feat(openai): OpenAI Fast/Flex Policy 完整实现（HTTP + WebSocket + Admin）

对称参照 Claude BetaPolicy 的 fast-mode 过滤实现，新增针对 OpenAI 上游 service_tier 字段（priority / flex，含客户端 "fast" → "priority" 归一化）的 pass / filter / block 三态策略，覆盖全部 OpenAI 入口 + admin 配置入口。后端核心 - 新增 SettingKeyOpenAIFastPolicySettings、OpenAIFastPolicyRule、 OpenAIFastPolicySettings 配置模型，含规则的 service_tier × action × scope × 模型白名单 × fallback action 维度。 - SettingService.Get/SetOpenAIFastPolicySettings；缺失时返回内置默认策略（所有模型的 priority 走 filter，whitelist 为空，fallback=pass）。设计依据：service_tier=fast 是用户级开关，与 model 字段正交，默认锁定特定 model slug 会留下"用 gpt-4 + fast 透传 priority 上游"的绕过路径。JSON 解析失败不再静默 fallback，slog.Warn 记录脏数据，便于运维定位。 - service_tier 归一化（trim + ToLower + fast→priority + 白名单 priority/flex）与策略评估（evaluateOpenAIFastPolicy）作为唯一真实来源，HTTP / WS 共用。抽出纯函数 evaluateOpenAIFastPolicyWithSettings，配合 ctx-bound settings 快照（withOpenAIFastPolicyContext / openAIFastPolicySettingsFromContext）， WS 长会话入口预取一次后所有帧复用，避免每帧打到 settingService。 HTTP 入口（4 个） - Chat Completions、Anthropic 兼容（Messages，含 BetaFastMode→priority 二次命中）、原生 Responses、Passthrough Responses 全部接入 applyOpenAIFastPolicyToBody，filter 走 sjson 顶层删除 service_tier，block 返回 403 forbidden_error JSON。 - 4 入口统一使用 upstream 视角的 model（GetMappedModel + normalizeOpenAIModelForUpstream + Codex OAuth normalize 后的 slug），避免 chat/messages/native /responses/passthrough 因为 model 维度不同造成 whitelist 命中差异。 - 在 pass 路径也把客户端 "fast" 别名归一化为 "priority" 写回 body，否则 native /responses 与 passthrough 入口会把 "fast" 原样透传给上游导致 400/拒绝（chat-completions 入口的 normalizeResponsesBodyServiceTier 此前已具备同等行为）。 WebSocket 入口 - 新增 applyOpenAIFastPolicyToWSResponseCreate：严格匹配 type="response.create"，仅处理顶层 service_tier；filter 用 sjson 删字段， block 返回 typed *OpenAIFastBlockedError。 - ingress 路径在 parseClientPayload 内调用，block 命中先 Write Realtime 风格 error event 再返回 OpenAIWSClientCloseError(StatusPolicyViolation =1008)，依赖底层 WebSocket Conn.Write 的同步 flush 保证 error 先于 close。 - passthrough 路径在 RunEntry 前对 firstClientMessage 应用策略，并通过 openAIWSPolicyEnforcingFrameConn 包装 ReadFrame 对每个 client→upstream 帧执行策略；后续帧无 model 字段时回退到 capturedSessionModel。 filter 闭包内同时侦测 session.update / session.created 帧的 session.model 字段刷新 capturedSessionModel，封堵"首帧 model=gpt-4o（pass）→ session.update 改为 gpt-5.5 → 不带 model 的 response.create fallback 到 gpt-4o"的 mid-session 绕过路径。 - passthrough billing：requestServiceTier 在策略 filter 之后再从 firstClientMessage 提取，filter 命中时 OpenAIForwardResult.ServiceTier 上报 nil（default tier），与 HTTP 入口（reqBody 来自 post-filter map） / WS ingress（payload 来自 post-filter bytes）的语义一致。 - 错误事件 schema：{event_id: "evt_<32hex>", type: "error", error: {type: "forbidden_error", code: "policy_violation", message}}，与 OpenAI codex 客户端 error event 解析兼容。 Admin / Frontend - dto.SystemSettings / UpdateSettingsRequest 新增 openai_fast_policy_settings 字段（omitempty），bulk GET/PUT 接入。 - Settings 页 Gateway 页签新增 Fast/Flex Policy 表单卡片： service_tier × action × scope × 模型白名单 × fallback action 全字段配置。 - 前端守门：openaiFastPolicyLoaded 标志仅在 GET 真带回字段时才允许回写，避免 rollout/错误把默认规则覆盖成空；saveSettings 回写循环 skip 该字段，由专用刷新逻辑处理；仅 action=block 时发送 error_message，匹配后端 omitempty 行为。测试 - HTTP 路径：openai_fast_policy_test.go 覆盖默认配置（whitelist=[]，所有模型 priority filter）/ block 自定义错误 / scope 区分 / filter 删字段 / block 不改 body / block 短路上游 / Anthropic BetaFastMode 触发 OpenAI fast policy 等场景。 - WebSocket 路径：openai_fast_policy_ws_test.go 覆盖 helper 单元（filter / fast→priority 归一化 / flex 透传 / block typed error / 无 service_tier 字节不变 / 非 response.create 帧不动 / 空 type 帧不动 / event_id+code 字段断言 / 非字符串 service_tier 容错）+ pass 路径 fast 别名归一化回归 + ingress 端到端（filter 后上游不含 service_tier / block 后客户端先收 error event 再收 close 1008 且上游 0 写）+ passthrough capturedSessionModel fallback 用例（whitelist 策略下首帧建立、缺 model 命中 fallback、缺少 fallback 时的 leak 文档化）+ passthrough session.update / session.created 旋转 capturedSessionModel 的 mid-session 绕过回归 + passthrough billing post-filter ServiceTier 与 idempotent filter 回归。 Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-28 00:34:23 +08:00
