@urun-sh/openai 0.5.6 → 0.6.0

package/dist/{ResponsesClient-BE3-hx3m.d.ts → ResponsesClient-CSSrOYD8.d.ts}

@@ -1,4 +1,4 @@
-import { b as UrunSessionLike } from './types-lsVTbNcH.js';
+import { b as UrunSessionLike } from './types-CHPtJx6d.js';
 
 interface ResponsesStream extends AsyncIterable<unknown> {
     readonly requestId: string;
@@ -11,6 +11,8 @@ declare class UrunResponses {
             model?: string;
             input: unknown;
             stream?: boolean;
+            
+            instructions?: string;
             tools?: unknown;
             tool_choice?: unknown;
             temperature?: number;
@@ -21,6 +23,8 @@ declare class UrunResponses {
             chat_template_kwargs?: Record<string, unknown>;
             
             response_format?: unknown;
+            
+            voice?: string;
         }) => Promise<ResponsesStream>;
     };
 }

package/dist/{ResponsesClient-DjZWjlFH.d.cts → ResponsesClient-_OZERUjH.d.cts}

@@ -1,4 +1,4 @@
-import { b as UrunSessionLike } from './types-lsVTbNcH.cjs';
+import { b as UrunSessionLike } from './types-CHPtJx6d.cjs';
 
 interface ResponsesStream extends AsyncIterable<unknown> {
     readonly requestId: string;
@@ -11,6 +11,8 @@ declare class UrunResponses {
             model?: string;
             input: unknown;
             stream?: boolean;
+            /** Per-request system instructions — forwarded verbatim to the serve envelope. */
+            instructions?: string;
             tools?: unknown;
             tool_choice?: unknown;
             temperature?: number;
@@ -21,6 +23,12 @@ declare class UrunResponses {
             chat_template_kwargs?: Record<string, unknown>;
             /** ENG-310 structured output: forwarded verbatim; the serve runtime validates. */
             response_format?: unknown;
+            /**
+             * ENG-331 speech lane: the per-request TTS voice, forwarded verbatim
+             * on the serve envelope (encode.ts `ServeEnvelope.voice`); absent →
+             * key absent — the engine's catalog-configured default voice applies.
+             */
+            voice?: string;
         }) => Promise<ResponsesStream>;
     };
 }

