@urun-sh/openai 0.5.6 → 0.6.1

package/dist/chunk-VLRMJRLS.js

@@ -1,6 +0,0 @@
-import{a as D,b as B}from"./chunk-HR5H6S7L.js";import{H as oe,I as C,J as se,K as ie,L as ae,M as le,N as ce,O as ue,P as de,b as Y,d as Q,e as X,g as Z,h as ee,i as te,j as re,w as ne,x}from"./chunk-NY23USZF.js";import{e as q}from"./chunk-3WAJD62J.js";import{a as V}from"./chunk-7M6LY6DR.js";import{c as p}from"./chunk-5CCJUNH7.js";p();import{Counter as R,Gauge as he,Histogram as pe,Registry as ke,collectDefaultMetrics as Oe}from"@prometheus-io/client";import{createServer as Ie}from"http";var xt=9464,ge="/metrics",S="(other)",Ne="(none)",Te=64,Pe=128,$e=["/v1/responses","/v1/chat/completions","/v1/messages","/v1/messages/count_tokens","/v1/models","/v1/images/models","/v1/images/generations","/v1/images/edits",oe,ie],Ue=new Set($e);function Le(t){return Ue.has(t)?t:S}function He(t){return t===null?"none":t>=500?"5xx":t>=400?"4xx":t>=300?"3xx":t>=200?"2xx":"1xx"}var A=class{constructor(e){this.max=e}max;admitted=new Set;overflows=0;of(e,r=!0){return e===null||e===""?Ne:this.admitted.has(e)?e:r?this.admitted.size>=this.max?(this.overflows+=1,S):(this.admitted.add(e),e):S}get size(){return this.admitted.size}get overflowCount(){return this.overflows}},k=class{registry=new ke;modelLabel=new A(Te);orgLabel=new A(Pe);requests;duration;ttft;streamErrors;undelivered;ledgerWrites;ledgerQueue;labelCardinality;labelOverflows;constructor(e={}){this.requests=new R({name:"urun_inference_requests_total",help:"Completed /v1 requests, by lane, method, HTTP status class, streaming, model and org.",labelNames:["lane","method","status_class","stream","model","org"],registers:[this.registry]}),this.duration=new pe({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 pe({name:"urun_inference_ttft_seconds",help:"Time to first output-text delta on a streamed response. The metric Hugging Face gates provider listing on (<5s for conversational models).",labelNames:["lane","model","org"],buckets:[.1,.25,.5,1,2,3,5,8,15,30,60],registers:[this.registry]}),this.streamErrors=new R({name:"urun_inference_stream_errors_total",help:"Responses that emitted an error frame AFTER the head flushed. These are HTTP 200 with a complete body, so no status-based metric can see them.",labelNames:["lane","model"],registers:[this.registry]}),this.undelivered=new R({name:"urun_inference_undelivered_total",help:"Responses whose socket did not finish flushing \u2014 the client vanished mid-answer.",labelNames:["lane","model"],registers:[this.registry]}),this.ledgerWrites=new R({name:"urun_inference_ledger_writes_total",help:"Terminal fate of the per-request ledger row for each billable /v1 request. outcome=written is the row landing (reason=priced|unpriced); outcome=skipped and outcome=failed are BOTH revenue that went unrecorded. failed/duplicate is the ENG-410 fabricated-row signal.",labelNames:["outcome","reason"],registers:[this.registry]}),this.ledgerQueue=new he({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"},ae()),this.ledgerQueue.set({measure:"oldest_age_seconds"},le()/1e3)}}),this.labelCardinality=new he({name:"urun_inference_label_cardinality",help:"Distinct values admitted for a bounded metric label. At its ceiling, further values report as (other).",labelNames:["label"],registers:[this.registry],collect:()=>{this.labelCardinality.set({label:"model"},this.modelLabel.size),this.labelCardinality.set({label:"org"},this.orgLabel.size)}}),this.labelOverflows=new R({name:"urun_inference_label_overflows_total",help:"Observations that would have taken a new label value but found the budget spent. Non-zero means a metric dimension is now partly blind.",labelNames:["label"],registers:[this.registry]}),e.defaultMetrics!