@urun-sh/openai 0.6.0 → 0.6.2

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-BfdBbey8.cjs';
-export { D as DialObserver, a as DialReport, g as ProxyIdentity, h as ProxyVideoOutLane, S as SessionGoneError, U as UnknownModelError } from '../responses-turn-BfdBbey8.cjs';
+import { f as ProxyHandlerOptions, P as ProxyClients } from '../responses-turn-vMefRdgJ.cjs';
+export { D as DialObserver, a as DialReport, g as ProxyIdentity, h as ProxyVideoOutLane, S as SessionGoneError, U as UnknownModelError } from '../responses-turn-vMefRdgJ.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';
@@ -285,7 +285,7 @@ declare const OPENAI_REALTIME_PATH = "/v1/realtime";
  * convention) or a `{ resolve(req, signal) }` object. `signal` aborts when
  * the CLIENT SOCKET closes/errors before the upgrade is accepted — the
  * resolver SHOULD abandon its admission wait on it (the backhaul's
- * `openAudio` rides the same signal into core `whenLive`'s own abort
+ * `openAudioOn` rides the same signal into core `whenLive`'s own abort
  * support); a resolver that ignores it still cannot admit an upgrade to a
  * vanished client: the settled binding is released, never upgraded.
  */
@@ -302,7 +302,11 @@ interface RealtimeBinding<S = unknown> {
      *   - `session.sendResponseCreate(params, input, requestId)` (the direct
      *     SdkTransport-style seam), or
      *   - `clients.createResponse(params)` — the canonical ProxyClients seam
-     *     the conversation factory hands out (`input` rides inside `params`).
+     *     shape (`input` rides inside `params`). The shared binding composition
+     *     (binding.ts, ENG-446) hands out a PINNED view of it: the call dials
+     *     `createResponseOn(handle, …)` on the ONE pooled session the binding
+     *     resolved with the audio bridge, so no turn on the connection can
+     *     reach a second pooled session behind the bridge's back.
      * A binding with neither is rejected at upgrade time, LOUD.
      */
     session?: S & {
@@ -321,8 +325,8 @@ interface RealtimeBinding<S = unknown> {
             appendInputAudio(base64Pcm16: string): void;
             onOutputAudio(handler: (b64: string) => void): () => void;
             /**
-             * The pooled uRun session this lane rides (backhaul `openAudio` reports
-             * core `Session.id`), when the backhaul knows it — the field the
+             * The pooled uRun session this lane rides (backhaul `openAudioOn`
+             * reports core `Session.id`), when the backhaul knows it — the field the
              * `realtime_upgrade` log line puts on every lifecycle event so an
              * orphaned session is traceable from the proxy log alone. Null is the
              * honest absence (an id-less session reports no id); sessionIdOf
@@ -395,6 +399,22 @@ declare function attachOpenAIRealtime<S = unknown>(server: Server, options: Open
  * Multi-attach on the SAME authorized session (explicit resume) is the one
  * sanctioned shared form: one bridge owner at a time, validated by claim.
  *
+ * ONE SESSION IDENTITY PER CONNECTION (ENG-446): the shared binding
+ * composition resolves the connection's pooled session ONCE and pins BOTH the
+ * audio bridge (`openAudioOn(handle)`) and every turn (`createResponseOn
+ * (handle, …)`, threaded through the binding's `clients` view) to that one
+ * handle. This adapter therefore treats a turn seam failure that means THE
+ * SESSION (not the request) is dead — a {@link SessionGoneError} (the pool
+ * replaced or dropped the pinned entry), or a dial that dies before any
+ * client-visible content — as the LOUD `session_gone` CLOSE of the whole
+ * connection: an `error` event naming the cause, then the 1011 close. The
+ * bridge stays bound to the dead session and NEVER rebinds, and the turn is
+ * never retried on a replacement behind the bridge's back (the two-session
+ * race the speech lane's #500 pin fixed for its lane). The ONE exception is
+ * core's own BOUNDED ADMISSION VERDICT (`SessionFailedError` never-live /
+ * create-failed — the platform's answer about this assignment, not a death):
+ * it stays the per-turn `response.failed`, exactly as before.
+ *
  * NO-OP POLICY: every unsupported client event or backend gap renders as a
  * LOUD `error` server event (or a close) — never a silently ignored frame.
  */
@@ -466,7 +486,9 @@ declare class OpenAIRealtimeConnection<S = unknown> {
      * The upstream turn seam, normalized: a binding carries EITHER
      * `session.sendResponseCreate(params, input, requestId)` (direct
      * SdkTransport-style) OR `clients.createResponse(params)` (the canonical
-     * ProxyClients seam the shared conversation factory hands out). A binding
+     * ProxyClients seam SHAPE — the shared binding composition hands out a
+     * PINNED view whose call dials `createResponseOn(handle, …)` on the ONE
+     * pooled session the audio bridge rides, ENG-446). A binding
      * with neither was already rejected at upgrade time; this is the loud
      * backstop. The ProxyClients seam is PROMISE-VALUED
      * (`rehomingCreateResponse` dials inside an async function — its
@@ -512,6 +534,45 @@ declare class OpenAIRealtimeConnection<S = unknown> {
      */
     private doneResponse;
     private upstreamErrorBody;
+    /**
+     * THE SESSION-GONE LOUD CLOSE (ENG-446): an `error` event naming the cause
+     * (`session_gone`), then the abnormal close and teardown. The connection's
+     * audio bridge is bound to the dead/replaced session and NEVER rebinds, and
+     * the turn is never retried on another session behind the bridge's back —
+     * the only honest moves are to say why and close.
+     */
+    private failSessionGone;
+    /**
+     * Classify ONE upstream turn failure (the dial throwing, or the stream
+     * throwing mid-read) — the two outcomes are the LOUD `session_gone` close
+     * and the per-turn fault the lane always had:
+     *
+     *   - a {@link SessionGoneError} (the pool no longer holds the handle's
