@urun-sh/openai 0.6.17 → 0.6.18

package/dist/proxy/index.d.cts

@@ -1,6 +1,6 @@
 import { Server, IncomingMessage } from 'node:http';
-import { f as ProxyHandlerOptions, P as ProxyClients } from '../responses-turn-Dd-h46sQ.cjs';
-export { D as DialObserver, a as DialReport, g as ProxyIdentity, h as ProxyVideoOutLane, S as SessionGoneError, U as UnknownModelError } from '../responses-turn-Dd-h46sQ.cjs';
+import { f as ProxyHandlerOptions, P as ProxyClients } from '../responses-turn-B91w2YKn.cjs';
+export { D as DialObserver, a as DialReport, g as ProxyIdentity, h as ProxyVideoOutLane, S as SessionGoneError, U as UnknownModelError } from '../responses-turn-B91w2YKn.cjs';
 export { h as GEMINI_LIVE_PATH, V as VIDEO_OUT_SETUP_KEY } from '../translator-CO8W_hmJ.cjs';
 import { WebSocket } from 'ws';
 import '../models-DUdx_Y6X.cjs';
@@ -339,6 +339,17 @@ interface RealtimeBinding<S = unknown> {
     };
     /** `model` the binding names for this conversation (overrides session config default). */
     model?: string;
+    /**
+     * The GA transcription-session intent (`?intent=transcription`): this
+     * connection is a TRANSCRIPTION session — transcription events flow
+     * without the client opting in, and `response.create` answers the loud
+     * `unsupported` error (the connection stays open) because there are no
+     * model turns to create. Set by the shared binding composition (binding.ts)
+     * only when the resolver's URL carried the intent AND the model's catalog
+     * row vouches stt (the binding carries `turns`). Absent = the realtime
+     * (voice) session, today's behaviour.
+     */
+    intent?: 'transcription';
     /** True when `session` is a resumed authorized session (vs a fresh private one). */
     resumed?: boolean;
     /**
@@ -350,8 +361,14 @@ interface RealtimeBinding<S = unknown> {
      * `turn_start` IS the native barge-in — the adapter drives server-VAD
      * emulation, real transcription events, and model-visible committed audio
      * from it. Absent = the backend wired no turn lane: manual mode only.
+     *
+     * The CONSUMER passes its own lifecycle `signal` (the realtime connection's
+     * turn-lane controller): the lane must END its stream subscription when it
+     * aborts — the transport races its pulls against it — so the subscription
+     * never outlives the consumer. A consumer that passes none gets an
+     * un-raced lane (its own teardown is then its own problem).
      */
-    turns?: () => AsyncIterable<{
+    turns?: (signal?: AbortSignal) => AsyncIterable<{
         kind: 'delta' | 'final' | 'turn_start' | 'turn_end' | 'turn_eager_end' | 'turn_resumed';
         text?: string;
         confidence?: number;
@@ -446,10 +463,39 @@ declare class OpenAIRealtimeConnection<S = unknown> {
     private closed;
     /** The backend's turn/transcript lane consumer (binding.turns), if wired. */
     private turnLaneAbort;
-    /** Accumulated transcript of the OPEN user turn (delta/final events). */
+    /**
+     * The CONSUMER-SIDE turn-lane iterator — held so teardown can end the lane
+     * subscription through the async-iterator return() protocol when it is
+     * parked at a YIELD (no pending pull for the abort to unwind; a plain
+     * for-await has no handle to close). Null once the consumer loop exits.
+     */
+    private turnLaneIterator;
+    /** Accumulated transcript of the OPEN user turn (delta/final events). The
+     * `final` flag says the engine finalized the accumulated text (its last
+     * transcript event was a `final`) — a manual commit may complete on it. */
     private turnTranscript;
     /** Transcript of the LAST committed turn (the model-visible commit). */
     private lastCommittedTranscript;
+    /**
+     * A manual transcription-session commit whose item is still awaiting the
+     * lane's final text (the item id of the committed item): interim deltas
+     * meanwhile carry THIS id, and the lane's next `final`/`turn_end`
+     * discharges exactly one `…transcription.completed` for it. No timers.
+     */
+    private pendingTranscriptItemId;
+    /**
+     * The engine turn a manual transcription-session commit ALREADY discharged
+     * (the discharged item's id): the lane's `final` completed the committed
+     * item (or the commit completed it on the lane's already-final text) while
+     * the engine's `turn_end` for THAT utterance is still to come. That
+     * `turn_end` is the boundary signal only — the server-side commit it
+     * would otherwise open mints a SECOND item for the same utterance and
+     * clears audio the client already appended for the next one. Consumed by
+     * that `turn_end` itself (one-shot); invalidated by the next `turn_start`
+     * (a new engine utterance opened — its boundary was not the discharged
+     * one).
+     */
+    private manuallyDischargedTurnItem;
     constructor(opts: {
         ws: WebSocket;
         binding: RealtimeBinding<S>;
@@ -464,8 +510,33 @@ declare class OpenAIRealtimeConnection<S = unknown> {
      * is the native barge-in, turn_end is the native commit, deltas/finals are
      * the REAL input transcription. This is the server-VAD primitive, not an
      * emulation gap: the voice engine itself decides turn boundaries.
+     *
+     * TEARDOWN ENDS THE SUBSCRIPTION FROM EVERY PARKING STATE, which a plain
+     * `for await` cannot do (its `break` only runs after the next event — a
+     * quiet engine parks the pull forever, and a `return()` enqueued behind a
+     * pending `next()` waits for it to settle):
+     *   - the `abort` threaded into `turns()` unwinds a PENDING pull (the
+     *     transport races its `messages()` pulls against it);
+     *   - the HELD consumer iterator's `return()` (teardown) closes a lane
+     *     parked at a YIELD with no pending pull.
      */
     private startTurnLane;
+    /** True when this connection IS a transcription session (the GA verdict:
+     * the binding's `?intent=transcription`, or the client's own
+     * `session.type: "transcription"` — the latter only accepted behind a
+     * turns lane, so a transcription session always has a transcript source). */
+    private isTranscriptionSession;
+    /** True when input-transcription events flow: the client opted in on a
+     * voice session, or ALWAYS on a transcription session (no opt-in needed). */
+    private transcriptionEventsOn;
+    /**
+     * Discharge a pending manual transcription commit: the ONE
+     * `…transcription.completed` for the committed item, carrying the text the
+     * lane delivered for its audio, and a reset shared accumulator (the
+     * utterance is client-delivered; the next utterance starts clean). The
+     * lane's own boundary events drive this — no timers anywhere.
+     */
+    private completePendingTranscription;
     /** One backend turn/transcript event → REAL GA protocol events. */
     private onTurnEvent;
     /** The committed turn's transcript, consumed once (null = none). */
@@ -647,8 +718,21 @@ declare function realtimeModelOf(req: IncomingMessage): string | undefined;
  *     (`sessionHandle` once → `openAudioOn(handle, signal)` +
  *     `createResponseOn(handle, …)`), so no turn on this connection can ever
  *     resolve a different pooled session than the bridge.
+ *   - An stt verdict (the catalog `task` gate `supportsTranscription`, the
+ *     SAME router-backed oracle the /v1/audio/transcriptions lane rides) with
+ *     no `openTurnsOn` seam ⇒ 501 — a transcription-capable binding without
+ *     its transcript source is never wired silently. A row that OVERRIDES
+ *     the platform's default transcript stream (`engine_args.transcript_stream`,
+ *     resolved through `transcriptStreamFor` when the seam is wired) is the
+ *     same-shape 501 — the lane never listens on a stream the engine does
+ *     not write.
+ *   - `?intent=transcription` on anything but an stt row ⇒ 400 naming the
+ *     model; any other intent value ⇒ 400 naming the accepted values. Every
+ *     refusal lands BEFORE the pooled session is resolved (`sessionHandle`)
+ *     and the bridge opens — a refused upgrade admits nothing. On an stt row
+ *     the intent is recorded on the binding (a transcription session).
  */
-declare function realtimeBindingFor(clients: Pick<ProxyClients, 'sessionHandle' | 'createResponseOn'> & Partial<Pick<ProxyClients, 'supportsAudio' | 'openAudioOn'>>, model: string | undefined, signal?: AbortSignal): Promise<RealtimeBinding<ProxyClients>>;
+declare function realtimeBindingFor(clients: Pick<ProxyClients, 'sessionHandle' | 'createResponseOn'> & Partial<Pick<ProxyClients, 'supportsAudio' | 'openAudioOn' | 'supportsTranscription' | 'openTurnsOn' | 'transcriptStreamFor'>>, model: string | undefined, signal?: AbortSignal, intent?: string): Promise<RealtimeBinding<ProxyClients>>;
 /**
  * The LOCAL lane resolver: fixed pooled clients, optional fixed apiKey (the
  * same shared-secret rule the local Gemini Live lane applies — a mismatched

