feat(runtime): 适配推理模式开关

This commit is contained in:
wangbo 2026-07-06 02:03:03 +08:00
parent 797edeedf4
commit b7351f3b9b
10 changed files with 547 additions and 17 deletions

View File

@ -0,0 +1,283 @@
package clients
import (
"fmt"
"math"
"strconv"
"strings"
"github.com/easyai/easyai-ai-gateway/apps/api/internal/store"
)
const OpenAIReasoningEffortValidationMessage = "reasoning_effort must be one of: none, minimal, low, medium, high, xhigh"
var (
openAIReasoningEfforts = map[string]struct{}{
"none": {},
"minimal": {},
"low": {},
"medium": {},
"high": {},
"xhigh": {},
}
volcesChatReasoningEfforts = map[string]struct{}{
"minimal": {},
"low": {},
"medium": {},
"high": {},
}
zhipuReasoningEfforts = map[string]struct{}{
"none": {},
"minimal": {},
"low": {},
"medium": {},
"high": {},
"xhigh": {},
}
)
func ValidateOpenAIReasoningEffort(value any) error {
effort := normalizedReasoningString(value)
if effort == "" {
return nil
}
if isOpenAIReasoningEffort(effort) {
return nil
}
return &ClientError{Code: "invalid_parameter", Message: OpenAIReasoningEffortValidationMessage, Retryable: false}
}
func applyOpenAIChatReasoningParams(body map[string]any, candidate store.RuntimeModelCandidate) {
effort := normalizedReasoningString(body["reasoning_effort"])
if effort == "" || !isOpenAIReasoningEffort(effort) {
return
}
body["reasoning_effort"] = effort
switch {
case isAliyunBailianOpenAI(candidate):
applyAliyunReasoning(body, candidate, effort)
case isDeepSeekOpenAI(candidate):
applyHighMaxThinkingReasoning(body, effort)
case isZhipuOpenAI(candidate):
applyZhipuReasoning(body, candidate, effort)
case isVolcesOpenAI(candidate):
applyVolcesReasoning(body, candidate, effort)
}
}
func applyAliyunReasoning(body map[string]any, candidate store.RuntimeModelCandidate, effort string) {
defer delete(body, "thinking_budget_tokens")
if effort == "none" {
body["enable_thinking"] = false
delete(body, "reasoning_effort")
return
}
body["enable_thinking"] = true
model := chatReasoningModelName(body, candidate)
if isAliyunHighMaxReasoningModel(model) {
body["reasoning_effort"] = highMaxReasoningEffort(effort)
return
}
delete(body, "reasoning_effort")
if isAliyunThinkingBudgetModel(model) {
if budget, ok := positiveIntFromAny(body["thinking_budget_tokens"]); ok {
body["thinking_budget"] = budget
}
}
}
func applyHighMaxThinkingReasoning(body map[string]any, effort string) {
if effort == "none" {
body["thinking"] = map[string]any{"type": "disabled"}
delete(body, "reasoning_effort")
return
}
body["thinking"] = map[string]any{"type": "enabled"}
body["reasoning_effort"] = highMaxReasoningEffort(effort)
}
func applyZhipuReasoning(body map[string]any, candidate store.RuntimeModelCandidate, effort string) {
if effort == "none" {
body["thinking"] = map[string]any{"type": "disabled"}
delete(body, "reasoning_effort")
return
}
body["thinking"] = map[string]any{"type": "enabled"}
if !isZhipuReasoningEffortModel(chatReasoningModelName(body, candidate)) {
delete(body, "reasoning_effort")
return
}
if mapped := zhipuReasoningEffort(effort); mapped != "" {
body["reasoning_effort"] = mapped
return
}
delete(body, "reasoning_effort")
}
func applyVolcesReasoning(body map[string]any, _ store.RuntimeModelCandidate, effort string) {
if effort == "none" {
body["thinking"] = map[string]any{"type": "disabled"}
delete(body, "reasoning_effort")
return
}
body["thinking"] = map[string]any{"type": "enabled"}
if mapped := volcesChatReasoningEffort(effort); mapped != "" {
body["reasoning_effort"] = mapped
return
}
delete(body, "reasoning_effort")
}
func normalizedReasoningString(value any) string {
