@oxyhq/contracts 0.26.0 → 0.28.0

package/dist/types/inference/errors.d.ts

@@ -0,0 +1,206 @@
+/**
+ * Inference errors and retryability.
+ *
+ * One closed set of codes, shared by the Oxy public edge, the data plane and
+ * every SDK.
+ * Closed because the alternative — a free-form `code` string — makes a client's
+ * error handling a guess about the producer's spelling, and makes "is this
+ * worth retrying" a decision every consumer re-derives from prose.
+ *
+ * Retryability is carried explicitly and is CONSTRAINED by the code: a code
+ * that can never succeed on a bare retry (an invalid request, a denied
+ * permission, an insufficient balance) cannot claim `retryable: true`. Without
+ * that constraint the field is advisory, and one producer setting it optimistically
+ * turns every client into a retry storm against a request that will never pass.
+ *
+ * The provider passthrough exists so a customer can see what the upstream said
+ * without Oxy having to interpret every provider's error vocabulary — but it is
+ * the single most likely place for an upstream credential to escape, because
+ * provider errors routinely echo the request that caused them. It is therefore
+ * a `.strict()` object of four fields with no room for headers or a request
+ * body, and its free text is refused if it looks like it contains a credential.
+ *
+ * Decided in: docs/adr/0010-public-api-compatibility.md.
+ */
+import { z } from 'zod';
+/**
+ * The closed set of inference error codes.
+ *
+ * Grouped by who must act: the caller (`invalid_request` … `idempotency_conflict`),
+ * the account owner (`insufficient_balance`, `spending_limit_exceeded`,
+ * `quota_exceeded`, `byok_credential_invalid`), routing/permission policy
+ * (`policy_violation`, `commercial_permission_denied`, `no_route_available`),
+ * and the platform or its upstreams (everything from `deployment_unavailable`).
+ *
+ * The platform group is NOT uniformly retryable, and that is the point of
+ * `provider_credential_invalid` sitting in it: an upstream that refuses the
+ * PLATFORM's own credential fails every identical retry until an operator
+ * rotates a key, so classifying it as `provider_error` would send every client
+ * into a retry loop against a request that cannot succeed.
+ */
+export declare const INFERENCE_ERROR_CODES: readonly ["invalid_request", "authentication_failed", "permission_denied", "insufficient_scope", "model_not_found", "unsupported_modality", "context_length_exceeded", "request_too_large", "output_limit_exceeded", "idempotency_conflict", "insufficient_balance", "spending_limit_exceeded", "quota_exceeded", "byok_credential_invalid", "policy_violation", "commercial_permission_denied", "no_route_available", "upstream_content_filtered", "cancelled", "rate_limited", "deployment_unavailable", "provider_error", "provider_timeout", "provider_overloaded", "provider_credential_invalid", "service_unavailable", "internal_error"];
+export declare const inferenceErrorCodeSchema: z.ZodEnum<["invalid_request", "authentication_failed", "permission_denied", "insufficient_scope", "model_not_found", "unsupported_modality", "context_length_exceeded", "request_too_large", "output_limit_exceeded", "idempotency_conflict", "insufficient_balance", "spending_limit_exceeded", "quota_exceeded", "byok_credential_invalid", "policy_violation", "commercial_permission_denied", "no_route_available", "upstream_content_filtered", "cancelled", "rate_limited", "deployment_unavailable", "provider_error", "provider_timeout", "provider_overloaded", "provider_credential_invalid", "service_unavailable", "internal_error"]>;
+/**
+ * Codes for which an identical retried request cannot succeed.
+ *
+ * `rate_limited` and `quota_exceeded` sit on opposite sides of this line
+ * deliberately: a rate limit clears on its own within the window the response
+ * names, while a quota is an account-level ceiling that only a human raises.
+ * `cancelled` is here because the caller already withdrew the request; a client
+ * that retries it is contradicting its own cancellation.
+ *
+ * `byok_credential_invalid` and `provider_credential_invalid` are the same
+ * failure seen from the two sides of the BYOK boundary — the customer's own
+ * upstream credential and the platform's — and they are two codes rather than
+ * one because only the first names an action the customer can take. Both are
+ * non-retryable for the same reason: a credential an upstream has refused keeps
+ * being refused until somebody replaces it.
