发布 0.2.0:模型池自愈、探测修正、任务异步触发与删除加固

.gitea/workflows/release.yml

@@ -23,7 +23,7 @@ jobs:
           echo "IMAGE=${REGISTRY}/${{ gitea.repository }}" >> "$GITHUB_ENV"
           echo "BUILD_TIME=$(date -u +%Y-%m-%dT%H:%M:%SZ)" >> "$GITHUB_ENV"
 
-      - name: 下载 DASH_VERSION 固定版本的前端 dist.zip,校验后解压进嵌入目录
+      - name: 下载前端 dist.zip
         env:
           TOKEN: ${{ secrets.BUILD_TOKEN }}
         run: |
@@ -36,7 +36,7 @@ jobs:
           unzip -q dist.zip -d internal/webui/dist
           test -f internal/webui/dist/index.html
 
-      - name: 构建双架构二进制(artifact stage 导出)
+      - name: 构建双架构二进制
         run: |
           docker buildx build --pull \
             --platform linux/amd64,linux/arm64 \

CHANGELOG.md

@@ -2,6 +2,24 @@
 
 格式遵循 [Keep a Changelog](https://keepachangelog.com/zh-CN/1.1.0/)，版本号遵循语义化版本。
 
+## [0.2.0] - 2026-07-10
+
+### Added
+
+- AI 网关模型池自愈：同步与每日探测后自动逐个试调验证模型可用性（maxTokens=1），不可按需调用的模型（微调基座 / 区域未开放按需供给）自动标记并从模型列表、路由与探测候选中剔除；已标记模型每 20 小时复检，OCI 恢复供给后自动解除标记。`ai_model_caches` 新增 `unusable` / `unusable_reason` / `checked_at`（启动自动迁移），渠道详情接口保留标记行与原因便于排查
+
+### Changed
+
+- 后台任务「立即执行」改异步触发：接口立即返回 202，执行结果经任务日志轮询呈现；同一任务在途时重复触发返回 409「任务正在执行中」，与 cron 重叠触发静默跳过（此前同步阻塞数十秒且同一任务可并发重复执行）
+- 渠道探测候选跨厂商分散（上限 8，voice 等不可对话形态排除）：单一厂商在区域内全为微调基座时不再拖垮整个渠道的探测结论
+- 网关调用（对话 / 流式 / 向量化）遇「模型不可按需调用」类错误（微调基座 400 / 实体不存在 404）自动标记该模型并换渠道重试，且不计入渠道熔断；仅鉴权类错误才判定租户无配额
+- 同名模型多条目去重优先保留非微调基座条目，降低缓存到不可调用 OCID 的概率
+
+### Fixed
+
+- 渠道「同步模型成功但探测报错不可用」：法兰克福等区域的微调基座模型占满探测候选导致的误报（探测状态与真实可用性不符的根因）
+- 租户删除：告警命中清理改子查询，日志事件数万条时不再超出 SQL 绑定变量上限导致删除失败；无法解析的任务 payload 记警告跳过并保留原任务，不再永久阻断租户删除
+
 ## [0.1.0] - 2026-07-10
 
 ### Added

docs/docs.go

@@ -5738,7 +5738,7 @@ const docTemplate = `{
                 "tags": [
                     "任务与日志回传"
                 ],
-                "summary": "立即执行任务",
+                "summary": "立即执行任务(异步触发,结果经任务日志轮询获取)",
                 "parameters": [
                     {
                         "type": "integer",
@@ -5749,8 +5749,15 @@ const docTemplate = `{
                     }
                 ],
                 "responses": {
-                    "200": {
+                    "202": {
-                        "description": "OK",
+                        "description": "Accepted",
+                        "schema": {
+                            "type": "object",
+                            "additionalProperties": true
+                        }
+                    },
+                    "409": {
+                        "description": "任务正在执行中",
                         "schema": {
                             "type": "object",
                             "additionalProperties": true

docs/swagger.json

@@ -5731,7 +5731,7 @@
                 "tags": [
                     "任务与日志回传"
                 ],
-                "summary": "立即执行任务",
+                "summary": "立即执行任务(异步触发,结果经任务日志轮询获取)",
                 "parameters": [
                     {
                         "type": "integer",
@@ -5742,8 +5742,15 @@
                     }
                 ],
                 "responses": {
-                    "200": {
+                    "202": {
-                        "description": "OK",
+                        "description": "Accepted",
+                        "schema": {
+                            "type": "object",
+                            "additionalProperties": true
+                        }
+                    },
+                    "409": {
+                        "description": "任务正在执行中",
                         "schema": {
                             "type": "object",
                             "additionalProperties": true

docs/swagger.yaml

@@ -4158,14 +4158,19 @@ paths:
         required: true
         type: integer
       responses:
-        "200":
+        "202":
-          description: OK
+          description: Accepted
+          schema:
+            additionalProperties: true
+            type: object
+        "409":
+          description: 任务正在执行中
           schema:
             additionalProperties: true
             type: object
       security:
       - BearerAuth: []
-      summary: 立即执行任务
+      summary: 立即执行任务(异步触发,结果经任务日志轮询获取)
       tags:
       - 任务与日志回传
   /api/v1/webhooks/oci-logs/{secret}:

internal/api/task.go

@@ -2,6 +2,7 @@ package api
 
 import (
 	"encoding/json"
+	"errors"
 	"net/http"
 	"strconv"
 
@@ -155,10 +156,11 @@ func (h *taskHandler) logs(c *gin.Context) {
 	c.JSON(http.StatusOK, logs)
 }
 
-// @Summary  立即执行任务
+// @Summary  立即执行任务(异步触发,结果经任务日志轮询获取)
 // @Tags     任务与日志回传
 // @Param    id path int true "配置 ID"
-// @Success  200 {object} map[string]any
+// @Success  202 {object} map[string]any
+// @Failure  409 {object} map[string]any "任务正在执行中"
 // @Security BearerAuth
 // @Router   /api/v1/tasks/{id}/run [post]
 func (h *taskHandler) run(c *gin.Context) {
@@ -166,10 +168,13 @@ func (h *taskHandler) run(c *gin.Context) {
 	if !ok {
 		return
 	}
-	entry, err := h.svc.RunTaskNow(c.Request.Context(), id)
+	if err := h.svc.TriggerTask(c.Request.Context(), id); err != nil {
-	if err != nil {
+		if errors.Is(err, service.ErrTaskRunning) {
+			c.JSON(http.StatusConflict, gin.H{"error": err.Error()})
+			return
+		}
 		respondError(c, err)
 		return
 	}
-	c.JSON(http.StatusOK, entry)
+	c.JSON(http.StatusAccepted, gin.H{"triggered": true})
 }

internal/model/models.go

@@ -309,6 +309,13 @@ type AiModelCache struct {
 	// 弃用后仍可调用,直到 RetiredAt(按需推理退役,同步层已剔除过期项)。
 	DeprecatedAt *time.Time `json:"deprecatedAt"`
 	RetiredAt    *time.Time `json:"retiredAt"`
+	// Unusable 标记该模型在此渠道不可按需调用(微调基座/区域未开放按需),
+	// 由探测、真实调用或同步后的后台验证习得:不参与网关列表/路由/探测候选;
+	// 任何同步都按 OCID 结转标记,已标记模型定期复检、恢复供给自动解除。
+	Unusable       bool   `json:"unusable"`
+	UnusableReason string `gorm:"size:200" json:"unusableReason,omitempty"`
+	// CheckedAt 是最近一次可用性验证时间;NULL 表示从未验证(同步后的后台验证以此为队列)。
+	CheckedAt *time.Time `json:"checkedAt"`
 }
 
 // AiCallLog 是网关调用日志:仅元数据与 token 用量,绝不记录 prompt / 响应正文。

internal/oci/errors.go

@@ -76,3 +76,34 @@ func ServiceStatus(err error) (int, bool) {
 	}
 	return 0, false
 }
+
+// IsOnDemandUnsupported 识别「模型在该区域仅为微调基座、不支持按需调用」的 400:
+// OCI 消息形如 "Not allowed to call finetune base model …, use Endpoint: false"。
+// ListModels 无字段可事先区分,只能在调用报错时识别并换渠道。
+func IsOnDemandUnsupported(err error) bool {
+	var svcErr common.ServiceError
+	if !errors.As(err, &svcErr) {
+		return false
+	}
+	return svcErr.GetHTTPStatusCode() == 400 &&
+		strings.Contains(strings.ToLower(svcErr.GetMessage()), "finetune base model")
+}
+
+// IsEntityNotFound 识别「实体不存在」404(消息 "Entity with key … not found"):
+// GenAI 对区域内无按需供给的模型 OCID 返回此类 404,属模型级错误;
+// 鉴权失败的 404 是 NotAuthorizedOrNotFound(消息为 Authorization failed…),不在此列。
+func IsEntityNotFound(err error) bool {
+	var svcErr common.ServiceError
+	if !errors.As(err, &svcErr) {
+		return false
+	}
+	msg := strings.ToLower(svcErr.GetMessage())
+	return svcErr.GetHTTPStatusCode() == 404 &&
+		strings.Contains(msg, "entity with key") && strings.Contains(msg, "not found")
+}
+
+// IsModelUnavailable 归并「模型×区域不可按需调用」两类报错:微调基座 400 与实体不存在 404;
+// 命中即应把该 (渠道, 模型) 从池中剔除并换候选/换渠道,而非定论租户配额问题。
+func IsModelUnavailable(err error) bool {
+	return IsOnDemandUnsupported(err) || IsEntityNotFound(err)
+}

