fix(openai): 原生 Responses 路径不再把上游确定性 400 归一成可重试 502

上游返回 400（invalid_function_parameters、missing_required_parameter 等）时，
`/v1/responses` 原生路径把它包成 `502 upstream_error / "Upstream request failed"`，
真实的 message/code/param 全部丢弃。

502 属于可重试类，下游网关会把这个永远不会成功的请求反复重放（issue #5479 实测
30 个失败请求被放大成 60 次上游调用），客户端也拿不到「哪个字段非法」的线索。

同一个 service 上的兄弟路径早已做对：
- `handleCompatErrorResponse`（ChatCompletions / Anthropic）回真实状态码 +
  invalid_request_error + 真实 message；
- `/v1/images/generations` 还额外透传 code/param。

原生 Responses 是唯一漏掉的一条，本次对齐。

改动只放行 400：走到该分支说明 `shouldFailoverOpenAIUpstreamResponse` 已判定这个
400 不可 failover，即 server_is_overloaded / at capacity 这类可重试的 400 不会到达。
401/402/403（运营方凭据问题）、404/405（可能是 base_url 配错）、429 一律维持原状。

Fixes #5479

backend/internal/service/openai_gateway_service_codex_cli_only_test.go

@@ -12,6 +12,7 @@ import (
 	"github.com/Wei-Shaw/sub2api/internal/config"
 	"github.com/gin-gonic/gin"
 	"github.com/stretchr/testify/require"
+	"github.com/tidwall/gjson"
 )
 
 type stubCodexRestrictionDetector struct {
@@ -365,7 +366,12 @@ func TestOpenAIGatewayService_Forward_LogsInstructionsRequiredDetails(t *testing
 
 	_, err := svc.Forward(context.Background(), c, account, body)
 	require.Error(t, err)
-	require.Equal(t, http.StatusBadGateway, rec.Code)
+	// missing_required_parameter 是确定性的请求错误：换账号、重试都不会变。按真实的
+	// 400 回写并保留 param/code，客户端才知道该补哪个字段（而不是收到可重试的 502）。
+	require.Equal(t, http.StatusBadRequest, rec.Code)
+	require.Equal(t, "invalid_request_error", gjson.Get(rec.Body.String(), "error.type").String())
+	require.Equal(t, "missing_required_parameter", gjson.Get(rec.Body.String(), "error.code").String())
+	require.Equal(t, "instructions", gjson.Get(rec.Body.String(), "error.param").String())
 	require.Contains(t, err.Error(), "upstream error: 400")
 
 	require.True(t, logSink.ContainsMessageAtLevel("OpenAI 上游返回 Instructions are required，已记录请求详情用于排查", "warn"))

backend/internal/service/openai_gateway_upstream_errors.go

@@ -488,6 +488,24 @@ func (s *OpenAIGatewayService) handleErrorResponse(
 
 	MarkResponseCommitted(c)
 
+	// 上游 400 是确定性的请求错误：同一份请求体换账号、重试多少次都会失败。归一成
+	// 502 upstream_error 会让下游网关把它当成可重试的上游故障反复重放（#5479 实测
+	// 30 个失败请求被放大成 60 次上游调用），同时抹掉客户端定位问题所需的 code/param。
+	//
+	// 走到这里说明 shouldFailoverOpenAIUpstreamResponse 已判定该 400 不可 failover，
+	// 即 server_is_overloaded / at capacity 这类可重试的 400 不会到达此处。
+	//
+	// 兄弟路径早已这么做：handleCompatErrorResponse（ChatCompletions / Anthropic）
+	// 回真实状态码 + invalid_request_error + 真实 message；/v1/images 还额外透传
+	// code/param。原生 Responses 是唯一漏掉的一条。
+	if isOpenAIDeterministicClientError(resp.StatusCode) {
+		writeOpenAIUpstreamClientError(c, resp.StatusCode, body, upstreamMsg)
+		if upstreamMsg == "" {
+			return nil, fmt.Errorf("upstream error: %d", resp.StatusCode)
+		}
+		return nil, fmt.Errorf("upstream error: %d message=%s", resp.StatusCode, upstreamMsg)
+	}
+
 	// Return appropriate error response
 	var errType, errMsg string
 	var statusCode int