+     *     session, or a replacement serves under it now) closes the connection
+     *     LOUDLY, pre- or post-content alike — and post-CANCEL alike: the
+     *     bridge is orphaned either way, and a cancelled turn must never mute
+     *     a dead pinned session (round-1 review, CodeRabbit + Greptile);
+     *   - a CANCELLED turn suppresses every OTHER late fault (round-1 review):
+     *     `response.cancel` does not abort a pending dial, and the dial's
+     *     rejection settles AFTER `cancelled` was set — the lane's pre-#511
+     *     behaviour stands (`response.cancelled` from `inFlight.finally`, the
+     *     connection stays open), so a generic late fault from the settling
+     *     dial/read is swallowed here rather than closing the connection as
+     *     `session_gone`;
+     *   - a dial that dies BEFORE any client-visible content (nothing flushed)
+     *     closes the connection LOUDLY too — under the pin there is no re-home,
+     *     and keeping the connection open would leave the bridge attached to a
+     *     session that cannot serve a turn, silently mute on the next
+     *     response.create. The ONE exception is core's own BOUNDED ADMISSION
+     *     VERDICT (`SessionFailedError` never-live / create-failed — the
+     *     platform's answer about this assignment, not a death), which keeps
+     *     the per-turn failure exactly as before;
+     *   - everything else (content already flowed; the upstream answered with
+     *     an `error` EVENT, handled elsewhere) is the per-turn fault — the
+     *     connection stays open, and a later turn on the same pinned handle
+     *     decides the connection's fate.
+     */
+    private upstreamTurnFault;
     private emitUpstreamFault;
     /**
      * Cancel the in-flight response. The upstream iterator is stopped through
@@ -540,11 +601,24 @@ declare class OpenAIRealtimeConnection<S = unknown> {
  * placements) through the SAME router-backed gate seam the Images lane uses
  * (`ProxyClients.supportsAudio` — `ModelRouter.supportsAudio`). A non-audio
  * model is refused LOUDLY with a 404 before anything is opened; a voice
- * binding acquires its PRIVATE audio bridge from `ProxyClients.openAudio`
- * (the pooled session's native `audio` lane at 24 kHz — the ONE canonical
- * session-audio wiring, `transport/media.ts` AUDIO_STREAM). There is no env-var
- * switch and no chat-default fall-through: an absent model resolves through
- * the router exactly like every other lane, and a chat default is refused
+ * binding rides ONE pooled-session identity for BOTH of its legs (ENG-446,
+ * the speech lane's #500 pin applied to the realtime lane): the connection
+ * resolves its pooled session ONCE (`ProxyClients.sessionHandle` — the same
+ * handle seam the speech lane pins its collector and turn with, here naming
+ * the INTERACTIVE pooled slot), opens its PRIVATE audio bridge on that exact
+ * handle (`openAudioOn(handle, signal)` — the pinned open), and drives every
+ * turn on the SAME handle (`createResponseOn(handle, …)`, threaded through
+ * the binding's `clients` view). The unpinned shape — the bridge via
+ * `openAudio(model)`, turns via the re-homing `createResponse` — let each leg
+ * resolve its OWN pooled session, so a dial that died before client-visible
+ * content re-homed the turn onto a REPLACEMENT session while the bridge
+ * stayed attached to the evicted one: the turn's audio was delivered to a
+ * session the client was not attached to, or a concurrent bridge on the
+ * replacement received another caller's audio. Behind the pin, a dead or
+ * replaced session is the connection's LOUD `session_gone` close
+ * (connection.ts), never a retry across sessions. There is no env-var switch
+ * and no chat-default fall-through: an absent model resolves through the
+ * router exactly like every other lane, and a chat default is refused
  * exactly like an explicitly named chat model.
  *
  * Used by BOTH compositions — the local `urun compat proxy` (fixed clients,
@@ -566,10 +640,15 @@ declare function realtimeModelOf(req: IncomingMessage): string | undefined;
  *   - {@link UnknownModelError} ⇒ 404 with the router's own loud message.
  *   - A non-audio verdict ⇒ 404 — the realtime lane never falls through to
  *     the chat lane, in either direction.
- *   - A voice verdict with no `openAudio` seam ⇒ 501 (a voice binding cannot
- *     be served without its bridge; a voiceless downgrade is never silent).
+ *   - A voice verdict with no `openAudioOn` seam ⇒ 501 (a voice binding cannot
+ *     be served without its pinned bridge; a voiceless downgrade is never
+ *     silent).
+ *   - ONE pooled-session identity for the bridge AND every turn
+ *     (`sessionHandle` once → `openAudioOn(handle, signal)` +
+ *     `createResponseOn(handle, …)`), so no turn on this connection can ever
+ *     resolve a different pooled session than the bridge.
  */
-declare function realtimeBindingFor(clients: Pick<ProxyClients, 'createResponse'> & Partial<Pick<ProxyClients, 'supportsAudio' | 'openAudio'>>, model: string | undefined, signal?: AbortSignal): Promise<RealtimeBinding<ProxyClients>>;
+declare function realtimeBindingFor(clients: Pick<ProxyClients, 'sessionHandle' | 'createResponseOn'> & Partial<Pick<ProxyClients, 'supportsAudio' | 'openAudioOn'>>, model: string | undefined, signal?: AbortSignal): 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-CdhIre_a.js';
-export { D as DialObserver, a as DialReport, g as ProxyIdentity, h as ProxyVideoOutLane, S as SessionGoneError, U as UnknownModelError } from '../responses-turn-CdhIre_a.js';
+import { f as ProxyHandlerOptions, P as ProxyClients } from '../responses-turn-EwxfRJJg.js';
+export { D as DialObserver, a as DialReport, g as ProxyIdentity, h as ProxyVideoOutLane, S as SessionGoneError, U as UnknownModelError } from '../responses-turn-EwxfRJJg.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';
@@ -171,6 +171,10 @@ declare class OpenAIRealtimeConnection<S = unknown> {
     