internal/oci/errors_test.go

@@ -109,3 +109,52 @@ func TestErrorHint(t *testing.T) {
 		})
 	}
 }
+
+func TestIsOnDemandUnsupported(t *testing.T) {
+	ft := fakeServiceError{status: 400, code: "InvalidParameter",
+		message: "Not allowed to call finetune base model ocid1.generativeaimodel.oc1.eu-frankfurt-1.tpel5q, use Endpoint: false"}
+	tests := []struct {
+		name string
+		err  error
+		want bool
+	}{
+		{"微调基座 400 命中(含包装链)", fmt.Errorf("genai chat: %w", ft), true},
+		{"同消息但非 400 不命中", fakeServiceError{status: 500, code: "InternalError", message: ft.message}, false},
+		{"普通 400 不命中", fakeServiceError{status: 400, code: "InvalidParameter", message: "bad request"}, false},
+		{"非服务端错误不命中", errors.New("finetune base model"), false},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := IsOnDemandUnsupported(tt.err); got != tt.want {
+				t.Errorf("IsOnDemandUnsupported() = %v, want %v", got, tt.want)
+			}
+		})
+	}
+}
+
+func TestIsModelUnavailable(t *testing.T) {
+	entity404 := fakeServiceError{status: 404, code: "NotFound",
+		message: "Entity with key ocid1.generativeaimodel.oc1.eu-frankfurt-1.2flsfq not found"}
+	auth404 := fakeServiceError{status: 404, code: "NotAuthorizedOrNotFound",
+		message: "Authorization failed or requested resource not found."}
+	ft400 := fakeServiceError{status: 400, code: "InvalidParameter",
+		message: "Not allowed to call finetune base model ocid1.generativeaimodel.oc1..x, use Endpoint: false"}
+	tests := []struct {
+		name string
+		err  error
+		want bool
+	}{
+		{"实体不存在 404 命中(含包装链)", fmt.Errorf("genai chat: %w", entity404), true},
+		{"鉴权类 404 不命中(仍属租户级)", auth404, false},
+		{"微调基座 400 命中", ft400, true},
+		{"其他 404 消息不命中", fakeServiceError{status: 404, code: "NotFound", message: "route not found"}, false},
+		{"非服务端错误不命中", errors.New("entity with key x not found"), false},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := IsModelUnavailable(tt.err); got != tt.want {
+				t.Errorf("IsModelUnavailable() = %v, want %v", got, tt.want)
+			}
+		})
+	}
+}

internal/oci/genai.go

@@ -17,11 +17,10 @@ import (
 
 // GenAiModel 是区域可用基础模型的摘要(管理面 ListModels)。
 type GenAiModel struct {
-	Ocid     string   `json:"ocid"`
+	Ocid   string   `json:"ocid"`
-	Name     string   `json:"name"`
+	Name   string   `json:"name"`
-	Vendor   string   `json:"vendor"`
+	Vendor string   `json:"vendor"`
-	ChatOnly bool     `json:"-"`
+	Caps   []string `json:"capabilities"`
-	Caps     []string `json:"capabilities"`
 	// Capability 是网关侧归一能力:CHAT / EMBEDDING(兼具时算 CHAT)
 	Capability string `json:"capability"`
 	// Deprecated 是 OCI 宣布的弃用时间(TimeDeprecated),nil 表示未宣布;
@@ -71,18 +70,39 @@ func (c *RealClient) ListGenAiModels(ctx context.Context, cred Credentials, regi
 	if err != nil {
 		return nil, fmt.Errorf("list genai models: %w", err)
 	}
-	seen := map[string]bool{}
+	return dedupGenAiModels(resp.Items, time.Now()), nil
-	now := time.Now()
+}
+
+// dedupGenAiModels 压平并按名称去重;同名多条目时优先保留不含 FINE_TUNE 能力的条目
+// (微调基座条目在部分区域不支持按需调用,缓存其 OCID 会导致调用 400)。
+func dedupGenAiModels(items []generativeai.ModelSummary, now time.Time) []GenAiModel {
+	seen := map[string]int{}
 	var out []GenAiModel
-	for _, m := range resp.Items {
+	for _, m := range items {
 		gm, ok := toGenAiModel(m, now)
-		if !ok || seen[gm.Name] {
+		if !ok {
 			continue
 		}
-		seen[gm.Name] = true
+		if i, dup := seen[gm.Name]; dup {
+			if hasFineTune(out[i].Caps) && !hasFineTune(gm.Caps) {
+				out[i] = gm
+			}
+			continue
+		}
+		seen[gm.Name] = len(out)
 		out = append(out, gm)
 	}
-	return out, nil
+	return out
+}
+
+// hasFineTune 判断能力列表是否含 FINE_TUNE(微调基座条目)。
+func hasFineTune(caps []string) bool {
+	for _, c := range caps {
+		if c == string(generativeai.ModelCapabilityFineTune) {
+			return true
+		}
+	}
+	return false
 }
 
 // toGenAiModel 压平模型摘要;无归一能力、无名称或按需推理已退役

internal/oci/genai_retire_test.go

@@ -38,3 +38,33 @@ func TestToGenAiModelRetiredFilter(t *testing.T) {
 		t.Error("正常模型应入池")
 	}
 }
+
+func TestDedupGenAiModelsFineTunePreference(t *testing.T) {
+	now := time.Date(2026, 7, 10, 0, 0, 0, 0, time.UTC)
+	chat := []generativeai.ModelCapabilityEnum{generativeai.ModelCapabilityChat}
+	chatFT := []generativeai.ModelCapabilityEnum{generativeai.ModelCapabilityChat, generativeai.ModelCapabilityFineTune}
+	mk := func(id, name string, caps []generativeai.ModelCapabilityEnum) generativeai.ModelSummary {
+		return generativeai.ModelSummary{
+			Id: common.String(id), DisplayName: common.String(name), Vendor: common.String("meta"),
+			Capabilities: caps, LifecycleState: generativeai.ModelLifecycleStateActive,
+		}
+	}
+	const name = "meta.llama-3-70b-instruct"
+	tests := []struct {
+		name  string
+		items []generativeai.ModelSummary
+		want  string // 期望保留的 OCID
+	}{
+		{"基座在前、纯对话在后:保留纯对话条目", []generativeai.ModelSummary{mk("o-ft", name, chatFT), mk("o-od", name, chat)}, "o-od"},
+		{"纯对话在前、基座在后:保留纯对话条目", []generativeai.ModelSummary{mk("o-od", name, chat), mk("o-ft", name, chatFT)}, "o-od"},
+		{"仅基座条目:保留不误删", []generativeai.ModelSummary{mk("o-ft", name, chatFT)}, "o-ft"},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := dedupGenAiModels(tt.items, now)
+			if len(got) != 1 || got[0].Ocid != tt.want {
+				t.Errorf("dedupGenAiModels() = %+v, want 仅保留 %s", got, tt.want)
+			}
+		})
+	}
+}

internal/service/aigateway.go

@@ -30,6 +30,10 @@ const (
 	aiBackoffCap    = 30 * time.Minute
 	// aiKeyTouchGap 是 LastUsedAt 的最小写库间隔,避免高频调用刷库
 	aiKeyTouchGap = time.Minute
+	// modelRecheckGap 是已标记不可用模型的复检间隔(恢复供给自动解除标记);
+	// validateBatchCap 限制单渠道单轮验证的试调次数
+	modelRecheckGap  = 20 * time.Hour
+	validateBatchCap = 32
 	// 内容日志(红线例外)约束:开启必须限时(上限 7 天),正文截断,短保留
 	aiContentLogMaxHours  = 168
 	aiContentLogRetention = 7 * 24 * time.Hour
@@ -55,13 +59,17 @@ type AiGatewayService struct {
 	// touchMu 保护各密钥的最近触达时间(内存节流,不追求跨实例精确)
 	touchMu   sync.Mutex
 	lastTouch map[uint]time.Time
+	// validateMu 保护 validating:同一渠道的模型验证不并发
+	validateMu sync.Mutex
+	validating map[uint]bool
 	// onChannelsChanged 在渠道增删后触发,由 main 装配为探测任务同步钩子
 	onChannelsChanged func(context.Context)
 }
 
 // NewAiGatewayService 组装依赖;调用 StartCleanup 后开始调用日志周期清理。
 func NewAiGatewayService(db *gorm.DB, configs *OciConfigService, client oci.Client) *AiGatewayService {
-	return &AiGatewayService{db: db, configs: configs, client: client, lastTouch: map[uint]time.Time{}}
+	return &AiGatewayService{db: db, configs: configs, client: client,
+		lastTouch: map[uint]time.Time{}, validating: map[uint]bool{}}
 }
 
 // SetOnChannelsChanged 注册渠道数量变化钩子(渠道创建/删除成功后调用)。
@@ -290,7 +298,7 @@ func (s *AiGatewayService) ProbeChannel(ctx context.Context, id uint) (*model.Ai
 	return &fresh, nil
 }
 
-// probe 执行探测并返回 (状态, 错误摘要);同时完成模型缓存同步。
+// probe 执行探测并返回 (状态, 错误摘要);同时完成模型缓存同步(保留不可用标记)。
 func (s *AiGatewayService) probe(ctx context.Context, cred oci.Credentials, ch *model.AiChannel) (string, string) {
 	models, err := s.client.ListGenAiModels(ctx, cred, ch.Region)
 	if err != nil {
@@ -303,43 +311,209 @@ func (s *AiGatewayService) probe(ctx context.Context, cred oci.Credentials, ch *
 	if err := s.replaceModels(ctx, ch.ID, models); err != nil {
 		return "error", truncateErr(err.Error())
 	}
-	return s.probeChat(ctx, cred, ch, models)
+	usable, err := s.usableModels(ctx, ch.ID, models)
+	if err != nil {
+		return "error", truncateErr(err.Error())
+	}
+	return s.probeChat(ctx, cred, ch, usable)
 }
 
