@bivelio/savings-layer 0.4.30 → 0.4.32

package/dist/index.d.ts

@@ -325,7 +325,7 @@ declare function precioDeLista(model: string, at?: Date | number): CostModel | u
  * on the customer's machine. `test/version-declarada.test.ts` ties this
  * constant to package.json so the two can never drift.
  */
-declare const BV_SALA_VERSION = "0.4.30";
+declare const BV_SALA_VERSION = "0.4.32";
 
 /**
  * Environment-agnostic UUID generation.
@@ -546,7 +546,11 @@ interface OutputContract {
     /** Stable identifier, e.g. "research-answer:v3". */
     contractId: string;
     schema: JsonSchema;
-    /** Hard ceiling on generated tokens. */
+    /**
+     * Hard ceiling on generated tokens for this contract. A free-text request (no
+     * schema) sets its ceiling with {@link SavingsRequest.maxOutputTokens}; when
+     * both are set, the smaller one is sent.
+     */
     maxOutputTokens?: number;
     /** Version marker included in the exact-cache key. */
     version?: string;
@@ -1103,17 +1107,58 @@ interface SavingsRequest {
     /** Sampling controls that participate in the exact-cache key. */
     temperature?: number;
     seed?: number;
+    /**
+     * How hard a reasoning model should think, sent as OpenAI's `reasoning_effort`
+     * by the OpenAI-compatible adapter (`"none"`, `"minimal"`, `"low"`,
+     * `"medium"`, `"high"`, `"xhigh"`; the value travels as written). Nothing is
+     * sent when absent: the layer never picks an effort for you.
+     *
+     * Required for the gpt-5.6 family with tools on Chat Completions: OpenAI
+     * rejects function tools there unless it is `"none"` («Function tools with
+     * reasoning_effort are not supported for gpt-5.6-sol in
+     * /v1/chat/completions … set reasoning_effort to 'none'»). Gemini's
+     * OpenAI-compatible endpoint maps it to its thinking level. A compatible
+     * endpoint that rejects the field by name gets the request once more without
+     * it (OpenAI's own API never: there its 400 is the answer). It is part of the
+     * cache identity, like `temperature` and `seed`. The Anthropic adapter
+     * ignores it.
+     */
+    reasoningEffort?: ReasoningEffort;
+    /**
+     * Hard ceiling on generated tokens for this request, with or without an
+     * output contract. It travels to every adapter: Anthropic `max_tokens`,
+     * Chat Completions `max_completion_tokens` on OpenAI's own API and
+     * `max_tokens` on any other OpenAI-compatible endpoint (see
+     * `OpenAICompatibleOptions.maxTokensParam`). When `response.maxOutputTokens`
+     * is also set, the smaller of the two is sent. Unset, nothing changes.
+     *
+     * It is a limit you set, not a saving: it never raises the estimated output
+     * of an avoided call and is never credited as output avoided. An answer cut
+     * by it (`finishReason: "length"`) is not cached. Must be a positive integer.
+     */
+    maxOutputTokens?: number;
     /** Tokenizer id to cost against; defaults to a model-derived estimate. */
     tokenizer?: string;
     /** Arbitrary structured payload to be serialized into the prompt. */
     structuredData?: unknown;
 }
 /** A tool invocation the model requested (the provider-neutral shape). */
+/** OpenAI's `reasoning_effort` values; any other string travels as written. */
+type ReasoningEffort = "none" | "minimal" | "low" | "medium" | "high" | "xhigh" | (string & {});
 interface ToolCallRequest {
     id?: string;
     name: string;
     /** Argumentos tal cual los serializó el modelo (JSON en crudo). */
     arguments: string;
+    /**
+     * Provider data attached to this call that must travel back, untouched, when
+     * the conversation returns it: Gemini 3's
+     * `extra_content.google.thought_signature` through its OpenAI-compatible
+     * endpoint, without which the next turn is rejected with a 400 («Function
+     * call is missing a thought_signature»). Opaque: pass it back as received in
+     * the assistant message's `toolCalls`. It is not part of any cache key.
