sub2api/backend/internal/service/openai_gateway_request_body.go

package service

import (
	"bytes"
	"context"
	"encoding/json"
	"errors"
	"fmt"
	"github.com/Wei-Shaw/sub2api/internal/pkg/openai"
	"net/http"
	"strings"
	"unicode/utf8"

	"github.com/Wei-Shaw/sub2api/internal/pkg/apicompat"
	"github.com/Wei-Shaw/sub2api/internal/pkg/ctxkey"
	"github.com/Wei-Shaw/sub2api/internal/util/urlvalidator"
	"github.com/gin-gonic/gin"
	"github.com/google/uuid"
	"github.com/tidwall/gjson"
	"github.com/tidwall/sjson"
)

func (s *OpenAIGatewayService) validateUpstreamBaseURL(raw string) (string, error) {
	normalized, err := s.validateOutboundURL(raw)
	if err != nil {
		return "", fmt.Errorf("invalid base_url: %w", err)
	}
	return normalized, nil
}

// validateOutboundURL 按 security.url_allowlist 策略校验网关主动连接的出站 URL。
func (s *OpenAIGatewayService) validateOutboundURL(raw string) (string, error) {
	if s.cfg == nil {
		return urlvalidator.ValidateURLFormat(raw, false)
	}
	if !s.cfg.Security.URLAllowlist.Enabled {
		return urlvalidator.ValidateURLFormat(raw, s.cfg.Security.URLAllowlist.AllowInsecureHTTP)
	}
	return urlvalidator.ValidateHTTPSURL(raw, urlvalidator.ValidationOptions{
		AllowedHosts:     s.cfg.Security.URLAllowlist.UpstreamHosts,
		RequireAllowlist: true,
		AllowPrivate:     s.cfg.Security.URLAllowlist.AllowPrivateHosts,
	})
}

// buildOpenAIResponsesURL 组装 OpenAI Responses 端点。
// - base 以 /v1 结尾：追加 /responses
// - base 以其他版本段结尾（如 /v4）：追加 /responses
// - base 已是 /responses：原样返回
// - 其他情况：追加 /v1/responses
func buildOpenAIResponsesURL(base string) string {
	return buildOpenAIEndpointURL(base, "/v1/responses")
}

// buildOpenAIResponsesURLForPlatform 组装 Responses 端点（平台感知）。
// 供应商 profile 声明了 ResponsesPath 时按其拼接（如 DeepSeek 为无 /v1 前缀的
// /responses）；其余平台维持 /v1/responses。
func buildOpenAIResponsesURLForPlatform(platform string, base string) string {
	if profile := LookupProviderProfile(platform); profile != nil && profile.ResponsesPath != "" {
		return buildOpenAIEndpointURL(base, profile.ResponsesPath)
	}
	return buildOpenAIResponsesURL(base)
}

func shouldPreserveOpenAIResponsesNoneReasoningEffort(account *Account) bool {
	if account == nil {
		return false
	}
	if account.IsOpenAIPassthroughEnabled() {
		return true
	}
	if account.IsOpenAIOAuthLike() {
		return true
	}
	if !account.IsOpenAIApiKey() {
		return false
	}
	baseURL := strings.TrimSpace(account.GetCredential("base_url"))
	return baseURL == "" || isOfficialOpenAIModelsBaseURL(baseURL)
}

// Codex 0.149.0 needs a single advertised effort to directly select a visible
// non-reasoning model. Treat that catalog-only "none" value as omission for
// compatible upstreams, while preserving official OpenAI request semantics.
func filterOpenAIResponsesNoneReasoningEffortForAccount(account *Account, body []byte) ([]byte, error) {
	if len(body) == 0 || shouldPreserveOpenAIResponsesNoneReasoningEffort(account) {
		return body, nil
	}

	out := body
	for _, path := range []string{"reasoning.effort", "reasoning_effort"} {
		effort := gjson.GetBytes(out, path)
		if effort.Type != gjson.String || !strings.EqualFold(strings.TrimSpace(effort.String()), "none") {
			continue
		}
		next, err := sjson.DeleteBytes(out, path)
		if err != nil {
			return body, fmt.Errorf("strip %s none placeholder: %w", path, err)
		}
		out = next
	}
	if reasoning := gjson.GetBytes(out, "reasoning"); reasoning.IsObject() && len(reasoning.Map()) == 0 {
		next, err := sjson.DeleteBytes(out, "reasoning")
		if err != nil {
			return body, fmt.Errorf("strip empty reasoning object: %w", err)
		}
		out = next
	}
	return out, nil
}

func deleteOpenAIResponsesNoneReasoningEffortFromObject(account *Account, body map[string]any) {
	if body == nil || shouldPreserveOpenAIResponsesNoneReasoningEffort(account) {
		return
	}
	if effort, ok := body["reasoning_effort"].(string); ok && strings.EqualFold(strings.TrimSpace(effort), "none") {
		delete(body, "reasoning_effort")
	}
	reasoning, ok := body["reasoning"].(map[string]any)
	if !ok {
		return
	}
	if effort, ok := reasoning["effort"].(string); ok && strings.EqualFold(strings.TrimSpace(effort), "none") {
		delete(reasoning, "effort")
	}
	if len(reasoning) == 0 {
		delete(body, "reasoning")
	}
}

// normalizeDeepSeekResponsesRequestBody 适配无状态 CN Responses 端点：
// 强制 store=false 并清除 previous_response_id（DeepSeek / Kimi 官方
// Responses 均不支持服务端状态存储，携带这些字段会被拒绝）。
//
// DeepSeek 另需把 Codex/OpenAI 的 input_image.image_url 写成线上 serde
// 要求的 url 字段；openai 平台但 base_url 指向 api.deepseek.com 的映射
// 账号同样走这条出站改写（Codex 贴图会 422 missing field url）。
// 非 CN / 非 DeepSeek 上游原样返回。
func normalizeDeepSeekResponsesRequestBody(account *Account, body []byte) []byte {
	if account == nil {
		return body
	}
	applyStateless := account.UsesNativeCNResponses()
	applyImages := shouldAliasDeepSeekResponsesInputImages(account)
	if !applyStateless && !applyImages {
		return body
	}

	normalized := body
	if applyStateless {
		patched, err := sjson.SetBytes(normalized, "store", false)
		if err != nil {
			return body
		}
		normalized = patched
		if stripped, err := sjson.DeleteBytes(normalized, "previous_response_id"); err == nil {
			normalized = stripped
		}
	}

	var requestBody map[string]any
	if err := decodeOpenAIJSONUseNumber(normalized, &requestBody); err != nil {
		return normalized
	}
	input, exists := requestBody["input"]
	if !exists {
		return normalized
	}

	changed := false
	if liftedInput, lifted := apicompat.LiftResponsesToolOutputMedia(input); lifted {
		requestBody["input"] = liftedInput
		input = liftedInput
		changed = true
	}
	if applyImages {
		if aliased, aliasedChanged := aliasDeepSeekResponsesInputImages(input); aliasedChanged {
			requestBody["input"] = aliased
			changed = true
		}
	}
	if !changed {
		return normalized
	}
	rebuilt, err := marshalOpenAIUpstreamJSON(requestBody)
	if err != nil {
		return normalized
	}
	return rebuilt
}

func shouldAliasDeepSeekResponsesInputImages(account *Account) bool {
	return targetsDeepSeekAPIHost(account)
}

// aliasDeepSeekResponsesInputImages 把图片 part 改写成 DeepSeek 能反序列化
// 的形状：type=input_image，同时带字符串字段 image_url 与 url。
// 文档与部分 400 提 image_url；线上 serde 要求 url。只发其中一个都会踩坑。
// 仅有 file_id、没有可用 URL 的 part 原样保留。
func aliasDeepSeekResponsesInputImages(input any) (any, bool) {
	items, ok := asDeepSeekResponsesSlice(input)
	if !ok {
		return input, false
	}
	changed := false
	for i, raw := range items {
		item, ok := raw.(map[string]any)
		if !ok {
			continue
		}
		if aliasDeepSeekResponsesInputItem(item) {
			items[i] = item
			changed = true
		}
	}
	if !changed {
		return input, false
	}
	return items, true
}

func aliasDeepSeekResponsesInputItem(item map[string]any) bool {
	changed := aliasDeepSeekResponsesImagePart(item)
	if content, exists := item["content"]; exists {
		if rewritten, did := aliasDeepSeekResponsesContent(content); did {
			item["content"] = rewritten
			changed = true
		}
	}
	if output, exists := item["output"]; exists {
		if rewritten, did := aliasDeepSeekResponsesContent(output); did {
			item["output"] = rewritten
			changed = true
		}
	}
	return changed
}

func aliasDeepSeekResponsesContent(content any) (any, bool) {
	parts, ok := asDeepSeekResponsesSlice(content)
	if !ok {
		return content, false
	}
	changed := false
	for i, raw := range parts {
		part, ok := raw.(map[string]any)
		if !ok {
			continue
		}
		if aliasDeepSeekResponsesImagePart(part) {
			parts[i] = part
			changed = true
		}
	}
	if !changed {
		return content, false
	}
	return parts, true
}

func aliasDeepSeekResponsesImagePart(part map[string]any) bool {
	partType := strings.TrimSpace(stringValue(part["type"]))
	switch partType {
	case "input_image", "image_url", "image":
	default:
		return false
	}
	imageURL := extractDeepSeekResponsesImageURL(part)
	if imageURL == "" {
		return false
	}
	changed := false
	if partType != "input_image" {
		part["type"] = "input_image"
		changed = true
	}
	if current, ok := part["image_url"].(string); !ok || strings.TrimSpace(current) != imageURL {
		part["image_url"] = imageURL
		changed = true
	}
	if current, ok := part["url"].(string); !ok || strings.TrimSpace(current) != imageURL {
		part["url"] = imageURL
		changed = true
	}
	return changed
}

