fix: 模型级强制参数(kimi-k2.7-code 锁 temperature=1)+ 测速超时/错误文案

- config: ModelInfo.params + MODEL_PARAM_KEYS 白名单 + _parse_model_params(写错即 ConfigError)
- adapters: _forced_params 在 payload 组装后强制覆盖;_learn_from_upstream_error 按上游 400
  要求修正并重发一次(非原样重试);流式已吐内容则不重发
- main: /v1/models 暴露 params;探针超时 15s→30s;_speedtest_error_text 说清超时/上游原文
- settings: 面板保存保留手写 models[].params(原先会被静默清掉)
- gateway.yaml: kimi-k2.7-code params.temperature=1
- 前端: Models 挂锁+参数固定标签;Chat 温度 chip 挂锁+抽屉提示;新增 lock 图标;中英对称
- tests: +10 → 168 passed / 0 failed
This commit is contained in:
api-gateway committed 2026-09-18 12:18:10 +08:00
1 parent e4065d31c3
commit 4507dc4b8e
14 files changed
+631 -13

No files matched your search

+253
View File
@@ -3779,6 +3779,259 @@ def test_opencode_session_header_from_conversation_id():
assert "x-opencode-session" not in ag._req_headers("conv-abc")
def test_api_conversation_id_flows_to_session_header():
"""API 调用(HTTP body)传 conversation_id → 适配器 session_id;且不透传上游 payload。
覆盖老板问的「非前端、纯 API 调用是否支持稳定会话」:只要请求体带
conversation_id,后端就把它作为 OpenCode Go 的 x-opencode-session 透传。
"""
from fastapi.testclient import TestClient
import app.main as m
captured = {}
class Capture:
async def chat(self, model=None, messages=None, session_id=None, extra=None, **kw):
captured["session_id"] = session_id
captured["extra"] = extra
return {"id": "c1", "object": "chat.completion", "model": model,
"choices": [{"message": {"role": "assistant", "content": "ok"},
"finish_reason": "stop"}],
"usage": {"prompt_tokens": 2, "completion_tokens": 1}}
orig = m._get_adapter
m._get_adapter = lambda model_id, timeout=None, cfg=None: (
Capture(), "agnes", m._cfg().providers["agnes"])
try:
client = TestClient(m.app)
# ① 带 conversation_id → 原样落为 session_id;且不作为上游参数透传
r = client.post("/v1/chat/completions", headers={"Authorization": _AUTOBEARER},
json={"model": "agnes-3.0-flash",
"messages": [{"role": "user", "content": "hi"}],
"conversation_id": "conv-api-42"})
assert r.status_code == 200, r.text
assert captured["session_id"] == "conv-api-42"
assert not captured["extra"] or "conversation_id" not in captured["extra"]
# ② 不传 → session_id 为 None(由适配器回退随机 UUID)
r2 = client.post("/v1/chat/completions", headers={"Authorization": _AUTOBEARER},
json={"model": "agnes-3.0-flash",
"messages": [{"role": "user", "content": "hi"}]})
assert r2.status_code == 200, r2.text
assert captured["session_id"] is None
finally:
m._get_adapter = orig
# ── 第二十五轮:模型级强制参数(上游硬约束)+ 测速超时/错误文案 ──
#
# 起因:kimi-k2.7-code 上游只接受 temperature=1,测速探针固定发 0 → 每次必 400;
# 而 mimo-v2.5 是慢(偶发冲破 15s 探针超时),却被显示成同样的「调用失败」。
# 两条修法各自的回归如下。
def test_model_params_parsed_from_config():
"""gateway.yaml 的 models[].params 要被解析出来,并随 all_models 暴露。"""
from app.config import parse_config
cfg = parse_config({"providers": {"p": {"base_url": "http://x/v1", "models": [
{"id": "m1", "params": {"temperature": 1, "top_p": 0.9}},
{"id": "m2"}]}}})
assert cfg.providers["p"].models[0].params == {"temperature": 1, "top_p": 0.9}
assert cfg.providers["p"].models[1].params == {}
am = {m["id"]: m for m in cfg.all_models()}
assert am["m1"]["params"] == {"temperature": 1, "top_p": 0.9}
def test_model_params_rejects_bad_keys_and_types():
"""params 的键必须在白名单内:结构性键(model/messages/stream)绝不允许 ——
