@usagetap/sdk 1.3.2 → 1.7.0

package/dist/adapters/openai.cjs

@@ -242,6 +242,11 @@ async function compressMessagesWithUsageTap(options) {
     aggressiveness,
     "UsageTap prompt message compression"
   );
+  if (options.latencyBudgetMs !== void 0 && (!Number.isFinite(options.latencyBudgetMs) || options.latencyBudgetMs < 0)) {
+    throw new Error(
+      "UsageTap prompt message compression latencyBudgetMs must be a non-negative number"
+    );
+  }
   const original = stableStringifyInput(options.input);
   const headers = {
     "content-type": "application/json"
@@ -256,7 +261,15 @@ async function compressMessagesWithUsageTap(options) {
       headers,
       body: JSON.stringify({
         ...cloneInputRecord(options.input),
-        compression_settings: { aggressiveness }
+        compression_settings: {
+          aggressiveness,
+          ...options.mode === void 0 ? {} : { mode: options.mode },
+          ...options.latencyBudgetMs === void 0 ? {} : { latency_budget_ms: options.latencyBudgetMs },
+          ...options.compactEmptyUserMessages === void 0 ? {} : { compact_empty_user_messages: options.compactEmptyUserMessages },
+          ...options.compactDuplicateUserTextParts === void 0 ? {} : {
+            compact_duplicate_user_text_parts: options.compactDuplicateUserTextParts
+          }
+        }
       }),
       signal: options.signal
     }
@@ -752,10 +765,369 @@ function scalarToToon(value) {
   return JSON.stringify(text);
 }
 
+// src/resources.ts
+var CANONICAL_MEDIA_TYPE = "application/vnd.usagetap.v1+json";
+var DEFAULT_GATEWAY_BASE_URL = "https://gateway.usagetap.com";
+function normalizedBaseUrl(value) {
+  return `${value.replace(/\/+$/, "")}/`;
+}
+function errorMessage(payload, status) {
+  if (payload && typeof payload === "object") {
+    const record = payload;
+    const error = record.error;
+    if (error && typeof error === "object") {
+      const message2 = error.message;
+      if (typeof message2 === "string" && message2) return message2;
+    }
+    const message = record.message;
+    if (typeof message === "string" && message) return message;
+  }
+  return `UsageTap request failed with HTTP ${status}`;
+}
+function errorCode(status) {
+  if (status === 401 || status === 403) return "USAGETAP_AUTH_ERROR";
+  if (status === 429) return "USAGETAP_RATE_LIMITED";
+  if (status >= 500) return "USAGETAP_SERVER_ERROR";
+  return "USAGETAP_BAD_REQUEST";
+}
+var ResourceTransport = class {
+  baseUrl;
+  apiKey;
+  fetchImpl;
+  defaultHeaders;
+  sdkVersion;
+  constructor(baseUrl, config) {
+    this.baseUrl = normalizedBaseUrl(baseUrl);
+    this.apiKey = config.apiKey;
+    this.fetchImpl = config.fetchImpl;
+    this.defaultHeaders = config.headers ?? {};
+    this.sdkVersion = config.sdkVersion;
+  }
+  async request(request) {
+    const body = request.body === void 0 ? void 0 : JSON.stringify(request.body);
+    const headers = {
+      ...this.defaultHeaders,
+      accept: request.response === "data" ? CANONICAL_MEDIA_TYPE : request.response === "ndjson" ? "application/x-ndjson" : "application/json",
+      authorization: `Bearer ${this.apiKey}`,
+      "x-usage-sdk": `js/${this.sdkVersion}`,
+      ...body ? { "content-type": "application/json" } : {},
+      ...request.options?.idempotencyKey ? { "idempotency-key": request.options.idempotencyKey } : {},
+      ...request.options?.headers
+    };
+    let response;
+    try {
+      response = await this.fetchImpl(
+        new URL(request.path.replace(/^\/+/, ""), this.baseUrl),
+        {
+          method: request.method,
+          headers,
+          body,
+          signal: request.options?.signal
+        }
+      );
+    } catch (error) {
+      throw new UsageTapError(
+        "USAGETAP_NETWORK_ERROR",
+        "Failed to reach UsageTap",
+        { retryable: true, cause: error }
+      );
+    }
+    const requestId = response.headers.get("x-request-id") ?? response.headers.get("x-usage-correlation-id") ?? void 0;
+    const text = await response.text();
+    let payload;
+    if (text && request.response !== "ndjson") {
+      try {
+        payload = JSON.parse(text);
+      } catch (error) {
+        throw new UsageTapError(
+          "USAGETAP_INVALID_RESPONSE",
+          "UsageTap returned invalid JSON",
+          { status: response.status, correlationId: requestId, cause: error }
+        );
+      }
+    }
+    if (!response.ok) {
+      throw new UsageTapError(
+        errorCode(response.status),
+        errorMessage(payload, response.status),
+        {
+          status: response.status,
+          retryable: response.status === 429 || response.status >= 500,
+          correlationId: requestId,
+          details: payload && typeof payload === "object" ? payload : void 0
+        }
+      );
+    }
+    if (request.response === "ndjson") {
+      if (!text.trim()) return [];
+      try {
+        return text.trim().split(/\r?\n/).filter(Boolean).map((line) => JSON.parse(line));
+      } catch (error) {
+        throw new UsageTapError(
+          "USAGETAP_INVALID_RESPONSE",
+          "UsageTap returned invalid NDJSON",
+          { status: response.status, correlationId: requestId, cause: error }
+        );
+      }
+    }
+    if (request.response === "data") {
+      if (!payload || typeof payload !== "object" || !("data" in payload)) {
+        throw new UsageTapError(
+          "USAGETAP_INVALID_RESPONSE",
+          "UsageTap response missing data",
+          { status: response.status, correlationId: requestId }
+        );
+      }
+      return payload.data;
+    }
+    if (payload === void 0) {
+      throw new UsageTapError(
+        "USAGETAP_INVALID_RESPONSE",
+        "UsageTap response was empty",
+        { status: response.status, correlationId: requestId }
+      );
+    }
+    return payload;
+  }
+};
+function resourceId(value, keys, label) {
+  const id = typeof value === "string" ? value : keys.map((key) => value[key]).find((candidate) => Boolean(candidate?.trim()));
+  if (!id?.trim()) {
+    throw new UsageTapError(
+      "USAGETAP_BAD_REQUEST",
+      `${label} requires a non-empty ID`
+    );
+  }
+  return id.trim();
+}
+function terminalSummary(status) {
+  return status === "COMPLETE" || status === "FAILED";
+}
+function terminalGatewayBatch(status) {
+  return ["completed", "failed", "expired", "cancelled"].includes(status);
+}
+function validateWaitOptions(options) {
+  const pollIntervalMs = options.pollIntervalMs ?? 1500;
+  const timeoutMs = options.timeoutMs ?? 3 * 6e4;
+  if (!Number.isFinite(pollIntervalMs) || pollIntervalMs < 0) {
+    throw new UsageTapError(
+      "USAGETAP_BAD_REQUEST",
+      "pollIntervalMs must be a non-negative number"
+    );
+  }
+  if (!Number.isFinite(timeoutMs) || timeoutMs < 1) {
+    throw new UsageTapError(
+      "USAGETAP_BAD_REQUEST",
+      "timeoutMs must be a positive number"
+    );
+  }
+  return { pollIntervalMs, timeoutMs };
+}
+var SummarizationResource = class {
+  summaries;
+  batches;
+  profiles;
+  measurements;
+  transport;
+  constructor(config) {
+    this.transport = new ResourceTransport(config.apiBaseUrl, config);
+    this.summaries = {
+      create: (params, options) => this.transport.request({
+        method: "POST",
+        path: "/v1/compression/summaries",
+        body: params,
+        options,
+        response: "data"
+      }),
+      retrieve: (jobId, options) => this.transport.request({
+        method: "GET",
+        path: `/v1/compression/jobs/${encodeURIComponent(
+          resourceId(jobId, ["jobId"], "summaries.retrieve")
+        )}`,