func extractDeepSeekResponsesImageURL(part map[string]any) string {
	if imageURL := deepSeekResponsesURLValue(part["url"]); imageURL != "" {
		return imageURL
	}
	if imageURL := deepSeekResponsesURLValue(part["image_url"]); imageURL != "" {
		return imageURL
	}
	if imageURL := deepSeekResponsesURLValue(part["image"]); imageURL != "" {
		return imageURL
	}
	source, ok := part["source"].(map[string]any)
	if !ok {
		return ""
	}
	if imageURL := deepSeekResponsesURLValue(source["url"]); imageURL != "" {
		return imageURL
	}
	data := strings.TrimSpace(stringValue(source["data"]))
	if data == "" {
		return ""
	}
	if strings.HasPrefix(strings.ToLower(data), "data:") {
		return data
	}
	mediaType := strings.TrimSpace(stringValue(source["media_type"]))
	if mediaType == "" {
		mediaType = "image/png"
	}
	return "data:" + mediaType + ";base64," + data
}

func deepSeekResponsesURLValue(value any) string {
	switch typed := value.(type) {
	case string:
		return strings.TrimSpace(typed)
	case map[string]any:
		if imageURL := strings.TrimSpace(stringValue(typed["url"])); imageURL != "" {
			return imageURL
		}
		return strings.TrimSpace(stringValue(typed["image_url"]))
	default:
		return ""
	}
}

func asDeepSeekResponsesSlice(value any) ([]any, bool) {
	switch items := value.(type) {
	case []any:
		return items, true
	case []map[string]any:
		out := make([]any, len(items))
		for i := range items {
			out[i] = items[i]
		}
		return out, true
	default:
		return nil, false
	}
}

func trimOpenAIEncryptedReasoningItems(reqBody map[string]any) bool {
	if len(reqBody) == 0 {
		return false
	}

	inputValue, has := reqBody["input"]
	if !has {
		return false
	}

	switch input := inputValue.(type) {
	case []any:
		filtered := input[:0]
		changed := false
		for _, item := range input {
			nextItem, itemChanged, keep := sanitizeEncryptedReasoningInputItem(item)
			if itemChanged {
				changed = true
			}
			if !keep {
				continue
			}
			filtered = append(filtered, nextItem)
		}
		if !changed {
			return false
		}
		if len(filtered) == 0 {
			delete(reqBody, "input")
			return true
		}
		reqBody["input"] = filtered
		return true
	case []map[string]any:
		filtered := input[:0]
		changed := false
		for _, item := range input {
			nextItem, itemChanged, keep := sanitizeEncryptedReasoningInputItem(item)
			if itemChanged {
				changed = true
			}
			if !keep {
				continue
			}
			nextMap, ok := nextItem.(map[string]any)
			if !ok {
				filtered = append(filtered, item)
				continue
			}
			filtered = append(filtered, nextMap)
		}
		if !changed {
			return false
		}
		if len(filtered) == 0 {
			delete(reqBody, "input")
			return true
		}
		reqBody["input"] = filtered
		return true
	case map[string]any:
		nextItem, changed, keep := sanitizeEncryptedReasoningInputItem(input)
		if !changed {
			return false
		}
		if !keep {
			delete(reqBody, "input")
			return true
		}
		nextMap, ok := nextItem.(map[string]any)
		if !ok {
			return false
		}
		reqBody["input"] = nextMap
		return true
	default:
		return false
	}
}

func sanitizeEncryptedReasoningInputItem(item any) (next any, changed bool, keep bool) {
	inputItem, ok := item.(map[string]any)
	if !ok {
		return item, false, true
	}

	itemType, _ := inputItem["type"].(string)
	switch strings.TrimSpace(itemType) {
	case "compaction", "compaction_summary":
		if _, encrypted := inputItem["encrypted_content"]; encrypted {
			return nil, true, false
		}
		return item, false, true
	case "reasoning":
	default:
		return item, false, true
	}

	if _, has := inputItem["encrypted_content"]; has {
		delete(inputItem, "encrypted_content")
		changed = true
	}

	// xAI 422: "content": null 导致 untagged enum 反序列化失败
	if v, has := inputItem["content"]; has && v == nil {
		delete(inputItem, "content")
		changed = true
	}

	if !changed {
		return item, false, true
	}
	if len(inputItem) == 1 {
		return nil, true, false
	}
	return inputItem, true, true
}

// SanitizeOpenAICrossModeFailoverReasoning derives a failover attempt body from
// the canonical request body by dropping provider-specific encrypted reasoning
// input items in full (encrypted_content plus the coupled id/summary shape).
//
// This is the proactive counterpart to the reactive same-account
// invalid_encrypted_content recovery in Forward: when a failover switches from an
// OpenAI passthrough account (which forwards upstream-native encrypted reasoning,
// e.g. Kiro) to a non-passthrough account (e.g. Bedrock Mantle) that rejects the
// provider-specific reasoning IDs/shape, the whole reasoning item must go before
// the request reaches the new upstream. Unlike trimOpenAIEncryptedReasoningItems,
// which only strips the encrypted_content / null-content fields while preserving
// the reasoning item's id and summary, this drops the entire item.
//
// The input slice is treated as immutable and is never mutated; a distinct slice
// is returned only when changed is true.
func SanitizeOpenAICrossModeFailoverReasoning(body []byte) (sanitized []byte, changed bool, err error) {
	if len(body) == 0 {
		return body, false, nil
	}
	if !gjson.GetBytes(body, "input").Exists() {
		return body, false, nil
	}
	var decoded map[string]any
	if err := decodeOpenAIJSONUseNumber(body, &decoded); err != nil {
		return body, false, fmt.Errorf("decode cross-mode failover body: %w", err)
	}
	if !dropOpenAIEncryptedReasoningInputItems(decoded) {
		return body, false, nil
	}
	out, marshalErr := marshalOpenAIUpstreamJSON(decoded)
	if marshalErr != nil {
		return body, false, fmt.Errorf("serialize cross-mode failover body: %w", marshalErr)
	}
	return out, true, nil
}

// dropOpenAIEncryptedReasoningInputItems removes reasoning input items that carry
// provider-specific encrypted_content in full — including their coupled id and
// summary — and reports whether anything changed. Contrast with
// trimOpenAIEncryptedReasoningItems, which only strips fields while keeping the
// reasoning item skeleton.
func dropOpenAIEncryptedReasoningInputItems(reqBody map[string]any) bool {
	if len(reqBody) == 0 {
		return false
	}
	inputValue, has := reqBody["input"]
	if !has {
		return false
	}
	switch input := inputValue.(type) {
	case []any:
		filtered := input[:0]
		changed := false
		for _, item := range input {
			if isOpenAIEncryptedReasoningInputItem(item) {
				changed = true
				continue
			}
			filtered = append(filtered, item)
		}
		if !changed {
			return false
		}
		if len(filtered) == 0 {
			delete(reqBody, "input")
			return true
		}
		reqBody["input"] = filtered
		return true
	case []map[string]any:
		filtered := input[:0]
		changed := false
		for _, item := range input {
			if isOpenAIEncryptedReasoningInputItem(item) {
				changed = true
				continue
			}
			filtered = append(filtered, item)
		}
		if !changed {
			return false
		}
		if len(filtered) == 0 {
			delete(reqBody, "input")
			return true
		}
		reqBody["input"] = filtered
		return true
	case map[string]any:
		if isOpenAIEncryptedReasoningInputItem(input) {
			delete(reqBody, "input")
			return true
		}
		return false
	default:
		return false
	}
}

func isOpenAIEncryptedReasoningInputItem(item any) bool {
	inputItem, ok := item.(map[string]any)
	if !ok {
		return false
	}
	if itemType, _ := inputItem["type"].(string); strings.TrimSpace(itemType) != "reasoning" {
		return false
	}
	_, has := inputItem["encrypted_content"]
	return has
}

// IsOpenAIResponsesCompactPath reports whether the request targets the legacy
// /responses/compact endpoint, including its forwardable subpaths.
func IsOpenAIResponsesCompactPath(c *gin.Context) bool {
	return isOpenAIResponsesCompactPath(c)
}

func OpenAICompactSessionSeedKeyForTest() string {
	return openAICompactSessionSeedKey
}

func NormalizeOpenAICompactRequestBodyForTest(body []byte) ([]byte, bool, error) {
	return normalizeOpenAICompactRequestBody(body)
}

func isOpenAIResponsesCompactPath(c *gin.Context) bool {
	suffix := strings.TrimSpace(openAIResponsesRequestPathSuffix(c))
	return suffix == "/compact" || strings.HasPrefix(suffix, "/compact/")
}

func normalizeOpenAICompactRequestBody(body []byte) ([]byte, bool, error) {
	if len(body) == 0 {
		return body, false, nil
	}
	normalized := []byte(`{}`)
	// Keep the current Codex /compact schema while still dropping request-scoped
	// fields such as prompt_cache_key, store, and stream.
	for _, field := range []string{
		"model",
		"input",
		"instructions",
		"tools",
		"parallel_tool_calls",
		"reasoning",
		"service_tier",
		"text",
		"previous_response_id",
	} {
		value := gjson.GetBytes(body, field)
		if !value.Exists() {
			continue
		}
		next, err := sjson.SetRawBytes(normalized, field, []byte(value.Raw))
		if err != nil {
			return body, false, fmt.Errorf("normalize compact body %s: %w", field, err)
		}
		normalized = next
	}
	if next, removed, err := normalizeOpenAIParallelToolCallsWithoutTools(normalized, false); err != nil {
		return body, false, err
	} else if removed {
		normalized = next
	}

	if bytes.Equal(bytes.TrimSpace(body), bytes.TrimSpace(normalized)) {
		return body, false, nil
	}
	return normalized, true, nil
}

func normalizeOpenAIParallelToolCallsWithoutTools(body []byte, responsesLite bool) ([]byte, bool, error) {
	if responsesLite {
		return body, false, nil
	}
	parallel := gjson.GetBytes(body, "parallel_tool_calls")
	if !parallel.Exists() {
		return body, false, nil
	}
	if openAIRequestBodyHasTools(body) {
		return body, false, nil
	}
	normalized, err := sjson.DeleteBytes(body, "parallel_tool_calls")
	if err != nil {
		return body, false, fmt.Errorf("normalize parallel_tool_calls without tools: %w", err)
	}
	return normalized, true, nil
}

