@urun-sh/openai 0.2.60 → 0.3.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/chunk-35LB5OHI.js +1 -0
- package/dist/chunk-3ED25N7E.js +1 -0
- package/dist/chunk-EFBD3CGZ.js +1 -0
- package/dist/chunk-F2TEK34X.js +1 -0
- package/dist/chunk-VG3X2LVG.js +41 -0
- package/dist/chunk-XHIIEA6Z.js +1 -0
- package/dist/chunk-XHRPJIEK.js +6 -0
- package/dist/chunk-YSFSRI3D.js +1 -0
- package/dist/hosted/bin.cjs +56 -0
- package/dist/hosted/bin.js +23 -0
- package/dist/hosted/index.cjs +34 -0
- package/dist/hosted/index.d.cts +620 -0
- package/dist/hosted/index.d.ts +225 -0
- package/dist/hosted/index.js +1 -0
- package/dist/index.js +1 -1
- package/dist/pi-extension/index.cjs +4 -4
- package/dist/pi-extension/index.js +7 -7
- package/dist/pi-extension/standalone.cjs +53 -53
- package/dist/proxy/cli.cjs +34 -34
- package/dist/proxy/cli.js +14 -19
- package/dist/proxy/index.cjs +21 -21
- package/dist/proxy/index.d.cts +4 -154
- package/dist/proxy/index.d.ts +4 -64
- package/dist/proxy/index.js +1 -1
- package/dist/server-BBKNv7hV.d.ts +78 -0
- package/dist/server-DiQUIDDg.d.cts +182 -0
- package/package.json +12 -2
- package/dist/chunk-4FUBGLFN.js +0 -41
- package/dist/chunk-OLE2YJO3.js +0 -1
- package/dist/chunk-SSZL77P5.js +0 -1
- package/dist/chunk-VFHQZ4OM.js +0 -1
- package/dist/chunk-VXTNG2TP.js +0 -1
|
@@ -0,0 +1,34 @@
|
|
|
1
|
+
"use strict";var vt=Object.create;var pe=Object.defineProperty;var xt=Object.getOwnPropertyDescriptor;var St=Object.getOwnPropertyNames;var bt=Object.getPrototypeOf,At=Object.prototype.hasOwnProperty;var Rt=(t,n)=>{for(var e in n)pe(t,e,{get:n[e],enumerable:!0})},Ke=(t,n,e,o)=>{if(n&&typeof n=="object"||typeof n=="function")for(let r of St(n))!At.call(t,r)&&r!==e&&pe(t,r,{get:()=>n[r],enumerable:!(o=xt(n,r))||o.enumerable});return t};var de=(t,n,e)=>(e=t!=null?vt(bt(t)):{},Ke(n||!t||!t.__esModule?pe(e,"default",{value:t,enumerable:!0}):e,t)),Et=t=>Ke(pe({},"__esModule",{value:!0}),t);var Rn={};Rt(Rn,{BIND_HOST:()=>gt,ControlPlaneUnavailableError:()=>V,DEFAULT_PORT:()=>Be,MAX_CACHED_KEYS:()=>Le,MAX_TENANTS:()=>He,ORG_BINDING_TTL_MS:()=>$e,OrgResolver:()=>ee,ProxyAuthError:()=>F,READINESS_TIMEOUT_MS:()=>qe,READINESS_TTL_MS:()=>Ge,SERVE_FUNCTION:()=>Se,TenantRegistry:()=>te,bearerFrom:()=>_e,createHostedProxy:()=>kt,resolveHostedConfig:()=>ht,tenantSubject:()=>je});module.exports=Et(Rn);var _t=require("http");var Ht=require("http");var Dt=require("crypto"),Bt=require("ws");function Ae(t){return t.toLowerCase().replace(/[^\p{L}\p{N}-]/gu,"-").replace(/^-+|-+$/g,"")}var U=class extends Error{},B=class extends Error{},Re="urs1.";function It(t,n){return Re+Buffer.from(JSON.stringify([t,n]),"utf8").toString("base64url")}function Ot(t){if(!t.startsWith(Re))return null;try{let n=JSON.parse(Buffer.from(t.slice(Re.length),"base64url").toString("utf8"));if(Array.isArray(n)&&typeof n[0]=="string"&&typeof n[1]=="string")return{app:n[0],key:n[1]}}catch{}return null}async function Ve(t){let n=t.fetchImpl??fetch,e=`${t.apiUrl.replace(/\/+$/,"")}/apps`,o=await n(e,{headers:{Authorization:`Bearer ${t.apiKey}`,Accept:"application/json"}});if(!o.ok)throw new Error(`org apps listing failed: GET ${e} \u2192 ${o.status}`);let r=await o.json();if(!Array.isArray(r.apps))throw new Error(`org apps listing returned no "apps" array (GET ${e})`);if(r.truncated===!0)throw new Error(`org apps listing was truncated (GET ${e} returned ${r.apps.length} of more) \u2014 model discovery would silently omit deployed apps`);return r.apps}var Tt=6e4,fe=class{constructor(n){this.opts=n}opts;pool=new Map;appsCache=null;seed(n,e){this.pool.set(n,Promise.resolve(e))}async deployedApps(){if(!this.opts.listApps)return[];let n=this.opts.appsTtlMs??Tt;if(this.appsCache&&Date.now()-this.appsCache.at<n)return this.appsCache.apps;let e=await this.opts.listApps();return this.appsCache={at:Date.now(),apps:e},e}servable(n){return n.filter(e=>e.function_name===this.opts.fnName&&e.deployment_status==="active").map(e=>e.app_slug)}async availableIds(){let n=this.opts.defaultApp,e=new Set(n===null?[]:[n]);if(this.opts.listApps)for(let r of this.servable(await this.deployedApps()))e.add(r);for(let r of this.pool.keys())e.add(r);let o=[...e].filter(r=>r!==n).sort();return n===null?o:[n,...o]}async noDefaultApp(n){let e=await this.availableIds(),o=e.length>0?e.join(", "):"(none deployed)";throw new U(n===""?`no "model" was given and this endpoint has no default model \u2014 name one of your org's deployed "${this.opts.fnName}" apps; available models: ${o}`:`model "${n}" does not name a deployed "${this.opts.fnName}" app in your org, and this endpoint has no default model to fall back to; available models: ${o}`)}async resolveApp(n){let e=(n??"").trim(),o=this.opts.defaultApp;if(!e||e==="urun")return o===null?this.noDefaultApp(e):o;let r=Ae(e);if(r===o)return o;if(this.pool.has(r))return r;if(this.opts.listApps){let i=await this.deployedApps(),s=this.servable(i);if(s.includes(r))return r;let a=s.filter(u=>u.startsWith(`${r}-`));if(a.length===1)return a[0];if(a.length>1)throw new U(`model "${e}" is ambiguous across deployed apps ${a.join(", ")} \u2014 name the exact app slug (or catalog id:variant)`);let l=i.find(u=>u.app_slug===r||u.app_slug.startsWith(`${r}-`));if(l)throw new U(`model "${e}" maps to app "${l.app_slug}", which is not an active "${this.opts.fnName}" app \u2014 deploy it with \`urun serve ${e}\`; available models: ${(await this.availableIds()).join(", ")}`)}if(this.opts.listCatalog){let s=(await this.opts.listCatalog()).find(a=>e===a.model_id||e===`${a.model_id}:${a.variant}`||r===Ae(a.model_id)||r===Ae(`${a.model_id}-${a.variant}`));if(s)throw new U(`model "${e}" is in the uRun catalog but not deployed \u2014 deploy it with \`urun serve ${s.model_id}\`; available models: ${(await this.availableIds()).join(", ")}`)}return o===null?this.noDefaultApp(e):o}async sessionFor(n){let e=await this.resolveApp(n),o=this.pool.get(e);return o||(o=Promise.resolve(this.opts.openSession(e)),this.pool.set(e,o),o.catch(()=>this.pool.delete(e))),{app:e,entry:await o}}keyOf(n){let e=this.opts.sessionKey;if(!e)throw new Error("ModelRouter: sessionKey is not configured \u2014 the session-identity seam (handleFor/sessionForHandle) is unavailable on this router");let o=e(n);if(typeof o!