+        options,
+        response: "data"
+      }),
+      wait: (job, options) => this.waitForSummary(job, options)
+    };
+    this.batches = {
+      create: (params, options) => this.transport.request({
+        method: "POST",
+        path: "/v1/compression/batches",
+        body: params,
+        options,
+        response: "data"
+      }),
+      retrieve: (batchId, options) => this.transport.request({
+        method: "GET",
+        path: `/v1/compression/batches/${encodeURIComponent(
+          resourceId(batchId, ["batchId"], "summarization.batches.retrieve")
+        )}`,
+        options,
+        response: "data"
+      }),
+      wait: (batch, options) => this.waitForBatch(batch, options)
+    };
+    this.profiles = {
+      retrieve: (profile, options) => this.transport.request({
+        method: "GET",
+        path: `/v1/compression/profiles/${encodeURIComponent(
+          resourceId(profile, [], "summarization.profiles.retrieve")
+        )}`,
+        options,
+        response: "data"
+      })
+    };
+    this.measurements = {
+      create: (params, options) => this.transport.request({
+        method: "POST",
+        path: "/v1/compression/measurements",
+        body: params,
+        options,
+        response: "data"
+      })
+    };
+  }
+  async waitForSummary(value, options = {}) {
+    const { pollIntervalMs, timeoutMs } = validateWaitOptions(options);
+    const deadline = Date.now() + timeoutMs;
+    let job = typeof value === "string" ? await this.summaries.retrieve(value, options) : value;
+    while (!terminalSummary(job.status)) {
+      if (Date.now() >= deadline) {
+        throw new UsageTapError(
+          "USAGETAP_RETRY_EXHAUSTED",
+          `Summarization job ${job.jobId} did not finish before timeout`,
+          { retryable: true }
+        );
+      }
+      await sleep(pollIntervalMs, options.signal);
+      job = await this.summaries.retrieve(job.jobId, options);
+    }
+    return job;
+  }
+  async waitForBatch(value, options = {}) {
+    const { pollIntervalMs, timeoutMs } = validateWaitOptions(options);
+    const deadline = Date.now() + timeoutMs;
+    let batch = typeof value === "string" ? await this.batches.retrieve(value, options) : value;
+    while (!terminalSummary(batch.status)) {
+      if (Date.now() >= deadline) {
+        throw new UsageTapError(
+          "USAGETAP_RETRY_EXHAUSTED",
+          `Summarization batch ${batch.batchId} did not finish before timeout`,
+          { retryable: true }
+        );
+      }
+      await sleep(pollIntervalMs, options.signal);
+      batch = await this.batches.retrieve(batch.batchId, options);
+    }
+    return batch;
+  }
+};
+var GatewayResource = class {
+  chat;
+  /** Buffered OpenAI Responses-compatible requests. */
+  responses;
+  models;
+  batches;
+  transport;
+  idempotencyGenerator;
+  constructor(config) {
+    this.transport = new ResourceTransport(
+      config.gatewayBaseUrl ?? DEFAULT_GATEWAY_BASE_URL,
+      config
+    );
+    this.idempotencyGenerator = config.idempotencyGenerator ?? createIdempotencyKey;
+    this.chat = {
+      completions: {
+        create: (params, options) => this.transport.request({
+          method: "POST",
+          path: "/v1/chat/completions",
+          body: params,
+          options,
+          response: "json"
+        })
+      }
+    };
+    this.responses = {
+      create: (params, options) => this.transport.request({
+        method: "POST",
+        path: "/v1/responses",
+        body: params,
+        options,
+        response: "json"
+      })
+    };
+    this.models = {
+      list: (options) => this.transport.request({
+        method: "GET",
+        path: "/v1/models",
+        options,
+        response: "json"
+      })
+    };
+    this.batches = {
+      create: (params, options = {}) => this.transport.request({
+        method: "POST",
+        path: "/v1/batches",
+        body: params,
+        options: {
+          ...options,
+          idempotencyKey: options.idempotencyKey ?? this.idempotencyGenerator()
+        },
+        response: "json"
+      }),
+      retrieve: (batchId, options) => this.transport.request({
+        method: "GET",
+        path: `/v1/batches/${encodeURIComponent(
+          resourceId(batchId, ["id"], "gateway.batches.retrieve")
+        )}`,
+        options,
+        response: "json"
+      }),
+      wait: (batch, options) => this.waitForBatch(batch, options),
+      cancel: (batchId, options) => this.transport.request({
+        method: "POST",
+        path: `/v1/batches/${encodeURIComponent(
+          resourceId(batchId, ["id"], "gateway.batches.cancel")
+        )}/cancel`,
+        options,
+        response: "json"
+      }),
+      results: (batchId, options) => this.transport.request({
+        method: "GET",
+        path: `/v1/batches/${encodeURIComponent(
+          resourceId(batchId, ["id"], "gateway.batches.results")
+        )}/results`,
+        options,
+        response: "ndjson"
+      })
+    };
+  }
+  async waitForBatch(value, options = {}) {
+    const { pollIntervalMs, timeoutMs } = validateWaitOptions(options);
+    const deadline = Date.now() + timeoutMs;
+    let batch = typeof value === "string" ? await this.batches.retrieve(value, options) : value;
+    while (!terminalGatewayBatch(batch.status)) {
+      if (Date.now() >= deadline) {
+        throw new UsageTapError(
+          "USAGETAP_RETRY_EXHAUSTED",
+          `Gateway batch ${batch.id} did not finish before timeout`,
+          { retryable: true }
+        );
+      }
+      await sleep(pollIntervalMs, options.signal);
+      batch = await this.batches.retrieve(batch.id, options);
+    }
+    return batch;
+  }
+};
+
 // src/client.ts
 var CALL_BEGIN_PATH = "call_begin";
 var CALL_END_PATH = "call_end";
 var COMPRESS_PROMPT_PATH = "compress_prompt";
+var SAMPLES_PATH = "samples";
+var SAMPLING_SETTINGS_PATH = "sampling/settings";
+var SAMPLING_DECIDE_PATH = "sampling/decide";
 var CHECK_USAGE_PATH = "customers/{customerId}/usage";
 var CREATE_CUSTOMER_PATH = "customers";
 var CHANGE_PLAN_PATH = "customers/{customerId}/change_plan";
@@ -766,11 +1138,16 @@ var CORRELATION_HEADER = "x-usage-correlation-id";
 var IDEMPOTENCY_HEADER = "idempotency-key";
 var SDK_HEADER = "x-usage-sdk";
 var USER_AGENT = "UsageTapClient";
-var CANONICAL_MEDIA_TYPE = "application/vnd.usagetap.v1+json";
+var CANONICAL_MEDIA_TYPE2 = "application/vnd.usagetap.v1+json";
 var DEFAULT_BASE_URL = "https://api.usagetap.com";
-var SDK_VERSION = "1.3.2" ;
+var DEFAULT_RUN_INACTIVITY_MS = 60 * 60 * 1e3;
+var SDK_VERSION = "1.7.0" ;
 var HAS_WINDOW = typeof globalThis !== "undefined" && typeof globalThis.window !== "undefined";
 var UsageTapClient = class {
+  /** OpenAI-compatible chat, model, and native batch operations. */
+  gateway;
+  /** Published-profile context summarization operations. */
+  summarization;
   apiKey;
   baseUrl;
   fetchImpl;
@@ -795,6 +1172,11 @@ var UsageTapClient = class {
   usageTapCompressionMessagesEndpoint;
   usageTapCompressionModel;
   usageTapCompressionAggressiveness;
+  sampling;
+  samplingSettingsCacheMs;
+  circuitBreaker;
+  circuitBreakerRuns = /* @__PURE__ */ new Map();
+  samplingSettingsCache;
   constructor(options = {}) {
     const apiKey = options.apiKey?.trim() || readEnvironmentVariable("USAGETAP_API_KEY");
     const baseUrl = options.baseUrl?.trim() || readEnvironmentVariable("USAGETAP_BASE_URL") || DEFAULT_BASE_URL;
@@ -817,10 +1199,21 @@ var UsageTapClient = class {
         "A global fetch implementation was not found. Pass fetchImpl in UsageTapClientOptions."