package/dist/proxy/index.d.ts

@@ -1,6 +1,6 @@
 import { Server, IncomingMessage } from 'node:http';
-import { f as ProxyHandlerOptions, P as ProxyClients } from '../responses-turn-B0zlsQGn.js';
-export { D as DialObserver, a as DialReport, g as ProxyIdentity, h as ProxyVideoOutLane, S as SessionGoneError, U as UnknownModelError } from '../responses-turn-B0zlsQGn.js';
+import { f as ProxyHandlerOptions, P as ProxyClients } from '../responses-turn-FcEDaymF.js';
+export { D as DialObserver, a as DialReport, g as ProxyIdentity, h as ProxyVideoOutLane, S as SessionGoneError, U as UnknownModelError } from '../responses-turn-FcEDaymF.js';
 export { h as GEMINI_LIVE_PATH, V as VIDEO_OUT_SETUP_KEY } from '../translator-Ddd65sXR.js';
 import { WebSocket } from 'ws';
 import '../models-DUdx_Y6X.js';
@@ -81,9 +81,11 @@ interface RealtimeBinding<S = unknown> {
     
     model?: string;
     
+    intent?: 'transcription';
+    
     resumed?: boolean;
     
-    turns?: () => AsyncIterable<{
+    turns?: (signal?: AbortSignal) => AsyncIterable<{
         kind: 'delta' | 'final' | 'turn_start' | 'turn_end' | 'turn_eager_end' | 'turn_resumed';
         text?: string;
         confidence?: number;
@@ -128,9 +130,15 @@ declare class OpenAIRealtimeConnection<S = unknown> {
     
     private turnLaneAbort;
     
+    private turnLaneIterator;
+    
     private turnTranscript;
     
     private lastCommittedTranscript;
+    
+    private pendingTranscriptItemId;
+    
+    private manuallyDischargedTurnItem;
     constructor(opts: {
         ws: WebSocket;
         binding: RealtimeBinding<S>;
@@ -142,6 +150,12 @@ declare class OpenAIRealtimeConnection<S = unknown> {
     
     private startTurnLane;
     
+    private isTranscriptionSession;
+    
+    private transcriptionEventsOn;
+    
+    private completePendingTranscription;
+    
     private onTurnEvent;
     
     private takeTurnTranscript;
@@ -188,7 +202,7 @@ declare class OpenAIRealtimeConnection<S = unknown> {
 
 declare function realtimeModelOf(req: IncomingMessage): string | undefined;
 
-declare function realtimeBindingFor(clients: Pick<ProxyClients, 'sessionHandle' | 'createResponseOn'> & Partial<Pick<ProxyClients, 'supportsAudio' | 'openAudioOn'>>, model: string | undefined, signal?: AbortSignal): Promise<RealtimeBinding<ProxyClients>>;
+declare function realtimeBindingFor(clients: Pick<ProxyClients, 'sessionHandle' | 'createResponseOn'> & Partial<Pick<ProxyClients, 'supportsAudio' | 'openAudioOn' | 'supportsTranscription' | 'openTurnsOn' | 'transcriptStreamFor'>>, model: string | undefined, signal?: AbortSignal, intent?: string): Promise<RealtimeBinding<ProxyClients>>;
 
 declare function fixedClientsResolver(clients: ProxyClients, apiKey?: string): {
     resolve(req: IncomingMessage, signal?: AbortSignal): Promise<RealtimeBinding<ProxyClients>>;

package/dist/proxy/index.js

@@ -1 +1 @@
-import{B as l,C as a,D as x,E as s,F as O,G as f,S as I,a as n,b as p,c as m}from"../chunk-U7EJSC2N.js";import{d as r,e as i}from"../chunk-TMQJY56F.js";import{a as o,h as t}from"../chunk-3JWYIKHM.js";import"../chunk-6M5JK4YC.js";import{c as e}from"../chunk-YSFSRI3D.js";e();export{o as GEMINI_LIVE_PATH,a as OPENAI_REALTIME_PATH,l as OpenAIRealtimeConnection,n as ResumptionRegistry,i as SessionGoneError,r as UnknownModelError,t as VIDEO_OUT_SETUP_KEY,p as attachGeminiLive,x as attachOpenAIRealtime,I as createOpenAIProxy,f as fixedClientsResolver,O as realtimeBindingFor,s as realtimeModelOf,m as serveGeminiLiveSocket};
+import{B as l,C as a,D as x,E as s,G as O,H as f,T as I,a as n,b as p,c as m}from"../chunk-3U7YQ7LE.js";import{f as r,g as i}from"../chunk-NYSZ3RGB.js";import{a as o,h as t}from"../chunk-3JWYIKHM.js";import"../chunk-6M5JK4YC.js";import{c as e}from"../chunk-YSFSRI3D.js";e();export{o as GEMINI_LIVE_PATH,a as OPENAI_REALTIME_PATH,l as OpenAIRealtimeConnection,n as ResumptionRegistry,i as SessionGoneError,r as UnknownModelError,t as VIDEO_OUT_SETUP_KEY,p as attachGeminiLive,x as attachOpenAIRealtime,I as createOpenAIProxy,f as fixedClientsResolver,O as realtimeBindingFor,s as realtimeModelOf,m as serveGeminiLiveSocket};