="string"||o.length===0)throw new Error("ModelRouter: sessionKey returned an empty identity for a pooled session");return o}async handleFor(n){let{app:e,entry:o}=await this.sessionFor(n);return{app:e,handle:It(e,this.keyOf(o))}}async sessionForHandle(n){let e=Ot(n);if(!e)throw new B(`malformed session handle ${JSON.stringify(n)} \u2014 not issued by this proxy`);let{app:o,key:r}=e,i=this.pool.get(o),s=i?await i.then(a=>a,()=>null):null;if(s===null)throw new B(`the session behind this handle (app "${o}") is gone \u2014 closed, evicted after its pod died, or the proxy restarted; it cannot be reattached. Reconnect without a handle to start a new session.`);if(this.keyOf(s)!==r)throw new B(`the session behind this handle (app "${o}") was replaced \u2014 the original backhaul died and a new session serves this app now; the handle's session cannot be reattached. Reconnect without a handle to start a new session.`);return{app:o,entry:s}}async evict(n,e){let o=this.pool.get(n);!o||await o.then(i=>i,()=>null)!==e||this.pool.get(n)!==o||(this.pool.delete(n),await this.opts.closeSession(e).catch(()=>{}))}async modelList(){return{object:"list",data:(await this.availableIds()).map(e=>({id:e,object:"model",created:0,owned_by:"urun"}))}}async closeAll(){let n=[...this.pool.values()];this.pool.clear();let e=[];if(await Promise.all(n.map(async o=>{try{await this.opts.closeSession(await o)}catch(r){e.push(String(r instanceof Error?r.message:r))}})),e.length>0)throw new Error(`failed to close ${e.length} pooled session(s): ${e.join("; ")}`)}};var Ee=class{constructor(n,e){this.backend=n;this.opts=e}backend;opts;source=null;sink=null;outHandlers=new Set;async startOutbound(){return this.source=this.backend.createSource(),this.source.track}appendInputAudio(n){if(!this.source)throw new Error("startOutbound() not called");let e=Buffer.from(n,"base64"),o=Math.floor(e.byteLength/2),r=new Int16Array(o);for(let i=0;i<o;i++)r[i]=e.readInt16LE(i*2);this.source.onData(r)}async startInbound(n){if(n==null)throw new Error('startInbound: no downstream audio track \u2014 the runtime audio producer has not been consumed yet (session.stream("rt-audio-out").track is null)');this.sink=this.backend.createSink(n),this.sink.onframe=e=>{let o=Buffer.from(e.samples.buffer,e.samples.byteOffset,e.samples.byteLength).toString("base64");for(let r of this.outHandlers)r(o)}}onOutputAudio(n){return this.outHandlers.add(n),()=>this.outHandlers.delete(n)}};async function Ye(t,n,e=1e4){let o=new Ee(n,{sampleRate:24e3}),r=await o.startOutbound();await t.stream("rt-audio-in").attach(r);let i=await Pt(t,e);return await o.startInbound(i),o}function Pt(t,n){let e=t.stream("rt-audio-out");return e.track!=null?Promise.resolve(e.track):new Promise((o,r)=>{let i,s=setTimeout(()=>{i?.(),r(new Error(`enableSessionAudio: no downstream audio track within ${n}ms (runtime produced no voice OUT / SFU audio consumer never attached)`))},n),a=l=>{l!=null&&(clearTimeout(s),i?.(),o(l))};i=e.on?.("track",a),e.track!=null&&a(e.track)})}var We="rt-video-in",ze=256e3;function Xe(t){let n=t.stream(We),e=n.emit;if(typeof e!="function")throw new Error(`enableSessionVideo: session.stream('${We}') has no emit() \u2014 the named-DATA stream produce seam (@urun-sh/core SessionStream.emit, contract \xA75) is required for the video frame lane`);return{sendInputFrame:async o=>{if(o.byteLength===0)throw new Error("enableSessionVideo: refusing to send an empty video frame");if(o.byteLength>ze)throw new Error(`enableSessionVideo: frame is ${o.byteLength} bytes \u2014 over the ${ze}-byte data-channel budget (the SFU's 262144-byte SCTP ceiling drops oversize messages SILENTLY; send smaller/lower-quality frames)`);await e.call(n,o)}}}var Ct=24e3,se=480,Ut=111;async function Nt(){let t=await import("opusscript"),n=t.default??t;return new n(Ct,1,2048)}async function Qe(){let t=await import("werift"),n=await Nt();return{createSource:()=>{let e=new t.MediaStreamTrack({kind:"audio"}),o=Math.random()*4294967295>>>0,r=Math.random()*65535&65535,i=Math.random()*4294967295>>>0,s=new Int16Array(0);return{track:e,onData:a=>{let l=new Int16Array(s.length+a.length);l.set(s,0),l.set(a,s.length);let u=0;for(;l.length-u>=se;){let d=l.subarray(u,u+se);u+=se;let f=Buffer.from(d.buffer,d.byteOffset,d.byteLength),m=n.encode(f,se);r=r+1&65535,i=i+se>>>0;let w=new t.RtpHeader({version:2,payloadType:Ut,sequenceNumber:r,timestamp:i,ssrc:o,marker:!1}),y=new t.RtpPacket(w,m);e.writeRtp(y)}s=l.slice(u)}}},createSink:e=>{let o={onframe:null};return e.onReceiveRtp.subscribe(i=>{let s;try{s=n.decode(i.payload)}catch{return}let a=new Int16Array(s.length/2);for(let l=0;l<a.length;l++)a[l]=s.readInt16LE(l*2);o.onframe?.({samples:a})}),o}}}var Lt=24e3,Un=`audio/pcm;rate=${Lt}`,Mt=Buffer.alloc(960).toString("base64");var Te=class t{startedAt=Date.now();requests={};models={};ttftMs=[];tokRates=[];tokensOut=0;charsOut=0;localInBytes=0;naiveInBytes=0;request(n,e,o){this.requests[n]=(this.requests[n]??0)+1,this.localInBytes+=e,this.naiveInBytes+=o}modelEntry(n){let e=this.models[n];if(e)return e;let o=Object.keys(this.models).length>=jt?"(other)":n;return this.models[o]??={requests:0,tokens_out:0}}model(n){this.modelEntry(n).requests+=1}track(n,e){let o=this,r=performance.now(),i=null,s=0;return(async function*(){for await(let l of n)l.type==="response.output_text.delta"&&typeof l.delta=="string"&&(i===null&&(i=performance.now(),o.ttftMs.push(i-r),o.ttftMs.length>512&&o.ttftMs.shift()),s+=1,o.tokensOut+=1,o.charsOut+=l.delta.length,e!