@urun-sh/openai 0.6.22 → 0.6.23

package/dist/{chunk-HCAKF2SQ.js → chunk-FHGSREBO.js}

@@ -1,4 +1,4 @@
-import{a as Se,c as z,d as Re,e as W}from"./chunk-KSC5Q6PZ.js";import{C as le,D as ce,E as ue,F as de,G as he,K as pe,L as T,M as fe,Q as ge,R as me,S as ye,T as we,U as be,V as ve,W as _e,X as Ee,b as oe,d as se,e as ie,q as ae,r as O}from"./chunk-4W3ZPIR7.js";import{c as te,g as ne,i as K}from"./chunk-CUDFHFNX.js";import{a as re}from"./chunk-3JWYIKHM.js";import{c as p}from"./chunk-YSFSRI3D.js";p();import{Counter as x,Gauge as xe,Histogram as ke,Registry as Ge,collectDefaultMetrics as Be}from"@prometheus-io/client";import{createServer as Ke}from"http";var Bt=9464,Ce="/metrics",A="(other)",ze="(none)",We=64,je=128,Je=["/v1/responses","/v1/chat/completions","/v1/systemone","/v1/messages","/v1/messages/count_tokens","/v1/models","/v1/images/models","/v1/images/generations","/v1/images/edits","/v1/audio/speech","/v1/audio/models","/v1/audio/transcriptions",pe,ge,te],Ve=new Set(Je);function qe(t){return Ve.has(t)?t:A}function Ye(t){return t===null?"none":t>=500?"5xx":t>=400?"4xx":t>=300?"3xx":t>=200?"2xx":"1xx"}var I=class{constructor(e){this.max=e}max;admitted=new Set;overflows=0;of(e,n=!0){return e===null||e===""?ze:this.admitted.has(e)?e:n?this.admitted.size>=this.max?(this.overflows+=1,A):(this.admitted.add(e),e):A}get size(){return this.admitted.size}get overflowCount(){return this.overflows}},P=class{registry=new Ge;modelLabel=new I(We);orgLabel=new I(je);requests;duration;ttft;streamErrors;undelivered;ledgerWrites;ledgerQueue;labelCardinality;labelOverflows;constructor(e={}){this.requests=new x({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 ke({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 ke({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 x({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 x({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 x({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 me)this.ledgerWrites.inc({outcome:n.outcome,reason:n.reason},0);this.ledgerQueue=new xe({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"},ye()),this.ledgerQueue.set({measure:"oldest_age_seconds"},we()/1e3)}}),this.labelCardinality=new xe({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 x({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&&Be({register:this.registry,prefix:"urun_inference_proxy_"})}observeLedger(e){this.ledgerWrites.inc({outcome:e.outcome,reason:e.reason})}observe(e){let n=qe(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 c=e.stream===!0?"true":"false";this.requests.inc({lane:n,method:e.method===""?A:e.method,status_class:Ye(e.status),stream:c,model:a,org:s}),this.duration.observe({lane:n,stream:c,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 Oe(t){return Ke((e,n)=>{(async()=>{let o=(e.url??"/").split("?")[0];if(e.method!=="GET"||o!==Ce){n.writeHead(404,{"content-type":"text/plain"}),n.end(`no route for ${e.method} ${o} \u2014 this port serves only GET ${Ce}
+import{a as Se,c as z,d as Re,e as W}from"./chunk-JOOR4RUT.js";import{C as le,D as ce,E as ue,F as de,G as he,K as pe,L as T,M as fe,Q as ge,R as me,S as ye,T as we,U as be,V as ve,W as _e,X as Ee,b as oe,d as se,e as ie,q as ae,r as O}from"./chunk-ERRG6R4T.js";import{c as te,g as ne,i as K}from"./chunk-CUDFHFNX.js";import{a as re}from"./chunk-3JWYIKHM.js";import{c as p}from"./chunk-YSFSRI3D.js";p();import{Counter as x,Gauge as xe,Histogram as ke,Registry as Ge,collectDefaultMetrics as Be}from"@prometheus-io/client";import{createServer as Ke}from"http";var Bt=9464,Ce="/metrics",A="(other)",ze="(none)",We=64,je=128,Je=["/v1/responses","/v1/chat/completions","/v1/systemone","/v1/messages","/v1/messages/count_tokens","/v1/models","/v1/images/models","/v1/images/generations","/v1/images/edits","/v1/audio/speech","/v1/audio/models","/v1/audio/transcriptions",pe,ge,te],Ve=new Set(Je);function qe(t){return Ve.has(t)?t:A}function Ye(t){return t===null?"none":t>=500?"5xx":t>=400?"4xx":t>=300?"3xx":t>=200?"2xx":"1xx"}var I=class{constructor(e){this.max=e}max;admitted=new Set;overflows=0;of(e,n=!0){return e===null||e===""?ze:this.admitted.has(e)?e:n?this.admitted.size>=this.max?