@urun-sh/openai 0.6.18 → 0.6.20

package/dist/hosted/index.d.cts

@@ -1,7 +1,7 @@
 import { Server } from 'node:http';
 import * as _prometheus_io_client from '@prometheus-io/client';
 import { Registry } from '@prometheus-io/client';
-import { c as LedgerOutcome, I as InferenceRecord, S as SessionGoneError, P as ProxyClients, M as ModelRouter, i as UsageRowReporter, b as InferenceUsageRecord, e as LedgerWrite, d as LedgerRecorded, L as LedgerFailure } from '../responses-turn-B91w2YKn.cjs';
+import { c as LedgerOutcome, I as InferenceRecord, S as SessionGoneError, P as ProxyClients, M as ModelRouter, i as UsageRowReporter, b as InferenceUsageRecord, e as LedgerWrite, d as LedgerRecorded, L as LedgerFailure } from '../responses-turn-Cb1qQBA7.cjs';
 import { createClientToken } from '@urun-sh/core';
 import { U as UrunResponses } from '../ResponsesClient-_OZERUjH.cjs';
 import { b as UrunSessionLike } from '../types-CHPtJx6d.cjs';
@@ -419,12 +419,13 @@ type PoolEntry = {
  *   1. Nothing here is authoritative. A replica's tenant map is a CACHE of
  *      backhauls it happens to have open; a cold replica rebuilds it from the
  *      caller's key on the first request. Losing it changes no answer.
- *   2. Replicas converge on the platform's OWN dedupe rather than on a
- *      registry of ours: client tokens are minted with a STABLE `subject`
- *      derived from the key, and the platform dedupes sessions per
- *      (org, actor, app, function) — so replica A and replica B dialing the
- *      same (key, app) coalesce onto ONE platform session. That is the
- *      library's own mechanism doing the sharding.
+ *   2. Every pool entry mints the SDK default credential — a fresh random
+ *      subject per token — so a dial is a FRESH platform session owned by
+ *      that dial alone. There is nothing to coalesce across replicas, and
+ *      no pinned name for one replica's actor to collide with another's
+ *      (`session_name_not_yours`, owner ruling 2026-10-01). The pool's
+ *      bounded/reuse semantics live per process; cross-replica coalescing
+ *      is NOT a pool property.
  * The yjs/valkey state plane in urun-infra#1490 §7 is therefore NOT built
  * here, and is not silently missing either — it is unnecessary for this
  * phase's correctness and is called out as such in the PR.
@@ -458,24 +459,6 @@ interface TenantRegistryOptions {
     apiUrl: string;
     /** The model-catalog oracle for routing rule 5, or null. */
     catalog: CatalogConfig | null;
-    /**
-     * THE STATELESS DIRECT PATH (urun-ts #497): present ⇒ every tenant's
-     * clients carry a per-caller decision dial — discovery over the CALLER'S
-     * own key (`GET {apiUrl}/runtimes`), tokens minted with the caller's
-     * org/app/fn, the flat body POSTed to the ready runtime's `:8099/v1/decision`.
-     * `shared` (the shared org's API key, when mounted) extends the same dial
-     * to SHARED-catalog models: discovery under the shared key, the token's
-     * `tenant` stays the caller and `provider` names the pod org from the
-     * block's `shared_org_id` (the intake's provider-aware org gate). Absent
-     * ⇒ shared models keep the session dial. Absent ⇒ no dial exists and every
-     * decision rides the session lane byte-for-byte.
-     */
-    decisionIntake?: {
-        secret: string;
-        shared?: {
-            apiKey: string;
-        };
-    };
     /** Injected in tests; production builds real uRun backhauls. */
     build?: (caller: CallerIdentity) => {
         router: ModelRouter<PoolEntry>;
@@ -921,39 +904,6 @@ interface HostedConfig {
     apiUrl: string;
     /** The `model_catalog` oracle for routing rule 5, or null when unconfigured. */
     catalog: CatalogConfig | null;
-    /**
-     * THE STATELESS DIRECT PATH's credential (URUN_SESSION_TOKEN_SECRET) — the
-     * runtime scoped-token signing secret the decision intake verifies
-     * (urun-python decision_intake). A SECRET input, not a behavior knob: it is
-     * the mandate's allowed env class, and it is the SAME predicate the local
-     * lane enables on (cli.ts statelessDialEnabled), so the two lanes can never
-     * disagree. Absent ('') the stateless lane does not exist and every
-     * decision rides the session lane byte-for-byte. Present, decisions for
-     * CALLER-ORG serves="openai" apps dispatch straight to a ready runtime's
-     * `:8099/v1/decision`: discovery rides the CALLER'S OWN key
-     * (`GET {apiUrl}/runtimes`), and the per-decision token is minted with the
-     * caller's org/app/fn identity. Shared-catalog models keep the session
-     * dial — the v1 boundary #497 declares (the dial's own targetFor refuses
-     * them).
-     *
-     * OPTIONAL on this bag because legacy constructors may omit it
-     * (resolveHostedConfig always supplies it — '' when the env is absent);
-     * consumers read it as `?? ''`.
-     */
-    decisionSecret?: string;
-    /**
-     * THE SHARED LANE'S credential (URUN_SHARED_ORG_API_KEY): the shared org's
-     * API key, so the decision dial can discover — and its tokens can name via
-     * the block's `shared_org_id` — the SHARED org's pods. Optional: absent ⇒
-     * shared models keep the session dial (the stateless boundary narrows to
-     * "no shared credential mounted"). URUN_SHARED_ORG_ID must ride WITH it
-     * (both-or-neither, loud) purely as the identity statement of the key —
-     * the dial never re-spells the pod org; it reads `shared_org_id` from the
-     * block that resolved the lane, so key and pod org cannot drift.
-     */
-    sharedOrgKey?: string;
-    /** Identity statement accompanying `sharedOrgKey` (URUN_SHARED_ORG_ID). */
-    sharedOrgId?: string;
 }
 /**
  * Read and VALIDATE the deployment's external inputs. Every failure is loud
@@ -1092,9 +1042,9 @@ declare function createHostedProxy(options: HostedProxyOptions): {
      */
     edgeIdentityServer: Server;
     /**
-     * The tenant registry this proxy built (production wiring). Tests assert
-     * the per-caller stateless dial through it; production code goes through
-     * `callerOf`, never here.
+     * The tenant registry this proxy built (production wiring). Also the ops
+     * surface for the tenant count. Production code goes through `callerOf`,
+     * never here.
      */
     registry: TenantRegistry;
     closeAll: () => Promise<void>;