       );
     }
-    const normalizedBaseUrl = normalizeBaseUrl(baseUrl);
-    this.baseUrl = new URL(normalizedBaseUrl);
+    const normalizedBaseUrl2 = normalizeBaseUrl(baseUrl);
+    this.baseUrl = new URL(normalizedBaseUrl2);
     this.apiKey = apiKey;
     this.fetchImpl = wrapFetchImplementation(fetchCandidate, !options.fetchImpl);
+    const resourceConfig = {
+      apiKey,
+      apiBaseUrl: normalizedBaseUrl2,
+      gatewayBaseUrl: options.gatewayBaseUrl?.trim() || readEnvironmentVariable("USAGETAP_GATEWAY_URL"),
+      fetchImpl: this.fetchImpl,
+      headers: options.headers,
+      sdkVersion: SDK_VERSION,
+      idempotencyGenerator: options.idempotencyGenerator
+    };
+    this.gateway = new GatewayResource(resourceConfig);
+    this.summarization = new SummarizationResource(resourceConfig);
     this.defaultFeature = options.defaultFeature;
     this.defaultTags = options.defaultTags?.length ? dedupeStrings(options.defaultTags) : void 0;
     this.defaultHeaders = options.headers ? normalizeHeaderDictionary(options.headers) : {};
@@ -842,11 +1235,122 @@ var UsageTapClient = class {
     this.usageTapCompressionMessagesEndpoint = options.usageTapCompressionMessagesEndpoint;
     this.usageTapCompressionModel = options.usageTapCompressionModel;
     this.usageTapCompressionAggressiveness = options.usageTapCompressionAggressiveness;
+    this.sampling = options.sampling;
+    this.samplingSettingsCacheMs = Number.isFinite(options.samplingSettingsCacheMs) ? Math.max(0, Number(options.samplingSettingsCacheMs)) : 5 * 60 * 1e3;
+    if (options.circuitBreaker) {
+      const maxCallsPerRun = options.circuitBreaker.maxCallsPerRun;
+      if (!Number.isInteger(maxCallsPerRun) || maxCallsPerRun < 1) {
+        throw new UsageTapError(
+          "USAGETAP_BAD_REQUEST",
+          "circuitBreaker.maxCallsPerRun must be a positive integer"
+        );
+      }
+      const runInactivityMs = options.circuitBreaker.runInactivityMs ?? DEFAULT_RUN_INACTIVITY_MS;
+      if (!Number.isFinite(runInactivityMs) || runInactivityMs < 1) {
+        throw new UsageTapError(
+          "USAGETAP_BAD_REQUEST",
+          "circuitBreaker.runInactivityMs must be a positive number"
+        );
+      }
+      this.circuitBreaker = {
+        maxCallsPerRun,
+        runInactivityMs
+      };
+    }
+  }
+  shouldSample(request, policy = this.sampling || void 0) {
+    if (!policy) return false;
+    const rate = Math.min(1, Math.max(0, Number(policy.rate) || 0));
+    if (rate <= 0) return false;
+    const customerId = request.customerId?.trim();
+    if (customerId && policy.customers?.exclude?.includes(customerId)) return false;
+    const feature = request.feature?.trim();
+    if (feature && policy.features?.exclude?.includes(feature)) return false;
+    const included = policy.features?.include?.filter(Boolean) ?? [];
+    if (included.length > 0 && (!feature || !included.includes(feature))) return false;
+    const minimum = Math.max(0, Math.round(policy.minInputTokens ?? 0));
+    if (minimum > 0 && estimatePromptTokens(request.input) < minimum) return false;
+    return (policy.random ?? Math.random)() < rate;
+  }
+  async getSamplingSettings(options = {}) {
+    const now = Date.now();
+    if (!options.forceRefresh && this.samplingSettingsCache && this.samplingSettingsCache.expiresAtMs > now) {
+      return {
+        result: { status: "ACCEPTED", code: "SAMPLING_SETTINGS_CACHED" },
+        data: this.samplingSettingsCache.settings,
+        correlationId: options.correlationId ?? "local-cache"
+      };
+    }
+    const response = await this.requestGet(
+      SAMPLING_SETTINGS_PATH,
+      {
+        signal: options.signal,
+        headers: options.headers,
+        retries: options.retries,
+        correlationId: options.correlationId
+      }
+    );
+    const serverCacheMs = Math.max(0, Number(response.data.cacheSeconds) || 0) * 1e3;
+    const cacheMs = Math.min(this.samplingSettingsCacheMs, serverCacheMs);
+    this.samplingSettingsCache = {
+      settings: response.data,
+      expiresAtMs: now + cacheMs
+    };
+    return response;
+  }
+  async shouldSampleAsync(request, policy) {
+    if (policy) return this.shouldSample(request, policy);
+    if (this.sampling === false) return false;
+    if (this.sampling) return this.shouldSample(request, this.sampling);
+    try {
+      const settings = await this.getSamplingSettings();
+      return this.shouldSample(request, settings.data);
+    } catch {
+      return false;
+    }
+  }
+  async decideSample(request, options = {}) {
+    const hasTokens = Number.isFinite(request.inputTokens) && Number(request.inputTokens) >= 0;
+    const hasCharacters = Number.isFinite(request.inputCharacters) && Number(request.inputCharacters) >= 0;
+    if (!hasTokens && !hasCharacters) {
+      throw new UsageTapError(
+        "USAGETAP_BAD_REQUEST",
+        "decideSample requires inputTokens or inputCharacters"
+      );
+    }
+    return this.request(
+      SAMPLING_DECIDE_PATH,
+      request,
+      options
+    );
+  }
+  async captureSample(request, options = {}) {
+    if (!request || request.input === void 0) {
+      throw new UsageTapError(
+        "USAGETAP_BAD_REQUEST",
+        "captureSample requires input"
+      );
+    }
+    if (!request.provider?.trim()) {
+      throw new UsageTapError(
+        "USAGETAP_BAD_REQUEST",
+        "captureSample requires provider"
+      );
+    }
+    const sampleId = request.sampleId?.trim() || this.idempotencyGenerator();
+    return this.request(
+      SAMPLES_PATH,
+      { ...request, sampleId },
+      { ...options, idempotencyKey: sampleId }
+    );
   }
   async beginCall(request, options = {}) {
     const idempotencyKey = request.idempotencyKey ?? request.idempotency ?? (this.autoIdempotency ? this.idempotencyGenerator() : void 0);
+    this.reserveRunCall(request, idempotencyKey);
+    const apiRequest = { ...request };
+    delete apiRequest.runId;
     const payload = {
-      ...request,
+      ...apiRequest,
       feature: request.feature ?? this.defaultFeature,
       tags: this.mergeTags(request.tags)
     };
@@ -864,6 +1368,28 @@ var UsageTapClient = class {
     );
     return response;
   }
+  /**
+   * Inspect a configured run circuit breaker without consuming another call.
+   */
+  canRunContinue(request) {
+    const identity = this.resolveRunIdentity(request);
+    if (!identity || !this.circuitBreaker) {
+      throw new UsageTapError(
+        "USAGETAP_BAD_REQUEST",
+        "canRunContinue requires circuitBreaker configuration and a non-empty runId"
+      );
+    }
+    this.expireInactiveRuns();
+    const calls = this.circuitBreakerRuns.get(identity.key)?.calls ?? 0;
+    return this.createCircuitBreakerDecision(identity.customerId, identity.runId, calls);
+  }
+  /**
+   * Release local state after a workflow finishes. Returns true when state existed.
+   */
+  resetRun(request) {
+    const identity = this.resolveRunIdentity(request);
+    return identity ? this.circuitBreakerRuns.delete(identity.key) : false;
+  }
   async promptCompress(request, options = {}) {
     if (!request?.callId) {
       throw new UsageTapError(
@@ -961,6 +1487,10 @@ var UsageTapClient = class {
       usageTapCompressionMessagesEndpoint: this.usageTapCompressionMessagesEndpoint,
       aggressiveness: options.aggressiveness ?? this.aggressiveness,
       usageTapCompressionAggressiveness: options.usageTapCompressionAggressiveness ?? this.usageTapCompressionAggressiveness,
+      mode: options.mode,
+      latencyBudgetMs: options.latencyBudgetMs,
+      compactEmptyUserMessages: options.compactEmptyUserMessages,
+      compactDuplicateUserTextParts: options.compactDuplicateUserTextParts,
       fetchImpl: this.fetchImpl,
       signal: options.signal,
       failOpen: options.failOpen
@@ -1002,14 +1532,29 @@ var UsageTapClient = class {
       callId: request.callId,
       feature: feature ?? this.defaultFeature,
       tags: tags ?? this.defaultTags,
+      providerUsed: request.providerUsed,
       modelUsed: request.modelUsed,
+      reasoningEffort: request.reasoningEffort,
+      reasoningEffortSource: request.reasoningEffortSource,
+      reasoningMode: request.reasoningMode,
+      reasoningBudgetTokens: request.reasoningBudgetTokens,
       metrics: {
         inputTokens: request.inputTokens,
         responseTokens: request.responseTokens,
         cachedInputTokens: request.cachedInputTokens,
+        cacheWriteInputTokens: request.cacheWriteInputTokens,
+        cacheWrite5mInputTokens: request.cacheWrite5mInputTokens,
+        cacheWrite1hInputTokens: request.cacheWrite1hInputTokens,
         reasoningTokens: request.reasoningTokens,
         searches: request.searches,
         audioSeconds: request.audioSeconds,
+        imageInputCount: request.imageInputCount,
+        imageInputTokens: request.imageInputTokens,
+        imageOutputCount: request.imageOutputCount,
+        imageOutputTokens: request.imageOutputTokens,
+        audioInputTokens: request.audioInputTokens,
+        cachedAudioInputTokens: request.cachedAudioInputTokens,
+        audioOutputTokens: request.audioOutputTokens,
         costUsd: response.data.costUSD
       },
       correlationId: response.correlationId
@@ -1096,10 +1641,10 @@ var UsageTapClient = class {
         "incrementCustomMeter requires meterSlot"