// openAIRequestBodyHasTools 同时识别顶层 tools 和 input[].additional_tools。
func openAIRequestBodyHasTools(body []byte) bool {
	if tools := gjson.GetBytes(body, "tools"); tools.IsArray() && len(tools.Array()) > 0 {
		return true
	}
	for _, item := range gjson.GetBytes(body, "input").Array() {
		if strings.TrimSpace(item.Get("type").String()) != "additional_tools" {
			continue
		}
		if tools := item.Get("tools"); tools.IsArray() && len(tools.Array()) > 0 {
			return true
		}
	}
	return false
}

// normalizeOpenAIResponsesReasoningContentReplay removes non-portable
// reasoning.content arrays before history is sent to a real OpenAI Responses
// endpoint. Compatible providers may return visible reasoning blocks there,
// while OpenAI accepts only an empty array when the item is replayed.
//
// Keep the reasoning item and its portable fields (summary, encrypted_content,
// ids, and opaque extensions). Callers scope this normalization to OpenAI
// destinations; compatible providers may still consume their own content.
func normalizeOpenAIResponsesReasoningContentReplay(body []byte) ([]byte, bool, error) {
	input := parseRawJSONView(body).Get("input")
	if !input.IsArray() {
		return body, false, nil
	}

	needsNormalization := false
	input.ForEach(func(_, item gjson.Result) bool {
		if strings.TrimSpace(item.Get("type").String()) != "reasoning" {
			return true
		}
		content := item.Get("content")
		if content.IsArray() && len(content.Array()) > 0 {
			needsNormalization = true
			return false
		}
		return true
	})
	if !needsNormalization {
		return body, false, nil
	}

	var reqBody map[string]any
	if err := decodeOpenAIJSONUseNumber(body, &reqBody); err != nil {
		return body, false, fmt.Errorf("normalize OpenAI reasoning content replay: %w", err)
	}
	items, ok := reqBody["input"].([]any)
	if !ok {
		return body, false, nil
	}
	changed := false
	for _, rawItem := range items {
		item, ok := rawItem.(map[string]any)
		if !ok || strings.TrimSpace(firstNonEmptyString(item["type"])) != "reasoning" {
			continue
		}
		content, ok := item["content"].([]any)
		if !ok || len(content) == 0 {
			continue
		}
		delete(item, "content")
		changed = true
	}
	if !changed {
		return body, false, nil
	}
	normalized, err := marshalOpenAIUpstreamJSON(reqBody)
	if err != nil {
		return body, false, fmt.Errorf("serialize normalized OpenAI reasoning content replay: %w", err)
	}
	return normalized, true, nil
}

func normalizeOpenAIAPIKeyStoreFalseReasoningReplay(body []byte, knownStoreFalse bool) ([]byte, bool, error) {
	if !knownStoreFalse && gjson.GetBytes(body, "store").Type != gjson.False {
		return body, false, nil
	}
	root := parseRawJSONView(body)
	input := root.Get("input")
	if !input.IsArray() {
		return body, false, nil
	}
	if !root.IsObject() || !gjson.ValidBytes(body) || !utf8.Valid(body) || hasDuplicateJSONObjectKeys(root) {
		return normalizeOpenAIAPIKeyStoreFalseReasoningReplayDecoded(body, knownStoreFalse)
	}

	// Only reasoning metadata needs decoding. Keep large image/tool results as
	// slices of the original JSON and copy them once into the final request.
	items := make([]string, 0)
	changed := false
	fallback := false
	var itemErr error
	input.ForEach(func(_, item gjson.Result) bool {
		if !item.IsObject() {
			items = append(items, item.Raw)
			return true
		}
		if hasDuplicateJSONObjectKeys(item) {
			fallback = true
			return false
		}
		typ := strings.TrimSpace(item.Get("type").String())
		id := strings.TrimSpace(item.Get("id").String())
		encrypted := item.Get("encrypted_content")
		if (typ == "reasoning" && (encrypted.Type != gjson.String || strings.TrimSpace(encrypted.Str) == "")) ||
			(typ == "item_reference" && strings.HasPrefix(id, "rs_")) {
			changed = true
			return true
		}
		stripID := typ == "reasoning" && strings.HasPrefix(id, "rs_")
		addSummary := typ == "reasoning" && item.Get("summary").Type == gjson.Null
		stripCallID := shouldStripOpenAIResponsesNonPairCallID(typ) && item.Get("call_id").Exists()
		if !stripID && !addSummary && !stripCallID {
			items = append(items, item.Raw)
			return true
		}
		var decoded map[string]any
		if err := decodeOpenAIJSONUseNumber([]byte(item.Raw), &decoded); err != nil {
			itemErr = err
			return false
		}
		if stripID {
			delete(decoded, "id")
		}
		if addSummary {
			decoded["summary"] = []any{}
		}
		if stripCallID {
			delete(decoded, "call_id")
		}
		encoded, err := marshalOpenAIUpstreamJSON(decoded)
		if err != nil {
			itemErr = err
			return false
		}
		items = append(items, string(encoded))
		changed = true
		return true
	})
	if fallback {
		return normalizeOpenAIAPIKeyStoreFalseReasoningReplayDecoded(body, knownStoreFalse)
	}
	if itemErr != nil {
		return body, false, fmt.Errorf("normalize API-key store=false reasoning replay: %w", itemErr)
	}
	if !changed {
		return body, false, nil
	}
	return replaceOpenAIRawInput(body, input, items), true, nil
}

// Preserve the decoder's handling of unusual or duplicate-key input objects.
func normalizeOpenAIAPIKeyStoreFalseReasoningReplayDecoded(body []byte, knownStoreFalse bool) ([]byte, bool, error) {
	if !knownStoreFalse && gjson.GetBytes(body, "store").Type != gjson.False {
		return body, false, nil
	}
	input := gjson.GetBytes(body, "input")
	if !input.IsArray() {
		return body, false, nil
	}

	var reqBody map[string]any
	if err := decodeOpenAIJSONUseNumber(body, &reqBody); err != nil {
		return body, false, fmt.Errorf("normalize API-key store=false reasoning replay: %w", err)
	}
	items, ok := reqBody["input"].([]any)
	if !ok {
		return body, false, nil
	}
	filtered := make([]any, 0, len(items))
	changed := false
	for _, rawItem := range items {
		item, ok := rawItem.(map[string]any)
		if !ok {
			filtered = append(filtered, rawItem)
			continue
		}
		typ := strings.TrimSpace(firstNonEmptyString(item["type"]))
		id := strings.TrimSpace(firstNonEmptyString(item["id"]))
		switch typ {
		case "reasoning":
			encryptedContent, hasEncryptedContent := item["encrypted_content"].(string)
			if !hasEncryptedContent || strings.TrimSpace(encryptedContent) == "" {
				changed = true
				continue
			}
			if strings.HasPrefix(id, "rs_") {
				delete(item, "id")
				changed = true
			}
			if summary, ok := item["summary"]; !ok || summary == nil {
				item["summary"] = []any{}
				changed = true
			}
		case "item_reference":
			if strings.HasPrefix(id, "rs_") {
				changed = true
				continue
			}
		}
		if shouldStripOpenAIResponsesNonPairCallID(typ) {
			if _, hasCallID := item["call_id"]; hasCallID {
				delete(item, "call_id")
				changed = true
			}
		}
		filtered = append(filtered, item)
	}
	if !changed {
		return body, false, nil
	}
	reqBody["input"] = filtered
	normalized, err := marshalOpenAIUpstreamJSON(reqBody)
	if err != nil {
		return body, false, fmt.Errorf("serialize API-key store=false reasoning replay: %w", err)
	}
	return normalized, true, nil
}

func normalizeOpenAICodexCompactReasoningEffortForAccount(c *gin.Context, account *Account, body []byte) ([]byte, bool, error) {
	if account == nil || !account.IsOpenAIOAuthLike() || !isOpenAIResponsesCompactPath(c) {
		return body, false, nil
	}

	requestedModel := strings.TrimSpace(gjson.GetBytes(body, "model").String())
	effectiveModel := account.GetMappedModel(requestedModel)
	return normalizeOpenAICodexCompactReasoningEffort(body, effectiveModel)
}

func normalizeOpenAICodexCompactReasoningEffort(body []byte, effectiveModel string) ([]byte, bool, error) {
	if !isOpenAIGPT56Model(effectiveModel) ||
		!strings.EqualFold(strings.TrimSpace(gjson.GetBytes(body, "reasoning.effort").String()), "max") {
		return body, false, nil
	}

	// Codex Ultra 在客户端编排层会下发 max；ChatGPT compact 端点目前只接受到
	// xhigh。这里只降级 OpenAI OAuth 的 GPT-5.6 compact 子请求，普通 Responses、
	// API Key 请求和其他平台的 OAuth 请求保留 max。
	normalized, err := sjson.SetBytes(body, "reasoning.effort", "xhigh")
	if err != nil {
		return body, false, fmt.Errorf("normalize codex compact reasoning effort: %w", err)
	}
	return normalized, true, nil
}

func resolveOpenAICompactSessionID(c *gin.Context) string {
	if c != nil {
		if sessionID := strings.TrimSpace(c.GetHeader("session_id")); sessionID != "" {
			return sessionID
		}
		if conversationID := strings.TrimSpace(c.GetHeader("conversation_id")); conversationID != "" {
			return conversationID
		}
		if seed, ok := c.Get(openAICompactSessionSeedKey); ok {
			if seedStr, ok := seed.(string); ok && strings.TrimSpace(seedStr) != "" {
				return strings.TrimSpace(seedStr)
			}
		}
	}
	return uuid.NewString()
}

// openAIResponsesRequestPathSuffix 返回可拼接到上游 /responses URL 后面的子路径。
// 不可转发的子路径返回空串（退化为裸 /responses）；真正的拒绝由入口守卫
// IsForwardableOpenAIResponsesRequestPath 负责。这样即便将来新增路由漏挂守卫，
// 拼进上游 URL 的也只会是合规片段。
func openAIResponsesRequestPathSuffix(c *gin.Context) string {
	suffix, ok := sanitizedUpstreamPathSuffix(rawOpenAIResponsesRequestPathSuffix(c))
	if !ok {
		return ""
	}
	return suffix
}