if value == nil {
return ""
}
text, ok := value.(string)
if !ok {
text = fmt.Sprint(value)
}
return strings.ToLower(strings.TrimSpace(text))
}
func isOpenAIReasoningEffort(effort string) bool {
_, ok := openAIReasoningEfforts[effort]
return ok
}
func chatReasoningModelName(body map[string]any, candidate store.RuntimeModelCandidate) string {
for _, value := range []any{
body["model"],
candidate.ProviderModelName,
candidate.ModelName,
candidate.ModelAlias,
} {
if text := normalizedReasoningString(value); text != "" {
return text
}
}
return ""
}
func providerCode(candidate store.RuntimeModelCandidate) string {
return normalizedReasoningString(candidate.Provider)
}
func baseURLCode(candidate store.RuntimeModelCandidate) string {
return normalizedReasoningString(candidate.BaseURL)
}
func isAliyunBailianOpenAI(candidate store.RuntimeModelCandidate) bool {
provider := providerCode(candidate)
baseURL := baseURLCode(candidate)
return provider == "aliyun-bailian-openai" || strings.Contains(baseURL, "dashscope.")
}
func isVolcesOpenAI(candidate store.RuntimeModelCandidate) bool {
provider := providerCode(candidate)
baseURL := baseURLCode(candidate)
return provider == "volces-openai" || strings.Contains(baseURL, "volces.com") || strings.Contains(baseURL, "byteplus.com")
}
func isDeepSeekOpenAI(candidate store.RuntimeModelCandidate) bool {
provider := providerCode(candidate)
baseURL := baseURLCode(candidate)
return provider == "deepseek-openai" || strings.Contains(baseURL, "api.deepseek.com")
}
func isZhipuOpenAI(candidate store.RuntimeModelCandidate) bool {
provider := providerCode(candidate)
baseURL := baseURLCode(candidate)
return provider == "zhipu-openai" || strings.Contains(baseURL, "bigmodel.cn") || strings.Contains(baseURL, "api.z.ai")
}
func isAliyunHighMaxReasoningModel(model string) bool {
return strings.Contains(model, "deepseek-v4") || strings.HasPrefix(model, "glm-")
}
func isAliyunThinkingBudgetModel(model string) bool {
return strings.Contains(model, "qwen") || strings.Contains(model, "qwq") || strings.Contains(model, "qvq") || strings.Contains(model, "kimi")
}
func isVolcesReasoningEffortModel(model string) bool {
return strings.HasPrefix(model, "doubao-seed-2-")
}
func isZhipuReasoningEffortModel(model string) bool {
return model == "" || strings.HasPrefix(model, "glm-5.2") || strings.HasPrefix(model, "glm-5-2")
}
func highMaxReasoningEffort(effort string) string {
if effort == "xhigh" {
return "max"
}
return "high"
}
func zhipuReasoningEffort(effort string) string {
if _, ok := zhipuReasoningEfforts[effort]; ok {
return effort
}
return ""
}
func volcesChatReasoningEffort(effort string) string {
switch effort {
case "none":
return "minimal"
case "xhigh":
return "high"
default:
if _, ok := volcesChatReasoningEfforts[effort]; ok {
return effort
}
return ""
}
}
func positiveIntFromAny(value any) (int, bool) {
if value == nil {
return 0, false
}
var number float64
switch typed := value.(type) {
case int:
number = float64(typed)
case int64:
number = float64(typed)
case float64:
number = typed
case float32:
number = float64(typed)
case jsonNumber:
parsed, err := typed.Float64()
if err != nil {
return 0, false
}
number = parsed
case string:
parsed := normalizedReasoningString(typed)
if parsed == "" {
return 0, false
}
value, err := strconv.ParseFloat(parsed, 64)
if err != nil {
return 0, false
}
number = value
default:
return 0, false
}
if math.IsNaN(number) || math.IsInf(number, 0) || number <= 0 {
return 0, false
}
return int(math.Floor(number)), true
}
type jsonNumber interface {
Float64() (float64, error)
}