       );
     }
-    if (!["CUSTOM1", "CUSTOM2"].includes(request.meterSlot)) {
+    if (!["CUSTOM1", "CUSTOM2", "AGENTIC_API"].includes(request.meterSlot)) {
       throw new UsageTapError(
         "USAGETAP_BAD_REQUEST",
-        "meterSlot must be CUSTOM1 or CUSTOM2"
+        "meterSlot must be CUSTOM1, CUSTOM2 or AGENTIC_API"
       );
     }
     if (typeof request.amount !== "number" || !Number.isFinite(request.amount) || request.amount <= 0) {
@@ -1114,6 +1659,15 @@ var UsageTapClient = class {
       meterSlot: request.meterSlot,
       amount: request.amount
     };
+    if (request.customerUserId) {
+      payload.customerUserId = request.customerUserId;
+    }
+    if (request.customerUserName) {
+      payload.customerUserName = request.customerUserName;
+    }
+    if (request.customerUserEmail) {
+      payload.customerUserEmail = request.customerUserEmail;
+    }
     if (request.feature) {
       payload.feature = request.feature;
     }
@@ -1150,6 +1704,11 @@ var UsageTapClient = class {
     const beginPayload = idempotencyKey ? { ...beginRequest, idempotencyKey, idempotency: idempotencyKey } : { ...beginRequest };
     const beginResponse = await this.beginCall(beginPayload, options);
     let usage = {};
+    const pricingMode = beginResponse.data.pricingMode ?? beginRequest.pricingMode ?? (beginRequest.batch === true ? "batch" : beginRequest.batch === false ? "standard" : void 0);
+    if (pricingMode) {
+      usage.pricingMode = pricingMode;
+      usage.batch = pricingMode === "batch";
+    }
     const initialStripeCustomerId = typeof beginResponse.data.stripeCustomerId === "string" ? beginResponse.data.stripeCustomerId : typeof beginRequest.stripeCustomerId === "string" ? beginRequest.stripeCustomerId : void 0;
     if (initialStripeCustomerId) {
       usage = { ...usage, stripeCustomerId: initialStripeCustomerId };
@@ -1158,6 +1717,28 @@ var UsageTapClient = class {
     let handlerResult;
     let handlerError;
     let endCallError;
+    let finalizationDeferred = false;
+    let finalizationPromise;
+    const finalize = () => {
+      if (!finalizationPromise) {
+        finalizationPromise = this.endCall(
+          {
+            callId: beginResponse.data.callId,
+            // Pass context for metric tracking
+            customerId: beginRequest.customerId,
+            feature: beginRequest.feature ?? this.defaultFeature,
+            tags: beginRequest.tags ?? this.defaultTags,
+            ...usage,
+            error: errorPayload
+          },
+          {
+            ...options,
+            correlationId: beginResponse.correlationId
+          }
+        ).then(() => void 0);
+      }
+      return finalizationPromise;
+    };
     const context = {
       begin: beginResponse,
       setUsage: (u) => {
@@ -1165,6 +1746,10 @@ var UsageTapClient = class {
       },
       setError: (err) => {
         errorPayload = err;
+      },
+      deferFinalization: () => {
+        finalizationDeferred = true;
+        return finalize;
       }
     };
     try {
@@ -1178,24 +1763,12 @@ var UsageTapClient = class {
         };
       }
     } finally {
-      try {
-        await this.endCall(
-          {
-            callId: beginResponse.data.callId,
-            // Pass context for metric tracking
-            customerId: beginRequest.customerId,
-            feature: beginRequest.feature ?? this.defaultFeature,
-            tags: beginRequest.tags ?? this.defaultTags,
-            ...usage,
-            error: errorPayload
-          },
-          {
-            ...options,
-            correlationId: beginResponse.correlationId
-          }
-        );
-      } catch (error) {
-        endCallError = error;
+      if (handlerError || !finalizationDeferred) {
+        try {
+          await finalize();
+        } catch (error) {
+          endCallError = error;
+        }
       }
     }
     if (handlerError) {
@@ -1217,17 +1790,86 @@ var UsageTapClient = class {
   toPromptCompressionTelemetry(result) {
     return {
       provider: result.provider,
-      originalCharacters: result.originalCharacters,
-      compressedCharacters: result.compressedCharacters,
-      savedCharacters: result.savedCharacters,
       originalTokens: result.originalTokens,
       compressedTokens: result.compressedTokens,
       savedTokens: result.savedTokens,
       tokenSavingsRatio: result.tokenSavingsRatio,
-      savingsRatio: result.savingsRatio,
       techniques: result.techniques
     };
   }
+  reserveRunCall(request, idempotencyKey) {
+    const identity = this.resolveRunIdentity(request);
+    if (!identity || !this.circuitBreaker) return;
+    this.expireInactiveRuns();
+    const now = Date.now();
+    const state = this.circuitBreakerRuns.get(identity.key) ?? {
+      calls: 0,
+      reservationKeys: /* @__PURE__ */ new Set(),
+      lastSeenAtMs: now
+    };
+    const reservationKey = idempotencyKey ?? this.idempotencyGenerator();
+    state.lastSeenAtMs = now;
+    if (state.reservationKeys.has(reservationKey)) {
+      this.circuitBreakerRuns.set(identity.key, state);
+      return;
+    }
+    const decision = this.createCircuitBreakerDecision(
+      identity.customerId,
+      identity.runId,
+      state.calls
+    );
+    if (!decision.allowed) {
+      throw new UsageTapError(
+        "USAGETAP_CIRCUIT_OPEN",
+        `Run ${identity.runId} reached its ${decision.limit}-call circuit-breaker limit`,
+        {
+          details: {
+            reason: decision.reason,
+            customerId: identity.customerId,
+            runId: identity.runId,
+            calls: decision.calls,
+            limit: decision.limit,
+            remaining: decision.remaining
+          }
+        }
+      );
+    }
+    state.calls += 1;
+    state.reservationKeys.add(reservationKey);
+    this.circuitBreakerRuns.set(identity.key, state);
+  }
+  resolveRunIdentity(request) {
+    const customerId = request.customerId?.trim();
+    const runId = request.runId?.trim();
+    if (!customerId || !runId) return void 0;
+    return {
+      key: `${customerId}\0${runId}`,
+      customerId,
+      runId
+    };
+  }
+  createCircuitBreakerDecision(customerId, runId, calls) {
+    const limit = this.circuitBreaker?.maxCallsPerRun ?? 0;
+    const allowed = calls < limit;
+    return {
+      allowed,
+      ...allowed ? {} : { reason: "max_calls_per_run" },
+      customerId,
+      runId,
+      calls,
+      limit,
+      remaining: Math.max(0, limit - calls)
+    };
+  }
+  expireInactiveRuns() {
+    if (!this.circuitBreaker || this.circuitBreakerRuns.size === 0) return;
+    const expiredBefore = Date.now() - this.circuitBreaker.runInactivityMs;
+    for (const [key, state] of this.circuitBreakerRuns) {
+      if (state.lastSeenAtMs < expiredBefore) {
+        this.circuitBreakerRuns.delete(key);
+      }
+    }
+  }
   async request(path, payload, options) {
     const url = new URL(path, this.baseUrl).toString();
     const body = payload !== void 0 ? JSON.stringify(payload) : void 0;
@@ -1415,7 +2057,7 @@ var UsageTapClient = class {
       ...this.defaultHeaders,
       [SDK_HEADER]: `js/${SDK_VERSION}`,
       "content-type": "application/json",
-      accept: CANONICAL_MEDIA_TYPE
+      accept: CANONICAL_MEDIA_TYPE2
     };
     if (!HAS_WINDOW) {
       headers["user-agent"] = `${USER_AGENT}/${SDK_VERSION}`;
@@ -1468,7 +2110,8 @@ var UsageTapClient = class {
   }
   toHttpError(status, payload, correlationId) {
     const code = mapStatusToErrorCode(status);
-    const retryable = isRetryableStatus(status);
+    const apiCode = payload?.error?.code ?? payload?.result?.code ?? "UNKNOWN";
+    const retryable = isRetryableStatus(status) || isRetryableApiCode(apiCode);
     const message = payload?.error?.message ?? payload?.result?.message ?? `UsageTap responded with HTTP ${status}`;
     return new UsageTapError(code, message, {
       status,
@@ -1501,7 +2144,7 @@ function isRetryableStatus(status) {
 }
 function isRetryableApiCode(code) {
   const normalized = code.toUpperCase();
-  return normalized.includes("TRANSIENT") || normalized.includes("RETRY") || normalized.includes("TIMEOUT") || normalized.includes("THROTTLE") || normalized.includes("RATE_LIMIT");
+  return normalized === "PAYG_CONFLICT" || normalized.includes("TRANSIENT") || normalized.includes("RETRY") || normalized.includes("TIMEOUT") || normalized.includes("THROTTLE") || normalized.includes("RATE_LIMIT");
 }
 function mapApiCodeToError(code) {
   const normalized = code.toUpperCase();
@@ -1616,17 +2259,18 @@ var OpenAIPromptCompressionStats = class {
   }
 };
 function createOpenAIAdapter(init) {
-  const { client, usageTap } = init;
+  const { client, usageTap, provider = "openai" } = init;
   return {
     async invoke(params) {
       const result = await usageTap.withUsage(
         params.begin,
         async (ctx) => {
+          ctx.setUsage({ providerUsed: provider });
           const response = await params.call(client, {
             hints: ctx.begin.data.vendorHints,
             begin: ctx.begin
           });
-          tryInferUsage(response, ctx.begin.data.vendorHints, params.extractUsage, ctx);
+          tryInferUsage(response, ctx.begin.data.vendorHints, params.extractUsage, ctx, provider);
           return {
             data: response,
             begin: ctx.begin