     private doneResponse;
     private upstreamErrorBody;
+    
+    private failSessionGone;
+    
+    private upstreamTurnFault;
     private emitUpstreamFault;
     
     private cancelResponse;
@@ -184,7 +188,7 @@ declare class OpenAIRealtimeConnection<S = unknown> {
 
 declare function realtimeModelOf(req: IncomingMessage): string | undefined;
 
-declare function realtimeBindingFor(clients: Pick<ProxyClients, 'createResponse'> & Partial<Pick<ProxyClients, 'supportsAudio' | 'openAudio'>>, model: string | undefined, signal?: AbortSignal): Promise<RealtimeBinding<ProxyClients>>;
+declare function realtimeBindingFor(clients: Pick<ProxyClients, 'sessionHandle' | 'createResponseOn'> & Partial<Pick<ProxyClients, 'supportsAudio' | 'openAudioOn'>>, model: string | undefined, signal?: AbortSignal): 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-TPH77ZOW.js";import{d as r,e as i}from"../chunk-WCMX5VOF.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,F as O,G as f,S as I,a as n,b as p,c as m}from"../chunk-URMOBMSP.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};

package/dist/{responses-turn-CdhIre_a.d.ts → responses-turn-EwxfRJJg.d.ts}

@@ -198,6 +198,10 @@ interface RouterModelEntry {
     warm?: number;
     
     voices?: string[];
+    
+    capabilities?: {
+        reasoning: boolean;
+    };
 }
 interface RouterModelList {
     object: 'list';
@@ -338,10 +342,14 @@ declare class ModelRouter<S> {
     
     supportsSpeech(model: string | undefined): Promise<boolean>;
     
