parts: [{ text: "hi" }] }] },
+    "gpt-4o",
+    { url: "/v1beta/models/gemini:generateContent" }
+  );
+  const forwarded = JSON.parse(r.fetchBody);
+  assert.equal(forwarded.stream, false);
+});