fix(providers): xAI translators drop legacy function_call and zero total_tokens (#12692, #12700) (#13753)

Merged in the 2026-09-16 sweep of the maintainer's own open PRs, at the owner's explicit instruction. No push was made to the PR branch: the merge took the head as the owning session left it (verified OPEN, non-draft and MERGEABLE against the release tip immediately before merging).

changelog.d/fixes/12692-xai-legacy-function-call.md

@@ -0,0 +1 @@
+- **fix(providers):** xAI requests no longer silently drop an assistant tool call sent in the legacy OpenAI `function_call` shape (instead of `tool_calls[]`) — the call is now translated into the xAI request the same way modern tool calls are (#12692) — thanks @soroush5

changelog.d/fixes/12700-xai-usage-total.md

@@ -0,0 +1 @@
+- **fix(providers):** xAI responses no longer report `total_tokens`/`totalTokenCount` as `0` when upstream usage uses the legacy `prompt_tokens`/`completion_tokens` names instead of `input_tokens`/`output_tokens` (#12700) — thanks @soroush5

src/lib/providers/xai/translators/gemini.ts

@@ -263,7 +263,7 @@ function toolsGeminiToXai(tools: GeminiTool[]): XaiTool[] | undefined {
  */
 export function geminiRequestToXaiResponses(
   req: GeminiRequest,
-  model: string | null = null,
+  model: string | null = null
 ): XaiResponsesRequest {
   if (!req || typeof req !== "object") return req as unknown as XaiResponsesRequest;
   const input: XaiInputItem[] = [];
@@ -275,9 +275,7 @@ export function geminiRequestToXaiResponses(
     if (fnItems.length) {
       for (const it of fnItems) input.push(it);
       // Filter remaining text/image parts
-      const remaining = (c.parts ?? []).filter(
-        (p) => !p?.functionCall && !p?.functionResponse,
-      );
+      const remaining = (c.parts ?? []).filter((p) => !p?.functionCall && !p?.functionResponse);
       if (remaining.length) input.push({ role, content: partsToXaiBlocks(remaining) });
     } else {
       input.push({ role, content: partsToXaiBlocks(c.parts ?? []) });
@@ -324,7 +322,7 @@ export function geminiRequestToXaiResponses(
  */
 export function xaiCompletedToGeminiJson(
   completed: XaiCompleted,
-  origReq: GeminiRequest | null = null,
+  origReq: GeminiRequest | null = null
 ): object {
   const parts: unknown[] = [];
   const finishReason = "STOP";
@@ -364,7 +362,8 @@ export function xaiCompletedToGeminiJson(
       promptTokenCount: u.input_tokens ?? u.prompt_tokens ?? 0,
       candidatesTokenCount: u.output_tokens ?? u.completion_tokens ?? 0,
       totalTokenCount:
-        u.total_tokens ?? ((u.input_tokens ?? 0) + (u.output_tokens ?? 0)),
+        u.total_tokens ??
+        (u.input_tokens ?? u.prompt_tokens ?? 0) + (u.output_tokens ?? u.completion_tokens ?? 0),
     };
   }
   return out;

src/lib/providers/xai/translators/openai-chat.ts

@@ -38,6 +38,7 @@ interface OpenAiMessage {
   content?: MessageContent;
   tool_calls?: OpenAiToolCall[];
   tool_call_id?: string;
+  function_call?: { name?: string; arguments?: string };
 }
 
 interface OpenAiChatRequest {
@@ -213,6 +214,22 @@ export function chatRequestToXaiResponses(req: OpenAiChatRequest): XaiResponsesR
       }
       continue;
     }
+    // Legacy OpenAI Chat Completions form: assistant tool-call carried as a top-level
+    // `function_call` field instead of `tool_calls[]`. Some OpenAI-compatible clients still
+    // emit this shape; without this branch the message falls through to the generic case
+    // below with empty content and the tool invocation is silently dropped (#12692).
+    if (m.role === "assistant" && m.function_call?.name) {
+      if (m.content) {
+        input.push({ role: "assistant", content: messageContentToXaiBlocks(m.content) });
+      }
+      input.push({
+        type: "function_call",
+        call_id: genId("call"),
+        name: m.function_call.name,
+        arguments: m.function_call.arguments ?? "",
+      });
+      continue;
+    }
     input.push({ role: m.role ?? "user", content: messageContentToXaiBlocks(m.content ?? "") });
   }
 