+    supportsTranscription(model: string | undefined): Promise<boolean>;
+    
     speechVoices(model: string | undefined): Promise<SpeechVoicesResolution>;
     
     imageModelList(): Promise<RouterImageModelList>;
     
+    speechToTextModelList(): Promise<RouterModelList>;
+    
     closeAll(): Promise<void>;
 }
 
@@ -555,6 +563,8 @@ interface ResponsesCreateParams {
     response_format?: unknown;
     
     voice?: string;
+    
+    language?: string;
 }
 
 interface SessionEndInfo {
@@ -580,10 +590,14 @@ interface ProxyClients {
     
     supportsSpeech?(model: string | undefined): Promise<boolean>;
     
+    supportsTranscription?(model: string | undefined): Promise<boolean>;
+    
     speechVoices?(model: string | undefined): Promise<SpeechVoicesResolution>;
     
     listImageModels?(): Promise<unknown>;
     
+    listSpeechToTextModels?(): Promise<unknown>;
+    
     catalogIdentity(model: string | undefined): Promise<{
         model_id: string;
         variant: string;
@@ -593,7 +607,7 @@ interface ProxyClients {
     
     openAudio?(model: string | undefined, signal?: AbortSignal): Promise<ProxyAudioLane>;
     
-    openAudioOn?(handle: string): Promise<PinnedAudioLane>;
+    openAudioOn?(handle: string, signal?: AbortSignal): Promise<PinnedAudioLane>;
     
     openVideo?(model: string | undefined): Promise<ProxyVideoLane>;
     

package/dist/{responses-turn-BfdBbey8.d.cts → responses-turn-vMefRdgJ.d.cts}

@@ -650,6 +650,23 @@ interface RouterModelEntry {
      * entry, and on tts entries whose voices the catalog does not vouch.
      */
     voices?: string[];
+    /**
+     * Declared lane capabilities, derived from the catalog row — never
+     * hand-set. `reasoning` is true/false exactly when the row's
+     * `engine_args.reasoning_parser` is configured/absent on EVERY GPU
+     * placement (the runtime's own per-row parser knob, urun-python
+     * engines_vllm_engine.py / engines_sglang_engine.py): true means the
+     * serve lane separates thinking text onto `reasoning_content` /
+     * reasoning items / `thinking` blocks and `content` carries only the
+     * answer; false means the model emits no separated reasoning. A variant
+     * 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).
+     */
+    capabilities?: {
+        reasoning: boolean;
+    };
 }
 interface RouterModelList {
     object: 'list';
@@ -1182,6 +1199,23 @@ declare class ModelRouter<S> {
      * never fabricates a `false` out of an infrastructure outage.
      */
     supportsSpeech(model: string | undefined): Promise<boolean>;
+    /**
+     * THE TRANSCRIPTION LANE'S GATE (proxy/transcriptions.ts,
+     * `POST /v1/audio/transcriptions`): does `model`'s deployed app carry
+     * catalog `task` **stt** — INVARIANTLY across every GPU-placement row
+     * ({@link audioModalityForRows} answering exactly `'stt'`)? The SAME
+     * shape as {@link supportsImages}/{@link supportsSpeech}, and deliberately
+     * NOT {@link supportsAudio} (which is modality-AGNISTIC — `!== null`
+     * accepts stt OR tts — because the realtime lane serves both): a `tts`
+     * model must never be asked to transcribe and a chat model is the loud
+     * 404 the images lane renders for non-image models. A SHARED-lane hit
+     * joins on the lane's exact catalog id + variant (the SAME join
+     * {@link imageModelList} uses for the shared block), exactly like
+     * {@link supportsSpeech}. A CONFIGURED catalog oracle that FAILS
+     * propagates its rejection — the gate never fabricates a `false` out of
+     * an infrastructure outage.
+     */
+    supportsTranscription(model: string | undefined): Promise<boolean>;
     /**
      * THE SPEECH LANE'S VOICE-SET RESOLUTION (proxy/speech.ts): the voice set
      * `/v1/models` publishes for `model` — resolved by the SAME resolver the
@@ -1213,6 +1247,21 @@ declare class ModelRouter<S> {
      * bare-entries degradation here.
      */
     imageModelList(): Promise<RouterImageModelList>;
+    /**
+     * `GET /v1/audio/models`: the TRANSCRIPTION-capable (task `stt`) slice of
+     * the model surface — the audio twin of {@link imageModelList}: the
+     * caller-org deployed apps whose catalog rows vouch `stt`
+     * ({@link audioModalityForRows} answering exactly `'stt'`, the SAME
+     * consensus {@link supportsTranscription} applies), plus the shared block's
+     * endpoints joined against each endpoint's OWN catalog row (exact model_id +
+     * variant; a definite non-stt task is simply skipped, never a loud defect —
+     * the task column IS the vouch, there is no expects_image-style flag to
+     * require). The catalog read propagates (this listing IS the capability
+     * oracle's answer), and the PRICE SURFACE IS NOT READ AT ALL — an stt model
+     * is not token-priced, so {@link UNPRICED_SURFACE} states that as a fact
+     * about the lane, exactly like the images listing.
+     */
+    speechToTextModelList(): Promise<RouterModelList>;
     /** Close every pooled session (Session.end() underneath) — proxy shutdown. */
     closeAll(): Promise<void>;
 }
@@ -1940,6 +1989,17 @@ interface ResponsesCreateParams {
      * default voice applies, today's behavior.
      */
     voice?: string;
+    /**
+     * The transcription lane's per-request LANGUAGE hint (ENG-332, additive
+     * like `voice`): the ISO code a `stt`-task engine transcribes under
+     * (LiveKit's `openai.STT` always sends one, so refusing the field would
+     * break the stock consumer). Forwarded VERBATIM on the serve envelope so
+     * the runtime validates it against the transcribe engine's declared
+     * `served_languages` — the proxy never validates languages itself (it
+     * cannot know the engine's set) and never silently drops the field.
+     * Absent -> key absent, never an empty string.
+     */
+    language?: string;
 }
 /** How one pinned session ended (the native phase machinery's terminal step). */
 interface SessionEndInfo {
@@ -2058,6 +2118,17 @@ interface ProxyClients {
      * silent allow-all, and never a chat model asked to speak.
      */
     supportsSpeech?(model: string | undefined): Promise<boolean>;
+    /**
+     * The transcription lane's capability gate (proxy/transcriptions.ts), backed
+     * by ModelRouter.supportsTranscription — the caller-org catalog's `task`
+     * column, which must vouch `stt` (NOT `tts`: a synthesis model must never
+     * be asked to transcribe; NOT chat) INVARIANTLY across placements. Optional
+     * at the seam exactly like {@link supportsSpeech}: hand-built clients may
+     * omit it, in which case the mounted transcription lane answers the loud
+     * 501 — never a silent allow-all, and never a chat model asked to
+     * transcribe.
+     */
+    supportsTranscription?(model: string | undefined): Promise<boolean>;
     /**
      * THE SPEECH LANE'S VOICE-SET RESOLUTION (proxy/speech.ts): the voices
      * `/v1/models` publishes for `model` (ModelRouter.speechVoices — the SAME
@@ -2080,6 +2151,14 @@ interface ProxyClients {
      * its absence is the loud 501, never a silently empty list.
      */
     listImageModels?(): Promise<unknown>;
+    /**
+     * `GET /v1/audio/models` — the transcription-capable (task `stt`) slice of
+     * the model surface (ModelRouter.speechToTextModelList: deployed stt apps +
+     * the shared block joined against the catalog), the audio twin of
+     * {@link listImageModels}. Optional at the seam for the same reason; its
+     * absence is the loud 501, never a silently empty list.
+     */
+    listSpeechToTextModels?(): Promise<unknown>;
     /**
      * THE PRE-DIAL DEFAULT FOR THE LEDGER'S BILLING IDENTITY
      * (ModelRouter.catalogIdentityFor): the CATALOG `(model_id, variant)` a
@@ -2146,28 +2225,46 @@ interface ProxyClients {
      * text-only embeddings exist — but a surface that RECEIVES audio while the
      * embedder wired no `openAudio` must fail LOUD, never drop chunks.
      *
-     * `signal` aborts when the CLIENT that asked for the lane went away while
-     * the lane's admission was still pending (the realtime lane's raw socket
-     * close) — the backhaul aborts its `whenLive` wait on it and hands the
-     * queued admission ticket back through core `Session.cancel()`, so a
-     * vanished client never holds the sense pod's one-session slot.
-     */
+  * `signal` aborts when the CLIENT that asked for the lane went away while
+      * the lane's admission was still pending (the realtime lane's raw socket
+      * close) — the backhaul aborts its `whenLive` wait on it and hands the
+      * queued admission ticket back through core `Session.cancel()`, so a
+      * vanished client never holds the sense pod's one-session slot. The
+      * client-gone wait is REFERENCE-COUNTED per pooled entry: several callers
+      * can wait on ONE queued admission, and the ticket goes back only when
+      * the LAST waiter aborts — one client's disconnect rejects only its own
+      * open, never another caller's admission (round-1 review, Greptile).
+      */
     openAudio?(model: string | undefined, signal?: AbortSignal): Promise<ProxyAudioLane>;
     /**
-     * THE PINNED AUDIO OPEN (speech lane): the NATIVE audio lane of the EXACT
-     * pooled session `handle` names (ModelRouter.sessionForHandle — a gone or
-     * replaced session throws {@link SessionGoneError} LOUDLY, never a fresh
-     * session opened behind the handle's back). The speech lane opens its
+     * THE PINNED AUDIO OPEN (speech + realtime lanes): the NATIVE audio lane of
+     * the EXACT pooled session `handle` names (ModelRouter.sessionForHandle — a
+     * gone or replaced session throws {@link SessionGoneError} LOUDLY, never a
+     * fresh session opened behind the handle's back). The speech lane opens its
      * audio collector here and dials its response turn on the SAME handle
      * ({@link createResponseOn}), so BOTH legs are one session identity by
      * construction: a re-home can never move the turn to a second session
      * while the collector stays subscribed to the first (which missed the
      * replacement's audio and answered a voiceless 502, or worse, served a
-     * concurrent collector another session's frames). Optional at the seam
-     * exactly like {@link openAudio}; the speech lane answers its absence
-     * with the loud 501 — never a voiceless downgrade.
-     */
-    openAudioOn?(handle: string): Promise<PinnedAudioLane>;
+     * concurrent collector another session's frames). The realtime binding
+     * (openai-realtime/binding.ts, ENG-446) rides the SAME pin: one handle per
+     * connection for the bridge AND every turn, so a re-home can never detach
+     * the turn from the bridge the audio rides. Optional at the seam exactly
+     * like {@link openAudio}; both lanes answer its absence with the loud 501 —
+     * never a voiceless downgrade.
+     *
+  * `signal` carries {@link openAudio}'s client-gone semantics onto the
+      * pinned open (ENG-446): the realtime lane threads the upgrade's
+      * client-socket abort down here — abort while the pinned session's
+      * admission is still queued abandons the queued session (core
+      * `Session.cancel()`), and the open rejects. A vanished client never holds
+      * the sense pod's one-session slot through a queue wait it is no longer
+      * party to. An already-cached lane is returned untouched — and the
+      * client-gone wait rides the SAME per-entry refcount as {@link openAudio}:
+      * one of several waiters aborting rejects only that caller's open, the
+      * queued admission going back only at the LAST waiter's abort.
+      */
+    openAudioOn?(handle: string, signal?: AbortSignal): Promise<PinnedAudioLane>;
     /**
      * Open (or reuse) the NATIVE video FRAME lane on the pooled session `model`
      * routes to (transport/media.ts `enableSessionVideo`: discrete JPEG frames →