View File

@ -343,6 +343,190 @@ func TestOpenAIClientChatContract(t *testing.T) {
}
}
func TestOpenAIClientChatReasoningParamsByProvider(t *testing.T) {
cases := []struct {
name string
provider string
model string
baseURL string
body map[string]any
assertion func(t *testing.T, body map[string]any)
}{
{
name: "generic openai keeps openai reasoning effort",
provider: "openai",
model: "gpt-5.5",
body: map[string]any{
"reasoning_effort": "xhigh",
},
assertion: func(t *testing.T, body map[string]any) {
if body["reasoning_effort"] != "xhigh" {
t.Fatalf("generic openai should keep xhigh, got %+v", body)
}
if _, ok := body["thinking"]; ok {
t.Fatalf("generic openai should not receive thinking object: %+v", body)
}
},
},
{
name: "aliyun qwen enables thinking and maps budget",
provider: "aliyun-bailian-openai",
model: "qwen3.7-plus",
body: map[string]any{
"reasoning_effort": "high",
"thinking_budget_tokens": 4096,
},
assertion: func(t *testing.T, body map[string]any) {
if body["enable_thinking"] != true || body["thinking_budget"] != float64(4096) {
t.Fatalf("aliyun qwen should use enable_thinking/thinking_budget, got %+v", body)
}
if _, ok := body["reasoning_effort"]; ok {
t.Fatalf("aliyun qwen should not receive reasoning_effort: %+v", body)
}
if _, ok := body["thinking_budget_tokens"]; ok {
t.Fatalf("internal thinking_budget_tokens should be removed: %+v", body)
}
},
},
{
name: "aliyun qwen disables thinking for none",
provider: "aliyun-bailian-openai",
model: "qwen3.7-plus",
body: map[string]any{
"reasoning_effort": "none",
},
assertion: func(t *testing.T, body map[string]any) {
if body["enable_thinking"] != false {
t.Fatalf("aliyun qwen none should disable thinking, got %+v", body)
}
if _, ok := body["reasoning_effort"]; ok {
t.Fatalf("aliyun qwen none should remove reasoning_effort: %+v", body)
}
},
},
{
name: "same deepseek model on aliyun uses aliyun protocol",
provider: "aliyun-bailian-openai",
model: "deepseek-v4",
body: map[string]any{
"reasoning_effort": "xhigh",
},
assertion: func(t *testing.T, body map[string]any) {
if body["enable_thinking"] != true || body["reasoning_effort"] != "max" {
t.Fatalf("aliyun deepseek should use enable_thinking and max, got %+v", body)
}
},
},
{
name: "deepseek official uses thinking object and high max effort",
provider: "deepseek-openai",
model: "deepseek-v4",
body: map[string]any{
"reasoning_effort": "xhigh",
},
assertion: func(t *testing.T, body map[string]any) {
thinking, _ := body["thinking"].(map[string]any)
if thinking["type"] != "enabled" || body["reasoning_effort"] != "max" {
t.Fatalf("deepseek official should use thinking enabled and max, got %+v", body)
}
},
},
{
name: "volces seed none disables thinking",
provider: "volces-openai",
model: "doubao-seed-2-0-pro-260215",
body: map[string]any{
"reasoning_effort": "none",
},
assertion: func(t *testing.T, body map[string]any) {
thinking, _ := body["thinking"].(map[string]any)
if thinking["type"] != "disabled" {
t.Fatalf("volces seed none should disable thinking, got %+v", body)
}
if _, ok := body["reasoning_effort"]; ok {
t.Fatalf("volces seed none should remove reasoning_effort: %+v", body)
}
},
},
{
name: "volces seed enables thinking and maps xhigh",
provider: "volces-openai",
model: "doubao-seed-2-0-pro-260215",
body: map[string]any{
"reasoning_effort": "xhigh",
},
assertion: func(t *testing.T, body map[string]any) {
thinking, _ := body["thinking"].(map[string]any)
if thinking["type"] != "enabled" || body["reasoning_effort"] != "high" {
t.Fatalf("volces seed xhigh should enable thinking and map to high, got %+v", body)
}
},
},
{
name: "zhipu preserves xhigh for glm 5.2",
provider: "zhipu-openai",
model: "glm-5.2",
body: map[string]any{
"reasoning_effort": "xhigh",
},
assertion: func(t *testing.T, body map[string]any) {
thinking, _ := body["thinking"].(map[string]any)
if thinking["type"] != "enabled" || body["reasoning_effort"] != "xhigh" {
t.Fatalf("zhipu glm should use thinking enabled and preserve xhigh, got %+v", body)
}
},
},
}
for _, tc := range cases {
t.Run(tc.name, func(t *testing.T) {
var captured map[string]any
server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
if err := json.NewDecoder(r.Body).Decode(&captured); err != nil {
t.Fatalf("decode request: %v", err)
}
_ = json.NewEncoder(w).Encode(map[string]any{
"id": "chatcmpl-reasoning",
"object": "chat.completion",
"model": captured["model"],
"choices": []any{map[string]any{
"message": map[string]any{"role": "assistant", "content": "ok"},
}},
})
}))
defer server.Close()
body := map[string]any{
"messages": []any{map[string]any{"role": "user", "content": "ping"}},
}
for key, value := range tc.body {
body[key] = value
}
baseURL := tc.baseURL
if baseURL == "" {
baseURL = server.URL
}
_, err := (OpenAIClient{HTTPClient: server.Client()}).Run(context.Background(), Request{
Kind: "chat.completions",
Model: tc.model,
Body: body,
Candidate: store.RuntimeModelCandidate{
Provider: tc.provider,
BaseURL: baseURL,
ProviderModelName: tc.model,
ModelName: tc.model,
Credentials: map[string]any{"apiKey": "test-key"},
},
})
if err != nil {
t.Fatalf("run openai client: %v", err)
}
tc.assertion(t, captured)
})
}
}
func TestUsageFromOpenAIUsageTracksKnownCachedInputZero(t *testing.T) {
knownZero := usageFromOpenAIUsage(map[string]any{
"prompt_tokens": 100,