-// probeChat 按偏好挑选至多 3 个模型依次试调:部分模型元数据标 CHAT 但实际
+// usableModels 过滤掉缓存中已标记不可按需调用的模型,探测候选不再反复踩坑。
-// 不可对话(如 voice agent),遇 400/5xx 换下一个;401/403/404 属租户级直接定论。
+func (s *AiGatewayService) usableModels(ctx context.Context, channelID uint, models []oci.GenAiModel) ([]oci.GenAiModel, error) {
+	marks, err := s.loadModelMarks(ctx, channelID)
+	if err != nil {
+		return nil, err
+	}
+	out := make([]oci.GenAiModel, 0, len(models))
+	for _, m := range models {
+		if !marks[m.Ocid].Unusable {
+			out = append(out, m)
+		}
+	}
+	return out, nil
+}
+
+// probeChat 按候选顺序试调(上限 8):遇「模型不可按需调用」标记剔除并换下一个;
+// 401/403 与鉴权类 404 属租户级直接定论 no_quota;其余错误(部分模型元数据标 CHAT
+// 但实际不可对话,如 voice agent)累计 3 次止损。
 func (s *AiGatewayService) probeChat(ctx context.Context, cred oci.Credentials, ch *model.AiChannel, models []oci.GenAiModel) (string, string) {
 	status, detail := "error", "无可试调对话模型"
+	errBudget := 3
 	for _, m := range probeCandidates(models) {
 		code, err := s.client.GenAiProbeChat(ctx, cred, ch.Region, m.Ocid, m.Name)
 		switch {
 		case code == 200 || code == 429:
 			return "ok", ""
+		case oci.IsModelUnavailable(err):
+			s.markModelUnusable(ctx, ch.ID, m.Ocid, oci.CompactError(err))
+			status, detail = "error", truncateErr(fmt.Sprintf("%s: 不可按需调用,已从模型池剔除", m.Name))
 		case code == 401 || code == 403 || code == 404:
 			return "no_quota", truncateErr(oci.CompactError(err))
 		default:
 			status, detail = "error", truncateErr(fmt.Sprintf("%s: %s", m.Name, oci.CompactError(err)))
+			if errBudget--; errBudget == 0 {
+				return status, detail
+			}
 		}
 	}
 	return status, detail
 }
 
-// probeCandidates 只取对话模型并按可靠度排序取前 3:主流文本模型优先,
+// markModelUnusable 把 (渠道, 模型OCID) 标记为不可按需调用;失败仅记日志不阻断主流程。
-// voice 等非常规形态殿后(embedding / rerank 已被能力筛选排除)。
+func (s *AiGatewayService) markModelUnusable(ctx context.Context, channelID uint, ocid, reason string) {
+	err := s.db.WithContext(ctx).Model(&model.AiModelCache{}).
+		Where("channel_id = ? AND model_ocid = ?", channelID, ocid).
+		Updates(map[string]any{"unusable": true, "unusable_reason": shortReason(reason), "checked_at": time.Now()}).Error
+	if err != nil {
+		log.Printf("mark model unusable: %v", err)
+	}
+}
+
+func shortReason(reason string) string {
+	if len(reason) > 200 {
+		return reason[:200]
+	}
+	return reason
+}
+
+// beginValidate 抢占渠道的验证执行权,同渠道同时只跑一轮。
+func (s *AiGatewayService) beginValidate(id uint) bool {
+	s.validateMu.Lock()
+	defer s.validateMu.Unlock()
+	if s.validating[id] {
+		return false
+	}
+	s.validating[id] = true
+	return true
+}
+
+func (s *AiGatewayService) endValidate(id uint) {
+	s.validateMu.Lock()
+	delete(s.validating, id)
+	s.validateMu.Unlock()
+}
+
+// validateTargets 取待验证行:从未验证的新模型,以及标记超过复检间隔的模型(仅对话能力,
+// 兼容存量空串);单轮上限 validateBatchCap 控制调用量。
+func (s *AiGatewayService) validateTargets(ctx context.Context, channelID uint) ([]model.AiModelCache, error) {
+	stale := time.Now().Add(-modelRecheckGap)
+	var rows []model.AiModelCache
+	err := s.db.WithContext(ctx).
+		Where("channel_id = ? AND capability IN ?", channelID, []string{"CHAT", ""}).
+		Where("checked_at IS NULL OR (unusable = ? AND checked_at < ?)", true, stale).
+		Order("id ASC").Limit(validateBatchCap).Find(&rows).Error
+	return rows, err
+}
+
+// validateChannelModels 逐个试调渠道内待验证模型并落结论,把不可按需调用的模型
+// 从池中剔除、把恢复供给的解除标记;ListModels 无字段可事先判别,只能试调习得。
+func (s *AiGatewayService) validateChannelModels(ctx context.Context, channelID uint) {
+	if !s.beginValidate(channelID) {
+		return
+	}
+	defer s.endValidate(channelID)
+	var ch model.AiChannel
+	if err := s.db.WithContext(ctx).First(&ch, channelID).Error; err != nil {
+		return
+	}
+	cred, err := s.configs.credentialsByID(ctx, ch.OciConfigID)
+	if err != nil {
+		return
+	}
+	rows, err := s.validateTargets(ctx, channelID)
+	if err != nil {
+		log.Printf("validate channel %d models: %v", channelID, err)
+		return
+	}
+	for _, r := range rows {
+		s.validateOne(ctx, cred, &ch, r)
+	}
+}
+
+// validateOne 试调单个模型并落库:可用(200/429)清除标记;模型级不可用标记剔除;
+// 其余 4xx 记录已检不改状态;5xx/网络类瞬态不落 checked_at,下轮再验。
+func (s *AiGatewayService) validateOne(ctx context.Context, cred oci.Credentials, ch *model.AiChannel, r model.AiModelCache) {
+	code, err := s.client.GenAiProbeChat(ctx, cred, ch.Region, r.ModelOcid, r.Name)
+	updates := map[string]any{"checked_at": time.Now()}
+	switch {
+	case code == 200 || code == 429:
+		updates["unusable"], updates["unusable_reason"] = false, ""
+	case oci.IsModelUnavailable(err):
+		updates["unusable"], updates["unusable_reason"] = true, shortReason(oci.CompactError(err))
+	case code == 0 || code >= 500:
+		return
+	}
+	s.db.WithContext(ctx).Model(&model.AiModelCache{}).Where("id = ?", r.ID).Updates(updates)
+}
+
+// validateModelsAsync 后台验证渠道模型池(手动同步后触发,不阻塞接口响应)。
+func (s *AiGatewayService) validateModelsAsync(ctx context.Context, channelID uint) {
+	s.wg.Add(1)
+	go func() {
+		defer s.wg.Done()
+		vctx, cancel := context.WithTimeout(context.WithoutCancel(ctx), 3*time.Minute)
+		defer cancel()
+		s.validateChannelModels(vctx, channelID)
+	}()
+}
+
+// probeCandidateCap 是单次探测的候选上限:模型级不可用会被标记剔除不再重试,
+// 放宽到 8 让一次探测有机会越过整批坏模型找到可用者;其他错误另有 3 次止损预算。
+const probeCandidateCap = 8
+
+// probeCandidates 只取对话模型,按可靠度排序后跨厂商取候选:主流文本模型优先;
+// voice 等负分形态(元数据标 CHAT 但实际不可对话)直接排除,不浪费试调预算;
+// 每厂商先取最高分再按分数补位——部分区域某厂商全为微调基座(调用必失败),
+// 不能让单一厂商占满候选名额拖垮整个渠道的探测结论。
 func probeCandidates(models []oci.GenAiModel) []oci.GenAiModel {
 	var sorted []oci.GenAiModel
 	for _, m := range models {
-		if m.Capability == "" || m.Capability == "CHAT" {
+		if (m.Capability == "" || m.Capability == "CHAT") && probeScore(m.Name) >= 0 {
 			sorted = append(sorted, m)
 		}
 	}
 	sort.SliceStable(sorted, func(i, j int) bool {
 		return probeScore(sorted[i].Name) > probeScore(sorted[j].Name)
 	})
-	if len(sorted) > 3 {
+	return diversifyByVendor(sorted, probeCandidateCap)
-		sorted = sorted[:3]
+}
+
+// diversifyByVendor 从已排序列表先每厂商各取一个,不足 limit 再按原序补位。
+func diversifyByVendor(sorted []oci.GenAiModel, limit int) []oci.GenAiModel {
+	picked := make([]oci.GenAiModel, 0, limit)
+	used := make(map[int]bool)
+	seenVendor := make(map[string]bool)
+	for i, m := range sorted {
+		if len(picked) >= limit {
+			break
+		}
+		if v := modelVendor(m); !seenVendor[v] {
+			seenVendor[v] = true
+			used[i] = true
+			picked = append(picked, m)
+		}
 	}
-	return sorted
+	for i, m := range sorted {
+		if len(picked) >= limit {
+			break
+		}
+		if !used[i] {
+			picked = append(picked, m)
+		}
+	}
+	return picked
+}
+
+// modelVendor 取厂商标识;OCI 未回填 vendor 时退化为模型名「.」前缀。
+func modelVendor(m oci.GenAiModel) string {
+	if m.Vendor != "" {
+		return strings.ToLower(m.Vendor)
+	}
+	name := strings.ToLower(m.Name)
+	if i := strings.Index(name, "."); i > 0 {
+		return name[:i]
+	}
+	return name
 }
 
 func probeScore(name string) int {
@@ -377,7 +551,8 @@ func truncateErr(msg string) string {
 	return msg
 }
 