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@urun-sh/openai",
-  "version": "0.6.0",
+  "version": "0.6.2",
   "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": "^0.6.0"
+    "@urun-sh/core": "^0.6.2"
   },
   "dependencies": {
     "@prometheus-io/client": "^0.16.1",

package/dist/chunk-45FJ5GTP.js

@@ -1,7 +0,0 @@
-import{a as j,b as W}from"./chunk-TDRZZGUK.js";import{C as re,D as oe,E as se,F as ie,H as ae,I as A,J as le,K as ce,L as ue,M as de,N as he,O as pe,P as ge,Q as fe,R as me,b as Z,d as ee,e as te,q as ne,r as k}from"./chunk-TPH77ZOW.js";import{e as Q,g as z}from"./chunk-WCMX5VOF.js";import{a as X}from"./chunk-3JWYIKHM.js";import{c as p}from"./chunk-YSFSRI3D.js";p();import{Counter as R,Gauge as ye,Histogram as we,Registry as Ue,collectDefaultMetrics as Pe}from"@prometheus-io/client";import{createServer as $e}from"http";var It=9464,be="/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",ae,ce],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 N=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}},T=class{registry=new Ue;modelLabel=new N(He);orgLabel=new N(Me);requests;duration;ttft;streamErrors;undelivered;ledgerWrites;ledgerQueue;labelCardinality;labelOverflows;constructor(e={}){this.requests=new R({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 we({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 we({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 R({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 R({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 R({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 ue)this.ledgerWrites.inc({outcome:n.outcome,reason:n.reason},0);this.ledgerQueue=new ye({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"},de()),this.ledgerQueue.set({measure:"oldest_age_seconds"},he()/1e3)}}),this.labelCardinality=new ye({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 R({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 ve(t){return $e((e,n)=>{(async()=>{let o=(e.url??"/").split("?")[0];if(e.method!=="GET"||o!==be){n.writeHead(404,{"content-type":"text/plain"}),n.end(`no route for ${e.method} ${o} \u2014 this port serves only GET ${be}
-`);return}let r=await t.scrape();n.writeHead(200,{"content-type":t.contentType}),n.end(r)})().catch(o=>{if(!n.headersSent){n.writeHead(500,{"content-type":"text/plain"}),n.end(`${String(o instanceof Error?o.message:o)}
-`);return}n.end()})})}p();import{createHash as Ee}from"crypto";import{createServer as Be}from"http";var $t=9465,ze="x-urun-key-id",je="x-urun-key-fingerprint",We="urun-ratelimit-bucket:",Je="urun-ratelimit-tier:",_e="anonymous",qe=/^urun_sk_[A-Za-z0-9_-]{43}$/;function Ve(t){return Ee("sha256").update(`${We}${t}`).digest("hex")}function Ye(t){return Ee("sha256").update(`${Je}${t}`).digest("hex")}function Qe(t){let e=/^Bearer[ \t]+(\S.*)$/i.exec((t??"").trim());if(!e)return null;let n=e[1].trim();return qe.test(n)?n:null}function Re(){return Be((t,e)=>{let n=Qe(t.headers.authorization);e.setHeader(ze,n===null?_e:Ve(n)),e.setHeader(je,n===null?_e:Ye(n)),e.writeHead(200),e.end()})}p();import{createClientToken as Xe,ClientTokenError as Ze}from"@urun-sh/core";var m=class extends Error{constructor(n,o=401){super(n);this.status=o;this.name="ProxyAuthError"}status},_=class extends Error{constructor(e){super(e),this.name="ControlPlaneUnavailableError"}};function S(t){let e=(t??"").trim();if(e==="")throw new m("missing Authorization header \u2014 send your uRun org API key as `Authorization: Bearer <key>`");let n=/^Bearer[ \t]+(\S.*)$/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}: ${r instanceof Error?r.message:String(r)}`)}if(!o)throw new _("the control plane verified the API key but returned no org_id \u2014 refusing to route a request with no tenancy (control-plane contract violation)");if(this.cache.size>=nt){let r=this.cache.keys().next();r.done||this.cache.delete(r.value)}return this.cache.set(e,{orgId:o,at:this.now()}),{apiKey:e,orgId:o}}clear(){this.cache.clear()}};p();var ot=8080,P="serve",Gt="0.0.0.0";function st(t,e,n){let o=(t[e]??"").trim();if(!o)throw new Error(`${e} is not set \u2014 ${n}`);return o}function Bt(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 ${JSON.stringify(e)}`);let o=ne(k(st(t,"URUN_BASE_URL","the hosted proxy needs the uRun session-gateway base it opens sessions and verifies API keys against (e.g. https://api.urun.sh)"),"URUN_BASE_URL")),r=(t.URUN_API_URL??"").trim(),a=r?k(r,"URUN_API_URL"):`${o.replace(/\/+$/,"")}/v1`,s=(t.URUN_CATALOG_URL??"").trim(),i=(t.URUN_CATALOG_ANON_KEY??"").trim();if(!!s!=!!i)throw new Error("URUN_CATALOG_URL and URUN_CATALOG_ANON_KEY must be set together (or both left unset) \u2014 exactly one of them is set, which would silently disable the model-catalog oracle");let l=s?{catalogUrl:k(s,"URUN_CATALOG_URL"),anonKey:i}:null,u=(t.URUN_SESSION_TOKEN_SECRET??"").trim(),f=(t.URUN_SHARED_ORG_API_KEY??"").trim(),x=(t.URUN_SHARED_ORG_ID??"").trim();if(!!f!