inference-relay 1.5.2 → 1.6.0
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/cli/main.js +7 -7
- package/dist/index.cjs +5 -5
- package/dist/index.js +5 -5
- package/package.json +2 -2
package/dist/cli/main.js
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
#!/usr/bin/env node
|
|
2
2
|
import {existsSync,mkdirSync,writeFileSync,readFileSync}from'fs';import {homedir}from'os';import {join,dirname}from'path';import {randomUUID,createHash}from'crypto';import {compileFunction}from'vm';import {createRequire}from'module';import {execSync}from'child_process';import {EventEmitter}from'events';/*! inference-relay CLI (c) 2026 L2B II LLC. Patent pending. */
|
|
3
|
-
var
|
|
3
|
+
var Ct=Object.defineProperty;var C=(n,e)=>()=>(n&&(e=n(n=0)),e);var De=(n,e)=>{for(var t in e)Ct(n,t,{get:e[t],enumerable:true});};var N,ge=C(()=>{N=class n extends Error{code="RELAY_CAP_EXCEEDED";tier;baseCap;topUps;effectiveCap;usedThisCycle;resolutionUrl;constructor(e){let t=e.resolutionUrl??"https://inference-relay.com/dashboard/billing?topup=1";super(`Cap reached (${e.usedThisCycle} / ${e.effectiveCap??"\u221E"} calls this cycle). Buy a top-up or upgrade at ${t}`),this.name="RelayCapExceededError",this.tier=e.tier,this.baseCap=e.baseCap,this.topUps=e.topUps,this.effectiveCap=e.effectiveCap,this.usedThisCycle=e.usedThisCycle,this.resolutionUrl=t,Object.setPrototypeOf(this,n.prototype);}};});function z(n,e){let t=new Map,r=null,s=null,o=new Promise((u,y)=>{r=u,s=y;});o.catch(()=>{});let i=[],a=null,l="",p="";function c(u,...y){let h=t.get(u);h&&h.forEach(g=>g(...y));}async function*d(){try{for await(let u of n)if(yield u,u.type==="content_block_start")i[u.index]=u.content_block;else if(u.type==="content_block_delta"){let y=i[u.index];y&&u.delta?.type==="text_delta"&&(y.text=(y.text||"")+u.delta.text,c("text",u.delta.text));}else if(u.type==="message_start")l=u.message?.model||"",a=u.message?.usage;else if(u.type==="message_delta")p=u.delta?.stop_reason||"",u.usage&&(a={...a,...u.usage});else if(u.type==="message_stop"){let y={role:"assistant",model:l,content:i.filter(Boolean),usage:a,stop_reason:p};c("message",y),r&&(r(y),r=null),e&&e(y);}if(r){let u={role:"assistant",model:l,content:i.filter(Boolean),usage:a,stop_reason:p||"end_turn"};c("message",u),r(u),r=null,e&&e(u);}}catch(u){throw c("error",u),s&&(s(u),s=null),u}}let f=d(),m=false,b={[Symbol.asyncIterator](){if(m)throw new Error("[inference-relay] RelayStream can only be iterated once");return m=true,f},on(u,y){return t.has(u)||t.set(u,[]),t.get(u).push(y),b},finalMessage(){let u,y=new Promise((h,g)=>{u=setTimeout(()=>g(new Error("[inference-relay] finalMessage() timed out \u2014 no message_stop received after 5 minutes")),3e5);});return Promise.race([o,y]).finally(()=>clearTimeout(u))}};return b}var fe=C(()=>{});var S,V=C(()=>{fe();S=class{priority=100;async streamRaw(e,t){throw new Error(`[inference-relay] Provider '${this.type}' does not implement streamRaw(). Override either stream() or streamRaw() in the provider class.`)}async stream(e,t){return z(await this.streamRaw(e,t))}};});function Oe(){At.clearCache();}var ye,Me,G,At,_e=C(()=>{ye=`-----BEGIN PUBLIC KEY-----
|
|
4
4
|
MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEAzk88Bt9LZSmuSd8ulKqy
|
|
5
5
|
EGNiPB+r8u1ZinBvMM4qX8dp4kzJEC/47L1RYrrxCphD05HToREjljbj+upH6rZH
|
|
6
6
|
t6D/5funYPi2TO1IOSC6BRVRwMqqTmJGVJcLcG/WF3gTE+2UQ3q5EppWJCEKcz6x
|
|
@@ -8,12 +8,12 @@ kfynZEzOcLikoQALIxEjxRfi6g83wNJS+WBwAJQ68s7N9FHE3rkLzEqv3FJRP+ZR
|
|
|
8
8
|
7lqwqIyuCLr9j9yc4XqF6VY9yooD2hyw/n1EVFejzjDgntWTqIkaCIXwh8lDzRrK
|
|
9
9
|
Rxn07gcI6ea2Eym9sXh4xZLFudXfAT2DFsH/XboDUuMKAV1YHc1NsIGApesWe2M+
|
|
10
10
|
UQIDAQAB
|
|
11
|
-
-----END PUBLIC KEY-----`,Le={version:"1.0.0",cliVersion:"2.1.87",eventTypes:{system:"system",streamEvent:"stream_event",assistant:"assistant",result:"result"},fieldMappings:{sessionId:"session_id",model:"model",cliVersion:"claude_code_version",content:"content",usage:"usage",inputTokens:"input_tokens",outputTokens:"output_tokens",stopReason:"stop_reason",isError:"is_error",totalCostUsd:"total_cost_usd",durationMs:"duration_ms",error:"error",event:"event"},syntheticModel:"<synthetic>"},V=class{verificationFailures=0;cachedManifest=null;currentSecurityState="SEC_NOMINAL";async verifyManifest(e){try{let{compactVerify:t,importSPKI:r}=await import('jose'),s=await r(ge,"RS256"),{payload:a}=await t(e,s),i=JSON.parse(new TextDecoder().decode(a));return this.verificationFailures=0,this.cachedManifest=i,this.currentSecurityState="SEC_NOMINAL",{manifest:i,securityState:"SEC_NOMINAL"}}catch{if(this.verificationFailures++,this.verificationFailures>=3){this.currentSecurityState="SEC_DEGRADED";let t=2e3+Math.random()*3e3;return await new Promise(r=>setTimeout(r,t)),console.warn("[inference-relay] Logic synchronization mismatch. Routing via local recovery."),{manifest:Le,securityState:"SEC_DEGRADED"}}return this.cachedManifest?(this.currentSecurityState="SEC_CACHE",{manifest:this.cachedManifest,securityState:"SEC_CACHE"}):(this.currentSecurityState="SEC_RECOVERY",{manifest:Le,securityState:"SEC_RECOVERY"})}}getSecurityState(){return this.currentSecurityState}clearCache(){this.cachedManifest=null,this.verificationFailures=0,this.currentSecurityState="SEC_NOMINAL";}},Ct=new V;});function Ee(n){return join(homedir(),".inference-relay","gateway",n)}function Ue(n){return join(Ee(n),"bundle.js")}function Ne(n){return join(Ee(n),"meta.json")}var wt,_,W,z,be=C(()=>{fe();wt=process.env.IR_GATEWAY_URL||"https://api.inference-relay.com";_=class extends Error{constructor(t,r,s){super(t);this.code=r;this.cause=s;this.name="GatewayLoaderError";}code;cause},W=new Map,z=class{static async load(e,t,r=wt){let s=`${t}:${e.slice(-8)}`;if(W.has(s))return W.get(s);try{let d=await this.tryLoadFromDisk(t);if(d)return W.set(s,d),d}catch{}let{bytes:a,sha256:i,signature:l,libraryVersion:o}=await this.fetchFromAuthority(e,t,r);if(createHash("sha256").update(a).digest("hex")!==i)throw new _("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: HASH_MISMATCH)","GATEWAY_VERIFICATION_FAILED");await this.verifySignature(l,o,i);try{let d=Ee(t);mkdirSync(d,{recursive:!0}),writeFileSync(Ue(t),a),writeFileSync(Ne(t),JSON.stringify({sha256:i,signature:l,libraryVersion:o}));}catch{}let c=this.loadIntoVm(a.toString("utf-8"));return W.set(s,c),c}static async loadFromDisk(e){if(!existsSync(e))throw new _(`[ir-system]: GATEWAY_LOAD_FAILED (REASON: BUNDLE_NOT_FOUND at ${e})`,"GATEWAY_LOAD_FAILED");let t=readFileSync(e,"utf-8");return this.loadIntoVm(t)}static async tryLoadFromDisk(e){let t=Ue(e),r=Ne(e);if(!existsSync(t)||!existsSync(r))return null;let s=readFileSync(t),a=JSON.parse(readFileSync(r,"utf-8"));return createHash("sha256").update(s).digest("hex")!==a.sha256?null:(await this.verifySignature(a.signature,a.libraryVersion,a.sha256),this.loadIntoVm(s.toString("utf-8")))}static async fetchFromAuthority(e,t,r){let s,a=[0,1e3,4e3],i;for(let c of a){c>0&&await new Promise(d=>setTimeout(d,c));try{let d=await fetch(`${r}/v1/gateway/fetch`,{method:"POST",headers:{"Content-Type":"application/json",Authorization:`Bearer ${e}`},body:JSON.stringify({libraryVersion:t})});if(d.status===401||d.status===403)throw new _("[ir-system]: LICENSE_REJECTED (REASON: GATEWAY_FETCH_UNAUTHORIZED)","LICENSE_REJECTED");if(d.status>=500&&d.status<600){i=new _(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: HTTP_${d.status})`,"GATEWAY_FETCH_FAILED");continue}if(!d.ok)throw new _(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: HTTP_${d.status})`,"GATEWAY_FETCH_FAILED");s=await d.json(),i=null;break}catch(d){if(d instanceof _&&d.code==="LICENSE_REJECTED")throw d;i=d;}}if(i||!s)throw i instanceof _?i:new _("[ir-system]: GATEWAY_FETCH_FAILED (REASON: NETWORK_ERROR)","GATEWAY_FETCH_FAILED",i);let l;try{l=await fetch(s.bundleUrl);}catch(c){throw new _("[ir-system]: GATEWAY_FETCH_FAILED (REASON: BUNDLE_DOWNLOAD_FAILED)","GATEWAY_FETCH_FAILED",c)}if(!l.ok)throw new _(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: BUNDLE_HTTP_${l.status})`,"GATEWAY_FETCH_FAILED");let o=await l.arrayBuffer();return {bytes:Buffer.from(o),sha256:s.sha256,signature:s.signature,libraryVersion:s.libraryVersion,bundleVersion:s.bundleVersion}}static async verifySignature(e,t,r){try{let{compactVerify:s,importSPKI:a}=await import('jose'),i=await a(ge,"RS256"),{payload:l}=await s(e,i),o=JSON.parse(new TextDecoder().decode(l));if(o.purpose!