parent c92b88e34a
commit 30f55a1f72
23 changed files with 2820 additions and 10 deletions
--- a/backend/internal/service/openai_gateway_service.go
+++ b/backend/internal/service/openai_gateway_service.go
@@ -334,6 +334,7 @@ type OpenAIGatewayService struct {
 	resolver              *ModelPricingResolver
 	channelService        *ChannelService
 	balanceNotifyService  *BalanceNotifyService
+	settingService        *SettingService

 	openaiWSPoolOnce              sync.Once
 	openaiWSStateStoreOnce        sync.Once
@@ -372,6 +373,7 @@ func NewOpenAIGatewayService(
 	resolver *ModelPricingResolver,
 	channelService *ChannelService,
 	balanceNotifyService *BalanceNotifyService,
+	settingService *SettingService,
 ) *OpenAIGatewayService {
 	svc := &OpenAIGatewayService{
 		accountRepo:         accountRepo,
@@ -402,6 +404,7 @@ func NewOpenAIGatewayService(
 		resolver:              resolver,
 		channelService:        channelService,
 		balanceNotifyService:  balanceNotifyService,
+		settingService:        settingService,
 		responseHeaderFilter:  compileResponseHeaderFilter(cfg),
 		codexSnapshotThrottle: newAccountWriteThrottle(openAICodexSnapshotPersistMinInterval),
 	}
@@ -2310,6 +2313,48 @@ func (s *OpenAIGatewayService) Forward(ctx context.Context, c *gin.Context, acco
 		disablePatch()
 	}

+	// Apply OpenAI fast policy (参照 Claude BetaPolicy 的 fast-mode 过滤)：
+	// 针对 body 的 service_tier 字段（"priority" 即 fast，"flex"），按策略
+	// 执行 filter（删除字段）或 block（拒绝请求）。对 gpt-5.5 等模型屏蔽
+	// fast 时在此生效。
+	//
+	// 注意：
+	//   1. 此处统一使用 upstreamModel（已经过 GetMappedModel +
+	//      normalizeOpenAIModelForUpstream + Codex OAuth normalize），与
+	//      chat-completions / messages 入口保持一致，避免不同入口因为模型
+	//      维度不同而出现 whitelist 命中差异。
+	//   2. action=pass 时也要把 raw "fast" 归一化为 "priority" 写回 body，
+	//      否则 native /responses 入口透传 "fast" 给上游会被拒。chat-
+	//      completions 入口由 normalizeResponsesBodyServiceTier 完成同一
+	//      行为，这里手工实现等效逻辑。
+	if rawTier, ok := reqBody["service_tier"].(string); ok {
+		if normTier := normalizedOpenAIServiceTierValue(rawTier); normTier != "" {
+			action, errMsg := s.evaluateOpenAIFastPolicy(ctx, account, upstreamModel, normTier)
+			switch action {
+			case BetaPolicyActionBlock:
+				msg := errMsg
+				if msg == "" {
+					msg = fmt.Sprintf("openai service_tier=%s is not allowed for model %s", normTier, upstreamModel)
+				}
+				blocked := &OpenAIFastBlockedError{Message: msg}
+				writeOpenAIFastPolicyBlockedResponse(c, blocked)
+				return nil, blocked
+			case BetaPolicyActionFilter:
+				delete(reqBody, "service_tier")
+				bodyModified = true
+				disablePatch()
+			default:
+				// pass：若客户端传的是别名 "fast"，归一化为 "priority"
+				// 后写回 body，确保上游收到的是其能识别的规范值。
+				if normTier != rawTier {
+					reqBody["service_tier"] = normTier
+					bodyModified = true
+					markPatchSet("service_tier", normTier)
+				}
+			}
+		}
+	}
+
 	// Re-serialize body only if modified
 	if bodyModified {
 		serializedByPatch := false
@@ -2758,6 +2803,26 @@ func (s *OpenAIGatewayService) forwardOpenAIPassthrough(
 		body = sanitizedBody
 	}

+	// Apply OpenAI fast policy to the passthrough body (filter/block by service_tier).