(this.overflows+=1,A):(this.admitted.add(e),e):A}get size(){return this.admitted.size}get overflowCount(){return this.overflows}},P=class{registry=new Ge;modelLabel=new I(We);orgLabel=new I(je);requests;duration;ttft;streamErrors;undelivered;ledgerWrites;ledgerQueue;labelCardinality;labelOverflows;constructor(e={}){this.requests=new x({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 ke({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 ke({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 x({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 x({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 x({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 me)this.ledgerWrites.inc({outcome:n.outcome,reason:n.reason},0);this.ledgerQueue=new xe({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"},ye()),this.ledgerQueue.set({measure:"oldest_age_seconds"},we()/1e3)}}),this.labelCardinality=new xe({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 x({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&&Be({register:this.registry,prefix:"urun_inference_proxy_"})}observeLedger(e){this.ledgerWrites.inc({outcome:e.outcome,reason:e.reason})}observe(e){let n=qe(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 c=e.stream===!0?"true":"false";this.requests.inc({lane:n,method:e.method===""?A:e.method,status_class:Ye(e.status),stream:c,model:a,org:s}),this.duration.observe({lane:n,stream:c,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 Oe(t){return Ke((e,n)=>{(async()=>{let o=(e.url??"/").split("?")[0];if(e.method!=="GET"||o!==Ce){n.writeHead(404,{"content-type":"text/plain"}),n.end(`no route for ${e.method} ${o} \u2014 this port serves only GET ${Ce}
 `);return}let r=await t.scrape();n.writeHead(200,{"content-type":t.contentType}),n.end(r)})().catch(o=>{if(!n.headersSent){n.writeHead(500,{"content-type":"text/plain"}),n.end(`${String(o instanceof Error?o.message:o)}
 `);return}n.end()})})}p();import{createHash as Ae}from"crypto";import{createServer as Qe}from"http";var Jt=9465,Xe="x-urun-key-id",Ze="x-urun-key-fingerprint",et="urun-ratelimit-bucket:",tt="urun-ratelimit-tier:",Te="anonymous",nt=/^urun_sk_[A-Za-z0-9_-]{43}$/;function rt(t){return Ae("sha256").update(`${et}${t}`).digest("hex")}function ot(t){return Ae("sha256").update(`${tt}${t}`).digest("hex")}function st(t){let e=/^Bearer[ \t]+(\S.*)$/i.exec((t??"").trim());if(!e)return null;let n=e[1].trim();return nt.test(n)?n:null}function Ie(){return Qe((t,e)=>{let n=st(t.headers.authorization);e.setHeader(Xe,n===null?Te:rt(n)),e.setHeader(Ze,n===null?Te:ot(n)),e.writeHead(200),e.end()})}p();import{createClientToken as it,ClientTokenError as at}from"@urun-sh/core";var g=class extends Error{constructor(n,o=401){super(n);this.status=o;this.name="ProxyAuthError"}status},S=class extends Error{constructor(e){super(e),this.name="ControlPlaneUnavailableError"}};function k(t){let e=(t??"").trim();if(e==="")throw new g("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 g("malformed Authorization header \u2014 expected `Authorization: Bearer <urun org api key>`");return n[1].trim()}var lt=60,ct=6e4,ut=4096;function dt(t){return t===401||t===403}var N=class{constructor(e){this.opts=e;this.mint=e.mint??it,this.now=e.now??Date.now}opts;cache=new Map;mint;now;async resolve(e){if(e==="")throw new g("empty Authorization bearer token");let n=this.cache.get(e);if(n&&this.now()-n.at<ct)return{apiKey:e,orgId:n.orgId};let o;try{o=(await this.mint(e,{baseUrl:this.opts.gatewayUrl,expiresIn:lt})).orgId}catch(r){throw r instanceof at&&dt(r.status)?(this.cache.delete(e),new g("the uRun org API key was rejected by the control plane \u2014 check the key and that it is not revoked")):new S(`could not verify the API key against the uRun control plane at ${this.opts.gatewayUrl}: ${r instanceof Error?r.message:String(r)}`)}if(!o)throw new S("the control plane verified the API key but returned no org_id \u2014 refusing to route a request with no tenancy (control-plane contract violation)");if(this.cache.size>=ut){let r=this.cache.keys().next();r.done||this.cache.delete(r.value)}return this.cache.set(e,{orgId:o,at:this.now()}),{apiKey:e,orgId:o}}clear(){this.cache.clear()}};p();var ht=8080,j="serve",en="0.0.0.0";function pt(t,e,n){let o=(t[e]??"").trim();if(!o)throw new Error(`${e} is not set \u2014 ${n}`);return o}function tn(t=process.env){let e=(t.PORT??"").trim(),n=e===""?ht:Number(e);if(!Number.isInteger(n)||n<0||n>65535)throw new Error(`PORT must be an integer in 0..65535, got ${JSON.stringify(e)}`);let o=ae(O(pt(t,"URUN_BASE_URL","the hosted proxy needs the uRun session-gateway base it opens sessions and verifies API keys against (e.g. https://api.urun.sh)"),"URUN_BASE_URL")),r=(t.URUN_API_URL??"").trim(),a=r?O(r,"URUN_API_URL"):`${o.replace(/\/+$/,"")}/v1`,s=(t.URUN_CATALOG_URL??"").trim(),i=(t.URUN_CATALOG_ANON_KEY??"").trim();if(!!s!=!!i)throw new Error("URUN_CATALOG_URL and URUN_CATALOG_ANON_KEY must be set together (or both left unset) \u2014 exactly one of them is set, which would silently disable the model-catalog oracle");let c=s?