+ */
+export declare const NON_RETRYABLE_INFERENCE_ERROR_CODES: readonly ["invalid_request", "authentication_failed", "permission_denied", "insufficient_scope", "model_not_found", "unsupported_modality", "context_length_exceeded", "request_too_large", "output_limit_exceeded", "idempotency_conflict", "insufficient_balance", "spending_limit_exceeded", "quota_exceeded", "byok_credential_invalid", "policy_violation", "commercial_permission_denied", "no_route_available", "upstream_content_filtered", "cancelled", "provider_credential_invalid"];
+/**
+ * Free text that is safe to hand a customer: bounded, and refused outright if a
+ * credential marker appears in it. Applied to BOTH the Oxy message and the
+ * upstream one — a leak is no less a leak for having been written by a provider.
+ */
+export declare const safeErrorTextSchema: z.ZodEffects<z.ZodString, string, string>;
+/**
+ * A coarse classification of an upstream failure (ADR 0010's `upstreamCategory`).
+ *
+ * Distinct from {@link providerErrorPassthroughSchema}, which carries the
+ * upstream's OWN code and text: this is Oxy's reading of what kind of failure it
+ * was, in a vocabulary that is the same across every provider, so a client can
+ * branch on it without knowing who served the request.
+ */
+export declare const upstreamErrorCategorySchema: z.ZodEnum<["rate_limit", "quota", "timeout", "overloaded", "server_error", "content_filter", "invalid_request", "authentication", "unknown"]>;
+/**
+ * What the upstream provider said, reduced to the four fields a customer can
+ * act on.
+ *
+ * `.strict()` is the security control here, not a tidiness preference: it means
+ * a producer cannot widen this by attaching `requestHeaders`, `curl`, `body` or
+ * `raw` and have it silently pass. Adding a field is a contract change with a
+ * version bump and a review, which is the point.
+ */
+export declare const providerErrorPassthroughSchema: z.ZodObject<{
+    provider: z.ZodString;
+    /** The upstream HTTP status, when the upstream spoke HTTP. */
+    status: z.ZodOptional<z.ZodNumber>;
+    /** The upstream's own error code, verbatim and uninterpreted. */
+    code: z.ZodOptional<z.ZodString>;
+    /** The upstream's message, subject to the same credential refusal. */
+    message: z.ZodOptional<z.ZodEffects<z.ZodString, string, string>>;
+}, "strict", z.ZodTypeAny, {
+    provider: string;
+    code?: string | undefined;
+    message?: string | undefined;
+    status?: number | undefined;
+}, {
+    provider: string;
+    code?: string | undefined;
+    message?: string | undefined;
+    status?: number | undefined;
+}>;
+/**
+ * The error body every inference surface returns and every stream error event
+ * carries.
+ *
+ * `requestId` is always present — an error a customer cannot correlate with a
+ * log line is an error they have to reproduce to report.
+ */
+export declare const inferenceErrorSchema: z.ZodEffects<z.ZodObject<{
+    /** See `version.ts`: this shape appears alone on the wire, so it is versioned. */
+    schemaVersion: z.ZodLiteral<1>;
+    code: z.ZodEnum<["invalid_request", "authentication_failed", "permission_denied", "insufficient_scope", "model_not_found", "unsupported_modality", "context_length_exceeded", "request_too_large", "output_limit_exceeded", "idempotency_conflict", "insufficient_balance", "spending_limit_exceeded", "quota_exceeded", "byok_credential_invalid", "policy_violation", "commercial_permission_denied", "no_route_available", "upstream_content_filtered", "cancelled", "rate_limited", "deployment_unavailable", "provider_error", "provider_timeout", "provider_overloaded", "provider_credential_invalid", "service_unavailable", "internal_error"]>;
+    message: z.ZodEffects<z.ZodString, string, string>;
+    retryable: z.ZodBoolean;
+    requestId: z.ZodString;
+    /** How long to wait before retrying. Only meaningful when `retryable`. */
+    retryAfterMs: z.ZodOptional<z.ZodNumber>;
+    /** The request field at fault, for `invalid_request`. */
+    param: z.ZodOptional<z.ZodString>;
+    /** Present only when an upstream provider was reached and failed. */