=="gateway-bundle")throw new _("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: PURPOSE_MISMATCH)","GATEWAY_VERIFICATION_FAILED");if(o.libraryVersion!==t)throw new _("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: LIBRARY_VERSION_MISMATCH)","GATEWAY_VERIFICATION_FAILED");if(o.sha256!==r)throw new _("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: SHA256_MISMATCH)","GATEWAY_VERIFICATION_FAILED")}catch(s){throw s instanceof _?s:new _("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: SIGNATURE_INVALID)","GATEWAY_VERIFICATION_FAILED",s)}}static loadIntoVm(e){try{let t=join(homedir(),".inference-relay","gateway","runtime.cjs"),r=dirname(t),s=createRequire(t),a={exports:{}};compileFunction(e,["exports","require","module","__filename","__dirname"],{filename:"<gateway-bundle>"})(a.exports,s,a,t,r);let l=a.exports;if(!l.ClaudeCLIProvider)throw new _("[ir-system]: GATEWAY_LOAD_FAILED (REASON: MISSING_CLAUDE_CLI_EXPORT)","GATEWAY_LOAD_FAILED");return l}catch(t){throw t instanceof _?t:new _("[ir-system]: GATEWAY_LOAD_FAILED (REASON: VM_EXECUTION_ERROR)","GATEWAY_LOAD_FAILED",t)}}static clearCache(){W.clear();}};});var He={};xe(He,{detectCLI:()=>N});function Lt(n){let e=process.env.PATH;if(!e)return null;let t=process.platform==="win32",r=t?";":":",s=t?(process.env.PATHEXT??".COM;.EXE;.BAT;.CMD").split(";").map(a=>a.toLowerCase()):[""];for(let a of e.split(r))if(a)for(let i of s){let l=join(a,n+i);if(existsSync(l))return l}return null}async function N(){if(!Lt("claude"))return {available:false,authenticated:false};let e=null;try{let r=execSync("claude --version",{encoding:"utf-8",timeout:5e3}).trim(),s=r.match(/(\d+\.\d+\.\d+)/);e=s?s[1]:r;}catch{}let t;return e&&!xt.some(r=>e.startsWith(r.split(".").slice(0,2).join(".")))&&(t=`CLI version ${e} is not tested with this version of inference-relay. Streaming may be unstable.`),{available:true,authenticated:true,cliVersion:e||void 0,versionWarning:t}}var xt,Q=C(()=>{xt=["2.1.87"];});var Mt,Z,Be=C(()=>{$();be();Mt="1.5.2",Z=class extends w{type="claude-cli";supportsStreaming=true;supportsMultimodal=true;platformRequirement="desktop";licenseKey;loadedProvider=null;loadingPromise=null;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.licenseKey=e;}async ensureLoaded(){return this.loadedProvider?this.loadedProvider:this.loadingPromise?this.loadingPromise:(this.loadingPromise=(async()=>{try{let t=(await z.load(this.licenseKey,Mt)).ClaudeCLIProvider,r=new t;return this.loadedProvider=r,r}catch(e){throw this.loadingPromise=null,this.health.status="unhealthy",this.health.errorCount++,e instanceof _?this.health.lastError=e.message:e instanceof Error&&(this.health.lastError=e.message),e}})(),this.loadingPromise)}async probe(){let{detectCLI:e}=await Promise.resolve().then(()=>(Q(),He));return e()}async execute(e,t){return (await this.ensureLoaded()).execute(e,t)}async streamRaw(e,t){return (await this.ensureLoaded()).streamRaw(e,t)}estimateCost(e){if(this.loadedProvider)return this.loadedProvider.estimateCost(e);let r=JSON.stringify(e.messages).length/4+5e3,s=(r*3+1e3*15)/1e6;return {estimatedInputTokens:Math.round(r),estimatedCost:Math.round(s*1e4)/1e4,model:e.model,provider:"claude-cli"}}getHealth(){return this.loadedProvider?this.loadedProvider.getHealth():{...this.health}}reportRateLimit(e){this.loadedProvider?this.loadedProvider.reportRateLimit(e):(this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now());}};});function Ve(n){return ee[n]||n}function ve(n){return $e.find(e=>e.canonical===n)?.provider}var $e,ee,Ce=C(()=>{$e=[{canonical:"claude-haiku-4-5",apiId:"claude-haiku-4-5-20251001",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[.8,4]},{canonical:"claude-sonnet-4-6",apiId:"claude-sonnet-4-20250514",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[3,15]},{canonical:"claude-opus-4-6",apiId:"claude-opus-4-20250514",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[15,75]},{canonical:"gpt-4.1",apiId:"gpt-4.1",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[2,8]},{canonical:"gpt-4.1-mini",apiId:"gpt-4.1-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[.4,1.6]},{canonical:"gpt-4.1-nano",apiId:"gpt-4.1-nano",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[.1,.4]},{canonical:"gpt-4o",apiId:"gpt-4o",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:128e3},pricing:[2.5,10]},{canonical:"gpt-4o-mini",apiId:"gpt-4o-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:128e3},pricing:[.15,.6]},{canonical:"o4-mini",apiId:"o4-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[1.1,4.4]},{canonical:"o3",apiId:"o3",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[.4,1.6]},{canonical:"o3-mini",apiId:"o3-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[1.1,4.4]},{canonical:"o1",apiId:"o1",provider:"openai-api",capabilities:{streaming:false,tools:false,multimodal:false,contextWindow:2e5},pricing:[15,60]},{canonical:"llama3.1",apiId:"llama3.1",provider:"ollama",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:128e3}},{canonical:"mistral",apiId:"mistral",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:false,contextWindow:32e3}},{canonical:"mistral-nemo",apiId:"mistral-nemo",provider:"ollama",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:128e3}},{canonical:"llava",apiId:"llava",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:true,contextWindow:32e3}},{canonical:"qwen2:0.5b",apiId:"qwen2:0.5b",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:false,contextWindow:32e3}}],ee=Object.fromEntries($e.filter(n=>n.canonical!==n.apiId).map(n=>[n.canonical,n.apiId]));ee["claude-sonnet-4"]="claude-sonnet-4-20250514";ee["claude-opus-4"]="claude-opus-4-20250514";ee["claude-haiku-4"]="claude-haiku-4-5-20251001";});function We(n){if(Ae[n])return Ae[n];for(let[e,t]of Object.entries(Ae))if(n.startsWith(e.split("-20")[0]))return t;return [3,15]}var Dt,Ot,Ae,te,Ge=C(()=>{$();Ce();Dt="https://api.anthropic.com/v1/messages",Ot="2023-06-01",Ae={"claude-sonnet-4-5-20250514":[3,15],"claude-sonnet-4-20250514":[3,15],"claude-haiku-4-5-20250414":[.8,4],"claude-opus-4-20250514":[15,75]};te=class extends w{type="anthropic-api";supportsStreaming=true;supportsMultimodal=true;platformRequirement="any";apiKey;baseUrl;dangerousDirectBrowserAccess;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.apiKey=e.apiKey,this.baseUrl=e.baseUrl?.replace(/\/$/,"")||Dt,this.dangerousDirectBrowserAccess=e.dangerousDirectBrowserAccess??false;}async probe(){try{let e=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify({model:"claude-haiku-4-5-20250414",max_tokens:1,messages:[{role:"user",content:"hi"}]})});return e.status===401?{available:!0,authenticated:!1}:e.ok||e.status===429?{available:!0,authenticated:!0,rateLimitTier:e.headers.get("anthropic-ratelimit-tier")??void 0}:e.status===529?{available:!0,authenticated:!0}:{available:!1,authenticated:!1}}catch{return {available:false,authenticated:false}}}async execute(e,t){let r=Date.now(),s=this.buildRequestBody(e,false),a=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(s),signal:t?.signal});if(!a.ok)throw this.handleErrorResponse(a),new Error("unreachable");let i=await a.json(),l=Date.now()-r;this.recordSuccess();let[o,p]=We(i.model||e.model),c=((i.usage?.input_tokens||0)*o+(i.usage?.output_tokens||0)*p)/1e6,d={},h=["request-id","x-request-id","anthropic-ratelimit-requests-limit","anthropic-ratelimit-requests-remaining","anthropic-ratelimit-requests-reset","anthropic-ratelimit-tokens-limit","anthropic-ratelimit-tokens-remaining","anthropic-ratelimit-tokens-reset","anthropic-ratelimit-tier"];for(let m of h){let E=a.headers.get(m);E&&(d[m]=E);}return {content:i.content||[],model:i.model||e.model,usage:{input_tokens:i.usage?.input_tokens||0,output_tokens:i.usage?.output_tokens||0},stop_reason:i.stop_reason||"end_turn",provider:"anthropic-api",costUsd:Math.round(c*1e6)/1e6,durationMs:l,fallbackChain:["anthropic-api:success"],headers:Object.keys(d).length>0?d:void 0}}async streamRaw(e,t){let r=this.buildRequestBody(e,true),s=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(r),signal:t?.signal});if(s.ok||this.handleErrorResponse(s),!s.body)throw new Error("ANTHROPIC_API_ERROR: Response body is null \u2014 streaming unavailable");let a=s.body.getReader(),i=new TextDecoder;return (async function*(){let l="",o=false;try{for(;;){let{done:p,value:c}=await a.read();if(p)break;l+=i.decode(c,{stream:!0});let d=l.split(`