==void 0&&(o.modelEntry(e).tokens_out+=1)),yield l;let a=(performance.now()-(i??r))/1e3;s>0&&a>0&&(o.tokRates.push(s/a),o.tokRates.length>512&&o.tokRates.shift())})()}static p50(n){if(!n.length)return null;let e=[...n].sort((o,r)=>o-r);return e[Math.floor(e.length/2)]}snapshot(){let n=Object.values(this.requests).reduce((e,o)=>e+o,0);return{uptime_s:Math.round((Date.now()-this.startedAt)/1e3),requests:this.requests,models:this.models,tokens_out:this.tokensOut,chars_out:this.charsOut,ttft_ms:{p50:t.p50(this.ttftMs),last:this.ttftMs.at(-1)??null},tok_per_s:{p50:t.p50(this.tokRates),last:this.tokRates.at(-1)??null},traffic:{local_request_bytes:this.localInBytes,naive_baseline:{request_bytes:this.naiveInBytes,connections:n},persistent_connections:1,history_bytes_saved:Math.max(0,this.naiveInBytes-this.localInBytes),history_bytes_saved_pct:this.naiveInBytes>0?Math.round((1-this.localInBytes/this.naiveInBytes)*1e3)/10:null}}}};function Ce(t){t.writeHead(200,{"content-type":"text/event-stream","cache-control":"no-cache",connection:"keep-alive"})}function j(t,n,e){t.writeHead(n,{"content-type":"application/json"}),t.end(JSON.stringify(e))}function J(t,n,e,o="invalid_request_error"){j(t,n,{error:{message:e,type:o}})}function H(t,n,e,o){j(t,n,{type:"error",error:{type:e,message:o}})}function Ue(t,n,e){if(n==="anthropic"){j(t,404,{type:"error",error:{type:"not_found_error",message:e.message}});return}j(t,404,{error:{message:e.message,type:"invalid_request_error",code:"model_not_found"}})}var jt=256,Ft=128;function Ne(t){return typeof t.model!="string"||!t.model.trim()?"(default)":t.model.trim().slice(0,Ft)}var Ze=64*1024*1024,me=class extends Error{};async function Gt(t){let n=[],e=0;for await(let r of t){if(e+=r.length,e>Ze)throw new me(`request body exceeds ${Ze} bytes`);n.push(r)}let o=Buffer.concat(n).toString("utf8");return o?JSON.parse(o):{}}function X(t,n,e,o){return{id:t,object:"chat.completion.chunk",created:Math.floor(Date.now()/1e3),model:n,choices:[{index:0,delta:e,finish_reason:o}]}}var qt=256;function Jt(t){if(typeof t=="string")return[{role:"user",content:t}];if(Array.isArray(t))return t;throw new Error("responses input must be a string or an array of messages/items")}var Pe=class{conversations=new Map;thread(n,e){let o=Jt(e);if(n==null)return o;let r=this.conversations.get(String(n));if(!r)throw new Error(`unknown previous_response_id ${String(n)} (proxy restarts drop stored responses)`);return[...r,...o]}remember(n,e,o){this.conversations.set(n,[...e,{role:"assistant",content:o}]);for(let r of this.conversations.keys()){if(this.conversations.size<=qt)break;this.conversations.delete(r)}}};function ie(t){let n=t?.output;return Array.isArray(n)?n.filter(e=>e.type==="function_call"):[]}function W(t){return t.find(n=>n.type==="response.completed")?.response??null}function ge(t){let n=t?.usage;if(!n||typeof n!="object")return null;let e=s=>typeof s=="number"&&Number.isFinite(s)&&s>=0?s:0,o=e(n.input_tokens),r=e(n.output_tokens);if(o===0&&r===0)return null;let i=e(n.total_tokens)||o+r;return{input_tokens:o,output_tokens:r,total_tokens:i}}function et(t){let n=ge(t);return n?{prompt_tokens:n.input_tokens,completion_tokens:n.output_tokens,total_tokens:n.total_tokens}:{prompt_tokens:0,completion_tokens:0,total_tokens:0}}function nt(t){let n=t?.output;if(!Array.isArray(n))return"";let e="";for(let o of n)if(!(o.type!=="reasoning"||!Array.isArray(o.content)))for(let r of o.content)r.type==="reasoning_text"&&(e+=String(r.text??""));return e}function ae(t){return t.type!=="response.function_call_arguments.delta"||typeof t.tool_index!="number"?null:{tool_index:t.tool_index,call_id:typeof t.call_id=="string"?t.call_id:void 0,name:typeof t.name=="string"?t.name:void 0,delta:typeof t.delta=="string"?t.delta:""}}var z=class{calls=new Map;add(n){let e=this.calls.get(n.tool_index),o=!e;return e||(e={args:""},this.calls.set(n.tool_index,e)),n.call_id&&(e.id=n.call_id),n.name&&(e.name=n.name),e.args+=n.delta,{call:e,isNew:o}}get size(){return this.calls.size}items(){return[...this.calls.entries()].map(([n,e])=>({type:"function_call",call_id:e.id??`call_${n}`,name:e.name??"",arguments:e.args}))}},Kt=["<tool_call>","<function="];function Vt(t){return Kt.find(n=>t.includes(n))}function Q(t,n){return`serving row emitted a prose tool call (literal ${JSON.stringify(n)} in the assistant text) for a request that carried tools, and no structured tool_calls arrived \u2014 the tool-call parser is not configured on the model row "${t}" (see catalog parser_defaults)`}function Z(t,n,e){if(!(!t||n>0))return Vt(e)}function Wt(t){if(Array.isArray(t))return t.map(n=>{let e=n.function;return!e||typeof e!="object"?n:{type:"function",name:e.name,description:e.description,parameters:e.parameters,strict:e.strict}})}var he=class extends Error{};function zt(t,n){return Array.isArray(n)?n.map(e=>{let o=e;if(o.type==="text")return{type:t==="assistant"?"output_text":"input_text",text:o.text};if(o.type==="image_url"){let r=o.image_url??{};return{type:"input_image",image_url:r.url,detail:r.detail}}throw new he(`unsupported chat content part type "${String(o.type)}" \u2014 the proxy translates text and image_url parts`)}):n}function Yt(t){let n=[];for(let e of t){if(e.role==="tool"){n.push({type:"function_call_output",call_id:e.tool_call_id,output:typeof e.content=="string"?e.content:JSON.stringify(e.content??"")});continue}if(e.role==="assistant"&&Array.isArray(e.tool_calls)){typeof e.content=="string"&&e.content&&n.push({role:"assistant",content:e.content});for(let o of e.tool_calls){let r=o.function??{};n.push({type:"function_call",call_id:o.id,name:r.name,arguments:r.arguments})}continue}n.push({...e,content:zt(e.role,e.content)})}return n}function Ie(t){return ie(t).map((n,e)=>({index:e,id:n.call_id??n.id??`call_${e}`,type:"function",function:{name:n.name??"",arguments:n.arguments??"{}"}}))}function tt(t,n){let e=n?.output_text;return typeof e=="string"?e:t.filter(o=>o.type==="response.output_text.delta"&&typeof o.delta=="string").map(o=>o.delta).join("")}async function Xt(t,n,e,o,r){let i=t.stream===!0,s;try{s=e.thread(t.previous_response_id,t.input)}catch(c){J(o,400,String(c instanceof Error?c.message:c));return}t.previous_response_id!=null&&(r.naiveInBytes+=Buffer.byteLength(JSON.stringify(s))-Buffer.byteLength(JSON.stringify(t.input)));let a=`resp_${Math.random().toString(36).slice(2,14)}`,l=t.store!==!1,u=Ne(t);r.model(u);let d=Array.isArray(t.tools)&&t.tools.length>0,f;try{f=await n.createResponse({model:t.model,input:s,stream:i,tools:t.tools,tool_choice:t.tool_choice,temperature:t.temperature,max_output_tokens:t.max_output_tokens,...t.reasoning_effort!==void 0?{reasoning_effort:t.reasoning_effort}:{},...t.chat_template_kwargs!==void 0?{chat_template_kwargs:t.chat_template_kwargs}:{}})}catch(c){if(c instanceof U){Ue(o,"openai",c);return}throw c}let m=[];if(i){Ce(o);let c=g=>{o.write(`event: ${String(g.type??"message")}