backend/internal/service/openai_upstream_client_error.go

@@ -0,0 +1,57 @@
+package service
+
+import (
+	"net/http"
+	"strings"
+
+	"github.com/gin-gonic/gin"
+	"github.com/tidwall/gjson"
+)
+
+// openAIUpstreamClientErrorFallbackType 是上游没给 error.type 时的兜底值。
+// 与 handleCompatErrorResponse 对 400 的取值保持一致。
+const openAIUpstreamClientErrorFallbackType = "invalid_request_error"
+
+// openAIUpstreamClientErrorFallbackMessage 是上游连 message 都没给时的兜底文案。
+// 仍然比 "Upstream request failed" 明确：它说明拒绝来自请求本身，而不是链路故障。
+const openAIUpstreamClientErrorFallbackMessage = "Upstream rejected the request"
+
+// isOpenAIDeterministicClientError 判断上游状态码是否表示「请求本身非法」。
+//
+// 只认 400：同一份请求体换任何账号、重试多少次都会得到同样的结果。
+//   - 401/402/403 是网关运营方的凭据/账单问题，继续包成 502，不向客户端暴露上游账号状态。
+//   - 404/405 既可能是模型不存在，也可能是上游 base_url 配错，同属运营方问题，保持现状。
+//   - 429 已有独立分支映射成 429。
+//   - 413 在更上面就按 request-body-too-large 走 failover 了。
+func isOpenAIDeterministicClientError(statusCode int) bool {
+	return statusCode == http.StatusBadRequest
+}
+
+// writeOpenAIUpstreamClientError 以 OpenAI 错误体形状回写确定性客户端错误。
+//
+// 保留上游的 type/code/param：客户端靠 param 定位是哪个字段非法（上游会给出形如
+// input[8].tools[1].tools[2].parameters 的路径），靠 code 判断是否值得重试。归一成
+// {type:"upstream_error", message:"Upstream request failed"} 会把这些信息全部抹掉。
+//
+// upstreamMsg 由调用方传入，调用方已做过 sanitizeUpstreamErrorMessage 与
+// redactAgentIdentitySensitiveBody；这里不重复清洗，也不回落读取原始 body 的
+// message，避免绕开那两道脱敏。
+func writeOpenAIUpstreamClientError(c *gin.Context, statusCode int, body []byte, upstreamMsg string) {
+	errorPayload := gin.H{"type": openAIUpstreamClientErrorFallbackType}
+	if errType := strings.TrimSpace(gjson.GetBytes(body, "error.type").String()); errType != "" {
+		errorPayload["type"] = errType
+	}
+	if code := strings.TrimSpace(extractUpstreamErrorCode(body)); code != "" {
+		errorPayload["code"] = code
+	}
+	if param := strings.TrimSpace(gjson.GetBytes(body, "error.param").String()); param != "" {
+		errorPayload["param"] = param
+	}
+	message := strings.TrimSpace(upstreamMsg)
+	if message == "" {
+		message = openAIUpstreamClientErrorFallbackMessage
+	}
+	errorPayload["message"] = message
+
+	c.JSON(statusCode, gin.H{"error": errorPayload})
+}