@@ -1640,24 +2284,29 @@ function createOpenAIAdapter(init) {
       const result = await usageTap.withUsage(
         params.begin,
         async (ctx) => {
+          const settle = deferUsageFinalization(ctx);
+          ctx.setUsage({ providerUsed: provider });
           const { stream, onComplete } = await params.call(client, {
             hints: ctx.begin.data.vendorHints,
             begin: ctx.begin
           });
-          const wrapped = wrapStreamForUsageTap(stream, async () => {
-            if (!onComplete) return;
+          const wrapped = wrapStreamForUsageTap(stream, async (termination) => {
             try {
-              const maybeUsage = await onComplete();
-              if (maybeUsage) {
-                ctx.setUsage(maybeUsage);
+              if (termination === "complete" && onComplete) {
+                const maybeUsage = await onComplete();
+                if (maybeUsage) {
+                  ctx.setUsage(maybeUsage);
+                }
               }
             } catch (error) {
               ctx.setError({
                 code: "USAGE_FINALIZE_ERROR",
                 message: error instanceof Error ? error.message : String(error)
               });
+              await settle();
               throw error;
             }
+            await settle();
           }, ctx);
           const finalize = async () => {
             await wrapped.__usageTapFinalize?.();
@@ -1765,6 +2414,181 @@ async function pipeToResponse(stream, res, options = {}) {
   }
 }
 var USAGETAP_CORRELATION_HEADER = "x-usage-correlation-id";
+function withSampling(client, options = {}) {
+  if (!client || !options) {
+    throw new UsageTapError(
+      "USAGETAP_BAD_REQUEST",
+      "withSampling requires an OpenAI-compatible client and sampling options"
+    );
+  }
+  const { apiKey, usageTapClient, provider = "openai", ...policy } = options;
+  const localPolicy = typeof policy.rate === "number" ? policy : void 0;
+  const usageTap = usageTapClient ?? new UsageTapClient({ apiKey, sampling: localPolicy });
+  const wrapCreate = (create) => async (params, requestOptions) => {
+    const { usageTap: callContextRaw, ...providerOptions } = requestOptions ?? {};
+    const callContext = isObjectRecord(callContextRaw) ? callContextRaw : {};
+    const streaming = params.stream === true;
+    const decision = streaming ? Promise.resolve(false) : usageTap.shouldSampleAsync({
+      customerId: readString(callContext.customerId),
+      feature: readString(callContext.feature),
+      input: params
+    }, localPolicy);
+    const startedAt = Date.now();
+    try {
+      const response = await create(
+        params,
+        Object.keys(providerOptions).length ? providerOptions : void 0
+      );
+      const selected = await decision;
+      if (selected) {
+        const record = isObjectRecord(response) ? response : {};
+        await usageTap.captureSample({
+          customerId: readString(callContext.customerId),
+          feature: readString(callContext.feature),
+          environment: readString(callContext.environment),
+          tags: readStringArray(callContext.tags),
+          provider,
+          model: readString(record.model) ?? readString(params.model),
+          input: params,
+          output: response,
+          usage: record.usage,
+          latencyMs: Date.now() - startedAt
+        }).catch(() => void 0);
+      }
+      return response;
+    } catch (error) {
+      const selected = await decision;
+      if (selected) {
+        await usageTap.captureSample({
+          customerId: readString(callContext.customerId),
+          feature: readString(callContext.feature),
+          environment: readString(callContext.environment),
+          tags: readStringArray(callContext.tags),
+          provider,
+          model: readString(params.model),
+          input: params,
+          latencyMs: Date.now() - startedAt,
+          error: serializeSamplingError(error)
+        }).catch(() => void 0);
+      }
+      throw error;
+    }
+  };
+  const chat = client.chat?.completions ? new Proxy(client.chat, {
+    get(target, prop, receiver) {
+      if (prop !== "completions") return safeReflectGet(target, prop, receiver);
+      const completions = target.completions;
+      return new Proxy(completions, {
+        get(completionTarget, completionProp, completionReceiver) {
+          if (completionProp === "create") {
+            return wrapCreate(
+              completionTarget.create.bind(completionTarget)
+            );
+          }
+          return safeReflectGet(
+            completionTarget,
+            completionProp,
+            completionReceiver
+          );
+        }
+      });
+    }
+  }) : void 0;
+  const responses = typeof client.responses !== "undefined" && client.responses ? new Proxy(client.responses, {
+    get(target, prop, receiver) {
+      if (prop === "create") {
+        const create = Reflect.get(target, prop, receiver);
+        return wrapCreate(create.bind(target));
+      }
+      return safeReflectGet(target, prop, receiver);
+    }
+  }) : void 0;
+  return new Proxy(client, {
+    get(target, prop, receiver) {
+      if (prop === "chat" && chat) return chat;
+      if (prop === "responses" && responses) return responses;
+      if (prop === "unwrap") return () => target;
+      return safeReflectGet(target, prop, receiver);
+    }
+  });
+}
+function safeReflectGet(target, prop, receiver) {
+  return Reflect.get(target, prop, receiver);
+}
+function readStringArray(value) {
+  if (!Array.isArray(value)) return void 0;
+  const strings = value.filter((item) => typeof item === "string");
+  return strings.length ? strings : void 0;
+}
+function readString(value) {
+  return typeof value === "string" && value.trim() ? value : void 0;
+}
+function serializeSamplingError(error) {
+  if (error instanceof Error) {
+    return { name: error.name, message: error.message };
+  }
+  return { message: String(error) };
+}
+function normalizeMeteredOpenAISampling(options) {
+  if (!options) return void 0;
+  if (options === true) return { provider: "openai" };
+  const { provider = "openai", ...policyFields } = options;
+  return {
+    provider,
+    policy: typeof policyFields.rate === "number" ? policyFields : void 0
+  };
+}
+function startMeteredOpenAISampleDecision({
+  usageTap,
+  sampling,
+  beginRequest,
+  input
+}) {
+  if (!sampling) return Promise.resolve(false);
+  return usageTap.shouldSampleAsync(
+    {
+      customerId: beginRequest.customerId,
+      feature: beginRequest.feature,
+      input
+    },
+    sampling.policy
+  );
+}
+async function captureMeteredOpenAISample({
+  usageTap,
+  sampling,
+  decision,
+  ctx,
+  beginRequest,
+  input,
+  response,
+  error,
+  startedAt
+}) {
+  if (!sampling) return;
+  let selected = false;
+  try {
+    selected = await decision;
+  } catch {
+    return;
+  }
+  if (!selected) return;
+  const record = isObjectRecord(response) ? response : {};
+  await usageTap.captureSample({
+    sampleId: ctx.begin.data.callId,
+    callId: ctx.begin.data.callId,
+    customerId: beginRequest.customerId,
+    feature: beginRequest.feature,
+    tags: beginRequest.tags,
+    provider: sampling.provider,
+    model: readString(record.model) ?? readString(input.model),
+    input,
+    ...response === void 0 ? {} : { output: response },
+    usage: record.usage,
+    latencyMs: Date.now() - startedAt,
+    ...error === void 0 ? {} : { error: serializeSamplingError(error) }
+  }).catch(() => void 0);
+}
 function withCompression(client, options = {}) {
   if (!client) {
     throw new UsageTapError(
@@ -1775,11 +2599,13 @@ function withCompression(client, options = {}) {
   const {
     apiKey,
     usageTapClient,
+    sampling,
     ...compressionOverrides
   } = options;
   const usageTap = usageTapClient ?? new UsageTapClient({ apiKey });
   const compression = {
     provider: "usagetap",
+    roles: { user: { enabled: true, aggressiveness: 0.2 } },
     minContextTokens: DEFAULT_PROMPT_COMPRESSION_MIN_CONTEXT_TOKENS,
     ...compressionOverrides
   };
@@ -1789,7 +2615,7 @@ function withCompression(client, options = {}) {
     usageTap,
     compression
   ) : void 0;
-  return new Proxy(client, {
+  const compressed = new Proxy(client, {
     get(target, prop, receiver) {
       if (prop === "chat" && proxiedChat) {
         return proxiedChat;
@@ -1803,6 +2629,11 @@ function withCompression(client, options = {}) {
       return Reflect.get(target, prop, receiver);
     }
   });
+  return sampling ? withSampling(compressed, {
+    ...sampling === true ? {} : sampling,
+    apiKey,
+    usageTapClient
+  }) : compressed;
 }
 function withMetering(client, customer) {
   const config = typeof customer === "string" ? { customerId: customer } : customer;
@@ -1817,6 +2648,8 @@ function withMetering(client, customer) {
     usageTapClient,
     applyVendorHints,
     promptCompression,
+    sampling,
+    provider,
     ...defaultContext
   } = config;
   const usageTap = usageTapClient ?? new UsageTapClient({ apiKey });
@@ -1824,7 +2657,9 @@ function withMetering(client, customer) {
   return wrapOpenAI(client, usageTap, {
     defaultContext,
     applyVendorHints,
-    promptCompression: normalizedCompression
+    promptCompression: normalizedCompression,
+    sampling,
+    provider
