fix(ai): merge parallel gemini tool results into one turn

Gemini requires all responses to a parallel function-call batch to
appear as functionResponse parts in a single user turn. Consecutive
tool messages each lowered to their own user turn, producing an
invalid continuation that providers reject.

Consecutive tool results now join the open function-response turn,
matching the existing rule that keeps system updates out of it.

Refs #43478

packages/ai/src/protocols/anthropic-messages.ts

@@ -502,7 +502,16 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
     if (message.role === "system") {
       if (splitsLocalToolResults(request.messages, index))
         return yield* invalid("Anthropic Messages system updates cannot split a local tool call from its tool result")
-      if (supportsNativeSystemUpdates(request) && canUseNativeSystemUpdate(request.messages, index)) {
+      // Vertex currently returns 404 for a terminal system message after local tool results.
+      const vertexToolResultUpdate =
+        request.model.provider === "google-vertex" &&
+        request.messages[index - 1]?.role === "tool" &&
+        request.messages[index + 1] === undefined
+      if (
+        supportsNativeSystemUpdates(request) &&
+        !vertexToolResultUpdate &&
+        canUseNativeSystemUpdate(request.messages, index)
+      ) {
         messages.push(yield* lowerNativeSystemUpdate(message, breakpoints))
         continue
       }

packages/ai/src/protocols/gemini.ts

@@ -289,7 +289,9 @@ const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMR
     if (message.role === "system") {
       const part = yield* ProviderShared.wrappedSystemUpdate("Gemini", message)
       const previous = contents.at(-1)
-      if (previous?.role === "user")
+      // Gemini rejects a continuation whose function-response turn carries extra
+      // parts, so an update after a tool result starts its own user turn.
+      if (previous?.role === "user" && !previous.parts.some((item) => "functionResponse" in item))
         contents[contents.length - 1] = { role: "user", parts: [...previous.parts, { text: part.text }] }
       else contents.push({ role: "user", parts: [{ text: part.text }] })
       continue
@@ -377,7 +379,12 @@ const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMR
         },
       })
     }
-    contents.push({ role: "user", parts })
+    // Gemini requires every response to a parallel call batch in one user turn,
+    // so consecutive tool results join the open function-response turn.
+    const previous = contents.at(-1)
+    if (previous?.role === "user" && previous.parts.some((item) => "functionResponse" in item))
+      contents[contents.length - 1] = { role: "user", parts: [...previous.parts, ...parts] }
+    else contents.push({ role: "user", parts })
   }
 
   return contents

packages/ai/test/provider/anthropic-messages.test.ts

@@ -18,6 +18,14 @@ const opus48 = AnthropicMessages.route
   .with({ endpoint: { baseURL: "https://api.anthropic.test/v1/" }, auth: Auth.header("x-api-key", "test") })
   .model({ id: "claude-opus-4-8" })
 
+const vertexOpus48 = AnthropicMessages.route
+  .with({
+    provider: "google-vertex",
+    endpoint: { baseURL: "https://vertex.test/v1/" },
+    auth: Auth.header("authorization", "Bearer test"),
+  })
+  .model({ id: "claude-opus-4-8" })
+
 const request = LLM.request({
   id: "req_1",
   model,
@@ -277,6 +285,105 @@ describe("Anthropic Messages route", () => {
     }),
   )
 
+  it.effect("keeps a terminal Vertex system update in the tool-result turn", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: vertexOpus48,
+          messages: [
+            Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: {} })]),
+            Message.tool({ id: "call_1", name: "lookup", result: "Done." }),
+            Message.system("Operator update."),
+          ],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [{ type: "tool_use", id: "call_1", name: "lookup", input: {} }],
+        },
+        {
+          role: "user",
+          content: [
+            {
+              type: "tool_result",
+              tool_use_id: "call_1",
+              content: '"Done."',
+              is_error: undefined,
+              cache_control: undefined,
+            },
+            {
+              type: "text",
+              text: "<system-update>\nOperator update.\n</system-update>",
+              cache_control: undefined,
+            },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("keeps a terminal direct Anthropic system update native", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: opus48,
+          messages: [
+            Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: {} })]),
+            Message.tool({ id: "call_1", name: "lookup", result: "Done." }),
+            Message.system("Operator update."),
+          ],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [{ type: "tool_use", id: "call_1", name: "lookup", input: {} }],
+        },
+        {
+          role: "user",
+          content: [
+            {
+              type: "tool_result",
+              tool_use_id: "call_1",
+              content: '"Done."',
+              is_error: undefined,
+              cache_control: undefined,
+            },
+          ],
+        },
+        {
+          role: "system",
+          content: [{ type: "text", text: "Operator update.", cache_control: undefined }],
+        },
+      ])
+    }),
+  )
+
+  it.effect("keeps an ordinary terminal Vertex system update native", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: vertexOpus48,
+          messages: [Message.user("Before."), Message.system("Operator update.")],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        { role: "user", content: [{ type: "text", text: "Before." }] },
+        {
+          role: "system",
+          content: [{ type: "text", text: "Operator update.", cache_control: undefined }],
+        },
+      ])
+    }),
+  )
+
   it.effect("rejects a system update between a local tool call and its result", () =>
     Effect.gen(function* () {
       const error = yield* compileRequest(

packages/ai/test/provider/gemini.test.ts

@@ -139,6 +139,95 @@ describe("Gemini route", () => {
     }),
   )
 
+  it.effect("keeps system updates separate from function responses", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([ToolCallPart.make({ id: "call_1", name: "lookup", input: { query: "weather" } })]),
+            Message.tool({ id: "call_1", name: "lookup", result: "done", resultType: "text" }),
+            Message.system("Update."),
+            Message.system("Later update."),
+          ],
+        }),
+      )
+
+      expect(prepared.body.contents).toEqual([
+        {
+          role: "model",
+          parts: [{ functionCall: { id: undefined, name: "lookup", args: { query: "weather" } } }],
+        },
+        {
+          role: "user",
+          parts: [
+            {
+              functionResponse: {
+                id: undefined,
+                name: "lookup",
+                response: { name: "lookup", content: "done" },
+              },
+            },
+          ],
+        },
+        {
+          role: "user",
+          parts: [
+            { text: "<system-update>\nUpdate.\n</system-update>" },
+            { text: "<system-update>\nLater update.\n</system-update>" },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("merges parallel tool results into one function-response turn", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([
+              ToolCallPart.make({ id: "call_1", name: "lookup", input: { query: "weather" } }),
+              ToolCallPart.make({ id: "call_2", name: "lookup", input: { query: "time" } }),
+            ]),
+            Message.tool({ id: "call_1", name: "lookup", result: "sunny", resultType: "text" }),
+            Message.tool({ id: "call_2", name: "lookup", result: "noon", resultType: "text" }),
+          ],
+        }),
+      )
+
+      expect(prepared.body.contents).toEqual([
+        {
+          role: "model",
+          parts: [
+            { functionCall: { id: undefined, name: "lookup", args: { query: "weather" } } },
+            { functionCall: { id: undefined, name: "lookup", args: { query: "time" } } },
+          ],
+        },
+        {
+          role: "user",
+          parts: [
+            {
+              functionResponse: {
+                id: undefined,
+                name: "lookup",
+                response: { name: "lookup", content: "sunny" },
+              },
+            },
+            {
+              functionResponse: {
+                id: undefined,
+                name: "lookup",
+                response: { name: "lookup", content: "noon" },
+              },
+            },
+          ],
+        },
+      ])
+    }),
+  )
+
   it.effect("prepares multimodal user input and tool history", () =>
     Effect.gen(function* () {
       const prepared = yield* compileRequest(