package/dist/{responses-turn-Dd-h46sQ.d.cts → responses-turn-B91w2YKn.d.cts}

@@ -381,7 +381,8 @@ interface ModelPriceRow {
 }
 
 /**
- * THE OPENROUTER PROVIDER DOCUMENT — `GET /v1/models?format=openrouter`.
+ * THE OPENROUTER PROVIDER DOCUMENT — `GET /partners/openrouter/models`
+ * (`GET /v1/models?format=openrouter` answers 308 here for one release).
  *
  * OpenRouter's provider monitor polls this document (schema 2.4, per
  * openrouter.ai/docs/guides/community/for-providers §1) to list a provider's
@@ -610,28 +611,30 @@ declare class UnknownModelError extends Error {
 declare class SessionGoneError extends Error {
 }
 /**
- * OpenAI-shaped model list (same shape as models.ts listModels). When the
- * catalog oracle is configured, each entry ALSO carries the catalog
- * enrichment fields (additive JSON — OpenAI clients ignore unknown fields;
- * the OpenRouter + Vercel AI Gateway provider listings need them). Entries
- * whose slug matches no catalog row stay at the bare four fields.
+ * One `/v1/models` row — a CLOSED allow-list of externally contracted keys
+ * (owner ruling, ENG-472: publish what an external contract names, never the
+ * implementation). The four OpenAI `Model` fields, plus — only when known —
+ * the Hugging Face pair (`context_length`, `pricing`), the TTS `voices` and
+ * the `capabilities` object. `name`, `description`, `task`, `engine`,
+ * `gpu_spec` and `warm` are GONE from the type so they cannot be set; every
+ * row builder goes through `publicModelEntry` (openrouter-doc.ts), which
+ * copies exactly these keys.
  */
 interface RouterModelEntry {
+    /** OpenAI `Model`: the model identifier callers reference in the API. */
     id: string;
+    /** OpenAI `Model`: always "model". */
     object: 'model';
+    /** OpenAI `Model`: the Unix time the model id first appeared in the uRun catalog (0 until the catalog carries it). */
     created: number;
+    /** OpenAI `Model`: the organization that owns the model. */
     owned_by: string;
-    /** Display name — the canonical `<model_id>:<variant>` catalog ref. */
-    name?: string;
-    /** User-facing description from the catalog (console Endpoints copy). */
-    description?: string;
-    /** Catalog modality (chat | code | agent | vl | audio | image | ...). */
-    task?: string;
-    /** Serving engine (vllm | sglang | llamacpp | ...). */
-    engine?: string;
-    /** The catalog lane this app deploys, e.g. 'rtx6000:1'. */
-    gpu_spec?: string;
-    /** Context window in tokens (chat rows carry it in engine_args). */
+    /**
+     * Hugging Face provider comparison table — its top-level `context_length`
+     * read (HF also gates `:fastest`/`:cheapest` on it). Only when the catalog
+     * states it and the placements agree (openrouter-doc.ts
+     * `placementsContextLength`).
+     */
     context_length?: number;
     /**
      * Per-token price, USD per MILLION tokens — the exact shape Hugging Face
@@ -641,8 +644,6 @@ interface RouterModelEntry {
      * a zero or an estimate here would be a silent lie on a public table.
      */
     pricing?: TokenPricing;
-    /** Idle-floor replica count for a shared-lane model (U6 reconciler). */
-    warm?: number;
     /**
      * The voices a `tts`-task model synthesizes under (ENG-331 — OpenAI's
      * speech API names a voice per request, so a caller needs the list).
@@ -664,7 +665,7 @@ interface RouterModelEntry {
      * whose placements disagree, a malformed value, or a slug with no catalog
      * row omit the object entirely — unknown, never guessed
      * (openrouter-doc.ts `placementsReasoning`, applied by both listing
-     * lanes).
+     * lanes). The conformance kit's voice_profile check reads it.
      */
     capabilities?: {
         reasoning: boolean;
@@ -674,10 +675,11 @@ interface RouterModelList {
     object: 'list';
     data: RouterModelEntry[];
 }
-/** One `/v1/images/models` entry: the OpenAI model object plus the image
+/** One `/v1/images/models` entry: the allow-listed model row plus the image
  * modes the catalog rows vouch for (the SAME per-placement consensus
  * {@link ModelRouter.supportsImages} applies — a row set that vouches
- * nothing is NOT listed). */
+ * nothing is NOT listed). `image_modes` is allow-listed for THIS listing
+ * only (publicModelEntry's image overload). */
 interface RouterImageModelEntry extends RouterModelEntry {
     image_modes: ImageCapabilityMode[];
 }
@@ -1102,7 +1104,8 @@ declare class ModelRouter<S> {
      */
     modelList(): Promise<RouterModelList>;
     /**
-     * The OpenRouter PROVIDER document (`GET /v1/models?format=openrouter`):
+     * The OpenRouter PROVIDER document (`GET /partners/openrouter/models`;
+     * `GET /v1/models?format=openrouter` 308-redirects there for one release):
      * schema 2.4 per openrouter.ai/docs/guides/community/for-providers §1 —
      * "an endpoint that returns all models that should be served by
      * OpenRouter". That is the SHARED block's sellable endpoints (the models
@@ -1218,6 +1221,25 @@ declare class ModelRouter<S> {
      * an infrastructure outage.
      */
     supportsTranscription(model: string | undefined): Promise<boolean>;
+    /**
+     * THE REALTIME TRANSCRIPTION LANE'S STREAM-NAME RESOLUTION (openai-realtime
+     * binding.ts, ENG-461): the §5 stream the model's engine emits its own
+     * transcript events on — `engine_args.transcript_stream`, defaulting to
+     * the platform's {@link STT_TRANSCRIPT_STREAM} when no row carries the
+     * field, over the SAME catalog join {@link supportsTranscription} uses
+     * (exact `(model_id, variant)` for a shared-lane hit,
+     * {@link catalogRowsForSlug} for a caller-org app). Deliberately NOT
+     * task-gated: the binding consults it only behind a `transcribes` verdict,
+     * and the resolution itself is a pure field consensus. The lane's consume
+     * side reads the PLATFORM DEFAULT only — a row that OVERRIDES the default
+     * is the binding's loud 501 naming the field, never a silent listen on a
+     * stream the engine does not write; a catalog that cannot vow ONE name
+     * (disagreeing placements, a non-string or empty value) is the loud
+     * defect, never a guess. Undeployed uRun-namespace models throw
+     * {@link UnknownModelError}, and a CONFIGURED catalog oracle that FAILS
+     * propagates — never a fabricated default.
+     */
+    transcriptStreamFor(model: string | undefined): Promise<string>;
     /**
      * THE SPEECH LANE'S VOICE-SET RESOLUTION (proxy/speech.ts): the voice set
      * `/v1/models` publishes for `model` — resolved by the SAME resolver the
@@ -1268,6 +1290,43 @@ declare class ModelRouter<S> {
     closeAll(): Promise<void>;
 }
 
+/**
+ * THE TURN/TRANSCRIPT LANE over the named §5 `stt` data stream — the
+ * engine's OWN transcript source for the OpenAI Realtime transcription
+ * intent (ENG-461).
+ *
+ * The delivery path is the platform's existing one (urun-python
+ * `urun.serve.transcribe_bridge`): the streaming STT engine emits its own
+ * transcript events on the session-scoped named stream {@link
+ * STT_TRANSCRIPT_STREAM} — the SAME §5 named-DATA mechanism the text-lane
+ * token deltas ride, in the text lane's own wire shape (`{"t": kind,
+ * "delta": text, "start_ms": int}`). Nothing here invents a framing: the
+ * lane is `session.stream('stt').messages()` (core `SessionStream.messages`,
+ * contract §5 — the downstream twin of the `emit` produce seam), and this
+ * module only DECODES each payload onto the `RealtimeBinding.turns` event
+ * shape the realtime adapter already consumes.
+ *
+ * `confidence` in the wire payload is SEMANTIC END-OF-TURN confidence (the
+ * engine's own VAD verdict that the user finished speaking), NOT
+ * transcription confidence — carrying it onto the GA surface would
+ * misrepresent it, so the decoder DROPS it by design (the adapter reads the
+ * boundary from `turn_end`, which is the same verdict as an event).
+ */
+
+/** The turn kinds the transcribe bridge emits (TranscriptEvent kinds). */
+type SessionTurnKind = 'delta' | 'final' | 'turn_start' | 'turn_end' | 'turn_eager_end' | 'turn_resumed';
+/**
+ * One decoded transcript-lane event — the `RealtimeBinding.turns` event
+ * shape. `text` is the wire `delta` field (a transcript PIECE — the adapter
+ * accumulates), `startMs` the additive `start_ms` clock. The wire
+ * `confidence` (semantic end-of-turn confidence) is deliberately absent.
+ */
+interface SessionTurnLaneEvent {
+    kind: SessionTurnKind;
+    text?: string;
+    startMs?: number;
+}
+
 /** One ready replica row from the presence read. */
 interface RuntimePresenceRow {
     runtime_id: string;
@@ -2106,9 +2165,10 @@ interface ProxyClients {
     /** `listModels(...)` result (an OpenAI model list object). */
     listModels(): Promise<unknown>;
     /**
-     * The OpenRouter provider document (`GET /v1/models?format=openrouter`).
+     * The OpenRouter provider document (`GET /partners/openrouter/models`; the
+     * old `GET /v1/models?format=openrouter` spelling 308-redirects there).
      * Optional at the seam: hand-built test clients may omit it, in which case
-     * the format=openrouter branch answers 501 — loud, never a silent empty.
+     * that route answers 501 — loud, never a silent empty.
      */
     openRouterModels?(): Promise<unknown>;
     /**
@@ -2286,6 +2346,46 @@ interface ProxyClients {
       * queued admission going back only at the LAST waiter's abort.
       */
     openAudioOn?(handle: string, signal?: AbortSignal): Promise<PinnedAudioLane>;
+    /**
+     * THE TURN/TRANSCRIPT LANE OPEN (the realtime transcription intent,
+     * ENG-461): the STT engine's OWN transcript events — the §5 `stt` named
+     * data stream (urun.serve.transcribe_bridge's wire shape: `t` =
+     * delta/final/turn_start/turn_end/turn_eager_end/turn_resumed, `delta`
+     * text pieces, additive `start_ms`), consumed off the EXACT pooled session
+     * the handle names (ModelRouter.sessionForHandle — a gone or replaced
+     * session throws {@link SessionGoneError} LOUDLY, never a fresh session
+     * opened behind the handle's back) and decoded onto the
+     * `RealtimeBinding.turns` event shape (the wire's semantic-VAD
+     * `confidence` is end-of-turn confidence, NOT transcription confidence —
+     * the lane deliberately does not carry it). One lane per call; the
+     * realtime binding calls it once per connection. The returned iterable is
+     * lazy. `signal` is the CONSUMER's own lifecycle (the realtime
+     * connection's turn-lane controller, aborted at its teardown): the
+     * transport races every `messages()` pull against it and ends the stream
+     * subscription on every exit path, so the lane never outlives its
+     * consumer — a quiet engine cannot pin it open behind a never-settling
+     * pull. Optional at the seam exactly like
+     * {@link openAudioOn}: the binding's stt verdict answers its absence with
+     * the loud 501 — never a transcription-capable binding with no transcript
+     * source. An already-aborted signal rejects on the first pull — never
+     * opens a lane for a vanished consumer.
+     */
+    openTurnsOn?(handle: string, signal?: AbortSignal): AsyncIterable<SessionTurnLaneEvent>;
+    /**
+     * THE REALTIME TRANSCRIPTION LANE'S STREAM-NAME RESOLUTION (openai-realtime
+     * binding.ts): the §5 transcript stream the model's catalog rows vow
+     * (ModelRouter.transcriptStreamFor — `engine_args.transcript_stream`,
+     * defaulting to the platform's `stt`). The lane's consume side reads the
+     * platform DEFAULT stream name only, so the binding answers a row that
+     * OVERRIDES the default with the loud 501 naming the field — never a
+     * silent listen on a stream the engine does not write, and never a
+     * transcript-less transcription session. Optional at the seam exactly
+     * like {@link supportsTranscription}: hand-built clients own their own
+     * lane (their `openTurnsOn` decides what to read — no catalog row exists
+     * to check), so its absence skips the catalog check; every router-backed
+     * backhaul wires it.
+     */
+    transcriptStreamFor?(model: string | undefined): Promise<string>;
     /**
      * Open (or reuse) the NATIVE video FRAME lane on the pooled session `model`
      * routes to (transport/media.ts `enableSessionVideo`: discrete JPEG frames →

