@broberg/ai-sdk 0.48.0 → 0.49.0

package/dist/index.d.ts

@@ -713,6 +713,19 @@ interface ClassifyInput {
     labels: string[];
     tier?: Tier;
     purpose?: string;
+    /** What to do when the model's reply contains no JSON at all (F059).
+     *
+     *  `"throw"` (the DEFAULT, and unchanged) is right in product use: a refusal or an
+     *  outage is not a classification, and a throw is the loudest honest answer.
+     *
+     *  `"value"` is for MEASURING. Requested by trail with the measurement behind it:
+     *  over 444 golden examples in one batch, a throw at example 212 is not informative,
+     *  it is destructive — the 232 that were never measured afterwards look like they
+     *  did not exist. Their alternative was to wrap every call in try/catch and count
+     *  the throws, which is this field built by hand, worse.
+     *
+     *  Opt-in on purpose: the caller who sets it is the caller who is reading for it. */
+    onUnparseable?: "throw" | "value";
 }
 interface ClassifyResult {
     /** The chosen label, or `null` when the model named a label that is not in `labels`
@@ -728,15 +741,55 @@ interface ClassifyResult {
      *  an autonomy level off this field — an unclassifiable ticket therefore landed on the
      *  tenant's first intent, chosen by the order of a config array, with no error and no
      *  log trace. `null` is the whole fix: there is no field to hardcode, because the
-     *  absence IS the signal. */
+     *  absence IS the signal.
+     *
+     *  **Expect `null` routinely since F060 (23 September 2026)** — the prompt now
+     *  explicitly lets the model say none of the labels fit. Before that it could not,
+     *  and trail measured the cost: 34 of 38 honest refusals came back as confident
+     *  WRONG labels. How OFTEN `null` now occurs under this exact prompt is not yet
+     *  measured (F060.2) — expect it to be common, not rare. If you route anything
+     *  automatic off this field, send `null` to a human. */
     label: string | null;
     /** The model's own answer when it did not match, so a caller can log or route what
-     *  actually came back instead of only knowing that something did not. */
+     *  actually came back instead of only knowing that something did not.
+     *
+     *  **It can contain RAW model output** — on the `unparseable` path it always does
+     *  (the first 200 characters of the reply), and on `out-of-set` it does whenever the
+     *  reply had no usable `label` string. A model that echoes part of your prompt can
+     *  therefore put part of YOUR INPUT here. On a path carrying personal or health data,
+     *  treat this field with the same care as the input before you log it. */
     rawLabel?: string;
     /** `null` when the model reported no confidence. `0` is a REAL confidence and stays
      *  `0` — the two used to be the same number, which made the field unusable as a
      *  signal even for a caller who wanted to check it. */
     confidence: number | null;
+    /** WHICH of the three things happened, as a value you must read rather than a shape
+     *  you might infer (F059).
+     *
+     *  `"answered"`     the model named a label from `labels`; `label` is it.
+     *  `"out-of-set"`   it named something else; `label` is null, `rawLabel` is its answer.
+     *  `"unparseable"`  we could not read the reply. Only reachable with
+     *                   `onUnparseable: "value"` — the default still throws. It covers
+     *                   BOTH ways parsing fails: no JSON in the reply at all, AND JSON
+     *                   that is present but malformed or truncated. Said explicitly
+     *                   because you are going to COUNT this, and a count that quietly
+     *                   includes a category the docs deny is a wrong number that reads
+     *                   as a right one.
+     *
+     *  It is NOT decoration. Once the throw is optional, `out-of-set` and `unparseable`
+     *  both yield `label: null` with `rawLabel` set, so without this field they cannot be
+     *  told apart — and telling them apart ("got it wrong" vs "did not answer" vs "could
+     *  not be read") is the whole reason the flag was asked for.
+     *
+     *  The form is components' argument, not ours: a boolean like `fallbackUsed` can be
+     *  destructured away as easily as a `confidence` field can be ignored, while a value
+     *  you must read to proceed cannot. `answered` ⟺ `label !== null`.
+     *
+     *  REQUIRED, not optional — additive for anyone READING a result, a compile fix for
+     *  anyone CONSTRUCTING one (a test stub or mock of `classify`). Saying it rather than
+     *  calling the change "purely additive": this package has shipped that exact
+     *  over-claim before, about `toolCall.arguments`, and it was wrong then. */
+    outcome: "answered" | "out-of-set" | "unparseable";
     usage: Usage;
 }
 interface RerankInput {
@@ -2268,7 +2321,6 @@ interface AiClient {
 
 declare function createAI(config?: AiConfig): AiClient;
 
