fix(billing): extend quantity validation and saturating conversions to remaining paths

Bound max-tokens fields across all relay format validators, saturate
tiered-expression rounding and audio/tool/task token conversions, and
route legacy remix ratios through the guarded setter.

AGENTS.md

@@ -100,9 +100,10 @@ Do NOT directly import or call `encoding/json` in business code. `json.RawMessag
 
 **Billing safety invariants:** Quota/billing code MUST never produce a negative charge (a credit) from arithmetic overflow or unvalidated input. Apply defense in depth:
 
-- Every user-controlled quantity that becomes a billing multiplier (image `n`, video `seconds`/`duration`, resolution/quality ratios, batch counts) MUST be bounded before it reaches quota calculation. Reject out-of-range values at request validation with a 400. Existing bounds: `dto.MaxImageN` for image generation count, `relaycommon.MaxTaskDurationSeconds` for task video duration. Reuse these constants instead of introducing new ad hoc limits for the same concepts.
+- Every user-controlled quantity that becomes a billing multiplier (image `n`, video `seconds`/`duration`, resolution/quality ratios, batch counts) MUST be bounded before it reaches quota calculation. Reject out-of-range values at request validation with a 400. Existing bounds: `dto.MaxImageN` for image generation count, `relaycommon.MaxTaskDurationSeconds` for task video duration, `maxTokensLimit` (`relay/helper/valid_request.go`) for `max_tokens`-family fields on every relay format (OpenAI, Claude, Gemini, Responses). Reuse these constants instead of introducing new ad hoc limits for the same concepts. When adding a new relay format or request DTO, bound its max-tokens and count fields in its validator from day one.
 - Watch for validation bypass paths: passthrough fields (e.g. `Extra["parameters"]`), task `metadata` maps, and multipart form fields can carry the same quantities around the standard DTO validation. Any adaptor that reads a multiplier from such a path must enforce the same bound (or clamp) locally.
-- Never convert a computed quota to `int` with a bare cast like `int(float64(quota) * ratio)` or `int(decimal.IntPart())`. Use the saturating converters: `common.QuotaFromFloat` for float products, `decimalToQuota` in `service` for decimal products. Saturation bounds are int32 because quota columns (user/token/log) are 32-bit integers in the database.
+- Durations parsed from media metadata are user/upstream-controlled too: audio file headers (transcription token counting, TTS response duration) and upstream deduction numbers (e.g. Kling `FinalUnitDeduction`) can claim absurd values. Convert them with saturation before they become token counts.
+- Never convert a computed quota or token count to `int` with a bare cast like `int(float64(quota) * ratio)`, `int(math.Round(...))` on unbounded input, or `int(decimal.IntPart())`. Use the saturating converters: `common.QuotaFromFloat` for float products, `decimalToQuota` in `service` for decimal products, `billingexpr.QuotaRound` (saturating) for tiered-expression results. Saturation bounds are int32 because quota columns (user/token/log) are 32-bit integers in the database.
 - Multiplier maps go through `types.PriceData.AddOtherRatio`, which rejects non-positive, NaN, and +Inf ratios. Do not write to `PriceData.OtherRatios` directly, and do not weaken these guards.
 - Pre-consume (预扣费) and settle (结算/差额) must both be safe: a saturated oversized quota must fail pre-consume with insufficient-quota, never silently wrap. When adding a new billing path (new relay format, new task platform, new adjustment hook), trace the full chain — validation → EstimateBilling/OtherRatios → quota conversion → pre-consume → settle/refund — and confirm each step preserves these invariants.
 - Fields parsed into unsigned types (`*uint`) accept huge positive JSON numbers (e.g. `18446744073686646784`, a wrapped negative); a `>= 0` check is not sufficient, an upper bound is mandatory.

pkg/billingexpr/billingexpr_test.go

@@ -291,6 +291,12 @@ func TestQuotaRound(t *testing.T) {
 		{999.4999, 999},
 		{999.5, 1000},
 		{1e9 + 0.5, 1e9 + 1},
+		// Oversized expression results must saturate at int32 bounds
+		// instead of wrapping into a negative charge.
+		{3.6893488147419103e19, math.MaxInt32},
+		{-3.6893488147419103e19, math.MinInt32},
+		{math.Inf(1), math.MaxInt32},
+		{math.NaN(), 0},
 	}
 	for _, tt := range tests {
 		got := billingexpr.QuotaRound(tt.in)

