Accumulate Codex stream output items so tool calls aren't lost

Root cause of the GPT-account "empty turn: no text and no tool calls"
failures, found by replaying the real stored response: the CEO actually
called hire_agent five times, but the ChatGPT/Codex backend delivers each
output item in a response.output_item.done event and then sends a
response.completed whose response.output is []. accumulateResponsesStream
read only the completion event, so every tool call was discarded and the
turn parsed as empty — the run failed and the company never staffed.

Collect items from output_item.done and fall back to them when the
completion carries no output of its own; keep using the completion output
when it is populated (non-Codex Responses backends). Verified against the
real captured stream (5 hires recovered) and unit-tested with the exact
event shape.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

AgentCompany-New/src/main/services/provider-tool-turns/responses-format.ts

@@ -74,24 +74,37 @@ function errorText(data: JsonRecord): string {
 }
 
 /**
- * Accumulate a Responses SSE stream (Codex/Copilot) into the final result. The
- * result is accepted from the completion event's `response.output`.
+ * Accumulate a Responses SSE stream (Codex/Copilot) into the final result.
+ *
+ * The ChatGPT/Codex backend streams each output item in a
+ * `response.output_item.done` event and then sends a `response.completed`
+ * whose `response.output` is EMPTY. Reading only the completion event there
+ * loses every tool call (the model acted, but the turn parses as empty). So
+ * collect items as they finish and fall back to them when the completion
+ * carries no output of its own.
  */
 export async function accumulateResponsesStream(messages: AsyncIterable<ProviderEventStreamMessage>): Promise<AgentTurnResult> {
   const raw: string[] = [];
+  const streamed: unknown[] = [];
   for await (const message of messages) {
     raw.push(message.rawData);
     const type = eventType(message);
     if (!type) continue;
     if (type === "error" || type === "response.failed") throw new Error(errorText(message.data));
+    if (type === "response.output_item.done" && isRecord(message.data.item)) {
+      streamed.push(message.data.item);
+    }
     if (type === "response.completed" || type === "response.done" || type === "response.incomplete") {
       const response = isRecord(message.data.response) ? message.data.response : null;
-      if (response && Array.isArray(response.output)) {
-        const parsed = parseResponsesOutput(response);
-        return { text: parsed.text, toolCalls: parsed.toolCalls, rawResponseJson: JSON.stringify(raw) };
-      }
-      throw new Error("Responses completion event was missing response.output.");
+      const completionOutput = response && Array.isArray(response.output) ? response.output : [];
+      const output = completionOutput.length > 0 ? completionOutput : streamed;
+      const parsed = parseResponsesOutput({ output });
+      return { text: parsed.text, toolCalls: parsed.toolCalls, rawResponseJson: JSON.stringify(raw) };
     }
   }
+  if (streamed.length > 0) {
+    const parsed = parseResponsesOutput({ output: streamed });
+    return { text: parsed.text, toolCalls: parsed.toolCalls, rawResponseJson: JSON.stringify(raw) };
+  }
   throw new Error("Responses stream ended before a completion event.");
 }

AgentCompany-New/tests/unit/provider-tool-turns.test.ts

@@ -60,6 +60,20 @@ describe("openai responses encoding", () => {
     ]));
     expect(result.toolCalls).toEqual([{ id: "c3", name: "write_deliverable", arguments: { path: "a" } }]);
   });
+
+  test("accumulates output_item.done when the completion output is empty (Codex/ChatGPT backend)", async () => {
+    // The Codex backend streams each tool call as an output_item.done event and
+    // then sends a response.completed whose response.output is [].
+    const result = await accumulateResponsesStream(stream([
+      { type: "response.output_item.added", item: { id: "fc1", type: "function_call", status: "in_progress", arguments: "", call_id: "call_1", name: "hire_agent" } },
+      { type: "response.function_call_arguments.done", arguments: '{"name":"Dr. Li"}', item_id: "fc1" },
+      { type: "response.output_item.done", item: { id: "fc1", type: "function_call", status: "completed", arguments: '{"name":"Dr. Li","department":"Engineering"}', call_id: "call_1", name: "hire_agent" } },
+      { type: "response.output_item.done", item: { id: "m1", type: "message", content: [{ type: "output_text", text: "Staffing the team." }] } },
+      { type: "response.completed", response: { output: [], status: "completed" } },
+    ]));
+    expect(result.toolCalls).toEqual([{ id: "call_1", name: "hire_agent", arguments: { name: "Dr. Li", department: "Engineering" } }]);
+    expect(result.text).toBe("Staffing the team.");
+  });
 });
 
 describe("openai chat completions encoding", () => {