-/** Pull the first JSON value out of a model reply (tolerates ```json fences + prose). */
 declare function parseJsonLoose(text: string): unknown;
 type ChatVision = Pick<AiClient, "chat" | "vision">;
 /** F052.2 — resolve the model's answer to one of the CALLER's labels, or to null.
@@ -2552,8 +2604,8 @@ declare const falStubAdapter: ProviderAdapter;
  *  wires the live adapters. */
 declare const stubProviders: Record<string, ProviderAdapter>;
 
-declare const VERSION: "0.48.0";
-declare const SDK_TAG: "@broberg/ai-sdk@0.48.0";
+declare const VERSION: "0.49.0";
+declare const SDK_TAG: "@broberg/ai-sdk@0.49.0";
 
 /** Built-in defaults. Every entry is overridable via AiConfig.defaults or a
  *  per-call override.
@@ -2758,10 +2810,46 @@ interface UpmetricsSinkConfig {
     complianceMode?: boolean;
     /** Injectable fetch for testing; defaults to global fetch. */
     fetch?: typeof fetch;
-    /** Optional error hook (errors are otherwise swallowed silently). */
+    /** Called when a record is actually LOST (dropped) or REFUSED (rejected) — not on a
+     *  transient failure that is still being retried. See F061 below. */
     onError?: (err: unknown) => void;
-}
-declare function upmetricsSink(config: UpmetricsSinkConfig): CostSink;
+    /** Retry transient failures in the background (F061). Default `true`.
+     *
+     *  SAFE ONLY IF THIS SINK'S RECEIVER DEDUPLICATES on `tags.idempotencyKey`. That is
+     *  a property of your configuration, not of this package: upmetrics does (live since
+     *  2026-09-08, measured by them — two deliveries of one payload → one row). If you
+     *  point `baseUrl` at something that does not, a retry trades a loss for a double
+     *  count there — set `retry: false`. We do not guess from the hostname: a sink that
+     *  takes a baseUrl cannot have its behaviour decided by a name. */
+    retry?: boolean;
+    /** Base backoff in ms (doubles per attempt, capped at 30 s). For tests. */
+    retryBaseMs?: number;
+}
+/** What the sink has done with what it was given — "we lost nothing" and "we do not
+ *  know whether we lost anything" are different statements, and only a count can make
+ *  the first one. In-process: it resets on deploy, so a zero is a claim about uptime,
+ *  not about history (trail's caveat, and it is right). */
+interface UpmetricsSinkStats {
+    /** Accepted by the receiver. */
+    sent: number;
+    /** Retry ATTEMPTS made (not records). */
+    retried: number;
+    /** Given up on: attempts exhausted, or pushed out of a full queue. LOST. */
+    dropped: number;
+    /** Permanently refused by the receiver (4xx other than 408/429). Not retried. */
+    rejected: number;
+    /** Waiting for a retry right now, including one in flight. */
+    queued: number;
+}
+interface UpmetricsSink extends CostSink {
+    /** Try everything waiting, once, now. Call before exit in a SHORT-LIVED process
+     *  (a script, a serverless function): the retry timer is unref'd so it never holds
+     *  a process open, which also means it will not finish on its own before exit.
+     *  Whatever still fails stays queued — check `stats().queued` afterwards. */
+    flush(): Promise<void>;
+    stats(): UpmetricsSinkStats;
+}
+declare function upmetricsSink(config: UpmetricsSinkConfig): UpmetricsSink;
 
 interface DiscordSinkConfig {
     webhookUrl: string;
@@ -3027,4 +3115,4 @@ interface StreamTransportRequest extends TransportRequest {
  */
 declare function streamTransport(req: StreamTransportRequest): AsyncIterable<string>;
 