+    upstreamCategory: z.ZodOptional<z.ZodEnum<["rate_limit", "quota", "timeout", "overloaded", "server_error", "content_filter", "invalid_request", "authentication", "unknown"]>>;
+    providerError: z.ZodOptional<z.ZodObject<{
+        provider: z.ZodString;
+        /** The upstream HTTP status, when the upstream spoke HTTP. */
+        status: z.ZodOptional<z.ZodNumber>;
+        /** The upstream's own error code, verbatim and uninterpreted. */
+        code: z.ZodOptional<z.ZodString>;
+        /** The upstream's message, subject to the same credential refusal. */
+        message: z.ZodOptional<z.ZodEffects<z.ZodString, string, string>>;
+    }, "strict", z.ZodTypeAny, {
+        provider: string;
+        code?: string | undefined;
+        message?: string | undefined;
+        status?: number | undefined;
+    }, {
+        provider: string;
+        code?: string | undefined;
+        message?: string | undefined;
+        status?: number | undefined;
+    }>>;
+}, "strip", z.ZodTypeAny, {
+    code: "invalid_request" | "authentication_failed" | "permission_denied" | "insufficient_scope" | "model_not_found" | "unsupported_modality" | "context_length_exceeded" | "request_too_large" | "output_limit_exceeded" | "idempotency_conflict" | "insufficient_balance" | "spending_limit_exceeded" | "quota_exceeded" | "byok_credential_invalid" | "policy_violation" | "commercial_permission_denied" | "no_route_available" | "upstream_content_filtered" | "cancelled" | "rate_limited" | "deployment_unavailable" | "provider_error" | "provider_timeout" | "provider_overloaded" | "provider_credential_invalid" | "service_unavailable" | "internal_error";
+    message: string;
+    requestId: string;
+    schemaVersion: 1;
+    retryable: boolean;
+    retryAfterMs?: number | undefined;
+    param?: string | undefined;
+    upstreamCategory?: "unknown" | "authentication" | "invalid_request" | "rate_limit" | "quota" | "timeout" | "overloaded" | "server_error" | "content_filter" | undefined;
+    providerError?: {
+        provider: string;
+        code?: string | undefined;
+        message?: string | undefined;
+        status?: number | undefined;
+    } | undefined;
+}, {
+    code: "invalid_request" | "authentication_failed" | "permission_denied" | "insufficient_scope" | "model_not_found" | "unsupported_modality" | "context_length_exceeded" | "request_too_large" | "output_limit_exceeded" | "idempotency_conflict" | "insufficient_balance" | "spending_limit_exceeded" | "quota_exceeded" | "byok_credential_invalid" | "policy_violation" | "commercial_permission_denied" | "no_route_available" | "upstream_content_filtered" | "cancelled" | "rate_limited" | "deployment_unavailable" | "provider_error" | "provider_timeout" | "provider_overloaded" | "provider_credential_invalid" | "service_unavailable" | "internal_error";
+    message: string;
+    requestId: string;
+    schemaVersion: 1;
+    retryable: boolean;
+    retryAfterMs?: number | undefined;
+    param?: string | undefined;
+    upstreamCategory?: "unknown" | "authentication" | "invalid_request" | "rate_limit" | "quota" | "timeout" | "overloaded" | "server_error" | "content_filter" | undefined;
+    providerError?: {
+        provider: string;
+        code?: string | undefined;
+        message?: string | undefined;
+        status?: number | undefined;
+    } | undefined;
+}>, {
+    code: "invalid_request" | "authentication_failed" | "permission_denied" | "insufficient_scope" | "model_not_found" | "unsupported_modality" | "context_length_exceeded" | "request_too_large" | "output_limit_exceeded" | "idempotency_conflict" | "insufficient_balance" | "spending_limit_exceeded" | "quota_exceeded" | "byok_credential_invalid" | "policy_violation" | "commercial_permission_denied" | "no_route_available" | "upstream_content_filtered" | "cancelled" | "rate_limited" | "deployment_unavailable" | "provider_error" | "provider_timeout" | "provider_overloaded" | "provider_credential_invalid" | "service_unavailable" | "internal_error";
+    message: string;
+    requestId: string;
+    schemaVersion: 1;
+    retryable: boolean;
