fix(sse): preserve Responses API events in stream translation

open-sse/utils/stream.ts

@@ -595,8 +595,12 @@ export function createSSEStream(options: StreamOptions = {}) {
               // When translating Responses -> Claude, `item` is already a Claude SSE event;
               // sanitizing it as an OpenAI chunk strips message_start/content_block_delta/message_stop
               // and causes Claude Code to drop the assistant message.
+              // #761: Responses API events have {event, data} structure — skip sanitization
+              // entirely as it strips them to {"object":"chat.completion.chunk"}, losing all content.
               let itemSanitized: Record<string, unknown> = item;
-              if (sourceFormat === FORMATS.OPENAI || sourceFormat === FORMATS.OPENAI_RESPONSES) {
+              const isResponsesEvent =
+                typeof item?.event === "string" && item.event.startsWith("response.");
+              if (sourceFormat === FORMATS.OPENAI && !isResponsesEvent) {
                 itemSanitized = sanitizeStreamingChunk(itemSanitized) as Record<string, unknown>;
 
                 // Extract reasoning tags from content if translation generated them

tests/unit/claude-code-rendering-fixes.test.mjs

@@ -150,6 +150,72 @@ test("Responses->Chat: empty-name tool call is dropped when done still has no va
   assert.equal(state.toolCallIndex, 0);
 });
 
+test("Claude->Responses: {event,data} items bypass sanitization in translate mode", async () => {
+  // Regression test: when translating Claude-format (GLM) to Responses API for Codex CLI,
+  // the sanitizer was stripping {event,data} items to {"object":"chat.completion.chunk"},
+  // losing all content and the critical response.completed event.
+  const encoder = new TextEncoder();
+  const decoder = new TextDecoder();
+
+  // Create stream translating claude → openai-responses (same path as GLM via Codex CLI)
+  const stream = createSSETransformStreamWithLogger(
+    FORMATS.CLAUDE,
+    FORMATS.OPENAI_RESPONSES,
+    "glm",
+    null,
+    null,
+    "glm-5.1",
+    "conn-test",
+    { messages: [{ role: "user", content: "hi" }] },
+    null,
+    null
+  );
+
+  const writer = stream.writable.getWriter();
+  // Simulate Claude-format SSE from GLM
+  await writer.write(
+    encoder.encode(
+      'event: message_start\ndata: {"type":"message_start","message":{"id":"msg_test","type":"message","role":"assistant","model":"glm-5.1","content":[],"stop_reason":null,"usage":{"input_tokens":10,"output_tokens":0}}}\n\n'
+    )
+  );
+  await writer.write(
+    encoder.encode(
+      'event: content_block_start\ndata: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}\n\n'
+    )
+  );
+  await writer.write(
+    encoder.encode(
+      'event: content_block_delta\ndata: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"hello"}}\n\n'
+    )
+  );
+  await writer.write(
+    encoder.encode(
+      'event: message_delta\ndata: {"type":"message_delta","delta":{"stop_reason":"end_turn"},"usage":{"output_tokens":5}}\n\n'
+    )
+  );
+  await writer.write(encoder.encode('event: message_stop\ndata: {"type":"message_stop"}\n\n'));
+  await writer.close();
+
+  const reader = stream.readable.getReader();
+  let output = "";
+  while (true) {
+    const { value, done } = await reader.read();
+    if (done) break;
+    output += decoder.decode(value, { stream: true });
+  }
+  output += decoder.decode();
+
+  // Must emit Responses API events (not sanitized chat.completion.chunk objects)
+  assert.match(output, /event: response\.created/);
+  assert.match(output, /event: response\.output_text\.delta/);
+  assert.match(output, /event: response\.completed/);
+  assert.match(output, /"delta":"hello"/);
+  assert.match(output, /"status":"completed"/);
+
+  // Must NOT contain sanitized empty chunks
+  assert.doesNotMatch(output, /data: \{"object":"chat\.completion\.chunk"\}\n\n/);
+});
+
 test("Responses->Claude: translated Claude SSE is not sanitized into empty OpenAI chunks", async () => {
   const encoder = new TextEncoder();
   const decoder = new TextDecoder();