mirror of
https://gitcode.com/JianFeeeee/ModelRouter.git
synced 2026-11-12 20:27:23 +00:00
feat(types): 统一格式落地为内核实际使用的类型 —— 私有字段泄漏归零
Phase 2 收尾（入向）。ChatRequest 从"客户端发什么就是什么"改为
UnifiedRequest（最大集合）的别名，内核真正用上了那个超集类型。
改掉的具体缺陷：
1. 私有字段泄漏归零（v1 的核心问题）
v1 的 disable_thinking 是内核私有开关，却被序列化进交给适配器的 body，
迫使 10 个适配器各自写 req.disable_thinking = nil 来清理。
现在它是中立的 thinking 字段（{enabled, budget_tokens}），客户端与
适配器都认这个词：
- gateway/chat.go：thinkingFromClient() 把 disable_thinking 翻译成
thinking.enabled=false；冲突时中立字段优先（它能带预算，布尔不能）
- deepseek.lua：thinking.enabled=false → extra_body.thinking.type=disabled
- anthropic.lua：thinking.enabled=true + budget → 原生 {type,budget_tokens}
（enabled=false 时不传该字段——Anthropic 没有 type=disabled 这个写法）
- 其余 7 个适配器：把"清理内核私有字段"改成"清理中立字段"
2. Tools/ToolChoice 改 json.RawMessage
原为 []interface{}（OpenAI 原始形状 = 内核在假设某个协议）。
内核从不解释工具 schema——它原样交给适配器，所以原始形状不损失任何
内核侧统一性。ToolDef 结构化类型保留在包内供新适配器选用。
3. 删除 MarshalJSON 后门
v1 手工 splice ExtraBody 进 OpenAI 形状的 body；现在 ExtraBody 是
UnifiedRequest 上一个正正经经声明的字段。
渐进迁移：ChatRequest 保留为别名（不是新类型），49 处引用零改动编译，
语义变更不被机械改名淹没。
测试：新增 thinking_test.go（翻译链 4 例 + 泄漏回归 + tools 透传）；
vm_test.go 两处迁移到中立语义。golden 字节级未漂移。
This commit is contained in:
JianFeeeee
@ -21,21 +21,79 @@ import (
|
||||
)
|
||||
|
||||
// chatRequest mirrors the OpenAI chat completions request the gateway accepts.
|
||||
// chatRequest is the gateway's IN-BOUND client shape for /v1/chat/completions.
|
||||
//
|
||||
// It is deliberately an OpenAI wire struct, NOT the internal unified format:
|
||||
// this is one client protocol, and the kernel parses it before anything else
|
||||
// knows which protocol the rest of the request will speak. What reaches an
|
||||
// adapter is types.UnifiedRequest (the maximal set), which Thinking populates.
|
||||
type chatRequest struct {
|
||||
Model string `json:"model"`
|
||||
Messages []types.ChatMessage `json:"messages"`
|
||||
Temperature *float64 `json:"temperature,omitempty"`
|
||||
MaxTokens int `json:"max_tokens,omitempty"`
|
||||
Stream bool `json:"stream,omitempty"`
|
||||
Tools []interface{} `json:"tools,omitempty"`
|
||||
ToolChoice interface{} `json:"tool_choice,omitempty"`
|
||||
Model string `json:"model"`
|
||||
Messages []types.ChatMessage `json:"messages"`
|
||||
Temperature *float64 `json:"temperature,omitempty"`
|
||||
MaxTokens int `json:"max_tokens,omitempty"`
|
||||
Stream bool `json:"stream,omitempty"`
|
||||
Tools []interface{} `json:"tools,omitempty"`
|
||||
ToolChoice interface{} `json:"tool_choice,omitempty"`
|
||||
// DisableThinking is the OpenAI-ecosystem spelling of "turn thinking off".
|
||||
// It is translated into the protocol-neutral Thinking field on the way in
|
||||
// (see thinkingFromClient), so that flag never travels under this name.