package/dist/{responses-turn-B0zlsQGn.d.ts → responses-turn-FcEDaymF.d.ts}

@@ -178,27 +178,19 @@ declare class SessionGoneError extends Error {
 }
 
 interface RouterModelEntry {
-    id: string;
-    object: 'model';
-    created: number;
-    owned_by: string;
     
-    name?: string;
-    
-    description?: string;
+    id: string;
     
-    task?: string;
+    object: 'model';
     
-    engine?: string;
+    created: number;
     
-    gpu_spec?: string;
+    owned_by: string;
     
     context_length?: number;
     
     pricing?: TokenPricing;
     
-    warm?: number;
-    
     voices?: string[];
     
     capabilities?: {
@@ -346,6 +338,8 @@ declare class ModelRouter<S> {
     
     supportsTranscription(model: string | undefined): Promise<boolean>;
     
+    transcriptStreamFor(model: string | undefined): Promise<string>;
+    
     speechVoices(model: string | undefined): Promise<SpeechVoicesResolution>;
     
     imageModelList(): Promise<RouterImageModelList>;
@@ -355,6 +349,14 @@ declare class ModelRouter<S> {
     closeAll(): Promise<void>;
 }
 
+type SessionTurnKind = 'delta' | 'final' | 'turn_start' | 'turn_end' | 'turn_eager_end' | 'turn_resumed';
+
+interface SessionTurnLaneEvent {
+    kind: SessionTurnKind;
+    text?: string;
+    startMs?: number;
+}
+
 interface RuntimePresenceRow {
     runtime_id: string;
     
@@ -615,6 +617,10 @@ interface ProxyClients {
     
     openAudioOn?(handle: string, signal?: AbortSignal): Promise<PinnedAudioLane>;
     
+    openTurnsOn?(handle: string, signal?: AbortSignal): AsyncIterable<SessionTurnLaneEvent>;
+    
+    transcriptStreamFor?(model: string | undefined): Promise<string>;
+    
     openVideo?(model: string | undefined): Promise<ProxyVideoLane>;
     
     openVideoOut?(model: string | undefined): Promise<ProxyVideoOutLane>;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@urun-sh/openai",
-  "version": "0.6.17",
+  "version": "0.6.18",
   "description": "OpenAI-compatible Realtime + Responses SDK over uRun session primitives (not websockets).",
   "type": "module",
   "main": "./dist/index.cjs",
@@ -70,7 +70,7 @@
     "build:bin": "bun build --compile --external @roamhq/wrtc src/proxy/bin.ts --outfile dist/bin/urun-openai && node scripts/assert-bin-transport.mjs"
   },
   "peerDependencies": {
-    "@urun-sh/core": "^0.6.17"
+    "@urun-sh/core": "^0.6.18"
   },
   "dependencies": {
     "@prometheus-io/client": "^0.16.1",