==!1&&Oe({register:this.registry,prefix:"urun_inference_proxy_"})}observeLedger(e){this.ledgerWrites.inc({outcome:e.outcome,reason:e.reason})}observe(e){let r=Le(e.path),n=e.status!==null&&!(e.status>=400&&e.status<500),o=this.modelLabel.overflowCount+this.orgLabel.overflowCount,a=this.modelLabel.of(e.model,n),s=this.orgLabel.of(e.org,n);this.modelLabel.overflowCount+this.orgLabel.overflowCount-o>0&&(this.modelLabel.overflowCount>o?this.labelOverflows.inc({label:"model"}):this.labelOverflows.inc({label:"org"}));let c=e.stream===!0?"true":"false";this.requests.inc({lane:r,method:e.method===""?S:e.method,status_class:He(e.status),stream:c,model:a,org:s}),this.duration.observe({lane:r,stream:c,model:a,org:s},e.duration_ms/1e3),e.ttft_ms!==null&&this.ttft.observe({lane:r,model:a,org:s},e.ttft_ms/1e3),e.stream_error&&this.streamErrors.inc({lane:r,model:a}),e.delivered||this.undelivered.inc({lane:r,model:a})}scrape(){return this.registry.metrics()}get contentType(){return this.registry.contentType}};function fe(t){return Ie((e,r)=>{(async()=>{let n=(e.url??"/").split("?")[0];if(e.method!=="GET"||n!==ge){r.writeHead(404,{"content-type":"text/plain"}),r.end(`no route for ${e.method} ${n} \u2014 this port serves only GET ${ge}
-`);return}let o=await t.scrape();r.writeHead(200,{"content-type":t.contentType}),r.end(o)})().catch(n=>{if(!r.headersSent){r.writeHead(500,{"content-type":"text/plain"}),r.end(`${String(n instanceof Error?n.message:n)}
-`);return}r.end()})})}p();import{createHash as ye}from"crypto";import{createServer as Me}from"http";var Ot=9465,Fe="x-urun-key-id",De="x-urun-key-fingerprint",Be="urun-ratelimit-bucket:",Ge="urun-ratelimit-tier:",me="anonymous",ze=/^urun_sk_[A-Za-z0-9_-]{43}$/;function Ke(t){return ye("sha256").update(`${Be}${t}`).digest("hex")}function je(t){return ye("sha256").update(`${Ge}${t}`).digest("hex")}function We(t){let e=/^Bearer[ \t]+(\S.*)$/i.exec((t??"").trim());if(!e)return null;let r=e[1].trim();return ze.test(r)?r:null}function we(){return Me((t,e)=>{let r=We(t.headers.authorization);e.setHeader(Fe,r===null?me:Ke(r)),e.setHeader(De,r===null?me:je(r)),e.writeHead(200),e.end()})}p();import{createClientToken as Je,ClientTokenError as qe}from"@urun-sh/core";var m=class extends Error{constructor(r,n=401){super(r);this.status=n;this.name="ProxyAuthError"}status},v=class extends Error{constructor(e){super(e),this.name="ControlPlaneUnavailableError"}};function E(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 r=/^Bearer[ \t]+(\S.*)$/i.exec(e);if(!r)throw new m("malformed Authorization header \u2014 expected `Authorization: Bearer <urun org api key>`");return r[1].trim()}var Ve=60,Ye=6e4,Qe=4096;function Xe(t){return t===401||t===403}var O=class{constructor(e){this.opts=e;this.mint=e.mint??Je,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 r=this.cache.get(e);if(r&&this.now()-r.at<Ye)return{apiKey:e,orgId:r.orgId};let n;try{n=(await this.mint(e,{baseUrl:this.opts.gatewayUrl,expiresIn:Ve})).orgId}catch(o){throw o instanceof qe&&Xe(o.