-export { AZURE_DANISH_VOICES, AZURE_DANISH_VOICE_LIST, type AiClient, type AiConfig, type AlignedWordTimings, type AzureVoiceInfo, type AzureWordBoundary, type BatchJob, type BatchRequestItem, type BatchResultItem, type BflAdapterConfig, type BflCredits, type BudgetConfig, BudgetExceededError, BudgetGuard, type BudgetStore, CONFIG_DERIVED_HOSTS, type CallOptions, type Capability, type ChatInput, type ChatRequest, type ChatResult, type ChatStreamEvent, type CheckVoiceOptions, type ClassifyInput, type ClassifyResult, type ContentPart, type Contracts, type CostQuery, type CostSink, type CostSummary, type CostSummaryQuery, type CostTimeseriesQuery, DEFAULT_BASE_URLS, DEFAULT_TIER_MAP, type DesignInput, type DesignResult, type DialogueRequest, type DialogueTurn, type DiscordSinkConfig, ELEVENLABS_DANISH_VOICES, type EmbeddingInput, type EmbeddingRequest, type EmbeddingResult, type ExtractInput, type ExtractResult, type FalAdapterConfig, type FixedHostProvider, type HttpResponse, type ImageInput, type ImageRequest, type ImageResult, LOCALLY_PINNED_HOSTS, type LoraWeight, type Message, type MockupInput, type MockupResult, type ModerationInput, type ModerationItem, type ModerationRequest, type ModerationResult, type OcrInput, type OcrPage, type OcrRequest, type OcrResult, type OpenAICompatibleConfig, type PodcastInput, type PodcastResult, type PricingEntry, type ProviderAdapter, type RefreshOptions, type RefreshResult, type Region, type RerankInput, type RerankResult, type Role, type RouteForecast, SDK_TAG, type SqliteBudgetStoreConfig, type SqliteSinkConfig, StreamHttpError, type SubprocessResponse, type Tier, type TierSpec, type Tool, type ToolCall, type TrainStyleInput, type TrainStyleRequest, type TrainStyleResult, type TranscribeInput, type TranscribeRequest, type TranscribeResult, type TranslateInput, type TranslateResult, type Transport, type TransportRequest, type TransportResponse, type TtsInput, type TtsRequest, type UpmetricsCostClientConfig, UpmetricsCostError, type UpmetricsCostRow, type UpmetricsCostSummary, type UpmetricsCostTimeseries, type UpmetricsSinkConfig, type Usage, VERSION, type VideoInput, type VisionInput, type VoiceInfo, type VoiceProvider, type VoiceResolveResult, type VoiceStatus, VoiceUnavailableError, type WordTiming, aiConfigSchema, alignWordTimings, anthropicAdapter, anthropicApiAdapter, anthropicSubprocessAdapter, azureAdapter, bflAdapter, bflCredits, chatInputSchema, checkVoice, classifyRegionName, computeCost, createAI, deepinfraAdapter, deeplAdapter, deepseekAdapter, defaultBaseUrl, defaultProviders, discordSink, elevenlabsAdapter, embeddingInputSchema, falAdapter, falStubAdapter, freshUsage, fromProviderToolCall, geminiAdapter, getCostSummary, getPrice, httpTransport, imageInputSchema, listAzureDanishVoices, listVoices, makeContracts, makeOpenAICompatibleAdapter, matchLabel, messageSchema, mistralAdapter, mistralStubAdapter, multiSink, noopSink, openaiAdapter, openaiStubAdapter, openrouterAdapter, parseClaudeCliJson, parseJsonLoose, refreshAvailability, regionOfHost, regionOfProvider, requestyAdapter, resetRefreshState, resetRegistry, resolveAzureVoice, resolveTier, resolveVoice, sqliteBudgetStore, sqliteSink, streamTransport, stubProviders, subprocessTransport, tierSpecSchema, toProviderTools, toolSchema, translateInputSchema, upmetricsCostClient, upmetricsSink, usdFromMicro, vertexAdapter, visionInputSchema, wouldProviderRouteTo, wouldRouteTo };
