mirror of
https://gitcode.com/JianFeeeee/ModelRouter.git
synced 2026-11-07 17:57:23 +00:00
Compare commits
1 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
| 2b8fcaf434 |
112
internal/lua/adapter_opencodezen_reasoning_test.go
Normal file
112
internal/lua/adapter_opencodezen_reasoning_test.go
Normal file
@ -0,0 +1,112 @@
|
||||
package lua
|
||||
|
||||
import (
|
||||
"encoding/json"
|
||||
"strings"
|
||||
"testing"
|
||||
)
|
||||
|
||||
// loadZenVM 装载 opencodezen 适配器(照 adapter_opencodezen_test.go 的方式)。
|
||||
func loadZenVM(t *testing.T) *VM {
|
||||
t.Helper()
|
||||
src, err := bundledAdapters.ReadFile("adapters/opencodezen.lua")
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
vm := NewVM(t.TempDir())
|
||||
t.Cleanup(vm.Stop)
|
||||
if err := vm.LoadAdapterSource("opencodezen", string(src)); err != nil {
|
||||
t.Fatalf("加载失败: %v", err)
|
||||
}
|
||||
return vm
|
||||
}
|
||||
|
||||
// ★ assistant 回合的 reasoning_content **必须回传**给 opencode 上游。
|
||||
//
|
||||
// 真实缺陷(生产实测,2026-10-16):
|
||||
//
|
||||
// provider localzen: api error 400: Upstream request failed: [invalid_request_error]
|
||||
// The `reasoning_content` in the thinking mode must be passed back to the API.
|
||||
//
|
||||
// 根因:transform_request 里写了 `msg.reasoning_content = nil`,理由是
|
||||
// "Zen 免费池不接受回传"。该假设**方向反了** —— 上游要的就是它。
|
||||
//
|
||||
// ★★ 这个缺陷极难归因,因为失败是**交替**出现的:
|
||||
// 只有当上一轮**真的产出了 reasoning**(thinking 模型 + reasoning_effort 开启)
|
||||
// 时,回放那段历史才会被拒;普通回合照常 200。审计里表现为成串 502 紧随一个
|
||||
// 成功的长上下文请求(实测 prompt_tokens=139330)之后。
|
||||
//
|
||||
// 判据必须**咬住「被剥掉」这个具体行为**,而不是"请求成功" —— 后者在
|
||||
// 非 thinking 回合下永远为真,测不出这个缺陷。
|
||||
func TestOpenCodeZenKeepsAssistantReasoningContent(t *testing.T) {
|
||||
vm := loadZenVM(t)
|
||||
|
||||
body := map[string]interface{}{
|
||||
"model": "deepseek-v4.1-flash-free",
|
||||
"stream": true,
|
||||
"messages": []map[string]interface{}{
|
||||
{"role": "user", "content": "1+1=?"},
|
||||
{
|
||||
"role": "assistant",
|
||||
"content": "2",
|
||||
// 这段正是 agent 回放历史时会带上的东西
|
||||
"reasoning_content": "一加一等于二。",
|
||||
},
|
||||
{"role": "user", "content": "再加一"},
|
||||
},
|
||||
}
|
||||
raw, err := json.Marshal(body)
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
out, err := vm.Transform("opencodezen", "transform_request", string(raw))
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
|
||||
// 解析出站体,找 assistant 回合
|
||||
var sent struct {
|
||||
Messages []map[string]interface{} `json:"messages"`
|
||||
}
|
||||
if err := json.Unmarshal([]byte(out), &sent); err != nil {
|
||||
t.Fatalf("出站体不是合法 JSON: %v\n%s", err, out)
|
||||
}
|
||||
var found bool
|
||||
for _, m := range sent.Messages {
|
||||
if m["role"] != "assistant" {
|
||||
continue
|
||||
}
|
||||
found = true
|
||||
rc, _ := m["reasoning_content"].(string)
|
||||
if rc == "" {
|
||||
t.Fatalf("assistant 回合的 reasoning_content 被剥掉了 ⇒ 上游会回 "+
|
||||
"`The reasoning_content in the thinking mode must be passed back to the API`\n"+
|
||||
"出站 messages: %s", out)
|
||||
}
|
||||
if rc != "一加一等于二。" {
|
||||
t.Errorf("reasoning_content 被改写:%q", rc)
|
||||
}
|
||||
}
|
||||
if !found {
|
||||
t.Fatalf("没找到 assistant 回合(判据失效)\n%s", out)
|
||||
}
|
||||
t.Log("✓ assistant 的 reasoning_content 已保留")
|
||||
}
|
||||
|
||||
// ★ 反向:**空的** reasoning_content 应被丢掉(不要把空串发给上游)。
|
||||
// 与上一条配对 —— 缺了这条,"改成无条件保留"这类过度修正不会被发现。
|
||||
func TestOpenCodeZenDropsEmptyReasoningContent(t *testing.T) {
|
||||
vm := loadZenVM(t)
|
||||
|
||||
raw := `{"model":"m","stream":true,"messages":[
|
||||
{"role":"assistant","content":"hi","reasoning_content":""}
|
||||
]}`
|
||||
out, err := vm.Transform("opencodezen", "transform_request", raw)
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if strings.Contains(out, `"reasoning_content":""`) {
|
||||
t.Errorf("空的 reasoning_content 不该发给上游(多余字段,部分上游会挑),实得: %s", out)
