@urun-sh/openai 0.6.18 → 0.6.20

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-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';
+import { f as ProxyHandlerOptions, P as ProxyClients } from '../responses-turn-Cb1qQBA7.cjs';
+export { D as DialObserver, a as DialReport, g as ProxyIdentity, h as ProxyVideoOutLane, S as SessionGoneError, U as UnknownModelError } from '../responses-turn-Cb1qQBA7.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';

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-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';
+import { f as ProxyHandlerOptions, P as ProxyClients } from '../responses-turn-DmzWIpmO.js';
+export { D as DialObserver, a as DialReport, g as ProxyIdentity, h as ProxyVideoOutLane, S as SessionGoneError, U as UnknownModelError } from '../responses-turn-DmzWIpmO.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';

package/dist/proxy/index.js

@@ -1 +1 @@
-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};
+import{B as l,C as a,D as x,E as s,G as O,H as f,S as I,a as n,b as p,c as m}from"../chunk-TF2IMK3T.js";import{f as r,g as i}from"../chunk-BAKPG5S3.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-B91w2YKn.d.cts → responses-turn-Cb1qQBA7.d.cts}

@@ -712,12 +712,10 @@ interface RouteTarget {
     variant?: string;
     /**
      * Shared lanes only: the pod org (`shared.shared_org_id` of the block that
-     * resolved the lane). THE STATELESS DECISION TOKEN'S PROVIDER CLAIM — the
-     * decision intake's provider-aware org gate accepts tenant OR provider
-     * matching the pod org, and on a shared pod the caller-org tenant differs
-     * from the shared-org pod. Sourced from the SAME block that resolved the
-     * lane, never a second env spelling, so key and pod-org identity cannot
-     * drift.
+     * resolved the lane). On a shared pod the caller-org tenant differs from
+     * the shared-org pod, and the org identity of the lane must come from the
+     * SAME block that resolved it — never a second env spelling, so key and
+     * pod-org identity cannot drift.
      */
     sharedOrgId?: string;
 }
@@ -982,15 +980,6 @@ declare class ModelRouter<S> {
      * Rides {@link resolveTarget}'s cached listings and OPENS NO SESSION.
      */
     servingAppSlugFor(model: string | undefined): Promise<string>;
-    /**
-     * THE STATELESS DIRECT PATH's resolution seam: the dial target for a
-     * model — the pure resolution (app slug, fn, shared flag) with NO session
-     * opened. The stateless decision dial (stateless-dispatch.ts) uses it to
-     * discover ready replicas for the resolved (org, app, fn) instead of
-     * minting a session; every public no-dial accessor above rides this same
-     * private resolution.
-     */
-    targetFor(model: string | undefined): Promise<RouteTarget>;
     /**
      * Resolve a request's `model` to its POOL KEY — the app slug for a
      * caller-org app, `shared:<slug>` for a shared-lane dial (the two never
@@ -1327,117 +1316,6 @@ interface SessionTurnLaneEvent {
     startMs?: number;
 }
 