+     */
+    extraContent?: Record<string, unknown>;
 }
 /** Result returned to the caller. */
 interface SavingsResult<T = unknown> {
@@ -1336,11 +1381,12 @@ interface QualityTruncationCheck {
     /** The stop reason exactly as the provider named it; absent when it reported none. */
     finishReason?: string;
     /**
-     * Whether you set the limit (`response.maxOutputTokens`). When false, any
+     * Whether you set the limit (`maxOutputTokens` on the request or on
+     * `response`). When false, any
      * limit that applied was a provider or adapter default.
      */
     capSetByCaller: boolean;
-    /** The limit you set, in output tokens. Present only when `capSetByCaller`. */
+    /** The limit that was sent (the smaller of the two when both were set). Present only when `capSetByCaller`. */
     outputLimit?: number;
 }
 /** How a cache hit matched this request. */
@@ -1902,6 +1948,8 @@ interface CanonicalRequest {
     contractId: string | null;
     temperature: number | null;
     seed: number | null;
+    /** `reasoning_effort` declared by the caller; absent when it declared none. */
+    reasoningEffort?: string;
     tenantId: string;
     authorizationScope: string;
     /**
@@ -2180,6 +2228,8 @@ interface SemanticRecord {
     temperature: number | null;
     /** Sampling seed of the request, or null when it set none. */
     seed: number | null;
+    /** `reasoning_effort` of the request; absent when it declared none. */
+    reasoningEffort?: string;
     storedAtMs: number;
     /**
      * With calibration on: the highest score at which this stored answer was
@@ -2240,8 +2290,13 @@ interface SemanticCacheOptions {
      *
      * Pass your own for true semantics; the local one only sees LEXICAL closeness.
      * Either way the deterministic guard below still rejects a candidate whose
-     * numbers, negation or entities differ, so a weaker vector can only cost you a
-     * cache miss — never a wrong answer.
+     * numbers (in digits or spelled out in en/es/ca: "three", "twice", "half"),
+     * negation parity or content words differ beyond its overlap floor. It is a
+     * lexical guard, not a proof: a one-word entity swap in a long query can
+     * still clear the overlap floor, and through 0.4.30 a number written in words
+     * cleared it too (GSM8K: 477 of 564 such variants got their original's
+     * answer). Calibrate on your own traffic before trusting it with answers
+     * that differ by one word.
      */
     embed?: EmbeddingFn;
     /**
@@ -2558,7 +2613,20 @@ interface ToolSelection {
     setIsStable: boolean;
 }
 interface ToolSelectorOptions {
-    /** Max tools to include (schema-token budget guard). */
+    /**
+     * Max tools the selector keeps when it has a SIGNAL to select on: a catalogue
+     * whose tools declare `intents` and/or `risk`, with the router on. Default 8.
+     *
+     * It never truncates a catalogue it cannot rank. A catalogue that declares no
+     * `intents` and no `risk` (the usual case: MCP and OpenAI tool lists) and every
+     * request with `savings: { router: false }` go to the provider WHOLE, exactly as
+     * the caller sent them, and the layer credits nothing for the tool block:
+     * cutting by position left the model unable to call the rest (on BFCL
+     * multi-turn, gpt-4o-mini hit the reference call 8/10 times directly and 1/10
+     * through the layer). When the selector does prune a typed catalogue, the
+     * report's `tool-selector` trace lists what was left out and the SDK warns
+     * (once per process, silenced by `silenceWarnings`).
+     */
     maxTools?: number;
     /** Tokenizer-aware scorer for tie-breaking (optional). */
     scoreIntentMatch?: (text: string, tool: ToolDescriptor) => number;
@@ -2604,16 +2672,9 @@ declare class ToolSelector {
     private readonly freezer;
     constructor(opts?: ToolSelectorOptions);
     /**
-     * Todas las herramientas que pasó el llamante, acotadas SOLO por el
-     * presupuesto de esquema (`maxTools`).
-     *
-     * Se usa cuando el llamante apaga el router (`savings:{router:false}`): sin
-     * clasificación no hay de dónde derivar el riesgo, y filtrar sobre el
-     * `defaultRoute()` tiraba en silencio las herramientas de escritura. Apagar el
-     * router significa "yo elijo las herramientas", así que la relevancia pasa a
-     * ser del llamante — pero el tope de presupuesto NO se levanta: mandar 200
-     * esquemas al proveedor es un 400 seguro y una petición que antes funcionaba
-     * pasaría a fallar en duro.
