@usagetap/sdk 1.4.0 → 1.7.0

package/dist/{client-CExQ8e1T.d.ts → client-C0UiaqVB.d.ts}

@@ -48,7 +48,7 @@ interface RollingCallsRateLimitState {
     source?: "current_rolling_check" | "call_records";
     nextAvailableAt?: string;
     scope?: {
-        callType?: "standard" | "API";
+        callType?: "standard";
         tier?: "standard" | "premium";
     };
 }
@@ -119,11 +119,6 @@ interface BeginCallRequest {
     requested?: RequestedEntitlements;
     feature?: string;
     tags?: string[];
-    /**
-     * Optional PAYG hold hint (USD) to authorize upfront during call_begin.
-     * When provided, overrides the plan's default hold for this call only.
-     */
-    holdUsd?: number;
     /**
      * Idempotency key for safe retries.
      * The SDK generates a unique key by default. Direct API callers should send
@@ -155,57 +150,10 @@ interface BeginCallRequest {
      * Explicit pricing mode for this call. Takes precedence over batch.
      */
     pricingMode?: "batch" | "standard";
-    /**
-     * Runtime Compression Policy resolution context. SDK adapters populate this
-     * automatically; direct integrations may provide it explicitly.
-     */
-    runtimeCompressionContext?: {
-        assignmentKey?: string;
-        projectId?: string;
-        provider?: string;
-        model?: string;
-    };
-}
-interface RuntimeCompressionPolicyDecision {
-    policyId: string;
-    policyVersionId: string;
-    version: number;
-    mode: "DRAFT" | "SHADOW" | "CANARY" | "ACTIVE" | "PAUSED" | "ARCHIVED";
-    selected: boolean;
-    snapshot: {
-        scope: {
-            minimumTokens: number;
-            messageRoles: string[];
-            contentTypes: string[];
-        };
-        methods: {
-            deterministic: {
-                enabled: true;
-            };
-            deepText: {
-                enabled: boolean;
-            };
-        };
-        gates: {
-            minimumTokensRemoved: number;
-            minimumReductionPercentage: number;
-            maximumLatencyMs: number;
-            maximumCompressionCostUsd?: number;
-            minimumProjectedNetSavingUsd?: number;
-        };
-        rollout: {
-            deterministicOnly: boolean;
-            deepTextCanaryAllowed: boolean;
-        };
-        capturePayloads: boolean;
-        payloadRetentionDays: number;
-    };
 }
 interface RuntimeCompressionMeasurement {
-    policyId: string;
-    policyVersionId: string;
-    policyVersion: number;
-    rolloutMode: RuntimeCompressionPolicyDecision["mode"];
+    status: "SHADOW" | "ACTIVE";
+    configurationUpdatedAt?: string;
     decision: "compressed" | "skipped" | "fallback" | "bypassed" | "error";
     reasonCode?: string;
     methodsAttempted: string[];
@@ -224,6 +172,7 @@ interface BalanceSummary {
     tokensRemaining?: number;
     searchesRemaining?: number;
     audioSecondsRemaining?: number;
+    agenticApiRemaining?: number;
     customMeter1Remaining?: number;
     customMeter2Remaining?: number;
 }
@@ -251,7 +200,7 @@ interface PlanSummary {
 }
 interface BeginCallResponseBody {
     callId: string;
-    callType?: "standard" | "API";
+    callType?: "standard";
     startTime: string;
     feature?: string;
     tags?: string[];
@@ -272,7 +221,6 @@ interface BeginCallResponseBody {
     stripeCustomerId?: string | null;
     batch?: boolean;
     pricingMode?: "batch" | "standard";
-    runtimeCompressionPolicy?: RuntimeCompressionPolicyDecision;
 }
 interface EndCallRequest {
     callId: string;
@@ -308,10 +256,21 @@ interface EndCallRequest {
      * inputTokens must include these tokens.
