Skip to content

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

C

fix(serve): a past tool_calls that is not an array answered 500 on Qwen and Kimi

kevin93277c65c8d8be95
diff --git a/c/openai_server.py b/c/openai_server.py
index 165eb73a..daa5fabb 100644
--- a/c/openai_server.py
+++ b/c/openai_server.py
@@ -1451,7 +1451,11 @@ def _k3_order_tool_results(messages):
msg = messages[i]
if isinstance(msg, dict) and msg.get("role") == "assistant":
index_map = {}
- for pos, tc in enumerate(msg.get("tool_calls") or [], start=1):
+ calls = msg.get("tool_calls")
+ if calls is not None and not isinstance(calls, list):
+ raise APIError(400, "`tool_calls` must be an array.",
+ f"messages.{i}.tool_calls")
+ for pos, tc in enumerate(calls or [], start=1):
if isinstance(tc, dict) and tc.get("id") is not None:
fn = tc.get("function", tc)
nm = fn.get("name") if isinstance(fn, dict) else None
@@ -1910,6 +1914,8 @@ def _qwen_tool_calls(tool_calls, has_content, index):
preceding content with a blank line only when that content is non-empty, and
every later call with a single newline; getting that wrong changes the prompt
the model is conditioned on."""
+ if tool_calls is not None and not isinstance(tool_calls, list):
+ raise APIError(400, "`tool_calls` must be an array.", f"messages.{index}.tool_calls")
out = []
for position, call in enumerate(tool_calls or []):
if not isinstance(call, dict):
diff --git a/c/tests/test_openai_server.py b/c/tests/test_openai_server.py
index 809d0371..74f12414 100644
--- a/c/tests/test_openai_server.py
+++ b/c/tests/test_openai_server.py
@@ -3294,6 +3294,39 @@ class HTTPTest(unittest.TestCase):
history({"name": "fn", "arguments": "{}"})) as response:
self.assertEqual(response.status, 200)
+ def test_past_tool_calls_that_are_not_an_array_are_a_client_error(self):
+ """A replayed assistant turn with `tool_calls: 5` answered HTTP 500 on Qwen and Kimi.
+
+ The other renderers already answer 400 "`tool_calls` must be an array.". The Qwen
+ renderer's _qwen_tool_calls and Kimi's _k3_order_tool_results iterated the value as
+ it came, and the TypeError became do_POST's 500 "The colibri engine failed to process the
+ request."
+ """
+ def history(tool_calls):
+ return {"model": "test-model", "messages": [
+ {"role": "user", "content": "run it"},
+ {"role": "assistant", "content": "", "tool_calls": tool_calls},
+ {"role": "tool", "tool_call_id": "x", "content": "done"},
+ {"role": "user", "content": "and now?"},
+ ]}
+ for arch in ("glm", "glm53", "qwen36", "qwen38", "kimi", "deepseek_v4",
+ "deepseek_v41", "mimo"):
+ for tool_calls in (5, 1.5, True, "call", {"id": "x"}):
+ with self.subTest(arch=arch, tool_calls=tool_calls):
+ with patch("openai_server.ARCH", arch):
+ with self.assertRaises(HTTPError) as caught:
+ self.request("/v1/chat/completions", history(tool_calls))
+ self.addCleanup(caught.exception.close)
+ self.assertEqual(caught.exception.code, 400)
+ self.assertEqual(json.loads(caught.exception.read())["error"]["param"],
+ "messages.1.tool_calls")
+ with self.subTest(arch=arch, tool_calls="well formed"):
+ with patch("openai_server.ARCH", arch):
+ with self.request("/v1/chat/completions", history([
+ {"id": "x", "type": "function",
+ "function": {"name": "fn", "arguments": "{}"}}])) as response:
+ self.assertEqual(response.status, 200)
+
def test_a_text_part_whose_text_is_not_a_string_is_a_client_error(self):
"""{"type": "text", "text": 5} answered HTTP 500 on GLM-5.3, Qwen3.8 and V4.1.