|
||||
DisableThinking bool `json:"disable_thinking"`
|
||||
ExtraBody map[string]interface{} `json:"extra_body,omitempty"`
|
||||
|
||||
// Thinking accepts the neutral shape directly, so a client already speaking
|
||||
// the unified format (or an adapter-served protocol that names it that way)
|
||||
// does not have to round-trip through disable_thinking.
|
||||
Thinking *types.ThinkingConfig `json:"thinking,omitempty"`
|
||||
|
||||
// PromptCacheKey 是 OpenAI 原生的缓存键；部分客户端用它携带会话标识。
|
||||
PromptCacheKey string `json:"prompt_cache_key,omitempty"`
|
||||
}
|
||||
|
||||
// thinkingFromClient folds the client's spelling of "no thinking" into the
|
||||
// neutral Thinking field.
|
||||
//
|
||||
// Precedence: an explicit `thinking` object wins, because it carries a budget
|
||||
// the boolean cannot express. disable_thinking only ever sets Enabled=false —
|
||||
// it says nothing about how much, and inventing a budget there would silently
|
||||
// override the upstream's default.
|
||||
func (r *chatRequest) thinkingFromClient() *types.ThinkingConfig {
|
||||
if r.Thinking != nil {
|
||||
return r.Thinking
|
||||
}
|
||||
if r.DisableThinking {
|
||||
no := false
|
||||
return &types.ThinkingConfig{Enabled: &no}
|
||||
}
|
||||
return nil
|
||||
}
|
||||
|
||||
// rawTools / rawToolChoice re-encode the loosely-typed client fields into the
|
||||
// RawMessage the internal format carries. The kernel never interprets them:
|
||||
// they are the client's payload, passed through to the adapter, which is the
|
||||
// only layer that knows the tool schema of any particular protocol.
|
||||
func (r *chatRequest) rawTools() json.RawMessage {
|
||||
if len(r.Tools) == 0 {
|
||||
return nil
|
||||
}
|
||||
b, err := json.Marshal(r.Tools)
|
||||
if err != nil {
|
||||
return nil
|
||||
}
|
||||
return b
|
||||
}
|
||||
|
||||
func (r *chatRequest) rawToolChoice() json.RawMessage {
|
||||
if r.ToolChoice == nil {
|
||||
return nil
|
||||
}
|
||||
b, err := json.Marshal(r.ToolChoice)
|
||||
if err != nil {
|
||||
return nil
|
||||
}
|
||||
return b
|
||||
}
|
||||
|
||||
// ChatCompletion is the non-streaming OpenAI response object.
|
||||
type ChatCompletion struct {
|
||||
ID string `json:"id"`
|
||||
@ -424,15 +482,15 @@ func (g *Gateway) handleChat(w http.ResponseWriter, r *http.Request) {
|
||||
done := g.stats.Begin()
|
||||
defer done()
|
||||
inner := &types.ChatRequest{
|
||||
Messages: req.Messages,
|
||||
Temperature: req.Temperature,
|
||||
MaxTokens: req.MaxTokens,
|
||||
Stream: req.Stream,
|
||||
Tools: req.Tools,
|
||||
ToolChoice: req.ToolChoice,
|
||||
DisableThinking: req.DisableThinking,
|
||||
ExtraBody: req.ExtraBody,
|
||||
ClientSession: clientSession,
|
||||
Messages: req.Messages,
|
||||
Temperature: req.Temperature,
|
||||
MaxTokens: req.MaxTokens,
|
||||
Stream: req.Stream,
|
||||
Tools: req.rawTools(),
|
||||
ToolChoice: req.rawToolChoice(),
|
||||
Thinking: req.thinkingFromClient(),
|
||||
ExtraBody: req.ExtraBody,
|
||||
ClientSession: clientSession,