      */
     cacheWriteInputTokens?: number;
+    /** Cache writes using the provider's five-minute cache lifetime. */
+    cacheWrite5mInputTokens?: number;
+    /** Cache writes using the provider's one-hour cache lifetime. */
+    cacheWrite1hInputTokens?: number;
     reasoningTokens?: number;
     searches?: number;
     audio?: number;
     audioSeconds?: number;
+    imageInputCount?: number;
+    imageInputTokens?: number;
+    imageOutputCount?: number;
+    imageOutputTokens?: number;
+    audioInputTokens?: number;
+    cachedAudioInputTokens?: number;
+    audioOutputTokens?: number;
     error?: {
         code: string;
         message: string;
@@ -344,6 +303,7 @@ interface EndCallRequest {
 }
 interface MeteredUsage {
     calls?: number;
+    /** Provider-reported output tokens used by plan limits and Stripe token meters. */
     tokens?: number;
     reasoningTokens?: number;
     searches?: number;
@@ -370,15 +330,24 @@ interface EndCallResponseBody {
     costBreakdown?: Record<string, unknown>;
     /** Model cost before a vendor-batch discount is applied. */
     standardCostUSD?: number;
-    /** Multiplier applied to standardCostUSD (0.5 for batch, 1 for standard). */
+    /** Effective multiplier relating final costUSD to standardCostUSD. */
     pricingMultiplier?: number;
     usage?: {
         inputTokens: number;
         cachedInputTokens: number;
         cacheWriteInputTokens: number;
+        cacheWrite5mInputTokens?: number;
+        cacheWrite1hInputTokens?: number;
         billableInputTokens: number;
         responseTokens: number;
         reasoningTokens: number;
+        imageInputCount?: number;
+        imageInputTokens?: number;
+        imageOutputCount?: number;
+        imageOutputTokens?: number;
+        audioInputTokens?: number;
+        cachedAudioInputTokens?: number;
+        audioOutputTokens?: number;
     };
     metered?: MeteredUsage;
     spendVelocity?: SpendVelocitySnapshot;
@@ -582,12 +551,13 @@ interface SamplingDecisionRequest {
 }
 interface SamplingDecisionResponseBody {
     sample: boolean;
-    reason: "selected" | "rate_disabled" | "customer_excluded" | "feature_excluded" | "feature_not_included" | "below_minimum_tokens" | "rate_not_selected";
+    reason: "selected" | "watch_selected" | "rate_disabled" | "customer_excluded" | "feature_excluded" | "feature_not_included" | "below_minimum_tokens" | "rate_not_selected";
     decisionId: string;
     samplingKey: string;
     policyVersion: string;
     settingsExpiresAt: string;
     score?: number;
+    watchId?: string;
 }
 interface SamplingDecisionOptions extends RequestOptions {
     correlationId?: string;
@@ -606,6 +576,7 @@ interface CaptureSampleRequest extends SamplingDecisionInput {
     error?: unknown;
     decisionId?: string;
     policyVersion?: string;
+    watchId?: string;
 }
 interface CaptureSampleOptions extends RequestOptions {
     correlationId?: string;
@@ -674,6 +645,12 @@ interface WithUsageContext {
     begin: UsageTapSuccessResponse<BeginCallResponseBody>;
     setUsage: (usage: Partial<Omit<EndCallRequest, "callId" | "error">>) => void;
     setError: (error: EndCallRequest["error"]) => void;
+    /**
+     * Defer call_end until a streaming operation has finished. The returned
+     * function settles the call exactly once using the latest usage and error.
+     * Adapters should call it on completion, cancellation, and failure.