// IsForwardableOpenAIResponsesRequestPath 判断入站请求携带的 /responses 子路径
// 是否可以安全转发。路由层用它在鉴权后、调度前直接拒绝畸形子路径。
func IsForwardableOpenAIResponsesRequestPath(c *gin.Context) bool {
	_, ok := sanitizedUpstreamPathSuffix(rawOpenAIResponsesRequestPathSuffix(c))
	return ok
}

// IsOpenAIResponsesInputTokensRequestPath reports whether the request targets
// the native Responses input-token counting endpoint.
func IsOpenAIResponsesInputTokensRequestPath(c *gin.Context) bool {
	return openAIResponsesRequestPathSuffix(c) == "/input_tokens"
}

// rawOpenAIResponsesRequestPathSuffix 仅做提取，不做任何安全判断。
func rawOpenAIResponsesRequestPathSuffix(c *gin.Context) string {
	if c == nil || c.Request == nil || c.Request.URL == nil {
		return ""
	}
	normalizedPath := strings.TrimRight(strings.TrimSpace(c.Request.URL.Path), "/")
	if normalizedPath == "" {
		return ""
	}
	idx := strings.LastIndex(normalizedPath, "/responses")
	if idx < 0 {
		return ""
	}
	suffix := normalizedPath[idx+len("/responses"):]
	if suffix == "" || suffix == "/" {
		return ""
	}
	if !strings.HasPrefix(suffix, "/") {
		return ""
	}
	return suffix
}

func appendOpenAIResponsesRequestPathSuffix(baseURL, suffix string) string {
	trimmedBase := strings.TrimRight(strings.TrimSpace(baseURL), "/")
	// 兜底：调用方漏了校验时，这里也不会把不合规的片段拼进上游 URL。
	trimmedSuffix, ok := sanitizedUpstreamPathSuffix(suffix)
	if !ok || trimmedBase == "" || trimmedSuffix == "" {
		return trimmedBase
	}
	return trimmedBase + trimmedSuffix
}

func (s *OpenAIGatewayService) replaceModelInResponseBody(body []byte, fromModel, toModel string) []byte {
	// A mapped request must retain its public name even when upstream uses an alias.
	if fromModel == "" || toModel == "" || fromModel == toModel || !gjson.ValidBytes(body) {
		return body
	}
	if m := gjson.GetBytes(body, "model"); m.Type == gjson.String {
		newBody, err := sjson.SetBytes(body, "model", toModel)
		if err != nil {
			return body
		}
		return newBody
	}
	return body
}

func getOpenAIReasoningEffortFromReqBody(reqBody map[string]any, requestedModel string) (value string, present bool) {
	if reqBody == nil {
		return "", false
	}

	// Primary: reasoning.effort
	if reasoning, ok := reqBody["reasoning"].(map[string]any); ok {
		if effort, ok := reasoning["effort"].(string); ok {
			return normalizeOpenAIReasoningEffortForModel(effort, requestedModel), true
		}
	}

	// Fallback: some clients may use a flat field.
	if effort, ok := reqBody["reasoning_effort"].(string); ok {
		return normalizeOpenAIReasoningEffortForModel(effort, requestedModel), true
	}

	return "", false
}

func deriveOpenAIReasoningEffortFromModel(model string) string {
	if strings.TrimSpace(model) == "" {
		return ""
	}

	modelID := strings.TrimSpace(model)
	if strings.Contains(modelID, "/") {
		parts := strings.Split(modelID, "/")
		modelID = parts[len(parts)-1]
	}

	parts := strings.FieldsFunc(strings.ToLower(modelID), func(r rune) bool {
		switch r {
		case '-', '_', ' ':
			return true
		default:
			return false
		}
	})
	if len(parts) == 0 {
		return ""
	}

	return normalizeOpenAIReasoningEffortForModel(parts[len(parts)-1], modelID)
}

// deriveOpenAIReasoningEffortFromModelCandidates 依次对每个候选模型做后缀推导，
// 返回第一个非空结果。
func deriveOpenAIReasoningEffortFromModelCandidates(models []string) string {
	for _, model := range models {
		if value := deriveOpenAIReasoningEffortFromModel(model); value != "" {
			return value
		}
	}
	return ""
}

type openAIRequestView struct {
	body               []byte
	Model              string
	Stream             bool
	PromptCacheKey     string
	PreviousResponseID string
	ServiceTier        string
	ReasoningEffort    string
	patches            []openAIRequestPatch
	patchesDisabled    bool
}

type openAIRequestPatch struct {
	path   string
	delete bool
	value  any
}

func newOpenAIRequestView(body []byte) openAIRequestView {
	if len(body) == 0 {
		return openAIRequestView{}
	}

	const (
		modelField uint8 = 1 << iota
		streamField
		promptCacheKeyField
		previousResponseIDField
		serviceTierField
		reasoningField
		allRequestViewFields = modelField | streamField | promptCacheKeyField |
			previousResponseIDField | serviceTierField | reasoningField
	)

	view := openAIRequestView{body: body}
	var seen uint8
	// parseRawJSONView reads body without copying; view keeps body alive for extracted strings.
	parseRawJSONView(body).ForEach(func(key, value gjson.Result) bool {
		switch key.Str {
		case "model":
			if seen&modelField == 0 {
				view.Model = strings.TrimSpace(value.String())
				seen |= modelField
			}
		case "stream":
			if seen&streamField == 0 {
				view.Stream = value.Bool()
				seen |= streamField
			}
		case "prompt_cache_key":
			if seen&promptCacheKeyField == 0 {
				view.PromptCacheKey = strings.TrimSpace(value.String())
				seen |= promptCacheKeyField
			}
		case "previous_response_id":
			if seen&previousResponseIDField == 0 {
				view.PreviousResponseID = strings.TrimSpace(value.String())
				seen |= previousResponseIDField
			}
		case "service_tier":
			if seen&serviceTierField == 0 {
				view.ServiceTier = strings.TrimSpace(value.String())
				seen |= serviceTierField
			}
		case "reasoning":
			if seen&reasoningField == 0 {
				view.ReasoningEffort = strings.TrimSpace(value.Get("effort").String())
				seen |= reasoningField
			}
		}
		return seen != allRequestViewFields
	})
	return view
}

// Decode 保留阶段一既有 full-map 行为；后续阶段会把调用点下沉到复杂分支。
func (v openAIRequestView) Decode(c *gin.Context) (map[string]any, error) {
	return getOpenAIRequestBodyMap(c, v.body)
}

func (v *openAIRequestView) MarkPatchSet(path string, value any) {
	if v == nil || v.patchesDisabled {
		return
	}
	path = strings.TrimSpace(path)
	if !isSimpleOpenAIRequestPatchPath(path) {
		v.DisablePatches()
		return
	}
	v.patches = append(v.patches, openAIRequestPatch{path: path, value: value})
}

func (v *openAIRequestView) MarkPatchDelete(path string) {
	if v == nil || v.patchesDisabled {
		return
	}
	path = strings.TrimSpace(path)
	if !isSimpleOpenAIRequestPatchPath(path) {
		v.DisablePatches()
		return
	}
	v.patches = append(v.patches, openAIRequestPatch{path: path, delete: true})
}

func isSimpleOpenAIRequestPatchPath(path string) bool {
	if path == "" || strings.ContainsRune(path, '\\') {
		return false
	}
	for _, part := range strings.Split(path, ".") {
		if strings.TrimSpace(part) == "" {
			return false
		}
	}
	return true
}

func (v *openAIRequestView) DisablePatches() {
	if v == nil {
		return
	}
	v.patchesDisabled = true
	v.patches = nil
}

func (v openAIRequestView) HasPatches() bool {
	return !v.patchesDisabled && len(v.patches) > 0
}

func (v openAIRequestView) ApplyPatches() ([]byte, error) {
	if v.patchesDisabled || len(v.patches) == 0 {
		return nil, errors.New("openai request patches disabled")
	}
	body := v.body
	for _, patch := range v.patches {
		var err error
		if patch.delete {
			body, err = sjson.DeleteBytes(body, patch.path)
		} else {
			body, err = sjson.SetBytes(body, patch.path, patch.value)
		}
		if err != nil {
			return nil, err
		}
	}
	return body, nil
}

func setOpenAIRequestMapPath(reqBody map[string]any, path string, value any) {
	path = strings.TrimSpace(path)
	if reqBody == nil || path == "" {
		return
	}
	parts := strings.Split(path, ".")
	current := reqBody
	for _, part := range parts[:len(parts)-1] {
		part = strings.TrimSpace(part)
		if part == "" {
			return
		}
		next, _ := current[part].(map[string]any)
		if next == nil {
			next = map[string]any{}
			current[part] = next
		}
		current = next
	}
	last := strings.TrimSpace(parts[len(parts)-1])
	if last != "" {
		current[last] = value
	}
}

func deleteOpenAIRequestMapPath(reqBody map[string]any, path string) {
	path = strings.TrimSpace(path)
	if reqBody == nil || path == "" {
		return
	}
	parts := strings.Split(path, ".")
	current := reqBody
	for _, part := range parts[:len(parts)-1] {
		part = strings.TrimSpace(part)
		if part == "" {
			return
		}
		next, _ := current[part].(map[string]any)
		if next == nil {
			return
		}
		current = next
	}
	last := strings.TrimSpace(parts[len(parts)-1])
	if last != "" {
		delete(current, last)
	}
}

func extractOpenAIRequestMetaFromBody(body []byte) (model string, stream bool, promptCacheKey string) {
	view := newOpenAIRequestView(body)
	return view.Model, view.Stream, view.PromptCacheKey
}