|
|
12
|
-
`);
|
|
13
|
-
`);
|
|
14
|
-
`);
|
|
11
|
+
-----END PUBLIC KEY-----`,Me={version:"1.0.0",cliVersion:"2.1.87",eventTypes:{system:"system",streamEvent:"stream_event",assistant:"assistant",result:"result"},fieldMappings:{sessionId:"session_id",model:"model",cliVersion:"claude_code_version",content:"content",usage:"usage",inputTokens:"input_tokens",outputTokens:"output_tokens",stopReason:"stop_reason",isError:"is_error",totalCostUsd:"total_cost_usd",durationMs:"duration_ms",error:"error",event:"event"},syntheticModel:"<synthetic>"},G=class{verificationFailures=0;cachedManifest=null;currentSecurityState="SEC_NOMINAL";async verifyManifest(e){try{let{compactVerify:t,importSPKI:r}=await import('jose'),s=await r(ye,"RS256"),{payload:o}=await t(e,s),i=JSON.parse(new TextDecoder().decode(o));return this.verificationFailures=0,this.cachedManifest=i,this.currentSecurityState="SEC_NOMINAL",{manifest:i,securityState:"SEC_NOMINAL"}}catch{if(this.verificationFailures++,this.verificationFailures>=3){this.currentSecurityState="SEC_DEGRADED";let t=2e3+Math.random()*3e3;return await new Promise(r=>setTimeout(r,t)),console.warn("[inference-relay] Logic synchronization mismatch. Routing via local recovery."),{manifest:Me,securityState:"SEC_DEGRADED"}}return this.cachedManifest?(this.currentSecurityState="SEC_CACHE",{manifest:this.cachedManifest,securityState:"SEC_CACHE"}):(this.currentSecurityState="SEC_RECOVERY",{manifest:Me,securityState:"SEC_RECOVERY"})}}getSecurityState(){return this.currentSecurityState}clearCache(){this.cachedManifest=null,this.verificationFailures=0,this.currentSecurityState="SEC_NOMINAL";}},At=new G;});function ve(n){return join(homedir(),".inference-relay","gateway",n)}function Fe(n){return join(ve(n),"bundle.js")}function He(n){return join(ve(n),"meta.json")}var wt,v,Y,X,Ce=C(()=>{_e();wt=process.env.IR_GATEWAY_URL||"https://api.inference-relay.com";v=class extends Error{constructor(t,r,s){super(t);this.code=r;this.cause=s;this.name="GatewayLoaderError";}code;cause},Y=new Map,X=class{static async load(e,t,r=wt){let s=`${t}:${e.slice(-8)}`;if(Y.has(s))return Y.get(s);try{let d=await this.tryLoadFromDisk(t);if(d)return Y.set(s,d),d}catch{}let{bytes:o,sha256:i,signature:a,libraryVersion:l}=await this.fetchFromAuthority(e,t,r);if(createHash("sha256").update(o).digest("hex")!==i)throw new v("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: HASH_MISMATCH)","GATEWAY_VERIFICATION_FAILED");await this.verifySignature(a,l,i);try{let d=ve(t);mkdirSync(d,{recursive:!0}),writeFileSync(Fe(t),o),writeFileSync(He(t),JSON.stringify({sha256:i,signature:a,libraryVersion:l}));}catch{}let c=this.loadIntoVm(o.toString("utf-8"));return Y.set(s,c),c}static async loadFromDisk(e){if(!existsSync(e))throw new v(`[ir-system]: GATEWAY_LOAD_FAILED (REASON: BUNDLE_NOT_FOUND at ${e})`,"GATEWAY_LOAD_FAILED");let t=readFileSync(e,"utf-8");return this.loadIntoVm(t)}static async tryLoadFromDisk(e){let t=Fe(e),r=He(e);if(!existsSync(t)||!existsSync(r))return null;let s=readFileSync(t),o=JSON.parse(readFileSync(r,"utf-8"));return createHash("sha256").update(s).digest("hex")!==o.sha256?null:(await this.verifySignature(o.signature,o.libraryVersion,o.sha256),this.loadIntoVm(s.toString("utf-8")))}static async fetchFromAuthority(e,t,r){let s,o=[0,1e3,4e3],i;for(let c of o){c>0&&await new Promise(d=>setTimeout(d,c));try{let d=await fetch(`${r}/v1/gateway/fetch`,{method:"POST",headers:{"Content-Type":"application/json",Authorization:`Bearer ${e}`},body:JSON.stringify({libraryVersion:t})});if(d.status===401||d.status===403)throw new v("[ir-system]: LICENSE_REJECTED (REASON: GATEWAY_FETCH_UNAUTHORIZED)","LICENSE_REJECTED");if(d.status>=500&&d.status<600){i=new v(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: HTTP_${d.status})`,"GATEWAY_FETCH_FAILED");continue}if(!d.ok)throw new v(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: HTTP_${d.status})`,"GATEWAY_FETCH_FAILED");s=await d.json(),i=null;break}catch(d){if(d instanceof v&&d.code==="LICENSE_REJECTED")throw d;i=d;}}if(i||!s)throw i instanceof v?i:new v("[ir-system]: GATEWAY_FETCH_FAILED (REASON: NETWORK_ERROR)","GATEWAY_FETCH_FAILED",i);let a;try{a=await fetch(s.bundleUrl);}catch(c){throw new v("[ir-system]: GATEWAY_FETCH_FAILED (REASON: BUNDLE_DOWNLOAD_FAILED)","GATEWAY_FETCH_FAILED",c)}if(!a.ok)throw new v(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: BUNDLE_HTTP_${a.status})`,"GATEWAY_FETCH_FAILED");let l=await a.arrayBuffer();return {bytes:Buffer.from(l),sha256:s.sha256,signature:s.signature,libraryVersion:s.libraryVersion,bundleVersion:s.bundleVersion}}static async verifySignature(e,t,r){try{let{compactVerify:s,importSPKI:o}=await import('jose'),i=await o(ye,"RS256"),{payload:a}=await s(e,i),l=JSON.parse(new TextDecoder().decode(a));if(l.purpose!=="gateway-bundle")throw new v("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: PURPOSE_MISMATCH)","GATEWAY_VERIFICATION_FAILED");if(l.libraryVersion!==t)throw new v("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: LIBRARY_VERSION_MISMATCH)","GATEWAY_VERIFICATION_FAILED");if(l.sha256!==r)throw new v("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: SHA256_MISMATCH)","GATEWAY_VERIFICATION_FAILED")}catch(s){throw s instanceof v?s:new v("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: SIGNATURE_INVALID)","GATEWAY_VERIFICATION_FAILED",s)}}static loadIntoVm(e){try{let t=join(homedir(),".inference-relay","gateway","runtime.cjs"),r=dirname(t),s=createRequire(t),o={exports:{}};compileFunction(e,["exports","require","module","__filename","__dirname"],{filename:"<gateway-bundle>"})(o.exports,s,o,t,r);let a=o.exports;if(!a.ClaudeCLIProvider)throw new v("[ir-system]: GATEWAY_LOAD_FAILED (REASON: MISSING_CLAUDE_CLI_EXPORT)","GATEWAY_LOAD_FAILED");return a}catch(t){throw t instanceof v?t:new v("[ir-system]: GATEWAY_LOAD_FAILED (REASON: VM_EXECUTION_ERROR)","GATEWAY_LOAD_FAILED",t)}}static clearCache(){Y.clear();}};});var $e={};De($e,{detectCLI:()=>F});function Dt(n){let e=process.env.PATH;if(!e)return null;let t=process.platform==="win32",r=t?";":":",s=t?(process.env.PATHEXT??".COM;.EXE;.BAT;.CMD").split(";").map(o=>o.toLowerCase()):[""];for(let o of e.split(r))if(o)for(let i of s){let a=join(o,n+i);if(existsSync(a))return a}return null}async function F(){if(!Dt("claude"))return {available:false,authenticated:false};let e=null;try{let r=execSync("claude --version",{encoding:"utf-8",timeout:5e3}).trim(),s=r.match(/(\d+\.\d+\.\d+)/);e=s?s[1]:r;}catch{}let t;return e&&!xt.some(r=>e.startsWith(r.split(".").slice(0,2).join(".")))&&(t=`CLI version ${e} is not tested with this version of inference-relay. Streaming may be unstable.