+     */
+    deferFinalization?: () => () => Promise<void>;
 }
 interface WithUsageOptions {
     signal?: AbortSignal;
@@ -734,7 +711,7 @@ interface ChangePlanResponseBody {
     success: boolean;
     subscription: SubscriptionSnapshot;
 }
-type CustomMeterSlot = "CUSTOM1" | "CUSTOM2";
+type CustomMeterSlot = "CUSTOM1" | "CUSTOM2" | "AGENTIC_API";
 interface IncrementCustomMeterRequest {
     customerId: string;
     /** Stable identifier for the end user responsible for this meter event. */
@@ -795,10 +772,19 @@ interface UsageMetricEvent {
         inputTokens?: number;
         cachedInputTokens?: number;
         cacheWriteInputTokens?: number;
+        cacheWrite5mInputTokens?: number;
+        cacheWrite1hInputTokens?: number;
         responseTokens?: number;
         reasoningTokens?: number;
         searches?: number;
         audioSeconds?: number;
+        imageInputCount?: number;
+        imageInputTokens?: number;
+        imageOutputCount?: number;
+        imageOutputTokens?: number;
+        audioInputTokens?: number;
+        cachedAudioInputTokens?: number;
+        audioOutputTokens?: number;
         costUsd?: number;
         /** Custom meter slot (for custom_meter events) */
         customMeterSlot?: CustomMeterSlot;
@@ -1006,6 +992,96 @@ interface GatewayChatCompletion {
     usagetap?: JsonRecord;
     [key: string]: unknown;
 }
+type GatewayResponseVerbosity = "low" | "medium" | "high";
+interface GatewayPromptCacheBreakpoint {
+    mode: "explicit";
+}
+interface GatewayResponseInputText {
+    type: "input_text";
+    text: string;
+    prompt_cache_breakpoint?: GatewayPromptCacheBreakpoint;
+    [key: string]: unknown;
+}
+interface GatewayResponseInputMessage {
+    type?: "message";
+    role: "system" | "developer" | "user" | "assistant";
+    content: string | Array<GatewayResponseInputText | JsonRecord>;
+    [key: string]: unknown;
+}
+type GatewayResponseInputItem = GatewayResponseInputMessage | ({
+    type: string;
+} & JsonRecord);
+type GatewayResponseTextFormat = {
+    type: "text";
+} | {
+    type: "json_object";
+} | {
+    type: "json_schema";
+    name: string;
+    schema: JsonRecord;
+    description?: string;
+    strict?: boolean;
+};
+interface GatewayResponseTextConfig {
+    format?: GatewayResponseTextFormat;
+    verbosity?: GatewayResponseVerbosity;
+}
+interface GatewayResponsePromptCacheOptions {
+    mode?: "implicit" | "explicit";
+    ttl?: "30m";
+}
+interface GatewayResponseCreateParams {
+    model?: string;
+    models?: string[];
+    input: string | GatewayResponseInputItem[];
+    instructions?: string;
+    tools?: Array<{
+        type: string;
+        [key: string]: unknown;
+    }>;
+    tool_choice?: unknown;
+    reasoning?: {
+        effort?: "none" | "minimal" | "low" | "medium" | "high" | "xhigh" | "max";
+        [key: string]: unknown;
+    };
+    text?: GatewayResponseTextConfig;
+    prompt_cache_key?: string;
+    prompt_cache_options?: GatewayResponsePromptCacheOptions;
+    max_output_tokens?: number;
+    stream?: false;
+    background?: false;
+    /** Ask the Gateway to compress eligible request text before provider invocation. */
+    compress?: boolean;
+    [key: string]: unknown;
+}
+interface GatewayResponse {
+    id: string;
+    object?: string;
+    status?: string;
+    model?: string;
+    output: Array<{
+        type: string;
+        [key: string]: unknown;
+    }>;
+    output_text?: string;
+    usage?: {
+        input_tokens?: number;
+        output_tokens?: number;
+        total_tokens?: number;
+        input_tokens_details?: {
+            cached_tokens?: number;
+            cache_write_tokens?: number;
+            [key: string]: unknown;
+        };
+        output_tokens_details?: {
+            reasoning_tokens?: number;
+            [key: string]: unknown;
+        };
+        [key: string]: unknown;
+    };
+    usagetap?: JsonRecord;
+    [key: string]: unknown;
+}
 interface GatewayModel {
     id: string;
     object?: string;
@@ -1108,6 +1184,10 @@ declare class GatewayResource {
