fix(responses): avoid invented empty assistant turns before tool replay

agent/codex_responses_adapter.py

@@ -571,12 +571,18 @@ def _chat_messages_to_responses_input(
             msg, is_github_responses=is_github_responses, current_issuer_kind=current_issuer_kind,
         )
         emit(message_items, msg)
+        fallback = None
         if not message_items:
-            # Every reasoning item needs a following item (else missing_following_item), hence the "" fallback.
             fallback = content_parts or (content_text if content_text.strip() else "" if reasoning_items else None)
-            if fallback is not None:
-                emit([{"role": "assistant", "content": fallback}], msg)
-        emit(_replay_tool_call_items(msg, start_index=len(items), wire_ids=wire_ids), msg)
+        tool_items = _replay_tool_call_items(msg, start_index=len(items) + (fallback is not None), wire_ids=wire_ids)
+        # A function_call already follows its reasoning. Inventing an empty assistant
+        # message between them changes the replayed turn (Muse can emit corrupt finals).
+        # Keep a follower only for reasoning with no other following item, and make it
+        # non-empty: strict Responses-compatible providers reject "" with 400.
+        if fallback is not None and not (fallback == "" and tool_items):
+            follower = " " if fallback == "" else fallback
+            emit([{"role": "assistant", "content": follower}], msg)
+        emit(tool_items, msg)
     # The server renders nothing placed before a compaction item, so pre-checkpoint history is
     # dead weight and plaintext asks / merged summaries silently vanish. Keep the newest checkpoint
     # first, retain pre-checkpoint USER and SUMMARY messages within a token budget, leave the tail.

tests/agent/test_responses_empty_tool_replay.py

@@ -0,0 +1,31 @@
+"""Responses replay must not invent a blank assistant turn before a tool call.
+
+Regression investigation for #103483: synthetic Muse tool-loop A/B reproduction.
+Also covers #75202: the lone-reasoning follower must be non-empty for strict
+Responses-compatible providers that reject "" with 400.
+"""
+import pytest
+from agent.codex_responses_adapter import _chat_messages_to_responses_input
+
+@pytest.mark.parametrize('content', ['', None])
+def test_reasoning_tool_round_has_no_invented_assistant_message(content):
+    messages = [{
+        'role': 'assistant', 'content': content,
+        'codex_reasoning_items': [{'type': 'reasoning', 'id': 'rs_test',
+                                  'encrypted_content': 'synthetic-encrypted-fixture', 'summary': []}],
+        'tool_calls': [{'id': 'call_test', 'type': 'function',
+                        'function': {'name': 'record_step', 'arguments': '{"step":1}'}}],
+    }, {'role': 'tool', 'tool_call_id': 'call_test', 'content': '{"ok":true}'}]
+    items = _chat_messages_to_responses_input(messages)
+    assert [item.get('type', item.get('role')) for item in items] == [
+        'reasoning', 'function_call', 'function_call_output']
+    assert items[1]['call_id'] == items[2]['call_id']
+
+
+def test_reasoning_without_tool_keeps_nonempty_following_item():
+    items = _chat_messages_to_responses_input([{
+        'role': 'assistant', 'content': '',
+        'codex_reasoning_items': [{'type': 'reasoning', 'id': 'rs_test',
+                                  'encrypted_content': 'synthetic-encrypted-fixture', 'summary': []}],
+    }])
+    assert items[-1] == {'role': 'assistant', 'content': ' '}