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 v(`could not verify the API key against the uRun control plane at ${this.opts.gatewayUrl}: ${o instanceof Error?o.message:String(o)}`)}if(!n)throw new v("the control plane verified the API key but returned no org_id \u2014 refusing to route a request with no tenancy (control-plane contract violation)");if(this.cache.size>=Qe){let o=this.cache.keys().next();o.done||this.cache.delete(o.value)}return this.cache.set(e,{orgId:n,at:this.now()}),{apiKey:e,orgId:n}}clear(){this.cache.clear()}};p();var Ze=8080,G="serve",Lt="0.0.0.0";function et(t,e,r){let n=(t[e]??"").trim();if(!n)throw new Error(`${e} is not set \u2014 ${r}`);return n}function Ht(t=process.env){let e=(t.PORT??"").trim(),r=e===""?Ze:Number(e);if(!Number.isInteger(r)||r<0||r>65535)throw new Error(`PORT must be an integer in 0..65535, got ${JSON.stringify(e)}`);let n=ne(x(et(t,"URUN_BASE_URL","the hosted proxy needs the uRun session-gateway base it opens sessions and verifies API keys against (e.g. https://api.urun.sh)"),"URUN_BASE_URL")),o=(t.URUN_API_URL??"").trim(),a=o?x(o,"URUN_API_URL"):`${n.replace(/\/+$/,"")}/v1`,s=(t.URUN_CATALOG_URL??"").trim(),i=(t.URUN_CATALOG_ANON_KEY??"").trim();if(!!s!=!!i)throw new Error("URUN_CATALOG_URL and URUN_CATALOG_ANON_KEY must be set together (or both left unset) \u2014 exactly one of them is set, which would silently disable the model-catalog oracle");let c=s?{catalogUrl:x(s,"URUN_CATALOG_URL"),anonKey:i}:null;return{port:r,baseUrl:n,apiUrl:a,catalog:c}}p();import{createHash as tt,randomUUID as rt}from"crypto";var nt=512;function P(t){return`urun-inference-proxy-${tt("sha256").update(t).digest("hex").slice(0,32)}`}var I=class{constructor(e){this.opts=e}opts;tenants=new Map;label(e,r){return Object.assign(Object.create(e),{tenant:r.orgId})}build(e){if(this.opts.build){let n=this.opts.build(e);return{router:n.router,clients:this.label(n.clients,e)}}let r=D({baseUrl:this.opts.baseUrl,orgId:e.orgId,defaultApp:null,fnName:G,apiUrl:this.opts.apiUrl,auth:{lane:"api-key",apiKey:e.apiKey,gatewayUrl:this.opts.baseUrl},catalog:this.opts.catalog,subject:P(e.apiKey),release:"detach"});return{router:r,clients:this.label(B(r),e)}}clientsFor(e){let r=this.tenants.get(e.apiKey);if(r)return this.tenants.delete(e.apiKey),this.tenants.set(e.apiKey,r),r.clients;for(;this.tenants.size>=nt;){let o=this.tenants.entries().next();if(o.done)break;let[a,s]=o.value;this.tenants.delete(a),s.router.closeAll().catch(()=>{})}let n=this.build(e);return this.tenants.set(e.apiKey,n),n.clients}get size(){return this.tenants.size}async closeAll(){let e=[...this.tenants.values()].map(n=>n.router);this.tenants.clear();let r=[];if(await Promise.all(e.map(async n=>{try{await n.closeAll()}catch(o){r.push(String(o instanceof Error?o.message:o))}})),r.length>0)throw new Error(`failed to close ${r.length} tenant backhaul(s): ${r.join("; ")}`)}},be=512,ot=1024,N=class extends Error{constructor(){super(`conversation capacity reached (${be} conversations on this replica) \u2014 every slot holds an active or resumable conversation; retry later or scale out`),this.name="ConversationCapacityError"}};function st(t,e,r,n){let o=D({baseUrl:t.baseUrl,orgId:e.orgId,defaultApp:null,fnName:G,apiUrl:t.apiUrl,auth:{lane:"api-key",apiKey:e.apiKey,gatewayUrl:t.baseUrl},catalog:t.catalog,subject:r.subject,release:"detach",sessionStartOptions:a=>{let