Merge pull request #909 from christopher-s/fix/responses-api-flush-total-tokens

fix(translator): emit response.completed with total_tokens for Responses API clients

open-sse/translator/index.ts

@@ -267,6 +267,15 @@ export function translateResponse(targetFormat, sourceFormat, chunk, state) {
           finalResults.push(...(Array.isArray(converted) ? converted : [converted]));
         }
       }
+      // Flush: pass null to source-format translator even when Step 1 produced no output.
+      // This is critical for formats like openai-responses that emit terminal events
+      // (e.g., response.completed with total_tokens) in their flush handler.
+      if (chunk === null && results.length === 0) {
+        const converted = fromOpenAI(null, state);
+        if (converted) {
+          finalResults.push(...(Array.isArray(converted) ? converted : [converted]));
+        }
+      }
       results = finalResults;
     }
   }

open-sse/translator/response/openai-responses.ts

@@ -14,11 +14,24 @@ export function openaiToOpenAIResponsesResponse(chunk, state) {
     return flushEvents(state);
   }
 
-  if (!chunk.choices?.length) {
-    // Capture usage from usage-only chunks (stream_options.include_usage)
-    if (chunk.usage) {
-      state.usage = chunk.usage;
+  // Capture usage from any chunk that carries it (usage-only chunks OR final chunks with finish_reason)
+  // Normalize Chat Completions format (prompt_tokens/completion_tokens) to Responses API format
+  // (input_tokens/output_tokens) so response.completed always has the fields Codex expects.
+  if (chunk.usage) {
+    const u = chunk.usage;
+    const input_tokens = u.input_tokens ?? u.prompt_tokens ?? 0;
+    const output_tokens = u.output_tokens ?? u.completion_tokens ?? 0;
+    state.usage = {
+      input_tokens,
+      output_tokens,
+      total_tokens: u.total_tokens ?? input_tokens + output_tokens,
+    };
+    if (u.prompt_tokens_details?.cached_tokens) {
+      state.usage.input_tokens_details = { cached_tokens: u.prompt_tokens_details.cached_tokens };
     }
+  }
+
+  if (!chunk.choices?.length) {
     return [];
   }
 

tests/unit/responses-translation-fixes.test.mjs

@@ -369,7 +369,9 @@ test("Chat→Responses streaming: usage-only chunk is captured (not dropped)", (
   const completedEvent = finishEvents.find((e) => e.event === "response.completed");
   assert.ok(completedEvent, "should have completed event");
   assert.ok(completedEvent.data.response.usage, "completed event should include usage");
-  assert.equal(completedEvent.data.response.usage.prompt_tokens, 10);
+  assert.equal(completedEvent.data.response.usage.input_tokens, 10);
+  assert.equal(completedEvent.data.response.usage.output_tokens, 5);
+  assert.equal(completedEvent.data.response.usage.total_tokens, 15);
 });
 
 test("Chat→Responses streaming: completed event includes accumulated output", () => {