package/dist/chunk-6SN4EJ7Q.js

@@ -1,7 +0,0 @@
-import{a as j,b as W}from"./chunk-VRCCLQ5P.js";import{C as oe,D as se,E as ie,F as ae,H as le,I as A,J as ce,K as ue,L as de,M as he,N as pe,O as ge,P as fe,Q as me,R as ye,b as ee,d as te,e as ne,q as re,r as O}from"./chunk-U7EJSC2N.js";import{e as X,g as z}from"./chunk-TMQJY56F.js";import{a as Z}from"./chunk-3JWYIKHM.js";import{c as p}from"./chunk-YSFSRI3D.js";p();import{Counter as S,Gauge as we,Histogram as be,Registry as Ue,collectDefaultMetrics as Pe}from"@prometheus-io/client";import{createServer as $e}from"http";var Tt=9464,ve="/metrics",I="(other)",Le="(none)",He=64,Me=128,De=["/v1/responses","/v1/chat/completions","/v1/systemone","/v1/messages","/v1/messages/count_tokens","/v1/models","/v1/images/models","/v1/images/generations","/v1/images/edits","/v1/audio/speech","/v1/audio/models","/v1/audio/transcriptions",le,ue],Fe=new Set(De);function Ke(t){return Fe.has(t)?t:I}function Ge(t){return t===null?"none":t>=500?"5xx":t>=400?"4xx":t>=300?"3xx":t>=200?"2xx":"1xx"}var T=class{constructor(e){this.max=e}max;admitted=new Set;overflows=0;of(e,n=!0){return e===null||e===""?Le:this.admitted.has(e)?e:n?this.admitted.size>=this.max?(this.overflows+=1,I):(this.admitted.add(e),e):I}get size(){return this.admitted.size}get overflowCount(){return this.overflows}},N=class{registry=new Ue;modelLabel=new T(He);orgLabel=new T(Me);requests;duration;ttft;streamErrors;undelivered;ledgerWrites;ledgerQueue;labelCardinality;labelOverflows;constructor(e={}){this.requests=new S({name:"urun_inference_requests_total",help:"Completed /v1 requests, by lane, method, HTTP status class, streaming, model and org.",labelNames:["lane","method","status_class","stream","model","org"],registers:[this.registry]}),this.duration=new be({name:"urun_inference_request_duration_seconds",help:"Wall time from request receipt to response close. For a stream this is the WHOLE generation, not the time to the head.",labelNames:["lane","stream","model","org"],buckets:[.1,.25,.5,1,2.5,5,10,30,60,120,300,600],registers:[this.registry]}),this.ttft=new be({name:"urun_inference_ttft_seconds",help:"Time to first output-text delta on a streamed response. The metric Hugging Face gates provider listing on (<5s for conversational models).",labelNames:["lane","model","org"],buckets:[.1,.25,.5,1,2,3,5,8,15,30,60],registers:[this.registry]}),this.streamErrors=new S({name:"urun_inference_stream_errors_total",help:"Responses that emitted an error frame AFTER the head flushed. These are HTTP 200 with a complete body, so no status-based metric can see them.",labelNames:["lane","model"],registers:[this.registry]}),this.undelivered=new S({name:"urun_inference_undelivered_total",help:"Responses whose socket did not finish flushing \u2014 the client vanished mid-answer.",labelNames:["lane","model"],registers:[this.registry]}),this.ledgerWrites=new S({name:"urun_inference_ledger_writes_total",help:"Terminal fate of the per-request ledger row for each billable /v1 request. outcome=written is the row landing (reason=priced|unpriced); outcome=skipped and outcome=failed are BOTH revenue that went unrecorded. failed/duplicate is the ENG-410 fabricated-row signal.",labelNames:["outcome","reason"],registers:[this.registry]});for(let n of de)this.ledgerWrites.inc({outcome:n.outcome,reason:n.reason},0);this.ledgerQueue=new we({name:"urun_inference_ledger_queue",help:'The ledger write queue right now. measure="depth" is rows waiting for a worker; measure="oldest_age_seconds" is how long the head has waited \u2014 approaching the dwell bound means rows are about to be given up on, which is revenue lost.',labelNames:["measure"],registers:[this.registry],collect:()=>{this.ledgerQueue.set({measure:"depth"},he()),this.ledgerQueue.set({measure:"oldest_age_seconds"},pe()/1e3)}}),this.labelCardinality=new we({name:"urun_inference_label_cardinality",help:"Distinct values admitted for a bounded metric label. At its ceiling, further values report as (other).",labelNames:["label"],registers:[this.registry],collect:()=>{this.labelCardinality.set({label:"model"},this.modelLabel.size),this.labelCardinality.set({label:"org"},this.orgLabel.size)}}),this.labelOverflows=new S({name:"urun_inference_label_overflows_total",help:"Observations that would have taken a new label value but found the budget spent. Non-zero means a metric dimension is now partly blind.",labelNames:["label"],registers:[this.registry]}),e.defaultMetrics!==!1&&Pe({register:this.registry,prefix:"urun_inference_proxy_"})}observeLedger(e){this.ledgerWrites.inc({outcome:e.outcome,reason:e.reason})}observe(e){let n=Ke(e.path),o=e.status!==null&&!(e.status>=400&&e.status<500),r=this.modelLabel.overflowCount+this.orgLabel.overflowCount,a=this.modelLabel.of(e.model,o),s=this.orgLabel.of(e.org,o);this.modelLabel.overflowCount+this.orgLabel.overflowCount-r>0&&(this.modelLabel.overflowCount>r?this.labelOverflows.inc({label:"model"}):this.labelOverflows.inc({label:"org"}));let l=e.stream===!0?"true":"false";this.requests.inc({lane:n,method:e.method===""?I:e.method,status_class:Ge(e.status),stream:l,model:a,org:s}),this.duration.observe({lane:n,stream:l,model:a,org:s},e.duration_ms/1e3),e.ttft_ms!==null&&this.ttft.observe({lane:n,model:a,org:s},e.ttft_ms/1e3),e.stream_error&&this.streamErrors.inc({lane:n,model:a}),e.delivered||this.undelivered.inc({lane:n,model:a})}scrape(){return this.registry.metrics()}get contentType(){return this.registry.contentType}};function _e(t){return $e((e,n)=>{(async()=>{let o=(e.url??"/").split("?")[0];if(e.method!=="GET"||o!==ve){n.writeHead(404,{"content-type":"text/plain"}),n.end(`no route for ${e.method} ${o} \u2014 this port serves only GET ${ve}
-`);return}let r=await t.scrape();n.writeHead(200,{"content-type":t.contentType}),n.end(r)})().catch(o=>{if(!n.headersSent){n.writeHead(500,{"content-type":"text/plain"}),n.end(`${String(o instanceof Error?o.message:o)}
-`);return}n.end()})})}p();import{createHash as Se}from"crypto";import{createServer as Be}from"http";var Lt=9465,ze="x-urun-key-id",je="x-urun-key-fingerprint",We="urun-ratelimit-bucket:",Je="urun-ratelimit-tier:",Ee="anonymous",qe=/^urun_sk_[A-Za-z0-9_-]{43}$/;function Ve(t){return Se("sha256").update(`${We}${t}`).digest("hex")}function Ye(t){return Se("sha256").update(`${Je}${t}`).digest("hex")}function Qe(t){let e=/^Bearer[ \t]+(\S.*)$/i.exec((t??"").trim());if(!e)return null;let n=e[1].trim();return qe.test(n)?n:null}function Re(){return Be((t,e)=>{let n=Qe(t.headers.authorization);e.setHeader(ze,n===null?Ee:Ve(n)),e.setHeader(je,n===null?Ee:Ye(n)),e.writeHead(200),e.end()})}p();import{createClientToken as Xe,ClientTokenError as Ze}from"@urun-sh/core";var m=class extends Error{constructor(n,o=401){super(n);this.status=o;this.name="ProxyAuthError"}status},_=class extends Error{constructor(e){super(e),this.name="ControlPlaneUnavailableError"}};function R(t){let e=(t??"").trim();if(e==="")throw new m("missing Authorization header \u2014 send your uRun org API key as `Authorization: Bearer <key>`");let n=/^Bearer[ \t]+(\S.*)$/i.exec(e);if(!n)throw new m("malformed Authorization header \u2014 expected `Authorization: Bearer <urun org api key>`");return n[1].trim()}var et=60,tt=6e4,nt=4096;function rt(t){return t===401||t===403}var U=class{constructor(e){this.opts=e;this.mint=e.mint??Xe,this.now=e.now??Date.now}opts;cache=new Map;mint;now;async resolve(e){if(e==="")throw new m("empty Authorization bearer token");let n=this.cache.get(e);if(n&&this.now()-n.at<tt)return{apiKey:e,orgId:n.orgId};let o;try{o=(await this.mint(e,{baseUrl:this.opts.gatewayUrl,expiresIn:et})).orgId}catch(r){throw r instanceof Ze&&rt(r.status)?