+    retryAfterMs?: number | undefined;
+    param?: string | undefined;
+    upstreamCategory?: "unknown" | "authentication" | "invalid_request" | "rate_limit" | "quota" | "timeout" | "overloaded" | "server_error" | "content_filter" | undefined;
+    providerError?: {
+        provider: string;
+        code?: string | undefined;
+        message?: string | undefined;
+        status?: number | undefined;
+    } | undefined;
+}, {
+    code: "invalid_request" | "authentication_failed" | "permission_denied" | "insufficient_scope" | "model_not_found" | "unsupported_modality" | "context_length_exceeded" | "request_too_large" | "output_limit_exceeded" | "idempotency_conflict" | "insufficient_balance" | "spending_limit_exceeded" | "quota_exceeded" | "byok_credential_invalid" | "policy_violation" | "commercial_permission_denied" | "no_route_available" | "upstream_content_filtered" | "cancelled" | "rate_limited" | "deployment_unavailable" | "provider_error" | "provider_timeout" | "provider_overloaded" | "provider_credential_invalid" | "service_unavailable" | "internal_error";
+    message: string;
+    requestId: string;
+    schemaVersion: 1;
+    retryable: boolean;
+    retryAfterMs?: number | undefined;
+    param?: string | undefined;
+    upstreamCategory?: "unknown" | "authentication" | "invalid_request" | "rate_limit" | "quota" | "timeout" | "overloaded" | "server_error" | "content_filter" | undefined;
+    providerError?: {
+        provider: string;
+        code?: string | undefined;
+        message?: string | undefined;
+        status?: number | undefined;
+    } | undefined;
+}>;
+export type InferenceErrorCode = z.infer<typeof inferenceErrorCodeSchema>;
+export type UpstreamErrorCategory = z.infer<typeof upstreamErrorCategorySchema>;
+export type ProviderErrorPassthrough = z.infer<typeof providerErrorPassthroughSchema>;
+export type InferenceError = z.infer<typeof inferenceErrorSchema>;

package/dist/types/inference/identifiers.d.ts

@@ -0,0 +1,157 @@
+/**
+ * Identifiers, references and wire primitives shared by every Oxy↔data-plane
+ * inference contract.
+ *
+ * Two kinds of identifier live here, and the difference matters:
+ *
+ *  - **Principal identifiers** owned by Oxy (`accountId`, `applicationId`,
+ *    `credentialId`, the optional delegated `userId`). The data plane may store
+ *    them as immutable references; it never owns, mints or mutates them.
+ *  - **Catalogue references** (`<publisher>/<model>`, `<publisher>/<model>@<revision>`,
+ *    a routing-profile slug). These are the strings a customer types, so their
+ *    grammar is part of the public contract, not an implementation detail.
+ *
+ * Platform-agnostic — zod only. Every regex here is plain ASCII: this package
+ * is imported by React Native apps running on Hermes, which rejects Unicode
+ * property escapes (`\p{…}`) at runtime.
+ *
+ * Decided in: docs/adr/0007-canonical-request-attribution.md, docs/adr/0008-catalogue-concept-separation.md.
+ */
+import { z } from 'zod';
+/**
+ * An Oxy account id — the account that owns the workload and is financially
+ * responsible for it.
+ *
+ * Branded, and that brand is load-bearing rather than decorative: it is the
+ * type-level half of the rule that a delegated end-user identity can never
+ * become the billing identity (see {@link delegatedUserIdSchema} and
+ * `billingPrincipalSchema`). A plain `string` — and therefore any id read out
+ * of a header, a JWT claim or a request body — is not assignable to it; the
+ * only way to obtain one is to parse a value through this schema.
+ */
+export declare const oxyAccountIdSchema: z.ZodBranded<z.ZodString, "OxyAccountId">;
+/**
+ * A delegated end-user identity (Alia's `X-Oxy-User-Id`), branded with a
+ * DIFFERENT brand from {@link oxyAccountIdSchema} so the two cannot be
+ * substituted for one another in either direction, in any consumer, without a
+ * cast that review would catch.
+ *
+ * It exists for attribution and product-side personalisation only. It is never
+ * a billing principal, never an access-control principal, and its presence
+ * never changes which account is charged.