pkg/billingexpr/round.go

@@ -5,6 +5,20 @@ import "math"
 // QuotaRound converts a float64 quota value to int using half-away-from-zero
 // rounding. Every tiered billing path (pre-consume, settlement, breakdown
 // validation, log fields) MUST use this function to avoid +-1 discrepancies.
+//
+// The result saturates at int32 bounds: quota columns are 32-bit integers in
+// the database, and an oversized expression result must never wrap around
+// and turn a charge into a credit.
 func QuotaRound(f float64) int {
-	return int(math.Round(f))
+	r := math.Round(f)
+	if math.IsNaN(r) {
+		return 0
+	}
+	if r >= math.MaxInt32 {
+		return math.MaxInt32
+	}
+	if r <= math.MinInt32 {
+		return math.MinInt32
+	}
+	return int(r)
 }

relay/channel/openai/audio.go

@@ -103,8 +103,9 @@ func OpenaiTTSHandler(c *gin.Context, resp *http.Response, info *relaycommon.Rel
 			usage.CompletionTokens = estimatedTokens
 			usage.CompletionTokenDetails.AudioTokens = estimatedTokens
 		} else if duration > 0 {
-			// 计算 token: ceil(duration) / 60.0 * 1000，即每分钟 1000 tokens
-			completionTokens := int(math.Round(math.Ceil(duration) / 60.0 * 1000))
+			// 计算 token: ceil(duration) / 60.0 * 1000，即每分钟 1000 tokens。
+			// duration 解析自上游返回的音频元数据，饱和转换防止 int 回绕。
+			completionTokens := common.QuotaFromFloat(math.Round(math.Ceil(duration) / 60.0 * 1000))
 			usage.CompletionTokens = completionTokens
 			usage.CompletionTokenDetails.AudioTokens = completionTokens
 		}

relay/channel/task/kling/adaptor.go

@@ -358,7 +358,8 @@ func (a *TaskAdaptor) ParseTaskResult(respBody []byte) (*relaycommon.TaskInfo, e
 			taskInfo.Url = video.Url
 		}
 		if tokens, err := strconv.ParseFloat(resPayload.Data.FinalUnitDeduction, 64); err == nil {
-			rounded := int(math.Ceil(tokens))
+			// 上游返回的扣费数值，饱和转换防止超大数值回绕成负数
+			rounded := common.QuotaFromFloat(math.Ceil(tokens))
 			if rounded > 0 {
 				taskInfo.CompletionTokens = rounded
 				taskInfo.TotalTokens = rounded

relay/helper/max_tokens_bounds_test.go

@@ -0,0 +1,72 @@
+package helper
+
+import (
+	"bytes"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	relayconstant "github.com/QuantumNous/new-api/relay/constant"
+	"github.com/gin-gonic/gin"
+	"github.com/stretchr/testify/require"
+)
+
+// TestMaxTokensBounds guards the billing invariant that user-supplied max
+// token fields are bounded on every relay format. These values feed
+// pre-consume quota math (preConsumedTokens * ratio); a huge or
+// wrapped-negative value (e.g. 18446744073686646784 parsed into *uint) must
+// be rejected at validation instead of corrupting the pre-charge.
+func TestMaxTokensBounds(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	newJSONContext := func(t *testing.T, body string) *gin.Context {
+		c, _ := gin.CreateTestContext(httptest.NewRecorder())
+		c.Request = httptest.NewRequest(http.MethodPost, "/relay", bytes.NewBufferString(body))
+		c.Request.Header.Set("Content-Type", "application/json")
+		return c
+	}
+
+	const hugeN = "18446744073686646784"
+
+	t.Run("openai max_tokens overflow rejected", func(t *testing.T) {
+		c := newJSONContext(t, `{"model":"gpt-4o","messages":[{"role":"user","content":"hi"}],"max_tokens":`+hugeN+`}`)
+		_, err := GetAndValidateTextRequest(c, relayconstant.RelayModeChatCompletions)
+		require.Error(t, err)
+		require.Contains(t, err.Error(), "max_tokens is invalid")
+	})
+
+	t.Run("openai max_completion_tokens overflow rejected", func(t *testing.T) {
+		c := newJSONContext(t, `{"model":"gpt-4o","messages":[{"role":"user","content":"hi"}],"max_completion_tokens":`+hugeN+`}`)
+		_, err := GetAndValidateTextRequest(c, relayconstant.RelayModeChatCompletions)
+		require.Error(t, err)
+		require.Contains(t, err.Error(), "max_tokens is invalid")
+	})
+
+	t.Run("claude max_tokens overflow rejected", func(t *testing.T) {
+		c := newJSONContext(t, `{"model":"claude-sonnet-4","messages":[{"role":"user","content":"hi"}],"max_tokens":`+hugeN+`}`)
+		_, err := GetAndValidateClaudeRequest(c)
+		require.Error(t, err)
+		require.Contains(t, err.Error(), "max_tokens is invalid")
+	})
+
+	t.Run("claude normal max_tokens accepted", func(t *testing.T) {
+		c := newJSONContext(t, `{"model":"claude-sonnet-4","messages":[{"role":"user","content":"hi"}],"max_tokens":8192}`)
+		req, err := GetAndValidateClaudeRequest(c)
+		require.NoError(t, err)
+		require.EqualValues(t, 8192, *req.MaxTokens)
+	})
+
+	t.Run("gemini maxOutputTokens overflow rejected", func(t *testing.T) {
+		c := newJSONContext(t, `{"contents":[{"parts":[{"text":"hi"}]}],"generationConfig":{"maxOutputTokens":`+hugeN+`}}`)
+		_, err := GetAndValidateGeminiRequest(c)
+		require.Error(t, err)
+		require.Contains(t, err.Error(), "maxOutputTokens is invalid")
+	})
+
+	t.Run("responses max_output_tokens overflow rejected", func(t *testing.T) {
+		c := newJSONContext(t, `{"model":"gpt-4o","input":"hi","max_output_tokens":`+hugeN+`}`)
+		_, err := GetAndValidateResponsesRequest(c)
+		require.Error(t, err)
+		require.Contains(t, err.Error(), "max_output_tokens is invalid")
+	})
+}