(this.cache.delete(e),new m("the uRun org API key was rejected by the control plane \u2014 check the key and that it is not revoked")):new _(`could not verify the API key against the uRun control plane at ${this.opts.gatewayUrl}: ${r instanceof Error?r.message:String(r)}`)}if(!o)throw new _("the control plane verified the API key but returned no org_id \u2014 refusing to route a request with no tenancy (control-plane contract violation)");if(this.cache.size>=nt){let r=this.cache.keys().next();r.done||this.cache.delete(r.value)}return this.cache.set(e,{orgId:o,at:this.now()}),{apiKey:e,orgId:o}}clear(){this.cache.clear()}};p();var ot=8080,P="serve",Bt="0.0.0.0";function st(t,e,n){let o=(t[e]??"").trim();if(!o)throw new Error(`${e} is not set \u2014 ${n}`);return o}function zt(t=process.env){let e=(t.PORT??"").trim(),n=e===""?ot:Number(e);if(!Number.isInteger(n)||n<0||n>65535)throw new Error(`PORT must be an integer in 0..65535, got ${JSON.stringify(e)}`);let o=re(O(st(t,"URUN_BASE_URL","the hosted proxy needs the uRun session-gateway base it opens sessions and verifies API keys against (e.g. https://api.urun.sh)"),"URUN_BASE_URL")),r=(t.URUN_API_URL??"").trim(),a=r?O(r,"URUN_API_URL"):`${o.replace(/\/+$/,"")}/v1`,s=(t.URUN_CATALOG_URL??"").trim(),i=(t.URUN_CATALOG_ANON_KEY??"").trim();if(!!s!=!!i)throw new Error("URUN_CATALOG_URL and URUN_CATALOG_ANON_KEY must be set together (or both left unset) \u2014 exactly one of them is set, which would silently disable the model-catalog oracle");let l=s?{catalogUrl:O(s,"URUN_CATALOG_URL"),anonKey:i}:null,u=(t.URUN_SESSION_TOKEN_SECRET??"").trim(),f=(t.URUN_SHARED_ORG_API_KEY??"").trim(),x=(t.URUN_SHARED_ORG_ID??"").trim();if(!!f!=!!x)throw new Error("URUN_SHARED_ORG_API_KEY and URUN_SHARED_ORG_ID must be set together (or both left unset) \u2014 exactly one of them is set, which would silently leave the shared endpoints on the session lane");return{port:n,baseUrl:o,apiUrl:a,catalog:l,decisionSecret:u,sharedOrgKey:f,sharedOrgId:x}}p();import{createHash as it,randomUUID as at}from"crypto";var lt=512;function M(t){return`urun-inference-proxy-${it("sha256").update(t).digest("hex").slice(0,32)}`}var $=class{constructor(e){this.opts=e}opts;tenants=new Map;label(e,n){return Object.assign(Object.create(e),{tenant:n.orgId})}build(e){if(this.opts.build){let i=this.opts.build(e);return{router:i.router,clients:this.label(i.clients,e)}}let n=0,o=M(e.apiKey),r=()=>{n+=1},a=j({baseUrl:this.opts.baseUrl,orgId:e.orgId,defaultApp:null,fnName:P,sharedOnlyListing:!0,apiUrl:this.opts.apiUrl,auth:{lane:"api-key",apiKey:e.apiKey,gatewayUrl:this.opts.baseUrl},catalog:this.opts.catalog,subject:()=>n===0?o:`${o}#${n}`,onPoolWedged:r,release:"end"}),s=this.opts.decisionIntake?new me({targetFor:i=>a.targetFor(i),apiUrl:this.opts.apiUrl,apiKey:e.apiKey,orgId:e.orgId,fnName:P,secret:this.opts.decisionIntake.secret,...this.opts.decisionIntake.shared?{shared:this.opts.decisionIntake.shared}:{}}):void 0;return{router:a,clients:this.label(W(a,void 0,r,s),e)}}clientsFor(e){let n=this.tenants.get(e.apiKey);if(n)return this.tenants.delete(e.apiKey),this.tenants.set(e.apiKey,n),n.clients;for(;this.tenants.size>=lt;){let r=this.tenants.entries().next();if(r.done)break;let[a,s]=r.value;this.tenants.delete(a),s.router.closeAll().catch(()=>{})}let o=this.build(e);return this.tenants.set(e.apiKey,o),o.clients}get size(){return this.tenants.size}async closeAll(){let e=[...this.tenants.values()].map(o=>o.router);this.tenants.clear();let n=[];if(await Promise.all(e.map(async o=>{try{await o.closeAll()}catch(r){n.push(String(r instanceof Error?r.message:r))}})),n.length>0)throw new Error(`failed to close ${n.length} tenant backhaul(s): ${n.join("; ")}`)}},xe=512,ct=1024,L=class extends Error{constructor(){super(`conversation capacity reached (${xe} conversations on this replica) \u2014 every slot holds an active or resumable conversation; retry later or scale out`),this.name="ConversationCapacityError"}};function ut(t){return{sessionStartOptions:e=>{if(e.startsWith(z))return;let n=t.get(e);return n===void 0?void 0:{sessionId:n}},onSessionOpen:(e,n)=>{e.startsWith(z)||t.set(e,n)}}}function dt(t,e,n,o){let r=j({baseUrl:t.baseUrl,orgId:e.orgId,defaultApp:null,fnName:P,sharedOnlyListing:!0,apiUrl:t.apiUrl,auth:{lane:"api-key",apiKey:e.apiKey,gatewayUrl:t.baseUrl},catalog:t.catalog,subject:n.subject,release:"detach",...ut(o)});return{clients:W(r),release:async()=>{await r.closeAll()}}}var H=class{constructor(e){this.opts=e}opts;byCaller=new Map;byHandle=new Map;build(e,n){let o={id:n,subject:`${M(e.apiKey)}:live:${n}`},r=new Map,a=this.opts.build?this.opts.build(e,o):dt(this.opts,e,o,r);return{id:n,backhaul:a,attachments:0,nativeRefs:r}}attachEntry(e,n){let o=this.byCaller.get(e.apiKey);o===void 0&&(o=new Map,this.byCaller.set(e.apiKey,o));let r=o.get(n);if(r!==void 0)return o.delete(n),o.set(n,r),r.attachments+=1,r;this.evictOverflow();let a=this.build(e,n);return a.attachments=1,o.set(n,a),a}get conversationCount(){let e=0;for(let n of this.byCaller.values())e+=n.size;return e}evictOverflow(){for(;this.conversationCount>=xe;){let e=this.byCaller.keys().next();if(e.done)throw new L;let n=e.value,o=this.byCaller.get(n),r=!1;for(let[a,s]of o)if(!(s.attachments>0)){o.delete(a),this.dropHandleIndexFor(n,a),s.backhaul.release().catch(()=>{}),r=!0;break}if(!r)throw o.size===0&&this.byCaller.delete(n),new L;o.size===0&&this.byCaller.delete(n)}}dropHandleIndexFor(e,n){for(let[o,r]of this.byHandle)r.apiKey===e&&r.conversationId===n&&this.byHandle.delete(o)}connectionFor(e){let n=null,o=!1,r=()=>(n===null&&(n=this.attachEntry(e,at())),n),a={createResponse:(s,i)=>r().backhaul.clients.createResponse(s,i),catalogIdentity:s=>r().backhaul.clients.catalogIdentity(s),servingAppSlug:s=>{let i=r().backhaul.clients.servingAppSlug;if(i===void 0)throw new Error("the conversation backhaul wires no servingAppSlug seam \u2014 the caller-org app a request resolves to cannot be known");return i(s)},listModels:()=>r().backhaul.clients.listModels(),sessionHandle:s=>r().backhaul.clients.sessionHandle(s),createResponseOn:(s,i)=>r().backhaul.clients.createResponseOn(s,i),onSessionEnd:(s,i)=>r().backhaul.clients.onSessionEnd(s,i)};return a.openAudio=async(s,i)=>{let l=r().backhaul.clients.openAudio;if(l===void 0)throw new Error("the conversation backhaul wires no openAudio seam \u2014 voice is unavailable on this deployment");return l(s,i)},a.openVideo=async s=>{let i=r().backhaul.clients.openVideo;if(i===void 0)throw new Error("the conversation backhaul wires no openVideo seam \u2014 video-in is unavailable on this deployment");return i(s)},a.openVideoOut=async s=>{let i=r().backhaul.clients.openVideoOut;if(i===void 0)throw new Error("the conversation backhaul wires no openVideoOut seam \u2014 video-out is unavailable on this deployment");return i(s)},a.supportsAudio=async s=>{let i=r().backhaul.clients.supportsAudio;if(i===void 0)throw new Error("the conversation backhaul wires no supportsAudio seam \u2014 the realtime lane cannot gate audio modality on this deployment");return i(s)},a.supportsSpeech=async s=>{let i=r().backhaul.clients.supportsSpeech;if(i===void 0)throw new Error("the conversation backhaul wires no supportsSpeech seam \u2014 the speech lane cannot gate model capability on this deployment");return i(s)},a.speechVoices=async s=>{let i=r().backhaul.clients.speechVoices;if(i===void 0)throw new Error("the conversation backhaul wires no speechVoices seam \u2014 the speech lane cannot resolve the voice set /v1/models publishes on this