|
||||
}
|
||||
t.Log("✓ 空 reasoning_content 已丢弃")
|
||||
}
|
||||
@ -11,8 +11,14 @@ adapter.endpoint = "/chat/completions"
|
||||
--
|
||||
-- * 本适配器(Zen)服务免费池 / 按量付费池。免费模型必须带 opencode 客户端
|
||||
-- 指纹(UA + x-opencode-*),否则上游拒绝:"free tier can only be used in
|
||||
-- OpenCode"。Zen 的免费池以非 thinking 模型为主,历史上回传
|
||||
-- reasoning_content 会引出上游报错,故这里剥掉(Go 侧相反,见 opencodego.lua)。
|
||||
-- OpenCode"。
|
||||
-- ★ reasoning_content:**保留**(与 Go 侧一致)。此前这里写「剥掉」是基于
|
||||
-- 「免费池以非 thinking 模型为主、回传会报错」的假设,**生产实测已证伪** ——
|
||||
-- 剥掉恰恰会引发
|
||||
-- The `reasoning_content` in the thinking mode must be passed back to the API
|
||||
-- (详见 transform_request 内的长注释:失败只在「上一轮真的产出了 reasoning」
|
||||
-- 时才出现,故表现为同一会话里成功/失败交替,极难归因)。
|
||||
-- 参考的共享序列化器(openai-compat.ts)也是「有就回传」。
|
||||
-- * Go 订阅读者请看 opencodego.lua:那边**必须**回传 reasoning_content,
|
||||
-- 否则 thinking 模式直接 400。
|
||||
--
|
||||
@ -219,26 +225,50 @@ function adapter.transform_request(raw_body)
|
||||
if type(msg.role) == "string" and not ROLE_WHITELIST[msg.role] then
|
||||
msg.role = "system"
|
||||
end
|
||||
-- Zen 免费池不接受回传 reasoning_content(Go 侧相反)
|
||||
msg.reasoning_content = nil
|
||||
-- ★ 必须**保留** reasoning_content(不再剥离)。
|
||||
--
|
||||
-- 历史:这里原先写 `msg.reasoning_content = nil`,理由是「Zen 免费池
|
||||
-- 不接受回传」。**生产实测证伪**(2026-10-16):
|
||||
-- provider localzen: api error 400: Upstream request failed:
|
||||
-- [invalid_request_error] The `reasoning_content` in the thinking
|
||||
-- mode must be passed back to the API.
|
||||
-- 失败形态很隐蔽:**同一模型同一会话里成功与失败交替出现** ——
|
||||
-- 因为只有当上一轮**真的产出了 reasoning**(thinking 模型、
|
||||
-- reasoning_effort 开启、长上下文)时,回放该历史才会被拒;
|
||||
-- 剥掉之后,客户端看到的 502 与「上一轮想过」这件事毫无关联。
|
||||
-- 审计文件里能看到成串 502 紧跟在一个 prompt_tokens=139330 的
|
||||
-- 成功请求之后 —— 那正是「想了、但下一轮我们没把它还回去」。
|
||||
--
|
||||
-- 参考实现(`openai-compat.ts` 的 serializeMessages,opencode 适配器
|
||||
-- 直接复用)在 assistant 回合携带 reasoning 时**就带上**该字段:
|
||||
-- `...reasoning.length > 0 ? { reasoning_content: reasoning } : {}`
|
||||
-- 即「有就回传」,没有额外条件 —— 与本适配器此前的假设方向相反。
|
||||
--
|
||||
-- ⚠️ 不改成「仅当 thinking 开启才回传」:我们**无法**从请求里可靠
|
||||
-- 判定上游是否处于 thinking 模式(真正知道的是上游)。而「有就回传」
|
||||
-- 是零风险方向 —— 上游不需要时忽略即可(实测传与不传同为 200),
|
||||
-- 需要时缺了就是硬 400。
|
||||
if type(msg.reasoning_content) == "string" and msg.reasoning_content == "" then
|
||||
msg.reasoning_content = nil
|
||||
end
|
||||
local drop = false
|
||||
if type(msg.content) == "table" then
|
||||
local parts = {}
|
||||
for _, part in ipairs(msg.content) do
|
||||
if type(part) == "table" and part.type ~= nil and part.type ~= "text" then
|
||||
-- multimodal part not supported by zen
|
||||
-- multimodal part not supported by zen
|
||||
else
|
||||
table.insert(parts, part)
|
||||
end
|
||||
end
|
||||
if #parts == 0 then
|
||||
-- Content collapsed to nothing after stripping unsupported
|
||||
-- parts. A message that still carries a tool call must
|
||||
-- NEVER be dropped: the very next message is its tool
|
||||
-- result, and dropping the call orphans that result. The
|
||||
-- model then sees a result for a call it never made and
|
||||
-- re-issues the same tool call on every turn (observed as
|
||||
-- an infinite "repeated tool call" loop).
|
||||
-- Content collapsed to nothing after stripping unsupported
|
||||
-- parts. A message that still carries a tool call must
|
||||
-- NEVER be dropped: the very next message is its tool
|
||||
-- result, and dropping the call orphans that result. The
|
||||
-- model then sees a result for a call it never made and
|
||||
-- re-issues the same tool call on every turn (observed as
|
||||
-- an infinite "repeated tool call" loop).
|
||||
if (type(msg.tool_calls) == "table" and #msg.tool_calls > 0)
|
||||
or msg.tool_call_id ~= nil then
|
||||
msg.content = ""
|
||||
|
||||
Reference in New Issue
Block a user