s=n.get(a);return s===void 0?void 0:{sessionId:s}},onSessionOpen:(a,s)=>{n.set(a,s)}});return{clients:B(o),release:async()=>{await o.closeAll()}}}var T=class{constructor(e){this.opts=e}opts;byCaller=new Map;byHandle=new Map;build(e,r){let n={id:r,subject:`${P(e.apiKey)}:live:${r}`},o=new Map,a=this.opts.build?this.opts.build(e,n):st(this.opts,e,n,o);return{id:r,backhaul:a,attachments:0,nativeRefs:o}}attachEntry(e,r){let n=this.byCaller.get(e.apiKey);n===void 0&&(n=new Map,this.byCaller.set(e.apiKey,n));let o=n.get(r);if(o!==void 0)return n.delete(r),n.set(r,o),o.attachments+=1,o;this.evictOverflow();let a=this.build(e,r);return a.attachments=1,n.set(r,a),a}get conversationCount(){let e=0;for(let r of this.byCaller.values())e+=r.size;return e}evictOverflow(){for(;this.conversationCount>=be;){let e=this.byCaller.keys().next();if(e.done)throw new N;let r=e.value,n=this.byCaller.get(r),o=!1;for(let[a,s]of n)if(!(s.attachments>0)){n.delete(a),this.dropHandleIndexFor(r,a),s.backhaul.release().catch(()=>{}),o=!0;break}if(!o)throw n.size===0&&this.byCaller.delete(r),new N;n.size===0&&this.byCaller.delete(r)}}dropHandleIndexFor(e,r){for(let[n,o]of this.byHandle)o.apiKey===e&&o.conversationId===r&&this.byHandle.delete(n)}connectionFor(e){let r=null,n=!1,o=()=>(r===null&&(r=this.attachEntry(e,rt())),r),a={createResponse:(s,i)=>o().backhaul.clients.createResponse(s,i),catalogIdentity:s=>o().backhaul.clients.catalogIdentity(s),servingAppSlug:s=>{let i=o().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:()=>o().backhaul.clients.listModels(),sessionHandle:s=>o().backhaul.clients.sessionHandle(s),createResponseOn:(s,i)=>o().backhaul.clients.createResponseOn(s,i),onSessionEnd:(s,i)=>o().backhaul.clients.onSessionEnd(s,i)};return a.openAudio=async s=>{let i=o().backhaul.clients.openAudio;if(i===void 0)throw new Error("the conversation backhaul wires no openAudio seam \u2014 voice is unavailable on this deployment");return i(s)},a.openVideo=async s=>{let i=o().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=o().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=o().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)},{clients:a,bindResumeHandle:s=>{let i=r;if(i===null)throw new Error("bindResumeHandle before the conversation opened \u2014 no conversation to bind to");for(;this.byHandle.size>=ot;){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 q("the conversation behind this resume handle is gone (evicted at capacity, or the replica restarted) or the handle is not yours \u2014 reconnect without a handle to start a new conversation");return r=u,u.attachments+=1,u.backhaul.clients},release:async()=>{if(n)return;n=!0;let s=r;r=null,s!==null&&(s.attachments-=1,!(s.attachments>0)&&await s.backhaul.release().catch(()=>{}))}}}async detachAll(){let e=[];for(let n of this.byCaller.values())for(let o of n.values())e.push(o.backhaul.release());this.byCaller.clear(),this.byHandle.clear();let r=[];if(await Promise.all(e.map(async n=>{try{await n}catch(o){r.push(String(o instanceof Error?o.message:o))}})),r.length>0)throw new Error(`failed to detach ${r.length} conversation backhaul(s): ${r.join("; ")}`)}};p();var it="/inference-ledger",at=1e4,$=class extends Error{ledgerFailureReason;constructor(e,r="duplicate"){super(e),this.name="LedgerDuplicateError",this.ledgerFailureReason=r}},K=class