func normalizeOpenAIOAuthResponsesCompatibilityFields(reqBody map[string]any) bool {
	if reqBody == nil {
		return false
	}
	changed := false
	if prompt, exists := reqBody["prompt"]; exists {
		if input, hasInput := reqBody["input"]; !hasInput || input == nil {
			if prompt != nil {
				reqBody["input"] = prompt
			}
		}
		delete(reqBody, "prompt")
		changed = true
	}
	if _, exists := reqBody["commands"]; exists {
		delete(reqBody, "commands")
		changed = true
	}
	// Codex can attach internal message metadata when a custom provider is
	// named OpenAI. ChatGPT rejects this field on input items (#7066).
	input, _ := reqBody["input"].([]any)
	for _, value := range input {
		item, ok := value.(map[string]any)
		if !ok {
			continue
		}
		if _, exists := item["internal_chat_message_metadata_passthrough"]; exists {
			delete(item, "internal_chat_message_metadata_passthrough")
			changed = true
		}
	}
	return changed
}

func normalizeOpenAIOAuthResponsesCompatibilityBody(body []byte) ([]byte, bool, error) {
	if len(body) == 0 {
		return body, false, nil
	}
	normalized := body
	changed := false
	prompt := gjson.GetBytes(normalized, "prompt")
	if prompt.Exists() {
		input := gjson.GetBytes(normalized, "input")
		if prompt.Type != gjson.Null && (!input.Exists() || input.Type == gjson.Null) {
			next, err := sjson.SetRawBytes(normalized, "input", []byte(prompt.Raw))
			if err != nil {
				return body, false, fmt.Errorf("normalize oauth responses prompt: %w", err)
			}
			normalized = next
		}
		next, err := sjson.DeleteBytes(normalized, "prompt")
		if err != nil {
			return body, false, fmt.Errorf("normalize oauth responses delete prompt: %w", err)
		}
		normalized = next
		changed = true
	}
	if gjson.GetBytes(normalized, "commands").Exists() {
		next, err := sjson.DeleteBytes(normalized, "commands")
		if err != nil {
			return body, false, fmt.Errorf("normalize oauth responses delete commands: %w", err)
		}
		normalized = next
		changed = true
	}
	// Only remove the input-item field, never same-named user content.
	input := gjson.GetBytes(normalized, "input")
	if !input.IsArray() {
		return normalized, changed, nil
	}
	for i, item := range input.Array() {
		if !item.IsObject() || !item.Get("internal_chat_message_metadata_passthrough").Exists() {
			continue
		}
		next, err := sjson.DeleteBytes(normalized, fmt.Sprintf("input.%d.internal_chat_message_metadata_passthrough", i))
		if err != nil {
			return body, false, fmt.Errorf("normalize oauth input metadata: %w", err)
		}
		normalized = next
		changed = true
	}
	return normalized, changed, nil
}

func normalizeGPT6ResponsesSampling(body []byte, model string) ([]byte, bool, error) {
	if (!openai.IsGPT6SolOrLunaModelSpelling(model) && !openai.IsGPT61SolModelSpelling(model)) || gjson.GetBytes(body, "reasoning.effort").String() == "none" {
		return body, false, nil
	}
	out := body
	changed := false
	for _, key := range []string{"temperature", "top_p", "top_logprobs", "logprobs"} {
		if !gjson.GetBytes(out, key).Exists() {
			continue
		}
		var err error
		out, err = sjson.DeleteBytes(out, key)
		if err != nil {
			return body, false, fmt.Errorf("remove GPT-6 sampling parameter %s: %w", key, err)
		}
		changed = true
	}
	if include := gjson.GetBytes(out, "include"); include.IsArray() {
		items := include.Array()
		for i := len(items) - 1; i >= 0; i-- {
			if items[i].String() != "message.output_text.logprobs" {
				continue
			}
			var err error
			out, err = sjson.DeleteBytes(out, fmt.Sprintf("include.%d", i))
			if err != nil {
				return body, false, fmt.Errorf("remove GPT-6 logprobs include: %w", err)
			}
			changed = true
		}
	}
	return out, changed, nil
}

func normalizeOpenAIResponsesReasoningMode(body []byte, model string) ([]byte, bool, error) {
	if len(body) == 0 {
		return body, false, nil
	}
	// GPT-6 treats reasoning.mode and reasoning.effort as independent
	// official fields. Preserve both verbatim; earlier models retain the
	// established mode stripping and pro-to-max compatibility behavior.
	if model == "" {
		model = gjson.GetBytes(body, "model").String()
	}
	if isOpenAIGPT6Model(model) {
		return normalizeGPT6ResponsesSampling(body, model)
	}
	mode := gjson.GetBytes(body, "reasoning.mode")
	if !mode.Exists() || mode.Type != gjson.String {
		return body, false, nil
	}
	updated := body
	effort := gjson.GetBytes(body, "reasoning.effort")
	if (!effort.Exists() || effort.Type == gjson.Null || strings.TrimSpace(effort.String()) == "") &&
		strings.EqualFold(strings.TrimSpace(mode.String()), "pro") {
		var err error
		updated, err = sjson.SetBytes(updated, "reasoning.effort", "max")
		if err != nil {
			return body, false, fmt.Errorf("set reasoning effort for mode=pro: %w", err)
		}
	}
	updated, err := sjson.DeleteBytes(updated, "reasoning.mode")
	if err != nil {
		return body, false, fmt.Errorf("delete unsupported reasoning.mode: %w", err)
	}
	if reasoning := gjson.GetBytes(updated, "reasoning"); reasoning.Exists() && reasoning.IsObject() && len(reasoning.Map()) == 0 {
		updated, err = sjson.DeleteBytes(updated, "reasoning")
		if err != nil {
			return body, false, fmt.Errorf("delete empty reasoning object: %w", err)
		}
	}
	return updated, true, nil
}

func normalizeOpenAIResponseFormatSchemasBody(body []byte) ([]byte, bool, error) {
	if len(body) == 0 {
		return body, false, nil
	}
	textFormat := strings.TrimSpace(gjson.GetBytes(body, "text.format.type").String())
	responseFormat := strings.TrimSpace(gjson.GetBytes(body, "response_format.type").String())
	if textFormat != "json_schema" && responseFormat != "json_schema" {
		return body, false, nil
	}
	var reqBody map[string]any
	if err := decodeOpenAIJSONUseNumber(body, &reqBody); err != nil {
		return body, false, fmt.Errorf("normalize responses schema body: %w", err)
	}
	if !normalizeOpenAIResponseFormatSchemas(reqBody) {
		return body, false, nil
	}
	normalized, err := json.Marshal(reqBody)
	if err != nil {
		return body, false, fmt.Errorf("serialize normalized responses schema body: %w", err)
	}
	return normalized, true, nil
}

func normalizeOpenAIResponsesWebSocketCompatibilityBody(body []byte, account *Account, responsesLite bool) ([]byte, bool, error) {
	return normalizeOpenAIResponsesCompatibilityBodyWithOptions(body, account, openAIResponsesCompatibilityOptions{ResponsesLite: responsesLite})
}

type openAIResponsesCompatibilityOptions struct {
	ResponsesLite bool
	// Compact marks the /responses/compact wire shape, which is left as-is
	// by request-shape compatibility rewrites such as web_search history.
	Compact bool
}

func normalizeOpenAIResponsesCompatibilityBodyWithOptions(body []byte, account *Account, opts openAIResponsesCompatibilityOptions) ([]byte, bool, error) {
	responsesLite := opts.ResponsesLite
	if account == nil || !account.IsOpenAI() {
		return body, false, nil
	}
	normalized := body
	changed := false
	if account.IsOpenAIOAuthLike() {
		var err error
		normalized, changed, err = normalizeOpenAIResponsesLegacyIngress(body)
		if err != nil {
			return body, false, err
		}
	}
	if next, normalizedReasoningContent, err := normalizeOpenAIResponsesReasoningContentReplay(normalized); err != nil {
		return body, false, err
	} else if normalizedReasoningContent {
		normalized = next
		changed = true
	}
	if account.IsOpenAIApiKey() {
		if next, normalizedParallel, err := normalizeOpenAIParallelToolCallsWithoutTools(normalized, responsesLite); err != nil {
			return body, false, err
		} else if normalizedParallel {
			normalized = next
			changed = true
		}
		if next, normalizedReasoning, err := normalizeOpenAIAPIKeyStoreFalseReasoningReplay(normalized, false); err != nil {
			return body, false, err
		} else if normalizedReasoning {
			normalized = next
			changed = true
		}
	}
	if sanitized, idsChanged, err := sanitizeOpenAIResponsesInputItemIDs(normalized); err != nil {
		return body, false, fmt.Errorf("sanitize websocket Responses input item IDs: %w", err)
	} else if idsChanged {
		normalized = sanitized
		changed = true
	}
	if account != nil && account.IsOpenAI() && account.IsOAuth() {
		if reasoningBody, reasoningChanged, err := normalizeOpenAIResponsesReasoningMode(normalized, account.GetMappedModel(gjson.GetBytes(normalized, "model").String())); err != nil {
			return body, false, err
		} else if reasoningChanged {
			normalized = reasoningBody
			changed = true
		}
	}
	if account != nil && account.IsOpenAIOAuthLike() {
		oauthBody, oauthChanged, err := normalizeOpenAIOAuthResponsesCompatibilityBody(normalized)
		if err != nil {
			return body, false, err
		}
		normalized = oauthBody
		changed = changed || oauthChanged
		for _, field := range openAIChatGPTInternalUnsupportedFields {
			if !gjson.GetBytes(normalized, field).Exists() {
				continue
			}
			next, deleteErr := sjson.DeleteBytes(normalized, field)
			if deleteErr != nil {
				return body, false, fmt.Errorf("normalize websocket body delete %s: %w", field, deleteErr)
			}
			normalized = next
			changed = true
		}
		if !opts.Compact {
			webSearchBody, webSearchChanged, err := ensureOpenAIOAuthWebSearchToolForHistoryBody(normalized, responsesLite)
			if err != nil {
				return body, false, fmt.Errorf("normalize websocket body: %w", err)
			}
			normalized = webSearchBody
			changed = changed || webSearchChanged
		}
	}
	needsOrphanCleanup := account != nil && account.IsOpenAIOAuthLike() &&
		gjson.GetBytes(normalized, "input").IsArray()
	if needsOrphanCleanup || openAIResponsesInputMayNeedTruncation(normalized) {
		var reqBody map[string]any
		if err := decodeOpenAIJSONUseNumber(normalized, &reqBody); err != nil {
			return body, false, fmt.Errorf("normalize websocket Responses body: %w", err)
		}
		mapChanged := false
		if needsOrphanCleanup {
			if input, ok := reqBody["input"].([]any); ok && sanitizeOpenAIResponsesOrphanToolOutputs(
				reqBody,
				input,
				strings.TrimSpace(firstNonEmptyString(reqBody["previous_response_id"])) != "",
			) {
				mapChanged = true
			}
		}
		if truncateOpenAIResponsesInputText(reqBody) {
			mapChanged = true
		}
		if mapChanged {
			next, err := marshalOpenAIUpstreamJSON(reqBody)
			if err != nil {
				return body, false, fmt.Errorf("serialize normalized websocket Responses body: %w", err)
			}
			normalized = next
			changed = true
		}
	}
	if schemaBody, schemaChanged, err := normalizeOpenAIResponseFormatSchemasBody(normalized); err != nil {
		return body, false, err
	} else if schemaChanged {
		normalized = schemaBody
		changed = true
	}
	if openAIRequestBodyImageGenerationToolNeedsNormalization(normalized) {
		var reqBody map[string]any
		if err := json.Unmarshal(normalized, &reqBody); err != nil {
			return body, false, fmt.Errorf("normalize websocket image tool body: %w", err)
		}
		if normalizeOpenAIResponsesImageGenerationTools(reqBody) {
			next, err := json.Marshal(reqBody)
			if err != nil {
				return body, false, fmt.Errorf("serialize normalized websocket image tool body: %w", err)
			}
			normalized = next
			changed = true
		}
	}
	if account != nil {
		if schemaBody, schemaChanged, err := sanitizeOpenAIResponsesToolSchemasForPlatform(normalized, account.Platform); err != nil {
			return body, false, fmt.Errorf("normalize websocket tool schemas: %w", err)
		} else if schemaChanged {
			normalized = schemaBody
			changed = true
		}
	}
	// Keep this last: earlier compatibility passes may filter or rebuild input.
	// Remote compaction v2 requires one trigger as the final input item.
	if triggerBody, triggerChanged, err := NormalizeCompactionTriggerInputOrder(normalized); err != nil {
		return body, false, fmt.Errorf("normalize websocket compaction trigger order: %w", err)
	} else if triggerChanged {
		normalized = triggerBody
		changed = true
	}
	return normalized, changed, nil
}