+     * Every tool the caller passed, as passed: none left out, none credited.
+     * Used when there is no signal to select on (`savings: { router: false }`,
+     * or a catalogue with no `intents` and no `risk`).
      */
     passthrough(tools: ToolDescriptor[] | undefined): ToolSelection;
     select(text: string, tools: ToolDescriptor[] | undefined, execution: ExecutionClass, 
@@ -3014,9 +3075,19 @@ interface ProviderRequest {
     tools?: ToolDescriptor[];
     /** Output contract to enforce via constrained decoding, when available. */
     schema?: JsonSchema;
+    /**
+     * Output cap chosen by the caller: the smaller of `SavingsRequest.maxOutputTokens`
+     * and `response.maxOutputTokens`. Each adapter maps it to its own field.
+     */
     maxOutputTokens?: number;
     temperature?: number;
     seed?: number;
+    /**
+     * `SavingsRequest.reasoningEffort`, as the caller declared it. The
+     * OpenAI-compatible adapter sends it as `reasoning_effort`; absent means
+     * nothing is sent.
+     */
+    reasoningEffort?: string;
     /** Backend id selected by the constraints layer, for observability. */
     constraintBackend?: string;
     /**
@@ -3028,6 +3099,9 @@ interface ProviderRequest {
      * boundary; adapters that cache automatically (OpenAI) may ignore it. The
      * count is a claim about byte-for-byte stability, so a caller that rewrites an
      * earlier turn must lower it.
+     *
+     * `0` claims only what renders BEFORE the messages: the tool catalogue. The
+     * Anthropic adapter then marks the end of the `tools` block (its last tool).
      */
     stablePrefixMessages?: number;
     /**
@@ -3192,6 +3266,15 @@ interface ProviderResponse {
      * podía distinguirlo y hacía exactamente eso.
      */
     finishReason?: "stop" | "length" | "content_filter" | "tool_calls" | (string & {});
+    /**
+     * The provider's own reason behind `finishReason`, when it gave one: Gemini's
+     * OpenAI-compatible endpoint stops a filtered answer with
+     * `finish_reason: "content_filter: OTHER"` (or `PROHIBITED_CONTENT`,
+     * `SAFETY`…), which the built-in adapter reports as `finishReason:
+     * "content_filter"` and `finishReasonDetail: "OTHER"`. Informational: the
+     * layer decides on `finishReason` alone.
+     */
+    finishReasonDetail?: string;
     /**
      * The refusal message the model returned INSTEAD of an answer, exactly as the
      * provider wrote it: OpenAI's `message.refusal` (Chat Completions with
@@ -3504,7 +3587,28 @@ interface SavingsLayerConfig {
     provider: Provider;
     tokenizer: TokenizerRegistry;
     exactCache?: ExactCache;
+    /**
+     * WHERE reworded questions are looked up and stored. Giving the layer a cache
+     * does NOT switch it on: a request still has to ask for it with
+     * `savings.semanticCache`, or you set {@link semanticCacheDefault} to ask for
+     * it on every request. A layer that has a cache and a request that neither
+     * asks for it nor turns it off gets a one-time console warning (silenced by
+     * `silenceWarnings`, or by setting `semanticCacheDefault`, `false` included).
+     */
     semanticCache?: InMemorySemanticCache;
+    /**
+     * Default for `savings.semanticCache` on every request of this layer. A
+     * request's own value always wins, `false` included. Unset, the semantic cache
+     * is used only by requests that ask for it.
+     *
+     * OPT-IN, and deliberately a separate field from {@link semanticCache}: the
+     * cache is the backend, this is the decision to serve answers to reworded
+     * questions, which can serve a wrong answer. Prefer it with a cache built with
+     * `calibration`, which serves nothing until your own traffic certifies a
+     * floor. `false` states that you switch it on per request, and silences the
+     * warning about a cache that no request uses.
+     */
+    semanticCacheDefault?: boolean | "safe-only" | "include-agentic";
     singleflight?: Singleflight;
     resolvers?: ResolverRegistry;
     router?: ExecutionRouter;
@@ -4307,6 +4411,13 @@ interface Policy {
      * así que sin esto "hour" y "day" no significaban nada operativamente.
