@urun-sh/openai 0.6.22 → 0.6.25

package/dist/{chunk-HCAKF2SQ.js → chunk-VY333OKA.js}

@@ -1,7 +1,7 @@
-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-PPPJ42Z7.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-SOAZAPCO.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 f}from"./chunk-YSFSRI3D.js";f();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} ${h.pathname} \u2014 this endpoint serves the OpenAI-compatible /v1 surface, the partner adapter surface (/partners/...), the WS upgrade paths (/ws/google.ai.generativelanguage.v1beta.GenerativeService.BidiGenerateContent and /v1/realtime), plus /healthz and /readyz`,"invalid_request_error");return}He(l,d)})().catch(y=>{if(!d.headersSent){Le(d,500,String(y instanceof Error?y.message:y),"proxy_error");return}d.end()})}),G=t.conversations??new L({baseUrl:t.baseUrl,apiUrl:t.apiUrl,catalog:t.catalog}),Me=[oe(R,{authenticate:async l=>{let d=await e.resolve(l),h=G.connectionFor(d);return{clients:h.clients,scope:J(d.apiKey),onBindResumeHandle:h.bindResumeHandle,clientsForResume:h.clientsForResume,release:h.release}}}),ie(R,{clients:async l=>n.clientsFor(await m(l)),statusOf:Z}).close,ce(R,{resolver:{resolve:async(l,d)=>{let h=await e.resolve(k(l.headers.authorization)),y=G.connectionFor(h),C;try{C=await he(y.clients,ue(l),d,de(l))}catch(w){throw y.release().catch(()=>{}),w}return{...C,release:y.release}}}})];R.on("upgrade",(l,d)=>{let h=new URL(l.url??"/","http://localhost");h.pathname===re||h.pathname===se||h.pathname===le||d.destroyed||(d.write(`HTTP/1.1 404 Not Found\r
+`);return}n.end()})})}f();import{createHash as Ae}from"crypto";import{createServer as Qe}from"http";var Jt=9465,Xe="x-urun-key-id",Ze="x-urun-key-fingerprint",et="urun-ratelimit-bucket:",tt="urun-ratelimit-tier:",Te="anonymous",nt=/^urun_sk_[A-Za-z0-9_-]{43}$/;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.*)$/i.exec((t??"").trim());if(!e)return null;let n=e[1].trim();return nt.test(n)?n:null}function Ie(){return Qe((t,e)=>{let n=st(t.headers.authorization);e.setHeader(Xe,n===null?Te:rt(n)),e.setHeader(Ze,n===null?Te:ot(n)),e.writeHead(200),e.end()})}f();import{createClientToken as it,ClientTokenError as at}from"@urun-sh/core";var m=class extends Error{constructor(n,o=401){super(n);this.status=o;this.name="ProxyAuthError"}status},S=class extends Error{constructor(e){super(e),this.name="ControlPlaneUnavailableError"}};function k(t){let e=(t??"").trim();if(e==="")throw new m("missing Authorization header \u2014 send your uRun org API key as `Authorization: Bearer <key>`");let n=/^Bearer[ \t]+(\S.*)$/i.exec(e);if(!n)throw new m("malformed Authorization header \u2014 expected `Authorization: Bearer <urun org api key>`");return n[1].trim()}var 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 m("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 m("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}: ${r instanceof Error?r.message:String(r)}`)}if(!o)throw new S("the control plane verified the API key but returned no org_id \u2014 refusing to route a request with no tenancy (control-plane contract violation)");if(this.cache.size>=ut){let r=this.cache.keys().next();r.done||this.cache.delete(r.value)}return this.cache.set(e,{orgId:o,at:this.now()}),{apiKey:e,orgId:o}}clear(){this.cache.clear()}};f();var ht=8080,j="serve",en="0.0.0.0";function pt(t,e,n){let o=(t[e]??"").trim();if(!o)throw new Error(`${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(/\/+$/,"")}/v1`,s=(t.URUN_CATALOG_URL??"").trim(),i=(t.URUN_CATALOG_ANON_KEY??"").trim();if(!!s!