deployment");return i(s)},a.speechSessionHandle=async s=>{let i=r().backhaul.clients.speechSessionHandle;if(i===void 0)throw new Error("the conversation backhaul wires no speechSessionHandle seam \u2014 the speech lane cannot resolve its own pooled session on this deployment");return i(s)},a.openAudioOn=async(s,i)=>{let l=r().backhaul.clients.openAudioOn;if(l===void 0)throw new Error("the conversation backhaul wires no openAudioOn seam \u2014 the speech and realtime lanes cannot open their pinned native session audio on this deployment");return l(s,i)},a.evictSpeechSession=async s=>{let i=r().backhaul.clients.evictSpeechSession;if(i===void 0)throw new Error("the conversation backhaul wires no evictSpeechSession seam \u2014 the speech lane cannot retire its pinned speech session after a failed turn on this deployment");return i(s)},a.supportsTranscription=async s=>{let i=r().backhaul.clients.supportsTranscription;if(i===void 0)throw new Error("the conversation backhaul wires no supportsTranscription seam \u2014 the transcription lane cannot gate model capability on this deployment");return i(s)},a.listSpeechToTextModels=async()=>{let s=r().backhaul.clients.listSpeechToTextModels;if(s===void 0)throw new Error("the conversation backhaul wires no listSpeechToTextModels seam \u2014 the transcription lane cannot list the stt models on this deployment");return s()},{clients:a,bindResumeHandle:s=>{let i=n;if(i===null)throw new Error("bindResumeHandle before the conversation opened \u2014 no conversation to bind to");for(;this.byHandle.size>=ct;){let l=this.byHandle.keys().next();if(l.done)break;this.byHandle.delete(l.value)}this.byHandle.set(s,{apiKey:e.apiKey,conversationId:i.id})},clientsForResume:async s=>{let i=this.byHandle.get(s),l=i===void 0?void 0:this.byCaller.get(i.apiKey),u=i===void 0||l===void 0?void 0:l.get(i.conversationId);if(i===void 0||u===void 0||i.apiKey!==e.apiKey)throw new X("the conversation behind this resume handle is gone (evicted at capacity, or the replica restarted) or the handle is not yours \u2014 reconnect without a handle to start a new conversation");return n=u,u.attachments+=1,u.backhaul.clients},release:async()=>{if(o)return;o=!0;let s=n;n=null,s!==null&&(s.attachments-=1,!(s.attachments>0)&&await s.backhaul.release().catch(()=>{}))}}}async detachAll(){let e=[];for(let o of this.byCaller.values())for(let r of o.values())e.push(r.backhaul.release());this.byCaller.clear(),this.byHandle.clear();let n=[];if(await Promise.all(e.map(async o=>{try{await o}catch(r){n.push(String(r instanceof Error?r.message:r))}})),n.length>0)throw new Error(`failed to detach ${n.length} conversation backhaul(s): ${n.join("; ")}`)}};p();var ht="/inference-ledger",pt=1e4,D=class extends Error{ledgerFailureReason;constructor(e,n="duplicate"){super(e),this.name="LedgerDuplicateError",this.ledgerFailureReason=n}},V=class extends Error{constructor(n,o){super(o);this.ledgerAlreadyWritten=n;this.name="LedgerAlreadyWrittenError"}ledgerAlreadyWritten},g=class extends Error{ledgerFailureReason="surface";constructor(e){super(e),this.name="LedgerSurfaceError"}},Y=class extends m{ledgerFailureReason="auth";constructor(e){super(e),this.name="LedgerAuthError"}},gt=[["inference_id",t=>typeof t=="string"&&/[^\s]/.test(t)],["model_id",t=>typeof t=="string"&&/[^\s]/.test(t)],["variant",t=>t===null||typeof t=="string"],["task",t=>typeof t=="string"&&/[^\s]/.test(t)],["http_status",t=>typeof t=="number"&&Number.isInteger(t)],["started_at",t=>typeof t=="string"&&!Number.isNaN(Date.parse(t))],["completed_at",t=>typeof t=="string"&&!Number.isNaN(Date.parse(t))],["prompt_tokens",t=>t===null||typeof t=="number"&&Number.isInteger(t)],["completion_tokens",t=>t===null||typeof t=="number"&&Number.isInteger(t)],["output_units",J],["output_minutes",J],["cost_nano_usd",J],["price_version",t=>t===null||typeof t=="string"]];function ft(t){for(let[e,n]of gt){if(!(e in t))return`${e} (absent)`;if(!n(t[e]))return`${e} (${JSON.stringify(t[e])})`}return null}function mt(t){let e;try{e=JSON.parse(t)}catch{return{status:"unavailable"}}if(e===null||typeof e!="object"||Array.isArray(e))return{status:"unavailable"};let n=e.existing_status;if(n==="not_in_this_org")return{status:"not_in_this_org"};if(n!=="found")return{status:"unavailable"};let o=e.existing;return o===null||typeof o!="object"||Array.isArray(o)?{status:"unavailable"}:ft(o)!==null?{status:"unavailable"}:{status:"found",row:o}}function J(t){return t===null?!0:typeof t=="number"?Number.isFinite(t):typeof t!="string"?!1:/[^\s]/.test(t)&&Number.isFinite(Number(t))}function q(t,e){return t===null?e===null:e===null?!1:Number(e)===t}function yt(t,e){return t===null?e===null:e===null?!1:Number(t.toFixed(9))===Number(e)}function Ce(t,e){if(typeof e!="string")return!1;let n=Date.parse(t),o=Date.parse(e);return Number.isFinite(n)&&Number.isFinite(o)&&n===o}function wt(t,e){let n=e;return n.inference_id===t.inference_id&&n.model_id===t.model_id&&(n.variant??null)===t.variant&&n.task===t.task&&n.http_status===t.http_status&&Ce(t.started_at,n.started_at)&&Ce(t.completed_at,n.completed_at)&&q(t.prompt_tokens,n.prompt_tokens)&&q(t.completion_tokens,n.completion_tokens)&&q(t.output_units,n.output_units)&&yt(t.output_minutes,n.output_minutes)}function bt(t,e){if(t===null||typeof t!="object"||Array.isArray(t))throw new g(`${e} answered a "recorded" that is not an object: ${JSON.stringify(t)}`);let n=t,o=n.inference_id;if(typeof o!="string"||o==="")throw new g(`${e} answered a row with no inference_id: ${JSON.stringify(t)}`);if(!("cost_nano_usd"in n))throw new g(`${e} answered a row with no cost_nano_usd field: ${JSON.stringify(t)}`);let r=n.cost_nano_usd;if(r!=null&&!(typeof r=="number"&&Number.isInteger(r)&&r>=0))throw new g(`${e} answered cost_nano_usd ${JSON.stringify(r)}, which is neither a non-negative integer nor null`);let a=u=>{if(!(u in n))throw new g(`${e} answered a row with no ${u} field: ${JSON.stringify(t)}`);let f=n[u];if(f===null)return null;if(typeof f!="string")throw new g(`${e} answered a non-string ${u}: ${JSON.stringify(f)}`);if(!/[^\s]/.test(f))throw new g(`${e} answered a BLANK ${u} (${JSON.stringify(f)}) \u2014 it satisfies "is a string" while naming nothing, which is exactly what inference_requests_priced_attribution_check exists to forbid`);return f},s=a("price_version"),i=a("unpriced_reason"),l=r!=null;if(l&&(s===null||i!==null))throw new g(`${e} answered a PRICED row (cost_nano_usd ${String(r)}) that is not cleanly attributed \u2014 price_version ${JSON.stringify(s)}, unpriced_reason ${JSON.stringify(i)}`);if(!l&&(s!==null||i===null))throw new g(`${e} answered an UNPRICED row that does not say why \u2014 price_version ${JSON.stringify(s)}, unpriced_reason ${JSON.stringify(i)}. NULL means NOT YET PRICED, and the reason is how an operator learns that traffic is being served which nobody can bill.`);return{inference_id:o,cost_nano_usd:r===void 0?null:r,price_version:s,unpriced_reason:i}}var F=class{constructor(e){this.opts=e;this.fetchImpl=e.fetchImpl??fetch}opts;fetchImpl;async write(e,n){let o=`${this.opts.apiUrl.replace(/\/+$/,"")}${ht}`,r;try{r=await this.fetchImpl(o,{method:"POST",headers:{"content-type":"application/json",authorization:`Bearer ${e}`},body:JSON.stringify(n),signal:AbortSignal.timeout(pt)})}catch(l){throw new g(`the uRun ledger surface at ${o} could not be reached: ${l instanceof Error?l.message:String(l)}`)}if(r.status===409){let l=await r.text().catch(()=>""),u=mt(l);throw u.status==="found"&&wt(n,u.row)?new V(u.row,`the uRun control plane already holds THIS PROXY'S OWN row for inference_id ${n.inference_id} \u2014 every billable field matches the row being re-sent, so an earlier attempt landed and this retry is a no-op. Nothing was written and nothing was overwritten.