mirror of
https://gitcode.com/JianFeeeee/ModelRouter.git
synced 2026-10-05 07:02:29 +00:00
gemini.lua 里 adapter.endpoint = "/v1/models",而它自己的注释写的是
POST /v1/models/{model}:generateContent
两者矛盾,而 Go 侧是静态拼接(provider.URL = base_url + endpoint),拼不出
模型名。预置模板 "Google Gemini"(base_url=.../v1beta)于是会 POST 到
https://generativelanguage.googleapis.com/v1beta/v1/models
既多一段 /v1,又缺 :generateContent——那是 Gemini 的模型**列表**端点,
对 POST 返 405。所以任何用户从模板建这个源,拿到的都是必定失败的源。
实测确认影响范围:线上 21 个源里没有 gemini(openai×15 / trae / sensenova /
opencodezen / deepseek / anthropic / agentrouter),所以是潜伏缺陷。
修法:endpoint 改成模板 `/v1beta/models/{model}:generateContent`,新增
provider.ChatURL(model, stream):
- 用 **PathEscape** 替换 {model}——模型 id 进的是 URL 路径,不转义的话
一个 "/" 就会静默指向另一个资源(判据里用 RequestURI 而非 URL.Path
断言,因为后者是解码后的,看不出 %2F);
- 流式把 ":generateContent" 换成 ":streamGenerateContent"(同一个路径、
不同动词,也在路径里)。替换刻意只认这个精确后缀,免得别的适配器
仅仅提到这个词就被改写;
- source 自己设的 endpoint: 仍然优先,模板被整体跳过。
Chat / ChatStream / probeChat 三处调用点改为传本次请求真实的 model——AUTO
按槽位把 req.Model 钉死,所以 URL 必须跟随**请求**的模型,用源默认模型会让
多模型源每次都打同一个(还记到别的模型的账上)。
判定静态 endpoint 的其他 10 个适配器零影响(TestNonGeminiEndpointsAreUntouched)。
顺带:Stats 的 mutex 不是可重入的,导出方法自己加锁、*Locked 后缀要求调用
方持锁。持锁调导出方法会死锁——我的探针真卡死过一次(直到 10 分钟超时)。
补上 LOCKING 注释,并加判据把这条规则钉住(含一个 20 秒上限的行为判据,
让未来的重构撞死锁时快速失败而不是拖满整个套件)。
271 lines
9.6 KiB
Go
271 lines
9.6 KiB
Go
package provider
|
|
|
|
import (
|
|
"context"
|
|
"encoding/json"
|
|
"fmt"
|
|
"io"
|
|
"net/http"
|
|
"net/http/httptest"
|
|
"path/filepath"
|
|
"strings"
|
|
"sync"
|
|
"testing"
|
|
|
|
"llmsproxy/internal/config"
|
|
"llmsproxy/internal/lua"
|
|
"llmsproxy/internal/types"
|
|
)
|
|
|
|
// Gemini is the only adapter whose endpoint carries the model in the PATH
|
|
// (POST /v1beta/models/{model}:generateContent), so it is the only one that
|
|
// needs the substitution this file covers.
|
|
//
|
|
// The defect it fixes was silent: the adapter shipped endpoint="/v1/models",
|
|
// which is Gemini's model-LIST endpoint and answers 405 to POST, so the preset
|
|
// template produced a source that could never work.
|
|
|
|
func geminiVM(t *testing.T) *lua.VM {
|
|
t.Helper()
|
|
// The dir must be named "adapters": the VM seeds its bundled adapters into
|
|
// <dir> on Start and a differently-named dir leaves it empty, which surfaces
|
|
// as "adapter gemini not loaded" rather than an obvious setup error.
|
|
vm := lua.NewVM(filepath.Join(t.TempDir(), "adapters"))
|
|
if err := vm.Start(); err != nil {
|
|
t.Fatalf("lua vm: %v", err)
|
|
}
|
|
t.Cleanup(vm.Stop)
|
|
return vm
|
|
}
|
|
|
|
// geminiProvider builds a provider on the gemini adapter against base.