`),{available:true,authenticated:true,cliVersion:e||void 0,versionWarning:t}}var xt,Z=C(()=>{xt=["2.1.87"];});var Mt,ee,Ve=C(()=>{V();Ce();Mt="1.6.0",ee=class extends S{type="claude-cli";supportsStreaming=true;supportsMultimodal=true;platformRequirement="desktop";licenseKey;loadedProvider=null;loadingPromise=null;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.licenseKey=e;}async ensureLoaded(){return this.loadedProvider?this.loadedProvider:this.loadingPromise?this.loadingPromise:(this.loadingPromise=(async()=>{try{let t=(await X.load(this.licenseKey,Mt)).ClaudeCLIProvider,r=new t;return this.loadedProvider=r,r}catch(e){throw this.loadingPromise=null,this.health.status="unhealthy",this.health.errorCount++,e instanceof v?this.health.lastError=e.message:e instanceof Error&&(this.health.lastError=e.message),e}})(),this.loadingPromise)}async probe(){let{detectCLI:e}=await Promise.resolve().then(()=>(Z(),$e));return e()}async execute(e,t){return (await this.ensureLoaded()).execute(e,t)}async streamRaw(e,t){return (await this.ensureLoaded()).streamRaw(e,t)}estimateCost(e){if(this.loadedProvider)return this.loadedProvider.estimateCost(e);let r=JSON.stringify(e.messages).length/4+5e3,s=(r*3+1e3*15)/1e6;return {estimatedInputTokens:Math.round(r),estimatedCost:Math.round(s*1e4)/1e4,model:e.model,provider:"claude-cli"}}getHealth(){return this.loadedProvider?this.loadedProvider.getHealth():{...this.health}}reportRateLimit(e){this.loadedProvider?this.loadedProvider.reportRateLimit(e):(this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now());}};});function Ye(n){return te[n]||n}function Ae(n){return Ge.find(e=>e.canonical===n)?.provider}var Ge,te,Re=C(()=>{Ge=[{canonical:"claude-haiku-4-5",apiId:"claude-haiku-4-5-20251001",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[.8,4]},{canonical:"claude-sonnet-4-6",apiId:"claude-sonnet-4-20250514",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[3,15]},{canonical:"claude-opus-4-6",apiId:"claude-opus-4-20250514",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[15,75]},{canonical:"gpt-4.1",apiId:"gpt-4.1",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[2,8]},{canonical:"gpt-4.1-mini",apiId:"gpt-4.1-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[.4,1.6]},{canonical:"gpt-4.1-nano",apiId:"gpt-4.1-nano",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[.1,.4]},{canonical:"gpt-4o",apiId:"gpt-4o",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:128e3},pricing:[2.5,10]},{canonical:"gpt-4o-mini",apiId:"gpt-4o-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:128e3},pricing:[.15,.6]},{canonical:"o4-mini",apiId:"o4-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[1.1,4.4]},{canonical:"o3",apiId:"o3",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[.4,1.6]},{canonical:"o3-mini",apiId:"o3-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[1.1,4.4]},{canonical:"o1",apiId:"o1",provider:"openai-api",capabilities:{streaming:false,tools:false,multimodal:false,contextWindow:2e5},pricing:[15,60]},{canonical:"llama3.1",apiId:"llama3.1",provider:"ollama",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:128e3}},{canonical:"mistral",apiId:"mistral",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:false,contextWindow:32e3}},{canonical:"mistral-nemo",apiId:"mistral-nemo",provider:"ollama",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:128e3}},{canonical:"llava",apiId:"llava",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:true,contextWindow:32e3}},{canonical:"qwen2:0.5b",apiId:"qwen2:0.5b",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:false,contextWindow:32e3}}],te=Object.fromEntries(Ge.filter(n=>n.canonical!==n.apiId).map(n=>[n.canonical,n.apiId]));te["claude-sonnet-4"]="claude-sonnet-4-20250514";te["claude-opus-4"]="claude-opus-4-20250514";te["claude-haiku-4"]="claude-haiku-4-5-20251001";});function We(n){if(ke[n])return ke[n];for(let[e,t]of Object.entries(ke))if(n.startsWith(e.split("-20")[0]))return t;return [3,15]}var Ot,Ut,ke,ne,Ke=C(()=>{V();Re();Ot="https://api.anthropic.com/v1/messages",Ut="2023-06-01",ke={"claude-sonnet-4-5-20250514":[3,15],"claude-sonnet-4-20250514":[3,15],"claude-haiku-4-5-20250414":[.8,4],"claude-opus-4-20250514":[15,75]};ne=class extends S{type="anthropic-api";supportsStreaming=true;supportsMultimodal=true;platformRequirement="any";apiKey;baseUrl;dangerousDirectBrowserAccess;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.apiKey=e.apiKey,this.baseUrl=e.baseUrl?.replace(/\/$/,"")||Ot,this.dangerousDirectBrowserAccess=e.dangerousDirectBrowserAccess??false;}async probe(){try{let e=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify({model:"claude-haiku-4-5-20250414",max_tokens:1,messages:[{role:"user",content:"hi"}]})});return e.status===401?{available:!0,authenticated:!1}:e.ok||e.status===429?{available:!0,authenticated:!0,rateLimitTier:e.headers.get("anthropic-ratelimit-tier")??void 0}:e.status===529?{available:!0,authenticated:!0}:{available:!1,authenticated:!1}}catch{return {available:false,authenticated:false}}}async execute(e,t){let r=Date.now(),s=this.buildRequestBody(e,false),o=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(s),signal:t?.signal});if(!o.ok)throw this.handleErrorResponse(o),new Error("unreachable");let i=await o.json(),a=Date.now()-r;this.recordSuccess();let[l,p]=We(i.model||e.model),c=((i.usage?.input_tokens||0)*l+(i.usage?.output_tokens||0)*p)/1e6,d={},f=["request-id","x-request-id","anthropic-ratelimit-requests-limit","anthropic-ratelimit-requests-remaining","anthropic-ratelimit-requests-reset","anthropic-ratelimit-tokens-limit","anthropic-ratelimit-tokens-remaining","anthropic-ratelimit-tokens-reset","anthropic-ratelimit-tier"];for(let m of f){let b=o.headers.get(m);b&&(d[m]=b);}return {content:i.content||[],model:i.model||e.model,usage:{input_tokens:i.usage?.input_tokens||0,output_tokens:i.usage?.output_tokens||0},stop_reason:i.stop_reason||"end_turn",provider:"anthropic-api",costUsd:Math.round(c*1e6)/1e6,durationMs:a,fallbackChain:["anthropic-api:success"],headers:Object.keys(d).length>0?d:void 0}}async streamRaw(e,t){let r=this.buildRequestBody(e,true),s=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(r),signal:t?.signal});if(s.ok||this.handleErrorResponse(s),!s.body)throw new Error("ANTHROPIC_API_ERROR: Response body is null \u2014 streaming unavailable");let o=s.body.getReader(),i=new TextDecoder;return (async function*(){let a="",l=false;try{for(;;){let{done:p,value:c}=await o.read();if(p)break;a+=i.decode(c,{stream:!0});let d=a.split(`
|
|
12
|
+
`);a=d.pop()||"";let f=null;for(let m of d)if(m.startsWith("event: "))f=m.slice(7).trim();else if(m.startsWith("data: ")){let b=m.slice(6);if(f&&b)try{let u=JSON.parse(b);u.type==="message_stop"&&(l=!0),yield u;}catch{console.warn("[inference-relay] Skipped malformed SSE data:",b.substring(0,100));}f=null;}else m.trim()===""&&(f=null);}l||(yield {type:"message_stop"});}finally{try{o.releaseLock();}catch{}}})()}estimateCost(e){let t=Math.round(JSON.stringify(e.messages).length/4),r=e.system?Math.round(e.system.length/4):0,s=t+r,[o,i]=We(e.model),a=(s*o+1e3*i)/1e6;return {estimatedInputTokens:s,estimatedCost:Math.round(a*1e4)/1e4,model:e.model,provider:"anthropic-api"}}getHealth(){return {...this.health}}reportRateLimit(e){this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now();}buildHeaders(){let e={"content-type":"application/json","x-api-key":this.apiKey,"anthropic-version":Ut};return this.dangerousDirectBrowserAccess&&(e["anthropic-dangerous-direct-browser-access"]="true"),e}buildRequestBody(e,t){let r={model:Ye(e.model),max_tokens:e.max_tokens,messages:e.messages};return t&&(r.stream=true),e.system&&(r.system=e.system),e.temperature!==void 0&&(r.temperature=e.temperature),e.tools&&e.tools.length>0&&(r.tools=e.tools),e.metadata&&(r.metadata=e.metadata),r}handleErrorResponse(e){let t=e.status;if(t===401)throw this.health.status="unhealthy",this.health.lastError="Authentication failed",this.health.errorCount++,this.health.lastCheckAt=Date.now(),new Error("ANTHROPIC_AUTH_ERROR: Invalid API key (401)");if(t===429){let r=e.headers.get("retry-after"),s=r?parseInt(r,10)*1e3:6e4;this.health.status="degraded",this.health.lastError=`Rate limited \u2014 retry after ${s}ms`,this.health.lastCheckAt=Date.now();let o=new Error(`ANTHROPIC_RATE_LIMIT: Too many requests (429). Retry after ${s}ms`);throw o.retryAfterMs=s,o}throw t===529?