上游会话 id:客户端显式带 x-opencode-session 时优先用它

老板要求:「x-opencode-session 如果 api 请求有带的话优先用请求上的,没有再用 apikey 的那个方式」。

取值优先级(_upstream_session):
  ① 入站 x-opencode-session 头(本轮新增) → sha256("s:{账号}:{客户端值}")[:32]
  ② conversation_id                          → sha256("{账号}:{conversation_id}")[:32]
  ③ 用这个 API Key 自己(一个 Key 一个稳定会话)→ sha256("k:{去前缀Key}")[:32]
  ④ 登录页会话(无 Key)                      → sha256("u:{账号}")[:32]

- 仍按账号哈希:保留第二十六轮的账号隔离(两个账号写同一个值也不会被绑到同一上游会话标识)
- 入站值过 conversation_id 同款入口校验(仅 ASCII、≤128 字符),非法值静默丢弃并回退
  (可选头不该把请求打回 400);三个调用点(直传 / managed 非流式 / managed 流式)全接
- 新增 _inbound_session(request) 与 SESSION_HEADER 常量

实测(生产):同一会话头连打两次同一段 4k 前缀 → cached 0 → 1920(稳定复用);不带头回退
Key 派生也命中。同时更正一条旧认知:上游缓存**并非严格按 session 隔离**(换会话头同前缀
照样命中 1920),会话 id 是「纳入缓存」的前提而非隔离边界 —— README 与代码注释已同步。

测试 204 → 206 全过(新增取值优先级+账号隔离、入站头端到端两条);ci_check 全绿。
This commit is contained in:
api-gateway committed 2026-09-28 19:35:59 +08:00
1 parent 3b35b14a8a
commit c68a76cb8c
4 files changed
+190 -15

No files matched your search

+86
View File
@@ -5473,6 +5473,92 @@ def test_metrics_record_tokens_breakdown_accumulates():
assert t["cached"] <= t["prompt"] and t["reasoning"] <= t["completion"]
def test_upstream_session_priority_and_account_isolation():
"""第三十五轮:上游会话 id 的取值优先级 + 账号隔离。
① 入站 `x-opencode-session` 优先(客户端最清楚哪几次算同一个会话);
② 没带才依次回退 conversation_id → API Key → 账号;
③ 带了客户端会话时仍**按账号哈希** —— 否则两个账号用同一个值就会共用上游缓存作用域
(第二十六轮定下的账号隔离被绕过)。
"""
import hashlib
import app.main as m
h = lambda x: hashlib.sha256(x.encode("utf-8")).hexdigest()[:32] # noqa: E731
# ① 客户端显式会话优先,且不受 conversation_id 影响
a = m._upstream_session(None, "acct", "tok", "conv-abc")
b = m._upstream_session("conv-xyz", "acct", "tok", "conv-abc")
assert a == b, "带了入站会话头时不该再被 conversation_id 改写"
assert a == h("s:acct:conv-abc"), a
# ③ 同值不同账号必须隔离
assert a != m._upstream_session(None, "other", "tok", "conv-abc"), "账号隔离不能绕过"
# ② 没带则按原优先级回退
assert m._upstream_session("c1", "acct", "tok") == h("acct:c1")
assert m._upstream_session(None, "acct", "tok") == h("k:tok")
assert m._upstream_session(None, "acct", None) == h("u:acct")
# 都稳定(缓存要靠它),且都是 32 位十六进制
for v in (a, m._upstream_session("c1", "acct", "tok"), m._upstream_session(None, "acct", None)):
assert len(v) == 32 and all(c in "0123456789abcdef" for c in v), v
def test_inbound_opencode_session_header_reaches_upstream():
"""第三十五轮端到端:入站的 `x-opencode-session` 真的决定了发给上游的会话 id。
- 同一个头 → 稳定(缓存前提);
- 换头 → 换会话;
- 不带头 → 回退到 Key 派生(结果与带头不同);
- 非法头(非 ASCII / 过长)→ 丢弃并回退,绝不把脏值带去上游。
"""
from fastapi.testclient import TestClient
import app.main as m
seen: list[str | None] = []
class FakeAdapter:
async def chat_stream(self, model=None, messages=None, session_id=None, **kw):
seen.append(session_id)
yield {"id": "c1", "choices": [{"index": 0, "delta": {"content": "hi"}}]}
yield {"done": True}
orig = m._get_adapter
m._get_adapter = lambda model_id, timeout=None, cfg=None: (
FakeAdapter(), "agnes", m._cfg().providers["agnes"])
body = {"model": "agnes-3.0-flash", "messages": [{"role": "user", "content": "hi"}],
"stream": True}
try:
client = TestClient(m.app)
def run(headers):
with client.stream("POST", "/v1/chat/completions", headers=headers, json=body) as r:
assert r.status_code == 200, r.read()
"".join(r.iter_text())
run({"x-opencode-session": "conv-A"})
run({"x-opencode-session": "conv-A"})
run({"x-opencode-session": "conv-B"})
run({})
run({"x-opencode-session": "x" * 500}) # 过长(ASCII)→ 丢弃并回退
assert len(seen) == 5, seen
assert seen[0] == seen[1], "同一个入站会话头必须映射到稳定的上游会话"
assert seen[0] != seen[2], "换会话头应换上游会话"
assert seen[0] != seen[3], "带了会话头就不该再用 Key 派生"
assert seen[4] == seen[3], "过长会话头应等同没带(回退到派生)"
assert all(isinstance(x, str) and len(x) == 32 for x in seen), seen
# 非 ASCII 值在 HTTP 层就发不出去(httpx 直接拒),所以用桩验门口的校验:
class _S:
rid = "t"
class _Req:
headers = {"x-opencode-session": "中文会话"}
state = _S()
assert m._inbound_session(_Req()) is None, "非 ASCII 会话头必须被丢弃(不能带进上游 header)"
finally:
m._get_adapter = orig
# 直接运行(非 pytest)—— 必须留在文件末尾,见上方 `test_image_request_goes_to_requested_model`
# 之前的旁注:运行器只会收集执行到这一行时已经存在的 globals。
if __name__ == "__main__":