{catalogUrl:O(s,"URUN_CATALOG_URL"),anonKey:i}:null;return{port:n,baseUrl:o,apiUrl:a,catalog:c}}p();import{createHash as ft,randomUUID as gt}from"crypto";var mt=512;function J(t){return`urun-inference-proxy-${ft("sha256").update(t).digest("hex").slice(0,32)}`}var U=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 a=this.opts.build(e);return a.decision&&this.ensureDecisionReaper(),{router:a.router,clients:this.label(a.clients,e),decision:a.decision}}let n={baseUrl:this.opts.baseUrl,orgId:e.orgId,defaultApp:null,fnName:j,sharedOnlyListing:!0,apiUrl:this.opts.apiUrl,auth:{lane:"api-key",apiKey:e.apiKey,gatewayUrl:this.opts.baseUrl},catalog:this.opts.catalog,release:"end",decisionPool:Se(process.env,process.env.HOSTNAME)??void 0},o=z(n),r=n.decisionPool?Re(n,o):null;return r&&this.ensureDecisionReaper(),{router:o,clients:this.label(W(o,void 0,void 0,r??void 0),e),decision:r??void 0}}ensureDecisionReaper(){if(this.decisionReaper!==null)return;let e=setInterval(()=>{for(let n of this.tenants.values())n.decision?.reap()},5e3);e.unref(),this.decisionReaper=e}decisionReaper=null;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>=mt;){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}*decisionPools(){for(let e of this.tenants.values())e.decision&&(yield e.decision)}decisionWarmerFor(e){return this.clientsFor(e),this.tenants.get(e.apiKey)?.decision??null}async closeAll(){this.decisionReaper!==null&&(clearInterval(this.decisionReaper),this.decisionReaper=null);let e=[...this.tenants.values()];this.tenants.clear();let n=[];if(await Promise.all(e.map(async o=>{let r=[o.router.closeAll().catch(a=>{n.push(String(a instanceof Error?a.message:a))})];o.decision&&r.push(o.decision.close().catch(a=>{n.push(String(a instanceof Error?a.message:a))})),await Promise.all(r)})),n.length>0)throw new Error(`failed to close ${n.length} tenant backhaul(s): ${n.join("; ")}`)}},Pe=512,yt=1024,$=class extends Error{constructor(){super(`conversation capacity reached (${Pe} conversations on this replica) \u2014 every slot holds an active or resumable conversation; retry later or scale out`),this.name="ConversationCapacityError"}};function wt(t){return{sessionStartOptions:e=>{if(e.startsWith(K))return;let n=t.get(e);return n===void 0?void 0:{sessionId:n}},onSessionOpen:(e,n)=>{e.startsWith(K)||t.set(e,n)}}}function bt(t,e,n,o){let r=z({baseUrl:t.baseUrl,orgId:e.orgId,defaultApp:null,fnName:j,sharedOnlyListing:!0,apiUrl:t.apiUrl,auth:{lane:"api-key",apiKey:e.apiKey,gatewayUrl:t.baseUrl},catalog:t.catalog,subject:n.subject,release:"detach",...wt(o)});return{clients:W(r),release:async()=>{await r.closeAll()}}}var L=class{constructor(e){this.opts=e}opts;byCaller=new Map;byHandle=new Map;build(e,n){let o={id:n,subject:`${J(e.apiKey)}:live:${n}`},r=new Map,a=this.opts.build?this.opts.build(e,o):bt(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>=Pe;){let e=this.byCaller.keys().next();if(e.done)throw new $;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 $;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,gt())),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 c=r().backhaul.clients.openAudio;if(c===void 0)throw new Error("the conversation backhaul wires no openAudio seam \u2014 voice is unavailable on this deployment");return c(s,i)},a.openVideo=async s=>{let i=r().backhaul.clients.openVideo;if(i===void 0)throw new Error("the conversation backhaul wires no openVideo seam \u2014 video-in is unavailable on this deployment");return i(s)},a.openVideoOut=async s=>{let i=r().backhaul.clients.openVideoOut;if(i===void 0)throw new Error("the conversation backhaul wires no openVideoOut seam \u2014 video-out is unavailable on this deployment");return i(s)},a.supportsAudio=async s=>{let i=r().backhaul.clients.supportsAudio;if(i===void 0)throw new Error("the conversation backhaul wires no supportsAudio seam \u2014 the realtime lane cannot gate audio modality on this deployment");return i(s)},a.supportsSpeech=async s=>{let i=r().backhaul.clients.supportsSpeech;if(i===void 0)throw new Error("the conversation backhaul wires no supportsSpeech seam \u2014 the speech lane cannot gate model capability on this deployment");return i(s)},a.speechVoices=async s=>{let i=r().backhaul.clients.speechVoices;if(i===void 0)throw new Error("the conversation backhaul wires no speechVoices seam \u2014 the speech lane cannot resolve the voice set /v1/models publishes on this deployment");return i(s)},a.speechSessionHandle=async s=>{let i=r().backhaul.clients.speechSessionHandle;if(i===void 0)throw new Error("the conversation backhaul wires no speechSessionHandle seam \u2014 the speech lane cannot resolve its own pooled session on this deployment");return i(s)},a.openAudioOn=async(s,i)=>{let c=r().backhaul.clients.openAudioOn;if(c===void 0)throw new Error("the