(this.health.status="degraded",this.health.lastError="API overloaded",this.health.lastCheckAt=Date.now(),new Error("ANTHROPIC_OVERLOADED: API is temporarily overloaded (529)")):(this.health.errorCount++,this.health.lastCheckAt=Date.now(),new Error(`ANTHROPIC_API_ERROR: Request failed with status ${t}`))}recordSuccess(){this.health.status="healthy",this.health.lastCheckAt=Date.now(),this.health.errorCount=0,this.health.lastError=void 0;}};});function qe(n){if(Te[n])return Te[n];for(let[e,t]of Object.entries(Te))if(n.startsWith(e))return t;return [2.5,10]}function Ft(n){return typeof n=="string"?n:n.filter(e=>e.type==="text"&&e.text).map(e=>e.text).join("")}function Ht(n,e){let t=[];e&&t.push({role:"system",content:e});for(let r of n){let s=Ft(r.content);if(r.role==="assistant"&&Array.isArray(r.content)){let o=r.content.filter(i=>i.type==="tool_use");if(o.length>0){t.push({role:"assistant",content:s||"",tool_calls:o.map(i=>({id:i.id||`call_${Math.random().toString(36).slice(2)}`,type:"function",function:{name:i.name,arguments:typeof i.input=="string"?i.input:JSON.stringify(i.input)}}))});continue}}t.push({role:r.role,content:s});}return t}function Je(n){switch(n){case "stop":return "end_turn";case "length":return "max_tokens";case "tool_calls":case "function_call":return "tool_use";default:return "end_turn"}}var Nt,je,Te,re,ze=C(()=>{V();Nt="https://api.openai.com/v1/chat/completions",je=["gpt-","o1","o3","o4","chatgpt-","text-","davinci","babbage","ft:"],Te={"gpt-4.1":[2,8],"gpt-4.1-mini":[.4,1.6],"gpt-4.1-nano":[.1,.4],"gpt-4o":[2.5,10],"gpt-4o-mini":[.15,.6],"o4-mini":[1.1,4.4],o3:[.4,1.6],"o3-mini":[1.1,4.4],o1:[15,60],"o1-mini":[3,12],"gpt-4-turbo":[10,30],"gpt-4":[30,60],"gpt-3.5-turbo":[.5,1.5]};re=class extends S{type="openai-api";supportsStreaming=true;supportsMultimodal=true;platformRequirement="any";apiKey;baseUrl;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.apiKey=e.apiKey,this.baseUrl=e.baseUrl?.replace(/\/$/,"")||Nt;}async probe(){try{let e=this.baseUrl.replace(/\/chat\/completions$/,"/models"),t=await fetch(e,{method:"GET",headers:this.buildHeaders()});return t.status===401?{available:!0,authenticated:!1}:t.ok||t.status===429?{available:!0,authenticated:!0}:{available:!1,authenticated:!1}}catch{return {available:false,authenticated:false}}}async execute(e,t){let r=Date.now(),s=this.buildRequestBody(e,false),o=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(s),signal:t?.signal});if(!o.ok)throw this.handleErrorResponse(o),new Error("unreachable");let i=await o.json(),a=Date.now()-r;this.recordSuccess();let l=i.choices?.[0],p=[];if(l?.message?.content&&p.push({type:"text",text:l.message.content}),l?.message?.tool_calls)for(let h of l.message.tool_calls)p.push({type:"tool_use",id:h.id,name:h.function?.name,input:h.function?.arguments?JSON.parse(h.function.arguments):{}});let c=i.usage?.prompt_tokens||0,d=i.usage?.completion_tokens||0,[f,m]=qe(i.model||e.model),b=(c*f+d*m)/1e6,u={},y=["x-request-id","x-ratelimit-limit-requests","x-ratelimit-remaining-requests","x-ratelimit-limit-tokens","x-ratelimit-remaining-tokens","x-ratelimit-reset-requests","x-ratelimit-reset-tokens"];for(let h of y){let g=o.headers.get(h);g&&(u[h]=g);}return {content:p,model:i.model||e.model,usage:{input_tokens:c,output_tokens:d},stop_reason:Je(l?.finish_reason),provider:"openai-api",costUsd:Math.round(b*1e6)/1e6,durationMs:a,fallbackChain:["openai-api:success"],headers:Object.keys(u).length>0?u:void 0}}async streamRaw(e,t){let r=this.buildRequestBody(e,true),s=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(r),signal:t?.signal});if(s.ok||this.handleErrorResponse(s),!s.body)throw new Error("OPENAI_API_ERROR: Response body is null \u2014 streaming unavailable");let o=s.body.getReader(),i=new TextDecoder,a=e.model;return (async function*(){let l="",p=false,c=false,d,f;yield {type:"message_start",message:{model:a,usage:{input_tokens:0,output_tokens:0}}};try{for(;;){let{done:m,value:b}=await o.read();if(m)break;l+=i.decode(b,{stream:!0});let u=l.split(`
|
|
13
|
+
`);l=u.pop()||"";for(let y of u){if(!y.startsWith("data: "))continue;let h=y.slice(6).trim();if(h==="[DONE]"){yield {type:"message_delta",delta:{stop_reason:f||"end_turn"},usage:d||{}},yield {type:"message_stop"},c=!0;return}try{let g=JSON.parse(h);g.usage&&(d={input_tokens:g.usage.prompt_tokens||0,output_tokens:g.usage.completion_tokens||0});let R=g.choices?.[0];if(!R)continue;let E=R.delta;E?.content&&!p&&(p=!0,yield {type:"content_block_start",index:0,content_block:{type:"text",text:""}}),E?.content&&(yield {type:"content_block_delta",index:0,delta:{type:"text_delta",text:E.content}}),R.finish_reason&&(p&&(yield {type:"content_block_stop",index:0}),f=Je(R.finish_reason));}catch{}}}c||(yield {type:"message_delta",delta:{stop_reason:f||"end_turn"},usage:d||{}},yield {type:"message_stop"});}finally{try{o.releaseLock();}catch{}}})()}estimateCost(e){let t=Math.round(JSON.stringify(e.messages).length/4),r=e.system?Math.round(e.system.length/4):0,s=t+r,[o,i]=qe(e.model),a=(s*o+1e3*i)/1e6;return {estimatedInputTokens:s,estimatedCost:Math.round(a*1e4)/1e4,model:e.model,provider:"openai-api"}}getHealth(){return {...this.health}}reportRateLimit(e){this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now();}buildHeaders(){return {"Content-Type":"application/json",Authorization:`Bearer ${this.apiKey}`}}buildRequestBody(e,t){let r=e.model.toLowerCase();if(!je.some(a=>r.startsWith(a)))throw new Error(`OPENAI_MODEL_TRANSLATION_REQUIRED: The OpenAI provider received the model name '${e.model}', which does not match any known OpenAI model prefix (${je.join(", ")}). The relay does NOT auto-translate model names across providers in v1.3.x. Either (a) pass an OpenAI model name like 'gpt-4o-mini' when routing to this provider, or (b) route non-OpenAI models through a provider that understands them (e.g. 'anthropic-api' for Claude models).`);let o=Ht(e.messages,e.system),i={model:e.model,max_tokens:e.max_tokens,messages:o};return t&&(i.stream=true,i.stream_options={include_usage:true}),e.temperature!==void 0&&(i.temperature=e.temperature),e.tools&&e.tools.length>0&&(i.tools=e.tools),i}handleErrorResponse(e){let t=e.status;if(t===401)throw this.health.status="unhealthy",this.health.lastError="Authentication failed",this.health.errorCount++,this.health.lastCheckAt=Date.now(),new Error("OPENAI_AUTH_ERROR: Invalid API key (401)");if(t===429){let r=e.headers.get("retry-after"),s=r?parseInt(r,10)*1e3:6e4;this.health.status="degraded",this.health.lastError=`Rate limited \u2014 retry after ${s}ms`,this.health.lastCheckAt=Date.now();let o=new Error(`OPENAI_RATE_LIMIT: Too many requests (429). Retry after ${s}ms`);throw o.retryAfterMs=s,o}throw this.health.errorCount++,this.health.lastCheckAt=Date.now(),new Error(`OPENAI_API_ERROR: Request failed with status ${t}`)}recordSuccess(){this.health.status="healthy",this.health.lastCheckAt=Date.now(),this.health.errorCount=0,this.health.lastError=void 0;}};});function $t(n){return typeof n=="string"?n:n.filter(e=>e.type==="text"&&e.text).map(e=>e.text).join("")}function Vt(n,e){let t=[];e&&t.push({role:"system",content:e});for(let r of n)t.push({role:r.role,content:$t(r.content)});return t}var Bt,se,Xe=C(()=>{V();Bt="http://localhost:11434";se=class extends S{type="ollama";supportsStreaming=true;supportsMultimodal=false;platformRequirement="desktop";baseUrl;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.baseUrl=e?.baseUrl?.replace(/\/$/,"")||Bt;}async probe(){try{return (await fetch(`${this.baseUrl}/api/tags`,{method:"GET"})).ok?