|
|
func geminiProvider(t *testing.T, base string, models ...string) *Provider {
|
|
t.Helper()
|
|
if len(models) == 0 {
|
|
models = []string{"gemini-3.6-flash"}
|
|
}
|
|
ms := make([]config.Model, 0, len(models))
|
|
for _, m := range models {
|
|
ms = append(ms, config.Model{ID: m, Kind: "chat"})
|
|
}
|
|
return New(config.Source{
|
|
Name: "gem",
|
|
BaseURL: base,
|
|
Adapter: "gemini",
|
|
APIKey: "k",
|
|
MaxConcurrent: 4,
|
|
Models: ms,
|
|
}, geminiVM(t))
|
|
}
|
|
|
|
// geminiRecorder captures the path the gateway actually requested and replies
|
|
// with a minimal Gemini-shaped payload.
|
|
type geminiRecorder struct {
|
|
mu sync.Mutex
|
|
paths []string
|
|
bodies []string
|
|
stream bool
|
|
}
|
|
|
|
func (g *geminiRecorder) handler(t *testing.T) http.HandlerFunc {
|
|
return func(w http.ResponseWriter, r *http.Request) {
|
|
raw, _ := io.ReadAll(r.Body)
|
|
g.mu.Lock()
|
|
// RequestURI, not URL.Path: URL.Path is the DECODED path, so an escaped
|
|
// %2F looks like a real "/" and the assertion could not tell an escaped
|
|
// id from an unescaped one.
|
|
g.paths = append(g.paths, r.RequestURI)
|
|
g.bodies = append(g.bodies, string(raw))
|
|
stream := g.stream
|
|
g.mu.Unlock()
|
|
|
|
if stream {
|
|
w.Header().Set("Content-Type", "text/event-stream")
|
|
w.WriteHeader(200)
|
|
fmt.Fprintf(w, "data: %s\n\n", `{"candidates":[{"content":{"parts":[{"text":"hi"}]},"finishReason":"STOP"}],"usageMetadata":{"promptTokenCount":3,"candidatesTokenCount":1,"totalTokenCount":4}}`)
|
|
fmt.Fprintln(w, "data: [DONE]")
|
|
return
|
|
}
|
|
w.Header().Set("Content-Type", "application/json")
|
|
fmt.Fprint(w, `{"candidates":[{"content":{"parts":[{"text":"pong"}]},"finishReason":"STOP"}],"usageMetadata":{"promptTokenCount":3,"candidatesTokenCount":1,"totalTokenCount":4}}`)
|
|
}
|
|
}
|
|
|
|
func (g *geminiRecorder) lastPath() string {
|
|
g.mu.Lock()
|
|
defer g.mu.Unlock()
|
|
if len(g.paths) == 0 {
|
|
return ""
|
|
}
|
|
return g.paths[len(g.paths)-1]
|
|
}
|
|
|
|
// TestGeminiEndpointCarriesModel: a non-streaming chat must POST to
|
|
// /v1beta/models/<model>:generateContent. The old static endpoint produced
|
|
// "/v1beta/v1/models" (a list endpoint, 405 on POST).
|
|
func TestGeminiEndpointCarriesModel(t *testing.T) {
|
|
rec := &geminiRecorder{}
|
|
srv := httptest.NewServer(rec.handler(t))
|
|
defer srv.Close()
|
|
|
|
p := geminiProvider(t, srv.URL, "gemini-3.6-flash")
|
|
if _, err := p.Chat(context.Background(), &types.ChatRequest{
|
|
Model: "gemini-3.6-flash",
|
|
Messages: []types.ChatMessage{{Role: "user", Content: types.StringContent("hi")}},
|
|
}); err != nil {
|
|
t.Fatalf("chat: %v (path was %q)", err, rec.lastPath())
|
|
}
|
|
if got, want := rec.lastPath(), "/v1beta/models/gemini-3.6-flash:generateContent"; got != want {
|
|
t.Errorf("path = %q, want %q", got, want)
|
|
}
|
|
}
|
|
|
|
// TestGeminiStreamUsesStreamGenerateContent: the streaming verb lives in the
|
|
// path too, so it must be swapped for stream requests only.