backend/internal/service/openai_upstream_client_error_test.go

@@ -0,0 +1,294 @@
+package service
+
+import (
+	"context"
+	"errors"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/Wei-Shaw/sub2api/internal/config"
+	"github.com/Wei-Shaw/sub2api/internal/model"
+	"github.com/gin-gonic/gin"
+	"github.com/stretchr/testify/require"
+	"github.com/tidwall/gjson"
+)
+
+// issue #5479 上游返回的原始错误体：Codex Desktop 的 automation_update 工具定义沉进
+// 会话历史后，OpenAI 每一轮都确定性地拒收。
+const openAIInvalidFunctionParametersBody = `{"error":{` +
+	`"message":"Invalid schema for function 'automation_update': schema must be a JSON Schema of 'type: \"object\"', got 'type: \"None\"'.",` +
+	`"type":"invalid_request_error",` +
+	`"param":"input[8].tools[1].tools[2].parameters",` +
+	`"code":"invalid_function_parameters"}}`
+
+func newOpenAIUpstreamErrorTestContext(t *testing.T) (*gin.Context, *httptest.ResponseRecorder) {
+	t.Helper()
+	gin.SetMode(gin.TestMode)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
+	return c, rec
+}
+
+func newOpenAIUpstreamErrorResponse(statusCode int, body string) *http.Response {
+	return &http.Response{
+		StatusCode: statusCode,
+		Header:     http.Header{"Content-Type": []string{"application/json"}},
+		Body:       io.NopCloser(strings.NewReader(body)),
+	}
+}
+
+func newOpenAIUpstreamErrorTestAccount() *Account {
+	return &Account{ID: 1, Platform: PlatformOpenAI, Type: AccountTypeOAuth, Name: "acct"}
+}
+
+// 主复现：原生 Responses 路径必须回真实的 400 与上游诊断信息，而不是可重试的 502。
+//
+// 归一成 502 时下游网关（CCH 等）会把确定性的 Schema 错误当成临时上游故障重试，
+// issue #5479 实测 30 个失败请求被放大成 60 次上游调用。
+func TestHandleErrorResponse_Deterministic400IsNotRewrappedAs502(t *testing.T) {
+	c, rec := newOpenAIUpstreamErrorTestContext(t)
+	svc := &OpenAIGatewayService{cfg: &config.Config{}}
+
+	_, err := svc.handleErrorResponse(
+		context.Background(),
+		newOpenAIUpstreamErrorResponse(http.StatusBadRequest, openAIInvalidFunctionParametersBody),
+		c, newOpenAIUpstreamErrorTestAccount(), nil,
+	)
+
+	require.Error(t, err)
+	require.Equal(t, http.StatusBadRequest, rec.Code, "确定性 400 不得被包成可重试的 502")
+
+	body := rec.Body.String()
+	require.Equal(t, "invalid_request_error", gjson.Get(body, "error.type").String())
+	require.Equal(t, "invalid_function_parameters", gjson.Get(body, "error.code").String())
+	require.Equal(t, "input[8].tools[1].tools[2].parameters", gjson.Get(body, "error.param").String(),
+		"param 是客户端定位哪个字段非法的唯一线索")
+	require.Contains(t, gjson.Get(body, "error.message").String(), "Invalid schema for function 'automation_update'")
+	require.NotContains(t, body, "Upstream request failed")
+
+	// 确定性请求错误不该换号重试——换任何账号都是同样的结果。
+	var failoverErr *UpstreamFailoverError
+	require.False(t, errors.As(err, &failoverErr), "400 不得触发 failover")
+}
+
+// 不变式：同一份上游错误体，原生 Responses 与 ChatCompletions/Anthropic 兼容路径
+// 必须给出同样的状态码和同样的 message。这两条路径在同一个 service 上，之前一条对
+// 一条错，正是本次修复的根因；锁死对称性避免将来只改一边。
+func TestHandleErrorResponse_MatchesCompatSiblingForDeterministic400(t *testing.T) {
+	svc := &OpenAIGatewayService{cfg: &config.Config{}}
+
+	nativeCtx, nativeRec := newOpenAIUpstreamErrorTestContext(t)
+	_, nativeErr := svc.handleErrorResponse(
+		context.Background(),
+		newOpenAIUpstreamErrorResponse(http.StatusBadRequest, openAIInvalidFunctionParametersBody),
+		nativeCtx, newOpenAIUpstreamErrorTestAccount(), nil,
+	)
+	require.Error(t, nativeErr)
+
+	compatCtx, _ := newOpenAIUpstreamErrorTestContext(t)
+	var compatStatus int
+	var compatType, compatMsg string
+	writeError := func(_ *gin.Context, statusCode int, errType, message string) {
+		compatStatus, compatType, compatMsg = statusCode, errType, message
+	}
+	_, compatErr := svc.handleCompatErrorResponse(
+		newOpenAIUpstreamErrorResponse(http.StatusBadRequest, openAIInvalidFunctionParametersBody),
+		compatCtx, newOpenAIUpstreamErrorTestAccount(), writeError,
+	)
+	require.Error(t, compatErr)