{available:!0,authenticated:!0}:{available:!1,authenticated:!1}}catch{return {available:false,authenticated:false}}}async execute(e,t){let r=Date.now(),s=this.buildRequestBody(e,false),o=await fetch(`${this.baseUrl}/api/chat`,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify(s),signal:t?.signal});if(!o.ok)throw this.handleErrorResponse(o),new Error("unreachable");let i=await o.json(),a=Date.now()-r;this.recordSuccess();let l=[];i.message?.content&&l.push({type:"text",text:i.message.content});let p=i.prompt_eval_count||0,c=i.eval_count||0;return {content:l,model:i.model||e.model,usage:{input_tokens:p,output_tokens:c},stop_reason:(i.done,"end_turn"),provider:"ollama",costUsd:0,durationMs:a,fallbackChain:["ollama:success"]}}async streamRaw(e,t){let r=this.buildRequestBody(e,true),s=await fetch(`${this.baseUrl}/api/chat`,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify(r),signal:t?.signal});if(s.ok||this.handleErrorResponse(s),!s.body)throw new Error("OLLAMA_ERROR: Response body is null \u2014 streaming unavailable");let o=s.body.getReader(),i=new TextDecoder,a=e.model;return (async function*(){let l="",p=false,c=false;yield {type:"message_start",message:{model:a,usage:{input_tokens:0,output_tokens:0}}};try{for(;;){let{done:d,value:f}=await o.read();if(d)break;l+=i.decode(f,{stream:!0});let m=l.split(`
|
|
14
|
+
`);l=m.pop()||"";for(let b of m){let u=b.trim();if(u)try{let y=JSON.parse(u);if(y.message?.content&&(p||(p=!0,yield {type:"content_block_start",index:0,content_block:{type:"text",text:""}}),yield {type:"content_block_delta",index:0,delta:{type:"text_delta",text:y.message.content}}),y.done){p&&(yield {type:"content_block_stop",index:0}),yield {type:"message_delta",delta:{stop_reason:"end_turn"},usage:{input_tokens:y.prompt_eval_count||0,output_tokens:y.eval_count||0}},yield {type:"message_stop"},c=!0;return}}catch{}}}c||(yield {type:"message_stop"});}finally{try{o.releaseLock();}catch{}}})()}estimateCost(e){let t=Math.round(JSON.stringify(e.messages).length/4),r=e.system?Math.round(e.system.length/4):0;return {estimatedInputTokens:t+r,estimatedCost:0,model:e.model,provider:"ollama"}}getHealth(){return {...this.health}}reportRateLimit(e){this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now();}buildRequestBody(e,t){if(e.tools&&e.tools.length>0)throw new Error(`OLLAMA_TOOLS_NOT_SUPPORTED: The Ollama provider received ${e.tools.length} tool definition(s) in params.tools, but this provider does not implement function calling in v1.3.x. For tool-use workflows route to 'anthropic-api' or 'openai-api' instead. Tracking issue: 1.4.0 ships an Ollama tool-use shim for models that support native function calling (llama3.1, mistral-nemo, etc.).`);let r=Vt(e.messages,e.system),s={model:e.model,messages:r,stream:t};return e.temperature!==void 0&&(s.options={temperature:e.temperature}),s}handleErrorResponse(e){let t=e.status;throw this.health.errorCount++,this.health.lastCheckAt=Date.now(),t===404?(this.health.status="unhealthy",this.health.lastError="Model not found",new Error("OLLAMA_MODEL_NOT_FOUND: Requested model is not available locally (404)")):(this.health.status="unhealthy",this.health.lastError=`Request failed with status ${t}`,new Error(`OLLAMA_ERROR: Request failed with status ${t}`))}recordSuccess(){this.health.status="healthy",this.health.lastCheckAt=Date.now(),this.health.errorCount=0,this.health.lastError=void 0;}};});var ie,Qe=C(()=>{ie=class extends EventEmitter{constructor(){super(),this.setMaxListeners(50);}emitTyped(e){return this.emit(e.kind,e)}onTyped(e,t){return this.on(e,t)}offTyped(e,t){return this.off(e,t)}};});function oe(n,e){if(e instanceof v)return `${n}:failed:${e.code}:${e.message}`;let t=e?.message;return `${n}:failed:${t?.slice(0,it)||"unknown"}`}function et(n){return n.messages.some(e=>typeof e.content=="string"?false:Array.isArray(e.content)&&e.content.some(t=>t.type==="image"||t.type==="image_url"))}function Ie(n){if(!(n instanceof Error))return "unknown";let e=n.code;if(typeof e=="string"){if(e==="LICENSE_REJECTED")return "auth_failed";if(e==="GATEWAY_FETCH_FAILED"||e==="GATEWAY_VERIFICATION_FAILED"||e==="GATEWAY_LOAD_FAILED")return "gateway_unavailable"}let t=n.message.toLowerCase(),r=n.name.toLowerCase();return t.startsWith("openai_model_translation_required:")||t.startsWith("ollama_tools_not_supported:")||t.startsWith("multimodal_cli_too_old:")||t.startsWith("stream_json_cli_too_old:")?"provider_incompatible":t.includes("rate limit")||t.includes("429")||r.includes("ratelimit")?"rate_limited":t.includes("auth")||t.includes("401")||t.includes("403")||t.includes("unauthorized")||t.includes("forbidden")?"auth_failed":t.includes("timeout")||t.includes("timed out")||t.includes("aborted")||r==="aborterror"||r==="timeouterror"?"timeout":"unknown"}function tt(n){return [...n].sort((e,t)=>e.priority-t.priority)}function nt(n){return {provider:n,inputTokens:0,outputTokens:0,costUsd:0,userCostUsd:0}}function Yt(n,e,t){typeof t.input_tokens=="number"&&(n.inputTokens=t.input_tokens),typeof t.output_tokens=="number"&&(e==="anthropic-api"||e==="claude-cli"?n.outputTokens+=t.output_tokens:n.outputTokens=t.output_tokens);}function rt(n){return {input_tokens:n[n.length-1]?.inputTokens??0,output_tokens:n.reduce((e,t)=>e+t.outputTokens,0)}}function Wt(n,e){if(e.type==="content_block_start"){let t=e.content_block;t?.type==="text"?n.content.push({type:"text",text:t.text||"",open:true}):t?.type==="tool_use"&&n.content.push({type:"tool_use",id:t.id,name:t.name,input:{},_partialJson:"",open:true});}else if(e.type==="content_block_delta"){let t=n.content[n.content.length-1];if(!t)return;t.type==="text"&&e.delta?.type==="text_delta"?t.text+=e.delta.text||"":t.type==="tool_use"&&e.delta?.type==="input_json_delta"&&(t._partialJson=(t._partialJson||"")+(e.delta.partial_json||""));}else if(e.type==="content_block_stop"){for(let t=n.content.length-1;t>=0;t--)if(n.content[t].open){if(n.content[t].open=false,n.content[t].type==="tool_use"){let r=n.content[t];try{r.input=JSON.parse(r._partialJson||"{}");}catch{}delete r._partialJson;}break}}}function Kt(n,e,t,r,s){if(s&&n.type==="content_block_delta"&&n.index===s.fromIndex)return {action:"yield",event:{...n,index:s.toIndex}};if(s&&n.type==="content_block_stop"&&n.index===s.fromIndex)return {action:"suppress",coalesceMode:null,coalesceModeCleared:true};if(r&&n.type==="content_block_start"&&n.content_block?.type==="text"){let i=t.content[t.content.length-1];if(i?.type==="text"&&i.open===true)return {action:"suppress",coalesceMode:{fromIndex:n.index+e,toIndex:t.content.length-1},clearedJustFailedOver:true}}let o=r&&n.type==="content_block_start";return typeof n.index=="number"&&e>0?{action:"yield",event:{...n,index:n.index+e},clearedJustFailedOver:o}:{action:"yield",event:n,clearedJustFailedOver:o}}function jt(n){let e=[];for(let t of n)t.type==="text"?t.text&&e.push({type:"text",text:t.text}):t.type==="tool_use"&&!t.open&&t.input&&Object.keys(t.input).length>0&&e.push({type:"tool_use",id:t.id,name:t.name,input:t.input});return e}function qt(n,e){return n?`${n}
|
|
15
15
|
|
|
16
|
-