+	// 统一使用 upstream 视角的 model：透传路径下 body 已经过 compact 映射 +
+	// OAuth normalize，body 中的 model 字段即上游真正会看到的 slug。
+	// 这样可以与 chat-completions / messages / native /responses 入口的
+	// upstreamModel 保持一致，避免 whitelist 命中差异。当 body 中没有
+	// model 字段时退回 reqModel。
+	policyModel := strings.TrimSpace(gjson.GetBytes(body, "model").String())
+	if policyModel == "" {
+		policyModel = reqModel
+	}
+	updatedBody, policyErr := s.applyOpenAIFastPolicyToBody(ctx, account, policyModel, body)
+	if policyErr != nil {
+		var blocked *OpenAIFastBlockedError
+		if errors.As(policyErr, &blocked) {
+			writeOpenAIFastPolicyBlockedResponse(c, blocked)
+		}
+		return nil, policyErr
+	}
+	body = updatedBody
+
 	logger.LegacyPrintf("service.openai_gateway",
 		"[OpenAI 自动透传] 命中自动透传分支: account=%d name=%s type=%s model=%s stream=%v",
 		account.ID,
@@ -5590,14 +5655,319 @@ func normalizeOpenAIServiceTier(raw string) *string {
 	if value == "fast" {
 		value = "priority"
 	}
+	// 放过 OpenAI 官方文档定义的所有合法 tier 值：priority/flex/auto/default/scale。
+	// 对 Codex 客户端零影响（Codex 只发 priority 或 flex，见 codex-rs/core/src/client.rs），
+	// 但能让直连 OpenAI SDK 的用户透传 auto/default/scale 以便抓包/调试。
+	// 真未知值仍返回 nil，由 normalizeResponsesBodyServiceTier 从 body 中删除。
 	switch value {
-	case "priority", "flex":
+	case "priority", "flex", "auto", "default", "scale":
 		return &value
 	default:
 		return nil
 	}
 }

+// OpenAIFastBlockedError indicates a request was rejected by the OpenAI fast
+// policy (action=block). Mirrors BetaBlockedError on the Claude side.
+type OpenAIFastBlockedError struct {
+	Message string
+}
+
+func (e *OpenAIFastBlockedError) Error() string { return e.Message }
+
+// evaluateOpenAIFastPolicy returns the action and error message that should be
+// applied for a request with the given account/model/service_tier. When the
+// policy service is unavailable or no rule matches, it returns
+// (BetaPolicyActionPass, "") so callers can short-circuit safely.