-// SyncModels 重新拉取渠道区域的模型列表并覆盖缓存。
+// SyncModels 重新拉取渠道区域的模型列表并覆盖缓存(标记按 OCID 结转),
+// 随后触发后台验证:新模型逐个试调,不可按需调用的数十秒内从池中剔除。
 func (s *AiGatewayService) SyncModels(ctx context.Context, id uint) ([]model.AiModelCache, error) {
 	var ch model.AiChannel
 	if err := s.db.WithContext(ctx).First(&ch, id).Error; err != nil {
@@ -394,16 +569,26 @@ func (s *AiGatewayService) SyncModels(ctx context.Context, id uint) ([]model.AiM
 	if err := s.replaceModels(ctx, id, models); err != nil {
 		return nil, err
 	}
+	s.validateModelsAsync(ctx, id)
 	return s.channelModels(ctx, id)
 }
 
-// replaceModels 以事务整组覆盖渠道模型缓存。
+// replaceModels 以事务整组覆盖渠道模型缓存,按 OCID 结转不可用标记与验证时间
+// (OCID 变化视为新条目,自然回到待验证状态)。
 func (s *AiGatewayService) replaceModels(ctx context.Context, channelID uint, models []oci.GenAiModel) error {
+	marks, err := s.loadModelMarks(ctx, channelID)
+	if err != nil {
+		return err
+	}
 	rows := make([]model.AiModelCache, 0, len(models))
 	now := time.Now()
 	for _, m := range models {
-		rows = append(rows, model.AiModelCache{ChannelID: channelID, ModelOcid: m.Ocid, Name: m.Name, Vendor: m.Vendor,
+		row := model.AiModelCache{ChannelID: channelID, ModelOcid: m.Ocid, Name: m.Name, Vendor: m.Vendor,
-			Capability: m.Capability, SyncedAt: now, DeprecatedAt: m.Deprecated, RetiredAt: m.Retired})
+			Capability: m.Capability, SyncedAt: now, DeprecatedAt: m.Deprecated, RetiredAt: m.Retired}
+		if prev, ok := marks[m.Ocid]; ok {
+			row.Unusable, row.UnusableReason, row.CheckedAt = prev.Unusable, prev.UnusableReason, prev.CheckedAt
+		}
+		rows = append(rows, row)
 	}
 	return s.db.WithContext(ctx).Transaction(func(tx *gorm.DB) error {
 		if err := tx.Where("channel_id = ?", channelID).Delete(&model.AiModelCache{}).Error; err != nil {
@@ -416,17 +601,33 @@ func (s *AiGatewayService) replaceModels(ctx context.Context, channelID uint, mo
 	})
 }
 
+// loadModelMarks 取渠道内带标记或已验证过的行(OCID → 行),供同步结转与候选过滤。
+func (s *AiGatewayService) loadModelMarks(ctx context.Context, channelID uint) (map[string]model.AiModelCache, error) {
+	var rows []model.AiModelCache
+	err := s.db.WithContext(ctx).Select("model_ocid", "unusable", "unusable_reason", "checked_at").
+		Where("channel_id = ? AND (unusable = ? OR checked_at IS NOT NULL)", channelID, true).Find(&rows).Error
+	if err != nil {
+		return nil, fmt.Errorf("load model marks: %w", err)
+	}
+	marks := make(map[string]model.AiModelCache, len(rows))
+	for _, r := range rows {
+		marks[r.ModelOcid] = r
+	}
+	return marks, nil
+}
+
 func (s *AiGatewayService) channelModels(ctx context.Context, channelID uint) ([]model.AiModelCache, error) {
 	var rows []model.AiModelCache
 	err := s.db.WithContext(ctx).Where("channel_id = ?", channelID).Order("name ASC").Find(&rows).Error
 	return rows, err
 }
 
-// GatewayModels 聚合启用渠道的模型(按名称去重),供 /ai/v1/models;
+// GatewayModels 聚合启用渠道的可用模型(按名称去重,剔除不可按需调用标记),
-// group 非空时仅聚合该分组渠道(与密钥分组路由口径一致)。
+// 供 /ai/v1/models;group 非空时仅聚合该分组渠道(与密钥分组路由口径一致)。
 func (s *AiGatewayService) GatewayModels(ctx context.Context, group string) (aiwire.ModelList, error) {
 	q := s.db.WithContext(ctx).
-		Joins("JOIN ai_channels ON ai_channels.id = ai_model_caches.channel_id AND ai_channels.enabled = ?", true)
+		Joins("JOIN ai_channels ON ai_channels.id = ai_model_caches.channel_id AND ai_channels.enabled = ?", true).
+		Where("(ai_model_caches.unusable = ? OR ai_model_caches.unusable IS NULL)", false)
 	if group != "" {
 		q = q.Where("ai_channels.channel_group = ?", group)
 	}
@@ -457,6 +658,7 @@ func (s *AiGatewayService) DeprecatingModels(ctx context.Context, within time.Du
 	err := s.db.WithContext(ctx).
 		Where("(retired_at IS NOT NULL AND retired_at > ? AND retired_at <= ?) OR (deprecated_at IS NOT NULL AND deprecated_at >= ? AND deprecated_at <= ?)",
 			now, deadline, now, deadline).
+		Where("(unusable = ? OR unusable IS NULL)", false).
 		Order("name ASC").Find(&rows).Error
 	if err != nil {
 		return nil, err
@@ -477,7 +679,7 @@ func (s *AiGatewayService) DeprecatingModels(ctx context.Context, within time.Du
 	return out, nil
 }
 
-// ProbeAll 逐个探测全部渠道,返回状态汇总;供 AI 探测后台任务调用。
+// ProbeAll 逐个探测全部渠道并顺带验证模型池,返回状态汇总;供 AI 探测后台任务调用。
 func (s *AiGatewayService) ProbeAll(ctx context.Context) (string, error) {
 	chs, err := s.Channels(ctx)
 	if err != nil {
@@ -493,6 +695,8 @@ func (s *AiGatewayService) ProbeAll(ctx context.Context) (string, error) {
 			continue
 		}
 		counts[fresh.ProbeStatus]++
+		// 后台任务里同步执行:新模型验证入池、坏模型定期复检,零人工收敛
+		s.validateChannelModels(ctx, ch.ID)
 	}
 	msg := fmt.Sprintf("probed %d: %d ok, %d no_service, %d no_quota, %d error",
 		len(chs), counts["ok"], counts["no_service"], counts["no_quota"], counts["error"])

internal/service/aigateway_chat.go

@@ -43,10 +43,13 @@ func (s *AiGatewayService) Chat(ctx context.Context, ir aiwire.ChatRequest, grou
 			s.markSuccess(ctx, cand.ch.ID)
 			return resp, meta, nil
 		}
-		if !retryable(err) {
+		retry, penalize := s.noteCallErr(ctx, cand, err)
+		if !retry {
 			return nil, meta, err
 		}
-		s.markFailure(ctx, cand.ch.ID)
+		if penalize {
+			s.markFailure(ctx, cand.ch.ID)
+		}
 		excluded[cand.ch.ID] = true
 		meta.Retries++
 		lastErr = err
@@ -83,10 +86,13 @@ func (s *AiGatewayService) OpenStream(ctx context.Context, ir aiwire.ChatRequest
 			s.markSuccess(ctx, cand.ch.ID)
 			return stream, meta, nil
 		}
-		if !retryable(err) {
+		retry, penalize := s.noteCallErr(ctx, cand, err)
+		if !retry {
 			return nil, meta, err
 		}
-		s.markFailure(ctx, cand.ch.ID)
+		if penalize {
+			s.markFailure(ctx, cand.ch.ID)
+		}
 		excluded[cand.ch.ID] = true
 		meta.Retries++
 		lastErr = err
@@ -102,12 +108,18 @@ func firstErr(lastErr, pickErr error) error {
 	return pickErr
 }
 
-// retryable 判定是否换渠道重试:429 / 5xx / 网络错误可重试,其余 4xx 直接透传。
+// noteCallErr 汇总一次调用失败:模型级不可用(微调基座 400 / 实体不存在 404)先标记
-func retryable(err error) bool {
+// 剔除该 (渠道, 模型),换渠道重试且不计熔断——这是模型×区域供给问题而非渠道健康问题;
-	if status, ok := oci.ServiceStatus(err); ok {
+// 429 / 5xx / 网络错误换渠道并计熔断;其余 4xx 直接透传。
-		return status == 429 || status >= 500
+func (s *AiGatewayService) noteCallErr(ctx context.Context, cand *aiCandidate, err error) (retry, penalize bool) {
+	if oci.IsModelUnavailable(err) {
+		s.markModelUnusable(ctx, cand.ch.ID, cand.modelOcid, oci.CompactError(err))
+		return true, false
 	}
-	return true
+	if status, ok := oci.ServiceStatus(err); ok {
+		return status == 429 || status >= 500, true
+	}
+	return true, true
 }
 
 // pick 选出支持该模型的最优渠道:能力匹配 → 启用 → 分组匹配 → 未熔断 → 最小优先级组 → 加权随机。
@@ -142,10 +154,11 @@ func (s *AiGatewayService) pick(ctx context.Context, modelName, group, capabilit
 	return &aiCandidate{ch: chosen, modelOcid: ocids[chosen.ID]}, nil
 }
 