${e}`:e}function nt(n){let e=0;for(let t of n.content)t.type==="text"?e+=t.text.length:t.type==="tool_use"&&(e+=(t._partialJson||"").length);return e}function qt(n){return Math.max(0,n.content.length-1)}async function*Jt(n){for await(let e of n)await new Promise(t=>setTimeout(t,50+Math.random()*50)),yield e;}var rt,Xe,F,ae,Ie=C(()=>{he();ze();Ce();be();rt=200;Xe=new Set(["anthropic-api","openai-api","ollama"]),F=class extends Error{code="ALL_PROVIDERS_FAILED";fallbackChain;constructor(e){super(`All providers failed. Chain: ${e.join(" \u2192 ")}`),this.name="RelayError",this.fallbackChain=e;}};ae=class n{healthMap=new Map;engineEvents=new se;static UNHEALTHY_TTL_MS=300*1e3;getHealthStatus(e){let t=this.healthMap.get(e);return t?t.status==="unhealthy"&&Date.now()-t.updatedAt>=n.UNHEALTHY_TTL_MS?(this.healthMap.delete(e),"healthy"):t.status:"healthy"}markUnhealthy(e,t){this.healthMap.set(e,{status:"unhealthy",lastError:t,updatedAt:Date.now()});}markDegraded(e,t){this.healthMap.set(e,{status:"degraded",lastError:t,updatedAt:Date.now()});}markHealthy(e){this.healthMap.set(e,{status:"healthy",updatedAt:Date.now()});}resetHealth(){this.healthMap.clear();}async execute(e,t,r){let s=Ze(t),a=[],i=Qe(e),l=randomUUID(),o=Date.now();for(let p of s){let c=p.type;if(this.getHealthStatus(c)==="unhealthy"){a.push(`${c}:skipped:unhealthy`);continue}if(i&&!p.supportsMultimodal){a.push(`${c}:skipped:no_multimodal`);continue}let h=ve(e.model);if(h&&h!==c&&Xe.has(c)){a.push(`${c}:skipped:wrong_provider`);continue}try{let m=await p.execute(e,{signal:r?.signal});return this.markHealthy(c),a.push(`${c}:success`),this.emitEngineEvent({kind:"stream_complete",callId:l,id:randomUUID(),ts:Date.now(),stitched:!1,segments:[{provider:c,inputTokens:m.usage.input_tokens,outputTokens:m.usage.output_tokens,costUsd:m.costUsd,userCostUsd:m.userCostUsd??0}],provider:c,model:m.model,inputTokens:m.usage.input_tokens,outputTokens:m.usage.output_tokens,costUsd:m.costUsd,userCostUsd:m.userCostUsd,durationMs:m.durationMs??Date.now()-o,stopReason:m.stop_reason,fallbackChain:a,rateLimitHit:!1,securityState:r?.licenseState}),{...m,fallbackChain:a}}catch(m){if(m?.name==="AbortError"||r?.signal?.aborted)throw m;switch(ke(m)){case "rate_limited":p.reportRateLimit({provider:c,utilization:1,resetsAt:Date.now()+6e4}),this.markDegraded(c,"rate_limited"),a.push(`${c}:failed:rate_limited`);break;case "auth_failed":this.markUnhealthy(c,"auth_failed"),a.push(`${c}:failed:auth_failed`);break;case "timeout":a.push(`${c}:failed:timeout`);break;case "provider_incompatible":a.push(`${c}:incompatible`);break;case "gateway_unavailable":this.markUnhealthy(c,"gateway_unavailable"),a.push(ie(c,m));break;default:a.push(ie(c,m));break}}}throw new F(a)}async stream(e,t,r){let s=randomUUID(),a=Ze(t),i=Qe(e),l=Date.now(),o=[],p=null,c=null;for(let g of a){let f=g.type;if(this.getHealthStatus(f)==="unhealthy"){o.push(`${f}:skipped:unhealthy`);continue}if(i&&!g.supportsMultimodal){o.push(`${f}:skipped:no_multimodal`);continue}if(!g.supportsStreaming){o.push(`${f}:skipped:no_streaming`);continue}let R=ve(e.model);if(R&&R!==f&&Xe.has(f)){o.push(`${f}:skipped:wrong_provider`);continue}try{c=await g.streamRaw(e,{signal:r?.signal}),p=g,o.push(`${f}:success`);break}catch(b){if(b?.name==="AbortError"||r?.signal?.aborted)throw b;switch(ke(b)){case "rate_limited":g.reportRateLimit({provider:f,utilization:1,resetsAt:Date.now()+6e4}),this.markDegraded(f,"rate_limited"),o.push(`${f}:failed:rate_limited`);break;case "auth_failed":this.markUnhealthy(f,"auth_failed"),o.push(`${f}:failed:auth_failed`);break;case "timeout":o.push(`${f}:failed:timeout`);break;case "provider_incompatible":o.push(`${f}:incompatible`);break;case "gateway_unavailable":this.markUnhealthy(f,"gateway_unavailable"),o.push(ie(f,b));break;default:o.push(ie(f,b));break}}}if(!p||!c)throw new F(o);let h=this,m=r?.maxMidStreamFailovers??2,E=new Set([p]);async function*u(){let g=p,f=c[Symbol.asyncIterator](),R={role:"assistant",content:[]},b={segments:[]},L=et(g.type);b.segments.push(L);let Te=0,K=null,j=false,pe=m,H="",B="";for(h.emitEngineEvent({kind:"segment_start",callId:s,id:randomUUID(),ts:Date.now(),provider:g.type});;)try{for(;;){let k=await f.next();if(k.done)break;let D=k.value,P=Yt(D,Te,R,j,K);if(P.action==="suppress"){"coalesceMode"in P&&(K=P.coalesceMode),P.clearedJustFailedOver&&(j=!1);continue}P.coalesceModeCleared&&(K=null),P.clearedJustFailedOver&&(j=!1);let A=P.event;if(Gt(R,A),A.type==="message_start"){if(b.segments.length>1)continue;H=A.message?.model||H,A.message?.usage?.input_tokens!=null&&(L.inputTokens=A.message.usage.input_tokens);}if(A.type==="message_delta"&&(A.delta?.stop_reason&&(B=A.delta.stop_reason),A.usage)){Wt(L,g.type,A.usage),yield {...A,usage:void 0};continue}if(A.type==="message_stop")break;yield A;}if(R.content.length===0&&L.outputTokens===0)throw new Error("stream ended without producing any content");h.markHealthy(g.type),h.emitEngineEvent({kind:"segment_complete",callId:s,id:randomUUID(),ts:Date.now(),provider:g.type,usage:{...L}});let I=tt(b.segments);yield {type:"message_delta",delta:{stop_reason:B||"end_turn"},usage:I},yield {type:"message_stop"},h.emitEngineEvent({kind:"stream_complete",callId:s,id:randomUUID(),ts:Date.now(),stitched:b.segments.length>1,segments:b.segments.map(k=>({...k})),provider:b.segments.map(k=>k.provider).join("+"),model:H,inputTokens:I.input_tokens,outputTokens:I.output_tokens,costUsd:b.segments.reduce((k,D)=>k+D.costUsd,0),userCostUsd:b.segments.reduce((k,D)=>k+D.userCostUsd,0),durationMs:Date.now()-l,stopReason:B||"end_turn",fallbackChain:o,rateLimitHit:!1,securityState:r?.licenseState});return}catch(I){if(I?.name==="AbortError"||r?.signal?.aborted)throw I;let k=ke(I);try{await f.return?.();}catch{}if(h.markUnhealthy(g.type,String(I)),h.emitEngineEvent({kind:"mid_stream_failover_attempt",callId:s,id:randomUUID(),ts:Date.now(),from:g.type,bytesBuffered:nt(R),errorClass:k,retriesLeft:pe}),o.push(`${g.type}:partial(${nt(R)}b)`),pe<=0)throw h.emitTerminalExhaustion(s,b,H,B,o,l,r?.licenseState),I;let D=Kt(R.content),P=[...e.messages,{role:"assistant",content:D}],A=null,q=null,Se=g;for(;;){let x=h.nextHealthyProvider(a,Se,i,E);if(!x)break;Se=x,E.add(x);let bt={...e,messages:P,...x.type!=="anthropic-api"&&x.type!=="claude-cli"?{system:jt(e.system,"[SYSTEM]: CONTINUE_LAST_RESPONSE_WITHOUT_REPEATING")}:{}};try{q=await x.streamRaw(bt,{signal:r?.signal}),A=x;break}catch(Pe){h.markUnhealthy(x.type,String(Pe)),o.push(`${x.type}:failed_connect:${String(Pe).slice(0,rt)}`);}}if(!A||!q)throw h.emitTerminalExhaustion(s,b,H,B,o,l,r?.licenseState),I;o.push(`${A.type}:stitched`),h.emitEngineEvent({kind:"mid_stream_failover_success",callId:s,id:randomUUID(),ts:Date.now(),from:g.type,to:A.type}),Te=qt(R)+1,pe--,g=A,f=q[Symbol.asyncIterator](),j=true,K=null,L=et(g.type),b.segments.push(L),h.emitEngineEvent({kind:"segment_start",callId:s,id:randomUUID(),ts:Date.now(),provider:g.type});}}let y=u();return r?.licenseState==="SEC_DEGRADED"&&(y=Jt(y)),J(y)}nextHealthyProvider(e,t,r,s){let a=e.indexOf(t);if(a===-1)return null;for(let i=a+1;i<e.length;i++){let l=e[i];if(!s.has(l)&&!(r&&!l.supportsMultimodal)&&l.supportsStreaming)return l}return null}emitEngineEvent(e){this.engineEvents.emitTyped(e);}emitTerminalExhaustion(e,t,r,s,a,i,l){let o=tt(t.segments);this.emitEngineEvent({kind:"stream_complete",callId:e,id:randomUUID(),ts:Date.now(),stitched:t.segments.length>1,segments:t.segments.map(p=>({...p})),provider:t.segments.map(p=>p.provider).join("+"),model:r,inputTokens:o.input_tokens,outputTokens:o.output_tokens,costUsd:t.segments.reduce((p,c)=>p+c.costUsd,0),userCostUsd:t.segments.reduce((p,c)=>p+c.userCostUsd,0),durationMs:Date.now()-i,stopReason:s||"error",fallbackChain:a,rateLimitHit:false,securityState:l});}};});function st(n,e){if(e==="prefer-user"){let s=n.filter(i=>i.type==="claude-cli"),a=n.filter(i=>i.type!=="claude-cli");return [...s,...a]}if(e==="prefer-app"){let s=n.filter(i=>i.type==="claude-cli");return [...n.filter(i=>i.type!=="claude-cli"),...s]}let t=n.filter(s=>s.type===e),r=n.filter(s=>s.type!==e);return [...t,...r]}function it(n){return (e,t)=>n.maxCostPerCall!