@kaana.ai/contracts 0.0.0-stage → 4.3.0

package/dist/public/money.js

@@ -0,0 +1,237 @@
+/**
+ * Money, prices and usage units.
+ *
+ * Moved from `@oxy.so/contracts` 4.12.0 (`src/inference/money.ts`) by Kaana ADR 0002.
+ * `priceSnapshotSchema` comes from `priceVersion.ts`; the price VERSION record
+ * itself stays in `@oxy.so/contracts`, because Oxy owns the price lists.
+ */
+import { z } from 'zod';
+/** ISO 4217 alpha-3 currency code, e.g. `USD`. */
+export const currencyCodeSchema = z
+    .string()
+    .regex(/^[A-Z]{3}$/, 'currency must be an ISO 4217 alpha-3 code');
+/**
+ * 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 $3 per million
+ * input tokens, one token costs `0.000003000000`, which this scale represents
+ * exactly. Amounts are compared and summed NUMERICALLY, never as text — `3.0`
+ * and `3.000000000000` are one amount written two ways.
+ */
+export const INFERENCE_MONEY_SCALE = 12;
+/**
+ * An exact non-negative decimal, carried as a STRING so no parse step can turn
+ * it into a float on the way past. Up to 18 integer digits and
+ * {@link INFERENCE_MONEY_SCALE} fractional digits.
+ *
+ * Non-negative: direction is carried by the SHAPE — a receipt debits, a refund
+ * credits — so a stray sign can never silently invert an entry.
+ *
+ * No exponent form: `1e-6` and `0.000001` are the same number, but only one of
+ * them survives a naive string comparison, a cache key or a log grep intact.
+ *
+ * Branded, so a bare `string` is not assignable and an amount cannot arrive
+ * from string concatenation that was never checked. Producers construct one
+ * with `exactDecimalSchema.parse(value)`.
+ */
+export const exactDecimalSchema = z
+    .string()
+    .regex(/^(?:0|[1-9][0-9]{0,17})(?:\.[0-9]{1,12})?$/, 'must be an exact non-negative decimal string without an exponent')
+    .brand();
+/**
+ * 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 const moneySchema = z
+    .object({
+    amount: exactDecimalSchema,
+    currency: currencyCodeSchema,
+})
+    .strict();
+/**
+ * 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
+ * ```
+ *
+ * **Audio tokens are siblings too** (added in contract set 3.2.0). A model that
+ * consumes or produces audio AS TOKENS — OpenAI's audio chat and Realtime
+ * models — reports them inside the same totals, and they are priced several
+ * times higher than text tokens, so folding them into `input_tokens` would
+ * under-charge by exactly the audio. They are subtracted out the same way:
+ *
+ * ```text
+ * cached_audio_input_tokens = prompt_tokens_details.cached_tokens_details.audio_tokens
+ * audio_input_tokens        = prompt_tokens_details.audio_tokens - cached_audio_input_tokens
+ * input_tokens              = prompt_tokens - (text AND audio cached) - audio_input_tokens
+ * audio_output_tokens       = completion_tokens_details.audio_tokens
+ * output_tokens             = completion_tokens - reasoning_tokens - audio_output_tokens
+ * ```
+ *
+ * `audio_input_milliseconds` and `audio_output_milliseconds` remain the units
+ * for audio metered by DURATION (file transcription, speech synthesis). A
+ * report carries one reading of a given audio, never both: the same audio
+ * reported as tokens and as milliseconds is the double charge this rule exists
+ * to prevent.
+ *
+ * **`session_milliseconds` is session wall-clock time** (added in contract set
+ * 3.3.0): wall-clock milliseconds a realtime session was open with the upstream
+ * provider; reported only by a provider that bills session time, and never
+ * together with a per-audio reading of the same span. xAI's Voice Agent API
+ * bills a session using server-side voice activity detection for its whole
+ * duration, speech or silence, which is neither audio sent nor audio received:
+ * reporting that clock as `audio_input_milliseconds` would misdescribe the
+ * charge, and reporting both would bill the same span twice. It is measured by
+ * the data plane from the upstream session's open to its close, never
+ * estimated from the audio, and a provider billed by the audio it carries (or
+ * by tokens) never reports it.
+ *
+ * **`input_transcription_milliseconds` is a second model's reading** (added in
+ * contract set 4.1.0): milliseconds of a realtime session's input audio that
+ * the deployment's own TRANSCRIPTION model transcribed, when the provider bills
+ * that model apart from the session model. OpenAI runs input transcription on
+ * a separate ASR model "billed according to the ASR model's pricing rather
+ * than the realtime model's", so the session model's `audio_input_tokens` and
+ * this unit are two charges for two models, each read once — which is why the
+ * partition rule holds: no other unit counts the transcription model's work.
+ * Reporting the ASR duration as `audio_input_milliseconds` beside the session
+ * model's audio tokens would be the "same audio as tokens and as milliseconds"
+ * this rule forbids, and folding it into the tokens would price it at the
+ * wrong model's rate. A provider whose session price already includes the
+ * transcription (xAI's Voice Agent) never reports it. A transcription model
+ * billed by TOKENS is a charge this unit does not express: the data plane
+ * declares only duration-billed transcription models and ends a session whose
+ * provider reports tokens for one, rather than approximate the charge.
+ *
+ * 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 const USAGE_UNITS = [
+    'input_tokens',
+    'cached_input_tokens',
+    'output_tokens',
+    'reasoning_tokens',
+    'requests',
+    'images',
+    'audio_input_milliseconds',
+    'audio_output_milliseconds',
+    'video_milliseconds',
+    'characters',
+    'embeddings',
+    'audio_input_tokens',
+    'cached_audio_input_tokens',
+    'audio_output_tokens',
+    'session_milliseconds',
+    'input_transcription_milliseconds',
+];
+export const usageUnitSchema = z.enum(USAGE_UNITS);
+/**
+ * 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 const usageQuantitySchema = z
+    .object({
+    unit: usageUnitSchema,
+    quantity: z.number().int().nonnegative().safe(),
+})
+    .strict();
+/**
+ * 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 const USAGE_SOURCES = ['provider_reported', 'oxy_measured', 'estimated'];
+export const usageSourceSchema = z.enum(USAGE_SOURCES);
+/**
+ * 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 const unitPriceSchema = z
+    .object({
+    unit: usageUnitSchema,
+    amount: exactDecimalSchema,
+    per: z.number().int().positive().safe(),
+    currency: currencyCodeSchema,
+})
+    .strict();
+/**
+ * The price snapshot a settled receipt keeps.
+ *
+ * The unit prices are COPIED onto the receipt, not just referenced, so a receipt
+ * remains readable even if the price version record is later archived, and so
+ * that a mistake in the copy is visible as a disagreement with the version it
+ * names rather than silently invisible.
+ */
+export const priceSnapshotSchema = z
+    .object({
+    priceVersionId: z.string().min(1).max(128),
+    currency: currencyCodeSchema,
+    unitPrices: z.array(unitPriceSchema).min(1),
+})
+    .strict();
+/**
+ * What a request consumed, as a public response reports it: one quantity per
+ * unit, each unit at most once, the units partitioning the request (see
+ * {@link USAGE_UNITS}). New in contract set 4.0.0.
+ */
+export const usageSchema = z.array(usageQuantitySchema).superRefine((units, ctx) => {
+    const names = units.map((quantity) => quantity.unit);
+    if (new Set(names).size !== names.length) {
+        ctx.addIssue({ code: 'custom', message: 'each unit is reported once, as a total' });
+    }
+});

