Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
"""Chat Completions tool results must follow call IDs before the template drops them."""
import copy
import http.client
import json
from pathlib import Path
import unittest
from serve.frontend import ChatTemplate, openai_to_messages
from serve.server import ByteTokenizer, MockEngine, Service, serve
TEMPLATE = ChatTemplate(Path(__file__).with_name("chat_template.jinja"))
def history():
return [
{"role": "user", "content": "Read a.txt and b.txt."},
{"role": "assistant", "content": "", "tool_calls": [
{"id": "call_a", "type": "function", "function": {
"name": "read_file", "arguments": '{"path":"a.txt"}'}},
{"id": "call_b", "type": "function", "function": {
"name": "read_file", "arguments": '{"path":"b.txt"}'}}]},
{"role": "tool", "tool_call_id": "call_b", "content": "RESULT_FROM_B"},
{"role": "tool", "tool_call_id": "call_a", "content": "RESULT_FROM_A"},
]
def normalized(messages):
return openai_to_messages({"messages": messages})[0]
def rendered(messages):
return TEMPLATE.render(normalized(messages), enable_thinking=False)
class ChatToolResults(unittest.TestCase):
def test_reversed_results_of_the_same_function_follow_call_ids(self):
raw = history()
before = copy.deepcopy(raw)
ordered = raw[:2] + list(reversed(raw[2:]))
self.assertEqual(rendered(raw), rendered(ordered))
self.assertLess(rendered(raw).index("RESULT_FROM_A"), rendered(raw).index("RESULT_FROM_B"))
self.assertEqual(raw, before)
messages = normalized(raw)
self.assertEqual([m["content"] for m in messages[2:]], ["RESULT_FROM_A", "RESULT_FROM_B"])
self.assertTrue(all(set(m) == {"role", "content"} for m in messages[2:]))
self.assertTrue(all(set(c) == {"function"} for c in messages[1]["tool_calls"]))
def test_already_ordered_and_legacy_prompts_are_identical(self):
ordered = history()
ordered[2:] = reversed(ordered[2:])
legacy = copy.deepcopy(ordered)
for call in legacy[1]["tool_calls"]:
call.pop("id")
for result in legacy[2:]:
result.pop("tool_call_id")
self.assertEqual(normalized(ordered), normalized(legacy))
self.assertEqual(rendered(ordered), rendered(legacy))
def test_ambiguous_or_incomplete_ids_preserve_received_order(self):
for case in ("missing_call", "missing_result", "duplicate_call", "duplicate_result", "unknown",
"list_call", "list_result", "empty_call", "empty_result", "partial"):
raw = history()
if case == "missing_call":
raw[1]["tool_calls"][0].pop("id")
elif case == "missing_result":
raw[2].pop("tool_call_id")
elif case == "duplicate_call":
raw[1]["tool_calls"][1]["id"] = "call_a"
elif case == "duplicate_result":
raw[3]["tool_call_id"] = "call_b"
elif case == "unknown":
raw[2]["tool_call_id"] = "unknown"
elif case == "list_call":
raw[1]["tool_calls"][0]["id"] = []
elif case == "list_result":
raw[2]["tool_call_id"] = []
elif case == "empty_call":
raw[1]["tool_calls"][0]["id"] = ""
elif case == "empty_result":
raw[2]["tool_call_id"] = ""
else:
raw.pop()
with self.subTest(case=case):
self.assertEqual([m["content"] for m in normalized(raw)[2:]],
[m["content"] for m in raw[2:]])
def test_each_tool_run_is_scoped_to_its_assistant_turn(self):
raw = history() + history()
self.assertEqual([m["content"] for m in normalized(raw) if m["role"] == "tool"],
["RESULT_FROM_A", "RESULT_FROM_B"] * 2)
raw = history()
raw.insert(2, {"role": "user", "content": "intervening message"})
self.assertEqual([m["content"] for m in normalized(raw) if m["role"] == "tool"],
["RESULT_FROM_B", "RESULT_FROM_A"])
def test_json_encoded_history_and_calls(self):
raw = history()
raw[1]["tool_calls"] = json.dumps(raw[1]["tool_calls"])
self.assertEqual(normalized(json.dumps(raw)), normalized(history()))
def test_multimodal_result_moves_with_its_call(self):
raw = history()
raw[2]["content"] = [{"type": "text", "text": "RESULT_FROM_B"},
{"type": "image_url", "image_url": {"url": "data:image/png;base64,AA=="}}]
messages = normalized(raw)
self.assertEqual(messages[2]["content"], "RESULT_FROM_A")
self.assertEqual(messages[3]["content"], [{"type": "text", "text": "RESULT_FROM_B"},
{"type": "image", "source": "data:image/png;base64,AA=="}])
def test_http_engine_receives_the_ordered_prompt(self):
tok = ByteTokenizer()
engine = MockEngine(tok, "ok", max_context=4096)
service = Service(engine, tok, TEMPLATE)
httpd = serve(service, port=0)
connection = http.client.HTTPConnection("127.0.0.1", httpd.server_address[1], timeout=10)
try:
connection.request("POST", "/v1/chat/completions", json.dumps({
"messages": history(), "max_tokens": 16, "reasoning_effort": "none"}),
{"Content-Type": "application/json"})
response = connection.getresponse()
body = response.read()
self.assertEqual(response.status, 200, body)
prompt = tok.decode(engine.last_prompt)
self.assertLess(prompt.index("RESULT_FROM_A"), prompt.index("RESULT_FROM_B"))
finally:
connection.close()
httpd.shutdown()
httpd.server_close()
if __name__ == "__main__":
unittest.main()