+//
+// Matching rules:
+//   - Scope filters by account type (all / oauth / apikey / bedrock)
+//   - ServiceTier must be empty (= any), "all", or equal the normalized tier
+//   - ModelWhitelist narrows the rule to specific models; FallbackAction
+//     handles the non-matching case (default: pass)
+//
+// 与 Claude BetaPolicy 的差异（保留首条匹配 short-circuit）：
+//   - BetaPolicy 处理的是 anthropic-beta header 中的 token 集合，不同
+//     规则可能针对不同 token，filter 需要累加成 set；block 则 first-match。
+//   - OpenAI fast policy 操作的是单个字段 service_tier：filter 即删字段，
+//     没有可累加的对象。一次请求只携带一个 service_tier，规则的 tier
+//     维度天然互斥；同一 (scope, tier) 下若多条规则的 model whitelist
+//     发生重叠，admin 可通过规则顺序明确意图。因此采用 first-match 而
+//     非 BetaPolicy 那样的"block 覆盖 filter 覆盖 pass"语义。
+func (s *OpenAIGatewayService) evaluateOpenAIFastPolicy(ctx context.Context, account *Account, model, serviceTier string) (action, errMsg string) {
+	if s == nil || s.settingService == nil {
+		return BetaPolicyActionPass, ""
+	}
+	tier := strings.ToLower(strings.TrimSpace(serviceTier))
+	if tier == "" {
+		return BetaPolicyActionPass, ""
+	}
+	settings := openAIFastPolicySettingsFromContext(ctx)
+	if settings == nil {
+		fetched, err := s.settingService.GetOpenAIFastPolicySettings(ctx)
+		if err != nil || fetched == nil {
+			return BetaPolicyActionPass, ""
+		}
+		settings = fetched
+	}
+	return evaluateOpenAIFastPolicyWithSettings(settings, account, model, tier)
+}
+
+// evaluateOpenAIFastPolicyWithSettings is the pure-function core extracted so
+// long-lived sessions (e.g. WS) can prefetch settings once and avoid hitting
+// the settingService on every frame. See WSSession entry and
+// openAIFastPolicySettingsFromContext for the caching glue.
+func evaluateOpenAIFastPolicyWithSettings(settings *OpenAIFastPolicySettings, account *Account, model, tier string) (action, errMsg string) {
+	if settings == nil {
+		return BetaPolicyActionPass, ""
+	}
+	isOAuth := account != nil && account.IsOAuth()
+	isBedrock := account != nil && account.IsBedrock()
+	for _, rule := range settings.Rules {
+		if !betaPolicyScopeMatches(rule.Scope, isOAuth, isBedrock) {
+			continue
+		}
+		ruleTier := strings.ToLower(strings.TrimSpace(rule.ServiceTier))
+		if ruleTier != "" && ruleTier != OpenAIFastTierAny && ruleTier != tier {
+			continue
+		}
+		eff := BetaPolicyRule{
+			Action:               rule.Action,
+			ErrorMessage:         rule.ErrorMessage,
+			ModelWhitelist:       rule.ModelWhitelist,
+			FallbackAction:       rule.FallbackAction,
+			FallbackErrorMessage: rule.FallbackErrorMessage,
+		}
+		return resolveRuleAction(eff, model)
+	}
+	return BetaPolicyActionPass, ""
+}
+
+// openAIFastPolicyCtxKey 是 context 中预取的 OpenAIFastPolicySettings 缓存
+// 键，仅用于 WebSocket 长会话内多帧复用同一份策略快照，避免每帧 DB 命中。
+//
+// Trade-off：策略变更不会影响当前 WS session（只影响新 session）。这是
+// 有意为之 —— 对长会话来说，"策略一致性"比"立刻生效"更重要，且 Claude
+// BetaPolicy 的 gin.Context 缓存也是同样取舍。需要 hot-reload 时管理员
+// 可以通过踢断 session 强制刷新。
+type openAIFastPolicyCtxKeyType struct{}
+
+var openAIFastPolicyCtxKey = openAIFastPolicyCtxKeyType{}
+
+// withOpenAIFastPolicyContext 将一份 settings 快照绑定到 context，供该 ctx
+// 衍生 goroutine 中的 evaluateOpenAIFastPolicy 复用。
+func withOpenAIFastPolicyContext(ctx context.Context, settings *OpenAIFastPolicySettings) context.Context {
+	if ctx == nil || settings == nil {
+		return ctx
+	}
+	return context.WithValue(ctx, openAIFastPolicyCtxKey, settings)
+}
+
+func openAIFastPolicySettingsFromContext(ctx context.Context) *OpenAIFastPolicySettings {
+	if ctx == nil {
+		return nil
+	}
+	if v, ok := ctx.Value(openAIFastPolicyCtxKey).(*OpenAIFastPolicySettings); ok {
+		return v
+	}
+	return nil
+}
+
+// applyOpenAIFastPolicyToBody applies the OpenAI fast policy to a raw request
+// body. When action=filter it removes the service_tier field; when
+// action=block it returns (body, *OpenAIFastBlockedError). On pass it
+// normalizes the service_tier value (e.g. client alias "fast" → "priority"),
+// rewriting the body so the upstream receives a slug it recognizes.
