tool call in-flight resume

server/src/llm/streaming.ts

@@ -58,7 +58,6 @@ export async function* runMultiplexStream(req: MultiplexRequest): AsyncGenerator
   let text = "";
   let usage: StreamEvent extends any ? any : never;
   let raw: unknown = { streamed: true };
-  let toolMessages: ReturnType<typeof buildToolLogMessageData>[] = [];
 
   try {
     if (req.provider === "openai" || req.provider === "xai") {
@@ -97,7 +96,16 @@ export async function* runMultiplexStream(req: MultiplexRequest): AsyncGenerator
         }
 
         if (ev.type === "tool_call") {
-          toolMessages.push(buildToolLogMessageData(chatId, ev.event));
+          const toolMessage = buildToolLogMessageData(chatId, ev.event);
+          await prisma.message.create({
+            data: {
+              chatId: toolMessage.chatId,
+              role: toolMessage.role as any,
+              content: toolMessage.content,
+              name: toolMessage.name,
+              metadata: toolMessage.metadata as any,
+            },
+          });
           yield { type: "tool_call", event: ev.event };
           continue;
         }
@@ -149,17 +157,6 @@ export async function* runMultiplexStream(req: MultiplexRequest): AsyncGenerator
     const latencyMs = Math.round(performance.now() - t0);
 
     await prisma.$transaction(async (tx) => {
-      if (toolMessages.length) {
-        await tx.message.createMany({
-          data: toolMessages.map((message) => ({
-            chatId: message.chatId,
-            role: message.role as any,
-            content: message.content,
-            name: message.name,
-            metadata: message.metadata as any,
-          })),
-        });
-      }
       await tx.message.create({
         data: { chatId, role: "assistant" as any, content: text },
       });