|
|
2
|
+
data: ${JSON.stringify(g)}
|
|
3
|
+
|
|
4
|
+
`)},R=g=>g.response&&typeof g.response=="object"?{...g,response:{...g.response,id:a}}:g,E=`item_${a.slice(5)}`,N=`rs_${a.slice(5)}`,P=!1,$=!1,p=0,v=()=>{P||(P=!0,c({type:"response.created",response:{id:a,object:"response",status:"in_progress"}}))},M=!1,T=0,L="",D=()=>{M&&(M=!1,c({type:"response.reasoning_text.done",item_id:N,output_index:T,content_index:0,text:L}),c({type:"response.output_item.done",output_index:T,item:{id:N,type:"reasoning",summary:[],content:[{type:"reasoning_text",text:L}],status:"completed"}}))},C=!1,k=0,A="",G=g=>{C&&(C=!1,c({type:"response.output_text.done",item_id:E,output_index:k,content_index:0,text:A}),c({type:"response.content_part.done",item_id:E,output_index:k,content_index:0,part:{type:"output_text",text:A}}),c({type:"response.output_item.done",output_index:k,item:{id:E,type:"message",role:"assistant",status:g,content:[{type:"output_text",text:A}]}}))},q=new z,x=null,Y=-1,ne=g=>{if(!x)return;let S=x;x=null,c({type:"response.function_call_arguments.done",item_id:S.fcId,output_index:S.outputIndex,arguments:S.args}),c({type:"response.output_item.done",output_index:S.outputIndex,item:{id:S.fcId,type:"function_call",call_id:S.callId??S.fcId,name:S.name??"",arguments:S.args,status:g}})},oe=!1;for await(let g of r.track(f,u)){m.push(g);let S=String(g.type??"");if(S==="response.created"||S==="response.in_progress"){P=!0,c(R(g));continue}if(S==="response.output_item.added"){$=!0,c(R(g));continue}if(S==="response.reasoning_text.delta"&&!$){v(),M||(M=!0,T=p++,c({type:"response.output_item.added",output_index:T,item:{id:N,type:"reasoning",summary:[],content:[],status:"in_progress"}})),L+=String(g.delta??""),c({type:"response.reasoning_text.delta",item_id:N,output_index:T,content_index:0,delta:g.delta});continue}if(S==="response.output_text.delta"&&!$){v(),D(),C||(C=!0,k=p++,c({type:"response.output_item.added",output_index:k,item:{id:E,type:"message",role:"assistant",status:"in_progress",content:[]}}),c({type:"response.content_part.added",item_id:E,output_index:k,content_index:0,part:{type:"output_text",text:""}})),A+=String(g.delta??""),c({type:S,item_id:E,output_index:k,content_index:0,delta:g.delta});continue}let I=$?null:ae(g);if(I){if(v(),D(),G("completed"),q.add(I),I.tool_index!==Y){ne("completed"),Y=I.tool_index;let h=p++;x={fcId:`fc_${a.slice(5)}_${I.tool_index}`,outputIndex:h,callId:I.call_id,name:I.name,args:""},c({type:"response.output_item.added",output_index:h,item:{id:x.fcId,type:"function_call",call_id:x.callId??x.fcId,name:x.name??"",arguments:"",status:"in_progress"}})}x&&(I.call_id&&(x.callId=I.call_id),I.name&&(x.name=I.name),x.args+=I.delta,c({type:"response.function_call_arguments.delta",item_id:x.fcId,output_index:x.outputIndex,delta:I.delta}));continue}if(S==="response.completed"&&!$){let h=ie(g.response),O=Z(d,q.size+h.length,A);if(O){oe=!0,c({type:"error",error:{type:"urun_error",code:"tool_call_parser_missing",message:Q(u,O)}});break}D(),G("completed"),ne("completed"),q.size===0&&h.forEach(K=>{let re=p++,ce=K.id??K.call_id??`fc_${re}`,be=K.arguments??"",Je={id:ce,type:"function_call",call_id:K.call_id??ce,name:K.name??""};c({type:"response.output_item.added",output_index:re,item:{...Je,arguments:"",status:"in_progress"}}),c({type:"response.function_call_arguments.delta",item_id:ce,output_index:re,delta:be}),c({type:"response.function_call_arguments.done",item_id:ce,output_index:re,arguments:be}),c({type:"response.output_item.done",output_index:re,item:{...Je,arguments:be,status:"completed"}})}),c(R(g));continue}c(R(g))}D(),G("incomplete"),ne("incomplete"),l&&!oe&&e.remember(a,s,tt(m,W(m))),o.write(`data: [DONE]
|
|
5
|
+
|
|
6
|
+
`),o.end();return}let w=null;for await(let c of r.track(f,u)){if(m.push(c),c.type==="error"){J(o,502,JSON.stringify(c),"upstream_error");return}c.type==="response.completed"&&(w=c.response)}if(w==null){J(o,502,"upstream produced no response.completed event","upstream_error");return}let y=tt(m,w),_=Z(d,ie(w).length,y);if(_){J(o,502,Q(u,_),"upstream_error");return}l&&e.remember(a,s,y),j(o,200,{...w,id:a})}async function Qt(t,n,e,o){let r=t.messages;if(!Array.isArray(r)){J(e,400,"chat/completions requires a messages array");return}let i=String(t.model??"urun"),s=`chatcmpl-${Math.random().toString(36).slice(2,14)}`,a;try{a=Yt(r)}catch(p){if(p instanceof he){J(e,400,p.message);return}throw p}let l=Ne(t);o.model(l);let u=Array.isArray(t.tools)&&t.tools.length>0,d;try{d=await n.createResponse({model:t.model,input:a,stream:!0,tools:Wt(t.tools),tool_choice:t.tool_choice,temperature:t.temperature,max_output_tokens:t.max_completion_tokens??t.max_tokens,...t.reasoning_effort!==void 0?{reasoning_effort:t.reasoning_effort}:{},...t.chat_template_kwargs!==void 0?{chat_template_kwargs:t.chat_template_kwargs}:{}})}catch(p){if(p instanceof U){Ue(e,"openai",p);return}throw p}if(t.stream===!0){Ce(e);let p=[],v=new z,M="";e.write(`data: ${JSON.stringify(X(s,i,{role:"assistant"},null))}
|
|
7
|
+
|
|
8
|
+
`);for await(let k of o.track(d,l)){p.push(k);let A=ae(k);if(k.type==="response.output_text.delta"&&typeof k.delta=="string")M+=k.delta,e.write(`data: ${JSON.stringify(X(s,i,{content:k.delta},null))}
|
|
9
|
+
|
|
10
|
+
`);else if(k.type==="response.reasoning_text.delta"&&typeof k.delta=="string")e.write(`data: ${JSON.stringify(X(s,i,{reasoning_content:k.delta},null))}
|
|
11
|
+
|
|
12
|
+
`);else if(A){let{call:G,isNew:q}=v.add(A),x=q?{index:A.tool_index,id:G.id??`call_${A.tool_index}`,type:"function",function:{name:G.name??"",arguments:A.delta}}:{index:A.tool_index,function:{arguments:A.delta}};e.write(`data: ${JSON.stringify(X(s,i,{tool_calls:[x]},null))}
|
|
13
|
+
|
|
14
|
+