package/dist/public/policy.d.ts

@@ -0,0 +1,19 @@
+/**
+ * The routing-policy reference a response and a receipt record.
+ *
+ * The policy itself is a Console resource and stays in `@oxy.so/contracts`; what
+ * a client reads back is only which of its own policy revisions served it.
+ *
+ * Moved from `@oxy.so/contracts` 4.12.0 (`src/inference/routingPolicy.ts`) by Kaana ADR 0002.
+ */
+import { z } from 'zod';
+/**
+ * The reference a request records: which policy, at which of the customer's own
+ * revisions. Embedded in the request envelope and in the settled receipt, so a
+ * charge can be explained against the exact configuration that produced it.
+ */
+export declare const routingPolicyReferenceSchema: z.ZodObject<{
+    routingPolicyId: z.ZodString;
+    policyVersion: z.ZodNumber;
+}, z.core.$strict>;
+export type RoutingPolicyReference = z.infer<typeof routingPolicyReferenceSchema>;

package/dist/public/policy.js

@@ -0,0 +1,20 @@
+/**
+ * The routing-policy reference a response and a receipt record.
+ *
+ * The policy itself is a Console resource and stays in `@oxy.so/contracts`; what
+ * a client reads back is only which of its own policy revisions served it.
+ *
+ * Moved from `@oxy.so/contracts` 4.12.0 (`src/inference/routingPolicy.ts`) by Kaana ADR 0002.
+ */
+import { z } from 'zod';
+/**
+ * The reference a request records: which policy, at which of the customer's own
+ * revisions. Embedded in the request envelope and in the settled receipt, so a
+ * charge can be explained against the exact configuration that produced it.
+ */
+export const routingPolicyReferenceSchema = z
+    .object({
+    routingPolicyId: z.string().min(1).max(128),
+    policyVersion: z.number().int().positive().safe(),
+})
+    .strict();