=!!x)throw new Error("URUN_SHARED_ORG_API_KEY and URUN_SHARED_ORG_ID must be set together (or both left unset) \u2014 exactly one of them is set, which would silently leave the shared endpoints on the session lane");return{port:n,baseUrl:o,apiUrl:a,catalog:l,decisionSecret:u,sharedOrgKey:f,sharedOrgId:x}}p();import{createHash as it,randomUUID as at}from"crypto";var lt=512;function M(t){return`urun-inference-proxy-${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 fe({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): ${n.join("; ")}`)}},Se=512,ct=1024,L=class extends Error{constructor(){super(`conversation capacity reached (${Se} 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>=Se;){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=>{let i=r().backhaul.clients.openAudioOn;if(i===void 0)throw new Error("the conversation backhaul wires no openAudioOn seam \u2014 the speech lane cannot open its pinned native session audio on this deployment");return i(s)},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)},{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 Q("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): ${n.join("; ")}`)}};p();var ht="/inference-ledger",pt=1e4,D=class extends Error{ledgerFailureReason;constructor(e,n="duplicate"){super(e),this.name="LedgerDuplicateError",this.ledgerFailureReason=n}},q=class extends Error{constructor(n,o){super(o);this.ledgerAlreadyWritten=n;this.name="LedgerAlreadyWrittenError"}ledgerAlreadyWritten},g=class extends Error{ledgerFailureReason="surface";constructor(e){super(e),this.name="LedgerSurfaceError"}},V=class extends m{ledgerFailureReason="auth";constructor(e){super(e),this.name="LedgerAuthError"}},gt=[["inference_id",t=>typeof t=="string"&&/[^\s]/.test(t)],["model_id",t=>typeof t=="string"&&/[^\s]/.test(t)],["variant",t=>t===null||typeof t=="string"],["task",t=>typeof t=="string"&&/[^\s]/.test(t)],["http_status",t=>typeof t=="number"&&Number.isInteger(t)],["started_at",t=>typeof t=="string"&&!Number.isNaN(Date.parse(t))],["completed_at",t=>typeof t=="string"&&!Number.isNaN(Date.parse(t))],["prompt_tokens",t=>t===null||typeof t=="number"&&Number.isInteger(t)],["completion_tokens",t=>t===null||typeof t=="number"&&Number.isInteger(t)],["output_units",xe],["cost_nano_usd",xe],["price_version",t=>t===null||typeof t=="string"]];function ft(t){for(let[e,n]of gt){if(!(e in t))return`${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 xe(t){return t===null?!0:typeof t=="number"?Number.isFinite(t):typeof t!="string"?!1:/[^\s]/.test(t)&&Number.isFinite(Number(t))}function J(t,e){return t===null?e===null:e===null?!1:Number(e)===t}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 yt(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)&&J(t.prompt_tokens,n.prompt_tokens)&&J(t.completion_tokens,n.completion_tokens)&&J(t.output_units,n.output_units)}function wt(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"&&yt(n,u.row)?new q(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 V("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=wt(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 vt}from"http";p();var K="/inference-usage",Oe=25e3;var bt=new Set(["Error","TypeError","SyntaxError","TimeoutError"]);function ke(t){if(!(t instanceof Error))return`a ${typeof t}`;let e=bt.has(t.name)?t.name:"Error",n=t.cause?.code;return typeof n=="string"&&/^[A-Z][A-Z0-9_]{0,31}$/.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(Oe)})}catch(s){throw new A(`the uRun usage surface (POST ${K}) could not be reached within ${Oe}ms: ${ke(s)}`)}if(r.status===401||r.status===403)throw await r.text().catch(()=>{}),new m("the uRun org API key was rejected by the uRun control plane \u2014 check the key and that it is not revoked");if(!r.ok){let s=await r.text().catch(()=>null),i=s===null?"body unreadable":`${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: ${ke(s)}`)}return le(a,this.opts.onRejectedRow)}};var _t=5e3,Et=3e3,Rt="/api/client-tokens";function Y(t,e,n){t.writeHead(e,{"content-type":"application/json"}),t.end(JSON.stringify(n))}function Ae(t,e,n,o){Y(t,e,{error:{message:n,type:o}})}async function St(t,e=fetch){let n=`${t.replace(/\/+$/,"")}${Rt}`;try{let o=await e(n,{method:"POST",headers:{"content-type":"application/json"},body:"{}",signal:AbortSignal.timeout(Et)});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 _n(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??(()=>St(t.baseUrl)),s=t.usageClient??new G({apiUrl:t.apiUrl}),i=t.ledgerClient??new F({apiUrl:t.apiUrl}),l=new T(t.defaultMetrics===void 0?{}:{defaultMetrics:t.defaultMetrics}),u=async c=>e.resolve(S(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=me({clients:async c=>r.clientsFor(await u(c)),usage:{authorize:c=>S(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(S(c.headers.authorization),d),observe:c=>l.observeLedger(c),...t.ledgerLog?{report:t.ledgerLog}:{}}}),v=null,E=vt((c,d)=>{let h=new URL(c.url??"/","http://localhost");(async()=>{if(c.method==="GET"&&h.pathname==="/healthz"){Y(d,200,{ok:!0});return}if(c.method==="GET"&&h.pathname==="/readyz"){let O=Date.now();if(!v||O-v.at>=_t){let Te=await a();v={at:O,...Te}}Y(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=[Z(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}}}),te(E,{clients:async c=>r.clientsFor(await u(c)),statusOf:f}).close,oe(E,{resolver:{resolve:async(c,d)=>{let h=await e.resolve(S(c.headers.authorization)),y=B.connectionFor(h),C;try{C=await ie(y.clients,se(c),d)}catch(O){throw y.release().catch(()=>{}),O}return{...C,release:y.release}}}})];E.on("upgrade",(c,d)=>{let h=new URL(c.url??"/","http://localhost");h.pathname===X||h.pathname===ee||h.pathname===re||d.destroyed||(d.write(`HTTP/1.1 404 Not Found\r
-Connection: close\r
-\r
-no WS route for ${h.pathname} \u2014 this endpoint serves only the Gemini Live, /v1/responses, and /v1/realtime upgrade paths`),d.destroy())});let Ne=async()=>{for(let c of Ie)await c();await B.detachAll()};return{server:E,metricsServer:ve(l),metrics:l,edgeIdentityServer:Re(),registry:r,closeAll:()=>r.closeAll(),closeWebSockets:Ne}}async function En(t,e,n){n&&await n();let o=new Promise(r=>t.close(()=>r()));t.closeAllConnections(),await o,await pe()||ge(),await e()}export{It as a,be as b,I as c,Le as d,He as e,Me as f,Ke as g,Ge as h,N as i,T as j,ve as k,$t as l,ze as m,je as n,We as o,Je as p,_e as q,Ve as r,Ye as s,Qe as t,Re as u,m as v,_ as w,S as x,tt as y,nt as z,U as A,ot as B,P as C,Gt as D,Bt as E,lt as F,M as G,$ as H,Se as I,ct as J,L as K,H as L,ht as M,pt as N,D as O,g as P,wt as Q,F as R,_t as S,Et as T,_n as U,En as V};