`)}else if(k.type==="error"){e.write(`data: ${JSON.stringify({error:k})}
|
|
15
|
+
|
|
16
|
+
`),e.end();return}}let T=Ie(W(p)),L=Z(u,v.size+T.length,M);if(L){e.write(`data: ${JSON.stringify({error:{type:"urun_error",code:"tool_call_parser_missing",message:Q(l,L)}})}
|
|
17
|
+
|
|
18
|
+
`),e.end();return}v.size===0&&T.length>0&&e.write(`data: ${JSON.stringify(X(s,i,{tool_calls:T},null))}
|
|
19
|
+
|
|
20
|
+
`);let D=v.size>0||T.length>0;e.write(`data: ${JSON.stringify(X(s,i,{},D?"tool_calls":"stop"))}
|
|
21
|
+
|
|
22
|
+
`),ge(W(p))&&e.write(`data: ${JSON.stringify({id:s,object:"chat.completion.chunk",created:Math.floor(Date.now()/1e3),model:i,choices:[],usage:et(W(p))})}
|
|
23
|
+
|
|
24
|
+
`),e.write(`data: [DONE]
|
|
25
|
+
|
|
26
|
+
`),e.end();return}let f="",m="",w=[],y=new z;for await(let p of o.track(d,l)){w.push(p),p.type==="response.output_text.delta"&&typeof p.delta=="string"&&(f+=p.delta),p.type==="response.reasoning_text.delta"&&typeof p.delta=="string"&&(m+=p.delta);let v=ae(p);if(v&&y.add(v),p.type==="error"){J(e,502,JSON.stringify(p),"upstream_error");return}}let _=W(w),c=Ie(_),R=c.length>0?c:Ie({output:y.items()}),E=Z(u,R.length,f);if(E){J(e,502,Q(l,E),"upstream_error");return}let N=R.map(({index:p,...v})=>v),P={role:"assistant",content:f||null},$=m||nt(_);$&&(P.reasoning_content=$),N.length>0&&(P.tool_calls=N),j(e,200,{id:s,object:"chat.completion",created:Math.floor(Date.now()/1e3),model:i,choices:[{index:0,message:P,finish_reason:N.length>0?"tool_calls":"stop"}],usage:et(_)})}function Zt(t){return typeof t=="string"?t:Array.isArray(t)?t.filter(n=>n.type==="text").map(n=>String(n.text??"")).join(""):""}var ue=class extends Error{};function ye(t){return new ue(`unsupported anthropic content block type "${t}" \u2014 the proxy translates text, tool_use and tool_result blocks`)}function en(t){if(Array.isArray(t))return t.map(n=>{let e=n;return{type:"function",name:e.name,description:e.description,parameters:e.input_schema}})}function tn(t){if(t==null)return;let n=t.type;if(n==="auto")return"auto";if(n==="none")return"none";if(n==="any")return"required";if(n==="tool")return{type:"function",name:t.name};throw new ue(`unsupported anthropic tool_choice type ${JSON.stringify(String(n))} \u2014 the proxy maps auto, none, any and tool`)}function nn(t){if(typeof t=="string")return t;if(t==null)return"";if(!Array.isArray(t))throw ye(typeof t);let n="";for(let e of t){let o=String(e.type??"");if(o!=="text")throw ye(`tool_result > ${o}`);n+=String(e.text??"")}return n}function on(t){let{role:n,content:e}=t;if(typeof e=="string")return[{role:n,content:e}];if(e==null)return[];if(!Array.isArray(e))throw ye(typeof e);let o=[],r="",i=()=>{r&&(o.push({role:n,content:r}),r="")};for(let s of e){let a=String(s.type??"");if(a==="text")r+=String(s.text??"");else if(a==="tool_use")i(),o.push({type:"function_call",call_id:s.id,name:s.name,arguments:JSON.stringify(s.input??{})});else if(a==="tool_result")i(),o.push({type:"function_call_output",call_id:s.tool_use_id,output:nn(s.content)});else{if(a==="thinking"||a==="redacted_thinking")continue;throw ye(a)}}return i(),o}function Oe(t){return ie(t).map((n,e)=>{let o;try{o=JSON.parse(n.arguments||"{}")}catch{throw new Error(`upstream function_call "${n.name??""}" arguments are not valid JSON`)}return{type:"tool_use",id:n.call_id??n.id??`toolu_${e}`,name:n.name??"",input:o}})}async function rn(t,n,e,o){let r=t.messages;if(!Array.isArray(r)){H(e,400,"invalid_request_error","messages requires a messages array");return}if(typeof t.max_tokens!="number"||!Number.isInteger(t.max_tokens)||t.max_tokens<1){H(e,400,"invalid_request_error","max_tokens is required and must be a positive integer");return}if(t.top_k!==void 0){H(e,400,"invalid_request_error","top_k is not supported: the uRun serve lane exposes temperature/top_p/stop_sequences only");return}if(t.stop_sequences!==void 0&&(!Array.isArray(t.stop_sequences)||t.stop_sequences.some(p=>typeof p!="string"))){H(e,400,"invalid_request_error","stop_sequences must be an array of strings");return}let i=String(t.model??"urun"),s=`msg_${Math.random().toString(36).slice(2,14)}`,a=[],l;try{t.system&&a.push({role:"system",content:Zt(t.system)});for(let p of r)a.push(...on(p));l=tn(t.tool_choice)}catch(p){if(p instanceof ue){H(e,400,"invalid_request_error",p.message);return}throw p}let u=Ne(t);o.model(u);let d=Array.isArray(t.tools)&&t.tools.length>0,f;try{f=await n.createResponse({model:t.model,input:a,stream:!0,tools:en(t.tools),tool_choice:l,temperature:t.temperature,max_output_tokens:t.max_tokens,top_p:t.top_p,stop:t.stop_sequences})}catch(p){if(p instanceof U){Ue(e,"anthropic",p);return}throw p}let m=(p,v)=>{e.destroyed||e.write(`event: ${p}
|
|
27
|
+
data: ${JSON.stringify({type:p,...v})}
|
|
28
|
+
|
|
29
|
+
`)};if(t.stream===!0){let p=o.track(f,u);Ce(e),m("message_start",{message:{id:s,type:"message",role:"assistant",content:[],model:i,stop_reason:null,usage:{input_tokens:0,output_tokens:0}}});let v=0,M="",T=0,L="none",D=()=>L,C=-1,k=-1,A=new z,G=[],q=()=>{L!=="none"&&(L="none",m("content_block_stop",{index:C}))},x=(h,O)=>{q(),L=h,C=T++,m("content_block_start",{index:C,content_block:O})};for await(let h of p){if(e.destroyed)break;G.push(h);let O=ae(h);if(h.type==="response.reasoning_text.delta"&&typeof h.delta=="string")D()!=="thinking"&&x("thinking",{type:"thinking",thinking:""}),m("content_block_delta",{index:C,delta:{type:"thinking_delta",thinking:h.delta}});else if(h.type==="response.output_text.delta"&&typeof h.delta=="string")D()!=="text"&&x("text",{type:"text",text:""}),v+=1,M+=h.delta,m("content_block_delta",{index:C,delta:{type:"text_delta",text:h.delta}});else if(O){let{call:K}=A.add(O);(D()!=="tool"||O.tool_index!==k)&&(k=O.tool_index,x("tool",{type:"tool_use",id:K.id??