feat: preserve governed OpenAI tool messages in 1.4.0
Add faithful single-choice chat results, tool stream activity, error-frame retries, and isolated protocol caching. Reject empty cached completions with the same value predicate as transport, add regression coverage, and adopt Apache-2.0.
This commit is contained in:
@@ -0,0 +1,354 @@
|
||||
"""OpenAI 消息保真回归：按网关 Chat Completions 形态构造。"""
|
||||
|
||||
import json
|
||||
|
||||
import httpx
|
||||
import pytest
|
||||
|
||||
from polygateway.errors import AllSourcesExhausted, TransientError
|
||||
from tests.unit.test_client import _client
|
||||
from tests.unit.test_openai_compat import _complete, _source, _sse_stream, _transport_for
|
||||
from tests.unit.test_openai_completion import accumulate, chunk, tool_frames
|
||||
|
||||
|
||||
async def test_error_before_done_is_failure():
|
||||
"""文本模式也不能接受 error 后的 DONE。"""
|
||||
transport = _transport_for(
|
||||
lambda _: _sse_stream(chunk({"content": "partial"}), 'data: {"error": null}\n\n')
|
||||
)
|
||||
try:
|
||||
with pytest.raises(TransientError):
|
||||
await _complete(transport, _source())
|
||||
finally:
|
||||
await transport.aclose()
|
||||
|
||||
|
||||
async def test_mixed_content_reasoning_survive():
|
||||
"""同一帧双文本不互斥。"""
|
||||
transport = _transport_for(
|
||||
lambda _: _sse_stream(chunk({"content": "answer", "reasoning_content": "reason"}))
|
||||
)
|
||||
try:
|
||||
result = await _complete(transport, _source())
|
||||
assert (result.content, result.thinking) == ("answer", "reason")
|
||||
finally:
|
||||
await transport.aclose()
|
||||
|
||||
|
||||
async def test_tools_roundtrip_and_retry():
|
||||
"""真实 transport 经过治理，失败重试不污染工具，消息可直接往返。"""
|
||||
requests = []
|
||||
|
||||
def handler(request):
|
||||
requests.append(json.loads(request.content))
|
||||
if len(requests) == 1:
|
||||
return _sse_stream(
|
||||
chunk({"content": "partial"}),
|
||||
'data: {"error": {"code": "upstream_stream_read_error"}}\n\n',
|
||||
)
|
||||
return _sse_stream(*tool_frames())
|
||||
|
||||
client = _client(handler=handler)
|
||||
try:
|
||||
response = await client.chat(
|
||||
[{"role": "user", "content": "hi"}],
|
||||
response_mode="openai",
|
||||
overlay={"reasoning_effort": "max"},
|
||||
)
|
||||
assert response.call_stats.attempts == 2
|
||||
message = response.completion.message
|
||||
assert message["role"] == "assistant" and message["content"] is None
|
||||
assert [t["id"] for t in message["tool_calls"]] == ["a", "b"]
|
||||
assert message["tool_calls"][0]["function"]["arguments"] == '{"v": "中" }'
|
||||
assert response.completion.response_id == "upstream-1"
|
||||
assert response.completion.finish_reason == "tool_calls"
|
||||
messages = [message, {"role": "tool", "tool_call_id": "a", "content": "ok"}]
|
||||
await client.chat(messages, response_mode="openai")
|
||||
assert requests[-1]["messages"] == messages
|
||||
assert requests[0]["reasoning_effort"] == "max"
|
||||
finally:
|
||||
await client._transport.aclose()
|
||||
|
||||
|
||||
@pytest.mark.parametrize("role", [None, "user"])
|
||||
async def test_final_role_required(role):
|
||||
"""非流式缺失或错误 role 必须协议失败。"""
|
||||
message = {"content": "ok"}
|
||||
if role:
|
||||
message["role"] = role
|
||||
client = _client(
|
||||
handler=lambda _: httpx.Response(200, json={"choices": [{"message": message}]})
|
||||
)
|
||||
try:
|
||||
with pytest.raises(AllSourcesExhausted):
|
||||
await client.chat(
|
||||
[{"role": "user", "content": "hi"}], stream=False, response_mode="openai"
|
||||
)
|
||||
finally:
|
||||
await client._transport.aclose()
|
||||
|
||||
|
||||
@pytest.mark.parametrize("stream", [True, False])
|
||||
async def test_completion_cache_roundtrip(stream):
|
||||
"""完整结果回放，修改返回对象不毒化缓存；模式隔离。"""
|
||||
from polygateway.backends.memory.cache import InMemoryCache
|
||||
|
||||
completion = accumulate().finish()
|
||||
|
||||
def handler(_):
|
||||
if stream:
|
||||
return _sse_stream(*tool_frames())
|
||||
return httpx.Response(
|
||||
200,
|
||||
json={
|
||||
"id": "upstream-1",
|
||||