package/dist/chunk-QTVTCMJU.js

@@ -1 +0,0 @@
-import{c as a}from"./chunk-YSFSRI3D.js";a();function _(e,t,o,n){let r=e.response??{},i={request_id:o,consumer_id:n,stream:!0,kind:"chat",messages:t};return typeof r.instructions=="string"&&(i.instructions=r.instructions),Array.isArray(r.modalities)&&(i.modalities=r.modalities),typeof r.temperature=="number"&&(i.temperature=r.temperature),typeof r.max_output_tokens=="number"&&(i.max_output_tokens=r.max_output_tokens),r.tools!==void 0&&(i.tools=r.tools),i}function f(e,t,o){let n={request_id:t,consumer_id:o,stream:!!e.stream,kind:"responses",input:e.input};return e.model&&(n.model=e.model),e.tools!==void 0&&(n.tools=e.tools),e.tool_choice!==void 0&&(n.tool_choice=e.tool_choice),typeof e.temperature=="number"&&(n.temperature=e.temperature),typeof e.max_output_tokens=="number"&&(n.max_output_tokens=e.max_output_tokens),typeof e.top_p=="number"&&(n.top_p=e.top_p),Array.isArray(e.stop)&&e.stop.length>0&&(n.stop=e.stop),typeof e.reasoning_effort=="string"&&(n.reasoning_effort=e.reasoning_effort),typeof e.instructions=="string"&&(n.instructions=e.instructions),e.chat_template_kwargs!==void 0&&(n.chat_template_kwargs=e.chat_template_kwargs),e.response_format!==void 0&&(n.response_format=e.response_format),typeof e.voice=="string"&&(n.voice=e.voice),n}a();function y(e){if(e instanceof Error)return{type:"error",error:{type:"urun_error",code:e.name||null,message:e.message}};if(e&&typeof e=="object"&&e.t==="error"){let t=e,o=t.body??{},n=o.error,r=n?.message??o.message??`unreadable serve error body: ${JSON.stringify(t.body??null)}`;return{type:"error",error:{type:n?.type??"urun_error",code:t.code??n?.code??null,message:r}}}return{type:"error",error:{type:"urun_error",code:null,message:String(e)}}}a();a();var w="llm-resp";async function*m(e,t){let o=`${w}:${t}`,n=`resp_${t}`;yield{type:"response.created",response:{id:n,status:"in_progress"}};let r=new Map;try{for await(let i of e.stream(o).messages()){let s=i;if(s.t==="delta"){if(typeof s.delta=="string"&&(yield{type:"response.output_text.delta",item_id:n,delta:s.delta}),typeof s.reasoning=="string"&&(yield{type:"response.reasoning_text.delta",item_id:n,delta:s.reasoning}),Array.isArray(s.tool_calls))for(let u of s.tool_calls){let d=typeof u.index=="number"?u.index:0,c=r.get(d)??{};u.id&&(c.id=u.id),u.function?.name&&(c.name=u.function.name),r.set(d,c),yield{type:"response.function_call_arguments.delta",item_id:`fc_${t}_${d}`,tool_index:d,call_id:c.id,name:c.name,delta:u.function?.arguments??""}}}else if(s.t==="response"){yield{type:"response.completed",response:{id:n,status:"completed",...s.body}};return}else if(s.t==="error"){yield y(s);return}}}finally{try{e.doc("llm").set({requests:{[t]:null}})}catch{}}}a();var k=class e{constructor(t,o=e.DEFAULT_DOC_SYNC_GATE_TIMEOUT_MS){this.session=t;this.docSyncTimeoutMs=o;this.sessionTag=t.sessionId??globalThis.crypto.randomUUID()}session;docSyncTimeoutMs;sessionTag;static DEFAULT_DOC_SYNC_GATE_TIMEOUT_MS=15e3;get consumerId(){return this.session.consumerId}async ensureDocSynced(){let t=this.session.doc("llm");if(t.synced===!0||typeof t.onSynced!="function")return;let o;await new Promise((r,i)=>{let s=t.onSynced(()=>{clearTimeout(o),r()});o=setTimeout(()=>{s(),i(new Error(`[urun] llm doc plane never synced within ${this.docSyncTimeoutMs}ms \u2014 the request envelope write is withheld rather than raced into a possibly-lost pre-sync write (resumed-attach first-write-lost, prod-usw2 2026-09-23)`))},this.docSyncTimeoutMs)})}write(t){t.session_tag=this.sessionTag,this.session.doc("llm").set({requests:{[t.request_id]:{payload:t,consumer_id:t.consumer_id,stream:t.stream}}})}sendResponses(t,o){return this.writeGated(o,()=>{let n=f(t,o,this.consumerId);return this.write(n),m(this.session,o)})}sendResponseCreate(t,o,n){return this.writeGated(n,()=>{let r=_(t,o,n,this.consumerId);return this.write(r),m(this.session,n)})}async writeGated(t,o){try{await this.ensureDocSynced()}catch(n){throw new Error(`[urun] responses dial ${t} aborted before its envelope write: ${n instanceof Error?n.message:String(n)}`)}return o()}};export{_ as a,f as b,y as c,m as d,k as e};