// normalizeOpenAIPassthroughOAuthBody 将透传 OAuth 请求体收敛为旧链路关键行为：
// 1) 删除 ChatGPT internal API 不支持的顶层 Responses 参数
// 2) store=false 3) 非 compact 保持 stream=true；compact 强制 stream=false
func normalizeOpenAIPassthroughOAuthBody(body []byte, compact bool) ([]byte, bool, error) {
	if len(body) == 0 {
		return body, false, nil
	}

	normalized, changed, err := normalizeOpenAIOAuthResponsesCompatibilityBody(body)
	if err != nil {
		return body, false, err
	}
	if reasoningBody, reasoningChanged, reasoningErr := normalizeOpenAIResponsesReasoningMode(normalized, ""); reasoningErr != nil {
		return body, false, reasoningErr
	} else if reasoningChanged {
		normalized = reasoningBody
		changed = true
	}

	for _, field := range openAIChatGPTInternalUnsupportedFields {
		if value := gjson.GetBytes(normalized, field); !value.Exists() {
			continue
		}
		next, err := sjson.DeleteBytes(normalized, field)
		if err != nil {
			return body, false, fmt.Errorf("normalize passthrough body delete %s: %w", field, err)
		}
		normalized = next
		changed = true
	}
	if schemaBody, schemaChanged, schemaErr := normalizeOpenAIResponseFormatSchemasBody(normalized); schemaErr != nil {
		return body, false, schemaErr
	} else if schemaChanged {
		normalized = schemaBody
		changed = true
	}

	if inputResult := gjson.GetBytes(normalized, "input"); inputResult.Exists() {
		switch {
		case inputResult.Type == gjson.String:
			text := inputResult.String()
			var inputValue any
			if strings.TrimSpace(text) != "" {
				inputValue = []any{map[string]any{
					"type": "message", "role": "user", "content": text,
				}}
			} else {
				inputValue = []any{}
			}
			next, err := sjson.SetBytes(normalized, "input", inputValue)
			if err != nil {
				return body, false, fmt.Errorf("normalize passthrough body input string: %w", err)
			}
			normalized = next
			changed = true
		case inputResult.Type == gjson.JSON && !inputResult.IsArray():
			next, err := sjson.SetRawBytes(normalized, "input", []byte("["+inputResult.Raw+"]"))
			if err != nil {
				return body, false, fmt.Errorf("normalize passthrough body input object: %w", err)
			}
			normalized = next
			changed = true
		}
	}

	if compact {
		if store := gjson.GetBytes(normalized, "store"); store.Exists() {
			next, err := sjson.DeleteBytes(normalized, "store")
			if err != nil {
				return body, false, fmt.Errorf("normalize passthrough body delete store: %w", err)
			}
			normalized = next
			changed = true
		}
		if stream := gjson.GetBytes(normalized, "stream"); stream.Exists() {
			next, err := sjson.DeleteBytes(normalized, "stream")
			if err != nil {
				return body, false, fmt.Errorf("normalize passthrough body delete stream: %w", err)
			}
			normalized = next
			changed = true
		}
	} else {
		if store := gjson.GetBytes(normalized, "store"); !store.Exists() || store.Type != gjson.False {
			next, err := sjson.SetBytes(normalized, "store", false)
			if err != nil {
				return body, false, fmt.Errorf("normalize passthrough body store=false: %w", err)
			}
			normalized = next
			changed = true
		}
		if stream := gjson.GetBytes(normalized, "stream"); !stream.Exists() || stream.Type != gjson.True {
			next, err := sjson.SetBytes(normalized, "stream", true)
			if err != nil {
				return body, false, fmt.Errorf("normalize passthrough body stream=true: %w", err)
			}
			normalized = next
			changed = true
		}
	}

	return normalized, changed, nil
}

func detectOpenAIPassthroughInstructionsRejectReason(reqModel string, body []byte) string {
	if !isOpenAICodexModel(reqModel) {
		return ""
	}

	instructions := gjson.GetBytes(body, "instructions")
	if !instructions.Exists() {
		return ""
	}
	if instructions.Type != gjson.String {
		return "instructions_not_string"
	}
	if strings.TrimSpace(instructions.String()) == "" {
		return "instructions_empty"
	}
	return ""
}

func isOpenAICodexModel(model string) bool {
	return strings.Contains(strings.ToLower(strings.TrimSpace(model)), "codex")
}

// extractOpenAIReasoningEffortFromBody 按优先级传入模型候选（如 upstreamModel,
// billingModel, originalModel）：显式 effort 的模型归一化（max 保留判定）用第一个
// 非空候选；body 未携带 effort 时的模型后缀推导依次尝试每个候选——OAuth 的
// normalizeCodexModel 会剥掉 upstreamModel 的 effort 后缀，只有原始模型名还留着。
func extractOpenAIReasoningEffortFromBody(body []byte, modelCandidates ...string) *string {
	reasoningEffort := strings.TrimSpace(gjson.GetBytes(body, "reasoning.effort").String())
	if reasoningEffort == "" {
		reasoningEffort = strings.TrimSpace(gjson.GetBytes(body, "reasoning_effort").String())
	}
	if reasoningEffort != "" {
		normalized := normalizeOpenAIReasoningEffortForModel(reasoningEffort, firstNonEmpty(modelCandidates...))
		if normalized == "" {
			return nil
		}
		return &normalized
	}

	value := deriveOpenAIReasoningEffortFromModelCandidates(modelCandidates)
	if value == "" {
		return nil
	}
	return &value
}

func explicitRequestedReasoningEffortFromBody(body []byte) string {
	raw := strings.TrimSpace(gjson.GetBytes(body, "reasoning.effort").String())
	if raw == "" {
		raw = strings.TrimSpace(gjson.GetBytes(body, "reasoning_effort").String())
	}
	if raw == "" {
		raw = strings.TrimSpace(gjson.GetBytes(body, "output_config.effort").String())
	}
	return raw
}

// CanonicalRequestedReasoningEffort extracts the client-requested effort before
// group policy rewriting and before model-family remapping (max -> xhigh).
// Empty or unknown values return nil. "max" is preserved even for models that
// later persist "xhigh".
func CanonicalRequestedReasoningEffort(body []byte, modelCandidates ...string) *string {
	if raw := explicitRequestedReasoningEffortFromBody(body); raw != "" {
		canonical := NormalizeMaxReasoningEffort(raw)
		if canonical == "" {
			return nil
		}
		return &canonical
	}
	for _, model := range modelCandidates {
		if value := canonicalReasoningEffortFromModelSuffix(model); value != "" {
			return &value
		}
	}
	if model := strings.TrimSpace(gjson.GetBytes(body, "model").String()); model != "" {
		if value := canonicalReasoningEffortFromModelSuffix(model); value != "" {
			return &value
		}
	}
	return nil
}

func canonicalReasoningEffortFromModelSuffix(model string) string {
	if strings.TrimSpace(model) == "" {
		return ""
	}
	modelID := strings.TrimSpace(model)
	if strings.Contains(modelID, "/") {
		parts := strings.Split(modelID, "/")
		modelID = parts[len(parts)-1]
	}
	parts := strings.FieldsFunc(strings.ToLower(modelID), func(r rune) bool {
		switch r {
		case '-', '_', ' ':
			return true
		default:
			return false
		}
	})
	if len(parts) == 0 {
		return ""
	}
	return NormalizeMaxReasoningEffort(parts[len(parts)-1])
}

func extractOpenAIServiceTier(reqBody map[string]any) *string {
	if reqBody == nil {
		return nil
	}
	raw, ok := reqBody["service_tier"].(string)
	if !ok {
		return nil
	}
	return normalizeOpenAIServiceTier(raw)
}

func extractOpenAIServiceTierFromBody(body []byte) *string {
	if len(body) == 0 {
		return nil
	}
	return normalizeOpenAIServiceTier(gjson.GetBytes(body, "service_tier").String())
}

func normalizeOpenAIServiceTier(raw string) *string {
	value := strings.ToLower(strings.TrimSpace(raw))
	if value == "" {
		return nil
	}
	if value == "fast" {
		value = "priority"
	}
	// 放过 OpenAI 官方文档定义的合法 tier 值，以及 Codex/API 新增的 ultrafast。
	// Codex 客户端会发 priority、flex 或 ultrafast；直连 OpenAI SDK 的用户还会
	// 透传 auto/default/scale。真未知值仍返回 nil，由
	// normalizeResponsesBodyServiceTier 从 body 中删除。
	switch value {
	case "priority", "flex", "auto", "default", "scale", OpenAIFastTierUltrafast:
		return &value
	default:
		return nil
	}
}