relay/helper/price.go

@@ -112,12 +112,12 @@ func ModelPriceHelper(c *gin.Context, info *relaycommon.RelayInfo, promptTokens
 		audioRatio = ratio_setting.GetAudioRatio(info.OriginModelName)
 		audioCompletionRatio = ratio_setting.GetAudioCompletionRatio(info.OriginModelName)
 		ratio := modelRatio * groupRatioInfo.GroupRatio
-		preConsumedQuota = int(float64(preConsumedTokens) * ratio)
+		preConsumedQuota = common.QuotaFromFloat(float64(preConsumedTokens) * ratio)
 	} else {
 		if meta.ImagePriceRatio != 0 {
 			modelPrice = modelPrice * meta.ImagePriceRatio
 		}
-		preConsumedQuota = int(modelPrice * common.QuotaPerUnit * groupRatioInfo.GroupRatio)
+		preConsumedQuota = common.QuotaFromFloat(modelPrice * common.QuotaPerUnit * groupRatioInfo.GroupRatio)
 	}
 
 	// check if free model pre-consume is disabled
@@ -194,7 +194,7 @@ func ModelPriceHelperPerCall(c *gin.Context, info *relaycommon.RelayInfo) (types
 	freeModel := false
 
 	if usePrice {
-		quota = int(modelPrice * common.QuotaPerUnit * groupRatioInfo.GroupRatio)
+		quota = common.QuotaFromFloat(modelPrice * common.QuotaPerUnit * groupRatioInfo.GroupRatio)
 		if !operation_setting.GetQuotaSetting().EnableFreeModelPreConsume {
 			if groupRatioInfo.GroupRatio == 0 || modelPrice == 0 {
 				quota = 0
@@ -203,7 +203,7 @@ func ModelPriceHelperPerCall(c *gin.Context, info *relaycommon.RelayInfo) (types
 		}
 	} else {
 		// 按量计费：以模型倍率的一半作为预扣额度
-		quota = int(modelRatio / 2 * common.QuotaPerUnit * groupRatioInfo.GroupRatio)
+		quota = common.QuotaFromFloat(modelRatio / 2 * common.QuotaPerUnit * groupRatioInfo.GroupRatio)
 		modelPrice = -1
 		if !operation_setting.GetQuotaSetting().EnableFreeModelPreConsume {
 			if groupRatioInfo.GroupRatio == 0 || modelRatio == 0 {

relay/helper/valid_request.go

@@ -114,6 +114,20 @@ func GetAndValidateEmbeddingRequest(c *gin.Context, relayMode int) (*dto.Embeddi
 	return embeddingRequest, nil
 }
 