-/** One ready replica row from the presence read. */
-interface RuntimePresenceRow {
-    runtime_id: string;
-    /** Base URL of the runtime's :8099 endpoint, e.g. `http://10.0.1.5:8099`. */
-    address: string;
-    ready: boolean;
-    /** ISO 8601 instant of the runtime's last presence heartbeat. */
-    last_seen: string;
-    status?: string | null;
-}
-/** One resolved dial target — the router's pure resolution (no session).
- * The shared arm's fields ride through so the dial can mint the
- * provider-aware token and the handlers can repin the CATALOG billing pair:
- * `sharedOrgId` is the shared block's `shared_org_id` (the pod org). */
-interface DialTarget {
-    appSlug: string;
-    fnName?: string;
-    sharedApp?: boolean;
-    modelId?: string;
-    variant?: string;
-    sharedOrgId?: string;
-}
-interface StatelessDecisionDialOptions {
-    /** Resolve a request's model to its dial target WITHOUT opening a session. */
-    targetFor: (model: string | undefined) => Promise<DialTarget>;
-    apiUrl: string;
-    apiKey: string;
-    /** The caller org's id — the minted token's `tenant` claim. */
-    orgId: string;
-    /** The serve function name the resolved app must expose (the backhaul's
-     * URUN_FUNCTION default, 'serve'). */
-    fnName: string;
-    /** The shared secret render injects (URUN_SESSION_TOKEN_SECRET). */
-    secret: string;
-    /** THE SHARED LANE'S credential (URUN_SHARED_ORG_API_KEY): the shared
-     * org's API key, so discovery and decisions can reach the SHARED org's
-     * pods too. The provider org id comes from the resolved target's
-     * `sharedOrgId` (the block's `shared_org_id`), never a second env spelling
-     * of it — key and pod-org identity cannot drift apart. Absent ⇒ shared
-     * models keep the session lane (the v1 boundary narrows to "no shared
-     * credential mounted" instead of "shared never"). */
-    shared?: {
-        apiKey: string;
-    };
-    fetchImpl?: typeof fetch;
-    nowMs?: () => number;
-    /** Per-attempt POST budget override (test seam; production rides
-     * DECISION_TIMEOUT_MS). */
-    timeoutMs?: number;
-}
-/**
- * The stateless decision dial. `decide` resolves the target, discovers ready
- * replicas, picks the least-outstanding one, mints the scoped token and POSTs
- * the flat body to the runtime's `/v1/decision` intake.
- */
-declare class StatelessDecisionDial {
-    private readonly opts;
-    private readonly replicas;
-    private readonly nowMs;
-    /** Per-(presence key, app, fn) fresh listing, TTL-bounded to
-     * PRESENCE_FRESH_S. A per-DECISION control-plane read would be the same
-     * design error the session dial was — a network round-trip for state the
-     * runtime plane re-proves every ~5s heartbeat. The cache serves the
-     * caller from memory, refreshes in the BACKGROUND past half the TTL, and
-     * evicts a replica (or the whole entry) the moment a dial to its cached
-     * address fails. */
-    private readonly presenceCache;
-    constructor(opts: StatelessDecisionDialOptions);
-    /** Enabled = the pod secret is configured. Absent ⇒ the session lane serves
-     * (byte-for-byte behavior). */
-    get enabled(): boolean;
-    /**
-     * Resolve a request's lane target WITHOUT dialing: `null` for a SHARED-lane
-     * model when the dial holds NO shared credential — the caller keeps the
-     * session lane. WITH the shared key, a shared target is dialable and rides
-     * through with its full identity (modelId/variant for the billing repin,
-     * `sharedOrgId` for the provider-aware token). The HTTP handlers call this
-     * FIRST and route non-null targets into decide; decide re-resolves its own
-     * target so the one it reports is the one it actually dialed.
-     */
-    targetFor(model: string | undefined): Promise<DialTarget | null>;
-    /**
-     * One decision. Returns the intake's `(status, body)` plus the SYNTHESIZED
-     * completed body the lane stamps its billing receipt from — the runtime's
-     * chat body carries `usage` (OpenAI shape) and `urun_timing`; both are the
-     * exact fields the ledger's receipt stamper reads, mapped once here so the
-     * stateless lane prices identically to the session lane. The RESOLVED
-     * target rides along: the handlers repin the billing identity from it
-     * (the dial's own answer, handed out — never asked for again) before
-     * stamping.
-     */
-    decide(body: Record<string, unknown>, model: string | undefined): Promise<{
-        status: number;
-        body: unknown;
-        completed: unknown;
-        target: DialTarget & {
-            fnName: string;
-        };
-    }>;
-    /** Serve the caller from the cache; a MISS (or a TTL-expired entry) reads
-     * synchronously — the cold path is the one decision that pays the hop —
-     * and an entry past half the TTL fires a single-flight BACKGROUND refresh
-     * that never blocks this caller. */
-    private ensureRows;
-    private readPresence;
-    /** LEAST-OUTSTANDING: pick the ready replica with the fewest in-flight
-     * decisions (ties → the first), skipping replicas in failure cooldown. */
-    pick(rows: RuntimePresenceRow[]): string | null;
-    private post;
-}
-
 /**
  * THE CANONICAL USAGE-QUERY LANE — `POST /v1/usage/requests`.
  *
@@ -2464,15 +2342,6 @@ interface ProxyClients {
     * the unsubscribe.
     */
     onSessionEnd(handle: string, cb: (end: SessionEndInfo) => void): Promise<() => void>;
-    /**
-     * THE STATELESS DIRECT PATH (stateless-dispatch.ts): the decision dial to
-     * the serving pods' `POST /v1/decision` intake, present only when the
-     * deployment carries the runtime scoped-token secret. Absent ⇒ every
-     * decision rides the session lane exactly as before (the staged rollout).
-     * Per-org: the HOSTED registry builds one per tenant so the minted token's
-     * tenant claim is always the CALLER's org.
-     */
-    readonly stateless?: StatelessDecisionDial;
 }
 /**
  * The audio lane handle `openAudio` returns — structurally the transport