`call_${O.tool_index}`,name:K.name??"",input:{}})),O.delta&&m("content_block_delta",{index:C,delta:{type:"input_json_delta",partial_json:O.delta}})}else if(h.type==="error"){m("error",{error:{type:"api_error",message:JSON.stringify(h)}}),e.end();return}}if(e.destroyed)return;let Y=W(G),ne=ie(Y),oe=Z(d,A.size+ne.length,M);if(oe){m("error",{error:{type:"api_error",message:Q(u,oe)}}),e.end();return}q();let g=[];if(A.size===0)try{g=Oe(Y)}catch(h){m("error",{error:{type:"api_error",message:String(h instanceof Error?h.message:h)}}),e.end();return}T===0&&g.length===0&&(x("text",{type:"text",text:""}),q()),g.forEach(h=>{let O=T++;m("content_block_start",{index:O,content_block:{...h,input:{}}}),m("content_block_delta",{index:O,delta:{type:"input_json_delta",partial_json:JSON.stringify(h.input??{})}}),m("content_block_stop",{index:O})});let S=A.size>0||g.length>0,I=ge(Y);m("message_delta",{delta:{stop_reason:S?"tool_use":"end_turn"},usage:I?{input_tokens:I.input_tokens,output_tokens:I.output_tokens}:{output_tokens:v}}),m("message_stop",{}),e.end();return}let w="",y="",_=[],c=new z;for await(let p of o.track(f,u)){_.push(p),p.type==="response.output_text.delta"&&typeof p.delta=="string"&&(w+=p.delta),p.type==="response.reasoning_text.delta"&&typeof p.delta=="string"&&(y+=p.delta);let v=ae(p);if(v&&c.add(v),p.type==="error"){H(e,502,"api_error",JSON.stringify(p));return}if(e.destroyed)break}if(e.destroyed)return;let R=W(_),E;try{E=Oe(R),E.length===0&&c.size>0&&(E=Oe({output:c.items()}))}catch(p){H(e,502,"api_error",String(p instanceof Error?p.message:p));return}let N=Z(d,E.length,w);if(N){H(e,502,"api_error",Q(u,N));return}let P=[],$=y||nt(R);$&&P.push({type:"thinking",thinking:$}),(w||E.length===0)&&P.push({type:"text",text:w}),P.push(...E),j(e,200,{id:s,type:"message",role:"assistant",content:P,model:i,stop_reason:E.length>0?"tool_use":"end_turn",usage:ge(R)??{input_tokens:0,output_tokens:0}})}function ot(t){let{clients:n,apiKey:e,identity:o,statusOf:r}=t,i=typeof n=="function"?n:()=>n,s=new WeakMap,a=u=>{let d=s.get(u);return d||(d=new Pe,s.set(u,d)),d},l=new Te;return(u,d)=>{let f=new URL(u.url??"/","http://localhost"),m=f.pathname==="/v1/messages"||f.pathname.startsWith("/v1/messages/"),w=(y,_,c,R)=>m?H(d,y,R,_):J(d,y,_,c);(async()=>{if(u.method==="GET"&&f.pathname==="/healthz"){j(d,200,{ok:!0});return}if(u.method==="GET"&&f.pathname==="/stats"){j(d,200,o?{identity:o,...l.snapshot()}:l.snapshot());return}if(e&&(u.headers.authorization??"")!==`Bearer ${e}`){w(401,"invalid local proxy api key","authentication_error","authentication_error");return}let y=await i(u);if(u.method==="GET"&&f.pathname==="/v1/models"){j(d,200,await y.listModels());return}if(u.method!=="POST"){w(404,`no route for ${u.method} ${f.pathname}`,"invalid_request_error","not_found_error");return}let _;try{_=await Gt(u)}catch(R){if(R instanceof me){w(413,R.message,"request_too_large","request_too_large");return}w(400,"request body is not valid JSON","invalid_request_error","invalid_request_error");return}let c=Buffer.byteLength(JSON.stringify(_));if(f.pathname==="/v1/responses"){l.request("responses",c,c),await Xt(_,y,a(y),d,l);return}if(f.pathname==="/v1/chat/completions"){l.request("chat_completions",c,c),await Qt(_,y,d,l);return}if(f.pathname==="/v1/messages"){l.request("messages",c,c),await rn(_,y,d,l);return}if(f.pathname==="/v1/messages/count_tokens"){l.request("count_tokens",c,c),H(d,404,"not_found_error","count_tokens is not implemented by the uRun proxy: the serve transport exposes no token-count lane, and the proxy will not fabricate counts");return}w(404,`no route for POST ${f.pathname}`,"invalid_request_error","not_found_error")})().catch(y=>{let _=String(y instanceof Error?y.message:y);if(!d.headersSent){let c=r?.(y)??null;if(c){w(c.status,_,c.openaiType,c.anthropicType);return}w(500,_,"proxy_error","api_error");return}d.destroyed||(m?d.write(`event: error
|
|
30
|
+
data: ${JSON.stringify({type:"error",error:{type:"api_error",message:_}})}
|
|
31
|
+
|
|
32
|
+
`):d.write(`data: ${JSON.stringify({error:{message:_,type:"upstream_error"}})}
|
|
33
|
+
|
|
34
|
+
`)),d.end()})}}var we=require("@urun-sh/core"),F=class extends Error{constructor(e,o=401){super(e);this.status=o;this.name="ProxyAuthError"}status},V=class extends Error{constructor(n){super(n),this.name="ControlPlaneUnavailableError"}};function _e(t){let n=(t??"").trim();if(n==="")throw new F("missing Authorization header \u2014 send your uRun org API key as `Authorization: Bearer <key>`");let e=/^Bearer[ \t]+(\S.*)$/i.exec(n);if(!e)throw new F("malformed Authorization header \u2014 expected `Authorization: Bearer <urun org api key>`");return e[1].trim()}var sn=60,$e=6e4,Le=4096;function an(t){return t===401||t===403}var ee=class{constructor(n){this.opts=n;this.mint=n.mint??we.createClientToken,this.now=n.now??Date.now}opts;cache=new Map;mint;now;async resolve(n){if(n==="")throw new F("empty Authorization bearer token");let e=this.cache.get(n);if(e&&this.now()-e.at<$e)return{apiKey:n,orgId:e.orgId};let o;try{o=(await this.mint(n,{baseUrl:this.opts.gatewayUrl,expiresIn:sn})).orgId}catch(r){throw r instanceof we.ClientTokenError&&an(r.status)?(this.cache.delete(n),new F("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}: ${r instanceof Error?r.message:String(r)}`)}if(!o)throw new V("the control plane verified the API key but returned no org_id \u2014 refusing to route a request with no tenancy (control-plane contract violation)");if(this.cache.size>=Le){let r=this.cache.keys().next();r.done||this.cache.delete(r.value)}return this.cache.set(n,{orgId:o,at:this.now()}),{apiKey:n,orgId:o}}clear(){this.cache.clear()}};var yt=require("crypto");var le=require("@urun-sh/core");async function rt(t){let e=await(t.fetchImpl??fetch)(`${t.catalogUrl}/model_catalog?select=model_id,variant`,{headers:{apikey:t.anonKey,Authorization:`Bearer ${t.anonKey}`}});if(!e.ok)throw new Error(`model_catalog fetch failed: ${e.status}`);return await e.json()}function st(t,n,e,o){let r=t.response??{},i={request_id:e,consumer_id:o,stream:!0,kind:"chat",messages:n};return typeof r.instructions=="string"&&(i.instructions=r.instructions),Array.isArray(r.modalities)&&(i.modalities=r.modalities),typeof r.temperature=="number"&&(i.temperature=r.temperature),typeof r.max_output_tokens=="number"&&(i.max_output_tokens=r.max_output_tokens),r.tools!==void 0&&(i.tools=r.tools),i}function it(t,n,e){let o={request_id:n,consumer_id:e,stream:!!t.stream,kind:"responses",input:t.input};return t.model&&(o.model=t.model),t.tools!