=!!i)throw new Error("URUN_CATALOG_URL and URUN_CATALOG_ANON_KEY must be set together (or both left unset) \u2014 exactly one of them is set, which would silently disable the model-catalog oracle");let c=s?{catalogUrl:O(s,"URUN_CATALOG_URL"),anonKey:i}:null;return{port:n,baseUrl:o,apiUrl:a,catalog:c}}f();import{createHash as ft,randomUUID as gt}from"crypto";var mt=512;function J(t){return`urun-inference-proxy-${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): ${n.join("; ")}`)}};f();var vt="/inference-ledger",_t=1e4,H=class extends Error{ledgerFailureReason;constructor(e,n="duplicate"){super(e),this.name="LedgerDuplicateError",this.ledgerFailureReason=n}},Y=class extends Error{constructor(n,o){super(o);this.ledgerAlreadyWritten=n;this.name="LedgerAlreadyWrittenError"}ledgerAlreadyWritten},g=class extends Error{ledgerFailureReason="surface";constructor(e){super(e),this.name="LedgerSurfaceError"}},Q=class extends m{ledgerFailureReason="auth";constructor(e){super(e),this.name="LedgerAuthError"}},Et=[["inference_id",t=>typeof t=="string"&&/[^\s]/.test(t)],["model_id",t=>typeof t=="string"&&/[^\s]/.test(t)],["variant",t=>t===null||typeof t=="string"],["task",t=>typeof t=="string"&&/[^\s]/.test(t)],["http_status",t=>typeof t=="number"&&Number.isInteger(t)],["started_at",t=>typeof t=="string"&&!Number.isNaN(Date.parse(t))],["completed_at",t=>typeof t=="string"&&!Number.isNaN(Date.parse(t))],["prompt_tokens",t=>t===null||typeof t=="number"&&Number.isInteger(t)],["completion_tokens",t=>t===null||typeof t=="number"&&Number.isInteger(t)],["output_units",V],["output_minutes",V],["cost_nano_usd",V],["price_version",t=>t===null||typeof t=="string"]];function St(t){for(let[e,n]of Et){if(!(e in t))return`${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 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: ${JSON.stringify(t)}`);let y=n[u];if(y===null)return null;if(typeof y!="string")throw new g(`${e} answered a non-string ${u}: ${JSON.stringify(y)}`);if(!/[^\s]/.test(y))throw new g(`${e} answered a BLANK ${u} (${JSON.stringify(y)}) \u2014 it satisfies "is a string" while naming nothing, which is exactly what inference_requests_priced_attribution_check exists to forbid`);return y},s=a("price_version"),i=a("unpriced_reason"),c=r!=null;if(c&&(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 ${JSON.stringify(i)}`);if(!c&&(s!==null||i===null))throw new g(`${e} answered an UNPRICED row that does not say why \u2014 price_version ${JSON.stringify(s)}, unpriced_reason ${JSON.stringify(i)}. NULL means NOT YET PRICED, and the reason is how an operator learns that traffic is being served which nobody can bill.`);return{inference_id:o,cost_nano_usd:r===void 0?null:r,price_version:s,unpriced_reason:i}}var M=class{constructor(e){this.opts=e;this.fetchImpl=e.fetchImpl??fetch}opts;fetchImpl;async write(e,n){let o=`${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 g(`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 g(`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 g(`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 g(`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 g(`the uRun ledger surface at ${o} answered for inference_id ${i.inference_id} when asked about ${n.inference_id}`);return i}};f();import{createServer as Tt}from"http";f();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 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 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 p of n.decisionPools()){l=p;break}if(!l)return null;let d=l.stats();if(d.some(p=>p.draining))return{ok:!1,detail:"decision pool draining"};if(d.every(p=>p.attached===0&&p.inFlight===0))return{ok:!0,detail:"no active decision slot; gateway accepts on-demand resume"};let h=await l.pingReady(Pt);return{ok:h,detail:h?"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): ${String(l instanceof Error?l.message:l)}