package/dist/{chunk-7M6LY6DR.js → chunk-3JWYIKHM.js}

@@ -1,2 +1,2 @@
-import{a as I,o as A}from"./chunk-BSHT6RZZ.js";import{c as v}from"./chunk-5CCJUNH7.js";v();var J="/ws/google.ai.generativelanguage.v1beta.GenerativeService.BidiGenerateContent",k=1007,F=1008,R=1011,V=1008,n=class extends Error{constructor(o,i=k){super(o);this.closeCode=i}closeCode},d=e=>typeof e=="object"&&e!==null&&!Array.isArray(e)?e:null,w=["setup","clientContent","realtimeInput","toolResponse"];function B(e){let t;try{t=JSON.parse(e)}catch{throw new n("client message is not valid JSON")}let o=d(t);if(!o)throw new n("client message must be a JSON object");let i=w.filter(c=>o[c]!==void 0),s=Object.keys(o).filter(c=>!w.includes(c));if(s.length>0)throw new n(`unknown client message field(s) ${JSON.stringify(s)} \u2014 expected exactly one of ${w.join(", ")}`);if(i.length!==1)throw new n(`client message must contain exactly one of ${w.join(", ")} (got ${i.length})`);let r=i[0],a=d(o[r]);if(!a)throw new n(`"${r}" must be a JSON object`);return{kind:r,[r]:a}}function C(e,t){if(typeof e=="string")return e;let o=d(e),i=o?o.parts:void 0;if(!Array.isArray(i))throw new n(`${t}: expected a Content with a parts array`);let s=[];for(let r of i){let a=d(r);if(a&&typeof a.text=="string"){s.push(a.text);continue}throw a&&(a.inlineData!==void 0||a.fileData!==void 0)?new n(`${t}: media parts (inlineData/fileData) are a declared follow-up slice (audio/video lanes over the proxy transport AudioBridge) \u2014 only text parts are supported in this slice`):a&&(a.functionCall!==void 0||a.functionResponse!==void 0)?new n(`${t}: functionCall/functionResponse parts are not valid here \u2014 the Live API exchanges tool traffic via the dedicated toolCall/toolResponse messages`):new n(`${t}: unsupported part ${JSON.stringify(r)}`)}return s.join(`
+import{a as I,o as A}from"./chunk-6M5JK4YC.js";import{c as v}from"./chunk-YSFSRI3D.js";v();var J="/ws/google.ai.generativelanguage.v1beta.GenerativeService.BidiGenerateContent",k=1007,F=1008,R=1011,V=1008,n=class extends Error{constructor(o,i=k){super(o);this.closeCode=i}closeCode},d=e=>typeof e=="object"&&e!==null&&!Array.isArray(e)?e:null,w=["setup","clientContent","realtimeInput","toolResponse"];function B(e){let t;try{t=JSON.parse(e)}catch{throw new n("client message is not valid JSON")}let o=d(t);if(!o)throw new n("client message must be a JSON object");let i=w.filter(c=>o[c]!==void 0),s=Object.keys(o).filter(c=>!w.includes(c));if(s.length>0)throw new n(`unknown client message field(s) ${JSON.stringify(s)} \u2014 expected exactly one of ${w.join(", ")}`);if(i.length!==1)throw new n(`client message must contain exactly one of ${w.join(", ")} (got ${i.length})`);let r=i[0],a=d(o[r]);if(!a)throw new n(`"${r}" must be a JSON object`);return{kind:r,[r]:a}}function C(e,t){if(typeof e=="string")return e;let o=d(e),i=o?o.parts:void 0;if(!Array.isArray(i))throw new n(`${t}: expected a Content with a parts array`);let s=[];for(let r of i){let a=d(r);if(a&&typeof a.text=="string"){s.push(a.text);continue}throw a&&(a.inlineData!==void 0||a.fileData!==void 0)?new n(`${t}: media parts (inlineData/fileData) are a declared follow-up slice (audio/video lanes over the proxy transport AudioBridge) \u2014 only text parts are supported in this slice`):a&&(a.functionCall!==void 0||a.functionResponse!==void 0)?new n(`${t}: functionCall/functionResponse parts are not valid here \u2014 the Live API exchanges tool traffic via the dedicated toolCall/toolResponse messages`):new n(`${t}: unsupported part ${JSON.stringify(r)}`)}return s.join(`
 `)}var _={manualActivity:!1,activityHandling:"START_OF_ACTIVITY_INTERRUPTS"},h="urun.videoOut";function G(e){return{framing:A,codec:e}}function Y(e){let{seq:t,ptsUs:o,codec:i,key:s}=e.header;return{[h]:{v:1,seq:t,ptsUs:o,codec:i,key:s,data:Buffer.from(e.au).toString("base64")}}}var x={inputAudioTranscription:"setup.inputAudioTranscription is part of the declared audio follow-up slice",outputAudioTranscription:"setup.outputAudioTranscription is part of the declared audio follow-up slice"},S=["model","generationConfig","systemInstruction","tools","realtimeInputConfig","sessionResumption",h],T=["START_OF_ACTIVITY_INTERRUPTS","NO_INTERRUPTION"];function E(e){let t=d(e);if(!t)throw new n("setup.realtimeInputConfig must be an object");for(let s of Object.keys(t))if(!(s==="automaticActivityDetection"||s==="activityHandling"))throw new n(s==="turnCoverage"?"realtimeInputConfig.turnCoverage is part of the declared audio follow-up slice":`unsupported realtimeInputConfig field "${s}"`);let o=d(t.automaticActivityDetection);if(t.automaticActivityDetection!==void 0&&!o)throw new n("realtimeInputConfig.automaticActivityDetection must be an object");for(let s of Object.keys(o??{}))if(s!=="disabled")throw new n(`automaticActivityDetection.${s} tunes AUTOMATIC (server-side) activity detection \u2014 the declared platform follow-up (serve turn_detection events over the session stream lane); only "disabled: true" (manual activity mode) is supported in this slice`);if(o?.disabled!