+export { AZURE_DANISH_VOICES, AZURE_DANISH_VOICE_LIST, type AiClient, type AiConfig, type AlignedWordTimings, type AzureVoiceInfo, type AzureWordBoundary, type BatchJob, type BatchRequestItem, type BatchResultItem, type BflAdapterConfig, type BflCredits, type BudgetConfig, BudgetExceededError, BudgetGuard, type BudgetStore, CONFIG_DERIVED_HOSTS, type CallOptions, type Capability, type ChatInput, type ChatRequest, type ChatResult, type ChatStreamEvent, type CheckVoiceOptions, type ClassifyInput, type ClassifyResult, type ContentPart, type Contracts, type CostQuery, type CostSink, type CostSummary, type CostSummaryQuery, type CostTimeseriesQuery, DEFAULT_BASE_URLS, DEFAULT_TIER_MAP, type DesignInput, type DesignResult, type DialogueRequest, type DialogueTurn, type DiscordSinkConfig, ELEVENLABS_DANISH_VOICES, type EmbeddingInput, type EmbeddingRequest, type EmbeddingResult, type ExtractInput, type ExtractResult, type FalAdapterConfig, type FixedHostProvider, type HttpResponse, type ImageInput, type ImageRequest, type ImageResult, LOCALLY_PINNED_HOSTS, type LoraWeight, type Message, type MockupInput, type MockupResult, type ModerationInput, type ModerationItem, type ModerationRequest, type ModerationResult, type OcrInput, type OcrPage, type OcrRequest, type OcrResult, type OpenAICompatibleConfig, type PodcastInput, type PodcastResult, type PricingEntry, type ProviderAdapter, type RefreshOptions, type RefreshResult, type Region, type RerankInput, type RerankResult, type Role, type RouteForecast, SDK_TAG, type SqliteBudgetStoreConfig, type SqliteSinkConfig, StreamHttpError, type SubprocessResponse, type Tier, type TierSpec, type Tool, type ToolCall, type TrainStyleInput, type TrainStyleRequest, type TrainStyleResult, type TranscribeInput, type TranscribeRequest, type TranscribeResult, type TranslateInput, type TranslateResult, type Transport, type TransportRequest, type TransportResponse, type TtsInput, type TtsRequest, type UpmetricsCostClientConfig, UpmetricsCostError, type UpmetricsCostRow, type UpmetricsCostSummary, type UpmetricsCostTimeseries, type UpmetricsSink, type UpmetricsSinkConfig, type UpmetricsSinkStats, type Usage, VERSION, type VideoInput, type VisionInput, type VoiceInfo, type VoiceProvider, type VoiceResolveResult, type VoiceStatus, VoiceUnavailableError, type WordTiming, aiConfigSchema, alignWordTimings, anthropicAdapter, anthropicApiAdapter, anthropicSubprocessAdapter, azureAdapter, bflAdapter, bflCredits, chatInputSchema, checkVoice, classifyRegionName, computeCost, createAI, deepinfraAdapter, deeplAdapter, deepseekAdapter, defaultBaseUrl, defaultProviders, discordSink, elevenlabsAdapter, embeddingInputSchema, falAdapter, falStubAdapter, freshUsage, fromProviderToolCall, geminiAdapter, getCostSummary, getPrice, httpTransport, imageInputSchema, listAzureDanishVoices, listVoices, makeContracts, makeOpenAICompatibleAdapter, matchLabel, messageSchema, mistralAdapter, mistralStubAdapter, multiSink, noopSink, openaiAdapter, openaiStubAdapter, openrouterAdapter, parseClaudeCliJson, parseJsonLoose, refreshAvailability, regionOfHost, regionOfProvider, requestyAdapter, resetRefreshState, resetRegistry, resolveAzureVoice, resolveTier, resolveVoice, sqliteBudgetStore, sqliteSink, streamTransport, stubProviders, subprocessTransport, tierSpecSchema, toProviderTools, toolSchema, translateInputSchema, upmetricsCostClient, upmetricsSink, usdFromMicro, vertexAdapter, visionInputSchema, wouldProviderRouteTo, wouldRouteTo };