// ErrInvalidOpenAIServiceTier indicates a request carried a service_tier value
// that is not a known OpenAI tier. HTTP handlers translate it into a 400
// invalid_request_error so malformed values are rejected up front instead of
// being silently stripped (which would mask the user's intent to use fast
// mode).
type ErrInvalidOpenAIServiceTier struct {
	Value string
}

func (e *ErrInvalidOpenAIServiceTier) Error() string {
	return fmt.Sprintf("invalid service_tier %q: must be one of auto, default, fast, flex, priority, scale, ultrafast", e.Value)
}

const invalidOpenAIServiceTierValueMaxLen = 64

func boundInvalidOpenAIServiceTierValue(raw string) string {
	if len(raw) <= invalidOpenAIServiceTierValueMaxLen {
		return raw
	}
	return raw[:invalidOpenAIServiceTierValueMaxLen] + "..."
}

// ValidateOpenAIServiceTierField validates the service_tier field of a raw
// OpenAI-compatible request body (/v1/responses and /v1/chat/completions).
//
//   - absent / null → valid, returns "" (field omitted keeps current behavior)
//   - "fast" → normalized to "priority" (the two are equivalent; the canonical
//     value is what reaches the OpenAI upstream)
//   - "priority" / "flex" / "auto" / "default" / "scale" / "ultrafast" → valid, returned as-is
//   - an explicitly present non-string value, an empty string, or any other
//     unknown value → *ErrInvalidOpenAIServiceTier (handler maps to HTTP 400),
//     matching OpenAI's enum validation semantics
func ValidateOpenAIServiceTierField(body []byte) (string, error) {
	tierResult := gjson.GetBytes(body, "service_tier")
	if !tierResult.Exists() || tierResult.Type == gjson.Null {
		return "", nil
	}
	if tierResult.Type != gjson.String {
		return "", &ErrInvalidOpenAIServiceTier{Value: "<non-string>"}
	}
	raw := strings.TrimSpace(tierResult.String())
	if raw == "" {
		return "", &ErrInvalidOpenAIServiceTier{Value: raw}
	}
	norm := normalizedOpenAIServiceTierValue(raw)
	if norm == "" {
		return "", &ErrInvalidOpenAIServiceTier{Value: boundInvalidOpenAIServiceTierValue(raw)}
	}
	return norm, 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)
//   - UserIDs, when present, filters by the trusted Sub2API user that owns the API key
//   - ServiceTier must be empty (= any), "all", "missing", or equal the normalized tier
//   - ModelWhitelist narrows the rule to specific models; FallbackAction
//     handles the non-matching case (default: pass)
//   - User-specific rules take precedence over global rules; each group keeps
//     the configured first-match order
//
// 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, openAIFastPolicyUserID(ctx), account, model, tier)
}

// shouldForceOpenAIFastPriorityForMissingTier reports whether a request that
// omitted service_tier should be upgraded to priority. This is opt-in through
// the dedicated "missing" tier matcher; legacy "all" rules continue to apply
// only to requests that explicitly selected a recognized tier.
func (s *OpenAIGatewayService) shouldForceOpenAIFastPriorityForMissingTier(ctx context.Context, account *Account, model string) bool {
	if account == nil || account.Platform != PlatformOpenAI {
		return false
	}
	action, _ := s.evaluateOpenAIFastPolicy(ctx, account, model, OpenAIFastTierMissing)
	return action == OpenAIFastPolicyActionForcePriority
}

// 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, userID int64, account *Account, model, tier string) (action, errMsg string) {
	if settings == nil {
		return BetaPolicyActionPass, ""
	}
	isOAuth := account != nil && account.IsOAuth()
	isBedrock := account != nil && account.IsBedrock()

	// 用户专属规则先于全局规则。规则组内仍按配置顺序首条命中，允许
	// 管理员为某位用户配置例外，而不被先出现的全局规则覆盖。
	for _, userScoped := range []bool{true, false} {
		for _, rule := range settings.Rules {
			if (len(rule.UserIDs) > 0) != userScoped || !openAIFastPolicyUserMatches(rule.UserIDs, userID) {
				continue
			}
			if !betaPolicyScopeMatches(rule.Scope, isOAuth, isBedrock) {
				continue
			}
			ruleTier := strings.ToLower(strings.TrimSpace(rule.ServiceTier))
			if tier == OpenAIFastTierMissing {
				if ruleTier != OpenAIFastTierMissing {
					continue
				}
			} else 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, ""
}

func openAIFastPolicyUserID(ctx context.Context) int64 {
	if ctx == nil {
		return 0
	}
	userID, _ := ctx.Value(ctxkey.UserID).(int64)
	if userID <= 0 {
		return 0
	}
	return userID
}

func openAIFastPolicyUserMatches(ruleUserIDs []int64, userID int64) bool {
	if len(ruleUserIDs) == 0 {
		return true
	}
	for _, ruleUserID := range ruleUserIDs {
		if ruleUserID == userID {
			return true
		}
	}
	return false
}

// 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
}

func openAIGroupForcesFast(ctx context.Context, account *Account) bool {
	if ctx == nil || account == nil || account.Platform != PlatformOpenAI {
		return false
	}
	group, _ := ctx.Value(ctxkey.Group).(*Group)
	return IsGroupContextValid(group) && groupSupportsOpenAIFast(group.Platform) && group.ForceOpenAIFast
}

// 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").
// action=force_priority rewrites any matched known tier to "priority". Before
// the global policy is evaluated, a trusted request Group with
// ForceOpenAIFast enabled unconditionally sets service_tier to "priority".
// The global policy remains authoritative and may still pass, filter, or block
// that final value.
//
// 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
	}
	if openAIGroupForcesFast(ctx, account) {
		updated, err := sjson.SetBytes(body, "service_tier", OpenAIFastTierPriority)
		if err != nil {
			return body, fmt.Errorf("force group service_tier priority on body: %w", err)
		}
		body = updated
	}
	rawTier := gjson.GetBytes(body, "service_tier").String()
	if rawTier == "" {
		if s.shouldForceOpenAIFastPriorityForMissingTier(ctx, account, model) {
			updated, err := sjson.SetBytes(body, "service_tier", OpenAIFastTierPriority)
			if err != nil {
				return body, fmt.Errorf("force missing service_tier priority on body: %w", err)
			}
			return updated, nil
		}
		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
	case OpenAIFastPolicyActionForcePriority:
		updated, err := sjson.SetBytes(body, "service_tier", OpenAIFastTierPriority)
		if err != nil {
			return body, fmt.Errorf("force service_tier priority on body: %w", err)
		}
		return updated, 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
	}
	MarkOpsClientBusinessLimited(c, OpsClientBusinessLimitedReasonLocalPolicyDenied)
	// body-signal compact 心跳可能已把响应头提交为 200（长排队后才进入
	// Forward），此时以 response.failed 终止事件回传；未提交时先停拍再写
	// JSON，保持原状态码语义（#3887）。
	if StopOpenAICompactSSEKeepaliveCommitted(c) {
		writeOpenAICompactSSEFailureMessage(c, http.StatusForbidden, "permission_error", err.Message)
		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: keeps service_tier, normalizing aliases such as "fast" to "priority"
//   - filter: returns a copy with top-level service_tier removed
//   - force_priority: keeps service_tier and rewrites it to "priority"
//   - block: returns (frame, *OpenAIFastBlockedError)
//   - Group ForceOpenAIFast: sets priority first, then applies the global rule
//
// 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
	}
	if openAIGroupForcesFast(ctx, account) {
		updated, err := sjson.SetBytes(frame, "service_tier", OpenAIFastTierPriority)
		if err != nil {
			return frame, nil, fmt.Errorf("force group service_tier priority in ws frame: %w", err)
		}
		frame = updated
	}
	rawTier := gjson.GetBytes(frame, "service_tier").String()
	if rawTier == "" {
		if s.shouldForceOpenAIFastPriorityForMissingTier(ctx, account, model) {
			updated, err := sjson.SetBytes(frame, "service_tier", OpenAIFastTierPriority)
			if err != nil {
				return frame, nil, fmt.Errorf("force missing service_tier priority in ws frame: %w", err)
			}
			return updated, nil, nil
		}
		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
	case OpenAIFastPolicyActionForcePriority:
		updated, err := sjson.SetBytes(frame, "service_tier", OpenAIFastTierPriority)
		if err != nil {
			return frame, nil, fmt.Errorf("force service_tier priority in ws frame: %w", err)
		}
		return updated, nil, nil
	default:
		if normTier == rawTier {
			return frame, nil, nil
		}
		updated, err := sjson.SetBytes(frame, "service_tier", normTier)
		if err != nil {
			return frame, nil, fmt.Errorf("normalize service_tier in ws frame: %w", err)
		}
		return updated, 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 openAIRequestBodyMayContainImageInput(body []byte) bool {
	if len(body) == 0 {
		return false
	}
	input := gjson.GetBytes(body, "input")
	messages := gjson.GetBytes(body, "messages.#-1")
	return openAIJSONValueMayContainImageInput(input) || openAIJSONValueMayContainImageInput(messages)
}