==!0)throw new n("automatic (server-side) activity detection is the declared platform follow-up (serve turn_detection turn_start/turn_end events over the session stream lane, not proxy DSP) \u2014 set realtimeInputConfig.automaticActivityDetection.disabled: true for manual activity mode");let i="START_OF_ACTIVITY_INTERRUPTS";if(t.activityHandling!==void 0){if(!T.includes(t.activityHandling))throw new n(`unsupported activityHandling ${JSON.stringify(t.activityHandling)} \u2014 expected one of ${T.join(", ")}`);i=t.activityHandling}return{manualActivity:!0,activityHandling:i}}var N=["temperature","maxOutputTokens","responseModalities"];function K(e){for(let[u,f]of Object.entries(x))if(e[u]!==void 0)throw new n(f);for(let u of Object.keys(e))if(!S.includes(u))throw new n(`unsupported setup field "${u}"`);if(typeof e.model!="string"||e.model.length===0)throw new n('setup.model is required (format "models/{model}")');let t=e.model.replace(/^models\//,""),o,i,s=!1,r=d(e.generationConfig);if(e.generationConfig!==void 0&&!r)throw new n("setup.generationConfig must be an object");if(r){for(let u of Object.keys(r))if(!N.includes(u))throw new n(`unsupported generationConfig field "${u}"`);if(r.temperature!==void 0){if(typeof r.temperature!="number")throw new n("generationConfig.temperature must be a number");o=r.temperature}if(r.maxOutputTokens!==void 0){if(typeof r.maxOutputTokens!="number")throw new n("generationConfig.maxOutputTokens must be a number");i=r.maxOutputTokens}if(r.responseModalities!==void 0){let u=Array.isArray(r.responseModalities)?r.responseModalities:[r.responseModalities];for(let f of u){let y=String(f).toUpperCase();if(y==="AUDIO"){s=!0;continue}if(y!=="TEXT")throw new n(`responseModalities ${JSON.stringify(f)} is not supported \u2014 this surface speaks TEXT and AUDIO (audio-out rides the native AudioBridge 'audio' lane)`)}}}let a=[];e.systemInstruction!==void 0&&a.push({role:"system",content:C(e.systemInstruction,"setup.systemInstruction")});let c;if(e.tools!==void 0){if(!Array.isArray(e.tools))throw new n("setup.tools must be an array");c=[];for(let u of e.tools){let f=d(u);if(!f)throw new n("setup.tools entries must be objects");for(let y of Object.keys(f))if(y!=="functionDeclarations")throw new n(`unsupported tool "${y}" \u2014 this slice translates functionDeclarations only (no googleSearch/codeExecution)`);if(!Array.isArray(f.functionDeclarations))throw new n("tool.functionDeclarations must be an array");for(let y of f.functionDeclarations){let g=d(y);if(!g||typeof g.name!="string")throw new n("functionDeclarations entries require a string name");c.push({type:"function",name:g.name,description:g.description,parameters:g.parameters})}}c.length===0&&(c=void 0)}let m={enabled:!1,handle:null};if(e.sessionResumption!==void 0){let u=d(e.sessionResumption);if(!u)throw new n("setup.sessionResumption must be an object (SessionResumptionConfig)");for(let f of Object.keys(u))if(f!=="handle")throw new n(`unsupported sessionResumption field "${f}" \u2014 this lane implements SessionResumptionConfig.handle only`);if(u.handle!==void 0&&(typeof u.handle!="string"||u.handle.length===0))throw new n("sessionResumption.handle must be a non-empty string when present");m={enabled:!0,handle:typeof u.handle=="string"?u.handle:null}}let p=e.realtimeInputConfig===void 0?_:E(e.realtimeInputConfig),l=!1;if(e[h]!==void 0){if(!d(e[h]))throw new n(`setup["${h}"] must be an object (the v1 opt-in is {})`);l=!0}return{model:t,systemItems:a,tools:c,temperature:o,maxOutputTokens:i,audioModality:s,activity:p,sessionResumption:m,videoOut:l}}function q(e){return{sessionResumptionUpdate:{newHandle:e,resumable:!0}}}function W(e){let t=Math.max(0,e??0)/1e3;return{goAway:{timeLeft:`${Math.round(t*1e3)/1e3}s`}}}function X(e){for(let i of Object.keys(e))if(i!=="turns"&&i!=="turnComplete")throw new n(`unsupported clientContent field "${i}"`);let t=e.turns===void 0?[]:e.turns;if(!Array.isArray(t))throw new n("clientContent.turns must be an array");let o=[];for(let i of t){let s=d(i);if(!s)throw new n("clientContent.turns entries must be Content objects");let r=s.role===void 0?"user":String(s.role).toLowerCase();if(r!=="user"&&r!=="model")throw new n(`clientContent turn role must be "user" or "model" (got ${JSON.stringify(s.role)})`);o.push({role:r==="model"?"assistant":"user",content:C(s,"clientContent.turns")})}return{items:o,turnComplete:e.turnComplete===!0}}var b=16e3,O=24e3,$=`audio/pcm;rate=${O}`,z=Buffer.alloc(960).toString("base64");function j(e,t,o){if(t===o)return e;let i=Math.floor(e.length*o/t),s=new Int16Array(i),r=t/o,a=e.length-1;for(let c=0;c<i;c++){let m=c*r,p=Math.floor(m);if(p>=a){s[c]=e[a];continue}let l=m-p;s[c]=Math.round(e[p]+(e[p+1]-e[p])*l)}return s}function U(e){let t=d(e);if(!t)throw new n("realtimeInput.video must be a Blob object {mimeType, data}");if(typeof t.mimeType!="string"||!/^image\/jpeg\s*(;|$)/.test(t.mimeType))throw new n(`realtimeInput.video.mimeType must be image/jpeg (the Live video input stream is individual JPEG frames), got ${JSON.stringify(t.mimeType)}`);if(typeof t.data!