+ */
+export declare const delegatedUserIdSchema: z.ZodBranded<z.ZodString, "DelegatedUserId">;
+/**
+ * An Oxy `Application._id`. Not branded: no invariant in this contract turns on
+ * confusing it with another id, and brand inflation costs every producer a
+ * parse call for no safety. The two ids above are branded because ADR 0007's
+ * rule is exactly that they must not be interchangeable.
+ */
+export declare const oxyApplicationIdSchema: z.ZodString;
+/** An Oxy `ApplicationCredential._id` — the credential used for this request. */
+export declare const oxyCredentialIdSchema: z.ZodString;
+/**
+ * A request id allocated by the Oxy EDGE, at admission and BEFORE
+ * authentication, so that a request rejected for a bad credential is as
+ * traceable as one that was served (ADR 0007, and step 1 of ADR 0010's edge
+ * order). It is required on the inbound envelope, which is what makes the data
+ * plane a consumer of this id rather than its source: the data plane echoes it
+ * on every stream event, on the usage report and on anything it can be asked
+ * about later.
+ *
+ * Correlates the Oxy edge, the data plane, the financial ledger and the
+ * customer-visible receipt, so it appears on every stream event and every
+ * ledger record.
+ */
+export declare const requestIdSchema: z.ZodString;
+/**
+ * A generation id generated by the data plane, present when a request produced
+ * a generation that can be looked up later (`GET /v1/generations/:id`).
+ */
+export declare const generationIdSchema: z.ZodString;
+/**
+ * A caller-supplied idempotency key. Every reserve/settle/refund call is keyed
+ * on one so a retry, a redelivered event or a duplicated webhook can never
+ * charge twice.
+ */
+export declare const idempotencyKeySchema: z.ZodString;
+/** Credential environments. A credential is issued into exactly one of them. */
+export declare const inferenceEnvironmentSchema: z.ZodEnum<["development", "staging", "production"]>;
+/**
+ * An instant, as an ISO 8601 string in UTC (`2026-08-15T09:41:00.000Z`).
+ *
+ * Validated rather than left free-form — unlike the session contracts, which
+ * carry legacy expiry strings no consumer interprets, every instant here is
+ * read: a reservation expires, a price version starts applying, a receipt is
+ * settled. One canonical spelling (UTC, `Z`) so two records that describe the
+ * same moment compare and sort as equal, which a mix of offsets would not.
+ */
+export declare const inferenceTimestampSchema: z.ZodString;
+/**
+ * A calendar DATE with no instant attached (`2026-05-01`) — a knowledge cutoff
+ * or a release date, which are published as days and become wrong when a
+ * timezone is invented for them.
+ */
+export declare const inferenceDateSchema: z.ZodString;
+/** An absolute https URL, for model cards, licenses and provider documentation. */
+export declare const inferenceHttpsUrlSchema: z.ZodString;
+/** A publisher slug, e.g. `openai`, `anthropic`, `meta`, `alia`. */
+export declare const publisherSlugSchema: z.ZodString;
+/** A model slug within its publisher's namespace, e.g. `gpt-5`, `llama-3.1-70b`. */
+export declare const modelSlugSchema: z.ZodString;
+/**
+ * A canonical model id, `<publisher>/<model>`. This names a MODEL — a
+ * long-lived product identity whose behaviour changes as revisions ship. It
+ * does not name a revision, a deployment or a provider.
+ */
+export declare const modelIdSchema: z.ZodString;
+/** An immutable revision label, unique within its model, e.g. `2026-05-01`. */
+export declare const modelRevisionLabelSchema: z.ZodString;
+/**
+ * A model reference as a customer writes it: `<publisher>/<model>` (the model's
+ * current revision, chosen by Oxy) or `<publisher>/<model>@<revision>` (an
+ * immutable revision the customer pinned).
+ *
+ * Both forms name a CONCRETE MODEL. Neither can name a routing profile — see
+ * {@link routingProfileSlugSchema} — which is what makes "a request for a
+ * concrete model is never silently replaced with a different model" a
+ * distinction the type system can carry rather than a convention.
+ */
+export declare const modelReferenceSchema: z.ZodString;
+/**
+ * A routing-profile slug, e.g. `auto`, `fast`, `quality`.
+ *
+ * Deliberately refuses a `/`, so a profile can never be written in the shape of
+ * a model id and no caller can be confused about whether they asked for a
+ * concrete model or for Oxy to choose one. Modes like `auto`/`fast`/`quality`
+ * are profiles or product presets; they are never model objects.