extends Error{constructor(r,n){super(n);this.ledgerAlreadyWritten=r;this.name="LedgerAlreadyWrittenError"}ledgerAlreadyWritten},g=class extends Error{ledgerFailureReason="surface";constructor(e){super(e),this.name="LedgerSurfaceError"}},j=class extends m{ledgerFailureReason="auth";constructor(e){super(e),this.name="LedgerAuthError"}},lt=[["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",ve],["cost_nano_usd",ve],["price_version",t=>t===null||typeof t=="string"]];function ct(t){for(let[e,r]of lt){if(!(e in t))return`${e} (absent)`;if(!r(t[e]))return`${e} (${JSON.stringify(t[e])})`}return null}function ut(t){let e;try{e=JSON.parse(t)}catch{return{status:"unavailable"}}if(e===null||typeof e!="object"||Array.isArray(e))return{status:"unavailable"};let r=e.existing_status;if(r==="not_in_this_org")return{status:"not_in_this_org"};if(r!=="found")return{status:"unavailable"};let n=e.existing;return n===null||typeof n!="object"||Array.isArray(n)?{status:"unavailable"}:ct(n)!==null?{status:"unavailable"}:{status:"found",row:n}}function ve(t){return t===null?!0:typeof t=="number"?Number.isFinite(t):typeof t!="string"?!1:/[^\s]/.test(t)&&Number.isFinite(Number(t))}function z(t,e){return t===null?e===null:e===null?!1:Number(e)===t}function _e(t,e){if(typeof e!="string")return!1;let r=Date.parse(t),n=Date.parse(e);return Number.isFinite(r)&&Number.isFinite(n)&&r===n}function dt(t,e){let r=e;return r.inference_id===t.inference_id&&r.model_id===t.model_id&&(r.variant??null)===t.variant&&r.task===t.task&&r.http_status===t.http_status&&_e(t.started_at,r.started_at)&&_e(t.completed_at,r.completed_at)&&z(t.prompt_tokens,r.prompt_tokens)&&z(t.completion_tokens,r.completion_tokens)&&z(t.output_units,r.output_units)}function ht(t,e){if(t===null||typeof t!="object"||Array.isArray(t))throw new g(`${e} answered a "recorded" that is not an object: ${JSON.stringify(t)}`);let r=t,n=r.inference_id;if(typeof n!="string"||n==="")throw new g(`${e} answered a row with no inference_id: ${JSON.stringify(t)}`);if(!("cost_nano_usd"in r))throw new g(`${e} answered a row with no cost_nano_usd field: ${JSON.stringify(t)}`);let o=r.cost_nano_usd;if(o!=null&&!(typeof o=="number"&&Number.isInteger(o)&&o>=0))throw new g(`${e} answered cost_nano_usd ${JSON.stringify(o)}, which is neither a non-negative integer nor null`);let a=u=>{if(!(u in r))throw new g(`${e} answered a row with no ${u} field: ${JSON.stringify(t)}`);let f=r[u];if(f===null)return null;if(typeof f!="string")throw new g(`${e} answered a non-string ${u}: ${JSON.stringify(f)}`);if(!/[^\s]/.test(f))throw new g(`${e} answered a BLANK ${u} (${JSON.stringify(f)}) \u2014 it satisfies "is a string" while naming nothing, which is exactly what inference_requests_priced_attribution_check exists to forbid`);return f},s=a("price_version"),i=a("unpriced_reason"),c=o!=null;if(c&&(s===null||i!==null))throw new g(`${e} answered a PRICED row (cost_nano_usd ${String(o)}) 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 ${JSON.stringify(i)}. NULL means NOT YET PRICED, and the reason is how an operator learns that traffic is being served which nobody can bill.