conversation backhaul wires no openAudioOn seam \u2014 the speech and realtime lanes cannot open their pinned native session audio on this deployment");return c(s,i)},a.openTurnsOn=(s,i)=>{let c=r().backhaul.clients.openTurnsOn;if(c===void 0)throw new Error("the conversation backhaul wires no openTurnsOn seam \u2014 the realtime transcription lane cannot open the backend turn/transcript lane on this deployment");return c(s,i)},a.evictSpeechSession=async s=>{let i=r().backhaul.clients.evictSpeechSession;if(i===void 0)throw new Error("the conversation backhaul wires no evictSpeechSession seam \u2014 the speech lane cannot retire its pinned speech session after a failed turn on this deployment");return i(s)},a.supportsTranscription=async s=>{let i=r().backhaul.clients.supportsTranscription;if(i===void 0)throw new Error("the conversation backhaul wires no supportsTranscription seam \u2014 the transcription lane cannot gate model capability on this deployment");return i(s)},a.transcriptStreamFor=async s=>{let i=r().backhaul.clients.transcriptStreamFor;if(i===void 0)throw new Error("the conversation backhaul wires no transcriptStreamFor seam \u2014 the realtime transcription lane cannot verify the model's engine_args.transcript_stream on this deployment");return i(s)},a.listSpeechToTextModels=async()=>{let s=r().backhaul.clients.listSpeechToTextModels;if(s===void 0)throw new Error("the conversation backhaul wires no listSpeechToTextModels seam \u2014 the transcription lane cannot list the stt models on this deployment");return s()},{clients:a,bindResumeHandle:s=>{let i=n;if(i===null)throw new Error("bindResumeHandle before the conversation opened \u2014 no conversation to bind to");for(;this.byHandle.size>=yt;){let c=this.byHandle.keys().next();if(c.done)break;this.byHandle.delete(c.value)}this.byHandle.set(s,{apiKey:e.apiKey,conversationId:i.id})},clientsForResume:async s=>{let i=this.byHandle.get(s),c=i===void 0?void 0:this.byCaller.get(i.apiKey),u=i===void 0||c===void 0?void 0:c.get(i.conversationId);if(i===void 0||u===void 0||i.apiKey!==e.apiKey)throw new ne("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): ${n.join("; ")}`)}};p();var vt="/inference-ledger",_t=1e4,H=class extends Error{ledgerFailureReason;constructor(e,n="duplicate"){super(e),this.name="LedgerDuplicateError",this.ledgerFailureReason=n}},Y=class extends Error{constructor(n,o){super(o);this.ledgerAlreadyWritten=n;this.name="LedgerAlreadyWrittenError"}ledgerAlreadyWritten},f=class extends Error{ledgerFailureReason="surface";constructor(e){super(e),this.name="LedgerSurfaceError"}},Q=class extends g{ledgerFailureReason="auth";constructor(e){super(e),this.name="LedgerAuthError"}},Et=[["inference_id",t=>typeof t=="string"&&/[^\s]/.test(t)],["model_id",t=>typeof t=="string"&&/[^\s]/.test(t)],["variant",t=>t===null||typeof t=="string"],["task",t=>typeof t=="string"&&/[^\s]/.test(t)],["http_status",t=>typeof t=="number"&&Number.isInteger(t)],["started_at",t=>typeof t=="string"&&!Number.isNaN(Date.parse(t))],["completed_at",t=>typeof t=="string"&&!Number.isNaN(Date.parse(t))],["prompt_tokens",t=>t===null||typeof t=="number"&&Number.isInteger(t)],["completion_tokens",t=>t===null||typeof t=="number"&&Number.isInteger(t)],["output_units",V],["output_minutes",V],["cost_nano_usd",V],["price_version",t=>t===null||typeof t=="string"]];function St(t){for(let[e,n]of Et){if(!(e in t))return`${e} (absent)`;if(!n(t[e]))return`${e} (${JSON.stringify(t[e])})`}return null}function Rt(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"}:St(o)!==null?{status:"unavailable"}:{status:"found",row:o}}function V(t){return t===null?!0:typeof t=="number"?Number.isFinite(t):typeof t!="string"?!1:/[^\s]/.test(t)&&Number.isFinite(Number(t))}function q(t,e){return t===null?e===null:e===null?!1:Number(e)===t}function xt(t,e){return t===null?e===null:e===null?!1:Number(t.toFixed(9))===Number(e)}function Ne(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 kt(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&&Ne(t.started_at,n.started_at)&&Ne(t.completed_at,n.completed_at)&&q(t.prompt_tokens,n.prompt_tokens)&&q(t.completion_tokens,n.completion_tokens)&&q(t.output_units,n.output_units)&&xt(t.output_minutes,n.output_minutes)}function Ct(t,e){if(t===null||typeof t!="object"||Array.isArray(t))throw new f(`${e} answered a "recorded" that is not an object: ${JSON.stringify(t)}`);let n=t,o=n.inference_id;if(typeof o!="string"||o==="")throw new f(`${e} answered a row with no inference_id: ${JSON.stringify(t)}`);if(!("cost_nano_usd"in n))throw new f(`${e} answered a row with no cost_nano_usd field: ${JSON.stringify(t)}`);let r=n.cost_nano_usd;if(r!=null&&!