="string"||t.data.length===0)throw new n("realtimeInput.video.data must be non-empty base64");let o=Buffer.from(t.data,"base64");if(o.length<3||o[0]!==255||o[1]!==216)throw new n("realtimeInput.video.data does not decode to a JPEG frame (missing FF D8 SOI marker)");return new Uint8Array(o)}function D(e){let t=d(e);if(!t)throw new n("realtimeInput.audio must be a Blob object {mimeType, data}");if(typeof t.mimeType!="string"||!/^audio\/pcm\s*(;|$)/.test(t.mimeType))throw new n(`realtimeInput.audio.mimeType must be audio/pcm (16-bit LE PCM @16kHz mono per the Live spec), got ${JSON.stringify(t.mimeType)}`);let o=/;\s*rate=(\d+)/.exec(t.mimeType);if(o&&Number(o[1])!==b)throw new n(`realtimeInput.audio.mimeType declares rate=${o[1]} \u2014 the Live spec input format is ${b} Hz (audio/pcm;rate=${b})`);if(typeof t.data!="string"||t.data.length===0)throw new n("realtimeInput.audio.data must be non-empty base64");let i=Buffer.from(t.data,"base64");if(i.length===0||i.length%2!==0)throw new n(`realtimeInput.audio.data must decode to whole 16-bit samples (got ${i.length} bytes \u2014 not even)`);let s=new Int16Array(i.length/2);for(let a=0;a<s.length;a++)s[a]=i.readInt16LE(a*2);let r=j(s,b,O);return Buffer.from(r.buffer,r.byteOffset,r.byteLength).toString("base64")}function Q(e){if(e.mediaChunks!==void 0)throw new n('realtimeInput.mediaChunks is deprecated by the Live API \u2014 send the audio stream via realtimeInput.audio ({mimeType:"audio/pcm;rate=16000", data}) instead');if(e.activityStart!==void 0||e.activityEnd!==void 0){if(Object.keys(e).length!==1||e.activityStart!==void 0&&e.activityEnd!==void 0)throw new n("realtimeInput must carry exactly one field when signaling activity (activityStart XOR activityEnd, nothing else)");let s=e.activityStart!==void 0?"activityStart":"activityEnd",r=d(e[s]);if(!r||Object.keys(r).length!==0)throw new n(`realtimeInput.${s} must be an empty object (ActivityStart/ActivityEnd are bare markers)`);return{kind:s}}let t=["text","audio","audioStreamEnd","video"];for(let i of Object.keys(e))if(!t.includes(i))throw new n(`unsupported realtimeInput field "${i}"`);let o=t.filter(i=>e[i]!==void 0);if(o.length!==1)throw new n(`realtimeInput must carry exactly one of ${t.join(", ")} (got ${o.length})`);if(e.audio!==void 0)return{kind:"audio",base64Pcm24k:D(e.audio)};if(e.video!==void 0)return{kind:"video",jpegFrame:U(e.video)};if(e.audioStreamEnd!==void 0){if(e.audioStreamEnd!==!0)throw new n("realtimeInput.audioStreamEnd must be true when present");return{kind:"audioStreamEnd"}}if(typeof e.text!="string"||e.text.length===0)throw new n("realtimeInput.text must be a non-empty string");return{kind:"text",text:e.text}}function Z(e){for(let t of Object.keys(e))if(t!=="functionResponses")throw new n(`unsupported toolResponse field "${t}"`);if(!Array.isArray(e.functionResponses)||e.functionResponses.length===0)throw new n("toolResponse.functionResponses must be a non-empty array");return e.functionResponses.map(t=>{let o=d(t);if(!o||typeof o.id!="string"||o.id.length===0)throw new n("functionResponses entries require the string id matched from the toolCall");return{type:"function_call_output",call_id:o.id,output:JSON.stringify(o.response??null)}})}function ee(e){return e.type!=="response.output_text.delta"||typeof e.delta!="string"?null:{serverContent:{modelTurn:{role:"model",parts:[{text:e.delta}]}}}}function te(e){return{serverContent:{modelTurn:{role:"model",parts:[{inlineData:{mimeType:$,data:e}}]}}}}function ne(e){let t=d(e)??{},o=Array.isArray(t.output)?t.output:[],i=[],s=[],r=typeof t.output_text=="string"?t.output_text:"";if(!r)for(let p of o){let l=d(p);if(!(l?.type!=="message"||!Array.isArray(l.content)))for(let u of l.content){let f=d(u);f?.type==="output_text"&&typeof f.text=="string"&&(r+=f.text)}}r&&s.push({role:"assistant",content:r});for(let p of o){let l=d(p);if(l?.type!=="function_call")continue;if(typeof l.call_id!="string"||typeof l.name!="string")throw new n(`upstream function_call item is missing call_id/name: ${JSON.stringify(p)}`,R);let u=typeof l.arguments=="string"&&l.arguments.length>0?l.arguments:"{}",f;try{f=JSON.parse(u)}catch{throw new n(`upstream function_call arguments are not valid JSON: ${u}`,R)}i.push({id:l.call_id,name:l.name,args:f}),s.push({type:"function_call",call_id:l.call_id,name:l.name,arguments:u})}let a=[];i.length>0&&a.push({toolCall:{functionCalls:i}}),a.push({serverContent:{generationComplete:!0}});let c=d(t.usage),m={serverContent:{turnComplete:!0}};if(c){let p={};typeof c.input_tokens=="number"&&(p.promptTokenCount=c.input_tokens),typeof c.output_tokens=="number"&&(p.responseTokenCount=c.output_tokens),typeof c.total_tokens=="number"&&(p.totalTokenCount=c.total_tokens),Object.keys(p).length>0&&(m.usageMetadata=p)}return a.push(m),{messages:a,assistantItems:s}}export{J as a,k as b,F as c,R as d,V as e,n as f,B as g,h,G as i,Y as j,K as k,q as l,W as m,X as n,b as o,O as p,$ as q,z as r,j as s,Q as t,Z as u,ee as v,te as w,ne as x};