==void 0&&(o.tools=t.tools),t.tool_choice!==void 0&&(o.tool_choice=t.tool_choice),typeof t.temperature=="number"&&(o.temperature=t.temperature),typeof t.max_output_tokens=="number"&&(o.max_output_tokens=t.max_output_tokens),typeof t.reasoning_effort=="string"&&(o.reasoning_effort=t.reasoning_effort),t.chat_template_kwargs!==void 0&&(o.chat_template_kwargs=t.chat_template_kwargs),o}function at(t){if(t instanceof Error)return{type:"error",error:{type:"urun_error",code:t.name||null,message:t.message}};if(t&&typeof t=="object"&&t.t==="error"){let n=t,e=n.body??{};return{type:"error",error:{type:"urun_error",code:n.code??null,message:e.message??"unknown error"}}}return{type:"error",error:{type:"urun_error",code:null,message:String(t)}}}var un="llm-resp";async function*Me(t,n){let e=`${un}:${n}`,o=`resp_${n}`;yield{type:"response.created",response:{id:o,status:"in_progress"}};let r=new Map;for await(let i of t.stream(e).messages()){let s=i;if(s.t==="delta"){if(typeof s.delta=="string"&&(yield{type:"response.output_text.delta",item_id:o,delta:s.delta}),typeof s.reasoning=="string"&&(yield{type:"response.reasoning_text.delta",item_id:o,delta:s.reasoning}),Array.isArray(s.tool_calls))for(let a of s.tool_calls){let l=typeof a.index=="number"?a.index:0,u=r.get(l)??{};a.id&&(u.id=a.id),a.function?.name&&(u.name=a.function.name),r.set(l,u),yield{type:"response.function_call_arguments.delta",item_id:`fc_${n}_${l}`,tool_index:l,call_id:u.id,name:u.name,delta:a.function?.arguments??""}}}else if(s.t==="response"){yield{type:"response.completed",response:{id:o,status:"completed",...s.body}};return}else if(s.t==="error"){yield at(s);return}}}var ln="llm",ke=class{constructor(n){this.session=n;this.sessionTag=n.sessionId??globalThis.crypto.randomUUID()}session;sessionTag;get consumerId(){return this.session.consumerId}write(n){n.session_tag=this.sessionTag,this.session.doc(ln).set({requests:{[n.request_id]:{payload:n,consumer_id:n.consumer_id,stream:n.stream}}})}sendResponses(n,e){let o=it(n,e,this.consumerId);return this.write(o),Me(this.session,e)}sendResponseCreate(n,e,o){let r=st(n,e,o,this.consumerId);return this.write(r),Me(this.session,o)}};var ut=0;function cn(){return ut+=1,`req_${Date.now().toString(36)}_${ut.toString(36)}`}var ve=class{transport;constructor(n){this.transport=new ke(n)}responses={create:async n=>{let e=cn(),o=this.transport.sendResponses(n,e);return Object.assign((async function*(){yield*o})(),{requestId:e})}}};function pn(t){let n=t?.type;return n==="response.created"||n==="response.in_progress"}function dn(t){let n=t?.type;return n==="response.completed"||n==="error"}function lt(t){return async function(e){return(async function*(){for(let r=1;r<=2;r++){let{app:i,entry:s}=await t.sessionFor(e.model),a=!1,l=!1;try{let u=await t.createStream(s,e);for await(let d of u){if(dn(d)&&(a=!0),pn(d)){if(r>1)continue}else l=!0;yield d}}catch(u){if(r<2&&!l){await t.evict(i,s);continue}throw u}if(a)return;if(r<2&&!l){await t.evict(i,s);continue}return}})()}}var fn=3,mn="llm",gn=new Set(["ended","expired","error"]);function ct(t,n){let e=t.doc(mn),o=typeof e.onConnectionState=="function",r=typeof t.onPhase=="function";if(!o&&!r)throw new Error(`session for app "${n}" exposes neither a doc connection-state surface nor onPhase \u2014 session-gone detection requires @urun-sh/core Session (SessionDocument.onConnectionState / Session.onPhase); pooling a session with no native lifecycle signal would silently swallow requests dispatched after the platform closes it`);let i=null,s=new Set,a=[],l=u=>{if(!i){i=u;for(let d of[...s])d(u);s.clear()}};return o&&a.push(e.onConnectionState(u=>{u.consecutiveFailures<fn||l(new B(`the pooled session for app "${n}" is gone: its doc backhaul failed ${u.consecutiveFailures} consecutive times (last close ${u.lastCloseCode===null?"by provider liveness reap":`code ${u.lastCloseCode}`}${u.lastCloseReason?`, reason ${JSON.stringify(u.lastCloseReason)}`:""}) \u2014 the platform closed the session (idle timeout or a serve-side crash). The entry is evicted; retry to get a fresh session.`))})),r&&a.push(t.onPhase(u=>{gn.has(u.name)&&l(new B(`the pooled session for app "${n}" is gone: it reached terminal phase "${u.name}". The entry is evicted; retry to get a fresh session.`))})),{gone:()=>i,onGone(u){return i?(u(i),()=>{}):(s.add(u),()=>s.delete(u))},dispose(){s.clear();for(let u of a.splice(0))u()}}}async function*De(t,n){let e=n.gone();if(e)throw e;let o=t[Symbol.asyncIterator](),r,i=new Promise((s,a)=>{r=n.onGone(l=>a(l))});i.catch(()=>{});try{for(;;){let s=await Promise.race([o.next(),i]);if(s.done)return;yield s.value}}finally{r?.(),o.return?.(void 0)}}function pt(t,n,e){let o=t[e];if(typeof o!="function")throw new Error(`app "${n}" has no function "${e}" (set URUN_FUNCTION)`);let r=o();if(typeof r.end!="function")throw new Error(`app "${n}" function "${e}" returned a session without end()`);return r}function hn(t,n){let{baseUrl:e,orgId:o,fnName:r,auth:i,subject:s}=t,a=f=>({session:f,responses:new ve(f),gone:ct(f,n)});if(i.lane==="jwt"){let f=(0,le.App)(n,{baseUrl:e,orgId:o,jwt:i.jwt});return a(pt(f,n,r))}let{apiKey:l,gatewayUrl:u}=i,d=(0,le.App)(n,{baseUrl:e,orgId:o,getAccessToken:async()=>(await(0,le.createClientToken)(l,{baseUrl:u,expiresIn:300,allowedFunctions:[`${n}/${r}`],...s===void 0?{}:{subject:s}})).token});return a(pt(d,n,r))}function yn(t){let n=t.session.id;if(typeof n!="string"||n.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 n}var wn=new Set(["error","ended","expired"]);function dt(t){let{auth:n,catalog:e}=t,o=null,r=s=>{let a=hn(t,s);return a.gone.onGone(()=>{o?.evict(s,a)}),a},i=new fe({defaultApp:t.defaultApp,fnName:t.fnName,openSession:r,closeSession:async s=>{s.gone.dispose(),await s.session.end()},listApps:n.lane==="api-key"?()=>Ve({apiUrl:t.apiUrl,apiKey:n.apiKey}):null,listCatalog:e?()=>rt(e):null,sessionKey:yn});return o=i,t.defaultApp!