   });
 }
 function wrapOpenAI(client, usageTap, options = {}) {
@@ -1834,6 +2669,8 @@ function wrapOpenAI(client, usageTap, options = {}) {
   const defaultContext = options.defaultContext;
   const applyVendorHints = options.applyVendorHints !== false;
   const defaultPromptCompression = normalizePromptCompressionOptions(options.promptCompression);
+  const defaultSampling = normalizeMeteredOpenAISampling(options.sampling);
+  const provider = options.provider ?? "openai";
   const promptCompressionStats = new OpenAIPromptCompressionStats();
   const proxiedChat = client.chat ? createChatProxy(
     client.chat,
@@ -1841,7 +2678,9 @@ function wrapOpenAI(client, usageTap, options = {}) {
     defaultContext,
     applyVendorHints,
     defaultPromptCompression,
-    promptCompressionStats
+    promptCompressionStats,
+    defaultSampling,
+    provider
   ) : void 0;
   const proxiedResponses = typeof client.responses !== "undefined" ? createResponsesProxy(
     client.responses,
@@ -1849,7 +2688,9 @@ function wrapOpenAI(client, usageTap, options = {}) {
     defaultContext,
     applyVendorHints,
     defaultPromptCompression,
-    promptCompressionStats
+    promptCompressionStats,
+    defaultSampling,
+    provider
   ) : void 0;
   const handler = {
     get(target, prop, receiver) {
@@ -1968,14 +2809,16 @@ function streamOpenAIRoute(usageTap, openai, options) {
     });
   };
 }
-function createChatProxy(resource, usageTap, defaultContext, applyVendorHints, defaultPromptCompression, promptCompressionStats) {
+function createChatProxy(resource, usageTap, defaultContext, applyVendorHints, defaultPromptCompression, promptCompressionStats, defaultSampling, provider) {
   const completions = createChatCompletionsProxy(
     resource.completions,
     usageTap,
     defaultContext,
     applyVendorHints,
     defaultPromptCompression,
-    promptCompressionStats
+    promptCompressionStats,
+    defaultSampling,
+    provider
   );
   const handler = {
     get(target, prop, receiver) {
@@ -1987,7 +2830,7 @@ function createChatProxy(resource, usageTap, defaultContext, applyVendorHints, d
   };
   return new Proxy(resource, handler);
 }
-function createResponsesProxy(resource, usageTap, defaultContext, applyVendorHints, defaultPromptCompression, promptCompressionStats) {
+function createResponsesProxy(resource, usageTap, defaultContext, applyVendorHints, defaultPromptCompression, promptCompressionStats, defaultSampling, provider) {
   if (!resource || typeof resource !== "object") {
     return void 0;
   }
@@ -2002,9 +2845,20 @@ function createResponsesProxy(resource, usageTap, defaultContext, applyVendorHin
       withUsage,
       promptCompression
     } = splitUsageOptions(options);
-    const beginRequest = resolveBeginRequest(defaultContext, usageContext);
+    const beginRequest = responsesBeginRequest(
+      resolveBeginRequest(defaultContext, usageContext),
+      params
+    );
     const wantsStream = isStreamingRequest(params);
     return usageTap.withUsage(beginRequest, async (ctx) => {
+      const settle = wantsStream ? deferUsageFinalization(ctx) : void 0;
+      const sampleStartedAt = Date.now();
+      const sampleDecision = wantsStream ? Promise.resolve(false) : startMeteredOpenAISampleDecision({
+        usageTap,
+        sampling: defaultSampling,
+        beginRequest,
+        input: params
+      });
       const hintedParams = applyVendorHints ? applyResponsesVendorHints(params, ctx.begin.data.vendorHints) : params;
       const finalParams = await compressResponsesParamsForCall({
         params: hintedParams,
@@ -2016,25 +2870,67 @@ function createResponsesProxy(resource, usageTap, defaultContext, applyVendorHin
         withUsage,
         operation: "responses.create"
       });
+      ctx.setUsage(openAIRequestExecutionMetadata(finalParams, provider));
       const request = attachCorrelationHeader(requestOptions, ctx.begin.correlationId);
       if (wantsStream) {
         const apiPromise2 = originalCreate(finalParams, request);
         const wrappedPromise2 = transformApiPromise(apiPromise2, (rawStream) => {
           ensureAsyncIterable(rawStream, "responses.create");
-          const wrappedStream = wrapStreamForUsageTap(rawStream, async () => {
-            const usage = await extractUsageFromStream(rawStream, ctx.begin.data.vendorHints);
-            if (usage) {
-              ctx.setUsage(usage);
+          const wrappedStream = wrapStreamForUsageTap(rawStream, async (termination) => {
+            try {
+              if (termination === "complete") {
+                const usage = await extractUsageFromStream(rawStream, ctx.begin.data.vendorHints, provider);
+                if (usage) {
+                  ctx.setUsage(usage);
+                }
+              }
+            } catch (error) {
+              ctx.setError({
+                code: "USAGE_FINALIZE_ERROR",
+                message: error instanceof Error ? error.message : String(error)
+              });
+              throw error;
+            } finally {
+              await settle?.();
             }
-          }, ctx);
+          }, ctx, (chunk) => {
+            tryInferUsageFromStreamChunk(
+              chunk,
+              ctx.begin.data.vendorHints,
+              ctx,
+              provider
+            );
+          });
           return wrappedStream;
         });
         return wrappedPromise2;
       }
       const apiPromise = originalCreate(finalParams, request);
-      const wrappedPromise = transformApiPromise(apiPromise, (response) => {
-        tryInferUsage(response, ctx.begin.data.vendorHints, void 0, ctx);
+      const wrappedPromise = transformApiPromise(apiPromise, async (response) => {
+        tryInferUsage(response, ctx.begin.data.vendorHints, void 0, ctx, provider);
+        await captureMeteredOpenAISample({
+          usageTap,
+          sampling: defaultSampling,
+          decision: sampleDecision,
+          ctx,
+          beginRequest,
+          input: params,
+          response,
+          startedAt: sampleStartedAt
+        });
         return response;
+      }, async (error) => {
+        await captureMeteredOpenAISample({
+          usageTap,
+          sampling: defaultSampling,
+          decision: sampleDecision,
+          ctx,
+          beginRequest,
+          input: params,
+          error,
+          startedAt: sampleStartedAt
+        });
+        throw error;
       });
       return wrappedPromise;
     }, withUsage);
@@ -2049,7 +2945,7 @@ function createResponsesProxy(resource, usageTap, defaultContext, applyVendorHin
   };
   return new Proxy(resource, handler);
 }
-function createChatCompletionsProxy(resource, usageTap, defaultContext, applyVendorHints, defaultPromptCompression, promptCompressionStats) {
+function createChatCompletionsProxy(resource, usageTap, defaultContext, applyVendorHints, defaultPromptCompression, promptCompressionStats, defaultSampling, provider) {
   const originalCreate = resource.create.bind(resource);
   const streamCandidate = resource.stream;
   const originalStream = typeof streamCandidate === "function" ? streamCandidate.bind(resource) : void 0;
@@ -2063,8 +2959,16 @@ function createChatCompletionsProxy(resource, usageTap, defaultContext, applyVen
     const beginRequest = resolveBeginRequest(defaultContext, usageContext);
     const wantsStream = isStreamingRequest(params);
     return usageTap.withUsage(beginRequest, async (ctx) => {
+      const settle = wantsStream ? deferUsageFinalization(ctx) : void 0;
+      const sampleStartedAt = Date.now();
+      const sampleDecision = wantsStream ? Promise.resolve(false) : startMeteredOpenAISampleDecision({
+        usageTap,
+        sampling: defaultSampling,
+        beginRequest,
+        input: params
+      });
       const hintedParams = applyVendorHints ? applyChatVendorHints(params, ctx.begin.data.vendorHints) : params;
-      const finalParams = await compressChatParamsForCall({
+      const compressedParams = await compressChatParamsForCall({
         params: hintedParams,
         usageTap,
         ctx,
@@ -2074,25 +2978,68 @@ function createChatCompletionsProxy(resource, usageTap, defaultContext, applyVen
         withUsage,
         operation: "chat.completions.create"
       });
+      const finalParams = wantsStream ? ensureOpenAIStreamUsage(compressedParams) : compressedParams;
+      ctx.setUsage(openAIRequestExecutionMetadata(finalParams, provider));
       const request = attachCorrelationHeader(requestOptions, ctx.begin.correlationId);
       if (wantsStream) {
         const apiPromise2 = originalCreate(finalParams, request);
         const wrappedPromise2 = transformApiPromise(apiPromise2, (rawStream) => {
           ensureAsyncIterable(rawStream, "chat.completions.create");
-          const wrappedStream2 = wrapStreamForUsageTap(rawStream, async () => {
-            const usage = await extractUsageFromStream(rawStream, ctx.begin.data.vendorHints);
-            if (usage) {
-              ctx.setUsage(usage);
+          const wrappedStream2 = wrapStreamForUsageTap(rawStream, async (termination) => {
+            try {
+              if (termination === "complete") {
+                const usage = await extractUsageFromStream(rawStream, ctx.begin.data.vendorHints, provider);
+                if (usage) {
+                  ctx.setUsage(usage);