package/dist/index.js

@@ -2866,6 +2866,8 @@ async function resolveAudio(audio, fetchImpl = fetch) {
 }
 
 // src/capabilities/contracts/index.ts
+var RAW_SNIPPET_CHARS = 200;
+var CLASSIFY_ABSTAIN_SENTENCE = 'If none of the labels fit, return {"label": null}.';
 function parseJsonLoose(text) {
   const fenced = text.replace(/```(?:json)?/gi, "").trim();
   const start = fenced.search(/[[{]/);
@@ -2946,7 +2948,19 @@ Your previous output was not valid JSON. Return ONLY parseable JSON.` : base,
     },
     async classify(input) {
       const res = await client.chat({
-        system: 'You are a zero-shot classifier. Choose exactly one label from the provided list. Return ONLY JSON: {"label": "<one of the labels>", "confidence": <0..1>}.',
+        // F060 — the model must be ALLOWED to say "none of these". Without that sentence,
+        // measured by trail on 444 golden examples (mistral-small-latest, temp 0): of 38
+        // inputs the model refused when invited to, this prompt turned 34 into a
+        // confident WRONG label inside the menu and let none through as a refusal. That
+        // undid F052 from the outside — the code stopped guessing labels[0], and the
+        // prompt made the model guess instead.
+        //
+        // It is exactly ONE sentence, the one trail measured. They isolated it: removing
+        // only this sentence reproduces the failure (35 wrong vs our 34), so it is the
+        // invitation to refuse that matters, not the confidence field or the wording
+        // around it. Do not "tidy" it away, and do not change the words — an unmeasured
+        // edit on top of a measured one makes the measurement say nothing.
+        system: "You are a zero-shot classifier. Choose exactly one label from the provided list. " + CLASSIFY_ABSTAIN_SENTENCE + ' Return ONLY JSON: {"label": "<one of the labels>", "confidence": <0..1>}.',
         prompt: `Labels: ${JSON.stringify(input.labels)}
 
 Text:
@@ -2954,13 +2968,26 @@ ${input.text}`,
         tier: input.tier ?? "cheap",
         purpose: input.purpose ?? "contract:classify"
       });
-      const parsed = parseJsonLoose(res.text);
+      let parsed;
+      try {
+        parsed = parseJsonLoose(res.text);
+      } catch (err) {
+        if (input.onUnparseable !== "value") throw err;
+        return {
+          label: null,
+          rawLabel: res.text.slice(0, RAW_SNIPPET_CHARS),
+          confidence: null,
+          outcome: "unparseable",
+          usage: res.usage
+        };
+      }
       const matched = matchLabel(parsed.label, input.labels);
       return {
         label: matched,
-        ...matched !== null ? {} : { rawLabel: typeof parsed.label === "string" ? parsed.label : res.text.slice(0, 200) },
+        ...matched !== null ? {} : { rawLabel: typeof parsed.label === "string" ? parsed.label : res.text.slice(0, RAW_SNIPPET_CHARS) },
         // 0 is a real confidence; "no confidence reported" is not 0.
         confidence: typeof parsed.confidence === "number" ? parsed.confidence : null,
+        outcome: matched !== null ? "answered" : "out-of-set",
         usage: res.usage
       };
     },
@@ -2977,7 +3004,7 @@ ${JSON.stringify(input.items)}`,
       const raw = parseJsonLoose(res.text);
       if (!Array.isArray(raw)) {
         throw new Error(
-          `ai.contracts.rerank: the model did not return a JSON array. Got: ${res.text.slice(0, 200)}${res.text.length > 200 ? "\u2026" : ""}`
+          `ai.contracts.rerank: the model did not return a JSON array. Got: ${res.text.slice(0, RAW_SNIPPET_CHARS)}${res.text.length > RAW_SNIPPET_CHARS ? "\u2026" : ""}`
         );
       }
       const scored = /* @__PURE__ */ new Map();
@@ -3378,83 +3405,201 @@ var aiConfigSchema = z.object({
   availability: availabilitySchema.optional()
 });
 
+// src/cost/sinks/upmetrics.ts
+import { randomUUID } from "crypto";
+
 // src/version.ts
-var VERSION = "0.48.0";
-var SDK_TAG = "@broberg/ai-sdk@0.48.0";
+var VERSION = "0.49.0";
+var SDK_TAG = "@broberg/ai-sdk@0.49.0";
 