package/dist/hosted/index.d.ts

@@ -1,7 +1,7 @@
 import { Server } from 'node:http';
 import * as _prometheus_io_client from '@prometheus-io/client';
 import { Registry } from '@prometheus-io/client';
-import { c as LedgerOutcome, I as InferenceRecord, S as SessionGoneError, P as ProxyClients, M as ModelRouter, i as UsageRowReporter, b as InferenceUsageRecord, e as LedgerWrite, d as LedgerRecorded, L as LedgerFailure } from '../responses-turn-FcEDaymF.js';
+import { c as LedgerOutcome, I as InferenceRecord, S as SessionGoneError, P as ProxyClients, M as ModelRouter, i as UsageRowReporter, b as InferenceUsageRecord, e as LedgerWrite, d as LedgerRecorded, L as LedgerFailure } from '../responses-turn-DmzWIpmO.js';
 import { createClientToken } from '@urun-sh/core';
 import { U as UrunResponses } from '../ResponsesClient-CSSrOYD8.js';
 import { b as UrunSessionLike } from '../types-CHPtJx6d.js';
@@ -152,13 +152,6 @@ interface TenantRegistryOptions {
     
     catalog: CatalogConfig | null;
     
-    decisionIntake?: {
-        secret: string;
-        shared?: {
-            apiKey: string;
-        };
-    };
-    
     build?: (caller: CallerIdentity) => {
         router: ModelRouter<PoolEntry>;
         clients: ProxyClients;
@@ -300,12 +293,6 @@ interface HostedConfig {
     apiUrl: string;
     
     catalog: CatalogConfig | null;
-    
-    decisionSecret?: string;
-    
-    sharedOrgKey?: string;
-    
-    sharedOrgId?: string;
 }
 
 declare function resolveHostedConfig(env?: NodeJS.ProcessEnv): HostedConfig;

