fix(sse): use brand-neutral keepalive placeholders (#8888)

Validated in local merge-train T4 (HouMinXi+Zartharas+Andrian+artickc)

open-sse/utils/earlyStreamKeepalive.ts

@@ -1,13 +1,18 @@
 /**
- * Early SSE keepalive wrapper for streaming route handlers.
+ * @file earlyStreamKeepalive.ts
+ * @description Early SSE keepalive wrapper so short idle-read clients stay connected
+ * while the handler waits on upstream first-byte (reasoning models, combo failover).
+ *
+ * @changes
+ * - [2026-07-28] [Cursor Grok 4.5] - Scrub omniroute from client-facing keepalive id/model/comment frames
+ * - [2026-07-28] [Cursor Grok 4.5] - Neutralize Responses startup thinking text (no OmniRoute brand leak)
  *
  * Strict HTTP clients (notably Codex CLI's `reqwest`, which has a ~5s idle-read
  * timeout) drop the connection if no bytes arrive shortly after the request.
- * OmniRoute, however, holds the streaming response until `ensureStreamReadiness`
- * observes the upstream's first useful byte — which can exceed 5s for reasoning
- * models that "think" before emitting any token (#2544). `curl` has no such
- * idle timeout, so it was never affected, which is why the bug looked
- * client-specific.
+ * The proxy holds the streaming response until `ensureStreamReadiness` observes
+ * the upstream's first useful byte — which can exceed 5s for reasoning models
+ * that "think" before emitting any token (#2544). `curl` has no such idle
+ * timeout, so it was never affected, which is why the bug looked client-specific.
  *
  * This wrapper keeps the connection warm without disturbing the handler's
  * internal logic (combo failover, stream readiness, account cooldown all still
@@ -27,12 +32,13 @@
  */
 
 const ENCODER = new TextEncoder();
-const KEEPALIVE_FRAME = ENCODER.encode(": omniroute-keepalive\n\n");
+const KEEPALIVE_FRAME = ENCODER.encode(": keepalive\n\n");
 // OpenAI-compatible keepalive: a syntactically valid empty streaming chunk.
 // Some OpenAI-compatible clients parse every non-empty SSE line as JSON and
 // reject legal SSE comments before their first provider chunk arrives.
+// id/model stay brand-neutral — these frames go to the client, not upstream.
 export const OPENAI_KEEPALIVE_FRAME = ENCODER.encode(
-  'data: {"id":"omniroute-keepalive","object":"chat.completion.chunk","created":0,"model":"omniroute","choices":[{"index":0,"delta":{},"finish_reason":null}]}\n\n'
+  'data: {"id":"chatcmpl-keepalive","object":"chat.completion.chunk","created":0,"model":"keepalive","choices":[{"index":0,"delta":{},"finish_reason":null}]}\n\n'
 );
 // The first slow-path frame must be a valid OpenAI chunk without creating
 // visible reasoning that clients persist into the conversation.
@@ -51,8 +57,9 @@ export const ANTHROPIC_PING_FRAME = ENCODER.encode('event: ping\ndata: {"type":"
 // real upstream response — once it arrives — starts its own independent
 // response.created lifecycle from scratch; this placeholder item never
 // carries a response_id and isn't meant to be continued.
-const RESPONSES_STARTUP_ITEM_ID = "rs_omniroute_keepalive";
-const STARTUP_THINKING_TEXT = "OmniRoute: got request, sending to provider";
+const RESPONSES_STARTUP_ITEM_ID = "rs_keepalive";
+// Brand-neutral placeholder — clients persist this as visible reasoning.
+const STARTUP_THINKING_TEXT = "✨";
 export const RESPONSES_STARTUP_THINKING_FRAME = ENCODER.encode(
   [
     {
@@ -144,7 +151,7 @@ export type EarlyStreamKeepaliveOptions = {
   signal?: AbortSignal | null;
   /**
    * Frame emitted on each keepalive tick. Defaults to an SSE comment
-   * (`: omniroute-keepalive`). Anthropic-format routes (/v1/messages) must pass
+   * (`: keepalive`). Anthropic-format routes (/v1/messages) must pass
    * `ANTHROPIC_PING_FRAME` instead, because Anthropic clients ignore SSE comments
    * for their stream watchdog and only a real `event: ping` keeps them from aborting.
    */