`):u.status==="not_in_this_org"?new D(`inference_id ${n.inference_id} is ALREADY TAKEN by a ledger row this org cannot read. The id space is a global primary key, so this is not a retry and cannot be this proxy's own write \u2014 either two orgs collided on a uuid, or a row was written for this id by somebody else.`,"duplicate_foreign"):new D(`the uRun control plane already holds a ledger row for inference_id ${n.inference_id} that is NOT the one being written${u.status==="unavailable"?" (and it could not be read back to compare)":""} \u2014 nothing was written and nothing was overwritten. Either this proxy wrote twice, or a row was fabricated ahead of it (this write is authenticated by the CALLER'S OWN org key). ${l.slice(0,400)}`)}if(r.status===401||r.status===403)throw await r.text().catch(()=>{}),new Y("the uRun org API key was rejected by the uRun control plane when writing the inference ledger \u2014 the request was served but produced no billable row");if(!r.ok){let l=await r.text().catch(u=>`<body unreadable: ${u instanceof Error?u.message:String(u)}>`);throw new g(`the uRun ledger surface at ${o} answered HTTP ${r.status}: ${l.slice(0,400)}`)}let a;try{a=await r.json()}catch(l){throw new g(`the uRun ledger surface at ${o} answered a body that is not JSON: ${l instanceof Error?l.message:String(l)}`)}let s=a?.recorded;if(s===void 0)throw new g(`the uRun ledger surface at ${o} answered a body with no "recorded" field`);let i=bt(s,`the uRun ledger surface at ${o}`);if(i.inference_id!==n.inference_id)throw new g(`the uRun ledger surface at ${o} answered for inference_id ${i.inference_id} when asked about ${n.inference_id}`);return i}};p();import{createServer as _t}from"http";p();var K="/inference-usage",ke=25e3;var vt=new Set(["Error","TypeError","SyntaxError","TimeoutError"]);function Oe(t){if(!(t instanceof Error))return`a ${typeof t}`;let e=vt.has(t.name)?t.name:"Error",n=t.cause?.code;return typeof n=="string"&&/^[A-Z][A-Z0-9_]{0,31}$/.test(n)?`${e} (${n})`:e}var G=class{constructor(e){this.opts=e;this.fetchImpl=e.fetchImpl??fetch}opts;fetchImpl;async lookup(e,n){let o=`${this.opts.apiUrl.replace(/\/+$/,"")}${K}`,r;try{r=await this.fetchImpl(o,{method:"POST",headers:{"content-type":"application/json",authorization:`Bearer ${e}`},body:JSON.stringify({inference_ids:n}),signal:AbortSignal.timeout(ke)})}catch(s){throw new A(`the uRun usage surface (POST ${K}) could not be reached within ${ke}ms: ${Oe(s)}`)}if(r.status===401||r.status===403)throw await r.text().catch(()=>{}),new m("the uRun org API key was rejected by the uRun control plane \u2014 check the key and that it is not revoked");if(!r.ok){let s=await r.text().catch(()=>null),i=s===null?"body unreadable":`${s.length}-character body`;throw new A(`the uRun usage surface (POST ${K}) answered HTTP ${r.status} (${i})`)}let a;try{a=await r.json()}catch(s){throw new A(`the uRun usage surface (POST ${K}) answered HTTP ${r.status} with a body that is not JSON: ${Oe(s)}`)}return ce(a,this.opts.onRejectedRow)}};var Et=5e3,St=3e3,Rt="/api/client-tokens";function Q(t,e,n){t.writeHead(e,{"content-type":"application/json"}),t.end(JSON.stringify(n))}function Ae(t,e,n,o){Q(t,e,{error:{message:n,type:o}})}async function xt(t,e=fetch){let n=`${t.replace(/\/+$/,"")}${Rt}`;try{let o=await e(n,{method:"POST",headers:{"content-type":"application/json"},body:"{}",signal:AbortSignal.timeout(St)});return o.status===404?{ready:!1,detail:`${n} answered 404 \u2014 this is not a uRun control plane`}:o.status>=500?{ready:!1,detail:`${n} answered HTTP ${o.status}`}:{ready:!0,detail:`${n} answered HTTP ${o.status}`}}catch(o){return{ready:!1,detail:`${n} unreachable: ${o instanceof Error?o.message:String(o)}`}}}function En(t){let e=t.resolver??new U({gatewayUrl:t.baseUrl}),n=t.decisionSecret??"",o=t.sharedOrgKey&&t.sharedOrgId?{apiKey:t.sharedOrgKey}:void 0,r=t.registry??new $({baseUrl:t.baseUrl,apiUrl:t.apiUrl,catalog:t.catalog,...n!==""?{decisionIntake:{secret:n,...o?{shared:o}:{}}}:{}});n!==""&&process.stderr.write(`inference-proxy: stateless direct path ENABLED \u2014 caller-org decisions dial ready runtimes directly (no session mint)
-`);let a=t.probeControlPlane??(()=>xt(t.baseUrl)),s=t.usageClient??new G({apiUrl:t.apiUrl}),i=t.ledgerClient??new F({apiUrl:t.apiUrl}),l=new N(t.defaultMetrics===void 0?{}:{defaultMetrics:t.defaultMetrics}),u=async c=>e.resolve(R(c.headers.authorization)),f=c=>c instanceof m?{status:c.status,openaiType:"authentication_error",anthropicType:"authentication_error"}:c instanceof _?{status:502,openaiType:"upstream_error",anthropicType:"api_error"}:null,x=ye({clients:async c=>r.clientsFor(await u(c)),usage:{authorize:c=>R(c.headers.authorization),lookup:(c,d)=>s.lookup(c,d)},statusOf:f,...t.requestLog?{requestLog:t.requestLog}:{},observe:c=>l.observe(c),ledger:{write:(c,d)=>i.write(R(c.headers.authorization),d),observe:c=>l.observeLedger(c),...t.ledgerLog?{report:t.ledgerLog}:{}}}),v=null,E=_t((c,d)=>{let h=new URL(c.url??"/","http://localhost");(async()=>{if(c.method==="GET"&&h.pathname==="/healthz"){Q(d,200,{ok:!0});return}if(c.method==="GET"&&h.pathname==="/readyz"){let k=Date.now();if(!v||k-v.at>=Et){let Ne=await a();v={at:k,...Ne}}Q(d,v.ready?200:503,{ok:v.ready,control_plane:v.detail});return}let y=h.pathname==="/v1"||h.pathname.startsWith("/v1/"),C=h.pathname.startsWith("/partners/");if(!y&&!C){Ae(d,404,`no route for ${c.method} ${h.pathname} \u2014 this endpoint serves the OpenAI-compatible /v1 surface, the partner adapter surface (/partners/...), the WS upgrade paths (/ws/google.ai.generativelanguage.v1beta.GenerativeService.BidiGenerateContent and /v1/realtime), plus /healthz and /readyz`,"invalid_request_error");return}x(c,d)})().catch(y=>{if(!d.headersSent){Ae(d,500,String(y instanceof Error?y.message:y),"proxy_error");return}d.end()})}),B=t.conversations??new H({baseUrl:t.baseUrl,apiUrl:t.apiUrl,catalog:t.catalog}),Ie=[ee(E,{authenticate:async c=>{let d=await e.resolve(c),h=B.connectionFor(d);return{clients:h.clients,scope:M(d.apiKey),onBindResumeHandle:h.bindResumeHandle,clientsForResume:h.clientsForResume,release:h.release}}}),ne(E,{clients:async c=>r.clientsFor(await u(c)),statusOf:f}).close,se(E,{resolver:{resolve:async(c,d)=>{let h=await e.resolve(R(c.headers.authorization)),y=B.connectionFor(h),C;try{C=await ae(y.clients,ie(c),d)}catch(k){throw y.release().catch(()=>{}),k}return{...C,release:y.release}}}})];E.on("upgrade",(c,d)=>{let h=new URL(c.url??"/","http://localhost");h.pathname===Z||h.pathname===te||h.pathname===oe||d.destroyed||(d.write(`HTTP/1.1 404 Not Found\r
-Connection: close\r
-\r
-no WS route for ${h.pathname} \u2014 this endpoint serves only the Gemini Live, /v1/responses, and /v1/realtime upgrade paths`),d.destroy())});let Te=async()=>{for(let c of Ie)await c();await B.detachAll()};return{server:E,metricsServer:_e(l),metrics:l,edgeIdentityServer:Re(),registry:r,closeAll:()=>r.closeAll(),closeWebSockets:Te}}async function Sn(t,e,n){n&&await n();let o=new Promise(r=>t.close(()=>r()));t.closeAllConnections(),await o,await ge()||fe(),await e()}export{Tt as a,ve as b,I as c,Le as d,He as e,Me as f,Ke as g,Ge as h,T as i,N as j,_e as k,Lt as l,ze as m,je as n,We as o,Je as p,Ee as q,Ve as r,Ye as s,Qe as t,Re as u,m as v,_ as w,R as x,tt as y,nt as z,U as A,ot as B,P as C,Bt as D,zt as E,lt as F,M as G,$ as H,xe as I,ct as J,L as K,H as L,ht as M,pt as N,D as O,g as P,bt as Q,F as R,Et as S,St as T,En as U,Sn as V};