package/dist/hosted/index.js

@@ -1 +1 @@
-import{A as x,B as v,C as m,D as y,E as B,F as H,G as P,H as u,I as G,J as U,K as F,L as X,M as Y,N as b,O as h,P as k,Q as K,R as V,S as j,T as W,U as q,a as r,b as o,c as t,d as E,e as n,f as _,g as a,h as T,i as O,j as i,k as s,l as I,m as p,n as A,o as R,p as S,q as N,r as C,s as D,t as L,u as M,v as d,w as f,x as l,y as c,z as g}from"../chunk-O4TKY7NE.js";import"../chunk-MLJ66LPI.js";import"../chunk-K23AZPI4.js";import"../chunk-3U7YQ7LE.js";import"../chunk-76BZLCGT.js";import"../chunk-NYSZ3RGB.js";import"../chunk-3JWYIKHM.js";import"../chunk-6M5JK4YC.js";import"../chunk-7V2JKPIG.js";import{c as e}from"../chunk-YSFSRI3D.js";e();export{N as ANONYMOUS_KEY_ID,y as BIND_HOST,R as BUCKET_DOMAIN,O as BoundedLabel,f as ControlPlaneUnavailableError,X as ConversationBackhauls,F as ConversationCapacityError,v as DEFAULT_PORT,I as EDGE_IDENTITY_PORT,A as EDGE_KEY_FINGERPRINT_HEADER,p as EDGE_KEY_ID_HEADER,i as InferenceMetrics,Y as LEDGER_FUNCTION_PATH,b as LEDGER_WRITE_TIMEOUT_MS,V as LedgerClient,h as LedgerDuplicateError,k as LedgerSurfaceError,g as MAX_CACHED_KEYS,G as MAX_CONVERSATIONS,U as MAX_CONVERSATION_HANDLES,n as MAX_MODEL_LABELS,_ as MAX_ORG_LABELS,H as MAX_TENANTS,o as METRICS_PATH,r as METRICS_PORT,E as NO_MODEL_LABEL,c as ORG_BINDING_TTL_MS,t as OTHER_LABEL,x as OrgResolver,d as ProxyAuthError,W as READINESS_TIMEOUT_MS,j as READINESS_TTL_MS,m as SERVE_FUNCTION,S as TIER_DOMAIN,u as TenantRegistry,l as bearerFrom,L as bearerTokenOf,M as createEdgeIdentityServer,q as createHostedProxy,s as createMetricsServer,C as edgeBucketId,D as edgeTierFingerprint,a as laneOf,K as parseLedgerRecorded,B as resolveHostedConfig,T as statusClassOf,P as tenantSubject};
+import{A as x,B as v,C as m,D as y,E as B,F as H,G as P,H as u,I as G,J as U,K as F,L as X,M as Y,N as b,O as h,P as k,Q as K,R as V,S as j,T as W,U as q,a as r,b as o,c as t,d as E,e as n,f as _,g as a,h as T,i as O,j as i,k as s,l as I,m as p,n as A,o as R,p as S,q as N,r as C,s as D,t as L,u as M,v as d,w as f,x as l,y as c,z as g}from"../chunk-64WZRHYB.js";import"../chunk-7ZH2UHK6.js";import"../chunk-K23AZPI4.js";import"../chunk-TF2IMK3T.js";import"../chunk-76BZLCGT.js";import"../chunk-BAKPG5S3.js";import"../chunk-3JWYIKHM.js";import"../chunk-6M5JK4YC.js";import"../chunk-7V2JKPIG.js";import{c as e}from"../chunk-YSFSRI3D.js";e();export{N as ANONYMOUS_KEY_ID,y as BIND_HOST,R as BUCKET_DOMAIN,O as BoundedLabel,f as ControlPlaneUnavailableError,X as ConversationBackhauls,F as ConversationCapacityError,v as DEFAULT_PORT,I as EDGE_IDENTITY_PORT,A as EDGE_KEY_FINGERPRINT_HEADER,p as EDGE_KEY_ID_HEADER,i as InferenceMetrics,Y as LEDGER_FUNCTION_PATH,b as LEDGER_WRITE_TIMEOUT_MS,V as LedgerClient,h as LedgerDuplicateError,k as LedgerSurfaceError,g as MAX_CACHED_KEYS,G as MAX_CONVERSATIONS,U as MAX_CONVERSATION_HANDLES,n as MAX_MODEL_LABELS,_ as MAX_ORG_LABELS,H as MAX_TENANTS,o as METRICS_PATH,r as METRICS_PORT,E as NO_MODEL_LABEL,c as ORG_BINDING_TTL_MS,t as OTHER_LABEL,x as OrgResolver,d as ProxyAuthError,W as READINESS_TIMEOUT_MS,j as READINESS_TTL_MS,m as SERVE_FUNCTION,S as TIER_DOMAIN,u as TenantRegistry,l as bearerFrom,L as bearerTokenOf,M as createEdgeIdentityServer,q as createHostedProxy,s as createMetricsServer,C as edgeBucketId,D as edgeTierFingerprint,a as laneOf,K as parseLedgerRecorded,B as resolveHostedConfig,T as statusClassOf,P as tenantSubject};