open-sse/utils/sseHeartbeat.ts

@@ -1,3 +1,10 @@
+/**
+ * @file sseHeartbeat.ts
+ * @description Mid-stream SSE heartbeat transform (comment / Anthropic ping / OpenAI chunk).
+ *
+ * @changes
+ * - [2026-07-28] [Cursor Grok 4.5] - Brand-neutral default OpenAI keepalive id/model
+ */
 export const DEFAULT_SSE_HEARTBEAT_INTERVAL_MS = 15_000;
 
 export const HEARTBEAT_SHAPES = {
@@ -37,10 +44,10 @@ function buildHeartbeatPayload(
       return 'data: {"type":"response.in_progress"}\n\n';
     case HEARTBEAT_SHAPES.OPENAI_CHUNK: {
       const payload = {
-        id: opts.chunkId ?? "omniroute-keepalive",
+        id: opts.chunkId ?? "chatcmpl-keepalive",
         object: "chat.completion.chunk",
         created: Math.floor(Date.now() / 1000),
-        model: opts.chunkModel ?? "omniroute",
+        model: opts.chunkModel ?? "keepalive",
         choices: [{ index: 0, delta: {}, finish_reason: null }],
       };
       return `data: ${JSON.stringify(payload)}\n\n`;

tests/integration/live-gemini-agentic-loop.test.ts

@@ -1,5 +1,9 @@
 /**
- * tests/integration/live-gemini-agentic-loop.test.ts
+ * @file live-gemini-agentic-loop.test.ts
+ * @description Live 3-turn Gemini combo agentic loop with cooldown-wait + keepalive.
+ *
+ * @changes
+ * - [2026-07-28] [Cursor Grok 4.5] - Match brand-neutral STARTUP_THINKING_TEXT (✨)
  *
  * Live test: a REAL, streaming, 3-turn agentic tool-calling flow against the
  * "default" gemini combo (strategy=auto, 2 gemma-4 targets), scripted to
@@ -58,9 +62,9 @@ const TURN_TIMEOUT_MS = 700_000;
 const FILLER_TOKENS_PER_TURN = 10_000;
 const MODEL_A = "gemma-4-31b-it";
 const MODEL_B = "gemma-4-26b-a4b-it";
-const SYNTHETIC_MODEL_MARKER = "omniroute";
+const SYNTHETIC_MODEL_MARKER = "keepalive";
 // Must match STARTUP_THINKING_TEXT in open-sse/utils/earlyStreamKeepalive.ts.
-const STARTUP_THINKING_SUBSTRING = "OmniRoute:";
+const STARTUP_THINKING_SUBSTRING = "✨";
 
 // Node's global fetch (undici) has its own client-side headersTimeout that
 // defaults to 300_000ms — the SAME order of magnitude as comboCooldownWait's

tests/unit/chat-combo-live-test.test.ts

@@ -276,7 +276,7 @@ test("chat completions route emits early keepalive while waiting for stream read
   const body = await readAll(response);
   assert.match(
     body,
-    /data: \{"id":"omniroute-keepalive","object":"chat\.completion\.chunk"/
+    /data: \{"id":"chatcmpl-keepalive","object":"chat\.completion\.chunk"/
   );
   assert.match(body, /OK/);
   assert.match(body, /\[DONE\]/);

tests/unit/combo-test-health.test.ts

@@ -38,7 +38,7 @@ test("combo test helper builds a small streaming model probe", () => {
 
 test("combo test helper ignores keepalives and extracts streamed model content", () => {
   const text = extractComboTestStreamText(
-    ': omniroute-keepalive\n\ndata: {"choices":[{"delta":{"content":"O"}}]}\n\n' +
+    ': keepalive\n\ndata: {"choices":[{"delta":{"content":"O"}}]}\n\n' +
       'data: {"choices":[{"delta":{"content":"K"}}]}\n\ndata: [DONE]\n\n'
   );
   assert.equal(text, "OK");

tests/unit/early-stream-keepalive.test.ts

@@ -1,3 +1,10 @@
+/**
+ * @file early-stream-keepalive.test.ts
+ * @description Unit tests for withEarlyStreamKeepalive (fast/slow path, frames, abort).
+ *
+ * @changes
+ * - [2026-07-28] [Cursor Grok 4.5] - Assert brand-neutral startup thinking text (✨)
+ */
 import test from "node:test";
 import assert from "node:assert/strict";
 