+ */
+export declare const routingProfileSlugSchema: z.ZodString;
+/** An inference provider slug, e.g. `openai`, `bedrock`, `oxy-hosted`. */
+export declare const inferenceProviderSlugSchema: z.ZodString;
+/**
+ * A deployment/endpoint id. Opaque to customers: which concrete endpoint served
+ * a request is the data plane's operational detail, and only the customer-safe
+ * subset of it is ever attributed back (see the catalogue's serving-boundary
+ * rules).
+ */
+export declare const deploymentIdSchema: z.ZodString;
+/**
+ * A region identifier, e.g. `us-west-2`, `eu-central-1`. Free-form rather than
+ * a closed enum because the set is provider-defined and grows without any
+ * contract change; residency policies match on exact strings.
+ */
+export declare const inferenceRegionSchema: z.ZodString;
+/**
+ * The publisher namespace reserved for models Alia actually owns or derives.
+ *
+ * `alia/*` is never a re-badged third-party route and never a prompt preset —
+ * enforcing that is the job of `modelSchema`'s provenance refinement, which
+ * this constant exists to be checked against.
+ */
+export declare const RESERVED_ALIA_PUBLISHER = "alia";
+export type OxyAccountId = z.infer<typeof oxyAccountIdSchema>;
+export type DelegatedUserId = z.infer<typeof delegatedUserIdSchema>;
+export type InferenceEnvironment = z.infer<typeof inferenceEnvironmentSchema>;
+export type ModelReference = z.infer<typeof modelReferenceSchema>;
+export type ModelId = z.infer<typeof modelIdSchema>;

package/dist/types/inference/money.d.ts

@@ -0,0 +1,185 @@
+/**
+ * Money and usage units for the inference contracts.
+ *
+ * The non-negotiable invariant this file exists to make structural: **customer
+ * charges never use floating-point values as the financial source of truth.**
+ * A JS `number` cannot represent `0.1 + 0.2` exactly, and an inference ledger
+ * adds millions of small amounts, so a float total is wrong by construction
+ * rather than by accident.
+ *
+ * Every amount and every price is one representation: {@link exactDecimalSchema},
+ * an exact decimal STRING at the scale ADR 0009 declares. Amounts are not
+ * integer minor units, and that is the ADR's decision rather than an oversight:
+ * one token costs several orders of magnitude less than one cent, so rounding
+ * per request would make a customer's bill depend on how their client chunked
+ * its work. Rounding happens ONCE, at the invoice boundary, and is itself a
+ * ledger entry.
+ *
+ * A string is also what the driver hands back — `postgres.js` decodes `NUMERIC`
+ * as a string — so keeping it a string on the wire means accidental JS
+ * arithmetic fails loudly instead of silently losing precision. Money
+ * arithmetic happens in SQL or in a decimal type, never in a JS `number`.
+ *
+ * Units are carried separately from money in every shape: a receipt says both
+ * "204 output tokens" and "0.003060000000 USD", and neither is derived from the
+ * other at read time. That separation is what lets a price version change
+ * without rewriting settled history.
+ *
+ * Decided in: docs/adr/0009-usage-reservation-and-settlement.md.
+ */
+import { z } from 'zod';
+/** ISO 4217 alpha-3 currency code, e.g. `USD`. */
+export declare const currencyCodeSchema: z.ZodString;
+/**
+ * The declared fractional scale of every amount and price in this contract, and
+ * of the `NUMERIC` columns the ledger stores them in.
+ *
+ * Twelve digits is sub-minor-unit precision by a wide margin: at $3 per million
+ * input tokens, one token costs `0.000003000000`, which this scale represents
+ * exactly. Amounts are compared and summed NUMERICALLY, never as text — `3.0`
+ * and `3.000000000000` are one amount written two ways.
+ */
+export declare const INFERENCE_MONEY_SCALE = 12;
+/**
+ * An exact non-negative decimal, carried as a STRING so no parse step can turn
+ * it into a float on the way past. Up to 18 integer digits and
+ * {@link INFERENCE_MONEY_SCALE} fractional digits.
+ *
+ * Non-negative: direction is carried by the SHAPE — a receipt debits, a refund
+ * credits — so a stray sign can never silently invert an entry.