package/dist/chunk-45FJ5GTP.js

@@ -0,0 +1,7 @@
+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-BSHT6RZZ.js → chunk-6M5JK4YC.js}

@@ -1 +1 @@
-import{c as m}from"./chunk-5CCJUNH7.js";m();var W=24e3,b="audio",w=class{constructor(r,t){this.backend=r;this.opts=t}backend;opts;source=null;sink=null;outHandlers=new Set;async startOutbound(){return this.source=this.backend.createSource(),this.source.track}appendInputAudio(r){if(!this.source)throw new Error("startOutbound() not called");let t=Buffer.from(r,"base64"),n=Math.floor(t.byteLength/2),s=new Int16Array(n);for(let a=0;a<n;a++)s[a]=t.readInt16LE(a*2);this.source.onData(s)}async startInbound(r){if(r==null)throw new Error('startInbound: no downstream audio track \u2014 the runtime audio producer has not been consumed yet (session.stream("audio").track is null)');this.sink=this.backend.createSink(r),this.sink.onframe=t=>{let n=Buffer.from(t.samples.buffer,t.samples.byteOffset,t.samples.byteLength).toString("base64");for(let s of this.outHandlers)s(n)}}onOutputAudio(r){return this.outHandlers.add(r),()=>this.outHandlers.delete(r)}};async function H(e,r,t=1e4){let n=new w(r,{sampleRate:24e3}),s=await n.startOutbound();await e.stream(b).attach(s);let a=await V(e,t);return await n.startInbound(a),n}function V(e,r){let t=e.stream(b);return t.track!=null?Promise.resolve(t.track):new Promise((n,s)=>{let a,c=setTimeout(()=>{a?.(),s(new Error(`enableSessionAudio: no downstream audio track within ${r}ms (runtime produced no voice OUT / SFU audio consumer never attached)`))},r),u=o=>{o!=null&&(clearTimeout(c),a?.(),n(o))};a=t.on?.("track",u),t.track!=null&&u(t.track)})}var h="rt-video-in",y=256e3;function N(e){let r=e.stream(h),t=r.emit;if(typeof t!="function")throw new Error(`enableSessionVideo: session.stream('${h}') has no emit() \u2014 the named-DATA stream produce seam (@urun-sh/core SessionStream.emit, contract \xA75) is required for the video frame lane`);return{sendInputFrame:async n=>{if(n.byteLength===0)throw new Error("enableSessionVideo: refusing to send an empty video frame");if(n.byteLength>y)throw new Error(`enableSessionVideo: frame is ${n.byteLength} bytes \u2014 over the ${y}-byte data-channel budget (the SFU's 262144-byte SCTP ceiling drops oversize messages SILENTLY; send smaller/lower-quality frames)`);await t.call(r,n)}}}var T=24e3,l=480,I=111;async function L(){let e=await import("opusscript"),r=e.default??e;return new r(T,1,2048)}async function $(){let e=await import("werift"),r=await L();return{createSource:()=>{let t=new e.MediaStreamTrack({kind:"audio"}),n=Math.random()*4294967295>>>0,s=Math.random()*65535&65535,a=Math.random()*4294967295>>>0,c=new Int16Array(0);return{track:t,onData:u=>{let o=new Int16Array(c.length+u.length);o.set(c,0),o.set(u,c.length);let i=0;for(;o.length-i>=l;){let d=o.subarray(i,i+l);i+=l;let f=Buffer.from(d.buffer,d.byteOffset,d.byteLength),S=r.encode(f,l);s=s+1&65535,a=a+l>>>0;let E=new e.RtpHeader({version:2,payloadType:I,sequenceNumber:s,timestamp:a,ssrc:n,marker:!1}),A=new e.RtpPacket(E,S);t.writeRtp(A)}c=o.slice(i)}}},createSink:t=>{let n={onframe:null};return t.onReceiveRtp.subscribe(a=>{let c;try{c=r.decode(a.payload)}catch{return}let u=new Int16Array(c.length/2);for(let o=0;o<u.length;o++)u[o]=c.readInt16LE(o*2);n.onframe?.({samples:u})}),n}}}async function M(){let e=await import("@roamhq/wrtc");return{createSource:()=>{let r=new e.nonstandard.RTCAudioSource;return{track:r.createTrack(),onData:n=>r.onData({samples:n,sampleRate:24e3})}},createSink:r=>{let t={onframe:null},n=new e.nonstandard.RTCAudioSink(r);return n.ondata=s=>{t.onframe?.(s)},t}}}m();import{WtControlDecoder as U,consumeWtDataStreams as _,encodeWtControlFrame as x,STREAM_END as D}from"@urun-sh/core";var p="rt-video-out",P="urun.videoOut.v1",g="h264";function k(e,r){if(!Number.isInteger(e.seq)||e.seq<0)throw new Error(`${r}: header seq must be a non-negative integer (got ${e.seq})`);if(!Number.isInteger(e.ptsUs))throw new Error(`${r}: header ptsUs must be an integer microsecond timestamp (got ${e.ptsUs})`);if(typeof e.codec!="string"||e.codec.length===0)throw new Error(`${r}: header codec must be a non-empty string`);if(typeof e.key!="boolean")throw new Error(`${r}: header key must be a boolean`)}function J(e,r){if(k(e,"encodeVideoOutRecord"),r.byteLength===0)throw new Error("encodeVideoOutRecord: refusing to encode an empty access unit");let t=new TextEncoder().encode(JSON.stringify({v:1,seq:e.seq,ptsUs:e.ptsUs,codec:e.codec,key:e.key}));if(t.byteLength>65535)throw new Error(`encodeVideoOutRecord: header JSON is ${t.byteLength} bytes \u2014 over the u16 length prefix`);let n=new Uint8Array(2+t.byteLength+r.byteLength);return n[0]=t.byteLength&255,n[1]=t.byteLength>>8,n.set(t,2),n.set(r,2+t.byteLength),n}function O(e){if(e.byteLength<2)throw new Error(`decodeVideoOutRecord: truncated record (${e.byteLength} byte(s), no header length)`);let r=e[0]|e[1]<<8;if(e.byteLength<2+r)throw new Error(`decodeVideoOutRecord: truncated record (header says ${r} bytes, ${e.byteLength-2} present)`);let t;try{t=JSON.parse(new TextDecoder().decode(e.subarray(2,2+r)))}catch(c){throw new Error(`decodeVideoOutRecord: header is not valid JSON (${String(c)})`)}let n=t;if(n===null||typeof n!="object"||n.v!==1)throw new Error(`decodeVideoOutRecord: header is not spec v1 (v=${JSON.stringify(n?.v)}) \u2014 a newer producer must be met with a newer consumer, never silently misread`);let s=n;k(s,"decodeVideoOutRecord");let a=e.subarray(2+r);if(a.byteLength===0)throw new Error("decodeVideoOutRecord: empty access unit \u2014 a record with no AU bytes is a producer bug");return{header:s,au:a.slice()}}async function Y(e){let r=e.wtDataEndpoint;if(typeof r!="function")throw new Error("negotiateSessionVideoOutQuic: the core session exposes no wtDataEndpoint() \u2014 the \xA75-over-QUIC dial seam (urun-infra#1608 TransportOfferResponse.webtransport, offered with wtData: true) is required for the video OUT lane; there is no WS fallback");let t=await r.call(e),n=t.sessionId??e.sessionId;if(typeof n!