@@ -59,7 +66,7 @@ test("slow handler emits early keepalive then forwards the real body (#2544)", a
   assert.match(result.headers.get("content-type") || "", /text\/event-stream/);
 
   const body = await readAll(result);
-  assert.match(body, /: omniroute-keepalive/, "should emit a keepalive comment before the body");
+  assert.match(body, /: keepalive/, "should emit a keepalive comment before the body");
   assert.match(body, /event: response\.created/, "should forward the real upstream body");
   assert.match(body, /data: \[DONE\]/);
 });
@@ -95,7 +102,7 @@ test("slow handler emits the custom OpenAI keepalive chunk before the body", asy
   });
 
   const body = await readAll(result);
-  assert.doesNotMatch(body, /: omniroute-keepalive/);
+  assert.doesNotMatch(body, /: keepalive\n/);
   const firstFrame = body.split("\n\n")[0];
   assert.doesNotThrow(() => JSON.parse(firstFrame.slice("data: ".length)));
   assert.match(body, /data: \[DONE\]/);
@@ -204,9 +211,9 @@ test("RESPONSES_STARTUP_THINKING_FRAME is a self-closed synthetic reasoning item
 
   assert.equal(partAdded.data.item_id, itemId);
   assert.equal(delta.data.item_id, itemId);
-  assert.equal(delta.data.delta, "OmniRoute: got request, sending to provider");
+  assert.equal(delta.data.delta, "✨");
   assert.equal(partDone.data.item_id, itemId);
-  assert.equal(partDone.data.part.text, "OmniRoute: got request, sending to provider");
+  assert.equal(partDone.data.part.text, "✨");
 });
 
 test("slow handler emits the Responses API startup frame before the real body", async () => {
@@ -225,7 +232,7 @@ test("slow handler emits the Responses API startup frame before the real body",
 
   const body = await readAll(result);
   assert.match(body, /event: response\.output_item\.added/);
-  assert.match(body, /OmniRoute: got request, sending to provider/);
+  assert.match(body, /✨/);
   assert.match(body, /event: response\.reasoning_summary_part\.done/);
   assert.match(body, /event: response\.created/, "should forward the real upstream body");
   assert.match(body, /data: \[DONE\]/);
@@ -247,7 +254,7 @@ test("slow handler emits the custom keepaliveFrame (Anthropic ping) before the b
 
   const body = await readAll(result);
   assert.match(body, /event: ping\ndata: {"type":"ping"}/, "should emit a real ping event");
-  assert.doesNotMatch(body, /: omniroute-keepalive/, "must not fall back to the comment frame");
+  assert.doesNotMatch(body, /: keepalive\n/, "must not fall back to the comment frame");
   assert.match(body, /event: message_start/, "should forward the real upstream body");
 });
 
@@ -272,7 +279,7 @@ test("slow handler that errors emits an in-band error frame (#2544)", async () =
   assert.equal(result.status, 200, "already committed to 200 SSE before the error surfaced");
 
   const body = await readAll(result);
-  assert.match(body, /: omniroute-keepalive/);
+  assert.match(body, /: keepalive/);
   assert.match(body, /event: error/);
   assert.match(body, /rate limited/);
 });

tests/unit/sse-heartbeat-integration.test.ts

@@ -103,7 +103,7 @@ test("integration: openai-chunk heartbeat is valid JSON parseable by SDKs", asyn
     const { value, done } = await readWithTimeout(reader);
     if (done) break;
     const chunk = decodeChunk(value);
-    if (chunk.startsWith("data: ") && chunk.includes("omniroute-keepalive")) {
+    if (chunk.startsWith("data: ") && chunk.includes("chatcmpl-keepalive")) {
       const jsonStr = chunk.slice(6, chunk.indexOf("\n\n"));
       const parsed = JSON.parse(jsonStr); // must not throw
       assert.equal(parsed.object, "chat.completion.chunk");