|
|
func TestGeminiStreamUsesStreamGenerateContent(t *testing.T) {
|
|
rec := &geminiRecorder{stream: true}
|
|
srv := httptest.NewServer(rec.handler(t))
|
|
defer srv.Close()
|
|
|
|
p := geminiProvider(t, srv.URL, "gemini-3.6-flash")
|
|
ch, err := p.ChatStream(context.Background(), &types.ChatRequest{
|
|
Model: "gemini-3.6-flash",
|
|
Messages: []types.ChatMessage{{Role: "user", Content: types.StringContent("hi")}},
|
|
})
|
|
if err != nil {
|
|
t.Fatalf("stream: %v (path was %q)", err, rec.lastPath())
|
|
}
|
|
for range ch {
|
|
}
|
|
if got, want := rec.lastPath(), "/v1beta/models/gemini-3.6-flash:streamGenerateContent"; got != want {
|
|
t.Errorf("path = %q, want %q", got, want)
|
|
}
|
|
}
|
|
|
|
// TestGeminiModelFollowsRequestNotSourceDefault is the part a naive fix gets
|
|
// wrong: substituting the source's default model would make every request on a
|
|
// multi-model source bill the wrong model. AUTO pins req.Model per slot, so the
|
|
// URL must follow the REQUEST's model.
|
|
func TestGeminiModelFollowsRequestNotSourceDefault(t *testing.T) {
|
|
rec := &geminiRecorder{}
|
|
srv := httptest.NewServer(rec.handler(t))
|
|
defer srv.Close()
|
|
|
|
p := geminiProvider(t, srv.URL, "gemini-3.6-flash", "gemini-3.1-pro")
|
|
for _, m := range []string{"gemini-3.1-pro", "gemini-3.6-flash"} {
|
|
if _, err := p.Chat(context.Background(), &types.ChatRequest{
|
|
Model: m,
|
|
Messages: []types.ChatMessage{{Role: "user", Content: types.StringContent("hi")}},
|
|
}); err != nil {
|
|
t.Fatalf("chat %s: %v", m, err)
|
|
}
|
|
if want := "/v1beta/models/" + m + ":generateContent"; rec.lastPath() != want {
|
|
t.Errorf("path = %q, want %q", rec.lastPath(), want)
|
|
}
|
|
}
|
|
}
|
|
|
|
// TestGeminiModelIsEscaped: a model id reaches a URL path, so an id carrying a
|
|
// slash or space must be escaped rather than silently addressing a different
|
|
// resource (or producing a request Go's http client rejects).
|
|
func TestGeminiModelIsEscaped(t *testing.T) {
|
|
rec := &geminiRecorder{}
|
|
srv := httptest.NewServer(rec.handler(t))
|
|
defer srv.Close()
|
|
|
|
// Configure the odd id as a real model so ModelFor resolves it.
|
|
p := geminiProvider(t, srv.URL, "vendor/model:v1 beta")
|
|
if _, err := p.Chat(context.Background(), &types.ChatRequest{
|
|
Model: "vendor/model:v1 beta",
|
|
Messages: []types.ChatMessage{{Role: "user", Content: types.StringContent("hi")}},
|
|
}); err != nil {
|
|
t.Fatalf("chat: %v", err)
|
|
}
|
|
got := rec.lastPath()
|
|
if strings.Contains(got, "/v1beta/models/vendor/model") {
|
|
t.Errorf("path %q was not escaped: the id's slash changed the resource", got)
|
|
}
|
|
if !strings.HasPrefix(got, "/v1beta/models/") || !strings.HasSuffix(got, ":generateContent") {
|
|
t.Errorf("path = %q, want an escaped id between the prefix and the verb", got)
|
|
}
|
|
}
|
|
|
|
// TestNonGeminiEndpointsAreUntouched: the substitution and the verb rewrite must
|
|
// be no-ops for every adapter whose endpoint is a fixed path. A regression here
|
|
// would silently rewrite OpenAI/Anthropic/… URLs.