-// modelChannels 查出提供该模型的渠道 ID 及各自的模型 OCID;
+// modelChannels 查出提供该模型的渠道 ID 及各自的模型 OCID(剔除不可按需调用标记);
-// capability=CHAT 时兼容存量空串(加列前只同步对话模型)。
+// capability=CHAT 时兼容存量空串(加列前只同步对话模型);unusable 需兼容 NULL
+// (AutoMigrate 加列后、首次重同步前的存量行)。
 func (s *AiGatewayService) modelChannels(ctx context.Context, modelName, capability string) (map[uint]string, []uint, error) {
-	q := s.db.WithContext(ctx).Where("name = ?", modelName)
+	q := s.db.WithContext(ctx).Where("name = ? AND (unusable = ? OR unusable IS NULL)", modelName, false)
 	if capability == "CHAT" {
 		q = q.Where("capability IN ?", []string{"CHAT", ""})
 	} else {
@@ -221,10 +234,13 @@ func (s *AiGatewayService) Embeddings(ctx context.Context, req aiwire.Embeddings
 			s.markSuccess(ctx, cand.ch.ID)
 			return resp, meta, nil
 		}
-		if !retryable(err) {
+		retry, penalize := s.noteCallErr(ctx, cand, err)
+		if !retry {
 			return nil, meta, err
 		}
-		s.markFailure(ctx, cand.ch.ID)
+		if penalize {
+			s.markFailure(ctx, cand.ch.ID)
+		}
 		excluded[cand.ch.ID] = true
 		meta.Retries++
 		lastErr = err

internal/service/aigateway_test.go

@@ -3,6 +3,7 @@ package service
 import (
 	"context"
 	"errors"
+	"fmt"
 	"strings"
 	"testing"
 	"time"
@@ -13,14 +14,28 @@ import (
 )
 
 // stubServiceError 实现 common.ServiceError,用于模拟带状态码的 OCI 服务端错误。
-type stubServiceError struct{ status int }
+type stubServiceError struct {
+	status int
+	msg    string
+}
 
-func (e stubServiceError) Error() string           { return "stub service error" }
+func (e stubServiceError) Error() string          { return "stub service error" }
-func (e stubServiceError) GetHTTPStatusCode() int  { return e.status }
+func (e stubServiceError) GetHTTPStatusCode() int { return e.status }
-func (e stubServiceError) GetMessage() string      { return "stub" }
+func (e stubServiceError) GetMessage() string {
+	if e.msg != "" {
+		return e.msg
+	}
+	return "stub"
+}
 func (e stubServiceError) GetCode() string         { return "Stub" }
 func (e stubServiceError) GetOpcRequestID() string { return "req-1" }
 
+// finetuneBaseErr 模拟「微调基座模型不可按需调用」的 OCI 400。
+func finetuneBaseErr() stubServiceError {
+	return stubServiceError{status: 400,
+		msg: "Not allowed to call finetune base model ocid1.generativeaimodel.oc1.eu-frankfurt-1.tpel5q, use Endpoint: false"}
+}
+
 // gatewayStubClient 覆写 GenAI 四方法;chatErrs 逐次弹出以模拟先失败后成功。
 type gatewayStubClient struct {
 	*fakeClient
@@ -29,6 +44,8 @@ type gatewayStubClient struct {
 	modelsErr error
 	probeCode int
 	probeErr  error
+	// probeSeq 非空时逐次弹出,模拟按候选依次试调;弹尽后回落 probeCode/probeErr
+	probeSeq  []probeResult
 	chatResp  *aiwire.ChatResponse
 	chatErrs  []error
 	chatCalls int
@@ -47,7 +64,18 @@ func (f *gatewayStubClient) ListGenAiModels(ctx context.Context, cred oci.Creden
 	return f.models, f.modelsErr
 }
 
+// probeResult 是 gatewayStubClient.probeSeq 的单次探测结果。
+type probeResult struct {
+	code int
+	err  error
+}
+
 func (f *gatewayStubClient) GenAiProbeChat(ctx context.Context, cred oci.Credentials, region, modelOcid, modelName string) (int, error) {
+	if len(f.probeSeq) > 0 {
+		r := f.probeSeq[0]
+		f.probeSeq = f.probeSeq[1:]
+		return r.code, r.err
+	}
 	return f.probeCode, f.probeErr
 }
 
