Merge pull request #7571 from SPHINX998/fix/openai-probe-model-unavailable-inconclusive

fix(openai): keep Responses support unknown when the probe model is missing

backend/internal/service/openai_apikey_responses_probe.go

@@ -77,7 +77,8 @@ func openaiResponsesProbePayload(modelID string) []byte {
 //
 // 工具能力探测必须用上游真实存在的模型——用占位模型(DefaultTestModel)打第三方
 // 上游只会拿到 400 model-not-found,无从判定工具能力。优先取账号 model_mapping
-// 的上游模型(值),按字典序取首个具体(非通配符)模型以保证可复现;无映射时回退
+// 的上游模型(值),其中优先通用 GPT 文本模型,同类按字典序取首个具体(非通配符)
+// 模型以保证可复现;无映射时回退
 // DefaultTestModel(适配 OpenAI 官方 APIKey 账号)。
 func selectResponsesProbeModel(account *Account) string {
 	mapping := account.GetModelMapping()
@@ -93,6 +94,13 @@ func selectResponsesProbeModel(account *Account) string {
 		return openai.DefaultTestModel
 	}
 	sort.Strings(candidates)
+	// Prefer a general text model over auxiliary entries such as
+	// codex-auto-review or image-only models, which cannot prove tool support.
+	for _, candidate := range candidates {
+		if strings.HasPrefix(candidate, "gpt-") && !isOpenAIImageGenerationModel(candidate) {
+			return candidate
+		}
+	}
 	return candidates[0]
 }
 
@@ -259,9 +267,13 @@ func (s *AccountTestService) ProbeOpenAIAPIKeyResponsesSupport(ctx context.Conte
 // 其余 2xx 一律可下结论——尤其 status=completed 却只回 reasoning 的上游（火山方舟
 // coding/v3 × kimi-k2.6），仍按原逻辑判为不支持。
 //
-// 非 2xx 的结论只看状态码、不依赖响应内容，恒可下结论。
+// 明确指向探测模型不可用的 400/404(model_not_found 等)只说明模型不存在,不说明端点
+// 能力,不下结论;其余非 2xx 的结论只看状态码、不依赖响应内容，恒可下结论。
 // 缺少 status 字段的响应体（含非 JSON）也按可下结论处理，保持既有行为。
 func responsesProbeVerdictIsConclusive(status int, body []byte) bool {
+	if isResponsesProbeModelUnavailable(status, body) {
+		return false
+	}
 	if status < 200 || status >= 300 {
 		return true
 	}
@@ -296,13 +308,17 @@ func isResponsesEndpointSupportedByStatus(status int) bool {
 // decideResponsesProbeSupport 依据探测响应判定上游 /v1/responses 是否真正可用于
 // 携带工具的请求。
 //
-//   - 404 / 405：端点不存在 → false
+//   - 探测模型不可用的 400/404：与端点能力无关 → true(调用方同时不下结论,不写入)
+//   - 其他 404 / 405：端点不存在 → false
 //   - 其他非 2xx（401/403/422/5xx 等）：端点存在,但本次无法判定工具能力
 //     （鉴权/校验/瞬时故障）→ 保守按 true,保持既有"端点存在即支持"行为
 //   - 2xx：探测以 tool_choice=required 强制工具调用,响应必须含 function_call
 //     输出项才算真正可用;否则(如火山方舟 coding/v3 × kimi-k2.6 仅回 reasoning)
 //     判为 false,使网关改走 /v1/chat/completions 直转路径。
 func decideResponsesProbeSupport(status int, body []byte) bool {
+	if isResponsesProbeModelUnavailable(status, body) {
+		return true // The model error says nothing negative about the endpoint.
+	}
 	if status == http.StatusNotFound || status == http.StatusMethodNotAllowed {
 		return false
 	}
@@ -312,6 +328,21 @@ func decideResponsesProbeSupport(status int, body []byte) bool {
 	return responsesProbeBodyHasFunctionCall(body)
 }
 
+// isResponsesProbeModelUnavailable 判断 400/404 是否只是在说探测模型不可用
+// (错误码/类型为 model_not_found 等,或错误文案明确说模型不存在/不可用)。
+func isResponsesProbeModelUnavailable(status int, body []byte) bool {
+	if status != http.StatusBadRequest && status != http.StatusNotFound {
+		return false
+	}
+	for _, path := range []string{"error.code", "error.type", "response.error.code", "response.error.type"} {
+		switch strings.ToLower(strings.TrimSpace(gjson.GetBytes(body, path).String())) {
+		case "model_not_found", "model_not_available", "unsupported_model", "invalid_model":
+			return true
+		}
+	}
+	return isExplicitOpenAIModelAvailabilityMessage(extractUpstreamErrorMessage(body))
+}
+
 // responsesProbeBodyHasFunctionCall 判断非流式 Responses 响应体的 output 数组里
 // 是否存在 function_call 输出项。
 func responsesProbeBodyHasFunctionCall(body []byte) bool {

backend/internal/service/openai_apikey_responses_probe_model_unavailable_test.go

@@ -0,0 +1,49 @@
+package service
+
+import (
+	"context"
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+
+	"github.com/Wei-Shaw/sub2api/internal/config"
+	"github.com/stretchr/testify/require"
+)
+
+// A 400/404 that only says the probe model does not exist says nothing about
+// the /v1/responses endpoint: the verdict must stay inconclusive and the
+// account capability must not be overwritten with "unsupported".
+func TestResponsesProbeModelUnavailableIsInconclusive(t *testing.T) {
+	for _, body := range []string{
+		`{"error":{"type":"model_not_found","message":"Model codex-auto-review is not supported by any configured account in this group"}}`,
+		`{"error":{"code":"model_not_available"}}`,
+		`{"error":{"message":"The model missing does not exist"}}`,
+	} {
+		require.False(t, responsesProbeVerdictIsConclusive(404, []byte(body)))
+		require.True(t, decideResponsesProbeSupport(404, []byte(body)))
+		updates := make(chan map[string]any, 1)
+		account := Account{ID: 96, Platform: PlatformOpenAI, Type: AccountTypeAPIKey, Credentials: map[string]any{"api_key": "test", "base_url": "https://upstream.example"}}
+		repo := &snapshotUpdateAccountRepo{stubOpenAIAccountRepo: stubOpenAIAccountRepo{accounts: []Account{account}}, updateExtraCalls: updates}
+		svc := &AccountTestService{accountRepo: repo, cfg: &config.Config{}, httpUpstream: &httpUpstreamRecorder{resp: &http.Response{
+			StatusCode: 404, Header: http.Header{}, Body: io.NopCloser(strings.NewReader(body)),
+		}}}
+		svc.ProbeOpenAIAPIKeyResponsesSupport(context.Background(), account.ID)
+		select {
+		case <-updates:
+			t.Fatal("a model failure must not overwrite the account endpoint capability")
+		default:
+		}
+	}
+	// A plain 404/405 still means the endpoint does not exist.
+	require.True(t, responsesProbeVerdictIsConclusive(404, []byte("404 page not found")))
+	require.False(t, decideResponsesProbeSupport(404, []byte("404 page not found")))
+	require.False(t, decideResponsesProbeSupport(405, nil))
+}
+
+func TestSelectResponsesProbeModelPrefersGeneralTextModel(t *testing.T) {
+	account := &Account{Credentials: map[string]any{"model_mapping": map[string]any{
+		"codex-auto-review": "codex-auto-review", "gpt-image-2": "gpt-image-2", "gpt-5.5": "gpt-5.5",
+	}}}
+	require.Equal(t, "gpt-5.5", selectResponsesProbeModel(account))
+}