|
|
func TestNonGeminiEndpointsAreUntouched(t *testing.T) {
|
|
for _, adapter := range []string{"openai", "deepseek", "anthropic", "ollama", "mistral", "trae", "sensenova"} {
|
|
vm := geminiVM(t)
|
|
p := New(config.Source{
|
|
Name: "s", BaseURL: "https://example.test", Adapter: adapter,
|
|
APIKey: "k", MaxConcurrent: 2,
|
|
Models: []config.Model{{ID: "m1", Kind: "chat"}},
|
|
}, vm)
|
|
for _, stream := range []bool{false, true} {
|
|
got := p.ChatURL("m1", stream)
|
|
want := "https://example.test" + p.Endpoint()
|
|
if got != want {
|
|
t.Errorf("adapter %s stream=%v: URL = %q, want %q", adapter, stream, got, want)
|
|
}
|
|
if strings.Contains(got, "streamGenerateContent") {
|
|
t.Errorf("adapter %s: streaming verb leaked into a fixed endpoint: %q", adapter, got)
|
|
}
|
|
}
|
|
vm.Stop()
|
|
}
|
|
}
|
|
|
|
// TestSourceEndpointOverridesTemplate: a source that spells its own endpoint
|
|
// must win outright — including when it spells one WITHOUT the placeholder.
|
|
func TestSourceEndpointOverridesTemplate(t *testing.T) {
|
|
vm := geminiVM(t)
|
|
defer vm.Stop()
|
|
p := New(config.Source{
|
|
Name: "s", BaseURL: "https://example.test", Adapter: "gemini",
|
|
Endpoint: "/custom/path", APIKey: "k", MaxConcurrent: 2,
|
|
Models: []config.Model{{ID: "m1", Kind: "chat"}},
|
|
}, vm)
|
|
if got, want := p.ChatURL("m1", false), "https://example.test/custom/path"; got != want {
|
|
t.Errorf("URL = %q, want %q (source endpoint must override the adapter template)", got, want)
|
|
}
|
|
// And the streaming verb rewrite must not damage an unrelated path.
|
|
if got := p.ChatURL("m1", true); strings.Contains(got, "streamGenerate") {
|
|
t.Errorf("streaming rewrite leaked into a user-spelled path: %q", got)
|
|
}
|
|
}
|
|
|
|
// TestGeminiResponseStillParses: the endpoint fix must not disturb the body
|
|
// transform — a request that reaches the right URL still has to come back as a
|
|
// unified response with Gemini's usageMetadata mapped.
|
|
func TestGeminiResponseStillParses(t *testing.T) {
|
|
rec := &geminiRecorder{}
|
|
srv := httptest.NewServer(rec.handler(t))
|
|
defer srv.Close()
|
|
|
|
p := geminiProvider(t, srv.URL, "gemini-3.6-flash")
|
|
resp, err := p.Chat(context.Background(), &types.ChatRequest{
|
|
Model: "gemini-3.6-flash",
|
|
Messages: []types.ChatMessage{{Role: "user", Content: types.StringContent("hi")}},
|
|
})
|
|
if err != nil {
|
|
t.Fatalf("chat: %v", err)
|
|
}
|
|
if resp.Content != "pong" {
|
|
t.Errorf("content = %q, want %q", resp.Content, "pong")
|
|
}
|
|
if resp.TokenUsage.Prompt != 3 || resp.TokenUsage.Completion != 1 || resp.TokenUsage.Total != 4 {
|
|
t.Errorf("usage = %+v, want prompt 3 / completion 1 / total 4", resp.TokenUsage)
|
|
}
|
|
// The body the adapter produced must still be Gemini-native.
|
|
rec.mu.Lock()
|
|
body := rec.bodies[0]
|
|
rec.mu.Unlock()
|
|
var sent map[string]interface{}
|
|
if err := json.Unmarshal([]byte(body), &sent); err != nil {
|
|
t.Fatalf("adapter sent invalid json: %v (%s)", err, body)
|
|
}
|
|
if _, ok := sent["contents"]; !ok {
|
|
t.Errorf("adapter did not convert to Gemini's contents[]: %s", body)
|
|
}
|
|
}
|