+
+	require.Equal(t, compatStatus, nativeRec.Code, "两条路径的状态码必须一致")
+	require.Equal(t, compatType, gjson.Get(nativeRec.Body.String(), "error.type").String(),
+		"两条路径的 error.type 必须一致")
+	require.Equal(t, compatMsg, gjson.Get(nativeRec.Body.String(), "error.message").String(),
+		"两条路径的 message 必须一致")
+}
+
+// 上游只给 message、没有 type/code/param 时，仍要回 400 + 真实 message，
+// 缺失字段用 OpenAI 惯例兜底，不得凭空编造 code/param。
+func TestHandleErrorResponse_Deterministic400WithoutUpstreamMetadata(t *testing.T) {
+	c, rec := newOpenAIUpstreamErrorTestContext(t)
+	svc := &OpenAIGatewayService{cfg: &config.Config{}}
+
+	_, err := svc.handleErrorResponse(
+		context.Background(),
+		newOpenAIUpstreamErrorResponse(http.StatusBadRequest, `{"error":{"message":"Invalid 'input': expected an array."}}`),
+		c, newOpenAIUpstreamErrorTestAccount(), nil,
+	)
+
+	require.Error(t, err)
+	require.Equal(t, http.StatusBadRequest, rec.Code)
+	body := rec.Body.String()
+	require.Equal(t, "invalid_request_error", gjson.Get(body, "error.type").String())
+	require.Equal(t, "Invalid 'input': expected an array.", gjson.Get(body, "error.message").String())
+	require.False(t, gjson.Get(body, "error.code").Exists(), "上游没给 code 就不要编一个")
+	require.False(t, gjson.Get(body, "error.param").Exists(), "上游没给 param 就不要编一个")
+}
+
+// 上游回非 JSON（反代的 HTML 错误页等）时不得 panic，也不得回空 message。
+func TestHandleErrorResponse_Deterministic400WithNonJSONBody(t *testing.T) {
+	c, rec := newOpenAIUpstreamErrorTestContext(t)
+	svc := &OpenAIGatewayService{cfg: &config.Config{}}
+
+	_, err := svc.handleErrorResponse(
+		context.Background(),
+		newOpenAIUpstreamErrorResponse(http.StatusBadRequest, `<html><body>400 Bad Request</body></html>`),
+		c, newOpenAIUpstreamErrorTestAccount(), nil,
+	)
+
+	require.Error(t, err)
+	require.Equal(t, http.StatusBadRequest, rec.Code)
+	body := rec.Body.String()
+	require.Equal(t, "invalid_request_error", gjson.Get(body, "error.type").String())
+	require.NotEmpty(t, gjson.Get(body, "error.message").String())
+}
+
+// 作用域守卫：本次只放行 400。其余落到 default 的状态码必须维持原样，
+// 避免后续有人顺手把 404/422/5xx 一起改掉。
+func TestHandleErrorResponse_NonDeterministicStatusesKeepGeneric502(t *testing.T) {
+	cases := []struct {
+		name       string
+		statusCode int
+		body       string
+		wantStatus int
+		wantType   string
+		wantMsg    string
+	}{
+		// 404/405 可能是上游 base_url 配错（运营方问题），不当成客户端错误暴露。
+		{"not_found", http.StatusNotFound, `{"error":{"message":"Unknown request URL"}}`,
+			http.StatusBadGateway, "upstream_error", "Upstream request failed"},
+		{"unprocessable", http.StatusUnprocessableEntity, `{"error":{"message":"Invalid schema for field messages"}}`,
+			http.StatusBadGateway, "upstream_error", "Upstream request failed"},
+		// 401/402/403 是网关运营方的凭据/账单问题，必须继续对客户端屏蔽上游账号状态。
+		{"unauthorized", http.StatusUnauthorized, `{"error":{"message":"Incorrect API key provided: sk-abc"}}`,
+			http.StatusBadGateway, "upstream_error", "Upstream authentication failed, please contact administrator"},
+		{"forbidden", http.StatusForbidden, `{"error":{"message":"Your account is deactivated"}}`,
+			http.StatusBadGateway, "upstream_error", "Upstream access forbidden, please contact administrator"},
+		// 429 保持独立映射。
+		{"rate_limited", http.StatusTooManyRequests, `{"error":{"message":"Rate limit reached"}}`,
+			http.StatusTooManyRequests, "rate_limit_error", "Upstream rate limit exceeded, please retry later"},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			c, rec := newOpenAIUpstreamErrorTestContext(t)
+			svc := &OpenAIGatewayService{cfg: &config.Config{}}
+
+			_, err := svc.handleErrorResponse(
+				context.Background(),
+				newOpenAIUpstreamErrorResponse(tc.statusCode, tc.body),
+				c, newOpenAIUpstreamErrorTestAccount(), nil,
+			)
+
+			require.Error(t, err)
+			require.Equal(t, tc.wantStatus, rec.Code)