否则等于让配置文件改写路由与协议语义(回到越权面)。写错必须报错,不能静默失效。"""
from app.config import parse_config, ConfigError
bads = [{"model": "evil"}, {"messages": []}, {"stream": True}, {"temperature": object()}, [1, 2]]
for bad in bads:
try:
parse_config({"providers": {"p": {"base_url": "http://x/v1",
"models": [{"id": "m", "params": bad}]}}})
except ConfigError:
continue
raise AssertionError(f"非法 params 未被拦截: {bad!r}")
def test_real_config_locks_kimi_temperature():
"""真配置口径:kimi-k2.7-code 只接受 temperature=1,必须在配置里写死。"""
cfg = load_config(CONFIG)
m = [x for x in cfg.providers["opencode"].models if x.id == "kimi-k2.7-code"][0]
assert m.params.get("temperature") == 1, m.params
def test_model_params_override_client_value():
"""配了 params 的模型:客户端传什么值都被覆盖,且只发一次请求(不做无谓重发)。"""
from app.adapters.openai_compat import OpenAICompatAdapter
seen = []
def handler(req):
seen.append(json.loads(req.content))
return httpx.Response(200, json={"id": "1", "choices": [], "usage": {}})
ad = OpenAICompatAdapter("t", "http://mock.test/v1",
model_params={"locked": {"temperature": 1}})
ad._client = httpx.AsyncClient(transport=httpx.MockTransport(handler),
base_url="http://mock.test/v1")
async def run():
await ad.chat("locked", [{"role": "user", "content": "hi"}], temperature=0)
await ad.close()
asyncio.run(run())
assert seen[0]["temperature"] == 1, seen[0]
assert len(seen) == 1
def test_model_param_constraint_auto_corrected_and_retried():
"""上游 400 说「只允许 temperature=1」→ 网关按 1 重发一次(而非原样重试)。"""
from app.adapters.openai_compat import OpenAICompatAdapter
seen = []
def handler(req):
seen.append(json.loads(req.content))
if len(seen) == 1:
return httpx.Response(400, json={"error": {"message":
"Error from provider (Console Go): [invalid_request_error] "
"invalid temperature: only 1 is allowed for this model"}})
return httpx.Response(200, json={"id": "1", "choices": [
{"message": {"role": "assistant", "content": "ok"}, "finish_reason": "stop"}], "usage": {}})
ad = OpenAICompatAdapter("t", "http://mock.test/v1")
ad._client = httpx.AsyncClient(transport=httpx.MockTransport(handler),
base_url="http://mock.test/v1")
async def run():
r = await ad.chat("kimi-k2.7-code", [{"role": "user", "content": "hi"}], temperature=0)
await ad.close()
return r
r = asyncio.run(run())
assert len(seen) == 2, seen
assert seen[0]["temperature"] == 0 and seen[1]["temperature"] == 1, seen
assert r["choices"][0]["message"]["content"] == "ok"
def test_model_param_constraint_no_infinite_retry():
"""上游始终 400 → 只修正重发一次就抛错(绝不能与同一约束来回打转)。"""
from app.adapters.openai_compat import OpenAICompatAdapter
seen = []
def handler(req):
seen.append(json.loads(req.content))
return httpx.Response(400, json={"error": {"message":
"invalid temperature: only 1 is allowed for this model"}})
ad = OpenAICompatAdapter("t", "http://mock.test/v1")
ad._client = httpx.AsyncClient(transport=httpx.MockTransport(handler),
base_url="http://mock.test/v1")
async def run():
try:
await ad.chat("m", [{"role": "user", "content": "hi"}], temperature=0)
finally:
await ad.close()
try:
asyncio.run(run())
raise AssertionError("应抛出 httpx.HTTPStatusError")
except httpx.HTTPStatusError:
pass
assert len(seen) == 2, f"应只发 2 次(原始 + 修正),实际 {len(seen)}"
def test_model_param_conflict_respects_configured_value():
"""管理员写死的值与上游要求冲突 → 尊重配置,不偷偷改、也不发无谓的第二枪。"""