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

@@ -290,8 +290,6 @@ declare class ModelRouter<S> {
     
     servingAppSlugFor(model: string | undefined): Promise<string>;
     
-    targetFor(model: string | undefined): Promise<RouteTarget>;
-    
     resolveApp(model: string | undefined): Promise<string>;
     
     sessionFor(model: string | undefined): Promise<{
@@ -357,73 +355,6 @@ interface SessionTurnLaneEvent {
     startMs?: number;
 }
 
-interface RuntimePresenceRow {
-    runtime_id: string;
-    
-    address: string;
-    ready: boolean;
-    
-    last_seen: string;
-    status?: string | null;
-}
-
-interface DialTarget {
-    appSlug: string;
-    fnName?: string;
-    sharedApp?: boolean;
-    modelId?: string;
-    variant?: string;
-    sharedOrgId?: string;
-}
-interface StatelessDecisionDialOptions {
-    
-    targetFor: (model: string | undefined) => Promise<DialTarget>;
-    apiUrl: string;
-    apiKey: string;
-    
-    orgId: string;
-    
-    fnName: string;
-    
-    secret: string;
-    
-    shared?: {
-        apiKey: string;
-    };
-    fetchImpl?: typeof fetch;
-    nowMs?: () => number;
-    
-    timeoutMs?: number;
-}
-
-declare class StatelessDecisionDial {
-    private readonly opts;
-    private readonly replicas;
-    private readonly nowMs;
-    
-    private readonly presenceCache;
-    constructor(opts: StatelessDecisionDialOptions);
-    
-    get enabled(): boolean;
-    
-    targetFor(model: string | undefined): Promise<DialTarget | null>;
-    
-    decide(body: Record<string, unknown>, model: string | undefined): Promise<{
-        status: number;
-        body: unknown;
-        completed: unknown;
-        target: DialTarget & {
-            fnName: string;
-        };
-    }>;
-    
-    private ensureRows;
-    private readPresence;
-    
-    pick(rows: RuntimePresenceRow[]): string | null;
-    private post;
-}
-
 interface InferenceUsageRecord {
     
     inference_id: string;
@@ -634,8 +565,6 @@ interface ProxyClients {
     createResponseOn(handle: string, params: ResponsesCreateParams): Promise<AsyncIterable<unknown>> | AsyncIterable<unknown>;
     
     onSessionEnd(handle: string, cb: (end: SessionEndInfo) => void): Promise<() => void>;
-    
-    readonly stateless?: StatelessDecisionDial;
 }
 
 type ProxyAudioLane = Pick<AudioBridge, 'appendInputAudio' | 'onOutputAudio'> & {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@urun-sh/openai",
-  "version": "0.6.18",
+  "version": "0.6.20",
   "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": "^0.6.18"
+    "@urun-sh/core": "^0.6.20"
   },
   "dependencies": {
     "@prometheus-io/client": "^0.16.1",