fix(gateway): 修复 Qwen3.8 推理参数并增加上游自愈

统一 Qwen3.8 固定思考、推理档位、预算互斥与温度下限规则，并覆盖 Chat Completions 和原生 Responses 请求。新增安全参数纠正、有限重试和并发安全的进程内 LRU 学习；补充共享行为向量、httptest 回归与真实模型验收用例。验证：go test ./... -count=1 通过，gofmt 检查通过；真实 Qwen3.8 流式、非流式及缓存重学习验收通过。
2026-08-01 22:50:45 +08:00
parent 50a13de70b
commit 9c093974a1
9 changed files with 1547 additions and 33 deletions
@@ -18,7 +18,8 @@ import (
 )

 type OpenAIClient struct {
-	HTTPClient *http.Client
+	HTTPClient  *http.Client
+	Corrections *ParameterCorrectionCache
 }

 func (c OpenAIClient) Run(ctx context.Context, request Request) (Response, error) {
@@ -56,6 +57,7 @@ func (c OpenAIClient) Run(ctx context.Context, request Request) (Response, error
 		applyOpenAIChatReasoningParams(body, request.Candidate)
 		body = FilterOpenAIChatRequestBody(body)
 	} else if request.Kind == "responses" {
+		applyOpenAIResponsesReasoningParams(body, request.Candidate)
 		body = FilterOpenAIResponsesRequestBody(body)
 		if _, hasInput := body["input"]; !hasInput {
 			if messages, hasMessages := request.Body["messages"]; hasMessages {
@@ -73,31 +75,64 @@ func (c OpenAIClient) Run(ctx context.Context, request Request) (Response, error
 	normalizeOpenAIImageRequestBody(endpointKind, body, request.OriginalBody)
 	stream := openAIEndpointSupportsStream(endpointKind) && (request.Stream || boolValue(body, "stream"))
 	ensureOpenAIStreamUsage(body, endpointKind, stream)
-	raw, contentType, err := openAIRequestPayload(endpointKind, body, request.Candidate)
-	if err != nil {
-		return Response{}, err
-	}
 	upstreamEndpoint := joinURL(openAIBaseURL(endpointKind, request.Candidate), endpoint)
-	req, err := http.NewRequestWithContext(ctx, http.MethodPost, upstreamEndpoint, bytes.NewReader(raw))
-	if err != nil {
-		return Response{}, err
-	}
-	req.Header.Set("Content-Type", contentType)
-	req.Header.Set("Authorization", "Bearer "+apiKey)
 	responseStartedAt := time.Now()
-	if err := notifySubmissionStarted(request); err != nil {
-		return Response{}, err
+	correctionScope := newParameterCorrectionScope(request, endpointKind)
+	correctionEnabled := endpointKind == "chat.completions" || endpointKind == "responses"
+	if correctionEnabled {
+		c.Corrections.apply(correctionScope, body)
 	}
-	resp, err := httpClient(request.HTTPClient, c.HTTPClient).Do(req)
-	if err != nil {
-		return Response{}, &ClientError{Code: "network", Message: err.Error(), Retryable: true}
+	provisionalRules := make([]parameterCorrectionRule, 0, 2)
+	seenCorrectionErrors := make(map[string]struct{})
+	var resp *http.Response
+	for correctionAttempt := 0; ; correctionAttempt++ {
+		raw, contentType, payloadErr := openAIRequestPayload(endpointKind, body, request.Candidate)
+		if payloadErr != nil {
+			return Response{}, payloadErr
+		}
+		req, requestErr := http.NewRequestWithContext(ctx, http.MethodPost, upstreamEndpoint, bytes.NewReader(raw))
+		if requestErr != nil {
+			return Response{}, requestErr
+		}
+		req.Header.Set("Content-Type", contentType)
+		req.Header.Set("Authorization", "Bearer "+apiKey)
+		if err := notifySubmissionStarted(request); err != nil {
+			return Response{}, err
+		}
+		resp, requestErr = httpClient(request.HTTPClient, c.HTTPClient).Do(req)
+		if requestErr != nil {
+			return Response{}, &ClientError{Code: "network", Message: requestErr.Error(), Retryable: true}
+		}
+		if err := notifyResponseReceived(request); err != nil {
+			_ = resp.Body.Close()
+			return Response{}, err
+		}
+		if !correctionEnabled || (resp.StatusCode != http.StatusBadRequest && resp.StatusCode != http.StatusUnprocessableEntity) {
+			break
+		}
+		_, _, upstreamErr := decodeHTTPResponseForProtocol(resp, openAIWireProtocol(endpointKind))
+		if upstreamErr == nil {
+			break
+		}
+		fingerprint := fmt.Sprintf("%d:%s", resp.StatusCode, strings.ToLower(upstreamErr.Error()))
+		if _, repeated := seenCorrectionErrors[fingerprint]; repeated || correctionAttempt >= 2 {
+			return Response{}, annotateResponseError(upstreamErr, requestIDFromParameterError(upstreamErr), responseStartedAt, time.Now())
+		}
+		seenCorrectionErrors[fingerprint] = struct{}{}
+		rule, safe := deriveParameterCorrection(upstreamErr, body, request.Candidate)
+		if !safe || !applyParameterCorrectionRule(body, rule) {
+			return Response{}, annotateResponseError(upstreamErr, requestIDFromParameterError(upstreamErr), responseStartedAt, time.Now())
+		}
+		c.Corrections.invalidate(correctionScope, rule.Param)
+		provisionalRules = append(provisionalRules, rule)
 	}
-	if err := notifyResponseReceived(request); err != nil {
-		return Response{}, err
+	if len(provisionalRules) > 0 && resp.StatusCode >= 200 && resp.StatusCode < 300 {
+		c.Corrections.commit(correctionScope, provisionalRules)
 	}
 	requestID := requestIDFromHTTPResponse(resp)
 	var result map[string]any
 	var wire *WireResponse
+	var err error
 	upstreamResponseID := ""
 	nativeStreamDelta := openAIWireStreamDelta(request.StreamDelta, openAIWireProtocol(endpointKind), resp)
 	if request.Kind == "responses" && protocol == ProtocolOpenAIResponses && stream {