="string"||n.length===0)throw new Error("negotiateSessionVideoOutQuic: no session id \u2014 the WT hello requires one and neither the negotiated endpoint nor the core session carries it");return{transport:"webtransport",url:t.url,certHashes:t.certHashes,token:t.token,sessionId:n,stream:p,codec:g,framing:P}}var G="urun.videoOut.v2-ws";function Q(e){let r=e.stream(p);if(typeof r.messages!="function")throw new Error(`openSessionVideoOutLane: session.stream('${p}') has no messages() \u2014 the named-DATA stream consume seam (@urun-sh/core SessionStream.messages, contract \xA75) is required for the video OUT lane`);let t=new Set,n=new Set,s=new Set,a=o=>{for(let i of n)i(o)},c=!1,u=async()=>{for await(let o of r.messages()){if(!(o instanceof Uint8Array)){a(new Error(`openSessionVideoOutLane: "${p}" delivered a non-BINARY \xA75 payload (${typeof o}) \u2014 spec v1 records ride BINARY envelopes only`));continue}let i;try{i=O(o)}catch(d){a(d instanceof Error?d:new Error(String(d)));continue}for(let d of t)d(i)}for(let o of s)o()};return{codec:g,onOutputFrame:o=>(t.add(o),c||(c=!0,u().catch(i=>a(i instanceof Error?i:new Error(String(i))))),()=>t.delete(o)),onLaneError:o=>(n.add(o),()=>n.delete(o)),onEnd:o=>(s.add(o),()=>s.delete(o))}}function B(e){if(e.length%2!==0||/[^0-9a-fA-F]/.test(e))throw new Error(`consumeVideoOut: certHashes entry is not hex sha-256 (${JSON.stringify(e)})`);let r=new Uint8Array(e.length/2);for(let t=0;t<r.length;t++)r[t]=parseInt(e.slice(t*2,t*2+2),16);return r}async function j(e,r,t){let n=t??(typeof globalThis.WebTransport=="function"?(i,d)=>new globalThis.WebTransport(i,d):void 0);if(!n)throw new Error("consumeVideoOut: no WebTransport implementation \u2014 this platform has no globalThis.WebTransport; inject a connect factory (the third argument)");let s=n(e.url,{serverCertificateHashes:e.certHashes.map(i=>({algorithm:"sha-256",value:B(i)}))});await s.ready;let a=await s.createBidirectionalStream();await a.writable.getWriter().write(x({t:"hello",token:e.token,sessionId:e.sessionId}));let c=a.readable.getReader(),u=new U,o=!1;for(;!o;){let{value:i,done:d}=await c.read();if(d)throw new Error("consumeVideoOut: gateway closed the control stream before hello-ack");if(i)for(let f of u.push(i)){if(f.t==="hello-ack"){o=!0;break}throw new Error(`consumeVideoOut: gateway rejected hello (got ${JSON.stringify(f.t)} before hello-ack)`)}}await _(s.incomingUnidirectionalStreams,{onData:(i,d)=>{if(i===e.stream){if(d===D){r.onEnd?.();return}if(!(d instanceof Uint8Array)){r.onLaneError(new Error(`consumeVideoOut: "${e.stream}" delivered a non-BINARY \xA75 payload (${typeof d}) \u2014 spec v1 records ride BINARY envelopes only`));return}try{r.onFrame(O(d))}catch(f){r.onLaneError(f instanceof Error?f:new Error(String(f)))}}},onLaneError:i=>r.onLaneError(i)})}export{W as a,w as b,H as c,h as d,y as e,N as f,$ as g,M as h,p as i,P as j,g as k,J as l,O as m,Y as n,G as o,Q as p,j as q};
+import{c as m}from"./chunk-YSFSRI3D.js";m();var W=24e3,b="audio",w=class{constructor(r,t){this.backend=r;this.opts=t}backend;opts;source=null;sink=null;outHandlers=new Set;async startOutbound(){return this.source=this.backend.createSource(),this.source.track}appendInputAudio(r){if(!this.source)throw new Error("startOutbound() not called");let t=Buffer.from(r,"base64"),n=Math.floor(t.byteLength/2),s=new Int16Array(n);for(let a=0;a<n;a++)s[a]=t.readInt16LE(a*2);this.source.onData(s)}async startInbound(r){if(r==null)throw new Error('startInbound: no downstream audio track \u2014 the runtime audio producer has not been consumed yet (session.stream("audio").track is null)');this.sink=this.backend.createSink(r),this.sink.onframe=t=>{let n=Buffer.from(t.samples.buffer,t.samples.byteOffset,t.samples.byteLength).toString("base64");for(let s of this.outHandlers)s(n)}}onOutputAudio(r){return this.outHandlers.add(r),()=>this.outHandlers.delete(r)}};async function H(e,r,t=1e4){let n=new w(r,{sampleRate:24e3}),s=await n.startOutbound();await e.stream(b).attach(s);let a=await V(e,t);return await n.startInbound(a),n}function V(e,r){let t=e.stream(b);return t.track!=null?Promise.resolve(t.track):new Promise((n,s)=>{let a,c=setTimeout(()=>{a?.(),s(new Error(`enableSessionAudio: no downstream audio track within ${r}ms (runtime produced no voice OUT / SFU audio consumer never attached)`))},r),u=o=>{o!=null&&(clearTimeout(c),a?.(),n(o))};a=t.on?.("track",u),t.track!=null&&u(t.track)})}var h="rt-video-in",y=256e3;function N(e){let r=e.stream(h),t=r.emit;if(typeof t!="function")throw new Error(`enableSessionVideo: session.stream('${h}') has no emit() \u2014 the named-DATA stream produce seam (@urun-sh/core SessionStream.emit, contract \xA75) is required for the video frame lane`);return{sendInputFrame:async n=>{if(n.byteLength===0)throw new Error("enableSessionVideo: refusing to send an empty video frame");if(n.byteLength>y)throw new Error(`enableSessionVideo: frame is ${n.byteLength} bytes \u2014 over the ${y}-byte data-channel budget (the SFU's 262144-byte SCTP ceiling drops oversize messages SILENTLY; send smaller/lower-quality frames)`);await t.call(r,n)}}}var T=24e3,l=480,I=111;async function L(){let e=await import("opusscript"),r=e.default??e;return new r(T,1,2048)}async function $(){let e=await import("werift"),r=await L();return{createSource:()=>{let t=new e.MediaStreamTrack({kind:"audio"}),n=Math.random()*4294967295>>>0,s=Math.random()*65535&65535,a=Math.random()*4294967295>>>0,c=new Int16Array(0);return{track:t,onData:u=>{let o=new Int16Array(c.length+u.length);o.set(c,0),o.set(u,c.length);let i=0;for(;o.length-i>=l;){let d=o.subarray(i,i+l);i+=l;let f=Buffer.from(d.buffer,d.byteOffset,d.byteLength),S=r.encode(f,l);s=s+1&65535,a=a+l>>>0;let E=new e.RtpHeader({version:2,payloadType:I,sequenceNumber:s,timestamp:a,ssrc:n,marker:!1}),A=new e.RtpPacket(E,S);t.writeRtp(A)}c=o.slice(i)}}},createSink:t=>{let n={onframe:null};return t.onReceiveRtp.subscribe(a=>{let c;try{c=r.decode(a.payload)}catch{return}let u=new Int16Array(c.length/2);for(let o=0;o<u.length;o++)u[o]=c.readInt16LE(o*2);n.onframe?.