func openAIJSONValueMayContainImageInput(value gjson.Result) bool {
	if !value.Exists() {
		return false
	}
	if value.IsArray() {
		found := false
		value.ForEach(func(_, item gjson.Result) bool {
			if openAIJSONValueMayContainImageInput(item) {
				found = true
				return false
			}
			return true
		})
		return found
	}
	if value.IsObject() {
		if strings.TrimSpace(value.Get("type").String()) == "input_image" || value.Get("image_url").Exists() {
			return true
		}
		return openAIJSONValueMayContainImageInput(value.Get("content"))
	}
	return false
}

func openAIRequestBodyMayContainEmptyBase64InputImage(body []byte) bool {
	if len(body) == 0 || !openAIRequestBodyMayContainInputImageToken(body) {
		return false
	}
	input := gjson.GetBytes(body, "input")
	if !input.Exists() {
		return false
	}
	return openAIJSONValueMayContainEmptyBase64InputImage(input)
}

func openAIRequestBodyMayContainInputImageToken(body []byte) bool {
	if bytes.Contains(body, []byte("input_image")) {
		return true
	}
	// JSON 字符串任意字符都可能被 unicode escape，遇到 \u 时交给 gjson 解码后的结构扫描兜底。
	return bytes.Contains(body, []byte("\\u"))
}

func openAIJSONValueMayContainEmptyBase64InputImage(value gjson.Result) bool {
	if !value.Exists() {
		return false
	}
	if value.IsArray() {
		found := false
		value.ForEach(func(_, item gjson.Result) bool {
			if openAIJSONValueMayContainEmptyBase64InputImage(item) {
				found = true
				return false
			}
			return true
		})
		return found
	}
	if value.IsObject() {
		if strings.TrimSpace(value.Get("type").String()) == "input_image" && isEmptyBase64DataURI(value.Get("image_url").String()) {
			return true
		}
		return openAIJSONValueMayContainEmptyBase64InputImage(value.Get("content"))
	}
	return false
}

func sanitizeEmptyBase64InputImagesInOpenAIBody(body []byte) ([]byte, bool, error) {
	if !openAIRequestBodyMayContainEmptyBase64InputImage(body) {
		return body, false, nil
	}

	var reqBody map[string]any
	if err := json.Unmarshal(body, &reqBody); err != nil {
		return body, false, fmt.Errorf("sanitize request body: %w", err)
	}
	if !sanitizeEmptyBase64InputImagesInOpenAIRequestBodyMap(reqBody) {
		return body, false, nil
	}
	normalized, err := marshalOpenAIUpstreamJSON(reqBody)
	if err != nil {
		return body, false, fmt.Errorf("serialize sanitized request body: %w", err)
	}
	return normalized, true, nil
}

func sanitizeEmptyBase64InputImagesInOpenAIRequestBodyMap(reqBody map[string]any) bool {
	if reqBody == nil {
		return false
	}
	input, ok := reqBody["input"]
	if !ok {
		return false
	}
	normalizedInput, changed := sanitizeEmptyBase64InputImagesInOpenAIInput(input)
	if !changed {
		return false
	}
	reqBody["input"] = normalizedInput
	return true
}

func sanitizeEmptyBase64InputImagesInOpenAIInput(input any) (any, bool) {
	items, ok := input.([]any)
	if !ok {
		return input, false
	}

	normalizedItems := make([]any, 0, len(items))
	changed := false
	for _, item := range items {
		itemMap, ok := item.(map[string]any)
		if !ok {
			normalizedItems = append(normalizedItems, item)
			continue
		}
		if shouldDropEmptyBase64InputImagePart(itemMap) {
			changed = true
			continue
		}
		content, ok := itemMap["content"]
		if !ok {
			normalizedItems = append(normalizedItems, itemMap)
			continue
		}
		parts, ok := content.([]any)
		if !ok {
			normalizedItems = append(normalizedItems, itemMap)
			continue
		}

		normalizedParts := make([]any, 0, len(parts))
		itemChanged := false
		for _, part := range parts {
			if shouldDropEmptyBase64InputImagePart(part) {
				changed = true
				itemChanged = true
				continue
			}
			normalizedParts = append(normalizedParts, part)
		}
		if itemChanged {
			if len(normalizedParts) == 0 {
				continue
			}
			itemMap["content"] = normalizedParts
		}
		normalizedItems = append(normalizedItems, itemMap)
	}
	if !changed {
		return input, false
	}
	return normalizedItems, true
}

func shouldDropEmptyBase64InputImagePart(part any) bool {
	partMap, ok := part.(map[string]any)
	if !ok {
		return false
	}
	typeValue, _ := partMap["type"].(string)
	if strings.TrimSpace(typeValue) != "input_image" {
		return false
	}
	imageURL, _ := partMap["image_url"].(string)
	return isEmptyBase64DataURI(imageURL)
}

func isEmptyBase64DataURI(raw string) bool {
	if !strings.HasPrefix(raw, "data:") {
		return false
	}
	rest := strings.TrimPrefix(raw, "data:")
	semicolonIdx := strings.Index(rest, ";")
	if semicolonIdx < 0 {
		return false
	}
	rest = rest[semicolonIdx+1:]
	if !strings.HasPrefix(rest, "base64,") {
		return false
	}
	return strings.TrimSpace(strings.TrimPrefix(rest, "base64,")) == ""
}

func getOpenAIRequestBodyMap(_ *gin.Context, body []byte) (map[string]any, error) {
	var reqBody map[string]any
	if err := decodeOpenAIJSONUseNumber(body, &reqBody); err != nil {
		return nil, fmt.Errorf("parse request: %w", err)
	}
	return reqBody, nil
}

// extractOpenAIReasoningEffort 的模型候选语义同 extractOpenAIReasoningEffortFromBody。
func extractOpenAIReasoningEffort(reqBody map[string]any, modelCandidates ...string) *string {
	if value, present := getOpenAIReasoningEffortFromReqBody(reqBody, firstNonEmpty(modelCandidates...)); present {
		if value == "" {
			return nil
		}
		return &value
	}

	value := deriveOpenAIReasoningEffortFromModelCandidates(modelCandidates)
	if value == "" {
		return nil
	}
	return &value
}

func CanonicalRequestedReasoningEffortFromReqBody(reqBody map[string]any, modelCandidates ...string) *string {
	if reqBody == nil {
		return CanonicalRequestedReasoningEffort(nil, modelCandidates...)
	}
	raw := ""
	if reasoning, ok := reqBody["reasoning"].(map[string]any); ok {
		if effort, ok := reasoning["effort"].(string); ok {
			raw = strings.TrimSpace(effort)
		}
	}
	if raw == "" {
		if effort, ok := reqBody["reasoning_effort"].(string); ok {
			raw = strings.TrimSpace(effort)
		}
	}
	if raw != "" {
		canonical := NormalizeMaxReasoningEffort(raw)
		if canonical == "" {
			return nil
		}
		return &canonical
	}
	return CanonicalRequestedReasoningEffort(nil, modelCandidates...)
}

func normalizeOpenAIReasoningEffort(raw string) string {
	value := strings.ToLower(strings.TrimSpace(raw))
	if value == "" {
		return ""
	}

	// Normalize separators for "x-high"/"x_high" variants.
	value = strings.NewReplacer("-", "", "_", "", " ", "").Replace(value)

	switch value {
	case "none", "minimal", "low", "medium", "high":
		return value
	case "xhigh", "extrahigh", "max":
		return "xhigh"
	default:
		// Only store known effort levels for now to keep UI consistent.
		return ""
	}
}

func normalizeOpenAIReasoningEffortForModel(raw, model string) string {
	if strings.EqualFold(strings.TrimSpace(raw), "none") && openai.IsGPT6SolOrLunaModelSpelling(model) {
		return "none"
	}
	if strings.EqualFold(strings.TrimSpace(raw), "max") && supportsOpenAIReasoningEffortMax(model) {
		return "max"
	}
	return normalizeOpenAIReasoningEffort(raw)
}

// supportsOpenAIReasoningEffortMax reports model families whose upstream scale
// has a distinct max level. Other models keep the legacy max -> xhigh behavior.
func supportsOpenAIReasoningEffortMax(model string) bool {
	if isOpenAIGPT6Model(model) || isOpenAIGPT56Model(model) {
		return true
	}

	normalized := strings.ToLower(lastOpenAIModelSegment(model))
	normalized = strings.ReplaceAll(normalized, "_", "-")
	switch {
	case strings.HasPrefix(normalized, "deepseek-v4"), strings.HasPrefix(normalized, "deepseek-flash"):
		// deepseek-flash（= DeepSeek-V4.1-Flash）与 v4 系同为 low/high/max 档位。
		return true
	case strings.HasPrefix(normalized, "glm-"):
		return true
	case strings.HasPrefix(normalized, "kimi-"), strings.HasPrefix(normalized, "moonshot-"):
		return true
	case normalized == "k3" || strings.HasPrefix(normalized, "k3-"):
		return true
	default:
		return false
	}
}

// validateGPT61SolCompatRequest runs after model mapping, before conversion can
// discard unsupported explicit effort selections or disabled thinking.
func validateGPT61SolCompatRequest(body []byte, model string) error {
	if !openai.IsGPT61SolModelSpelling(model) {
		return nil
	}
	for _, path := range []string{"reasoning.effort", "reasoning_effort", "output_config.effort"} {
		if err := openai.ValidateGPT61SolReasoningEffort(model, gjson.GetBytes(body, path).String()); err != nil {
			return err
		}
	}
	if gjson.GetBytes(body, "thinking.type").String() == "disabled" {
		return openai.ValidateGPT61SolReasoningEffort(model, "none")
	}
	requestedModel := gjson.GetBytes(body, "model").String()
	for _, effort := range []string{"none", "minimal"} {
		if strings.HasSuffix(strings.ToLower(requestedModel), "-"+effort) {
			return openai.ValidateGPT61SolReasoningEffort(model, effort)
		}
	}
	return nil
}