 // src/cost/sinks/upmetrics.ts
+function isRetryableStatus(status) {
+  return status === 408 || status === 429 || status >= 500;
+}
+var MAX_QUEUE = 30;
+var MAX_ATTEMPTS = 5;
+var MAX_BACKOFF_MS = 3e4;
 function upmetricsSink(config) {
   const doFetch = config.fetch ?? fetch;
   const url = `${config.baseUrl.replace(/\/$/, "")}/api/agent`;
+  const retry = config.retry ?? true;
+  const baseMs = config.retryBaseMs ?? 1e3;
+  const queue = [];
+  const counts = { sent: 0, retried: 0, dropped: 0, rejected: 0 };
+  let inFlight = 0;
+  let timer = null;
+  function buildBody(usage) {
+    const startedAt = usage.ts || (/* @__PURE__ */ new Date()).toISOString();
+    const endedAt = new Date(
+      new Date(startedAt).getTime() + (usage.latencyMs || 0)
+    ).toISOString();
+    const agentKind = config.agentKind ?? (usage.capability === "embedding" ? "embedding" : "chatbot");
+    const body = {
+      mode: "record",
+      agent_kind: agentKind,
+      agent_name: config.agentName,
+      provider: usage.provider,
+      model: usage.model,
+      status: "success",
+      input_tokens: usage.inputTokens,
+      output_tokens: usage.outputTokens,
+      cache_read_tokens: usage.cacheReadTokens,
+      cache_creation_tokens: usage.cacheCreationTokens,
+      cost_usd: usage.costUsd,
+      duration_ms: usage.latencyMs,
+      started_at: startedAt,
+      ended_at: endedAt,
+      tags: {
+        // Consumer attribution labels (e.g. tenantId) ride in tags so no new
+        // top-level field risks the strict-shape ingest schema (F011). The
+        // SDK-owned keys win — a label can never clobber capability/transport/sdk.
+        ...usage.labels,
+        capability: usage.capability,
+        transport: usage.transport,
+        // F042: data residency of the route that answered. Rides in tags like
+        // capability/transport — no ingest-schema change, and without it the one
+        // field built for auditability existed only in memory.
+        region: usage.region,
+        // F050: HOW cost_usd was arrived at. upmetrics already distinguishes
+        // reported / computed / unpriced — we were sending an assumed number in
+        // the same field as a measured one, so their labels could not be right
+        // about our rows however carefully they were applied.
+        cost_basis: usage.costBasis ?? "computed",
+        sdk: SDK_TAG,
+        // F061: the receiver's dedupe key. IDENTICAL across every retry of this record
+        // — the body is built once and the same string is resent — and different
+        // between records. Sent on the FIRST attempt too: a first attempt can reach the
+        // server and lose its answer, and the retry must then dedupe against it. Placed
+        // after the labels spread, so a consumer label cannot overwrite it.
+        idempotencyKey: randomUUID()
+      }
+    };
+    if (usage.tier !== void 0) body.tier = usage.tier;
+    if (usage.purpose !== void 0) body.purpose = usage.purpose;
+    if (usage.toolCalls) {
+      body.tool_calls = usage.toolCalls.map((t) => ({
+        name: t.name,
+        count: t.count,
+        error_count: t.errorCount ?? 0
+      }));
+    }
+    void config.complianceMode;
+    return JSON.stringify(body);
+  }
+  async function send(body) {
+    try {
+      const res = await doFetch(url, {
+        method: "POST",
+        headers: {
+          "content-type": "application/json",
+          "X-Upmetrics-Key": config.apiKey
+        },
+        body
+      });
+      if (res.ok) return { kind: "ok" };
+      const text = await res.text().catch(() => "");
+      const err = new Error(`upmetricsSink: ingest returned ${res.status}: ${text.slice(0, 200)}`);
+      return { kind: isRetryableStatus(res.status) ? "retry" : "reject", err };
+    } catch (err) {
+      return { kind: "retry", err };
+    }
+  }
+  function lose(err) {
+    counts.dropped += 1;
+    config.onError?.(err);
+  }
+  function enforceCap() {
+    while (queue.length > MAX_QUEUE) {
+      queue.shift();
+      lose(new Error(`upmetricsSink: retry queue full (${MAX_QUEUE}) \u2014 dropped the oldest pending record`));
+    }
+  }
+  function settle(p, r) {
+    if (r.kind === "ok") {
+      counts.sent += 1;
+      return false;
+    }
+    if (r.kind === "reject") {
+      counts.rejected += 1;
+      config.onError?.(r.err);
+      return false;
+    }
+    if (p.attempts >= MAX_ATTEMPTS) {
+      lose(r.err);
+      return false;
+    }
+    return true;
+  }
+  function schedule() {