`);return{inference_id:n,cost_nano_usd:o===void 0?null:o,price_version:s,unpriced_reason:i}}var U=class{constructor(e){this.opts=e;this.fetchImpl=e.fetchImpl??fetch}opts;fetchImpl;async write(e,r){let n=`${this.opts.apiUrl.replace(/\/+$/,"")}${it}`,o;try{o=await this.fetchImpl(n,{method:"POST",headers:{"content-type":"application/json",authorization:`Bearer ${e}`},body:JSON.stringify(r),signal:AbortSignal.timeout(at)})}catch(c){throw new g(`the uRun ledger surface at ${n} could not be reached: ${c instanceof Error?c.message:String(c)}`)}if(o.status===409){let c=await o.text().catch(()=>""),u=ut(c);throw u.status==="found"&&dt(r,u.row)?new K(u.row,`the uRun control plane already holds THIS PROXY'S OWN row for inference_id ${r.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 $(`inference_id ${r.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 $(`the uRun control plane already holds a ledger row for inference_id ${r.inference_id} that is NOT the one being written${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(o.status===401||o.status===403)throw await o.text().catch(()=>{}),new j("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(!o.ok){let c=await o.text().catch(u=>`<body unreadable: ${u instanceof Error?u.message:String(u)}>`);throw new g(`the uRun ledger surface at ${n} answered HTTP ${o.status}: ${c.slice(0,400)}`)}let a;try{a=await o.json()}catch(c){throw new g(`the uRun ledger surface at ${n} 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 ${n} answered a body with no "recorded" field`);let i=ht(s,`the uRun ledger surface at ${n}`);if(i.inference_id!==r.inference_id)throw new g(`the uRun ledger surface at ${n} answered for inference_id ${i.inference_id} when asked about ${r.inference_id}`);return i}};p();import{createServer as gt}from"http";p();var L="/inference-usage",Re=25e3;var pt=new Set(["Error","TypeError","SyntaxError","TimeoutError"]);function Ee(t){if(!(t instanceof Error))return`a ${typeof t}`;let e=pt.has(t.name)?t.name:"Error",r=t.cause?.code;return typeof r=="string"&&/^[A-Z][A-Z0-9_]{0,31}$/.test(r)?`${e} (${r})`:e}var H=class{constructor(e){this.opts=e;this.fetchImpl=e.fetchImpl??fetch}opts;fetchImpl;async lookup(e,r){let n=`${this.opts.apiUrl.replace(/\/+$/,"")}${L}`,o;try{o=await this.fetchImpl(n,{method:"POST",headers:{"content-type":"application/json",authorization:`Bearer ${e}`},body:JSON.stringify({inference_ids:r}),signal:AbortSignal.timeout(Re)})}catch(s){throw new C(`the uRun usage surface (POST ${L}) could not be reached within ${Re}ms: ${Ee(s)}`)}if(o.status===401||o.status===403)throw await o.text().catch(()=>{}),new m("the uRun org API key was rejected by the uRun control plane \u2014 check the key and that it is not revoked");if(!o.ok){let s=await o.text().catch(()=>null),i=s===null?"body unreadable":`${s.length}-character body`;throw new C(`the uRun usage surface (POST ${L}) answered HTTP ${o.status} (${i})`)}let a;try{a=await o.json()}catch(s){throw new C(`the uRun usage surface (POST ${L}) answered HTTP ${o.status} with a body that is not JSON: ${Ee(s)}`)}return se(a,this.opts.onRejectedRow)}};var ft=5e3,mt=3e3,yt="/api/client-tokens";function W(t,e,r){t.writeHead(e,{"content-type":"application/json"}),t.end(JSON.stringify(r))}function xe(t,e,r,n){W(t,e,{error:{message:r,type:n}})}async function wt(t,e=fetch){let r=`${t.replace(/\/+$/,"")}${yt}`;try{let n=await e(r,{method:"POST",headers:{"content-type":"application/json"},body:"{}",signal:AbortSignal.timeout(mt)});return n.status===404?{ready:!1,detail:`${r} answered 404 \u2014 this is not a uRun control plane`}:n.status>=500?