View File

@ -27,6 +27,7 @@ func (c OpenAIClient) Run(ctx context.Context, request Request) (Response, error
body := cloneBody(request.Body)
if request.Kind == "chat.completions" {
body = NormalizeChatCompletionRequestBody(body)
applyOpenAIChatReasoningParams(body, request.Candidate)
}
body["model"] = upstreamModelName(request.Candidate)
stream := openAIEndpointSupportsStream(request.Kind) && (request.Stream || boolValue(body, "stream"))

View File

@ -138,6 +138,27 @@ func TestWriteCompatibleTaskResponseReturnsJSONWhenStreamIsFalse(t *testing.T) {
}
}
func TestWriteCompatibleTaskResponseMapsInvalidParameterToBadRequest(t *testing.T) {
executor := &fakeTaskExecutor{
runErr: &clients.ClientError{
Code: "invalid_parameter",
Message: "reasoning_effort must be one of: none, minimal, low, medium, high, xhigh",
Retryable: false,
},
}
req := httptest.NewRequest(http.MethodPost, "/api/v1/chat/completions", nil)
recorder := httptest.NewRecorder()
writeCompatibleTaskResponse(context.Background(), recorder, req, executor, "chat.completions", "gpt-test", store.GatewayTask{ID: "task-test"}, &auth.User{}, false, false)
if recorder.Code != http.StatusBadRequest {
t.Fatalf("status=%d want=%d body=%s", recorder.Code, http.StatusBadRequest, recorder.Body.String())
}
if !strings.Contains(recorder.Body.String(), "invalid_parameter") {
t.Fatalf("response should include invalid_parameter code: %s", recorder.Body.String())
}
}
func TestWriteCompatibleTaskResponseReturnsSSEWhenStreamIsTrue(t *testing.T) {
executor := &fakeTaskExecutor{
deltas: []clients.StreamDeltaEvent{{Text: "hel"}, {Text: "lo"}},
@ -284,15 +305,22 @@ type fakeTaskExecutor struct {
streamCalls int
deltas []clients.StreamDeltaEvent
output map[string]any
runErr error
}
func (f *fakeTaskExecutor) Execute(context.Context, store.GatewayTask, *auth.User) (runner.Result, error) {
f.executeCalls++
if f.runErr != nil {
return runner.Result{}, f.runErr
}
return runner.Result{Output: f.output}, nil
}
func (f *fakeTaskExecutor) ExecuteStream(_ context.Context, _ store.GatewayTask, _ *auth.User, onDelta clients.StreamDelta) (runner.Result, error) {
f.streamCalls++
if f.runErr != nil {
return runner.Result{}, f.runErr
}
for _, delta := range f.deltas {
if err := onDelta(delta); err != nil {
return runner.Result{}, err