+// maxTokensLimit bounds user-supplied max token fields. These values feed
+// pre-consume quota math (preConsumedTokens * ratio); an unbounded value can
+// overflow the conversion and corrupt billing.
+const maxTokensLimit = math.MaxInt32 / 2
+
+func exceedsMaxTokensLimit(values ...*uint) bool {
+	for _, v := range values {
+		if lo.FromPtrOr(v, uint(0)) > maxTokensLimit {
+			return true
+		}
+	}
+	return false
+}
+
 func GetAndValidateResponsesRequest(c *gin.Context) (*dto.OpenAIResponsesRequest, error) {
 	request := &dto.OpenAIResponsesRequest{}
 	err := common.UnmarshalBodyReusable(c, request)
@@ -126,6 +140,9 @@ func GetAndValidateResponsesRequest(c *gin.Context) (*dto.OpenAIResponsesRequest
 	if request.Input == nil {
 		return nil, errors.New("input is required")
 	}
+	if exceedsMaxTokensLimit(request.MaxOutputTokens) {
+		return nil, errors.New("max_output_tokens is invalid")
+	}
 	return request, nil
 }
 
@@ -259,6 +276,9 @@ func GetAndValidateClaudeRequest(c *gin.Context) (textRequest *dto.ClaudeRequest
 	if textRequest.Model == "" {
 		return nil, errors.New("field model is required")
 	}
+	if exceedsMaxTokensLimit(textRequest.MaxTokens, textRequest.MaxTokensToSample) {
+		return nil, errors.New("max_tokens is invalid")
+	}
 
 	//if textRequest.Stream {
 	//	relayInfo.IsStream = true
@@ -281,7 +301,7 @@ func GetAndValidateTextRequest(c *gin.Context, relayMode int) (*dto.GeneralOpenA
 		textRequest.Model = c.Param("model")
 	}
 
-	if lo.FromPtrOr(textRequest.MaxTokens, uint(0)) > math.MaxInt32/2 {
+	if exceedsMaxTokensLimit(textRequest.MaxTokens, textRequest.MaxCompletionTokens) {
 		return nil, errors.New("max_tokens is invalid")
 	}
 	if textRequest.Model == "" {
@@ -334,6 +354,9 @@ func GetAndValidateGeminiRequest(c *gin.Context) (*dto.GeminiChatRequest, error)
 	if len(request.Contents) == 0 && len(request.Requests) == 0 {
 		return nil, errors.New("contents is required")
 	}
+	if exceedsMaxTokensLimit(request.GenerationConfig.MaxOutputTokens) {
+		return nil, errors.New("maxOutputTokens is invalid")
+	}
 
 	//if c.Query("alt") == "sse" {
 	//	relayInfo.IsStream = true

relay/relay_task.go

@@ -121,14 +121,15 @@ func ResolveOriginTask(c *gin.Context, info *relaycommon.RelayInfo) *dto.TaskErr
 			if seconds <= 0 {
 				seconds = 4
 			}
-			sizeStr, _ := taskData["size"].(string)
-			if info.PriceData.OtherRatios == nil {
-				info.PriceData.OtherRatios = map[string]float64{}
+			// 历史任务数据可能包含未经校验的时长，作为计费乘数前必须钳制
+			if seconds > relaycommon.MaxTaskDurationSeconds {
+				seconds = relaycommon.MaxTaskDurationSeconds
 			}
-			info.PriceData.OtherRatios["seconds"] = float64(seconds)
-			info.PriceData.OtherRatios["size"] = 1
+			sizeStr, _ := taskData["size"].(string)
+			info.PriceData.AddOtherRatio("seconds", float64(seconds))
+			info.PriceData.AddOtherRatio("size", 1)
 			if sizeStr == "1792x1024" || sizeStr == "1024x1792" {
-				info.PriceData.OtherRatios["size"] = 1.666667
+				info.PriceData.AddOtherRatio("size", 1.666667)
 			}
 		}
 	}

service/quota.go

@@ -54,7 +54,7 @@ func calculateAudioQuota(info QuotaInfo) int {
 		groupRatio := decimal.NewFromFloat(info.GroupRatio)
 
 		quota := modelPrice.Mul(quotaPerUnit).Mul(groupRatio)
-		return int(quota.IntPart())
+		return decimalToQuota(quota)
 	}
 
 	completionRatio := decimal.NewFromFloat(ratio_setting.GetCompletionRatio(info.ModelName))
@@ -83,7 +83,7 @@ func calculateAudioQuota(info QuotaInfo) int {
 		quota = decimal.NewFromInt(1)
 	}
 