@@ -303,7 +320,9 @@ export function xaiCompletedToChatJson(
     out.usage = {
       prompt_tokens: u.input_tokens ?? u.prompt_tokens ?? 0,
       completion_tokens: u.output_tokens ?? u.completion_tokens ?? 0,
-      total_tokens: u.total_tokens ?? (u.input_tokens ?? 0) + (u.output_tokens ?? 0),
+      total_tokens:
+        u.total_tokens ??
+        (u.input_tokens ?? u.prompt_tokens ?? 0) + (u.output_tokens ?? u.completion_tokens ?? 0),
     };
   }
   return out;

tests/unit/xai-translators.test.ts

@@ -193,6 +193,51 @@ test("chatRequestToXaiResponses: maps max_tokens to max_output_tokens", () => {
   assert.equal(out.max_output_tokens, 512);
 });
 
+test("#12692: chatRequestToXaiResponses maps legacy assistant function_call to a function_call item", () => {
+  const req = {
+    model: "grok-4",
+    messages: [
+      {
+        role: "assistant",
+        content: null,
+        function_call: { name: "get_weather", arguments: '{"city":"Paris"}' },
+      },
+    ],
+  };
+  const out = chatRequestToXaiResponses(req);
+  const calls = (out.input as Array<{ type: string; name?: string; arguments?: string }>).filter(
+    (i) => i.type === "function_call"
+  );
+  assert.equal(calls.length, 1, "expected a function_call item to be present in xAI input");
+  assert.equal(calls[0]?.name, "get_weather");
+  assert.equal(calls[0]?.arguments, '{"city":"Paris"}');
+});
+
+test("#12692: chatRequestToXaiResponses preserves leading text alongside legacy function_call", () => {
+  const req = {
+    model: "grok-4",
+    messages: [
+      {
+        role: "assistant",
+        content: "Let me check that for you.",
+        function_call: { name: "get_weather", arguments: '{"city":"Paris"}' },
+      },
+    ],
+  };
+  const out = chatRequestToXaiResponses(req);
+  const items = out.input as Array<{
+    type?: string;
+    role?: string;
+    content?: unknown;
+    name?: string;
+  }>;
+  const textItem = items.find((i) => i.role === "assistant");
+  assert.ok(textItem, "expected the leading assistant text block to be preserved");
+  const calls = items.filter((i) => i.type === "function_call");
+  assert.equal(calls.length, 1);
+  assert.equal(calls[0]?.name, "get_weather");
+});
+
 // ─── xaiCompletedToChatJson ──────────────────────────────────────────────────
 
 test("xaiCompletedToChatJson: extracts output_text content into message", () => {
@@ -237,6 +282,21 @@ test("xaiCompletedToChatJson: maps function_call to tool_calls with finish_reaso
   assert.equal(fn.name, "get_weather");
 });
 
+test("#12700: xaiCompletedToChatJson sums legacy prompt_tokens/completion_tokens into total_tokens", () => {
+  const completed = {
+    output: [{ type: "message", content: [{ type: "output_text", text: "hi" }] }],
+    usage: { prompt_tokens: 10, completion_tokens: 5 },
+  };
+  const result = xaiCompletedToChatJson(completed) as { usage?: Record<string, unknown> };
+  assert.equal(result.usage?.prompt_tokens, 10);
+  assert.equal(result.usage?.completion_tokens, 5);
+  assert.equal(
+    result.usage?.total_tokens,
+    15,
+    "total_tokens should sum legacy fields, not report 0"
+  );
+});
+
 // ─── openaiResponsesRequestToXai ─────────────────────────────────────────────
 
 test("openaiResponsesRequestToXai: drops service_tier", () => {
@@ -521,3 +581,19 @@ test("xaiCompletedToGeminiJson: maps usage to usageMetadata", () => {
   assert.equal(meta.candidatesTokenCount, 20);
   assert.equal(meta.totalTokenCount, 30);
 });
+
+test("#12700: xaiCompletedToGeminiJson sums legacy prompt_tokens/completion_tokens into totalTokenCount", () => {
+  const completed = {
+    model: "grok-4",
+    output: [],
+    usage: { prompt_tokens: 10, completion_tokens: 5 },
+  };
+  const result = xaiCompletedToGeminiJson(completed) as { usageMetadata?: Record<string, unknown> };
+  assert.equal(result.usageMetadata?.promptTokenCount, 10);
+  assert.equal(result.usageMetadata?.candidatesTokenCount, 5);
+  assert.equal(
+    result.usageMetadata?.totalTokenCount,
+    15,
+    "totalTokenCount should sum legacy fields, not report 0"
+  );
+});