+                }
+              }
+            } catch (error) {
+              ctx.setError({
+                code: "USAGE_FINALIZE_ERROR",
+                message: error instanceof Error ? error.message : String(error)
+              });
+              throw error;
+            } finally {
+              await settle?.();
             }
-          }, ctx);
+          }, ctx, (chunk) => {
+            tryInferUsageFromStreamChunk(
+              chunk,
+              ctx.begin.data.vendorHints,
+              ctx,
+              provider
+            );
+          });
           return wrappedStream2;
         });
         return wrappedPromise2;
       }
       const apiPromise = originalCreate(finalParams, request);
-      const wrappedPromise = transformApiPromise(apiPromise, (response) => {
-        tryInferUsage(response, ctx.begin.data.vendorHints, void 0, ctx);
+      const wrappedPromise = transformApiPromise(apiPromise, async (response) => {
+        tryInferUsage(response, ctx.begin.data.vendorHints, void 0, ctx, provider);
+        await captureMeteredOpenAISample({
+          usageTap,
+          sampling: defaultSampling,
+          decision: sampleDecision,
+          ctx,
+          beginRequest,
+          input: params,
+          response,
+          startedAt: sampleStartedAt
+        });
         return response;
+      }, async (error) => {
+        await captureMeteredOpenAISample({
+          usageTap,
+          sampling: defaultSampling,
+          decision: sampleDecision,
+          ctx,
+          beginRequest,
+          input: params,
+          error,
+          startedAt: sampleStartedAt
+        });
+        throw error;
       });
       return wrappedPromise;
     }, withUsage);
@@ -2106,8 +3053,9 @@ function createChatCompletionsProxy(resource, usageTap, defaultContext, applyVen
     } = splitUsageOptions(options);
     const beginRequest = resolveBeginRequest(defaultContext, usageContext);
     return usageTap.withUsage(beginRequest, async (ctx) => {
+      const settle = deferUsageFinalization(ctx);
       const hintedParams = applyVendorHints ? applyChatVendorHints(params, ctx.begin.data.vendorHints) : params;
-      const finalParams = await compressChatParamsForCall({
+      const compressedParams = await compressChatParamsForCall({
         params: hintedParams,
         usageTap,
         ctx,
@@ -2117,16 +3065,41 @@ function createChatCompletionsProxy(resource, usageTap, defaultContext, applyVen
         withUsage,
         operation: "chat.completions.stream"
       });
+      const finalParams = ensureOpenAIStreamUsage(compressedParams);
+      ctx.setUsage(openAIRequestExecutionMetadata(finalParams, provider));
       const request = attachCorrelationHeader(requestOptions, ctx.begin.correlationId);
       const apiPromise = originalStream(finalParams, request);
       const wrappedPromise = transformApiPromise(apiPromise, (rawStream) => {
         ensureAsyncIterable(rawStream, "chat.completions.stream");
-        const wrappedStreamInner = wrapStreamForUsageTap(rawStream, async () => {
-          const usage = await extractUsageFromStream(rawStream, ctx.begin.data.vendorHints);
-          if (usage) {
-            ctx.setUsage(usage);
+        const wrappedStreamInner = wrapStreamForUsageTap(rawStream, async (termination) => {
+          try {
+            if (termination === "complete") {
+              const usage = await extractUsageFromStream(
+                rawStream,
+                ctx.begin.data.vendorHints,
+                provider
+              );
+              if (usage) {
+                ctx.setUsage(usage);
+              }
+            }
+          } catch (error) {
+            ctx.setError({
+              code: "USAGE_FINALIZE_ERROR",
+              message: error instanceof Error ? error.message : String(error)
+            });
+            throw error;
+          } finally {
+            await settle();
           }
-        }, ctx);
+        }, ctx, (chunk) => {
+          tryInferUsageFromStreamChunk(
+            chunk,
+            ctx.begin.data.vendorHints,
+            ctx,
+            provider
+          );
+        });
         return wrappedStreamInner;
       });
       return wrappedPromise;
@@ -2248,6 +3221,10 @@ async function compressChatParams(params, usageTap, compression, signal) {
     const result = await usageTap.compressPromptMessages(source, {
       provider: "usagetap",
       failOpen: compression.failOpen,
+      mode: compression.mode,
+      latencyBudgetMs: compression.latencyBudgetMs,
+      compactEmptyUserMessages: compression.compactEmptyUserMessages,
+      compactDuplicateUserTextParts: compression.compactDuplicateUserTextParts,
       aggressiveness: resolveMessageEndpointAggressiveness(compression),
       signal
     });
@@ -2717,19 +3694,52 @@ function resolveBeginRequest(defaults, override) {
   if (requested) begin.requested = requested;
   const feature = current.feature ?? base.feature;
   if (feature) begin.feature = feature;
+  const runId = current.runId ?? base.runId;
+  if (runId) begin.runId = runId;
   const idempotency = current.idempotency ?? base.idempotency;
   if (idempotency) begin.idempotency = idempotency;
   const customerName = current.customerName ?? base.customerName;
   if (customerName) begin.customerName = customerName;
   const customerEmail = current.customerEmail ?? base.customerEmail;
   if (customerEmail) begin.customerEmail = customerEmail;
+  const customerUserId = current.customerUserId ?? base.customerUserId;
+  if (customerUserId) begin.customerUserId = customerUserId;
+  const customerUserName = current.customerUserName ?? base.customerUserName;
+  if (customerUserName) begin.customerUserName = customerUserName;
+  const customerUserEmail = current.customerUserEmail ?? base.customerUserEmail;
+  if (customerUserEmail) begin.customerUserEmail = customerUserEmail;
+  const stripeCustomerId = current.stripeCustomerId ?? base.stripeCustomerId;
+  if (stripeCustomerId) begin.stripeCustomerId = stripeCustomerId;
+  const batch = current.batch ?? base.batch;
+  if (typeof batch === "boolean") begin.batch = batch;
+  const pricingMode = current.pricingMode ?? base.pricingMode;
+  if (pricingMode) begin.pricingMode = pricingMode;
   if (tags?.length) {
     begin.tags = tags;
   }
   return begin;
 }
-function transformApiPromise(apiPromise, onResolve) {
-  const resolvedPromise = Promise.resolve(apiPromise).then(onResolve);
+function responsesBeginRequest(begin, params) {
+  if (!responsesRequestUsesWebSearch(params)) return begin;
+  return {
+    ...begin,
+    requested: {
+      ...begin.requested ?? {},
+      search: true
+    }
+  };
+}
+function responsesRequestUsesWebSearch(params) {
+  if (!params || typeof params !== "object") return false;
+  const tools = params.tools;
+  return Array.isArray(tools) && tools.some((tool) => {
+    if (!tool || typeof tool !== "object") return false;
+    const type = tool.type;
+    return type === "web_search" || type === "web_search_preview";
+  });
+}
+function transformApiPromise(apiPromise, onResolve, onReject) {
+  const resolvedPromise = Promise.resolve(apiPromise).then(onResolve, onReject);
   if (isObjectRecord(apiPromise)) {
     const proto = Object.getPrototypeOf(apiPromise);
     if (proto) {
@@ -2864,12 +3874,12 @@ function applyResponsesVendorHints(params, hints) {
   }
   return next;
 }
-async function extractUsageFromStream(stream, hints) {
+async function extractUsageFromStream(stream, hints, provider = "openai") {
   const finalPayload = await resolveStreamFinalPayload(stream);
   if (!finalPayload) {
     return void 0;
   }
-  return inferUsageFromResponse(finalPayload, hints);
+  return inferUsageFromResponse(finalPayload, hints, provider);
 }
 async function resolveStreamFinalPayload(stream) {
   if (!stream || typeof stream !== "object") {
@@ -2942,14 +3952,14 @@ function setHeaderIfPossible(res, key, value) {
     res.setHeader(key, value);
   }
 }
-function tryInferUsage(response, hints, extractor, ctx) {
+function tryInferUsage(response, hints, extractor, ctx, provider = "openai") {
   const explicit = extractor?.(response);
-  const inferred = explicit ?? inferUsageFromResponse(response, hints);
+  const inferred = explicit ?? inferUsageFromResponse(response, hints, provider);
   if (inferred) {
     ctx.setUsage(inferred);
   }
 }
-function inferUsageFromResponse(response, hints) {
+function inferUsageFromResponse(response, hints, provider = "openai") {
   if (!response || typeof response !== "object") {
     return void 0;
   }
@@ -2957,32 +3967,110 @@ function inferUsageFromResponse(response, hints) {
   if (!candidate.usage) {
     return void 0;
   }
-  const cachedInputTokens = candidate.usage.prompt_tokens_details?.cached_tokens ?? candidate.usage.cached_tokens;
+  const cachedInputTokens = candidate.usage.prompt_tokens_details?.cached_tokens ?? candidate.usage.input_tokens_details?.cached_tokens ?? candidate.usage.cached_tokens;
+  const cacheWriteInputTokens = candidate.usage.prompt_tokens_details?.cache_write_tokens ?? candidate.usage.prompt_tokens_details?.cache_creation_tokens ?? candidate.usage.input_tokens_details?.cache_write_tokens ?? candidate.usage.input_tokens_details?.cache_creation_tokens ?? candidate.usage.cache_creation_input_tokens ?? candidate.usage.cache_write_input_tokens ?? candidate.usage.cache_write_tokens;