package/dist/pi-extension/index.js

@@ -1,4 +1,4 @@
-import{a as z}from"../chunk-76BZLCGT.js";import{e as W,h as Q}from"../chunk-NYSZ3RGB.js";import"../chunk-7V2JKPIG.js";import{c as C}from"../chunk-YSFSRI3D.js";C();import{createAssistantMessageEventStream as he}from"@earendil-works/pi-ai";import{mkdirSync as we,openSync as _e,writeSync as ke}from"fs";import{dirname as ve}from"path";C();import{appendFileSync as de,mkdirSync as pe}from"fs";import fe from"os";import{dirname as ge,join as me}from"path";import{format as ye}from"util";function J(t=fe.homedir()){return me(t,".urun","logs","pi-extension.log")}var V=!1;function te(t,e){let n=J();try{pe(ge(n),{recursive:!0}),de(n,`${new Date().toISOString()} [console.${t}] [urun] ${ye(...e)}
+import{a as z}from"../chunk-76BZLCGT.js";import{e as W,h as Q}from"../chunk-BAKPG5S3.js";import"../chunk-7V2JKPIG.js";import{c as C}from"../chunk-YSFSRI3D.js";C();import{createAssistantMessageEventStream as he}from"@earendil-works/pi-ai";import{mkdirSync as we,openSync as _e,writeSync as ke}from"fs";import{dirname as ve}from"path";C();import{appendFileSync as de,mkdirSync as pe}from"fs";import fe from"os";import{dirname as ge,join as me}from"path";import{format as ye}from"util";function J(t=fe.homedir()){return me(t,".urun","logs","pi-extension.log")}var V=!1;function te(t,e){let n=J();try{pe(ge(n),{recursive:!0}),de(n,`${new Date().toISOString()} [console.${t}] [urun] ${ye(...e)}
 `)}catch(o){if(!V){V=!0;try{process.stderr.write(`[urun] pi-extension quiet console: cannot write ${n} (${String(o)}); further urun-side console output will be dropped
 `)}catch{}}}}var f=t=>(...e)=>te(t,e),j={log:f("log"),info:f("info"),warn:f("warn"),error:f("error"),debug:f("debug"),trace:f("trace"),dir:f("dir"),table:f("table"),group:f("group"),groupCollapsed:f("groupCollapsed"),groupEnd:f("groupEnd"),count:f("count"),countReset:f("countReset"),time:f("time"),timeLog:f("timeLog"),timeEnd:f("timeEnd"),assert:(t,...e)=>{t||te("assert",e.length>0?e:["Assertion failed"])}};function q(t){try{process.stderr.write(t)}catch{}}C();async function ne(t){let e=await Q(t),n=e.filter(o=>o.function_name===t.fnName&&o.deployment_status==="active").map(o=>o.app_slug);if(n.length===0)throw new Error(`urun pi extension: the org has no active deployed app exposing "${t.fnName}" (GET ${t.apiUrl}/apps returned ${e.length} app(s), none servable) \u2014 deploy one with \`urun serve <model>\` first`);return n.sort()}var ae="urun-serve",xe="serve",Re=131072,be=16384,oe=3,Te=42e4,Ee=6e4,re=2,Se=15e3;function Ae(t){let e=(t.URUN_API_KEY??"").trim();if(!e)throw new Error("urun pi extension: URUN_API_KEY is required \u2014 model discovery lists the org's deployed serve apps via the org API (URUN_JWT alone cannot list apps)");return{apiKey:e,apiUrl:(t.URUN_API_URL??"").trim()||W,fnName:(t.URUN_FUNCTION??"").trim()||xe}}function Ue(t){let e=(t.URUN_BASE_URL??"").trim(),n=(t.URUN_ORG_ID??"").trim();if(!e)throw new Error("urun pi extension: URUN_BASE_URL is required to open a session");if(/\/v1\/*$/.test(new URL(e).pathname))throw new Error(`urun pi extension: URUN_BASE_URL must be the session-gateway base (e.g. https://api.urun.sh), got ${e} \u2014 a trailing /v1 is the org API form (URUN_API_URL); session allocation 404s against it. Drop the /v1.`);if(!n)throw new Error("urun pi extension: URUN_ORG_ID is required to open a session");let o=(t.URUN_JWT??"").trim();if(o)return{baseUrl:e,orgId:n,auth:{lane:"jwt",jwt:o}};let s=(t.URUN_API_KEY??"").trim();if(!s)throw new Error("urun pi extension: URUN_JWT or URUN_API_KEY is required to open a session");let l=(t.URUN_GATEWAY_URL??"").trim()||void 0;return{baseUrl:e,orgId:n,auth:{lane:"api-key",apiKey:s,gatewayUrl:l}}}function Ie(t){let e=null;return n=>{e===null&&(we(ve(t),{recursive:!0}),e=_e(t,"a"));let o=n.detail?` ${JSON.stringify(n.detail)}`:"";ke(e,`${new Date().toISOString()} [${n.level}] [urun] ${n.message}${o}