+ *
+ * No exponent form: `1e-6` and `0.000001` are the same number, but only one of
+ * them survives a naive string comparison, a cache key or a log grep intact.
+ *
+ * Branded, so a bare `string` is not assignable and an amount cannot arrive
+ * from string concatenation that was never checked. Producers construct one
+ * with `exactDecimalSchema.parse(value)`.
+ */
+export declare const exactDecimalSchema: z.ZodBranded<z.ZodString, "ExactDecimal">;
+/**
+ * An amount of money: the exact decimal plus the currency it is in.
+ *
+ * `.strict()` so a payload carrying a convenience float beside the exact value
+ * (`{ amount: '18.06', amountFloat: 18.06 }`) is REJECTED rather than stripped.
+ * A stripped float is the more dangerous outcome: it disappears silently here
+ * and survives in the producer, where it is the value somebody eventually
+ * displays.
+ */
+export declare const moneySchema: z.ZodObject<{
+    amount: z.ZodBranded<z.ZodString, "ExactDecimal">;
+    currency: z.ZodString;
+}, "strict", z.ZodTypeAny, {
+    amount: string & z.BRAND<"ExactDecimal">;
+    currency: string;
+}, {
+    amount: string;
+    currency: string;
+}>;
+/**
+ * The closed set of units inference is metered in.
+ *
+ * **The units PARTITION a request: every unit counts material no other unit
+ * counts.** `cached_input_tokens` is not part of `input_tokens`, and
+ * `reasoning_tokens` is not part of `output_tokens` — they are siblings, not
+ * subsets. A request whose 10 000-token prompt was served 9 000 tokens from
+ * cache is reported as `input_tokens: 1000` beside `cached_input_tokens: 9000`,
+ * never as `input_tokens: 10000` beside it.
+ *
+ * That belongs to the definition rather than to a convention somewhere else,
+ * because settlement applies a price to EVERY reported unit and sums them
+ * (`inferenceLedger.service.ts`'s `computeCharge`). Under the partition rule
+ * that sum IS the request's cost, and a cached token can carry its own — lower
+ * — price. Under the nested reading the same sum charges the cached and
+ * reasoning tokens twice: once inside their parent and once on their own line.
+ * It fails silently, because every total still looks plausible and the receipt
+ * is still internally consistent, and on a reasoning model the reasoning tokens
+ * can dominate the completion, so the error is not marginal.
+ *
+ * **Every OpenAI-compatible provider reports the other way round**:
+ * `prompt_tokens` INCLUDES `prompt_tokens_details.cached_tokens`, and
+ * `completion_tokens` INCLUDES `completion_tokens_details.reasoning_tokens`.
+ * Normalising is the data plane's job and it is subtraction:
+ *
+ * ```text
+ * input_tokens  = prompt_tokens     - prompt_tokens_details.cached_tokens
+ * output_tokens = completion_tokens - completion_tokens_details.reasoning_tokens
+ * ```
+ *
+ * No refinement in this package can enforce it, and saying so is part of the
+ * rule: a nested report and a disjoint one are the same four non-negative
+ * integers, so no predicate over a single report can tell them apart. The two
+ * structural guards that DO exist — refining `cached <= input` and
+ * `reasoning <= output`, or deriving the parents instead of reporting them —
+ * both encode the nested reading, which is the one this rule rejects. What IS
+ * enforceable is the arithmetic that depends on the rule, and that is where the
+ * enforcement lives — `inferenceLedger.service.test.ts` prices a report in
+ * which cached and reasoning tokens are both non-zero and asserts the exact
+ * total, which the nested reading cannot produce.
+ *
+ * Where the public surface has to speak a nested dialect, the sum is put back
+ * at the boundary rather than the internal reading being bent to it
+ * (`routes/inferenceEdge.ts` renders `prompt_tokens` as
+ * `input_tokens + cached_input_tokens`).
+ *
+ * Time is carried in integer MILLISECONDS rather than seconds so that no unit
+ * quantity is ever fractional: a 12.5-second transcription is `12500`, exactly,
+ * and the "units are integers" rule holds for every modality instead of holding
+ * for tokens and being quietly broken by audio.