package/dist/chunk-TDRZZGUK.js

@@ -1 +0,0 @@
-import{a as C}from"./chunk-K23AZPI4.js";import{a as L}from"./chunk-YXK42SKC.js";import{a as x,b as I,e as v,f as G,g as D,h as U,j as M}from"./chunk-WCMX5VOF.js";import{c as E,f as k,g as R,p as F}from"./chunk-6M5JK4YC.js";import{c as w}from"./chunk-YSFSRI3D.js";w();import{App as H,createClientToken as Y}from"@urun-sh/core";w();import{SessionFailedError as W}from"@urun-sh/core";var b=2;function q(e){let s=e?.type;return s==="response.created"||s==="response.in_progress"}function K(e){let s=e?.type;return s==="response.completed"||s==="error"}function $(e){return async function(i,p){return(async function*(){for(let r=1;r<=b;r++){let{app:d,entry:c,identity:l}=await e.sessionFor(i.model);p?.({sessionId:e.sessionIdOf?.(c)??null,identity:l});let h=!1,a=!1;try{let n=await e.createStream(c,i);for await(let o of n){if(K(o)&&(h=!0),q(o)){if(r>1)continue}else a=!0;yield o}}catch(n){if(n instanceof W&&(n.kind==="never-live"||n.kind==="create-failed")){try{e.onAdmissionVerdict?.(d,c,n)}catch{}throw n}if(r<b&&!a){await e.evict(d,c);continue}throw n}if(h)return;if(r<b&&!a){await e.evict(d,c);continue}return}})()}}w();var J=3,Q="llm",X=new Set(["paused","error"]);function j(e,s){let i=e.doc(Q),p=typeof i.onConnectionState=="function",u=typeof e.onPhase=="function";if(!p&&!u)throw new Error(`session for app "${s}" exposes neither a doc connection-state surface nor onPhase \u2014 session-gone detection requires @urun-sh/core Session (SessionDocument.onConnectionState / Session.onPhase); pooling a session with no native lifecycle signal would silently swallow requests dispatched after the platform closes it`);let r=null,d=!1,c=new Set,l=[],h=a=>{if(!r){r=a;for(let n of[...c])n(a);c.clear()}};return p&&l.push(i.onConnectionState(a=>{d&&(a.consecutiveFailures<J||h(new v(`the pooled session for app "${s}" is gone: its doc backhaul failed ${a.consecutiveFailures} consecutive times (last close ${a.lastCloseCode===null?"by provider liveness reap":`code ${a.lastCloseCode}`}${a.lastCloseReason?`, reason ${JSON.stringify(a.lastCloseReason)}`:""}) \u2014 the platform closed the session (idle timeout or a serve-side crash). The entry is evicted; retry to get a fresh session.`)))})),u&&l.push(e.onPhase(a=>{if(a.name==="live"){d=!0;return}d&&X.has(a.name)&&h(new v(`the pooled session for app "${s}" lost its compute binding (phase "${a.name}"). The entry is evicted; retry to get a fresh session. The named session itself is not over \u2014 it stays attachable by name.`))})),{gone:()=>r,onGone(a){return r?(a(r),()=>{}):(c.add(a),()=>c.delete(a))},dispose(){c.clear();for(let a of l.splice(0))a()}}}async function*A(e,s){let i=s.gone();if(i)throw i;let p=e[Symbol.asyncIterator](),u,r=new Promise((d,c)=>{u=s.onGone(l=>c(l))});r.catch(()=>{});try{for(;;){let d=await Promise.race([p.next(),r]);if(d.done)return;yield d.value}}finally{u?.(),p.return?.(void 0)}}function V(e,s,i,p){let u=e[i];if(typeof u!="function")throw new Error(`app "${s}" has no function "${i}" (set URUN_FUNCTION)`);let r=p?u(void 0,p):u();if(typeof r.end!="function")throw new Error(`app "${s}" function "${i}" returned a session without end()`);return r}function z(e,s,i){let{baseUrl:p,orgId:u,auth:r}=e,d=typeof e.subject=="function"?e.subject():e.subject,c=i.startsWith(D)&&d!==void 0?`${d}:speech`:d,l=s.fnName??e.fnName,h=e.sessionStartOptions?.(i),a=s.sharedApp?{sharedApp:!0,...h}:h,n=m=>({session:m,responses:new L(m),gone:j(m,i)});if(r.lane==="jwt"){let m=H(s.appSlug,{baseUrl:p,orgId:u,jwt:r.jwt});return n(V(m,s.appSlug,l,a))}let{apiKey:o,gatewayUrl:t}=r,f=H(s.appSlug,{baseUrl:p,orgId:u,getAccessToken:async()=>(await Y(o,{baseUrl:t,expiresIn:300,allowedFunctions:[`${s.appSlug}/${l}`],...c===void 0?{}:{subject:c},...s.sharedApp?{sharedApp:!0}:{}})).token});return n(V(f,s.appSlug,l,a))}function Z(e){let s=e.session.id;if(typeof s!="string"||s.length===0)throw new Error("pooled session exposes no stable session id \u2014 the session-identity seam requires @urun-sh/core Session (session.id)");return s}function ee(e){if(e.kind==="never-live")return!0;if(e.kind!=="create-failed")return!1;let s=e.status.error?.httpStatus,i=e.status.error?.code;return s===410||s===409&&(i==="cancelled"||i==="expired")?!0:i==="client_ended"}var ne=new Set(["error","paused"]);function ge(e){let{auth:s,catalog:i}=e,p=null,u=(d,c)=>{let l=z(e,d,c);return e.onSessionOpen&&typeof l.session.id=="string"&&l.session.id!==""&&e.onSessionOpen(c,l.session.id),l.gone.onGone(()=>{p?.evict(c,l)}),l},r=new M({sharedOnlyListing:e.sharedOnlyListing===!0,defaultApp:e.defaultApp,fnName:e.fnName,openSession:u,closeSession:async d=>{if(d.gone.dispose(),e.release==="detach"){if(typeof d.session.detach!="function")throw new Error('pooled session exposes no native detach() \u2014 the release:"detach" lane requires @urun-sh/core Session.detach');d.session.detach();return}await d.session.end()},listApps:s.lane==="api-key"?()=>G({apiUrl:e.apiUrl,apiKey:s.apiKey}):null,listCatalog:i?()=>C(i):null,listPrices:i?()=>I(i):null,shared:s.lane==="api-key"?()=>x({apiUrl:e.apiUrl,apiKey:s.apiKey}):null,sessionKey:Z});return p=r,e.defaultApp!==null&&e.statelessDirect!==!0&&r.seed(e.defaultApp,u({appSlug:e.defaultApp},U({appSlug:e.defaultApp}))),r}async function P(e,s){let i=e.whenLive;if(typeof i!="function")throw new Error("pooled session exposes no native whenLive() \u2014 the audio lane must wait for the session's own admission (cold wake: queued \u2192 demand \u2192 promote) before wiring media; a session without the core Session.whenLive surface cannot be served audio");await i.call(e,s===void 0?void 0:{signal:s})}async function N(e){if(typeof e.cancel=="function")try{await e.cancel();return}catch{}if(typeof e.detach!="function")throw new Error("pooled session exposes neither cancel() nor detach() \u2014 the client-gone abandon path requires the @urun-sh/core Session native lifecycle surface");e.detach()}function B(e){return e.reason instanceof Error?e.reason:new Error("the client socket closed")}function ve(e,s=R,i,p){let u=new WeakMap,r=new WeakMap,d=new WeakMap,c=new WeakMap,l=n=>{let o=n.session.id;return typeof o=="string"&&o.length>0?o:null},h=(n,o)=>{let t=u.get(n);return t||(t=(async()=>{await P(n.session,o);let f=await E(n.session,await s());return Object.assign(f,{sessionId:n.session.id})})(),u.set(n,t),t.catch(()=>u.delete(n))),t},a=n=>{let o=c.get(n);return o||(o=(async()=>{let t=await h(n);return{appendInputAudio:f=>t.appendInputAudio(f),onOutputAudio:f=>t.onOutputAudio(f),sessionId:l(n)}})(),c.set(n,o),o.catch(()=>c.delete(n))),o};return{createResponse:$({sessionFor:n=>e.sessionFor(n),sessionIdOf:n=>l(n),evict:(n,o)=>e.evict(n,o),onAdmissionVerdict:(n,o,t)=>{ee(t)&&(e.evict(n,o),i?.(n,t.kind==="never-live"?"never-live":"create-failed"))},createStream:async(n,o)=>(await P(n.session),A(await n.responses.responses.create(o),n.gone))}),listModels:()=>e.modelList(),openRouterModels:()=>e.openRouterModels(),supportsImages:(n,o)=>e.supportsImages(n,o),listImageModels:()=>e.imageModelList(),catalogIdentity:n=>e.catalogIdentityFor(n),servingAppSlug:async n=>await e.servingAppSlugFor(n),supportsAudio:n=>e.supportsAudio(n),supportsSpeech:n=>e.supportsSpeech(n),speechVoices:n=>e.speechVoices(n),openAudio:async(n,o)=>{let{entry:t}=await e.sessionFor(n);if(o?.aborted)throw await N(t.session),B(o);let f=h(t,o);return o===void 0?f:new Promise((m,S)=>{let y=()=>{N(t.session).catch(()=>{}),S(B(o))};o.addEventListener("abort",y,{once:!0});let O=()=>o.removeEventListener("abort",y);f.then(g=>{O(),m(g)},g=>{O(),S(g)})})},openAudioOn:async n=>a((await e.sessionForHandle(n)).entry),openVideo:async n=>{let{entry:o}=await e.sessionFor(n),t=r.get(o);return t||(t=k(o.session),r.set(o,t)),t},openVideoOut:async n=>{let{entry:o}=await e.sessionFor(n),t=d.get(o);return t||(t=F(o.session),d.set(o,t)),t},sessionHandle:async n=>(await e.handleFor(n)).handle,speechSessionHandle:async n=>(await e.speechHandleFor(n)).handle,evictSpeechSession:async n=>{let{app:o,entry:t}=await e.sessionForHandle(n);await e.evict(o,t)},createResponseOn:async(n,o)=>{let{entry:t}=await e.sessionForHandle(n);return await P(t.session),A(await t.responses.responses.create(o),t.gone)},onSessionEnd:async(n,o)=>{let{entry:t}=await e.sessionForHandle(n),f=t.session;if(typeof f.onPhase!="function")throw new Error("pooled session exposes no onPhase \u2014 session-end (goAway) wiring requires @urun-sh/core Session");let m=!1;return f.onPhase(S=>{if(m||!ne.has(S.name))return;m=!0;let y=f.endsAt??null;o({timeLeftMs:y?Math.max(0,y.getTime()-Date.now()):null,reason:S.reason??`session ${S.name}`})})},...p?{stateless:p}:{}}}export{ge as a,ve as b};