-	return int(quota.Round(0).IntPart())
+	return decimalToQuota(quota)
 }
 
 func PreWssConsumeQuota(ctx *gin.Context, relayInfo *relaycommon.RelayInfo, usage *dto.RealtimeUsage) error {

service/text_quota.go

@@ -145,15 +145,13 @@ func composeTieredTextQuota(relayInfo *relaycommon.RelayInfo, summary textQuotaS
 
 	if tieredResult != nil {
 		if snap := relayInfo.TieredBillingSnapshot; snap != nil {
-			return int(decimal.NewFromFloat(tieredResult.ActualQuotaBeforeGroup).
+			return decimalToQuota(decimal.NewFromFloat(tieredResult.ActualQuotaBeforeGroup).
 				Mul(decimal.NewFromFloat(snap.GroupRatio)).
-				Add(summary.ToolCallSurchargeQuota).
-				Round(0).
-				IntPart())
+				Add(summary.ToolCallSurchargeQuota))
 		}
 	}
 
-	return tieredQuota + int(summary.ToolCallSurchargeQuota.Round(0).IntPart())
+	return tieredQuota + decimalToQuota(summary.ToolCallSurchargeQuota)
 }
 
 func calculateTextQuotaSummary(ctx *gin.Context, relayInfo *relaycommon.RelayInfo, usage *dto.Usage) textQuotaSummary {

service/token_counter.go

@@ -208,8 +208,14 @@ func EstimateRequestToken(c *gin.Context, meta *types.TokenCountMeta, info *rela
 			if err != nil {
 				return 0, fmt.Errorf("error getting audio duration: %v", err)
 			}
-			// 一分钟 1000 token，与 $price / minute 对齐
-			totalAudioToken += int(math.Round(math.Ceil(duration) / 60.0 * 1000))
+			// 一分钟 1000 token，与 $price / minute 对齐。
+			// duration 来自用户上传文件的元数据，可被伪造成天文数字，
+			// 必须饱和转换防止 int 回绕成负数 token。
+			audioTokens := common.QuotaFromFloat(math.Round(math.Ceil(duration) / 60.0 * 1000))
+			if audioTokens < 0 {
+				audioTokens = 0
+			}
+			totalAudioToken += audioTokens
 		}
 		return totalAudioToken, nil
 	}
@@ -377,7 +383,8 @@ func CountAudioTokenInput(audioBase64 string, audioFormat string) (int, error) {
 	if err != nil {
 		return 0, err
 	}
-	return int(duration / 60 * 100 / 0.06), nil
+	// duration 来自用户提供的音频元数据，饱和转换防止 int 回绕
+	return common.QuotaFromFloat(duration / 60 * 100 / 0.06), nil
 }
 
 func CountAudioTokenOutput(audioBase64 string, audioFormat string) (int, error) {
@@ -388,7 +395,8 @@ func CountAudioTokenOutput(audioBase64 string, audioFormat string) (int, error)
 	if err != nil {
 		return 0, err
 	}
-	return int(duration / 60 * 200 / 0.24), nil
+	// duration 来自上游返回的音频元数据，饱和转换防止 int 回绕
+	return common.QuotaFromFloat(duration / 60 * 200 / 0.24), nil
 }
 
 // CountTextToken 统计文本的token数量，仅OpenAI模型使用tokenizer，其余模型使用估算

service/tool_billing.go

@@ -49,7 +49,7 @@ func ComputeToolCallQuota(usage ToolCallUsage, groupRatio float64) ToolCallResul
 			return
 		}
 		totalPrice := pricePer1K * float64(count) / 1000
-		quota := int(math.Round(totalPrice * common.QuotaPerUnit * groupRatio))
+		quota := common.QuotaFromFloat(math.Round(totalPrice * common.QuotaPerUnit * groupRatio))
 		items = append(items, ToolCallItem{
 			Name:       toolName,
 			CallCount:  count,
@@ -70,7 +70,7 @@ func ComputeToolCallQuota(usage ToolCallUsage, groupRatio float64) ToolCallResul
 
 	if usage.ImageGenerationCall {
 		price := operation_setting.GetGPTImage1PriceOnceCall(usage.ImageGenerationQuality, usage.ImageGenerationSize)
-		quota := int(math.Round(price * common.QuotaPerUnit * groupRatio))
+		quota := common.QuotaFromFloat(math.Round(price * common.QuotaPerUnit * groupRatio))
 		items = append(items, ToolCallItem{
 			Name:       "image_generation",
 			CallCount:  1,