+    if (timer || queue.length === 0) return;
+    const head = queue[0];
+    const delay = Math.min(baseMs * 2 ** (head.attempts - 1), MAX_BACKOFF_MS);
+    timer = setTimeout(() => {
+      timer = null;
+      void drainOne();
+    }, delay);
+    timer.unref?.();
+  }
+  async function drainOne() {
+    const p = queue.shift();
+    if (!p) return;
+    inFlight += 1;
+    p.attempts += 1;
+    counts.retried += 1;
+    const r = await send(p.body);
+    inFlight -= 1;
+    if (settle(p, r)) {
+      queue.unshift(p);
+      enforceCap();
+    }
+    schedule();
+  }
   return {
     async record(usage) {
       try {
-        const startedAt = usage.ts || (/* @__PURE__ */ new Date()).toISOString();
-        const endedAt = new Date(
-          new Date(startedAt).getTime() + (usage.latencyMs || 0)
-        ).toISOString();
-        const agentKind = config.agentKind ?? (usage.capability === "embedding" ? "embedding" : "chatbot");
-        const body = {
-          mode: "record",
-          agent_kind: agentKind,
-          agent_name: config.agentName,
-          provider: usage.provider,
-          model: usage.model,
-          status: "success",
-          input_tokens: usage.inputTokens,
-          output_tokens: usage.outputTokens,
-          cache_read_tokens: usage.cacheReadTokens,
-          cache_creation_tokens: usage.cacheCreationTokens,
-          cost_usd: usage.costUsd,
-          duration_ms: usage.latencyMs,
-          started_at: startedAt,
-          ended_at: endedAt,
-          tags: {
-            // Consumer attribution labels (e.g. tenantId) ride in tags so no new
-            // top-level field risks the strict-shape ingest schema (F011). The
-            // SDK-owned keys win — a label can never clobber capability/transport/sdk.
-            ...usage.labels,
-            capability: usage.capability,
-            transport: usage.transport,
-            // F042: data residency of the route that answered. Rides in tags like
-            // capability/transport — no ingest-schema change, and without it the one
-            // field built for auditability existed only in memory.
-            region: usage.region,
-            // F050: HOW cost_usd was arrived at. upmetrics already distinguishes
-            // reported / computed / unpriced — we were sending an assumed number in
-            // the same field as a measured one, so their labels could not be right
-            // about our rows however carefully they were applied.
-            cost_basis: usage.costBasis ?? "computed",
-            sdk: SDK_TAG
-          }
-        };
-        if (usage.tier !== void 0) body.tier = usage.tier;
-        if (usage.purpose !== void 0) body.purpose = usage.purpose;
-        if (usage.toolCalls) {
-          body.tool_calls = usage.toolCalls.map((t) => ({
-            name: t.name,
-            count: t.count,
-            error_count: t.errorCount ?? 0
-          }));
+        const body = buildBody(usage);
+        const r = await send(body);
+        if (r.kind === "ok") {
+          counts.sent += 1;
+          return;
         }
-        void config.complianceMode;
-        const res = await doFetch(url, {
-          method: "POST",
-          headers: {
-            "content-type": "application/json",
-            "X-Upmetrics-Key": config.apiKey
-          },
-          body: JSON.stringify(body)
-        });
-        if (!res.ok) {
-          const text = await res.text().catch(() => "");
-          config.onError?.(
-            new Error(`upmetricsSink: ingest returned ${res.status}: ${text.slice(0, 200)}`)
-          );
+        if (r.kind === "reject") {
+          counts.rejected += 1;
+          config.onError?.(r.err);
+          return;
         }
+        if (!retry) {
+          lose(r.err);
+          return;
+        }
+        queue.push({ body, attempts: 1 });
+        enforceCap();
+        schedule();
       } catch (err) {
         config.onError?.(err);
       }
+    },
+    async flush() {
+      if (timer) {
+        clearTimeout(timer);
+        timer = null;
+      }
+      const batch = queue.splice(0);
+      const stillPending = [];
+      for (const p of batch) {
+        inFlight += 1;
+        p.attempts += 1;
+        counts.retried += 1;
+        const r = await send(p.body);
+        inFlight -= 1;
+        if (settle(p, r)) stillPending.push(p);
+      }
+      queue.unshift(...stillPending);
+      enforceCap();
+      schedule();
+    },
+    stats() {
+      return { ...counts, queued: queue.length + inFlight };
     }
   };
 }