+			require.Equal(t, tc.wantType, gjson.Get(rec.Body.String(), "error.type").String())
+			require.Equal(t, tc.wantMsg, gjson.Get(rec.Body.String(), "error.message").String())
+		})
+	}
+}
+
+// 顺序守卫：管理员配置的错误透传规则在更上游命中，新分支不得抢在它前面。
+func TestHandleErrorResponse_PassthroughRuleStillWinsOver400Branch(t *testing.T) {
+	c, rec := newOpenAIUpstreamErrorTestContext(t)
+	ruleSvc := &ErrorPassthroughService{}
+	ruleSvc.setLocalCache([]*model.ErrorPassthroughRule{
+		newNonFailoverPassthroughRule(http.StatusBadRequest, "automation_update", http.StatusTeapot, "自定义文案"),
+	})
+	BindErrorPassthroughService(c, ruleSvc)
+	svc := &OpenAIGatewayService{cfg: &config.Config{}}
+
+	_, err := svc.handleErrorResponse(
+		context.Background(),
+		newOpenAIUpstreamErrorResponse(http.StatusBadRequest, openAIInvalidFunctionParametersBody),
+		c, newOpenAIUpstreamErrorTestAccount(), nil,
+	)
+
+	require.Error(t, err)
+	require.Equal(t, http.StatusTeapot, rec.Code, "命中透传规则时必须按规则的状态码回写")
+	require.Equal(t, "自定义文案", gjson.Get(rec.Body.String(), "error.message").String())
+}
+
+func TestIsOpenAIDeterministicClientError(t *testing.T) {
+	require.True(t, isOpenAIDeterministicClientError(http.StatusBadRequest))
+	for _, status := range []int{
+		http.StatusUnauthorized, http.StatusPaymentRequired, http.StatusForbidden,
+		http.StatusNotFound, http.StatusMethodNotAllowed, http.StatusRequestEntityTooLarge,
+		http.StatusUnprocessableEntity, http.StatusTooManyRequests,
+		http.StatusInternalServerError, http.StatusBadGateway, http.StatusServiceUnavailable,
+	} {
+		require.False(t, isOpenAIDeterministicClientError(status), "status %d", status)
+	}
+}
+
+func TestWriteOpenAIUpstreamClientError_PayloadShape(t *testing.T) {
+	cases := []struct {
+		name        string
+		body        string
+		upstreamMsg string
+		wantType    string
+		wantCode    string
+		wantParam   string
+		wantMessage string
+	}{
+		{
+			name:        "full_metadata",
+			body:        openAIInvalidFunctionParametersBody,
+			upstreamMsg: "Invalid schema for function 'automation_update'",
+			wantType:    "invalid_request_error",
+			wantCode:    "invalid_function_parameters",
+			wantParam:   "input[8].tools[1].tools[2].parameters",
+			wantMessage: "Invalid schema for function 'automation_update'",
+		},
+		{
+			name:        "upstream_type_preserved",
+			body:        `{"error":{"type":"invalid_prompt","message":"blocked"}}`,
+			upstreamMsg: "blocked",
+			wantType:    "invalid_prompt",
+			wantMessage: "blocked",
+		},
+		{
+			name:        "empty_body_falls_back",
+			body:        ``,
+			upstreamMsg: "",
+			wantType:    "invalid_request_error",
+			wantMessage: openAIUpstreamClientErrorFallbackMessage,
+		},
+		{
+			// 调用方传入的 message 已脱敏，必须原样使用，不得回落读取原始 body。
+			name:        "sanitized_message_wins_over_raw_body",
+			body:        `{"error":{"message":"failed for key=secret123"}}`,
+			upstreamMsg: "failed for key=***",
+			wantType:    "invalid_request_error",
+			wantMessage: "failed for key=***",
+		},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			c, rec := newOpenAIUpstreamErrorTestContext(t)
+
+			writeOpenAIUpstreamClientError(c, http.StatusBadRequest, []byte(tc.body), tc.upstreamMsg)
+
+			require.Equal(t, http.StatusBadRequest, rec.Code)
+			body := rec.Body.String()
+			require.Equal(t, tc.wantType, gjson.Get(body, "error.type").String())
+			require.Equal(t, tc.wantMessage, gjson.Get(body, "error.message").String())
+			if tc.wantCode == "" {
+				require.False(t, gjson.Get(body, "error.code").Exists())
+			} else {
+				require.Equal(t, tc.wantCode, gjson.Get(body, "error.code").String())
+			}
+			if tc.wantParam == "" {
+				require.False(t, gjson.Get(body, "error.param").Exists())
+			} else {
+				require.Equal(t, tc.wantParam, gjson.Get(body, "error.param").String())
+			}
+			require.NotContains(t, body, "secret123", "原始 body 里的敏感串不得泄漏")
+		})
+	}
+}