+ */
+export declare const USAGE_UNITS: readonly ["input_tokens", "cached_input_tokens", "output_tokens", "reasoning_tokens", "requests", "images", "audio_input_milliseconds", "audio_output_milliseconds", "video_milliseconds", "characters", "embeddings"];
+export declare const usageUnitSchema: z.ZodEnum<["input_tokens", "cached_input_tokens", "output_tokens", "reasoning_tokens", "requests", "images", "audio_input_milliseconds", "audio_output_milliseconds", "video_milliseconds", "characters", "embeddings"]>;
+/**
+ * A metered quantity of ONE unit. Never money — a quantity carries no price and
+ * no currency, so a consumer cannot mistake a token count for an amount owed.
+ */
+export declare const usageQuantitySchema: z.ZodObject<{
+    unit: z.ZodEnum<["input_tokens", "cached_input_tokens", "output_tokens", "reasoning_tokens", "requests", "images", "audio_input_milliseconds", "audio_output_milliseconds", "video_milliseconds", "characters", "embeddings"]>;
+    quantity: z.ZodNumber;
+}, "strict", z.ZodTypeAny, {
+    unit: "input_tokens" | "cached_input_tokens" | "output_tokens" | "reasoning_tokens" | "requests" | "images" | "audio_input_milliseconds" | "audio_output_milliseconds" | "video_milliseconds" | "characters" | "embeddings";
+    quantity: number;
+}, {
+    unit: "input_tokens" | "cached_input_tokens" | "output_tokens" | "reasoning_tokens" | "requests" | "images" | "audio_input_milliseconds" | "audio_output_milliseconds" | "video_milliseconds" | "characters" | "embeddings";
+    quantity: number;
+}>;
+/**
+ * Where a metered quantity came from.
+ *
+ * Kept explicit because the three are not interchangeable when a charge is
+ * disputed: `provider_reported` is the upstream's own count, `oxy_measured` is
+ * counted by the platform (streamed bytes, wall-clock milliseconds), and
+ * `estimated` is a reconstruction used when a provider returned no usage at
+ * all. An estimate that is indistinguishable from a reported number is an
+ * estimate nobody can later reconcile or refund against.
+ */
+export declare const USAGE_SOURCES: readonly ["provider_reported", "oxy_measured", "estimated"];
+export declare const usageSourceSchema: z.ZodEnum<["provider_reported", "oxy_measured", "estimated"]>;
+/**
+ * A price for one unit, as `amount` per `per` units — `per` because a price
+ * quoted per single token would need more fractional digits than it is worth
+ * ("$3.00 per 1000000 input_tokens" is how every provider quotes it, and how
+ * every customer reads it).
+ */
+export declare const unitPriceSchema: z.ZodObject<{
+    unit: z.ZodEnum<["input_tokens", "cached_input_tokens", "output_tokens", "reasoning_tokens", "requests", "images", "audio_input_milliseconds", "audio_output_milliseconds", "video_milliseconds", "characters", "embeddings"]>;
+    amount: z.ZodBranded<z.ZodString, "ExactDecimal">;
+    per: z.ZodNumber;
+    currency: z.ZodString;
+}, "strict", z.ZodTypeAny, {
+    amount: string & z.BRAND<"ExactDecimal">;
+    currency: string;
+    unit: "input_tokens" | "cached_input_tokens" | "output_tokens" | "reasoning_tokens" | "requests" | "images" | "audio_input_milliseconds" | "audio_output_milliseconds" | "video_milliseconds" | "characters" | "embeddings";
+    per: number;
+}, {
+    amount: string;
+    currency: string;
+    unit: "input_tokens" | "cached_input_tokens" | "output_tokens" | "reasoning_tokens" | "requests" | "images" | "audio_input_milliseconds" | "audio_output_milliseconds" | "video_milliseconds" | "characters" | "embeddings";
+    per: number;
+}>;
+export type CurrencyCode = z.infer<typeof currencyCodeSchema>;
+export type ExactDecimal = z.infer<typeof exactDecimalSchema>;
+export type Money = z.infer<typeof moneySchema>;
+export type UsageUnit = z.infer<typeof usageUnitSchema>;
+export type UsageSource = z.infer<typeof usageSourceSchema>;
+export type UsageQuantity = z.infer<typeof usageQuantitySchema>;
+export type UnitPrice = z.infer<typeof unitPriceSchema>;