fix malformed handling

Author: connortbot

packages/cost/models/providers/base.ts

@@ -55,8 +55,12 @@ export abstract class BaseProvider {
     endpoint: Endpoint,
     context: RequestBodyContext
   ): string | Promise<string> {
+    let updatedBody = context.parsedBody;
+    if (context.bodyMapping === "RESPONSES") {
+      updatedBody = context.toChatCompletions(updatedBody);
+    }
     return JSON.stringify({
-      ...context.parsedBody,
+      ...updatedBody,
       model: endpoint.providerModelId,
     });
   }

packages/cost/models/providers/index.ts

@@ -40,7 +40,34 @@ export const providers = {
 export type ModelProviderName = keyof typeof providers;
 
 // TODO: temporarily whitelist responses API providers until all mappings are done
-export const ResponsesAPIEnabledProviders: ModelProviderName[] = ["openai", "helicone", "vertex"];
+export const ResponsesAPIEnabledProviders: ModelProviderName[] = [
+  "openai",
+  "helicone",
+
+  // chat completions only
+  "azure",
+  "chutes",
+  "cohere",
+  "deepinfra",
+  "deepseek",
+  
+  // has known issues with returning structured JSONS
+  // should be okay to enable, but its not stable enough to add without request
+  // "google-ai-studio",
+  
+  "groq",
+  "nebius",
+  "novita",
+  "openrouter",
+  "perplexity",
+  "xai",
+
+  // anthropic and chat completions provider
+  "vertex"
+
+  // anthropic only
+  // none right now, need anthropic mapper
+];
 
 // Re-export base for extending
 export { BaseProvider } from "./base";

packages/llm-mapper/transform/providers/responses/streamedResponse/toResponses.ts

@@ -19,6 +19,7 @@ export class ChatToResponsesStreamConverter {
   private itemAdded: boolean = false;
   private partAdded: boolean = false;
   private emittedFunctionItems: Set<string> = new Set();
+  private completedEmitted: boolean = false;
 
   constructor() {
     this.toolCalls = new Map();
@@ -139,28 +140,27 @@ export class ChatToResponsesStreamConverter {
         }
       }
 
-      // if finish reason was sent for this choice, emit text.done
+      // if finish reason was sent for this choice, emit done + completed
       if (choice?.finish_reason) {
-        const doneEvt: ResponseOutputTextDoneEvent = {
-          type: "response.output_text.done",
-          item_id: `msg_${this.responseId}`,
-          output_index: 0,
-          content_index: 0,
-          text: this.textBuffer,
-        };
-        events.push(doneEvt);
-
-        // close part and item
-        if (this.partAdded) {
-          events.push({
-            type: "response.content_part.done",
+        if (this.itemAdded) {
+          const doneEvt: ResponseOutputTextDoneEvent = {
+            type: "response.output_text.done",
             item_id: `msg_${this.responseId}`,
             output_index: 0,
             content_index: 0,
-            part: { type: "output_text", text: this.textBuffer, annotations: [] },
-          } as any);
-        }
-        if (this.itemAdded) {
+            text: this.textBuffer,
+          };
+          events.push(doneEvt);
+
+          if (this.partAdded) {
+            events.push({
+              type: "response.content_part.done",
+              item_id: `msg_${this.responseId}`,
+              output_index: 0,
+              content_index: 0,
+              part: { type: "output_text", text: this.textBuffer, annotations: [] },
+            } as any);
+          }
           events.push({
             type: "response.output_item.done",
             output_index: 0,
@@ -178,14 +178,12 @@ export class ChatToResponsesStreamConverter {
 
         // Finalize any function calls
         this.toolCalls.forEach((tc) => {
-          // Done event with full arguments
           events.push({
             type: "response.function_call_arguments.done",
             item_id: tc.item_id,
             output_index: 0,
             arguments: tc.arguments || "{}",
           } as any);
-          // Output item done
           events.push({
             type: "response.output_item.done",
             output_index: 0,
@@ -200,11 +198,50 @@ export class ChatToResponsesStreamConverter {
             },
           } as any);
         });
+
+        // Emit completed now if usage not provided
+        const usage = this.finalUsage || undefined;
+        const completed: ResponseCompletedEvent = {
+          type: "response.completed",
+          response: {
+            id: this.responseId,
+            object: "response",
+            created: this.created,
+            created_at: this.created as any,
+            status: "completed" as any,
+            model: this.model,
+            output: [
+              ...(this.itemAdded
+                ? ([
+                    {
+                      type: "message" as const,
+                      role: "assistant" as const,
+                      content: [
+                        { type: "output_text" as const, text: this.textBuffer },
+                      ],
+                    },
+                  ] as any)
+                : []),
+              ...Array.from(this.toolCalls.values()).map((tc) => ({
+                id: `fc_${tc.id}`,
+                type: "function_call" as const,
+                status: "completed" as const,
+                name: tc.name || "",
+                call_id: tc.id,
+                arguments: tc.arguments || "{}",
+                parsed_arguments: null,
+              })),
+            ],
+            ...(usage ? { usage } : {}),
+          },
+        };
+        events.push(completed);
+        this.completedEmitted = true;
       }
     }
 
-    // usage usagealley (haha...?) arrives in the final chunk with empty choices
-    if (c.usage) {
+    // usage arrives in the final chunk with empty choices
+    if (c.usage && !this.completedEmitted) {
       const usage: ResponsesUsage = {
         input_tokens: c.usage.prompt_tokens,
         output_tokens: c.usage.completion_tokens,
@@ -253,6 +290,7 @@ export class ChatToResponsesStreamConverter {
         },
       };
       events.push(completed);
+      this.completedEmitted = true;
     }
 
     return events;

worker/src/lib/clients/llmmapper/router/oaiChat2responses/stream.ts

@@ -37,18 +37,12 @@ export function oaiChat2responsesStream(
             if (line.startsWith("data: ")) {
               const data = line.slice(6);
               if (data.trim() === "[DONE]") {
-                // Ensure stream termination; Responses API may end with response.completed
-                if (!emittedCompleted) {
-                  // If the converter hasn't emitted a response.completed (e.g., no usage)
-                  // we just end the stream. Clients also listen for "event: done".
-                  controller.enqueue(encoder.encode("event: done\n\n"));
-                }
+                // end of upstream stream, so response.completed should have been emitted by converter.
                 continue;
               }
 
               try {
                 const chunk = JSON.parse(data) as ChatCompletionChunk;
-                console.log("chunk", JSON.stringify(chunk, null, 2));
                 const events = converter.convert(chunk);
                 for (const ev of events) {
                   const type = (ev as any).type;