@@ -400,6 +428,263 @@ func TestProbeCandidates(t *testing.T) {
 	}
 }
 
+func TestProbeCandidatesVendorDiversity(t *testing.T) {
+	// 部分区域单一厂商全为微调基座:候选须跨厂商分散,不能被 3 个 llama 占满前排
+	models := []oci.GenAiModel{
+		{Ocid: "l1", Name: "meta.llama-3-70b-instruct"},
+		{Ocid: "l2", Name: "meta.llama-3.1-405b-instruct"},
+		{Ocid: "l3", Name: "meta.llama-3.3-70b-instruct"},
+		{Ocid: "c1", Name: "cohere.command-a-03-2025"},
+		{Ocid: "g1", Name: "xai.grok-4"},
+	}
+	got := probeCandidates(models)
+	if len(got) != 5 || got[0].Name != "meta.llama-3-70b-instruct" {
+		t.Fatalf("上限内全量返回且最高分居首: %+v", got)
+	}
+	vendors := map[string]bool{}
+	for _, m := range got[:3] {
+		vendors[modelVendor(m)] = true
+	}
+	if len(vendors) != 3 {
+		t.Errorf("前 3 个候选应覆盖 3 个厂商: %+v", got)
+	}
+	// 超过上限时截断到 probeCandidateCap
+	var many []oci.GenAiModel
+	for i := 0; i < 12; i++ {
+		many = append(many, oci.GenAiModel{Ocid: fmt.Sprintf("m%d", i), Name: fmt.Sprintf("meta.llama-%d", i)})
+	}
+	if capped := probeCandidates(many); len(capped) != probeCandidateCap {
+		t.Errorf("候选应截断到 %d: got %d", probeCandidateCap, len(capped))
+	}
+}
+
+// entityNotFoundErr 模拟「实体不存在」404(模型在区域内无按需供给)。
+func entityNotFoundErr() stubServiceError {
+	return stubServiceError{status: 404,
+		msg: "Entity with key ocid1.generativeaimodel.oc1.eu-frankfurt-1.2flsfq not found"}
+}
+
+func TestProbeSkipsUnavailableModels(t *testing.T) {
+	// 首选 llama 不可按需调用(基座 400 / 实体 404)→ 标记剔除 → 换候选成功 → 渠道判可用
+	tests := []struct {
+		name string
+		bad  probeResult
+	}{
+		{"微调基座 400", probeResult{400, finetuneBaseErr()}},
+		{"实体不存在 404", probeResult{404, entityNotFoundErr()}},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			client := &gatewayStubClient{
+				fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
+				models: []oci.GenAiModel{
+					{Ocid: "m1", Name: "meta.llama-3-70b-instruct", Vendor: "meta"},
+					{Ocid: "m2", Name: "meta.llama-3.1-70b-instruct", Vendor: "meta"},
+					{Ocid: "m3", Name: "cohere.command-a-03-2025", Vendor: "cohere"},
+				},
+				probeSeq: []probeResult{tt.bad, {200, nil}},
+			}
+			gw, svc := newTestGateway(t, client)
+			cfg := importAliveConfig(t, svc)
+			ctx := context.Background()
+
+			ch, err := gw.CreateChannel(ctx, ChannelInput{OciConfigID: cfg.ID, Region: "eu-frankfurt-1"})
+			if err != nil {
+				t.Fatalf("CreateChannel: %v", err)
+			}
+			probed, err := gw.ProbeChannel(ctx, ch.ID)
+			if err != nil || probed.ProbeStatus != "ok" {
+				t.Fatalf("坏候选后应换候选并判可用: %+v, %v", probed, err)
+			}
+			var row model.AiModelCache
+			if err := gw.db.Where("channel_id = ? AND model_ocid = ?", ch.ID, "m1").First(&row).Error; err != nil || !row.Unusable {
+				t.Fatalf("首个坏候选应被标记不可用: %+v, %v", row, err)
+			}
+			// 再次探测:同步保留标记,m1 不再进候选(probeSeq 只需一次 200)
+			client.probeSeq = []probeResult{{200, nil}}
+			probed, err = gw.ProbeChannel(ctx, ch.ID)
+			if err != nil || probed.ProbeStatus != "ok" {
+				t.Fatalf("复测应跳过已标记模型: %+v, %v", probed, err)
+			}
+			var again model.AiModelCache
+			gw.db.Where("channel_id = ? AND model_ocid = ?", ch.ID, "m1").First(&again)
+			if !again.Unusable {
+				t.Error("探测触发的同步应保留不可用标记")
+			}
+			// 手动同步同样保留标记(坏模型不随重新同步复活);近期已检的标记不被后台验证翻转
+			client.probeCode = 200
+			if _, err := gw.SyncModels(ctx, ch.ID); err != nil {
+				t.Fatalf("SyncModels: %v", err)
+			}
+			gw.Wait()
+			var kept model.AiModelCache
+			gw.db.Where("channel_id = ? AND model_ocid = ?", ch.ID, "m1").First(&kept)
+			if !kept.Unusable {
+				t.Error("手动同步不应清除不可用标记")
+			}
+		})
+	}
+}
+
+func TestValidateModelsAfterSync(t *testing.T) {
+	// 同步后后台验证:坏模型标记剔除、好模型记录已检、其他 4xx 不改状态、非对话模型不试调
+	client := &gatewayStubClient{
+		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
+		models: []oci.GenAiModel{
+			{Ocid: "v1", Name: "meta.llama-4-maverick", Vendor: "meta", Capability: "CHAT"},
+			{Ocid: "v2", Name: "xai.grok-4", Vendor: "xai", Capability: "CHAT"},
+			{Ocid: "v3", Name: "xai.grok-voice-agent", Vendor: "xai", Capability: "CHAT"},
+			{Ocid: "v4", Name: "cohere.embed-v4.0", Vendor: "cohere", Capability: "EMBEDDING"},
+		},
+		probeSeq: []probeResult{{404, entityNotFoundErr()}, {200, nil}, {400, stubServiceError{status: 400}}},
+	}
+	gw, svc := newTestGateway(t, client)
+	cfg := importAliveConfig(t, svc)
+	ctx := context.Background()
+	ch, err := gw.CreateChannel(ctx, ChannelInput{OciConfigID: cfg.ID, Region: "us-ashburn-1"})
+	if err != nil {
+		t.Fatalf("CreateChannel: %v", err)
+	}
+	if _, err := gw.SyncModels(ctx, ch.ID); err != nil {
+		t.Fatalf("SyncModels: %v", err)
+	}
+	gw.Wait()
+
+	want := map[string]struct {
+		unusable bool
+		checked  bool
+	}{
+		"v1": {true, true},   // 实体 404 → 标记剔除
+		"v2": {false, true},  // 200 → 可用已检
+		"v3": {false, true},  // 普通 400 → 已检不标记
+		"v4": {false, false}, // EMBEDDING 不试调
+	}
+	rows, _ := gw.channelModels(ctx, ch.ID)
+	for _, r := range rows {
+		w := want[r.ModelOcid]
+		if r.Unusable != w.unusable || (r.CheckedAt != nil) != w.checked {
+			t.Errorf("%s: unusable=%v checked=%v, want %+v", r.ModelOcid, r.Unusable, r.CheckedAt != nil, w)
+		}
+	}
+	list, err := gw.GatewayModels(ctx, "")
+	if err != nil || len(list.Data) != 3 {
+		t.Errorf("网关列表应只剔除被标记的坏模型(余 v2/v3/v4), got %+v, %v", list.Data, err)
+	}
+}
+
+func TestValidateRecheckUnmarksRecovered(t *testing.T) {
+	// 已标记模型超过复检间隔后重验:恢复供给(200)自动解除标记
+	client := &gatewayStubClient{fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}}, probeCode: 200}
+	gw, svc := newTestGateway(t, client)
+	cfg := importAliveConfig(t, svc)
+	ch := seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
+	ctx := context.Background()
+	old := time.Now().Add(-25 * time.Hour)
+	gw.db.Model(&model.AiModelCache{}).Where("channel_id = ?", ch.ID).
+		Updates(map[string]any{"unusable": true, "unusable_reason": "x", "checked_at": old})
+
+	gw.validateChannelModels(ctx, ch.ID)
+	var row model.AiModelCache
+	gw.db.Where("channel_id = ?", ch.ID).First(&row)
+	if row.Unusable || row.UnusableReason != "" || row.CheckedAt == nil || !row.CheckedAt.After(old) {
+		t.Errorf("超期复检应解除标记并刷新已检时间: %+v", row)
+	}
+	// 未超期的标记不复检(probeSeq 为空、fallback 200 也不会被消费)
+	fresh := time.Now()
+	gw.db.Model(&model.AiModelCache{}).Where("channel_id = ?", ch.ID).
+		Updates(map[string]any{"unusable": true, "checked_at": fresh})
+	gw.validateChannelModels(ctx, ch.ID)
+	gw.db.Where("channel_id = ?", ch.ID).First(&row)
+	if !row.Unusable {
+		t.Error("未超期的标记不应被复检翻转")
+	}
+}
+
+func TestProbeAuth404StillNoQuota(t *testing.T) {
+	// 鉴权类 404(NotAuthorizedOrNotFound)仍属租户级,直接定论 no_quota 且不标记模型
+	client := &gatewayStubClient{
+		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
+		models:     []oci.GenAiModel{{Ocid: "m1", Name: "meta.llama-3.3-70b-instruct", Vendor: "meta"}},
+		probeCode:  404,
+		probeErr:   stubServiceError{status: 404, msg: "Authorization failed or requested resource not found."},
+	}
+	gw, svc := newTestGateway(t, client)
+	cfg := importAliveConfig(t, svc)
+	ctx := context.Background()
+
+	ch, _ := gw.CreateChannel(ctx, ChannelInput{OciConfigID: cfg.ID, Region: "eu-frankfurt-1"})
+	probed, _ := gw.ProbeChannel(ctx, ch.ID)
+	if probed.ProbeStatus != "no_quota" {
+		t.Errorf("鉴权 404 status = %q, want no_quota", probed.ProbeStatus)
+	}
+	var marked int64
+	gw.db.Model(&model.AiModelCache{}).Where("unusable = ?", true).Count(&marked)
+	if marked != 0 {
+		t.Errorf("鉴权 404 不应标记模型, marked=%d", marked)
+	}
+}
+
+func TestAiChatFinetuneSwitchesChannelWithoutPenalty(t *testing.T) {
+	// 微调基座 400 换渠道重试成功,且不计入熔断失败
+	client := &gatewayStubClient{
+		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
+		chatResp:   &aiwire.ChatResponse{Model: "meta.llama-3.3-70b-instruct", Choices: []aiwire.Choice{{Message: aiwire.ChatMessage{Role: "assistant", Content: aiwire.NewTextContent("hi")}, FinishReason: "stop"}}},
+		chatErrs:   []error{finetuneBaseErr()},
+	}
+	gw, svc := newTestGateway(t, client)
+	cfg := importAliveConfig(t, svc)
+	seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
+	seedChannel(t, gw, cfg.ID, "us-chicago-1", 1, 1)
+
+	resp, meta, err := gw.Chat(context.Background(), aiwire.ChatRequest{Model: "meta.llama-3.3-70b-instruct", Messages: []aiwire.ChatMessage{{Role: "user", Content: aiwire.NewTextContent("你好")}}}, "")
+	if err != nil || resp == nil {
+		t.Fatalf("Chat = %v, %v", resp, err)
+	}
+	if meta.Retries != 1 || client.chatCalls != 2 {
+		t.Errorf("应换渠道重试一次: retries=%d calls=%d", meta.Retries, client.chatCalls)
+	}
+	var chs []model.AiChannel
+	gw.db.Find(&chs)
+	for _, ch := range chs {
+		if ch.FailCount != 0 {
+			t.Errorf("微调基座 400 不应计入熔断: 渠道 %s failCount=%d", ch.Name, ch.FailCount)
+		}
+	}
+	// 失败渠道的该模型被标记,不再参与路由;成功渠道不受影响
+	region := client.regions[0]
+	var row model.AiModelCache
+	gw.db.Where("model_ocid = ?", "ocid1..m-"+region).First(&row)
+	if !row.Unusable {
+		t.Errorf("失败渠道的模型应被标记不可用: %+v", row)
+	}
+	var usable int64
+	gw.db.Model(&model.AiModelCache{}).Where("unusable = ?", false).Count(&usable)
+	if usable != 1 {
+		t.Errorf("成功渠道模型应保持可用, usable=%d", usable)
+	}
+}
+
+func TestUnusableModelExcludedFromPoolAndRouting(t *testing.T) {
+	// 唯一渠道的模型被标记后:网关列表不再展示,路由按未知模型拒绝,渠道详情仍可见标记
+	gw, svc := newTestGateway(t, &gatewayStubClient{fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}}})
+	cfg := importAliveConfig(t, svc)
+	ch := seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
+	ctx := context.Background()
+
+	gw.markModelUnusable(ctx, ch.ID, "ocid1..m-eu-frankfurt-1", "Entity with key … not found")
+	list, err := gw.GatewayModels(ctx, "")
+	if err != nil || len(list.Data) != 0 {
+		t.Errorf("已标记模型不应出现在网关列表: %+v, %v", list.Data, err)
+	}
+	if _, _, err := gw.Chat(ctx, aiwire.ChatRequest{Model: "meta.llama-3.3-70b-instruct"}, ""); !errors.Is(err, ErrAiUnknownModel) {
+		t.Errorf("已标记模型路由应拒绝: %v", err)
+	}
+	rows, err := gw.channelModels(ctx, ch.ID)
+	if err != nil || len(rows) != 1 || !rows[0].Unusable || rows[0].UnusableReason == "" {
+		t.Errorf("渠道详情应保留标记行便于排查: %+v, %v", rows, err)
+	}
+}
+
 func TestDeprecatingModels(t *testing.T) {
 	gw, _ := newTestGateway(t, &gatewayStubClient{fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}}})
 	ctx := context.Background()
@@ -507,3 +792,24 @@ func TestAiContentLogSwitch(t *testing.T) {
 		t.Fatalf("关闭失败: %+v, %v", fresh.ContentLogUntil, err)
 	}
 }
+
+func TestLegacyNullUnusableRowsStayVisible(t *testing.T) {
+	// 升级路径回归:AutoMigrate 加列后、首次重同步前,存量行 unusable 为 NULL,
+	// 网关列表与路由必须照常包含这些行,不能因 unusable = false 过滤而整池消失
+	gw, svc := newTestGateway(t, &gatewayStubClient{fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}}})
+	cfg := importAliveConfig(t, svc)
+	ch := seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
+	ctx := context.Background()
+	if err := gw.db.Exec("UPDATE ai_model_caches SET unusable = NULL WHERE channel_id = ?", ch.ID).Error; err != nil {
+		t.Fatalf("set legacy null: %v", err)
+	}
+
+	list, err := gw.GatewayModels(ctx, "")
+	if err != nil || len(list.Data) != 1 {
+		t.Errorf("存量 NULL 行应仍在网关列表: %+v, %v", list.Data, err)
+	}
+	_, ids, err := gw.modelChannels(ctx, "meta.llama-3.3-70b-instruct", "CHAT")
+	if err != nil || len(ids) != 1 {
+		t.Errorf("存量 NULL 行应仍参与路由: %v, %v", ids, err)
+	}
+}

internal/service/ociconfig_test.go

@@ -137,6 +137,13 @@ func newTestService(t *testing.T, client oci.Client) *OciConfigService {
 	if err != nil {
 		t.Fatalf("open in-memory sqlite: %v", err)
 	}
+	// :memory: 库每条连接各自独立,后台 goroutine(模型验证等)触发第二条连接会看到空库;
+	// 收敛到单连接,与 tenantdelete 测试环境一致
+	sqlDB, err := db.DB()
+	if err != nil {
+		t.Fatalf("database handle: %v", err)
+	}
+	sqlDB.SetMaxOpenConns(1)
 	if err := db.AutoMigrate(
 		&model.OciConfig{}, &model.RegionCache{}, &model.CompartmentCache{}, &model.Proxy{},
 	); err != nil {

internal/service/task.go

@@ -3,6 +3,7 @@ package service
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"log"
 	"strings"
@@ -46,6 +47,12 @@ type TaskService struct {
 	entries map[uint]cron.EntryID
 	// runMu 让租户删除等待在途执行结束,并阻止新执行读取删除前的任务快照。
 	runMu sync.RWMutex
+	// runningMu 保护 running:同一任务不允许并发重复执行
+	// (cron 重叠触发静默跳过,手动触发返回 ErrTaskRunning)
+	runningMu sync.Mutex
+	running   map[uint]bool
+	// runWG 追踪手动触发的后台执行,Stop 时等待收尾
+	runWG sync.WaitGroup
 