+//
+// Rationale for normalize-on-pass: chat-completions / messages 入口在调用本
+// 函数之前已经通过 normalizeResponsesBodyServiceTier 把 service_tier 归一化
+// 到了上游可识别值；passthrough（OpenAI 自动透传） / native /responses 等
+// 入口没有这一前置步骤，pass 路径下若不在此处归一化，"fast" 就会被原样
+// 透传到 OpenAI 上游导致 400/拒绝。把归一化收敛到本函数，所有入口行为一致。
+func (s *OpenAIGatewayService) applyOpenAIFastPolicyToBody(ctx context.Context, account *Account, model string, body []byte) ([]byte, error) {
+	if len(body) == 0 {
+		return body, nil
+	}
+	rawTier := gjson.GetBytes(body, "service_tier").String()
+	if rawTier == "" {
+		return body, nil
+	}
+	normTier := normalizedOpenAIServiceTierValue(rawTier)
+	if normTier == "" {
+		return body, nil
+	}
+	action, errMsg := s.evaluateOpenAIFastPolicy(ctx, account, model, normTier)
+	switch action {
+	case BetaPolicyActionBlock:
+		msg := errMsg
+		if msg == "" {
+			msg = fmt.Sprintf("openai service_tier=%s is not allowed for model %s", normTier, model)
+		}
+		return body, &OpenAIFastBlockedError{Message: msg}
+	case BetaPolicyActionFilter:
+		trimmed, err := sjson.DeleteBytes(body, "service_tier")
+		if err != nil {
+			return body, fmt.Errorf("strip service_tier from body: %w", err)
+		}
+		return trimmed, nil
+	default:
+		// pass：把别名（如 "fast"）写回为规范值（"priority"）。
+		if normTier == rawTier {
+			return body, nil
+		}
+		updated, err := sjson.SetBytes(body, "service_tier", normTier)
+		if err != nil {
+			return body, fmt.Errorf("normalize service_tier on pass: %w", err)
+		}
+		return updated, nil
+	}
+}
+
+// writeOpenAIFastPolicyBlockedResponse writes a 403 JSON response for a
+// request blocked by the OpenAI fast policy.
+func writeOpenAIFastPolicyBlockedResponse(c *gin.Context, err *OpenAIFastBlockedError) {
+	if c == nil || err == nil {
+		return
+	}
+	c.JSON(http.StatusForbidden, gin.H{
+		"error": gin.H{
+			"type":    "permission_error",
+			"message": err.Message,
+		},
+	})
+}
+
+// applyOpenAIFastPolicyToWSResponseCreate evaluates the OpenAI fast policy
+// against a single client→upstream WebSocket frame whose top-level
+// "type"=="response.create". It mirrors the HTTP-side
+// applyOpenAIFastPolicyToBody contract but operates on a Realtime/Responses
+// WS payload:
+//
+//   - pass: returns frame unchanged (newBytes == frame, blocked == nil)
+//   - filter: returns a copy with top-level service_tier removed
+//   - block: returns (frame, *OpenAIFastBlockedError)
+//
+// Only frames whose "type" field strictly equals "response.create" are
+// inspected/mutated. Any other frame type — including the empty string —
+// passes through untouched. The OpenAI Realtime client-event spec requires
+// "type" to be set, so an empty type is treated as a malformed frame we do
+// not police; the upstream is the source of truth for rejecting it.
+//
+// service_tier lives at the top level of response.create — same as the
+// Responses HTTP body shape (see openai_gateway_chat_completions.go:304 +
+// extractOpenAIServiceTierFromBody at line 5593, and the test fixture at
+// openai_ws_forwarder_ingress_session_test.go:402). We therefore only need
+// to inspect / strip the top-level field; there is no nested form in the
+// schema today.
+//
+// The caller is responsible for choosing the upstream model passed in —
+// this helper does not re-derive it.