+  const cacheWrite5mInputTokens = candidate.usage.cache_write_5m_input_tokens ?? candidate.usage.cache_creation?.ephemeral_5m_input_tokens;
+  const cacheWrite1hInputTokens = candidate.usage.cache_write_1h_input_tokens ?? candidate.usage.cache_creation?.ephemeral_1h_input_tokens;
+  const outputSearches = Array.isArray(candidate.output) ? candidate.output.filter((item) => item?.type === "web_search_call").length : 0;
+  const searches = candidate.usage.searches ?? candidate.usage.web_search_queries ?? candidate.usage.server_tool_use?.web_search_requests ?? outputSearches;
+  const responseEffort = normalizeExecutionReasoningEffort(
+    candidate.reasoning?.effort
+  );
   return {
+    providerUsed: provider,
     modelUsed: candidate.model ?? hints?.preferredModel,
-    inputTokens: candidate.usage.prompt_tokens,
-    responseTokens: candidate.usage.completion_tokens,
-    cachedInputTokens
+    inputTokens: candidate.usage.prompt_tokens ?? candidate.usage.input_tokens,
+    responseTokens: candidate.usage.completion_tokens ?? candidate.usage.output_tokens,
+    cachedInputTokens,
+    cacheWriteInputTokens,
+    cacheWrite5mInputTokens,
+    cacheWrite1hInputTokens,
+    audioInputTokens: candidate.usage.prompt_tokens_details?.audio_tokens ?? candidate.usage.input_tokens_details?.audio_tokens,
+    cachedAudioInputTokens: candidate.usage.prompt_tokens_details?.cached_audio_tokens ?? candidate.usage.input_tokens_details?.cached_audio_tokens ?? candidate.usage.prompt_tokens_details?.cached_tokens_details?.audio_tokens ?? candidate.usage.input_tokens_details?.cached_tokens_details?.audio_tokens,
+    imageInputTokens: candidate.usage.prompt_tokens_details?.image_tokens ?? candidate.usage.input_tokens_details?.image_tokens,
+    imageOutputTokens: candidate.usage.completion_tokens_details?.image_tokens ?? candidate.usage.output_tokens_details?.image_tokens,
+    audioOutputTokens: candidate.usage.completion_tokens_details?.audio_tokens ?? candidate.usage.output_tokens_details?.audio_tokens,
+    reasoningTokens: candidate.usage.completion_tokens_details?.reasoning_tokens ?? candidate.usage.output_tokens_details?.reasoning_tokens,
+    ...responseEffort ? {
+      reasoningEffort: responseEffort,
+      reasoningEffortSource: "provider_response"
+    } : {},
+    ...typeof candidate.reasoning?.type === "string" ? { reasoningMode: candidate.reasoning.type } : typeof candidate.reasoning?.mode === "string" ? { reasoningMode: candidate.reasoning.mode } : {},
+    ...typeof searches === "number" && searches > 0 ? { searches } : {}
   };
 }
-function wrapStreamForUsageTap(source, finalize, ctx) {
+function normalizeExecutionReasoningEffort(value) {
+  return value === "none" || value === "minimal" || value === "low" || value === "medium" || value === "high" || value === "xhigh" || value === "max" ? value : void 0;
+}
+function openAIRequestExecutionMetadata(params, provider) {
+  const record = params && typeof params === "object" ? params : {};
+  const reasoning = record.reasoning && typeof record.reasoning === "object" ? record.reasoning : void 0;
+  const effort = normalizeExecutionReasoningEffort(
+    record.reasoning_effort ?? reasoning?.effort ?? record.thinking_level
+  );
+  const mode = typeof reasoning?.type === "string" ? reasoning.type : typeof reasoning?.mode === "string" ? reasoning.mode : void 0;
+  const rawBudget = reasoning?.budget_tokens ?? record.thinking_budget ?? record.thinking_budget_tokens;
+  const budget = typeof rawBudget === "number" && Number.isInteger(rawBudget) && rawBudget >= 0 ? rawBudget : void 0;
+  return {
+    providerUsed: provider,
+    ...typeof record.model === "string" ? { modelUsed: record.model } : {},
+    ...effort ? { reasoningEffort: effort, reasoningEffortSource: "provider_request" } : {},
+    ...mode ? { reasoningMode: mode } : {},
+    ...budget !== void 0 ? { reasoningBudgetTokens: budget } : {}
+  };
+}
+function tryInferUsageFromStreamChunk(chunk, hints, ctx, provider) {
+  const payload = isObjectRecord(chunk) && isObjectRecord(chunk.response) ? chunk.response : chunk;
+  const inferred = inferUsageFromResponse(payload, hints, provider);
+  if (inferred) {
+    ctx.setUsage(inferred);
+  }
+}
+function ensureOpenAIStreamUsage(params) {
+  if (!isObjectRecord(params)) {
+    return params;
+  }
+  const streamOptions = isObjectRecord(params.stream_options) ? params.stream_options : {};
+  return {
+    ...params,
+    stream_options: {
+      ...streamOptions,
+      include_usage: true
+    }
+  };
+}
+function deferUsageFinalization(ctx) {
+  return ctx.deferFinalization?.() ?? (() => Promise.resolve());
+}
+function wrapStreamForUsageTap(source, finalize, ctx, onChunk) {
   const getIterator = source[Symbol.asyncIterator];
   if (typeof getIterator !== "function") {
     throw new TypeError("Stream is not async iterable");
   }
   const iterator = getIterator.call(source);
   let completed = false;
-  const invokeFinalize = async () => {
+  const invokeFinalize = async (termination, error) => {
     if (completed) return;
     completed = true;
+    if (termination === "cancel" || termination === "manual") {
+      ctx.setError({
+        code: "STREAM_ABORTED",
+        message: termination === "cancel" ? "Provider stream consumption was cancelled before completion" : "Provider stream was finalized before completion"
+      });
+    } else if (termination === "error") {
+      ctx.setError({
+        code: "VENDOR_ERROR",
+        message: error instanceof Error ? error.message : String(error)
+      });
+    }
     try {
-      await finalize();
-    } catch (error) {
+      await finalize(termination);
+    } catch (error2) {
       ctx.setError({
         code: "USAGE_FINALIZE_ERROR",
-        message: error instanceof Error ? error.message : String(error)
+        message: error2 instanceof Error ? error2.message : String(error2)
       });
-      throw error;
+      throw error2;
     }
   };
   const prototype = Object.getPrototypeOf(source) ?? Object.prototype;
@@ -3006,12 +4094,15 @@ function wrapStreamForUsageTap(source, finalize, ctx) {
     value: async (...args) => {
       try {
         const result = await iterator.next(...args);
+        if (!result.done) {
+          onChunk?.(result.value);
+        }
         if (result.done) {
-          await invokeFinalize();
+          await invokeFinalize("complete");
         }
         return result;
       } catch (error) {
-        await invokeFinalize().catch(() => void 0);
+        await invokeFinalize("error", error).catch(() => void 0);
         throw error;
       }
     },
@@ -3020,39 +4111,49 @@ function wrapStreamForUsageTap(source, finalize, ctx) {
   });
   Object.defineProperty(wrapped, "return", {
     value: async (value) => {
-      if (typeof iterator.return === "function") {
-        const rawResult = await iterator.return(value);
-        if (!isIteratorResult(rawResult)) {
-          throw new TypeError("Iterator.return() returned an invalid result");
+      try {
+        if (typeof iterator.return === "function") {
+          const rawResult = await iterator.return(value);
+          if (!isIteratorResult(rawResult)) {
+            throw new TypeError("Iterator.return() returned an invalid result");
+          }
+          await invokeFinalize("cancel");
+          return rawResult;
         }
-        await invokeFinalize();
-        return rawResult;
+        await invokeFinalize("cancel");
+        return { done: true, value };
+      } catch (error) {
+        await invokeFinalize("error", error).catch(() => void 0);
+        throw error;
       }
-      await invokeFinalize();
-      return { done: true, value };
     },
     configurable: true,
     writable: true
   });
   Object.defineProperty(wrapped, "throw", {
     value: async (error) => {
-      if (typeof iterator.throw === "function") {
-        const rawResult = await iterator.throw(error);
-        if (!isIteratorResult(rawResult)) {
-          throw new TypeError("Iterator.throw() returned an invalid result");
+      try {
+        if (typeof iterator.throw === "function") {
+          const rawResult = await iterator.throw(error);
+          if (!isIteratorResult(rawResult)) {
+            throw new TypeError("Iterator.throw() returned an invalid result");
+          }
+          await invokeFinalize("error", error);
+          return rawResult;
         }
-        await invokeFinalize();
-        return rawResult;
+        await invokeFinalize("error", error);
+        throw error;
+      } catch (thrownError) {
+        await invokeFinalize("error", thrownError).catch(() => void 0);
+        throw thrownError;
       }
-      await invokeFinalize();
-      throw error;
     },
     configurable: true,
     writable: true
   });
   Object.defineProperty(wrapped, "__usageTapFinalize", {
     value: async () => {
-      await invokeFinalize();
+      await invokeFinalize("manual");
     },
     configurable: true
   });
@@ -3069,6 +4170,7 @@ exports.streamOpenAIRoute = streamOpenAIRoute;
 exports.toNextResponse = toNextResponse;
 exports.withCompression = withCompression;
 exports.withMetering = withMetering;
+exports.withSampling = withSampling;
 exports.wrapOpenAI = wrapOpenAI;
 //# sourceMappingURL=openai.cjs.map
 //# sourceMappingURL=openai.cjs.map