      */
     maxCacheAgeMs: number | null;
+    /**
+     * The user turns touch a sensitive topic (self-harm or suicide, a mental
+     * health crisis, a medical emergency), or null. With a topic, neither the
+     * exact nor the semantic cache is used and no reference record is left: the
+     * answer is measured and served, never replayed or credited as a saving.
+     */
+    sensitiveTopic: "self-harm" | "mental-health-crisis" | "medical-emergency" | null;
     reasons: string[];
 }
 declare function evaluatePolicy(c: CanonicalRequest): Policy;
@@ -4556,6 +4667,21 @@ declare function createMockProvider(options?: MockProviderOptions): Provider;
  * `constraints.backend: "none"` rather than claiming a guarantee the provider
  * never gave.
  *
+ * OpenAI-only request fields are sent only to OpenAI's own API
+ * (`prompt_cache_key`, see {@link OpenAICompatibleOptions.sendPromptCacheKey}).
+ * An optional field an endpoint still rejects by name with a 400/422
+ * (`prompt_cache_key`, `service_tier`, `seed`) is dropped and the request
+ * repeated once, and the model is remembered, as with `response_format`.
+ *
+ * Output is counted as the provider bills it: where `total_tokens` exceeds
+ * `prompt_tokens + completion_tokens` (Gemini's OpenAI-compatible endpoint
+ * leaves its thinking tokens out of `completion_tokens`), the difference is
+ * added to the output and reported as reasoning. Against OpenAI's own API,
+ * whose `completion_tokens` already include reasoning, nothing is added. A
+ * content-filter stop with a reason (Gemini's `"content_filter: OTHER"`) is
+ * reported as `finishReason: "content_filter"`, the reason in
+ * `finishReasonDetail`.
+ *
  * Network calls use global `fetch` (Node >= 18) so there is no SDK dependency.
  */
 interface OpenAICompatibleOptions {
@@ -4565,8 +4691,11 @@ interface OpenAICompatibleOptions {
      * Strip the `provider/` prefix from model ids before sending (`openai/gpt-4o-mini`
      * → `gpt-4o-mini`).
      *
-     * Default (undefined): strip when `baseUrl` is OpenAI's own API or an Azure
-     * OpenAI endpoint, keep otherwise. OpenAI rejects a prefixed id with a 404
+     * Default (undefined): strip when `baseUrl` is OpenAI's own API, an Azure
+     * OpenAI endpoint or Google's own Gemini API
+     * (https://generativelanguage.googleapis.com, whose OpenAI-compatible
+     * endpoint answers 404 «models/google/gemini-… is not found» to a prefixed
+     * id), keep otherwise. OpenAI rejects a prefixed id with a 404
      * ("model does not exist"), so the qualified ids the layer uses everywhere
      * else — pricing, cache keys, the report — would break the very first call
      * that reaches the provider. A gateway or a self-hosted server with a
@@ -4587,6 +4716,39 @@ interface OpenAICompatibleOptions {
      * saving. Benchmarks that make real cost claims must set this.
      */
     requireUsage?: boolean;
+    /**
+     * Which Chat Completions field carries the output cap (`maxOutputTokens`).
+     *
+     * Default (undefined): `max_completion_tokens` when `baseUrl` is OpenAI's own
+     * API (https://api.openai.com), `max_tokens` everywhere else. OpenAI
+     * deprecated `max_tokens` when it introduced `max_completion_tokens`
+     * (2024-09-12) and its reasoning models (o-series, gpt-5.x) reject it with a
+     * 400; `max_completion_tokens` is accepted by every current OpenAI chat
+     * model. OpenAI-compatible servers (proxies, self-hosted inference servers,
+     * routers, Azure with an older api-version, a local gateway) understand
+     * `max_tokens`, and some silently IGNORE a field they do not know, which
+     * would drop the cap without an error.
+     *
+     * With the default, a 400/422 that names the field sent as unsupported is
+     * retried once with the other one (it is rejected before any token is
+     * generated, so nothing is billed twice) and the model is remembered. An
+     * explicit value always wins and is never switched.
+     */
+    maxTokensParam?: "max_tokens" | "max_completion_tokens";
+    /**
+     * Whether to send `prompt_cache_key`, the routing hint OpenAI uses to land
+     * requests that share a prefix on the same cache instance.