 `)}}function Pe(t){let e=Ie(J());return async(n,o,s)=>{let{baseUrl:l,orgId:i,auth:p}=Ue(n),{App:k,createClientToken:a}=await t(),m=(p.lane==="jwt"?k(o,{baseUrl:l,orgId:i,jwt:p.jwt,diagnosticSink:e}):k(o,{baseUrl:l,orgId:i,diagnosticSink:e,getAccessToken:async()=>(await a(p.apiKey,{baseUrl:p.gatewayUrl,expiresIn:300,allowedFunctions:[`${o}/${s}`]})).token}))[s];if(typeof m!="function")throw new Error(`urun pi extension: app "${o}" has no function "${s}"`);let h=m();if(typeof h.end!="function")throw new Error(`urun pi extension: app "${o}" function "${s}" returned a session without end()`);let A=h.connect;return typeof A=="function"&&await A.call(h),h}}var Ce=Pe(async()=>{let{createRequire:t}=await import("module"),e=t(import.meta.url);if(typeof globalThis.RTCPeerConnection>"u")try{e.resolve("werift")}catch{throw new Error('urun pi extension: token streams require a WebRTC data channel, and this Node runtime has no RTCPeerConnection and no "werift" backend installed (@urun-sh/core\'s optional Node WebRTC backend \u2014 it was probably stripped by a --no-optional install). Install werift (`npm i werift`) and retry.')}return e("@urun-sh/core")});function Ne(t,e){try{return Promise.resolve(t.end()).then(()=>{},n=>e(n))}catch(n){return e(n),Promise.resolve()}}function $e(t){if(t?.hasUI===!1)return!0;if(t?.hasUI===!0)return!1;let e=process.argv.slice(2);if(e.includes("-p")||e.includes("--print"))return!0;let n=e.indexOf("--mode");return n!==-1&&(e[n+1]==="json"||e[n+1]==="rpc")?!0:!process.stdin.isTTY||!process.stdout.isTTY}function le(t,e){return`uRun: releasing the pooled "${t}" session failed \u2014 the backend session may stay allocated until the platform reclaims it: ${e instanceof Error?e.message:String(e)}`}var H=class{constructor(e,n,o,s=(l,i)=>j.warn(le(l,i))){this.open=e;this.env=n;this.fnName=o;this.onReleaseError=s}open;env;fnName;onReleaseError;pool=new Map;releasing=new Set;acquire(e){let n=this.pool.get(e);if(!n){let o=Promise.resolve(this.open(this.env,e,this.fnName)).then(s=>({session:s,responses:new z(s)}));n=o,this.pool.set(e,o),o.catch(()=>{this.pool.get(e)===o&&this.pool.delete(e)})}return n}evict(e){let n=this.pool.get(e);if(!n)return Promise.resolve();this.pool.delete(e);let o=n.then(s=>Ne(s.session,l=>this.onReleaseError(e,l)),()=>{});return this.releasing.add(o),o.then(()=>this.releasing.delete(o)),o}closeAll(){let e=[...this.pool.keys()].map(n=>this.evict(n));return Promise.all([...e,...this.releasing]).then(()=>{})}};function Le(t){let e=[];t.systemPrompt&&e.push({role:"system",content:t.systemPrompt});for(let n of t.messages){if(n.role==="toolResult"){e.push({type:"function_call_output",call_id:n.toolCallId,output:De(n.content)});continue}if(n.role==="assistant"&&Array.isArray(n.content)){let s="",l=()=>{s&&(e.push({role:"assistant",content:s}),s="")};for(let i of n.content)if(i.type==="text")s+=String(i.text??"");else if(i.type==="toolCall")l(),e.push({type:"function_call",call_id:String(i.id??""),name:String(i.name??""),arguments:JSON.stringify(i.arguments??{})});else{if(i.type==="thinking")continue;throw new Error(`urun pi extension: unsupported assistant content part "${String(i.type)}"`)}l();continue}let o=n.role==="assistant"?"assistant":"user";e.push({role:o,content:Me(n.content)})}return Oe(e)}function Oe(t){let e=[];for(let n of t){let o=e[e.length-1];o&&"role"in o&&"role"in n&&o.role===n.role?o.content=[o.content,n.content].filter(s=>s.length>0).join(`