({samples:u})}),n}}}async function M(){let e=await import("@roamhq/wrtc");return{createSource:()=>{let r=new e.nonstandard.RTCAudioSource;return{track:r.createTrack(),onData:n=>r.onData({samples:n,sampleRate:24e3})}},createSink:r=>{let t={onframe:null},n=new e.nonstandard.RTCAudioSink(r);return n.ondata=s=>{t.onframe?.(s)},t}}}m();import{WtControlDecoder as U,consumeWtDataStreams as _,encodeWtControlFrame as x,STREAM_END as D}from"@urun-sh/core";var p="rt-video-out",P="urun.videoOut.v1",g="h264";function k(e,r){if(!Number.isInteger(e.seq)||e.seq<0)throw new Error(`${r}: header seq must be a non-negative integer (got ${e.seq})`);if(!Number.isInteger(e.ptsUs))throw new Error(`${r}: header ptsUs must be an integer microsecond timestamp (got ${e.ptsUs})`);if(typeof e.codec!="string"||e.codec.length===0)throw new Error(`${r}: header codec must be a non-empty string`);if(typeof e.key!="boolean")throw new Error(`${r}: header key must be a boolean`)}function J(e,r){if(k(e,"encodeVideoOutRecord"),r.byteLength===0)throw new Error("encodeVideoOutRecord: refusing to encode an empty access unit");let t=new TextEncoder().encode(JSON.stringify({v:1,seq:e.seq,ptsUs:e.ptsUs,codec:e.codec,key:e.key}));if(t.byteLength>65535)throw new Error(`encodeVideoOutRecord: header JSON is ${t.byteLength} bytes \u2014 over the u16 length prefix`);let n=new Uint8Array(2+t.byteLength+r.byteLength);return n[0]=t.byteLength&255,n[1]=t.byteLength>>8,n.set(t,2),n.set(r,2+t.byteLength),n}function O(e){if(e.byteLength<2)throw new Error(`decodeVideoOutRecord: truncated record (${e.byteLength} byte(s), no header length)`);let r=e[0]|e[1]<<8;if(e.byteLength<2+r)throw new Error(`decodeVideoOutRecord: truncated record (header says ${r} bytes, ${e.byteLength-2} present)`);let t;try{t=JSON.parse(new TextDecoder().decode(e.subarray(2,2+r)))}catch(c){throw new Error(`decodeVideoOutRecord: header is not valid JSON (${String(c)})`)}let n=t;if(n===null||typeof n!="object"||n.v!==1)throw new Error(`decodeVideoOutRecord: header is not spec v1 (v=${JSON.stringify(n?.v)}) \u2014 a newer producer must be met with a newer consumer, never silently misread`);let s=n;k(s,"decodeVideoOutRecord");let a=e.subarray(2+r);if(a.byteLength===0)throw new Error("decodeVideoOutRecord: empty access unit \u2014 a record with no AU bytes is a producer bug");return{header:s,au:a.slice()}}async function Y(e){let r=e.wtDataEndpoint;if(typeof r!="function")throw new Error("negotiateSessionVideoOutQuic: the core session exposes no wtDataEndpoint() \u2014 the \xA75-over-QUIC dial seam (urun-infra#1608 TransportOfferResponse.webtransport, offered with wtData: true) is required for the video OUT lane; there is no WS fallback");let t=await r.call(e),n=t.sessionId??e.sessionId;if(typeof n!="string"||n.length===0)throw new Error("negotiateSessionVideoOutQuic: no session id \u2014 the WT hello requires one and neither the negotiated endpoint nor the core session carries it");return{transport:"webtransport",url:t.url,certHashes:t.certHashes,token:t.token,sessionId:n,stream:p,codec:g,framing:P}}var G="urun.videoOut.v2-ws";function Q(e){let r=e.stream(p);if(typeof r.messages!="function")throw new Error(`openSessionVideoOutLane: session.stream('${p}') has no messages() \u2014 the named-DATA stream consume seam (@urun-sh/core SessionStream.messages, contract \xA75) is required for the video OUT lane`);let t=new Set,n=new Set,s=new Set,a=o=>{for(let i of n)i(o)},c=!1,u=async()=>{for await(let o of r.messages()){if(!(o instanceof Uint8Array)){a(new Error(`openSessionVideoOutLane: "${p}" delivered a non-BINARY \xA75 payload (${typeof o}) \u2014 spec v1 records ride BINARY envelopes only`));continue}let i;try{i=O(o)}catch(d){a(d instanceof Error?d:new Error(String(d)));continue}for(let d of t)d(i)}for(let o of s)o()};return{codec:g,onOutputFrame:o=>(t.add(o),c||(c=!0,u().catch(i=>a(i instanceof Error?i:new Error(String(i))))),()=>t.delete(o)),onLaneError:o=>(n.add(o),()=>n.delete(o)),onEnd:o=>(s.add(o),()=>s.delete(o))}}function B(e){if(e.length%2!==0||/[^0-9a-fA-F]/.test(e))throw new Error(`consumeVideoOut: certHashes entry is not hex sha-256 (${JSON.stringify(e)})`);let r=new Uint8Array(e.length/2);for(let t=0;t<r.length;t++)r[t]=parseInt(e.slice(t*2,t*2+2),16);return r}async function j(e,r,t){let n=t??(typeof globalThis.WebTransport=="function"?(i,d)=>new globalThis.WebTransport(i,d):void 0);if(!n)throw new Error("consumeVideoOut: no WebTransport implementation \u2014 this platform has no globalThis.WebTransport; inject a connect factory (the third argument)");let s=n(e.url,{serverCertificateHashes:e.certHashes.map(i=>({algorithm:"sha-256",value:B(i)}))});await s.ready;let a=await s.createBidirectionalStream();await a.writable.getWriter().write(x({t:"hello",token:e.token,sessionId:e.sessionId}));let c=a.readable.getReader(),u=new U,o=!1;for(;!o;){let{value:i,done:d}=await c.read();if(d)throw new Error("consumeVideoOut: gateway closed the control stream before hello-ack");if(i)for(let f of u.push(i)){if(f.t==="hello-ack"){o=!0;break}throw new Error(`consumeVideoOut: gateway rejected hello (got ${JSON.stringify(f.t)} before hello-ack)`)}}await _(s.incomingUnidirectionalStreams,{onData:(i,d)=>{if(i===e.stream){if(d===D){r.onEnd?.();return}if(!(d instanceof Uint8Array)){r.onLaneError(new Error(`consumeVideoOut: "${e.stream}" delivered a non-BINARY \xA75 payload (${typeof d}) \u2014 spec v1 records ride BINARY envelopes only`));return}try{r.onFrame(O(d))}catch(f){r.onLaneError(f instanceof Error?f:new Error(String(f)))}}},onLaneError:i=>r.onLaneError(i)})}export{W as a,w as b,H as c,h as d,y as e,N as f,$ as g,M as h,p as i,P as j,g as k,J as l,O as m,Y as n,G as o,Q as p,j as q};