View File

@ -1306,7 +1306,7 @@ func scopeForTaskKind(kind string) string {
func statusFromRunError(err error) int {
switch {
case clients.ErrorCode(err) == "bad_request" || clients.ErrorCode(err) == "cloned_voice_expired" || clients.ErrorCode(err) == "cloned_voice_unavailable" || clients.ErrorCode(err) == "cloned_voice_platform_unavailable" || clients.ErrorCode(err) == "unsupported_operation" || clients.ErrorCode(err) == "invalid_proxy":
case clients.ErrorCode(err) == "bad_request" || clients.ErrorCode(err) == "invalid_parameter" || clients.ErrorCode(err) == "cloned_voice_expired" || clients.ErrorCode(err) == "cloned_voice_unavailable" || clients.ErrorCode(err) == "cloned_voice_platform_unavailable" || clients.ErrorCode(err) == "unsupported_operation" || clients.ErrorCode(err) == "invalid_proxy":
return http.StatusBadRequest
case clients.ErrorCode(err) == "cloned_voice_not_found":
return http.StatusNotFound

View File

@ -190,7 +190,7 @@ type TaskRequest struct {
Stream bool `json:"stream,omitempty" example:"false"`
RunMode string `json:"runMode,omitempty" example:"simulation"`
MaxTokens int `json:"max_tokens,omitempty" example:"512"`
// ReasoningEffort 推理深度OpenAI-compatible 请求字段;开放字符串,取值随 provider 和模型能力而定,常见值为 none、minimal、low、medium、high、xhigh也可配置 max 等供应商自定义值
// ReasoningEffort 推理强度OpenAI-compatible 请求字段;仅支持 none、minimal、low、medium、high、xhigh。供应商自定义取值由网关按平台适配
ReasoningEffort string `json:"reasoning_effort,omitempty" example:"medium"`
Size string `json:"size,omitempty" example:"1024x1024"`
Duration int `json:"duration,omitempty" example:"5"`
@ -216,7 +216,7 @@ type ChatCompletionRequest struct {
Messages []ChatMessage `json:"messages"`
Temperature float64 `json:"temperature,omitempty" example:"0.7"`
MaxTokens int `json:"max_tokens,omitempty" example:"512"`
// ReasoningEffort 推理深度OpenAI-compatible 请求字段;开放字符串,取值随 provider 和模型能力而定,常见值为 none、minimal、low、medium、high、xhigh也可配置 max 等供应商自定义值
// ReasoningEffort 推理强度OpenAI-compatible 请求字段;仅支持 none、minimal、low、medium、high、xhigh。供应商自定义取值由网关按平台适配
ReasoningEffort string `json:"reasoning_effort,omitempty" example:"medium"`
Stream bool `json:"stream,omitempty" example:"false"`
RunMode string `json:"runMode,omitempty" example:"simulation"`

View File

@ -1295,6 +1295,9 @@ func validateRequest(kind string, body map[string]any) error {
if body["messages"] == nil {
return errors.New("messages is required")
}
if err := clients.ValidateOpenAIReasoningEffort(body["reasoning_effort"]); err != nil {
return err
}
case "responses":
if body["input"] == nil && body["messages"] == nil {
return errors.New("input or messages is required")