|
||||
}
|
||||
rec := &Req{
|
||||
Key: keyID(reqKey(ctx)),
|
||||
@ -484,16 +542,16 @@ func (g *Gateway) handleChat(w http.ResponseWriter, r *http.Request) {
|
||||
defer done()
|
||||
|
||||
inner := &types.ChatRequest{
|
||||
Model: effective,
|
||||
Messages: req.Messages,
|
||||
Temperature: req.Temperature,
|
||||
MaxTokens: req.MaxTokens,
|
||||
Stream: req.Stream,
|
||||
Tools: req.Tools,
|
||||
ToolChoice: req.ToolChoice,
|
||||
DisableThinking: req.DisableThinking,
|
||||
ExtraBody: req.ExtraBody,
|
||||
ClientSession: clientSession,
|
||||
Model: effective,
|
||||
Messages: req.Messages,
|
||||
Temperature: req.Temperature,
|
||||
MaxTokens: req.MaxTokens,
|
||||
Stream: req.Stream,
|
||||
Tools: req.rawTools(),
|
||||
ToolChoice: req.rawToolChoice(),
|
||||
Thinking: req.thinkingFromClient(),
|
||||
ExtraBody: req.ExtraBody,
|
||||
ClientSession: clientSession,
|
||||
}
|
||||
rec := &Req{
|
||||
Key: keyID(reqKey(ctx)),
|
||||
@ -586,7 +644,7 @@ func estimatePromptTokens(req *types.ChatRequest) int64 {
|
||||
}
|
||||
b, _ := json.Marshal(struct {
|
||||
Messages []types.ChatMessage `json:"messages"`
|
||||
Tools []interface{} `json:"tools,omitempty"`
|
||||
Tools json.RawMessage `json:"tools,omitempty"`
|
||||
}{Messages: req.Messages, Tools: req.Tools})
|
||||
if len(b) == 0 {
|
||||
return 0
|
||||
|
||||
133
internal/gateway/thinking_test.go
Normal file
133
internal/gateway/thinking_test.go
Normal file
@ -0,0 +1,133 @@
|
||||
package gateway
|
||||
|
||||
import (
|
||||
"encoding/json"
|
||||
"testing"
|
||||
|
||||
"llmsproxy/internal/types"
|
||||
)
|
||||
|
||||
// TestThinkingFromClient 把「客户端措辞 → 中立语义」这条翻译链钉死。
|
||||
//
|
||||
// v1 的问题：disable_thinking 是内核私有字段，被序列化进交给适配器的 body，
|
||||
// 于是 10 个适配器各自写 req.disable_thinking = nil 来清理它。v2 把它换成
|
||||
// 双方都认可的 thinking 字段，这条测试确保翻译只发生一次、且在适配器之前。
|
||||
func TestThinkingFromClient(t *testing.T) {
|
||||
cases := []struct {
|
||||
name string
|
||||
req chatRequest
|
||||
want *types.ThinkingConfig
|
||||
}{
|
||||
{
|
||||
name: "无任何字段 → 不带 thinking（上游用自己默认）",
|
||||
req: chatRequest{},
|
||||
want: nil,
|
||||
},
|
||||
{
|
||||
name: "disable_thinking=true → enabled=false",
|
||||
req: chatRequest{DisableThinking: true},
|
||||
want: &types.ThinkingConfig{Enabled: boolPtr(false)},
|
||||
},
|
||||
{
|
||||
name: "中立 thinking 显式给出 → 原样透传",
|
||||
req: chatRequest{Thinking: &types.ThinkingConfig{Enabled: boolPtr(true), BudgetTokens: 4096}},
|
||||
want: &types.ThinkingConfig{Enabled: boolPtr(true), BudgetTokens: 4096},
|
||||
},
|
||||
{
|
||||
name: "两者冲突时，中立的 thinking 优先（它带预算，布尔表达不了）",
|
||||
req: chatRequest{
|
||||
DisableThinking: true,
|
||||
Thinking: &types.ThinkingConfig{Enabled: boolPtr(true), BudgetTokens: 1024},
|
||||
},
|
||||
want: &types.ThinkingConfig{Enabled: boolPtr(true), BudgetTokens: 1024},
|
||||
},
|
||||
}
|
||||
for _, c := range cases {
|
||||
t.Run(c.name, func(t *testing.T) {
|
||||
got := c.req.thinkingFromClient()
|
||||
if c.want == nil {
|
||||
if got != nil {