(typeof r=="number"&&Number.isInteger(r)&&r>=0))throw new f(`${e} answered cost_nano_usd ${JSON.stringify(r)}, which is neither a non-negative integer nor null`);let a=u=>{if(!(u in n))throw new f(`${e} answered a row with no ${u} field: ${JSON.stringify(t)}`);let m=n[u];if(m===null)return null;if(typeof m!="string")throw new f(`${e} answered a non-string ${u}: ${JSON.stringify(m)}`);if(!/[^\s]/.test(m))throw new f(`${e} answered a BLANK ${u} (${JSON.stringify(m)}) \u2014 it satisfies "is a string" while naming nothing, which is exactly what inference_requests_priced_attribution_check exists to forbid`);return m},s=a("price_version"),i=a("unpriced_reason"),c=r!=null;if(c&&(s===null||i!==null))throw new f(`${e} answered a PRICED row (cost_nano_usd ${String(r)}) that is not cleanly attributed \u2014 price_version ${JSON.stringify(s)}, unpriced_reason ${JSON.stringify(i)}`);if(!c&&(s!==null||i===null))throw new f(`${e} answered an UNPRICED row that does not say why \u2014 price_version ${JSON.stringify(s)}, unpriced_reason ${JSON.stringify(i)}. NULL means NOT YET PRICED, and the reason is how an operator learns that traffic is being served which nobody can bill.`);return{inference_id:o,cost_nano_usd:r===void 0?null:r,price_version:s,unpriced_reason:i}}var M=class{constructor(e){this.opts=e;this.fetchImpl=e.fetchImpl??fetch}opts;fetchImpl;async write(e,n){let o=`${this.opts.apiUrl.replace(/\/+$/,"")}${vt}`,r;try{r=await this.fetchImpl(o,{method:"POST",headers:{"content-type":"application/json",authorization:`Bearer ${e}`},body:JSON.stringify(n),signal:AbortSignal.timeout(_t)})}catch(c){throw new f(`the uRun ledger surface at ${o} could not be reached: ${c instanceof Error?c.message:String(c)}`)}if(r.status===409){let c=await r.text().catch(()=>""),u=Rt(c);throw u.status==="found"&&kt(n,u.row)?new Y(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 H(`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 H(`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). ${c.slice(0,400)}`)}if(r.status===401||r.status===403)throw await r.text().catch(()=>{}),new Q("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 c=await r.text().catch(u=>`<body unreadable: ${u instanceof Error?u.message:String(u)}>`);throw new f(`the uRun ledger surface at ${o} answered HTTP ${r.status}: ${c.slice(0,400)}`)}let a;try{a=await r.json()}catch(c){throw new f(`the uRun ledger surface at ${o} answered a body that is not JSON: ${c instanceof Error?c.message:String(c)}`)}let s=a?.recorded;if(s===void 0)throw new f(`the uRun ledger surface at ${o} answered a body with no "recorded" field`);let i=Ct(s,`the uRun ledger surface at ${o}`);if(i.inference_id!==n.inference_id)throw new f(`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 Tt}from"http";p();var D="/inference-usage",Ue=25e3;var Ot=new Set(["Error","TypeError","SyntaxError","TimeoutError"]);function $e(t){if(!(t instanceof Error))return`a ${typeof t}`;let e=Ot.has(t.name)?t.name:"Error",n=t.cause?.code;return typeof n=="string"&&/^[A-Z][A-Z0-9_]{0,31}$/.test(n)?`${e} (${n})`:e}var F=class{constructor(e){this.opts=e;this.fetchImpl=e.fetchImpl??fetch}opts;fetchImpl;async lookup(e,n){let o=`${this.opts.apiUrl.replace(/\/+$/,"")}${D}`,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(Ue)})}catch(s){throw new T(`the uRun usage surface (POST ${D}) could not be reached within ${Ue}ms: ${$e(s)}`)}if(r.status===401||r.status===403)throw await r.text().catch(()=>{}),new g("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 T(`the uRun usage surface (POST ${D}) answered HTTP ${r.status} (${i})`)}let a;try{a=await r.json()}catch(s){throw new T(`the uRun usage surface (POST ${D}) answered HTTP ${r.status} with a body that is not JSON: ${$e(s)}`)}return fe(a,this.opts.onRejectedRow)}};var At=5e3,It=1e3,Pt=1500,Nt=3e3,Ut="/api/client-tokens";function X(t,e,n){t.writeHead(e,{"content-type":"application/json"}),t.end(JSON.stringify(n))}function Le(t,e,n,o){X(t,e,{error:{message:n,type:o}})}async function $t(t,e=fetch){let n=`${t.replace(/\/+$/,"")}${Ut}`;try{let o=await e(n,{method:"POST",headers:{"content-type":"application/json"},body:"{}",signal:AbortSignal.timeout(Nt)});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: ${o instanceof Error?o.message:String(o)}`}}}function Pn(t){let e=t.resolver??new N({gatewayUrl:t.baseUrl}),n=t.registry??new U({baseUrl:t.baseUrl,apiUrl:t.apiUrl,catalog:t.catalog}),o=t.probeControlPlane??(()=>$t(t.baseUrl)),r=async()=>{let l=null;for(let h of n.decisionPools()){l=h;break}if(!l)return null;let d=await l.pingReady(Pt);return{ok:d,detail:d?"a decision slot answered its session ping":"no decision slot answered its session ping"}},a=t.usageClient??new F({apiUrl:t.apiUrl}),s=t.ledgerClient??new M({apiUrl:t.apiUrl}),i=process.env.URUN_SHARED_ORG_ID?.trim(),c=process.env.URUN_SHARED_ORG_API_KEY?.trim();i&&c&&(async()=>{try{await n.decisionWarmerFor({orgId:i,apiKey:c})?.warmFor(_e)}catch(l){process.stderr.write(`[decision-warm] out-of-band warm failed (best-effort, retrying on next demand): ${String(l instanceof Error?l.message:l)}