             create: <T extends GatewayChatCompletion = GatewayChatCompletion>(params: GatewayChatCompletionCreateParams, options?: ResourceRequestOptions) => Promise<T>;
         };
     };
+    /** Buffered OpenAI Responses-compatible requests. */
+    readonly responses: {
+        create: <T extends GatewayResponse = GatewayResponse>(params: GatewayResponseCreateParams, options?: ResourceRequestOptions) => Promise<T>;
+    };
     readonly models: {
         list: (options?: ResourceRequestOptions) => Promise<GatewayModelList>;
     };
@@ -1222,4 +1302,4 @@ declare class UsageTapClient {
     private toApiError;
 }
 
-export { type PromptCompressionResponseBody as $, compressPrompt as A, type BeginCallRequest as B, compressPromptMessages as C, compressPromptHeuristic as D, type EndCallRequest as E, compressPromptToon as F, type GatewayBatch as G, estimatePromptTokens as H, protect as I, protectPromptText as J, type PromptCompressionInput as K, type PromptCompressionMessagesAggressiveness as L, type MeterRequest as M, type PromptCompressionMessagesInput as N, type PromptCompressionMessageRole as O, type PromptCompressionProvider as P, type PromptCompressionResult as Q, type ResourceRequestOptions as R, type SamplingOptions as S, type PromptCompressionRoleAggressiveness as T, UsageTapClient as U, type VendorHints as V, type WithUsageOptions as W, type BeginCallOptions as X, type PromptCompressionRequest as Y, type PromptCompressionOptions as Z, type PromptCompressionMessagesOptions as _, type PromptCompressionMode as a, type PromptCompressionStandaloneOptions as a0, type RecordPromptCompressionRequest as a1, type CreateCustomerOptions as a2, type CreateCustomerRequest as a3, type CreateCustomerResponseBody as a4, type CheckUsageOptions as a5, type CheckUsageRequest as a6, type CheckUsageResponseBody as a7, type ChangePlanOptions as a8, type ChangePlanRequest as a9, type UsageTapClientOptions as aA, type UsageTapErrorResponse as aB, type UsageTapResultEnvelope as aC, type UsageTapResultStatus as aD, type UsageTapLogEntry as aE, type WithUsageContext as aF, type LimitType as aG, type SubscriptionSnapshot as aH, type ModelHints as aI, type IdempotencyMetadata as aJ, type UsageMetricEvent as aK, type SamplingDecisionInput as aL, type CaptureSampleRequest as aM, type CaptureSampleOptions as aN, type CaptureSampleResponseBody as aO, type SamplingSettings as aP, type GetSamplingSettingsOptions as aQ, type SamplingDecisionRequest as aR, type SamplingDecisionResponseBody as aS, type SamplingDecisionOptions as aT, type CircuitBreakerOptions as aU, type CircuitBreakerDecision as aV, type ChangePlanResponseBody as aa, type ChangePlanStrategy as ab, type IncrementCustomMeterOptions as ac, type IncrementCustomMeterRequest as ad, type IncrementCustomMeterResponseBody as ae, type CustomMeterSlot as af, type EndCallOptions as ag, type EndCallResponseBody as ah, type BalanceSummary as ai, type SpendVelocitySnapshot as aj, type SpendVelocityWindow as ak, type SpendVelocityWindowKey as al, type EntitlementHints as am, type PlanSummary as an, type MeterSummary as ao, type MeterSnapshot as ap, type MeteredUsage as aq, type RateLimitsSnapshot as ar, type RemainingRatios as as, type RollingCallsRateLimitState as at, type RequestedEntitlements as au, type AllowedEntitlements as av, type ReasoningLevel as aw, type ReasoningEffort as ax, type ReasoningEffortSource as ay, type RetryOptions as az, type PromptCompressionTelemetry as b, type UsageTapSuccessResponse as c, type BeginCallResponseBody as d, type UsageTapCompressOptions as e, type UsageTapCompressResult as f, type GatewayBatchCreateParams as g, type GatewayBatchRequest as h, type GatewayBatchResult as i, type GatewayBatchStatus as j, type GatewayChatCompletion as k, type GatewayChatCompletionCreateParams as l, type GatewayChatMessage as m, type GatewayModel as n, type GatewayModelList as o, type SummarizationBatch as p, type SummarizationBatchCreateParams as q, type SummarizationBatchJob as r, type SummarizationContext as s, type SummarizationCreateParams as t, type SummarizationJob as u, type SummarizationMeasurement as v, type SummarizationMeasurementCreateParams as w, type SummarizationProfile as x, type SummarizationStatus as y, type WaitOptions as z };