|
||||
t.Fatalf("want nil, got %+v", got)
|
||||
}
|
||||
return
|
||||
}
|
||||
if got == nil {
|
||||
t.Fatal("want non-nil thinking, got nil")
|
||||
}
|
||||
if (got.Enabled == nil) != (c.want.Enabled == nil) {
|
||||
t.Fatalf("Enabled presence mismatch: got %v want %v", got.Enabled, c.want.Enabled)
|
||||
}
|
||||
if got.Enabled != nil && *got.Enabled != *c.want.Enabled {
|
||||
t.Errorf("Enabled = %v, want %v", *got.Enabled, *c.want.Enabled)
|
||||
}
|
||||
if got.BudgetTokens != c.want.BudgetTokens {
|
||||
t.Errorf("BudgetTokens = %d, want %d", got.BudgetTokens, c.want.BudgetTokens)
|
||||
}
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
// TestDisableThinkingNeverReachesAdapterBody 是那条泄漏的回归测试：内核私有
|
||||
// 字段不得出现在交给适配器的 body 里。
|
||||
func TestDisableThinkingNeverReachesAdapterBody(t *testing.T) {
|
||||
var r chatRequest
|
||||
if err := json.Unmarshal([]byte(`{"model":"m","messages":[],"disable_thinking":true}`), &r); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
inner := &types.ChatRequest{
|
||||
Model: "m",
|
||||
Thinking: r.thinkingFromClient(),
|
||||
Tools: r.rawTools(),
|
||||
ToolChoice: r.rawToolChoice(),
|
||||
ExtraBody: r.ExtraBody,
|
||||
ClientSession: "",
|
||||
}
|
||||
body, err := json.Marshal(inner)
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
var m map[string]json.RawMessage
|
||||
if err := json.Unmarshal(body, &m); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if _, leaked := m["disable_thinking"]; leaked {
|
||||
t.Errorf("内核私有字段 disable_thinking 泄漏进适配器 body（v1 回归）: %s", body)
|
||||
}
|
||||
if _, ok := m["thinking"]; !ok {
|
||||
t.Errorf("中立 thinking 字段缺失: %s", body)
|
||||
}
|
||||
}
|
||||
|
||||
// TestToolsPassthroughRaw 确认 tools/tool_choice 以原始形状交给适配器：
|
||||
// 内核不解释工具 schema，那是适配器的职责。
|
||||
func TestToolsPassthroughRaw(t *testing.T) {
|
||||
var r chatRequest
|
||||
body := `{"model":"m","messages":[],"tools":[{"type":"function","function":{"name":"f","parameters":{"type":"object"}}}],"tool_choice":"auto"}`
|
||||
if err := json.Unmarshal([]byte(body), &r); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
tools := r.rawTools()
|
||||
if len(tools) == 0 {
|
||||
t.Fatal("tools 未透传")
|
||||
}
|
||||
var decoded []map[string]interface{}
|
||||
if err := json.Unmarshal(tools, &decoded); err != nil {
|
||||
t.Fatalf("tools 不是合法 JSON: %v (%s)", err, tools)
|
||||
}
|
||||
if len(decoded) != 1 || decoded[0]["type"] != "function" {
|
||||
t.Errorf("tools 形状被内核改写: %s", tools)
|
||||
}
|
||||
if tc := string(r.rawToolChoice()); tc != `"auto"` {
|
||||
t.Errorf("tool_choice = %s, want \"auto\"", tc)
|
||||
}
|
||||
// 空值不应变成 "null" 出现在 body 里
|
||||
var empty chatRequest
|
||||
if got := empty.rawTools(); got != nil {
|
||||
t.Errorf("无 tools 时应返回 nil, got %s", got)
|
||||
}
|
||||
if got := empty.rawToolChoice(); got != nil {
|
||||
t.Errorf("无 tool_choice 时应返回 nil, got %s", got)
|
||||
}
|
||||
}
|
||||
|
||||
func boolPtr(v bool) *bool { return &v }
|
||||
Reference in New Issue
Block a user