+     *
+     * Default (undefined): only when `baseUrl` is OpenAI's own API
+     * (https://api.openai.com). The field is OpenAI's, and strict endpoints
+     * reject what they do not know: Gemini's OpenAI-compatible endpoint answers
+     * 400 «Unknown name "prompt_cache_key"», so from 0.4.1 to 0.4.31 every
+     * request with a cached prefix failed there. Set `true` for a proxy or router
+     * in front of OpenAI that forwards it (if the endpoint rejects it anyway,
+     * the request is repeated once without it and the model is remembered), or
+     * `false` to never send it.
+     */
+    sendPromptCacheKey?: boolean;
 }
 /**
  * ¿Cumple este esquema las reglas del modo `strict` de OpenAI?
@@ -4604,8 +4766,8 @@ interface OpenAICompatibleOptions {
 declare function isStrictCompatible(schema: unknown): boolean;
 /**
  * Decides the id that travels on the wire. Explicit option wins; otherwise the
- * endpoint decides: OpenAI/Azure get the bare id, everything else keeps what
- * the caller wrote. Exported for the pipeline-level test that pins the README
+ * endpoint decides: OpenAI/Azure and Google's own Gemini API get the bare id,
+ * everything else keeps what the caller wrote. Exported for the pipeline-level test that pins the README
  * snippet; not part of the public API.
  */
 declare function quitarPrefijoDeProveedor(model: string, explicito: boolean | undefined, baseUrl: string): string;
@@ -4616,6 +4778,8 @@ declare class OpenAICompatibleProvider implements Provider {
     /** Models of this endpoint already known to reject `response_format`. */
     private readonly sinResponseFormat;
     private streamOptionsRejected;
+    private readonly campoDeTopeAprendido;
+    private readonly camposRechazados;
     private readonly origen;
     constructor(opts: OpenAICompatibleOptions);
     generate(req: ProviderRequest, options?: ProviderStreamOptions): Promise<ProviderResponse>;
@@ -4629,6 +4793,10 @@ declare class OpenAICompatibleProvider implements Provider {
      * as one text event.
      */
     generateStream(req: ProviderRequest, options?: ProviderStreamOptions): AsyncGenerator<ProviderStreamEvent, void, undefined>;
+    private campoDeTope;
+    private corregirCampoDeTope;
+    private corregirCampoOpcional;
+    private campoAdmitido;
     private enviar;
     private construirCuerpo;
     /** Translate a 2xx endpoint response into the provider-neutral shape. */
@@ -4732,6 +4900,8 @@ declare class AnthropicProvider implements Provider {
     private readonly opts;
     private readonly fetchImpl;
     private readonly busquedas;
+    private readonly sinToolChoiceForzado;
+    private readonly sinMuestreo;
     constructor(opts?: AnthropicOptions);
     /** The id Anthropic's API sees: bare at `api.anthropic.com`, as written elsewhere. */
     private modeloEnElCable;
@@ -4746,6 +4916,7 @@ declare class AnthropicProvider implements Provider {
      * as one text event.