"choices": [{"message": completion.message, "finish_reason": "tool_calls"}],
|
||||
},
|
||||
)
|
||||
|
||||
client = _client(
|
||||
handler=handler, cache=InMemoryCache(), cache_namespace="tools", cache_ttl_s=60
|
||||
)
|
||||
try:
|
||||
first = await client.chat([], stream=stream, response_mode="openai")
|
||||
second = await client.chat([], stream=stream, response_mode="openai")
|
||||
assert not first.cache_hit and second.cache_hit
|
||||
assert second.completion == first.completion
|
||||
assert second.call_id != first.call_id and second.call_stats.attempts == 0
|
||||
second.completion.message["tool_calls"].clear()
|
||||
third = await client.chat([], stream=stream, response_mode="openai")
|
||||
assert len(third.completion.message["tool_calls"]) == 2
|
||||
finally:
|
||||
await client._transport.aclose()
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"kwargs",
|
||||
[
|
||||
{"response_mode": "bad"},
|
||||
{"response_mode": "openai", "structured": "json"},
|
||||
{"response_mode": "openai", "overlay": {"n": 2}},
|
||||
{"response_mode": "openai", "overlay": {"n": True}},
|
||||
],
|
||||
)
|
||||
async def test_invalid_mode_combinations_fail_before_http(kwargs):
|
||||
"""组合错误入口拒绝。"""
|
||||
|
||||
def unexpected(_):
|
||||
pytest.fail("不应发送 HTTP")
|
||||
|
||||
client = _client(handler=unexpected)
|
||||
try:
|
||||
with pytest.raises(ValueError):
|
||||
await client.chat([], **kwargs)
|
||||
finally:
|
||||
await client._transport.aclose()
|
||||
|
||||
|
||||
async def test_openai_never_salvages():
|
||||
"""已收完整工具也必须有 DONE。"""
|
||||
transport = _transport_for(lambda _: _sse_stream(*tool_frames(), done=False))
|
||||
try:
|
||||
with pytest.raises(TransientError, match="DONE"):
|
||||
await transport.complete(
|
||||
messages=[],
|
||||
source=_source(missing_done="salvage"),
|
||||
stream=True,
|
||||
overlay={},
|
||||
call_id="c",
|
||||
reasoning_effort=None,
|
||||
first_token_event=None,
|
||||
content_event=None,
|
||||
response_mode="openai",
|
||||
)
|
||||
finally:
|
||||
await transport.aclose()
|
||||
|
||||
|
||||
async def test_tool_progress_sets_events():
|
||||
"""工具首进展驱动两个事件与时间观测。"""
|
||||
import asyncio
|
||||
|
||||
first, progress = asyncio.Event(), asyncio.Event()
|
||||
transport = _transport_for(lambda _: _sse_stream(*tool_frames()))
|
||||
try:
|
||||
result = await transport.complete(
|
||||
messages=[],
|
||||
source=_source(),
|
||||
stream=True,
|
||||
overlay={},
|
||||
call_id="c",
|
||||
reasoning_effort=None,
|
||||
first_token_event=first,
|
||||
content_event=progress,
|
||||
response_mode="openai",
|
||||
)
|
||||
assert first.is_set() and progress.is_set()
|
||||
assert result.ttft_ms is not None and result.stall_count == 0
|
||||
finally:
|
||||
await transport.aclose()
|
||||
|
||||
|
||||
async def test_real_transport_stall_hedge_cancels_loser():
|
||||
"""真实工具流沉默触发异源对冲，连接关闭且赢家不混入半成品。"""
|
||||
import asyncio
|
||||
import time
|
||||
|
||||
from polygateway.sources import RoundRobinSelector
|
||||
from polygateway.types import ChatRequest, _CallContext
|
||||
from tests.unit.test_retry import _src
|
||||
from tests.unit.test_stall_hedge import _harness
|
||||
|
||||
entered, closed = asyncio.Event(), asyncio.Event()
|
||||
|
||||
class Hanging(httpx.AsyncByteStream):
|
||||
"""工具首帧后阻塞，取消由真实 httpx 流处理。"""
|
||||
|
||||
async def __aiter__(self):
|
||||
yield tool_frames()[0].encode()
|
||||
yield tool_frames()[1].encode()
|
||||
entered.set()
|
||||
await asyncio.Event().wait()
|
||||
|
||||
async def aclose(self):
|
||||
closed.set()
|
||||
|
||||
count = 0
|
||||
|
||||
def handler(_):
|
||||
nonlocal count
|
||||
count += 1
|
||||
if count == 1:
|
||||
return httpx.Response(200, stream=Hanging())
|
||||
return _sse_stream(*tool_frames())
|
||||
|
||||
transport = _transport_for(handler)
|
||||
sources = [_src("s1"), _src("s2")]
|
||||
mw, limiter, *_ = _harness(
|
||||
sources, transport, stall_after_s=0.05, selector=RoundRobinSelector()
|
||||
)
|
||||
context = _CallContext(now=time.monotonic)
|
||||
try:
|
||||
response = await asyncio.wait_for(