 `)}})();let u=new P(t.defaultMetrics===void 0?{}:{defaultMetrics:t.defaultMetrics}),m=async l=>e.resolve(k(l.headers.authorization)),Z=l=>l instanceof g?{status:l.status,openaiType:"authentication_error",anthropicType:"authentication_error"}:l instanceof S?{status:502,openaiType:"upstream_error",anthropicType:"api_error"}:null,He=Ee({clients:async l=>n.clientsFor(await m(l)),usage:{authorize:l=>k(l.headers.authorization),lookup:(l,d)=>a.lookup(l,d)},statusOf:Z,...t.requestLog?{requestLog:t.requestLog}:{},observe:l=>u.observe(l),ledger:{write:(l,d)=>s.write(k(l.headers.authorization),d),observe:l=>u.observeLedger(l),...t.ledgerLog?{report:t.ledgerLog}:{}}}),_=null,E=null,R=Tt((l,d)=>{let h=new URL(l.url??"/","http://localhost");(async()=>{if(l.method==="GET"&&h.pathname==="/healthz"){X(d,200,{ok:!0});return}if(l.method==="GET"&&h.pathname==="/readyz"){let w=Date.now();if((!_||w-_.at>=At)&&(_={at:w,...await o()}),!E||w-E.at>=It){let B=await r();E=B===null?{at:w,ok:!0,detail:null}:{at:w,ok:B.ok,detail:B.detail}}let ee=(_?.ready??!1)&&E.ok,Fe=E.detail===null?_?.detail??"control plane not yet probed":`${_?.detail??"control plane not yet probed"}; ${E.detail}`;X(d,ee?200:503,{ok:ee,control_plane:Fe});return}let y=h.pathname==="/v1"||h.pathname.startsWith("/v1/"),C=h.pathname.startsWith("/partners/");if(!y&&!C){Le(d,404,`no route for ${l.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}He(l,d)})().catch(y=>{if(!d.headersSent){Le(d,500,String(y instanceof Error?y.message:y),"proxy_error");return}d.end()})}),G=t.conversations??new L({baseUrl:t.baseUrl,apiUrl:t.apiUrl,catalog:t.catalog}),Me=[oe(R,{authenticate:async l=>{let d=await e.resolve(l),h=G.connectionFor(d);return{clients:h.clients,scope:J(d.apiKey),onBindResumeHandle:h.bindResumeHandle,clientsForResume:h.clientsForResume,release:h.release}}}),ie(R,{clients:async l=>n.clientsFor(await m(l)),statusOf:Z}).close,ce(R,{resolver:{resolve:async(l,d)=>{let h=await e.resolve(k(l.headers.authorization)),y=G.connectionFor(h),C;try{C=await he(y.clients,ue(l),d,de(l))}catch(w){throw y.release().catch(()=>{}),w}return{...C,release:y.release}}}})];R.on("upgrade",(l,d)=>{let h=new URL(l.url??"/","http://localhost");h.pathname===re||h.pathname===se||h.pathname===le||d.destroyed||(d.write(`HTTP/1.1 404 Not Found\r

package/dist/chunk-JOOR4RUT.js

@@ -0,0 +1,2 @@
+import{a as M}from"./chunk-K23AZPI4.js";import{I as j,J as B,N as _,O as K,P as q}from"./chunk-ERRG6R4T.js";import{a as F}from"./chunk-76BZLCGT.js";import{a as x,b as G,d as $,g as A,h as W,j as E,k as H,l as V}from"./chunk-CUDFHFNX.js";import{c as R,f as D,g as N,p as U}from"./chunk-6M5JK4YC.js";import{c as v}from"./chunk-YSFSRI3D.js";v();import{App as z,createClientToken as de,COLD_BOOT_WHEN_LIVE_TIMEOUT_MS as ue,DEFAULT_WHEN_LIVE_TIMEOUT_MS as Z,isWakingPhase as pe}from"@urun-sh/core";v();import{SessionFailedError as te}from"@urun-sh/core";var L=2;function se(e){let s=e?.type;return s==="response.created"||s==="response.in_progress"}function ie(e){let s=e?.type;return s==="response.completed"||s==="error"}function Q(e){return async function(a,d){return(async function*(){for(let l=1;l<=L;l++){let{app:p,entry:f,identity:t}=await e.sessionFor(a.model);d?.({sessionId:e.sessionIdOf?.(f)??null,identity:t});let i=!1,c=!1;try{let m=await e.createStream(f,a);for await(let g of m){if(ie(g)&&(i=!0),se(g)){if(l>1)continue}else c=!0;yield g}}catch(m){if(m instanceof te&&(m.kind==="never-live"||m.kind==="create-failed")){try{e.onAdmissionVerdict?.(p,f,m)}catch{}throw m}if(l<L&&!c){await e.evict(p,f);continue}throw m}if(i)return;if(l<L&&!c){await e.evict(p,f);continue}return}})()}}import{randomUUID as fe}from"crypto";v();var re=3,ae="llm",le=new Set(["paused","error"]);function Y(e,s){let a=e.doc(ae),d=typeof a.onConnectionState=="function",u=typeof e.onPhase=="function";if(!d&&!u)throw new Error(`session for app "${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 l=null,p=!1,f=new Set,t=[],i=c=>{if(!l){l=c;for(let m of[...f])m(c);f.clear()}};return d&&t.push(a.onConnectionState(c=>{p&&(c.consecutiveFailures<re||i(new A(`the pooled session for app "${s}" is gone: its doc backhaul failed ${c.consecutiveFailures} consecutive times (last close ${c.lastCloseCode===null?"by provider liveness reap":`code ${c.lastCloseCode}`}${c.lastCloseReason?