 	// snatchVarsMu 保护 snatchVars:runSnatch 抢满时写入成功通知变量,
 	// execute 组装执行后快照时取走(一次性)
@@ -63,6 +70,7 @@ func NewTaskService(db *gorm.DB, configs *OciConfigService, notifier *Notifier,
 		settings:     settings,
 		cron:         cron.New(),
 		entries:      map[uint]cron.EntryID{},
+		running:      map[uint]bool{},
 		snatchIPWait: defaultSnatchIPWait,
 		snatchVars:   map[uint]map[string]string{},
 	}
@@ -86,10 +94,11 @@ func (s *TaskService) Start() error {
 	return nil
 }
 
-// Stop 停止调度并等待执行中的任务收尾，
+// Stop 停止调度并等待执行中的任务收尾(含手动触发的后台执行),
 // 保证任务产生的异步通知都已进入 Notifier 的等待队列。
 func (s *TaskService) Stop() {
 	<-s.cron.Stop().Done()
+	s.runWG.Wait()
 }
 
 // healthCheckPayload 是测活任务参数；ociConfigIds 为空表示全部配置。
@@ -356,7 +365,7 @@ func (s *TaskService) TaskLogs(ctx context.Context, id uint, limit int) ([]model
 	return logs, nil
 }
 
-// RunTaskNow 立即执行一次任务并返回本次日志。
+// RunTaskNow 同步执行一次任务并返回本次日志(内部与测试使用;API 走 TriggerTask)。
 func (s *TaskService) RunTaskNow(ctx context.Context, id uint) (*model.TaskLog, error) {
 	if _, err := s.GetTask(ctx, id); err != nil {
 		return nil, err
@@ -364,6 +373,43 @@ func (s *TaskService) RunTaskNow(ctx context.Context, id uint) (*model.TaskLog,
 	return s.execute(id), nil
 }
 
+// ErrTaskRunning 表示任务已有一次执行在途,拒绝重复触发。
+var ErrTaskRunning = errors.New("任务正在执行中,请稍候")
+
+// TriggerTask 异步触发一次任务执行并立即返回;执行结果照常落任务日志与通知,
+// 由前端轮询呈现。同一任务在途时返回 ErrTaskRunning。
+func (s *TaskService) TriggerTask(ctx context.Context, id uint) error {
+	if _, err := s.GetTask(ctx, id); err != nil {
+		return err
+	}
+	if !s.beginRun(id) {
+		return ErrTaskRunning
+	}
+	s.runWG.Add(1)
+	go func() {
+		defer s.runWG.Done()
+		s.runHeld(id)
+	}()
+	return nil
+}
+
+// beginRun 抢占任务的执行权;在途时返回 false。
+func (s *TaskService) beginRun(id uint) bool {
+	s.runningMu.Lock()
+	defer s.runningMu.Unlock()
+	if s.running[id] {
+		return false
+	}
+	s.running[id] = true
+	return true
+}
+
+func (s *TaskService) endRun(id uint) {
+	s.runningMu.Lock()
+	delete(s.running, id)
+	s.runningMu.Unlock()
+}
+
 // schedule 把任务注册进 cron 调度。
 func (s *TaskService) schedule(task *model.Task) error {
 	s.mu.Lock()
@@ -412,9 +458,18 @@ func (s *TaskService) lockTenantCleanup() func() {
 	return s.runMu.Unlock
 }
 
-// execute 执行一次任务：加载 → 分派 → 更新任务状态并写日志，
+// execute 执行一次任务:同一任务已有执行在途时静默跳过(cron 重叠触发防抖)。
-// 前后状态交给通知判定，只在状态变化时推送。
 func (s *TaskService) execute(taskID uint) *model.TaskLog {
+	if !s.beginRun(taskID) {
+		return nil
+	}
+	return s.runHeld(taskID)
+}
+
+// runHeld 在已持有执行权的前提下完成一次执行:加载 → 分派 → 更新任务状态并写日志,
+// 前后状态交给通知判定,只在状态变化时推送。
+func (s *TaskService) runHeld(taskID uint) *model.TaskLog {
+	defer s.endRun(taskID)
 	s.runMu.RLock()
 	defer s.runMu.RUnlock()
 	return s.executeLocked(taskID)

internal/service/task_test.go

@@ -3,6 +3,7 @@ package service
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"strings"
 	"sync"
@@ -801,3 +802,36 @@ func TestNotifyEventsSnatchVarsMerged(t *testing.T) {
 		}
 	}
 }
+
+func TestTriggerTaskAsyncAndDedup(t *testing.T) {
+	client := &blockingClient{fakeClient: &fakeClient{}, started: make(chan struct{}, 1), release: make(chan struct{})}
+	configs, tasks, db := newTenantDeleteEnv(t, client)
+	target, _ := seedDeleteTenants(t, db)
+	setTenantPrivateKey(t, configs, target.ID)
+	task := createHealthTask(t, tasks, target.ID)
+	ctx := context.Background()
+
+	// 触发立即返回,执行在后台开始
+	if err := tasks.TriggerTask(ctx, task.ID); err != nil {
+		t.Fatalf("TriggerTask: %v", err)
+	}
+	<-client.started
+	// 在途重复触发被拒;cron 重叠执行静默跳过
+	if err := tasks.TriggerTask(ctx, task.ID); !errors.Is(err, ErrTaskRunning) {
+		t.Errorf("在途重复触发应返回 ErrTaskRunning: %v", err)
+	}
+	if entry := tasks.execute(task.ID); entry != nil {
+		t.Error("在途时 cron 重叠执行应静默跳过")
+	}
+	close(client.release)
+	tasks.Stop() // 等待后台执行收尾
+	assertCount(t, db, &model.TaskLog{}, "task_id = ?", []any{task.ID}, 1)
+	// 执行权已释放,可再次触发;不存在的任务报错
+	if !tasks.beginRun(task.ID) {
+		t.Error("执行结束后应可重新获得执行权")
+	}
+	tasks.endRun(task.ID)
+	if err := tasks.TriggerTask(ctx, 9999); err == nil {
+		t.Error("不存在的任务应报错")
+	}
+}

internal/service/tenantdelete.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"log"
 	"strings"
 
 	"gorm.io/gorm"
@@ -104,7 +105,8 @@ func planTenantTask(task model.Task, id uint) (tenantTaskAction, bool, error) {
 func planSnatchTask(task model.Task, id uint) (tenantTaskAction, bool, error) {
 	var payload snatchPayload
 	if err := json.Unmarshal([]byte(task.Payload), &payload); err != nil {
-		return tenantTaskAction{}, false, fmt.Errorf("parse task %d payload: %w", task.ID, err)
+		logSkippedTask(task, err)
+		return tenantTaskAction{}, false, nil
 	}
 	if payload.OciConfigID != id {
 		return tenantTaskAction{}, false, nil
@@ -112,10 +114,17 @@ func planSnatchTask(task model.Task, id uint) (tenantTaskAction, bool, error) {
 	return tenantTaskAction{task: task, deleteTask: true}, true, nil
 }
 
+// logSkippedTask 记录 payload 无法解析而被跳过的任务:坏数据无法归属租户,
+// fail-closed 会永久阻断删除且只能手工修库;改为保留原任务(不删不改)并放行删除。
+func logSkippedTask(task model.Task, err error) {
+	log.Printf("tenant delete: 任务 %d(%s)payload 无法解析,跳过处理: %v", task.ID, task.Type, err)
+}
+
 func planMultiTenantTask(task model.Task, id uint) (tenantTaskAction, bool, error) {
 	payload, err := decodeConfigIDs(task)
 	if err != nil {
-		return tenantTaskAction{}, false, err
+		logSkippedTask(task, err)
+		return tenantTaskAction{}, false, nil
 	}
 	if len(payload.OciConfigIDs) == 0 {
 		return tenantTaskAction{task: task, payload: task.Payload}, true, nil
@@ -197,11 +206,11 @@ func resetTenantTask(tx *gorm.DB, action tenantTaskAction) error {
 }
 
 func deleteTenantEvents(tx *gorm.DB, id uint, result *tenantDeleteResult) error {
-	ruleIDs, eventIDs, affectedRules, err := loadTenantEventRefs(tx, id)
+	ruleIDs, affectedRules, err := loadTenantEventRefs(tx, id)
 	if err != nil {
 		return err
 	}
-	if err := deleteAlertHits(tx, ruleIDs, eventIDs); err != nil {
+	if err := deleteAlertHits(tx, id); err != nil {
 		return err
 	}
 	result.alertRuleIDs = mergeIDs(ruleIDs, affectedRules)
@@ -214,20 +223,19 @@ func deleteTenantEvents(tx *gorm.DB, id uint, result *tenantDeleteResult) error
 	return nil
 }
 
