fix(ai): ignore Vertex keepalives sent as SSE data (#50954)

packages/ai/src/protocols/shared.ts

@@ -253,9 +253,9 @@ export const errorText = (error: unknown) => {
 
 /**
  * `framing` step for Server-Sent Events. Decodes UTF-8, runs the SSE channel
- * decoder, optionally filters named events, and drops empty and bare `null`
- * events. `[DONE]` is dropped by default or retained for protocols that use it
- * as their stream boundary. Retry control events are ignored without
+ * decoder, optionally filters named events, and drops empty events and known
+ * keepalives that proxies send as data. `[DONE]` is dropped by default or
+ * retained for protocols that use it as their stream boundary. Retry control events are ignored without
  * interrupting the stream. Decoder failures become provider output errors so
  * the public error channel stays `AIError`.
  */
@@ -291,6 +291,9 @@ export const sseFraming = (
         // `data: null`, between events or after `[DONE]`. No protocol has a
         // null event, so it carries nothing and must not abort the stream.
         event.data !== "null" &&
+        // Vertex AI partner models (e.g. `xai/grok-4.6`) send their SSE
+        // keepalive comment as `data: : keepalive` while reasoning.
+        event.data !== ": keepalive" &&
         (event.data !== "[DONE]" || includeDone || (events !== undefined && event.event !== "message")),
     ),
     Stream.map((event) => event.data),

packages/ai/test/provider/google-vertex.test.ts

@@ -343,6 +343,34 @@ describe("Google Vertex providers", () => {
     }),
   )
 
+  // Captured from xai/grok-4.6 on Vertex: one keepalive every 15s until the first token.
+  it.effect("ignores keepalives sent as data while a partner model reasons", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(
+        LLM.request({
+          model: GoogleVertexChat.configure({
+            accessToken: "vertex-token",
+            location: "global",
+            project: "vertex-project",
+          }).model("xai/grok-4.6"),
+          prompt: "Say hello.",
+        }),
+      ).pipe(
+        Effect.provide(
+          fixedResponse(
+            `data: : keepalive\n\ndata: : keepalive\n\n${sseEvents(
+              deltaChunk({ role: "assistant", content: "Hello." }),
+              finishChunk("stop"),
+            )}`,
+          ),
+        ),
+      )
+
+      expect(response.text).toBe("Hello.")
+      expect(response.finishReason).toEqual({ normalized: "stop", raw: "stop" })
+    }),
+  )
+
   it.effect("sends Grok requests through Vertex Responses", () =>
     Effect.gen(function* () {
       const response = yield* LLMClient.generate(

packages/ai/test/schema.test.ts

@@ -192,6 +192,20 @@ describe("AI.Usage", () => {
     }),
   )
 
+  it.effect("sseFraming drops keepalive comments sent as data and keeps other payloads", () =>
+    Effect.gen(function* () {
+      const frames = yield* ProviderShared.sseFraming(
+        Stream.make(
+          new TextEncoder().encode(
+            'data: {"first":true}\n\ndata: : keepalive\n\n: keepalive\n\ndata: : ping\n\ndata: {"second":true}\n\n',
+          ),
+        ),
+      ).pipe(Stream.runCollect)
+
+      expect(Array.from(frames)).toEqual(['{"first":true}', ": ping", '{"second":true}'])
+    }),
+  )
+
   test("visibleOutputTokens clamps reasoning > output to zero", () => {
     expect(new Usage({ outputTokens: 10, reasoningTokens: 4 }).visibleOutputTokens).toBe(6)
     expect(new Usage({ outputTokens: 10 }).visibleOutputTokens).toBe(10)