{ready:!1,detail:`${r} answered HTTP ${n.status}`}:{ready:!0,detail:`${r} answered HTTP ${n.status}`}}catch(n){return{ready:!1,detail:`${r} unreachable: ${n instanceof Error?n.message:String(n)}`}}}function pr(t){let e=t.resolver??new O({gatewayUrl:t.baseUrl}),r=t.registry??new I({baseUrl:t.baseUrl,apiUrl:t.apiUrl,catalog:t.catalog}),n=t.probeControlPlane??(()=>wt(t.baseUrl)),o=t.usageClient??new H({apiUrl:t.apiUrl}),a=t.ledgerClient??new U({apiUrl:t.apiUrl}),s=new k(t.defaultMetrics===void 0?{}:{defaultMetrics:t.defaultMetrics}),i=async l=>e.resolve(E(l.headers.authorization)),c=l=>l instanceof m?{status:l.status,openaiType:"authentication_error",anthropicType:"authentication_error"}:l instanceof v?{status:502,openaiType:"upstream_error",anthropicType:"api_error"}:null,u=de({clients:async l=>r.clientsFor(await i(l)),usage:{authorize:l=>E(l.headers.authorization),lookup:(l,d)=>o.lookup(l,d)},statusOf:c,...t.requestLog?{requestLog:t.requestLog}:{},observe:l=>s.observe(l),ledger:{write:(l,d)=>a.write(E(l.headers.authorization),d),observe:l=>s.observeLedger(l),...t.ledgerLog?{report:t.ledgerLog}:{}}}),f=null,_=gt((l,d)=>{let h=new URL(l.url??"/","http://localhost");(async()=>{if(l.method==="GET"&&h.pathname==="/healthz"){W(d,200,{ok:!0});return}if(l.method==="GET"&&h.pathname==="/readyz"){let J=Date.now();if(!f||J-f.at>=ft){let Ae=await n();f={at:J,...Ae}}W(d,f.ready?200:503,{ok:f.ready,control_plane:f.detail});return}let y=h.pathname==="/v1"||h.pathname.startsWith("/v1/"),F=h.pathname.startsWith("/partners/");if(!y&&!F){xe(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}u(l,d)})().catch(y=>{if(!d.headersSent){xe(d,500,String(y instanceof Error?y.message:y),"proxy_error");return}d.end()})}),M=t.conversations??new T({baseUrl:t.baseUrl,apiUrl:t.apiUrl,catalog:t.catalog}),Ce=[Y(_,{authenticate:async l=>{let d=await e.resolve(l),h=M.connectionFor(d);return{clients:h.clients,scope:P(d.apiKey),onBindResumeHandle:h.bindResumeHandle,clientsForResume:h.clientsForResume,release:h.release}}}),X(_,{clients:async l=>r.clientsFor(await i(l)),statusOf:c}).close,ee(_,{resolver:{resolve:async l=>{let d=await e.resolve(E(l.headers.authorization)),h=M.connectionFor(d),y;try{y=await re(h.clients,te(l))}catch(F){throw h.release().catch(()=>{}),F}return{...y,release:h.release}}}})];_.on("upgrade",(l,d)=>{let h=new URL(l.url??"/","http://localhost");h.pathname===V||h.pathname===Q||h.pathname===Z||d.destroyed||(d.write(`HTTP/1.1 404 Not Found\r
-Connection: close\r
-\r
-no WS route for ${h.pathname} \u2014 this endpoint serves only the Gemini Live, /v1/responses, and /v1/realtime upgrade paths`),d.destroy())});let Se=async()=>{for(let l of Ce)await l();await M.detachAll()};return{server:_,metricsServer:fe(s),metrics:s,edgeIdentityServer:we(),closeAll:()=>r.closeAll(),closeWebSockets:Se}}async function gr(t,e,r){r&&await r();let n=new Promise(o=>t.close(()=>o()));t.closeAllConnections(),await n,await ce()||ue(),await e()}export{xt as a,ge as b,S as c,Ne as d,Te as e,Pe as f,Le as g,He as h,A as i,k as j,fe as k,Ot as l,Fe as m,De as n,Be as o,Ge as p,me as q,Ke as r,je as s,We as t,we as u,m as v,v as w,E as x,Ye as y,Qe as z,O as A,Ze as B,G as C,Lt as D,Ht as E,nt as F,P as G,I as H,be as I,ot as J,N as K,T as L,it as M,at as N,$ as O,g as P,ht as Q,U as R,ft as S,mt as T,pr as U,gr as V};