package/dist/{chunk-5HKWNK3O.js → chunk-K23AZPI4.js}

@@ -1 +1 @@
-import{c as n}from"./chunk-5CCJUNH7.js";n();async function o(e){let t=await(e.fetchImpl??fetch)(`${e.catalogUrl}/model_catalog?select=model_id,variant,gpu_spec,engine,task,description,size_gb,hf_repo,engine_args,updated_at`,{headers:{apikey:e.anonKey,Authorization:`Bearer ${e.anonKey}`}});if(!t.ok)throw new Error(`model_catalog fetch failed: ${t.status}`);return await t.json()}async function l(e){return{object:"list",data:(await o(e)).map(t=>({id:`${t.model_id}:${t.variant}`,object:"model",created:0,owned_by:"urun"}))}}export{o as a,l as b};
+import{c as n}from"./chunk-YSFSRI3D.js";n();async function o(e){let t=await(e.fetchImpl??fetch)(`${e.catalogUrl}/model_catalog?select=model_id,variant,gpu_spec,engine,task,description,size_gb,hf_repo,engine_args,updated_at`,{headers:{apikey:e.anonKey,Authorization:`Bearer ${e.anonKey}`}});if(!t.ok)throw new Error(`model_catalog fetch failed: ${t.status}`);return await t.json()}async function l(e){return{object:"list",data:(await o(e)).map(t=>({id:`${t.model_id}:${t.variant}`,object:"model",created:0,owned_by:"urun"}))}}export{o as a,l as b};

package/dist/chunk-QTVTCMJU.js

@@ -0,0 +1 @@
+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

@@ -0,0 +1 @@
+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};