|
||||
mw(ChatRequest(messages=[], response_mode="openai", call_context=context)), 3
|
||||
)
|
||||
assert entered.is_set() and closed.is_set()
|
||||
assert response.source_name == "s2" and len(response.completion.message["tool_calls"]) == 2
|
||||
assert context.snapshot().hedges == 1 and context.snapshot().attempts == 2
|
||||
finally:
|
||||
await transport.aclose()
|
||||
|
||||
|
||||
@pytest.mark.parametrize("stream", [True, False])
|
||||
async def test_bad_usage_is_domain_failure(stream):
|
||||
"""坏 usage 不能穿透成裸 AttributeError，包括默认文本模式。"""
|
||||
transport = _transport_for(
|
||||
lambda _: (
|
||||
_sse_stream(chunk({"content": "ok"}), 'data: {"choices": [], "usage": [1]}\n\n')
|
||||
if stream
|
||||
else httpx.Response(
|
||||
200, json={"choices": [{"message": {"content": "ok"}}], "usage": [1]}
|
||||
)
|
||||
)
|
||||
)
|
||||
try:
|
||||
with pytest.raises(TransientError):
|
||||
await _complete(transport, _source(), stream=stream)
|
||||
finally:
|
||||
await transport.aclose()
|
||||
|
||||
|
||||
async def test_nonstream_tool_result_has_no_liveness_events():
|
||||
"""非流式工具不制造流式观测。"""
|
||||
import asyncio
|
||||
|
||||
completion = accumulate().finish()
|
||||
first, progress = asyncio.Event(), asyncio.Event()
|
||||
transport = _transport_for(
|
||||
lambda _: httpx.Response(
|
||||
200, json={"choices": [{"message": completion.message, "finish_reason": "tool_calls"}]}
|
||||
)
|
||||
)
|
||||
try:
|
||||
result = await transport.complete(
|
||||
messages=[],
|
||||
source=_source(),
|
||||
stream=False,
|
||||
overlay={},
|
||||
call_id="c",
|
||||
reasoning_effort=None,
|
||||
first_token_event=first,
|
||||
content_event=progress,
|
||||
response_mode="openai",
|
||||
)
|
||||
assert result.completion.message == completion.message
|
||||
assert not first.is_set() and not progress.is_set()
|
||||
assert result.ttft_ms is None and result.stall_ms is None
|
||||
finally:
|
||||
await transport.aclose()
|
||||
|
||||
|
||||
async def test_openai_does_not_bypass_managed_conflict():
|
||||
"""保真模式不静默撤销源级受管意图。"""
|
||||
from polygateway.errors import RequestRejectedError
|
||||
|
||||
def unexpected(_):
|
||||
pytest.fail("冲突请求不应发送 HTTP")
|
||||
|
||||
transport = _transport_for(unexpected)
|
||||
try:
|
||||
with pytest.raises(RequestRejectedError):
|
||||
await transport.complete(
|
||||
messages=[],
|
||||
source=_source(enable_thinking=True),
|
||||
stream=True,
|
||||
overlay={"reasoning_effort": "max"},
|
||||
call_id="c",
|
||||
reasoning_effort=None,
|
||||
first_token_event=None,
|
||||
content_event=None,
|
||||
response_mode="openai",
|
||||
)
|
||||
finally:
|
||||
await transport.aclose()
|
||||
|
||||
|
||||
async def test_cancel_tool_stream_releases_resources():
|
||||
"""显式取消真实工具流，连接关闭且配额归还。"""
|
||||
import asyncio
|
||||
|
||||
from polygateway.backends.memory.limiter import InMemoryLimiter
|
||||
from polygateway.types import GlobalLimits
|
||||
|
||||
entered, closed = asyncio.Event(), asyncio.Event()
|
||||
|
||||
class Hanging(httpx.AsyncByteStream):
|
||||
"""可确定取消窗口的 HTTP 流。"""
|
||||
|
||||
async def __aiter__(self):
|
||||
for frame in tool_frames()[:2]:
|
||||
yield frame.encode()
|
||||
entered.set()
|
||||
await asyncio.Event().wait()
|
||||
|
||||
async def aclose(self):
|
||||
closed.set()
|
||||
|
||||
source = _source(max_concurrency=1)
|
||||
limiter = InMemoryLimiter(
|
||||
scope="llm", sources={source.name: source}, global_limits=GlobalLimits(1, 0, 0)
|
||||
)
|
||||
client = _client(
|
||||
sources=[source], limiter=limiter, handler=lambda _: httpx.Response(200, stream=Hanging())
|
||||
)
|
||||
task = asyncio.create_task(client.chat([], response_mode="openai"))
|
||||
try:
|
||||
await asyncio.wait_for(entered.wait(), 2)
|
||||
task.cancel()
|
||||
with pytest.raises(asyncio.CancelledError):
|
||||
await task
|
||||
assert closed.is_set()
|
||||
assert (await limiter.source_stats(source.name)).inflight == 0
|
||||
finally:
|
||||
task.cancel()
|
||||
await client._transport.aclose()
|
||||
Reference in New Issue
Block a user