      */
     generateStream(req: ProviderRequest, options?: ProviderStreamOptions): AsyncGenerator<ProviderStreamEvent, void, undefined>;
+    private enviarConCorrecciones;
     private enviar;
     private construirCuerpo;
     private reenvioDeBusquedas;
@@ -5630,4 +5801,4 @@ declare function toCsv(rows: readonly Readonly<Record<string, CsvValue>>[], opti
     columns?: readonly string[];
 }): string;
 
-export { type AnswerJudge, type AnswerSource, type AnthropicOptions, AnthropicProvider, BV_SALA_VERSION, type BatchCounts, type BatchItemResult, type BatchJob, type BatchJobStatus, type BatchRequest, type BatchResults, BatchSubmitError, type BatchSubmitOptions, type BatchSummary, type BivelioHandoff, CATALOGO, COMPACTION_TRANSFORMS, COMPACT_ROWS_INSTRUCTION, type CacheBreakpointStore, type CacheOutcome, type CachePlacement, type CacheTier, type CacheWriteReport, type CachedAnswer, type CalibrationObservation, CalibrationStateError, type CanonicalRequest, type ChatMessage, type CompactRowsDecodeOptions, type CompactRowsOptions, type CompactVerification, type CompactionTransform, type CompressionEconomics, type CompressionPolicyInput, type CompressionRequest, type CompressionResult, type Compressor, type ConstraintBackend, type ConstraintBackendId, type ConstraintPlan, type ConversationState, type CostBasis, type CostModel, type CsvValue, DEFAULT_COST_MODEL, DEFAULT_MAX_STALENESS_SECONDS, DEFAULT_MIN_SAVED_CHARS, type DeferredExecutionReport, type DeterministicResolver, EMBEDDED_PUBLIC_KEY, type EmbeddingFn, type EvidenceDoc, type ExactCache, ExactTokenizer, type ExecutionClass, ExecutionRouter, type FocusExportOptions, type FocusRow, type FormatChoice, FormatSelector, type FormatSelectorOptions, type HandoffFact, HeuristicEstimator, InMemoryCacheBreakpointStore, InMemoryExactCache, InMemorySemanticCache, InMemoryToolSetFreezer, type JsonSchema, type LangfuseGenerationUpdate, LicenseError, type LicenseGate, type LicenseKeyShape, type LicenseKeyset, type LicenseStatus, type LicenseTokenPayload, type LocalClassifier, MODELOS_POR_CODIFICACION, MODELS_BY_ENCODING, type MemoryDelta, type MessageRole, MockProvider, type MockProviderOptions, type ModelCascadeMode, type ModelCascadeReport, type ModelCascadeSettings, type OpenAICompatibleOptions, OpenAICompatibleProvider, type OutputContract, PRECIOS_AL_DIA, type Policy, type PriceLeverSaving, type PriceLeverSettings, type ProtectedSpan, type Provider, type ProviderRequest, type ProviderResponse, type ProviderStreamEvent, type ProviderStreamOptions, type ProviderToolSearch, type ProviderToolSearchUsage, type ProviderUsage, QUALITY_STATUSES, type QualityCacheMatch, type QualityCascadeCheck, type QualityLedger, type QualityRepairCheck, type QualitySchemaCheck, type QualityStatus, type QualityTruncationCheck, REFERENCE_MAX_AGE_MS, REFERENCE_MESSAGE_ALLOWANCE, REFERENCE_OUTPUT_SHARE, type ReferenceCredit, type ReferenceRejection, type RequestContext, type Reranker, type ResolvedDeps, type ResolvedGate, type ResolvedLicense, ResolverRegistry, type ResolverResult, type RetrievalOptions, type RetrievalOutcome, type RetrievalPlan, type Retriever, type RevocationListBody, type RevocationListVerification, type RevocationPollEvent, RevocationPoller, type RevocationPollerOptions, RevocationSnapshot, type RouteDecision, type SavingsAdapterLevers, type SavingsAdapterOptions, type SavingsAttributeOptions, type SavingsBatches, type SavingsCallOptions, type SavingsLayer, type SavingsLayerConfig, SavingsLedger, type SavingsRecord, type SavingsReport, type SavingsReportForExport, type SavingsRequest, type SavingsResult, type SavingsSeriesOptions, type SavingsSeriesRow, type SavingsSpanAttributes, type SavingsSpanOptions, type SavingsStream, type SavingsStreamEvent, type SavingsStreamOptions, type SavingsToggles, type SavingsTracingOptions, SchemaRegistry, type SemanticCacheOptions, type SemanticCalibrationBound, type SemanticCalibrationOptions, type SemanticCalibrationSegmentState, type SemanticCalibrationState, type SemanticCalibrationStatus, type