+export { type PromptCompressionInput as $, type SummarizationBatchCreateParams as A, type BeginCallRequest as B, type SummarizationBatchJob as C, type SummarizationContext as D, type EndCallRequest as E, type SummarizationCreateParams as F, type GatewayBatch as G, type SummarizationJob as H, type SummarizationMeasurement as I, type SummarizationMeasurementCreateParams as J, type SummarizationProfile as K, type SummarizationStatus as L, type MeterRequest as M, type WaitOptions as N, compressPrompt as O, type PromptCompressionProvider as P, compressPromptMessages as Q, type ResourceRequestOptions as R, type SamplingOptions as S, compressPromptHeuristic as T, UsageTapClient as U, type VendorHints as V, type WithUsageOptions as W, compressPromptToon as X, estimatePromptTokens as Y, protect as Z, protectPromptText as _, type PromptCompressionMode as a, type SamplingDecisionRequest as a$, type PromptCompressionMessagesAggressiveness as a0, type PromptCompressionMessagesInput as a1, type PromptCompressionMessageRole as a2, type PromptCompressionResult as a3, type PromptCompressionRoleAggressiveness as a4, type BeginCallOptions as a5, type PromptCompressionRequest as a6, type PromptCompressionOptions as a7, type PromptCompressionMessagesOptions as a8, type PromptCompressionResponseBody as a9, type MeteredUsage as aA, type RateLimitsSnapshot as aB, type RemainingRatios as aC, type RollingCallsRateLimitState as aD, type RequestedEntitlements as aE, type AllowedEntitlements as aF, type ReasoningLevel as aG, type ReasoningEffort as aH, type ReasoningEffortSource as aI, type RetryOptions as aJ, type UsageTapClientOptions as aK, type UsageTapErrorResponse as aL, type UsageTapResultEnvelope as aM, type UsageTapResultStatus as aN, type UsageTapLogEntry as aO, type WithUsageContext as aP, type LimitType as aQ, type SubscriptionSnapshot as aR, type ModelHints as aS, type IdempotencyMetadata as aT, type UsageMetricEvent as aU, type SamplingDecisionInput as aV, type CaptureSampleRequest as aW, type CaptureSampleOptions as aX, type CaptureSampleResponseBody as aY, type SamplingSettings as aZ, type GetSamplingSettingsOptions as a_, type PromptCompressionStandaloneOptions as aa, type RecordPromptCompressionRequest as ab, type CreateCustomerOptions as ac, type CreateCustomerRequest as ad, type CreateCustomerResponseBody as ae, type CheckUsageOptions as af, type CheckUsageRequest as ag, type CheckUsageResponseBody as ah, type ChangePlanOptions as ai, type ChangePlanRequest as aj, type ChangePlanResponseBody as ak, type ChangePlanStrategy as al, type IncrementCustomMeterOptions as am, type IncrementCustomMeterRequest as an, type IncrementCustomMeterResponseBody as ao, type CustomMeterSlot as ap, type EndCallOptions as aq, type EndCallResponseBody as ar, type BalanceSummary as as, type SpendVelocitySnapshot as at, type SpendVelocityWindow as au, type SpendVelocityWindowKey as av, type EntitlementHints as aw, type PlanSummary as ax, type MeterSummary as ay, type MeterSnapshot as az, type PromptCompressionTelemetry as b, type SamplingDecisionResponseBody as b0, type SamplingDecisionOptions as b1, type CircuitBreakerOptions as b2, type CircuitBreakerDecision as b3, type UsageTapSuccessResponse as c, type BeginCallResponseBody as d, type UsageTapCompressOptions as e, type UsageTapCompressResult as f, type GatewayBatchCreateParams as g, type GatewayBatchRequest as h, type GatewayBatchResult as i, type GatewayBatchStatus as j, type GatewayChatCompletion as k, type GatewayChatCompletionCreateParams as l, type GatewayChatMessage as m, type GatewayPromptCacheBreakpoint as n, type GatewayResponse as o, type GatewayResponseCreateParams as p, type GatewayResponseInputItem as q, type GatewayResponseInputMessage as r, type GatewayResponseInputText as s, type GatewayResponsePromptCacheOptions as t, type GatewayResponseTextConfig as u, type GatewayResponseTextFormat as v, type GatewayResponseVerbosity as w, type GatewayModel as x, type GatewayModelList as y, type SummarizationBatch as z };