==null&&i.seed(t.defaultApp,r(t.defaultApp)),i}function ft(t,n=Qe){let e=new WeakMap,o=new WeakMap;return{createResponse:lt({sessionFor:r=>t.sessionFor(r),evict:(r,i)=>t.evict(r,i),createStream:async(r,i)=>De(await r.responses.responses.create(i),r.gone)}),listModels:()=>t.modelList(),openAudio:async r=>{let{entry:i}=await t.sessionFor(r),s=e.get(i);return s||(s=(async()=>Ye(i.session,await n()))(),e.set(i,s),s.catch(()=>e.delete(i))),s},openVideo:async r=>{let{entry:i}=await t.sessionFor(r),s=o.get(i);return s||(s=Xe(i.session),o.set(i,s)),s},sessionHandle:async r=>(await t.handleFor(r)).handle,createResponseOn:async(r,i)=>{let{entry:s}=await t.sessionForHandle(r);return De(await s.responses.responses.create(i),s.gone)},onSessionEnd:async(r,i)=>{let{entry:s}=await t.sessionForHandle(r),a=s.session;if(typeof a.onPhase!="function")throw new Error("pooled session exposes no onPhase \u2014 session-end (goAway) wiring requires @urun-sh/core Session");let l=!1;return a.onPhase(u=>{if(l||!wn.has(u.name))return;l=!0;let d=a.endsAt??null;i({timeLeftMs:d?Math.max(0,d.getTime()-Date.now()):null,reason:u.name==="expired"?"session reached its maximum length":`session ${u.name}`})})}}}var _n=de(require("os"),1),kn=de(require("path"),1),vn=require("url"),xn=require("util");function mt(t,n="URUN_BASE_URL"){if(/\/v1\/*$/.test(new URL(t).pathname))throw new Error(`${n} must be the session-gateway base (e.g. https://api.urun.sh), got ${t} \u2014 a trailing /v1 is the org control-plane API form (that belongs in URUN_API_URL); session allocation 404s against it. Drop the /v1.`);return t}function xe(t,n="URUN_GATEWAY_URL"){let e=new URL(t),o=e.hostname==="localhost"||e.hostname==="127.0.0.1"||e.hostname==="[::1]";if(e.protocol!=="https:"&&!o)throw new Error(`${n} must be https (got ${t}) \u2014 the org API key rides the request's Authorization header`);return t}var Be=8080,Se="serve",gt="0.0.0.0";function Sn(t,n,e){let o=(t[n]??"").trim();if(!o)throw new Error(`${n} is not set \u2014 ${e}`);return o}function ht(t=process.env){let n=(t.PORT??"").trim(),e=n===""?Be:Number(n);if(!Number.isInteger(e)||e<0||e>65535)throw new Error(`PORT must be an integer in 0..65535, got ${JSON.stringify(n)}`);let o=mt(xe(Sn(t,"URUN_BASE_URL","the hosted proxy needs the uRun session-gateway base it opens sessions and verifies API keys against (e.g. https://api.urun.sh)"),"URUN_BASE_URL")),r=(t.URUN_API_URL??"").trim(),i=r?xe(r,"URUN_API_URL"):`${o.replace(/\/+$/,"")}/v1`,s=(t.URUN_CATALOG_URL??"").trim(),a=(t.URUN_CATALOG_ANON_KEY??"").trim();if(!!s!=!!a)throw new Error("URUN_CATALOG_URL and URUN_CATALOG_ANON_KEY must be set together (or both left unset) \u2014 exactly one of them is set, which would silently disable the model-catalog oracle");let l=s?{catalogUrl:xe(s,"URUN_CATALOG_URL"),anonKey:a}:null;return{port:e,baseUrl:o,apiUrl:i,catalog:l}}var He=512;function je(t){return`urun-inference-proxy-${(0,yt.createHash)("sha256").update(t).digest("hex").slice(0,32)}`}var te=class{constructor(n){this.opts=n}opts;tenants=new Map;build(n){if(this.opts.build)return this.opts.build(n);let e=dt({baseUrl:this.opts.baseUrl,orgId:n.orgId,defaultApp:null,fnName:Se,apiUrl:this.opts.apiUrl,auth:{lane:"api-key",apiKey:n.apiKey,gatewayUrl:this.opts.baseUrl},catalog:this.opts.catalog,subject:je(n.apiKey)});return{router:e,clients:ft(e)}}clientsFor(n){let e=this.tenants.get(n.apiKey);if(e)return this.tenants.delete(n.apiKey),this.tenants.set(n.apiKey,e),e.clients;for(;this.tenants.size>=He;){let r=this.tenants.entries().next();if(r.done)break;let[i,s]=r.value;this.tenants.delete(i),s.router.closeAll().catch(()=>{})}let o=this.build(n);return this.tenants.set(n.apiKey,o),o.clients}get size(){return this.tenants.size}async closeAll(){let n=[...this.tenants.values()].map(o=>o.router);this.tenants.clear();let e=[];if(await Promise.all(n.map(async o=>{try{await o.closeAll()}catch(r){e.push(String(r instanceof Error?r.message:r))}})),e.length>0)throw new Error(`failed to close ${e.length} tenant backhaul(s): ${e.join("; ")}`)}};var Ge=5e3,qe=3e3,bn="/api/client-tokens";function Fe(t,n,e){t.writeHead(n,{"content-type":"application/json"}),t.end(JSON.stringify(e))}function wt(t,n,e,o){Fe(t,n,{error:{message:e,type:o}})}async function An(t,n=fetch){let e=`${t.replace(/\/+$/,"")}${bn}`;try{let o=await n(e,{method:"POST",headers:{"content-type":"application/json"},body:"{}",signal:AbortSignal.timeout(qe)});return o.status===404?{ready:!1,detail:`${e} answered 404 \u2014 this is not a uRun control plane`}:o.status>=500?{ready:!1,detail:`${e} answered HTTP ${o.status}`}:{ready:!0,detail:`${e} answered HTTP ${o.status}`}}catch(o){return{ready:!1,detail:`${e} unreachable: ${o instanceof Error?o.message:String(o)}`}}}function kt(t){let n=t.resolver??new ee({gatewayUrl:t.baseUrl}),e=t.registry??new te({baseUrl:t.baseUrl,apiUrl:t.apiUrl,catalog:t.catalog}),o=t.probeControlPlane??(()=>An(t.baseUrl)),r=async l=>n.resolve(_e(l.headers.authorization)),i=ot({clients:async l=>e.clientsFor(await r(l)),statusOf:l=>l instanceof F?{status:l.status,openaiType:"authentication_error",anthropicType:"authentication_error"}:l instanceof V?{status:502,openaiType:"upstream_error",anthropicType:"api_error"}:null}),s=null;return{server:(0,_t.createServer)((l,u)=>{let d=new URL(l.url??"/","http://localhost");(async()=>{if(l.method==="GET"&&d.pathname==="/healthz"){Fe(u,200,{ok:!0});return}if(l.method==="GET"&&d.pathname==="/readyz"){let f=Date.now();if(!s||f-s.at>=Ge){let m=await o();s={at:f,...m}}Fe(u,s.ready?200:503,{ok:s.ready,control_plane:s.detail});return}if(d.pathname!=="/v1"&&!d.pathname.startsWith("/v1/")){wt(u,404,`no route for ${l.method} ${d.pathname} \u2014 this endpoint serves the OpenAI-compatible /v1 surface plus /healthz and /readyz`,"invalid_request_error");return}i(l,u)})().catch(f=>{if(!u.headersSent){wt(u,500,String(f instanceof Error?f.message:f),"proxy_error");return}u.end()})}),closeAll:()=>e.closeAll()}}0&&(module.exports={BIND_HOST,ControlPlaneUnavailableError,DEFAULT_PORT,MAX_CACHED_KEYS,MAX_TENANTS,ORG_BINDING_TTL_MS,OrgResolver,ProxyAuthError,READINESS_TIMEOUT_MS,READINESS_TTL_MS,SERVE_FUNCTION,TenantRegistry,bearerFrom,createHostedProxy,resolveHostedConfig,tenantSubject});
|