from app.adapters.openai_compat import OpenAICompatAdapter
seen = []
def handler(req):
seen.append(json.loads(req.content))
return httpx.Response(400, json={"error": {"message":
"invalid temperature: only 1 is allowed for this model"}})
ad = OpenAICompatAdapter("t", "http://mock.test/v1", model_params={"m": {"temperature": 0.5}})
ad._client = httpx.AsyncClient(transport=httpx.MockTransport(handler),
base_url="http://mock.test/v1")
async def run():
try:
await ad.chat("m", [{"role": "user", "content": "hi"}], temperature=0.5)
finally:
await ad.close()
try:
asyncio.run(run())
raise AssertionError("应抛出 httpx.HTTPStatusError")
except httpx.HTTPStatusError:
pass
assert len(seen) == 1, seen
assert seen[0]["temperature"] == 0.5, seen[0]
def test_stream_param_constraint_auto_corrected():
"""流式路径同样自修复:开局 400 → 按上游要求修正后重发。"""
from app.adapters.openai_compat import OpenAICompatAdapter
seen = []
def handler(req):
seen.append(json.loads(req.content))
if len(seen) == 1:
return httpx.Response(400, json={"error": {"message":
"invalid temperature: only 1 is allowed for this model"}})
return httpx.Response(200, text='data: {"choices":[{"delta":{"content":"hi"}}]}\n\ndata: [DONE]\n\n',
headers={"content-type": "text/event-stream"})
ad = OpenAICompatAdapter("t", "http://mock.test/v1")
ad._client = httpx.AsyncClient(transport=httpx.MockTransport(handler),
base_url="http://mock.test/v1")
async def run():
out = [c async for c in ad.chat_stream("kimi-k2.7-code",
[{"role": "user", "content": "hi"}], temperature=0)]
await ad.close()
return out
out = asyncio.run(run())
assert len(seen) == 2, seen
assert seen[0]["temperature"] == 0 and seen[1]["temperature"] == 1, seen
assert out, "流式应至少收到一个数据块"
def test_speedtest_error_text_surfaces_timeout_and_upstream_detail():
"""测速失败文案:超时要说清是超时,400 要透出上游原文。
(本轮排查时,"参数被上游拒绝" 与 "模型太慢" 在旧文案里长得一模一样。)"""
import app.main as m
t = m._speedtest_error_text(httpx.ReadTimeout("timed out"), 30.0)
assert "超时" in t and "30" in t, t
req = httpx.Request("POST", "http://x/v1/chat/completions")
resp = httpx.Response(400, request=req, json={"error": {
"message": "invalid temperature: only 1 is allowed for this model"}})
e = httpx.HTTPStatusError("Client error '400 Bad Request'", request=req, response=resp)
t2 = m._speedtest_error_text(e, 30.0)
assert "only 1 is allowed" in t2, t2
def test_settings_panel_preserves_model_params():
"""管理页面保存配置时,手写在 YAML 里的 models[].params 不能被面板冲掉。
(面板的模型编辑器只认 id/display_name/type/vision,照它重建就会静默丢掉 params。)"""
from app import settings as gw_settings
from app.config import parse_config
raw = {"providers": {"p": {"name": "P", "base_url": "http://x/v1", "models": [
{"id": "m", "display_name": "M", "type": "text", "vision": False,
"params": {"temperature": 1}}]}}}
cfg = parse_config(raw)
view = gw_settings.config_to_view(cfg)
# 面板读出来就带着 params
assert view["providers"]["p"]["models"][0]["params"] == {"temperature": 1}
# 提交回去(模拟面板原样回传)→ params 必须还在
out = gw_settings.view_to_raw(cfg, raw, {"providers": view["providers"]})
assert out["providers"]["p"]["models"][0]["params"] == {"temperature": 1}, \
out["providers"]["p"]["models"][0]
# 直接运行(非 pytest)
if __name__ == "__main__":
fns = [v for k, v in list(globals().items()) if k.startswith("test_") and callable(v)]