-func loadTenantEventRefs(tx *gorm.DB, id uint) ([]uint, []uint, []uint, error) {
+func loadTenantEventRefs(tx *gorm.DB, id uint) ([]uint, []uint, error) {
 	ruleIDs, err := lockedTenantRuleIDs(tx, id)
 	if err != nil {
-		return nil, nil, nil, fmt.Errorf("load tenant alert rules: %w", err)
+		return nil, nil, fmt.Errorf("load tenant alert rules: %w", err)
 	}
-	eventIDs, err := lockedTenantEventIDs(tx, id)
+	if err := lockTenantEventRows(tx, id); err != nil {
+		return nil, nil, fmt.Errorf("load tenant log events: %w", err)
+	}
+	affected, err := alertHitRuleIDs(tx, id)
 	if err != nil {
-		return nil, nil, nil, fmt.Errorf("load tenant log events: %w", err)
+		return nil, nil, err
 	}
-	affected, err := alertHitRuleIDs(tx, eventIDs)
+	return ruleIDs, affected, nil
-	if err != nil {
-		return nil, nil, nil, err
-	}
-	return ruleIDs, eventIDs, affected, nil
 }
 
 func lockedTenantRuleIDs(tx *gorm.DB, id uint) ([]uint, error) {
@@ -241,23 +249,24 @@ func lockedTenantRuleIDs(tx *gorm.DB, id uint) ([]uint, error) {
 	return ids, err
 }
 
-func lockedTenantEventIDs(tx *gorm.DB, id uint) ([]uint, error) {
+// lockTenantEventRows 对租户全部日志事件行加 FOR UPDATE 锁(SQLite 忽略,
+// MySQL/PG 阻塞并发命中写入);事件可达数万,ID 不再回传拼接 SQL,
+// 后续删除与关联查询一律用子查询,避免绑定变量上限。
+func lockTenantEventRows(tx *gorm.DB, id uint) error {
 	var rows []model.LogEvent
-	err := tx.Clauses(clause.Locking{Strength: "UPDATE"}).Select("id").
+	return tx.Clauses(clause.Locking{Strength: "UPDATE"}).Select("id").
 		Where("oci_config_id = ?", id).Order("id").Find(&rows).Error
-	ids := make([]uint, 0, len(rows))
-	for _, row := range rows {
-		ids = append(ids, row.ID)
-	}
-	return ids, err
 }
 
-func alertHitRuleIDs(tx *gorm.DB, eventIDs []uint) ([]uint, error) {
+// tenantEventIDs 构造「本租户日志事件 ID」子查询,供 IN (?) 内联。
-	if len(eventIDs) == 0 {
+func tenantEventIDs(tx *gorm.DB, id uint) *gorm.DB {
-		return nil, nil
+	return tx.Model(&model.LogEvent{}).Select("id").Where("oci_config_id = ?", id)
-	}
+}
+
+// alertHitRuleIDs 找出命中引用了本租户事件的规则 ID(去重,可能含他租户/全局规则)。
+func alertHitRuleIDs(tx *gorm.DB, id uint) ([]uint, error) {
 	ids := make([]uint, 0)
-	err := tx.Model(&model.AlertRuleHit{}).Where("log_event_id IN ?", eventIDs).
+	err := tx.Model(&model.AlertRuleHit{}).Where("log_event_id IN (?)", tenantEventIDs(tx, id)).
 		Distinct().Pluck("rule_id", &ids).Error
 	if err != nil {
 		return nil, fmt.Errorf("load affected alert rules: %w", err)
@@ -280,19 +289,13 @@ func mergeIDs(groups ...[]uint) []uint {
 	return out
 }
 
-func deleteAlertHits(tx *gorm.DB, ruleIDs, eventIDs []uint) error {
+// deleteAlertHits 删除本租户相关的全部命中:规则属于本租户,或命中引用了
-	query := tx.Model(&model.AlertRuleHit{})
+// 本租户的日志事件(他租户/全局规则命中本租户事件的行一并清)。
-	switch {
+func deleteAlertHits(tx *gorm.DB, id uint) error {
-	case len(ruleIDs) > 0 && len(eventIDs) > 0:
+	rules := tx.Model(&model.AlertRule{}).Select("id").Where("oci_config_id = ?", id)
-		query = query.Where("rule_id IN ? OR log_event_id IN ?", ruleIDs, eventIDs)
+	err := tx.Where("rule_id IN (?) OR log_event_id IN (?)", rules, tenantEventIDs(tx, id)).
-	case len(ruleIDs) > 0:
+		Delete(&model.AlertRuleHit{}).Error
-		query = query.Where("rule_id IN ?", ruleIDs)
+	if err != nil {
-	case len(eventIDs) > 0:
-		query = query.Where("log_event_id IN ?", eventIDs)
-	default:
-		return nil
-	}
-	if err := query.Delete(&model.AlertRuleHit{}).Error; err != nil {
 		return fmt.Errorf("delete tenant alert hits: %w", err)
 	}
 	return nil

internal/service/tenantdelete_test.go

@@ -108,7 +108,10 @@ func tenantTaskCases() []tenantTaskCase {
 		{name: "测活多租户", task: taskOf(model.TaskTypeHealthCheck, `{"ociConfigIds":[1,2]}`), wantOK: true, wantPayload: `{"ociConfigIds":[2]}`},
 		{name: "成本去重命中", task: taskOf(model.TaskTypeCost, `{"ociConfigIds":[1,1,2]}`), wantOK: true, wantPayload: `{"ociConfigIds":[2]}`},
 		{name: "成本未命中", task: taskOf(model.TaskTypeCost, `{"ociConfigIds":[2]}`)},
-		{name: "非法 JSON", task: taskOf(model.TaskTypeCost, `{`), wantErr: true},
+		// 坏 payload 记警告跳过(保留原任务),不再 fail-closed 阻断整个租户删除
+		{name: "非法 JSON 跳过", task: taskOf(model.TaskTypeCost, `{`)},
+		{name: "抢机坏 payload 跳过", task: taskOf(model.TaskTypeSnatch, `not-json`)},
+		{name: "抢机空 payload 跳过", task: taskOf(model.TaskTypeSnatch, ``)},
 	}
 }
 
@@ -607,3 +610,67 @@ func assertCount(t *testing.T, db *gorm.DB, value any, query string, args []any,
 		t.Errorf("count %T = %d, want %d", value, got, want)
 	}
 }
+
+func TestDeleteTenantSkipsCorruptTaskPayload(t *testing.T) {
+	configs, _, db := newTenantDeleteEnv(t, &fakeClient{})
+	target, _ := seedDeleteTenants(t, db)
+	corrupt := model.Task{Name: "corrupt-snatch", Type: model.TaskTypeSnatch, Payload: `{broken`}
+	mustCreate(t, db, &corrupt)
+	mustCreate(t, db, &model.TaskLog{TaskID: corrupt.ID, Message: "keep"})
+
+	if err := configs.Delete(context.Background(), target.ID); err != nil {
+		t.Fatalf("坏 payload 不应阻断租户删除: %v", err)
+	}
+	var kept model.Task
+	if err := db.First(&kept, corrupt.ID).Error; err != nil || kept.Payload != `{broken` {
+		t.Errorf("坏任务应原样保留: %+v, %v", kept, err)
+	}
+	assertCount(t, db, &model.TaskLog{}, "task_id = ?", []any{corrupt.ID}, 1)
+	assertCount(t, db, &model.OciConfig{}, "id = ?", []any{target.ID}, 0)
+}
+
+func TestDeleteTenantEventsAffectedRules(t *testing.T) {
+	_, _, db := newTenantDeleteEnv(t, &fakeClient{})
+	target, other := seedDeleteTenants(t, db)
+	seedTenantEvents(t, db, target.ID, other.ID)
+
+	// 规则 ID 须在删除前取:target-rule 会随租户一并删除
+	var targetRule, globalRule model.AlertRule
+	db.Where("name = ?", "target-rule").First(&targetRule)
+	db.Where("name = ?", "global-rule").First(&globalRule)
+
+	result := &tenantDeleteResult{}
+	err := db.Transaction(func(tx *gorm.DB) error {
+		return deleteTenantEvents(tx, target.ID, result)
+	})
+	if err != nil {
+		t.Fatalf("deleteTenantEvents: %v", err)
+	}
+	// target-rule 属本租户,global-rule 的命中引用了本租户事件:都应回收冷却
+	got := map[uint]bool{}
+	for _, id := range result.alertRuleIDs {
+		got[id] = true
+	}
+	if len(got) != 2 || !got[targetRule.ID] || !got[globalRule.ID] {
+		t.Errorf("alertRuleIDs = %v, want 含 target-rule(%d) 与 global-rule(%d)", result.alertRuleIDs, targetRule.ID, globalRule.ID)
+	}
+	// 仅保留 global-rule × other-event 一条命中
+	assertCount(t, db, &model.AlertRuleHit{}, "", nil, 1)
+}
+
+func TestDeleteTenantManyEventsNoVarLimit(t *testing.T) {
+	// 回归:事件数超 SQLite 绑定变量上限(32766)时删除仍成功(旧实现 IN 展开必失败)
+	configs, _, db := newTenantDeleteEnv(t, &fakeClient{})
+	target, _ := seedDeleteTenants(t, db)
+	events := make([]model.LogEvent, 0, 33000)
+	for i := 0; i < 33000; i++ {
+		events = append(events, model.LogEvent{OciConfigID: target.ID, MessageID: fmt.Sprintf("bulk-%d", i)})
+	}
+	if err := db.CreateInBatches(&events, 500).Error; err != nil {
+		t.Fatalf("seed events: %v", err)
+	}
+	if err := configs.Delete(context.Background(), target.ID); err != nil {
+		t.Fatalf("数万事件时删除不应受绑定变量上限影响: %v", err)
+	}
+	assertCount(t, db, &model.LogEvent{}, "oci_config_id = ?", []any{target.ID}, 0)
+}