`, reason ${JSON.stringify(c.lastCloseReason)}`:""}) \u2014 the platform closed the session (idle timeout or a serve-side crash). The entry is evicted; retry to get a fresh session.`)))})),u&&t.push(e.onPhase(c=>{if(c.name==="live"){p=!0;return}p&&le.has(c.name)&&i(new A(`the pooled session for app "${s}" lost its compute binding (phase "${c.name}"). The entry is evicted; retry to get a fresh session. The named session itself is not over \u2014 it stays attachable by name.`))})),{gone:()=>l,onGone(c){return l?(c(l),()=>{}):(f.add(c),()=>f.delete(c))},dispose(){f.clear();for(let c of t.splice(0))c()}}}async function*C(e,s){let a=s.gone();if(a)throw a;let d=e[Symbol.asyncIterator](),u,l=new Promise((p,f)=>{u=s.onGone(t=>f(t))});l.catch(()=>{});try{for(;;){let p=await Promise.race([d.next(),l]);if(p.done)return;yield p.value}}finally{u?.(),d.return?.(void 0)}}var me=1,ge=4,he=4;function He(e,s){let a=(p,f)=>{let t=(e[p]??"").trim();if(!t)return f;let i=Number.parseInt(t,10);if(!Number.isInteger(i)||i<1)throw new Error(`${p} must be a positive integer (got ${JSON.stringify(e[p])})`);return i},d=(e.URUN_DECISION_POOL_MIN??"").trim();if(d==="0")return null;let u=d?Number.parseInt(d,10):me;if(!Number.isInteger(u)||u<0)throw new Error(`URUN_DECISION_POOL_MIN must be a non-negative integer (got ${JSON.stringify(e.URUN_DECISION_POOL_MIN)})`);let l=a("URUN_DECISION_POOL_MAX",Math.max(u,ge));return{min:u,max:l,idleGraceMs:3e4,inFlightCap:a("URUN_DECISION_POOL_INFLIGHT_CAP",he),replicaId:(e.URUN_REPLICA_ID??"").trim()||s||"proxy"}}function Ve(e){let s=(e.URUN_DECISION_STREAM_IDLE_TIMEOUT_S??"").trim();if(!s)return _;let a=Number.parseInt(s,10);if(!Number.isInteger(a)||a<=0)throw new Error(`URUN_DECISION_STREAM_IDLE_TIMEOUT_S must be a positive integer (got ${JSON.stringify(e.URUN_DECISION_STREAM_IDLE_TIMEOUT_S)})`);return a*1e3}function ee(e,s,a,d){let u=e[a];if(typeof u!="function")throw new Error(`app "${s}" has no function "${a}" (set URUN_FUNCTION)`);let l=d?u(void 0,d):u();if(typeof l.end!="function")throw new Error(`app "${s}" function "${a}" returned a session without end()`);return l}function oe(e,s,a,d){let{baseUrl:u,orgId:l,auth:p}=e,f=s.fnName??e.fnName,t=e.sessionStartOptions?.(a),i=d||t||s.sharedApp?{...s.sharedApp?{sharedApp:!0}:{},...t,...d}:void 0,c=o=>({session:o,responses:new F(o),gone:Y(o,a)});if(p.lane==="jwt"){let o=z(s.appSlug,{baseUrl:u,orgId:l,jwt:p.jwt});return c(ee(o,s.appSlug,f,i))}let{apiKey:m,gatewayUrl:g}=p,w=(typeof e.subject=="function"?e.subject():e.subject)??fe(),n=z(s.appSlug,{baseUrl:u,orgId:l,getAccessToken:async()=>(await de(m,{baseUrl:g,expiresIn:300,allowedFunctions:[`${s.appSlug}/${f}`],subject:w,...s.sharedApp?{sharedApp:!0}:{}})).token});return c(ee(n,s.appSlug,f,i))}function we(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 Se(e){if(e.kind==="never-live")return!0;if(e.kind!=="create-failed")return!1;let s=e.status.error?.httpStatus,a=e.status.error?.code;return s===410||s===409&&(a==="cancelled"||a==="expired")?!0:a==="client_ended"}var ye=new Set(["error","paused"]);function je(e){let{auth:s,catalog:a}=e,d=null,u=(p,f)=>{let t=oe(e,p,f);return e.onSessionOpen&&typeof t.session.id=="string"&&t.session.id!==""&&e.onSessionOpen(f,t.session.id),t.gone.onGone(()=>{d?.evict(f,t)}),t},l=new V({sharedOnlyListing:e.sharedOnlyListing===!0,defaultApp:e.defaultApp,fnName:e.fnName,openSession:u,closeSession:async p=>{if(p.gone.dispose(),e.release==="detach"){if(typeof p.session.detach!="function")throw new Error('pooled session exposes no native detach() \u2014 the release:"detach" lane requires @urun-sh/core Session.detach');p.session.detach();return}await p.session.end()},listApps:s.lane==="api-key"?()=>W({apiUrl:e.apiUrl,apiKey:s.apiKey}):null,listCatalog:a?()=>M(a):null,listPrices:a?()=>G(a):null,shared:s.lane==="api-key"?()=>x({apiUrl:e.apiUrl,apiKey:s.apiKey}):null,sessionKey:we});return d=l,e.defaultApp!==null&&l.seed(e.defaultApp,u({appSlug:e.defaultApp},E({appSlug:e.defaultApp}))),l}function Be(e,s){let a=e.decisionPool;if(!a)throw new Error("wireDecisionPool called without BackhaulConfig.decisionPool");let d=a.replicaId??"proxy",u=new Map,l=new Map,p=t=>{process.stderr.write(`[decision-pool] ${t}