View File

@ -2,6 +2,7 @@ package runner
import (
"context"
"strings"
"testing"
"github.com/easyai/easyai-ai-gateway/apps/api/internal/clients"
@ -14,6 +15,37 @@ func (namedClient) Run(context.Context, clients.Request) (clients.Response, erro
return clients.Response{}, nil
}
func TestValidateRequestAcceptsOpenAIReasoningEffort(t *testing.T) {
for _, effort := range []string{"none", "minimal", "low", "medium", "high", "xhigh"} {
t.Run(effort, func(t *testing.T) {
err := validateRequest("chat.completions", map[string]any{
"messages": []any{map[string]any{"role": "user", "content": "ping"}},
"reasoning_effort": effort,
})
if err != nil {
t.Fatalf("expected %s to be accepted: %v", effort, err)
}
})
}
}
func TestValidateRequestRejectsNonOpenAIReasoningEffort(t *testing.T) {
for _, effort := range []string{"max", "auto"} {
t.Run(effort, func(t *testing.T) {
err := validateRequest("chat.completions", map[string]any{
"messages": []any{map[string]any{"role": "user", "content": "ping"}},
"reasoning_effort": effort,
})
if err == nil {
t.Fatalf("expected %s to be rejected", effort)
}
if clients.ErrorCode(err) != "invalid_parameter" || !strings.Contains(err.Error(), clients.OpenAIReasoningEffortValidationMessage) {
t.Fatalf("unexpected validation error for %s: %T %v", effort, err, err)
}
})
}
}
func TestClientForMapsGoogleGeminiSpecToGeminiClient(t *testing.T) {
service := &Service{clients: map[string]clients.Client{
"gemini": namedClient("gemini"),

View File

@ -160,18 +160,17 @@ WHERE p.status = 'enabled'
)
)
OR (
COALESCE(m.model_alias, '') = ''
AND (
m.model_name = $1::text
OR b.canonical_model_key = $1::text
OR b.provider_model_name = $1::text
OR (
NULLIF($3::text, '') IS NOT NULL
AND (
regexp_replace(COALESCE(m.model_name, ''), '[[:space:]]+', '', 'g') = $3::text
OR regexp_replace(COALESCE(b.canonical_model_key, ''), '[[:space:]]+', '', 'g') = $3::text
OR regexp_replace(COALESCE(b.provider_model_name, ''), '[[:space:]]+', '', 'g') = $3::text
)
m.model_name = $1::text
OR COALESCE(NULLIF(m.provider_model_name, ''), m.model_name) = $1::text
OR b.canonical_model_key = $1::text
OR b.provider_model_name = $1::text
OR (
NULLIF($3::text, '') IS NOT NULL
AND (
regexp_replace(COALESCE(m.model_name, ''), '[[:space:]]+', '', 'g') = $3::text
OR regexp_replace(COALESCE(NULLIF(m.provider_model_name, ''), m.model_name), '[[:space:]]+', '', 'g') = $3::text
OR regexp_replace(COALESCE(b.canonical_model_key, ''), '[[:space:]]+', '', 'g') = $3::text
OR regexp_replace(COALESCE(b.provider_model_name, ''), '[[:space:]]+', '', 'g') = $3::text
)
)
)

View File

@ -47,7 +47,7 @@ const textFields: FieldDefinition[] = [
{ key: 'max_input_tokens', label: '最大输入 Token', placeholder: '64000', type: 'number' },
{ key: 'max_output_tokens', label: '最大输出 Token', placeholder: '8192', type: 'number' },
{ key: 'max_thinking_tokens', label: '最大思考 Token', placeholder: '32768', type: 'number' },
{ key: 'thinkingEffortLevels', label: '推理深度', hint: '声明模型支持的 reasoning_effort 取值,可填写 max 等供应商自定义值', placeholder: 'none, minimal, low, medium, high, xhigh, max', type: 'list' },
{ key: 'thinkingEffortLevels', label: '推理深度', hint: '声明模型支持的 OpenAI reasoning_effort 取值,平台差异由网关适配', placeholder: 'none, minimal, low, medium, high, xhigh', type: 'list' },
];
const embeddingFields: FieldDefinition[] = [
@ -536,7 +536,7 @@ const imageAspectRatioOptions = [
'7:4',
'4:7',
];
const thinkingEffortOptions = ['none', 'minimal', 'low', 'medium', 'high', 'xhigh', 'max'];
const thinkingEffortOptions = ['none', 'minimal', 'low', 'medium', 'high', 'xhigh'];
const omniVideoModeOptions = ['text_to_video', 'image_reference', 'element_reference', 'first_last_frame', 'video_reference', 'video_edit', 'multi_shot'];
const durationOptionValues = ['1', '2', '3', '4', '5', '6', '8', '10', '15', '20', '25', '30'];
const exclusiveCapabilityFields: Record<string, string> = {