+`)}})();let u=new P(t.defaultMetrics===void 0?{}:{defaultMetrics:t.defaultMetrics}),y=async l=>e.resolve(k(l.headers.authorization)),Z=l=>l instanceof m?{status:l.status,openaiType:"authentication_error",anthropicType:"authentication_error"}:l instanceof S?{status:502,openaiType:"upstream_error",anthropicType:"api_error"}:null,He=Ee({clients:async l=>n.clientsFor(await y(l)),usage:{authorize:l=>k(l.headers.authorization),lookup:(l,d)=>a.lookup(l,d)},statusOf:Z,...t.requestLog?{requestLog:t.requestLog}:{},observe:l=>u.observe(l),ledger:{write:(l,d)=>s.write(k(l.headers.authorization),d),observe:l=>u.observeLedger(l),...t.ledgerLog?{report:t.ledgerLog}:{}}}),_=null,E=null,R=Tt((l,d)=>{let h=new URL(l.url??"/","http://localhost");(async()=>{if(l.method==="GET"&&h.pathname==="/healthz"){X(d,200,{ok:!0});return}if(l.method==="GET"&&h.pathname==="/readyz"){let w=Date.now();if((!_||w-_.at>=At)&&(_={at:w,...await o()}),!E||w-E.at>=It){let B=await r();E=B===null?{at:w,ok:!0,detail:null}:{at:w,ok:B.ok,detail:B.detail}}let ee=(_?.ready??!1)&&E.ok,Fe=E.detail===null?_?.detail??"control plane not yet probed":`${_?.detail??"control plane not yet probed"}; ${E.detail}`;X(d,ee?200:503,{ok:ee,control_plane:Fe});return}let p=h.pathname==="/v1"||h.pathname.startsWith("/v1/"),C=h.pathname.startsWith("/partners/");if(!p&&!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(p=>{if(!d.headersSent){Le(d,500,String(p instanceof Error?p.message:p),"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 y(l)),statusOf:Z}).close,ce(R,{resolver:{resolve:async(l,d)=>{let h=await e.resolve(k(l.headers.authorization)),p=G.connectionFor(h),C;try{C=await he(p.clients,ue(l),d,de(l))}catch(w){throw p.release().catch(()=>{}),w}return{...C,release:p.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
 Connection: close\r
 \r
-no WS route for ${h.pathname} \u2014 this endpoint serves only the Gemini Live, /v1/responses, and /v1/realtime upgrade paths`),d.destroy())});let De=async()=>{for(let l of Me)await l();await G.detachAll()};return{server:R,metricsServer:Oe(u),metrics:u,edgeIdentityServer:Ie(),registry:n,closeAll:()=>n.closeAll(),closeWebSockets:De}}async function Nn(t,e,n){n&&await n();let o=new Promise(r=>t.close(()=>r()));t.closeAllConnections(),await o,await be()||ve(),await e()}export{Bt as a,Ce as b,A as c,ze as d,We as e,je as f,qe as g,Ye as h,I as i,P as j,Oe as k,Jt as l,Xe as m,Ze as n,et as o,tt as p,Te as q,rt as r,ot as s,st as t,Ie as u,g as v,S as w,k as x,ct as y,ut as z,N as A,ht as B,j as C,en as D,tn as E,mt as F,J as G,U as H,Pe as I,yt as J,$ as K,L,vt as M,_t as N,H as O,f as P,Ct as Q,M as R,At as S,Nt as T,Pn as U,Nn as V};
+no WS route for ${h.pathname} \u2014 this endpoint serves only the Gemini Live, /v1/responses, and /v1/realtime upgrade paths`),d.destroy())});let De=async()=>{for(let l of Me)await l();await G.detachAll()};return{server:R,metricsServer:Oe(u),metrics:u,edgeIdentityServer:Ie(),registry:n,closeAll:()=>n.closeAll(),closeWebSockets:De}}async function Nn(t,e,n){n&&await n();let o=new Promise(r=>t.close(()=>r()));t.closeAllConnections(),await o,await be()||ve(),await e()}export{Bt as a,Ce as b,A as c,ze as d,We as e,je as f,qe as g,Ye as h,I as i,P as j,Oe as k,Jt as l,Xe as m,Ze as n,et as o,tt as p,Te as q,rt as r,ot as s,st as t,Ie as u,m as v,S as w,k as x,ct as y,ut as z,N as A,ht as B,j as C,en as D,tn as E,mt as F,J as G,U as H,Pe as I,yt as J,$ as K,L,vt as M,_t as N,H as O,g as P,Ct as Q,M as R,At as S,Nt as T,Pn as U,Nn as V};