diff --git a/tests/agent/test_responses_empty_tool_replay.py b/tests/agent/test_responses_empty_tool_replay.py
index 452f419221..caad38b906 100644
--- a/tests/agent/test_responses_empty_tool_replay.py
+++ b/tests/agent/test_responses_empty_tool_replay.py
@@ -4,6 +4,9 @@ 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 json
+from pathlib import Path
+
 import pytest
 from agent.codex_responses_adapter import _chat_messages_to_responses_input
 
@@ -29,3 +32,32 @@ def test_reasoning_without_tool_keeps_nonempty_following_item():
                                   'encrypted_content': 'synthetic-encrypted-fixture', 'summary': []}],
     }])
     assert items[-1] == {'role': 'assistant', 'content': ' '}
+
+
+def test_failing_turn_fixture_emits_no_invented_carrier():
+    """Real failing-turn shape from #103483 (fixture by @cristianbdev, see
+    https://gist.github.com/cristianbdev/6036f7aa3838935adfaeb9b6800f7450 —
+    3 reasoning+tool assistant rows, 5 tool calls/outputs, content redacted).
+
+    The adapter must invent no blank carrier, must keep every reasoning item
+    directly followed by a function_call, and must preserve all call/output
+    pairings.
+    """
+    fixture_path = (
+        Path(__file__).resolve().parent.parent
+        / "fixtures" / "spark_failing_turn_shape.json"
+    )
+    messages = json.loads(fixture_path.read_text(encoding="utf-8"))
+    items = _chat_messages_to_responses_input(messages)
+    assert {"role": "assistant", "content": ""} not in items
+    for i, item in enumerate(items):
+        if item.get("type") == "reasoning":
+            # Reasoning items may cluster; the next non-reasoning item after
+            # each of them must be the function_call that follows it — never
+            # an invented blank assistant message.
+            nxt = next(it for it in items[i + 1:] if it.get("type") != "reasoning")
+            assert nxt.get("type") == "function_call"
+    calls = [item["call_id"] for item in items if item.get("type") == "function_call"]
+    outputs = [item["call_id"] for item in items if item.get("type") == "function_call_output"]
+    assert len(calls) == 5
+    assert sorted(calls) == sorted(outputs)
diff --git a/tests/fixtures/spark_failing_turn_shape.json b/tests/fixtures/spark_failing_turn_shape.json
new file mode 100644
index 0000000000..c6f3c3d45e
--- /dev/null
+++ b/tests/fixtures/spark_failing_turn_shape.json
@@ -0,0 +1,91 @@
+[
+  {
+    "role": "user",
+    "content": "Investigate the installed 'agy' CLI and summarize its version and capabilities in one report."
+  },
+  {
+    "role": "assistant",
+    "content": "",
+    "codex_reasoning_items": [
+      {"type": "reasoning", "id": "rs_fixture_01", "summary": [{"type": "summary_text", "text": "Batched three tool calls to inspect the CLI."}], "encrypted_content": "synthetic-encrypted-content-placeholder-0001-AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA"},
+      {"type": "reasoning", "id": "rs_fixture_02", "summary": [], "encrypted_content": "synthetic-encrypted-content-placeholder-0002-BBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB"}
+    ],
+    "tool_calls": [
+      {
+        "id": "call_fixture_01",
+        "call_id": "call_fixture_01",
+        "type": "function",
+        "function": {"name": "terminal", "arguments": "{\"command\": \"which agy; agy --version\"}"}
+      }
+    ]
+  },
+  {
+    "role": "tool",
+    "tool_call_id": "call_fixture_01",
+    "content": "[tool output elided — 11,269 chars in the original failing run]"
+  },
+  {
+    "role": "assistant",
+    "content": "",
+    "codex_reasoning_items": [
+      {"type": "reasoning", "id": "rs_fixture_03", "summary": [], "encrypted_content": "synthetic-encrypted-content-placeholder-0003-CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC"},
+      {"type": "reasoning", "id": "rs_fixture_04", "summary": [], "encrypted_content": "synthetic-encrypted-content-placeholder-0004-DDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDD"}
+    ],
+    "tool_calls": [
+      {
+        "id": "call_fixture_02",
+        "call_id": "call_fixture_02",
+        "type": "function",
+        "function": {"name": "terminal", "arguments": "{\"command\": \"agy agent --help\"}"}
+      }
+    ]
+  },
+  {
+    "role": "tool",
+    "tool_call_id": "call_fixture_02",
+    "content": "[tool output elided — 3,169 chars in the original failing run]"
+  },
+  {
+    "role": "assistant",
+    "content": "",
+    "codex_reasoning_items": [
+      {"type": "reasoning", "id": "rs_fixture_05", "summary": [], "encrypted_content": "synthetic-encrypted-content-placeholder-0005-EEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEE"}
+    ],
+    "tool_calls": [
+      {
+        "id": "call_fixture_03",
+        "call_id": "call_fixture_03",
+        "type": "function",
+        "function": {"name": "terminal", "arguments": "{\"command\": \"agy mcp --help\"}"}
+      },
+      {
+        "id": "call_fixture_04",
+        "call_id": "call_fixture_04",
+        "type": "function",
+        "function": {"name": "terminal", "arguments": "{\"command\": \"agy changelog | head -40\"}"}
+      },
+      {
+        "id": "call_fixture_05",
+        "call_id": "call_fixture_05",
+        "type": "function",
+        "function": {"name": "terminal", "arguments": "{\"command\": \"ls ~/.config/agy/ 2>&1 || true\"}"}
+      }
+    ]
+  },
+  {
+    "role": "tool",
+    "tool_call_id": "call_fixture_03",
+    "content": "[tool output elided — 1,494 chars in the original failing run]"
+  },
+  {
+    "role": "tool",
+    "tool_call_id": "call_fixture_04",
+    "content": "[tool output elided — 32,686 chars in the original failing run]"
+  },
+  {
+    "role": "tool",
+    "tool_call_id": "call_fixture_05",
+    "content": "[tool output elided — 598 chars in the original failing run]"
+  }
+]
+