SemanticHit, type SemanticLookup, type SemanticProbe, type SemanticRecord, type SemanticVerifier, type SerializationFormat, type Serializer, Singleflight, type SpanAttributeValue, type SpanLike, type SpanStartOptionsLike, StreamAbortedError, type StreamingReport, type TokenEstimator, TokenizerRegistry, type ToolCallRequest, type ToolDescriptor, type ToolOutputCompaction, type ToolOutputCompactionSettings, type ToolRisk, type ToolSearchMode, type ToolSearchOptions, type ToolSelection, ToolSelector, type ToolSelectorOptions, type ToolSetFreezer, TraceCollector, type TraceEvent, type TracerLike, type TracerProviderLike, UMBRAL_LOCAL, UsageEmitter, type UsageEmitterOptions, type ValidationError, type ValidationResult, type VerificationCheck, type VerificationExclusion, type VerificationOutcome, type VerificationReport, type VerificationSettings, type VerificationSide, type VerificationStratum, adaptiveK, arithmeticResolver, buildExactKey, buildInflightKey, calibrateSemanticFloor, canonicalize, compactRowsSerializer, compactToolOutput, compresionFiel, computeNextDelay, cosine, createCompactRowsSerializer, createDefaultResolvers, createHandoff, createMockProvider, createSavingsLayer, csvSerializer, decodeCompactRows, dedupe, defaultTokenizerRegistry, deriveSeatId, deriveStatus, detectCi, embebedorLocal, emptyState, encodeCompactRows, evalArithmetic, evaluatePolicy, executeRetrievalPlan, gateFromKey, heuristicEstimator, instrumentSavingsLayer, isKeyset, isStrictCompatible, jsonCompactSerializer, keyTokensMatch, licenseKeyShape, liveGateFromKey, mergeDelta, modelFamilyOf, multisetEqual, negationCount, noopCompressor, normalizeLicenseKey, normalizeText, numericTokens, precioDeLista, projectContext, providerNativeBackend, quitarPrefijoDeProveedor, randomId, recordSavingsOnSpan, registerExactTokenizer, registrarTokenizadorExacto, renderState, resolveConstraintBackend, resolveDependencies, resolveLicense, sameAnswer, savingsSpanAttributes, sha256, shouldCompress, stableStringify, toCsv, toFocusRows, toLangfuseGeneration, toSavingsSeries, uniformColumns, validateAgainstSchema, vectorLocal, verifyCompactSignature, verifyLicenseToken, verifyRevocationList, withProtectedSpans, words };
+export { type AnswerJudge, type AnswerSource, type AnthropicOptions, AnthropicProvider, BV_SALA_VERSION, type BatchCounts, type BatchItemResult, type BatchJob, type BatchJobStatus, type BatchRequest, type BatchResults, BatchSubmitError, type BatchSubmitOptions, type BatchSummary, type BivelioHandoff, CATALOGO, COMPACTION_TRANSFORMS, COMPACT_ROWS_INSTRUCTION, type CacheBreakpointStore, type CacheOutcome, type CachePlacement, type CacheTier, type CacheWriteReport, type CachedAnswer, type CalibrationObservation, CalibrationStateError, type CanonicalRequest, type ChatMessage, type CompactRowsDecodeOptions, type CompactRowsOptions, type CompactVerification, type CompactionTransform, type CompressionEconomics, type CompressionPolicyInput, type CompressionRequest, type CompressionResult, type Compressor, type ConstraintBackend, type ConstraintBackendId, type ConstraintPlan, type ConversationState, type CostBasis, type CostModel, type CsvValue, DEFAULT_COST_MODEL, DEFAULT_MAX_STALENESS_SECONDS, DEFAULT_MIN_SAVED_CHARS, type DeferredExecutionReport, type DeterministicResolver, EMBEDDED_PUBLIC_KEY, type EmbeddingFn, type EvidenceDoc, type ExactCache, ExactTokenizer, type ExecutionClass, ExecutionRouter, type FocusExportOptions, type FocusRow, type FormatChoice, FormatSelector, type FormatSelectorOptions, type HandoffFact, HeuristicEstimator, InMemoryCacheBreakpointStore, InMemoryExactCache, InMemorySemanticCache, InMemoryToolSetFreezer, type JsonSchema, type LangfuseGenerationUpdate, LicenseError, type LicenseGate, type LicenseKeyShape, type LicenseKeyset, type LicenseStatus, type LicenseTokenPayload, type LocalClassifier, MODELOS_POR_CODIFICACION, MODELS_BY_ENCODING, type MemoryDelta, type MessageRole, MockProvider, type MockProviderOptions, type ModelCascadeMode, type ModelCascadeReport, type