+`)},f=(t,i)=>{let c=u.get(i);if(c)return c;let m=new B({min:a.min,max:a.max,idleGraceMs:a.idleGraceMs,inFlightCap:a.inFlightCap,replicaId:d,openSlot:async g=>{let w=oe(e,t,i,{concurrencyKey:g.concurrencyKey});return w.gone.onGone(()=>m.evict(w)),await b(w.session),p(`slot attached pool=${i} slot=${g.concurrencyKey} session=${w.session.id??"(no id)"}`),w},closeSlot:async g=>{g.gone.dispose(),await g.session.end()},onDetach:(g,w)=>p(`slot detached pool=${i} slot=${g.concurrencyKey} reason=${w}`)});return u.set(i,m),m};return{async boot(){if(e.defaultApp===null)return;let t={appSlug:e.defaultApp};await f(t,E(t)).boot()},async sessionFor(t){let i=await s.resolveTargetFor(t),c=E(i),m=f(i,c).tryAcquireReady();if(!m)throw new j(f(i,c).stats().lastMissAgeMs);let g=l.get(m.entry)??[];return g.push(m),l.set(m.entry,g),{app:c,entry:m.entry,identity:H(i)}},async warmFor(t){let i=await s.resolveTargetFor(t);await f(i,E(i)).boot()},end(t){let i=l.get(t)?.shift();i&&(l.get(t)?.length===0&&l.delete(t),i.release())},evict(t){for(let i of u.values())i.evict(t)},reap(){for(let t of u.values())t.reap()},async pingReady(t){for(let i of u.values())if(await i.pingReady(t))return!0;return!1},stats(){return[...u.entries()].map(([t,i])=>({poolKey:t,...i.stats()}))},async drain(t){await Promise.all([...u.values()].map(i=>i.drain({timeoutMs:t})))},async close(){await Promise.all([...u.values()].map(t=>t.close()))}}}function ve(e){return pe(e?.name)?e?.runtime?.wake?.expectedWakePath==="portable-golden-restore"?Z:ue:Z}async function b(e,s){let a=e.whenLive;if(typeof a!="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 a.call(e,{timeout:ve(e.phase),signal:s})}async function ne(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 O(e){return e.reason instanceof Error?e.reason:new Error("the client socket closed")}function Ke(e,s=N,a,d,u=_){let l=new WeakMap,p=new WeakMap,f=new WeakMap,t=new WeakMap,i=new WeakMap,c=n=>{let o=n.session.id;return typeof o=="string"&&o.length>0?o:null},m=n=>{let o=l.get(n);if(!o){let r={waiting:0,gone:new AbortController};i.set(n,r),o=(async()=>{if(await b(n.session,r.gone.signal),r.gone.signal.aborted)throw O(r.gone.signal);i.get(n)===r&&i.delete(n);let h=await R(n.session,await s());return Object.assign(h,{sessionId:n.session.id})})(),l.set(n,o),o.catch(()=>{l.get(n)===o&&l.delete(n)}),o.then(()=>{i.get(n)===r&&i.delete(n)},()=>{i.get(n)===r&&i.delete(n)})}return o},g=n=>{let o=t.get(n);return o||(o=(async()=>{let r=await m(n);return{appendInputAudio:h=>r.appendInputAudio(h),onOutputAudio:h=>r.onOutputAudio(h),sessionId:c(n)}})(),t.set(n,o),o.catch(()=>{t.get(n)===o&&t.delete(n)})),o},w=async(n,o,r)=>{if(r?.aborted){let S=i.get(n);throw(l.get(n)===void 0||S!==void 0&&S.waiting===0)&&await ne(n.session),O(r)}let h=o(n),y=i.get(n);return y!==void 0&&y.waiting++,r===void 0?h:new Promise((S,P)=>{let T=()=>{y!==void 0&&--y.waiting===0&&i.get(n)===y&&(y.gone.abort(),i.delete(n),l.delete(n),t.delete(n),ne(n.session).catch(()=>{})),P(O(r))};r.addEventListener("abort",T,{once:!0});let k=()=>r.removeEventListener("abort",T);h.then(I=>{k(),S(I)},I=>{k(),P(I)})})};return{createResponse:Q({sessionFor:d?n=>d.sessionFor(n):n=>e.sessionFor(n),sessionIdOf:n=>c(n),evict:(n,o)=>{d?d.evict(o):e.evict(n,o)},onAdmissionVerdict:(n,o,r)=>{Se(r)&&(d?d.evict(o):e.evict(n,o),a?.(n,r.kind==="never-live"?"never-live":"create-failed"))},createStream:async(n,o)=>{try{await b(n.session);let r=K(C(await n.responses.responses.create(o),n.gone),u);if(!d)return r;let h=n;return(async function*(){try{yield*r}catch(S){throw q(S)&&d.evict(h),S}finally{d.end(h)}})()}catch(r){throw d&&d.end(n),r}}}),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),supportsTranscription:n=>e.supportsTranscription(n),transcriptStreamFor:n=>e.transcriptStreamFor(n),listSpeechToTextModels:()=>e.speechToTextModelList(),openAudio:async(n,o)=>{let{entry:r}=await e.sessionFor(n);return await w(r,m,o)},openAudioOn:async(n,o)=>{let{entry:r}=await e.sessionForHandle(n);return await w(r,g,o)},openTurnsOn:async function*(n,o){if(o?.aborted)throw O(o);let{entry:r}=await e.sessionForHandle(n);yield*$(r.session,o)},openVideo:async n=>{let{entry:o}=await e.sessionFor(n),r=p.get(o);return r||(r=D(o.session),p.set(o,r)),r},openVideoOut:async n=>{let{entry:o}=await e.sessionFor(n),r=f.get(o);return r||(r=U(o.session),f.set(o,r)),r},sessionHandle:async n=>(await e.handleFor(n)).handle,speechSessionHandle:async n=>(await e.speechHandleFor(n)).handle,evictSpeechSession:async n=>{let{app:o,entry:r}=await e.sessionForHandle(n);await e.evict(o,r)},createResponseOn:async(n,o)=>{let{entry:r}=await e.sessionForHandle(n);return await b(r.session),C(await r.responses.responses.create(o),r.gone)},onSessionEnd:async(n,o)=>{let{entry:r}=await e.sessionForHandle(n),h=r.session;if(typeof h.onPhase!="function")throw new Error("pooled session exposes no onPhase \u2014 session-end (goAway) wiring requires @urun-sh/core Session");let y=!1;return h.onPhase(S=>{if(y||!ye.has(S.name))return;y=!0;let P=h.endsAt??null;o({timeLeftMs:P?Math.max(0,P.getTime()-Date.now()):null,reason:S.reason??`session ${S.name}`})})}}}export{He as a,Ve as b,je as c,Be as d,Ke as e};