+func (s *OpenAIGatewayService) applyOpenAIFastPolicyToWSResponseCreate(
+	ctx context.Context,
+	account *Account,
+	model string,
+	frame []byte,
+) ([]byte, *OpenAIFastBlockedError, error) {
+	if len(frame) == 0 {
+		return frame, nil, nil
+	}
+	if !gjson.ValidBytes(frame) {
+		return frame, nil, nil
+	}
+	frameType := strings.TrimSpace(gjson.GetBytes(frame, "type").String())
+	// Strict match: only response.create is policy-checked. Empty / other
+	// types pass through untouched so we never accidentally strip fields
+	// from response.cancel, conversation.item.create, or any future
+	// client-event the spec adds. The Realtime spec requires "type" on
+	// every client event, so an empty type is malformed input — let the
+	// upstream reject it rather than guessing at our layer.
+	if frameType != "response.create" {
+		return frame, nil, nil
+	}
+	rawTier := gjson.GetBytes(frame, "service_tier").String()
+	if rawTier == "" {
+		return frame, nil, nil
+	}
+	normTier := normalizedOpenAIServiceTierValue(rawTier)
+	if normTier == "" {
+		return frame, nil, nil
+	}
+	action, errMsg := s.evaluateOpenAIFastPolicy(ctx, account, model, normTier)
+	switch action {
+	case BetaPolicyActionBlock:
+		msg := errMsg
+		if msg == "" {
+			msg = fmt.Sprintf("openai service_tier=%s is not allowed for model %s", normTier, model)
+		}
+		return frame, &OpenAIFastBlockedError{Message: msg}, nil
+	case BetaPolicyActionFilter:
+		trimmed, err := sjson.DeleteBytes(frame, "service_tier")
+		if err != nil {
+			return frame, nil, fmt.Errorf("strip service_tier from ws frame: %w", err)
+		}
+		return trimmed, nil, nil
+	default:
+		return frame, nil, nil
+	}
+}
+
+// newOpenAIFastPolicyWSEventID returns a Realtime-style event_id for a
+// server-emitted error event. Matches the loose "evt_<rand>" convention used
+// by upstream Realtime servers; the exact value is not load-bearing and is
+// only required for client-side log correlation. We reuse the existing
+// google/uuid dependency rather than pulling a new one.
+func newOpenAIFastPolicyWSEventID() string {
+	id, err := uuid.NewRandom()
+	if err != nil {
+		// Extremely unlikely; fall back to a fixed prefix so the field is
+		// still non-empty and the schema stays self-consistent.
+		return "evt_openai_fast_policy"
+	}
+	// Strip dashes so it visually matches "evt_<hex>" rather than UUID v4
+	// canonical form, mirroring what real Realtime traces look like.
+	return "evt_" + strings.ReplaceAll(id.String(), "-", "")
+}
+
+// buildOpenAIFastPolicyBlockedWSEvent renders an OpenAI Realtime/Responses
+// style "error" event payload for a request blocked by the OpenAI fast
+// policy. The shape mirrors Realtime error events as observed in upstream
+// traces and per the spec's server "error" event:
+//
+//	{
+//	  "event_id": "evt_<random>",
+//	  "type": "error",
+//	  "error": {
+//	    "type": "invalid_request_error",
+//	    "code": "policy_violation",
+//	    "message": "..."
+//	  }
+//	}
+//
+// event_id lets clients correlate the rejection in their logs; "code" gives
+// programmatic clients a stable identifier (HTTP-side equivalent is the
+// 403 permission_error JSON body).
+func buildOpenAIFastPolicyBlockedWSEvent(err *OpenAIFastBlockedError) []byte {
+	if err == nil {
+		return nil
+	}
+	eventID := newOpenAIFastPolicyWSEventID()
+	payload, mErr := json.Marshal(map[string]any{
+		"event_id": eventID,
+		"type":     "error",
+		"error": map[string]any{
+			"type":    "invalid_request_error",
+			"code":    "policy_violation",
+			"message": err.Message,
+		},
+	})
+	if mErr != nil {
+		// Fallback to a minimal hand-rolled payload; Marshal of the literal
+		// shape above should never fail in practice.
+		return []byte(`{"event_id":"` + eventID + `","type":"error","error":{"type":"invalid_request_error","code":"policy_violation","message":"openai fast policy blocked this request"}}`)
+	}
+	return payload
+}
+
 func sanitizeEmptyBase64InputImagesInOpenAIBody(body []byte) ([]byte, bool, error) {
 	if len(body) == 0 || !bytes.Contains(body, []byte(`"image_url"`)) || !bytes.Contains(body, []byte(`base64,`)) {
 		return body, false, nil