ModelCascadeSettings, type OpenAICompatibleOptions, OpenAICompatibleProvider, type OutputContract, PRECIOS_AL_DIA, type Policy, type PriceLeverSaving, type PriceLeverSettings, type ProtectedSpan, type Provider, type ProviderRequest, type ProviderResponse, type ProviderStreamEvent, type ProviderStreamOptions, type ProviderToolSearch, type ProviderToolSearchUsage, type ProviderUsage, QUALITY_STATUSES, type QualityCacheMatch, type QualityCascadeCheck, type QualityLedger, type QualityRepairCheck, type QualitySchemaCheck, type QualityStatus, type QualityTruncationCheck, REFERENCE_MAX_AGE_MS, REFERENCE_MESSAGE_ALLOWANCE, REFERENCE_OUTPUT_SHARE, type ReasoningEffort, type ReferenceCredit, type ReferenceRejection, type RequestContext, type Reranker, type ResolvedDeps, type ResolvedGate, type ResolvedLicense, ResolverRegistry, type ResolverResult, type RetrievalOptions, type RetrievalOutcome, type RetrievalPlan, type Retriever, type RevocationListBody, type RevocationListVerification, type RevocationPollEvent, RevocationPoller, type RevocationPollerOptions, RevocationSnapshot, type RouteDecision, type SavingsAdapterLevers, type SavingsAdapterOptions, type SavingsAttributeOptions, type SavingsBatches, type SavingsCallOptions, type SavingsLayer, type SavingsLayerConfig, SavingsLedger, type SavingsRecord, type SavingsReport, type SavingsReportForExport, type SavingsRequest, type SavingsResult, type SavingsSeriesOptions, type SavingsSeriesRow, type SavingsSpanAttributes, type SavingsSpanOptions, type SavingsStream, type SavingsStreamEvent, type SavingsStreamOptions, type SavingsToggles, type SavingsTracingOptions, SchemaRegistry, type SemanticCacheOptions, type SemanticCalibrationBound, type SemanticCalibrationOptions, type SemanticCalibrationSegmentState, type SemanticCalibrationState, type SemanticCalibrationStatus, type SemanticHit, type SemanticLookup, type SemanticProbe, type SemanticRecord, type SemanticVerifier, type SerializationFormat, type Serializer, Singleflight, type SpanAttributeValue, type SpanLike, type SpanStartOptionsLike, StreamAbortedError, type StreamingReport, type TokenEstimator, TokenizerRegistry, type ToolCallRequest, type ToolDescriptor, type ToolOutputCompaction, type ToolOutputCompactionSettings, type ToolRisk, type ToolSearchMode, type ToolSearchOptions, type ToolSelection, ToolSelector, type ToolSelectorOptions, type ToolSetFreezer, TraceCollector, type TraceEvent, type TracerLike, type TracerProviderLike, UMBRAL_LOCAL, UsageEmitter, type UsageEmitterOptions, type ValidationError, type ValidationResult, type VerificationCheck, type VerificationExclusion, type VerificationOutcome, type VerificationReport, type VerificationSettings, type VerificationSide, type VerificationStratum, adaptiveK, arithmeticResolver, buildExactKey, buildInflightKey, calibrateSemanticFloor, canonicalize, compactRowsSerializer, compactToolOutput, compresionFiel, computeNextDelay, cosine, createCompactRowsSerializer, createDefaultResolvers, createHandoff, createMockProvider, createSavingsLayer, csvSerializer, decodeCompactRows, dedupe, defaultTokenizerRegistry, deriveSeatId, deriveStatus, detectCi, embebedorLocal, emptyState, encodeCompactRows, evalArithmetic, evaluatePolicy, executeRetrievalPlan, gateFromKey, heuristicEstimator, instrumentSavingsLayer, isKeyset, isStrictCompatible, jsonCompactSerializer, keyTokensMatch, licenseKeyShape, liveGateFromKey, mergeDelta, modelFamilyOf, multisetEqual, negationCount, noopCompressor, normalizeLicenseKey, normalizeText, numericTokens, precioDeLista, projectContext, providerNativeBackend, quitarPrefijoDeProveedor, randomId, recordSavingsOnSpan, registerExactTokenizer, registrarTokenizadorExacto, renderState, resolveConstraintBackend, resolveDependencies, resolveLicense, sameAnswer, savingsSpanAttributes, sha256, shouldCompress, stableStringify, toCsv, toFocusRows, toLangfuseGeneration, toSavingsSeries, uniformColumns, validateAgainstSchema, vectorLocal, verifyCompactSignature, verifyLicenseToken, verifyRevocationList, withProtectedSpans, words };