==void 0&&t.estimatedCost>n.maxCostPerCall?"prefer-user":n.preferLocal?"ollama":"prefer-user"}var at=C(()=>{});function ot(n){if(we[n])return we[n];for(let[e,t]of Object.entries(we))if(n.startsWith(e))return t;return [3,15]}var we,lt=C(()=>{we={"claude-opus-4":[15,75],"claude-sonnet-4":[3,15],"claude-haiku-4":[.8,4],"gpt-4o":[2.5,10],"gpt-4o-mini":[.15,.6],"gpt-4-turbo":[10,30],"gpt-3.5-turbo":[.5,1.5],llama:[0,0],mistral:[0,0],codellama:[0,0]};});function ct(n){let e=0;for(let l of n.messages)if(typeof l.content=="string")e+=l.content.length;else if(Array.isArray(l.content))for(let o of l.content)o.type==="text"&&o.text&&(e+=o.text.length);n.system&&(e+=n.system.length);let t=Math.ceil(e/4),r=Math.min(n.max_tokens||1024,4096),[s,a]=ot(n.model),i=(t*s+r*a)/1e6;return {estimatedInputTokens:t,estimatedCost:Math.round(i*1e4)/1e4,model:n.model,provider:""}}var ut=C(()=>{lt();});var oe,dt=C(()=>{oe=class{endpoint;enabled;keyId;lineageBuffer=new Map;unsubscribers=[];constructor(e){this.endpoint=e.endpoint||"https://api.inference-relay.com/v1/events",this.enabled=e.enabled??true,this.keyId=e.keyId||"";}subscribeToEngine(e){let t=s=>this.bufferLineageEvent(s),r=s=>{s.kind==="stream_complete"&&this.flushLineageForCall(s);};e.engineEvents.onTyped("segment_start",t),e.engineEvents.onTyped("segment_complete",t),e.engineEvents.onTyped("mid_stream_failover_attempt",t),e.engineEvents.onTyped("mid_stream_failover_success",t),e.engineEvents.onTyped("stream_complete",r),this.unsubscribers.push(()=>{e.engineEvents.offTyped("segment_start",t),e.engineEvents.offTyped("segment_complete",t),e.engineEvents.offTyped("mid_stream_failover_attempt",t),e.engineEvents.offTyped("mid_stream_failover_success",t),e.engineEvents.offTyped("stream_complete",r);});}unsubscribe(){for(let e of this.unsubscribers)e();this.unsubscribers=[],this.lineageBuffer.clear();}bufferLineageEvent(e){let t=this.lineageBuffer.get(e.callId)??[];t.push(e),this.lineageBuffer.set(e.callId,t);}flushLineageForCall(e){let t=this.lineageBuffer.get(e.callId)??[];this.lineageBuffer.delete(e.callId);let r={timestamp:e.ts,provider:e.provider,model:e.model,inputTokens:e.inputTokens,outputTokens:e.outputTokens,costUsd:e.costUsd,userCostUsd:e.userCostUsd,durationMs:e.durationMs,stopReason:e.stopReason,fallbackChain:e.fallbackChain,rateLimitHit:e.rateLimitHit,securityState:e.securityState,stitched:e.stitched,segments:e.segments,lineage:t.map(s=>s.id),promptContent:false,completionContent:false};this.report(r);}report(e){!this.enabled||!this.keyId||fetch(this.endpoint,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({keyId:this.keyId,provider:e.provider,model:e.model,inputTokens:e.inputTokens,outputTokens:e.outputTokens,costUsd:e.costUsd,...e.userCostUsd!==void 0?{userCostUsd:e.userCostUsd}:{},durationMs:e.durationMs,fallback:e.fallbackChain.length>1,timestamp:e.timestamp,...e.stitched?{stitched:true,segmentCount:e.segments?.length??0}:{},...e.lineage?{lineageIds:e.lineage}:{}})}).catch(()=>{});}};});async function le(n,e){if(!n)return {valid:false,error:"License key required. Get one at inference-relay.com"};let t=O.get(n);if(t&&Date.now()-t.cachedAt<t.ttl)return t.result;let r=new AbortController,a=!O.has(n)?3e3:1e3,i=setTimeout(()=>r.abort(),a),l;try{let p=(e||zt)+Xt;l=await(await fetch(p,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({key:n}),signal:r.signal})).json();}catch{return clearTimeout(i),t&&Date.now()-t.cachedAt<Zt?t.result:{valid:false,error:"License validation failed (network error). Visit inference-relay.com"}}clearTimeout(i);let o;if(l.jws)try{let{compactVerify:p,importSPKI:c}=await import('jose'),d=await c(en,"RS256"),{payload:h}=await p(l.jws,d);o=JSON.parse(new TextDecoder().decode(h));}catch{return O.delete(n),{valid:false,error:"License validation signature invalid. Possible MITM \u2014 refusing to start."}}else if(o=l,o.valid)return O.delete(n),{valid:false,error:"License validation response not signed. Possible MITM."};return O.set(n,{result:o,cachedAt:Date.now(),ttl:o.valid?Qt:6e4}),o}var zt,Xt,Qt,Zt,en,O,ce,pt=C(()=>{me();Be();Ge();qe();Je();Ie();Q();at();ut();fe();dt();zt="https://api.inference-relay.com",Xt="/v1/validate",Qt=1440*60*1e3,Zt=10080*60*1e3,en=`-----BEGIN PUBLIC KEY-----
|
|
16
|
+
${e}`:e}function st(n){let e=0;for(let t of n.content)t.type==="text"?e+=t.text.length:t.type==="tool_use"&&(e+=(t._partialJson||"").length);return e}function Jt(n){return Math.max(0,n.content.length-1)}async function*zt(n){for await(let e of n)await new Promise(t=>setTimeout(t,50+Math.random()*50)),yield e;}var it,Ze,H,ae,we=C(()=>{fe();Qe();Re();Ce();it=200;Ze=new Set(["anthropic-api","openai-api","ollama"]),H=class extends Error{code="ALL_PROVIDERS_FAILED";fallbackChain;constructor(e){super(`All providers failed. Chain: ${e.join(" \u2192 ")}`),this.name="RelayError",this.fallbackChain=e;}};ae=class n{healthMap=new Map;engineEvents=new ie;static UNHEALTHY_TTL_MS=300*1e3;getHealthStatus(e){let t=this.healthMap.get(e);return t?t.status==="unhealthy"&&Date.now()-t.updatedAt>=n.UNHEALTHY_TTL_MS?(this.healthMap.delete(e),"healthy"):t.status:"healthy"}markUnhealthy(e,t){this.healthMap.set(e,{status:"unhealthy",lastError:t,updatedAt:Date.now()});}markDegraded(e,t){this.healthMap.set(e,{status:"degraded",lastError:t,updatedAt:Date.now()});}markHealthy(e){this.healthMap.set(e,{status:"healthy",updatedAt:Date.now()});}resetHealth(){this.healthMap.clear();}async execute(e,t,r){let s=tt(t),o=[],i=et(e),a=randomUUID(),l=Date.now();for(let p of s){let c=p.type;if(this.getHealthStatus(c)==="unhealthy"){o.push(`${c}:skipped:unhealthy`);continue}if(i&&!p.supportsMultimodal){o.push(`${c}:skipped:no_multimodal`);continue}let f=Ae(e.model);if(f&&f!==c&&Ze.has(c)){o.push(`${c}:skipped:wrong_provider`);continue}try{let m=await p.execute(e,{signal:r?.signal});return this.markHealthy(c),o.push(`${c}:success`),this.emitEngineEvent({kind:"stream_complete",callId:a,id:randomUUID(),ts:Date.now(),stitched:!1,segments:[{provider:c,inputTokens:m.usage.input_tokens,outputTokens:m.usage.output_tokens,costUsd:m.costUsd,userCostUsd:m.userCostUsd??0}],provider:c,model:m.model,inputTokens:m.usage.input_tokens,outputTokens:m.usage.output_tokens,costUsd:m.costUsd,userCostUsd:m.userCostUsd,durationMs:m.durationMs??Date.now()-l,stopReason:m.stop_reason,fallbackChain:o,rateLimitHit:!1,securityState:r?.licenseState}),{...m,fallbackChain:o}}catch(m){if(m?.name==="AbortError"||r?.signal?.aborted)throw m;switch(Ie(m)){case "rate_limited":p.reportRateLimit({provider:c,utilization:1,resetsAt:Date.now()+6e4}),this.markDegraded(c,"rate_limited"),o.push(`${c}:failed:rate_limited`);break;case "auth_failed":this.markUnhealthy(c,"auth_failed"),o.push(`${c}:failed:auth_failed`);break;case "timeout":o.push(`${c}:failed:timeout`);break;case "provider_incompatible":o.push(`${c}:incompatible`);break;case "gateway_unavailable":this.markUnhealthy(c,"gateway_unavailable"),o.push(oe(c,m));break;default:o.push(oe(c,m));break}}}throw new H(o)}async stream(e,t,r){let s=randomUUID(),o=tt(t),i=et(e),a=Date.now(),l=[],p=null,c=null;for(let h of o){let g=h.type;if(this.getHealthStatus(g)==="unhealthy"){l.push(`${g}:skipped:unhealthy`);continue}if(i&&!h.supportsMultimodal){l.push(`${g}:skipped:no_multimodal`);continue}if(!h.supportsStreaming){l.push(`${g}:skipped:no_streaming`);continue}let R=Ae(e.model);if(R&&R!==g&&Ze.has(g)){l.push(`${g}:skipped:wrong_provider`);continue}try{c=await h.streamRaw(e,{signal:r?.signal}),p=h,l.push(`${g}:success`);break}catch(E){if(E?.name==="AbortError"||r?.signal?.aborted)throw E;switch(Ie(E)){case "rate_limited":h.reportRateLimit({provider:g,utilization:1,resetsAt:Date.now()+6e4}),this.markDegraded(g,"rate_limited"),l.push(`${g}:failed:rate_limited`);break;case "auth_failed":this.markUnhealthy(g,"auth_failed"),l.push(`${g}:failed:auth_failed`);break;case "timeout":l.push(`${g}:failed:timeout`);break;case "provider_incompatible":l.push(`${g}:incompatible`);break;case "gateway_unavailable":this.markUnhealthy(g,"gateway_unavailable"),l.push(oe(g,E));break;default:l.push(oe(g,E));break}}}if(!p||!c)throw new H(l);let f=this,m=r?.maxMidStreamFailovers??2,b=new Set([p]);async function*u(){let h=p,g=c[Symbol.asyncIterator](),R={role:"assistant",content:[]},E={segments:[]},k=nt(h.type);E.segments.push(k);let Pe=0,j=null,q=false,he=m,B="",$="";for(f.emitEngineEvent({kind:"segment_start",callId:s,id:randomUUID(),ts:Date.now(),provider:h.type});;)try{for(;;){let I=await g.next();if(I.done)break;let O=I.value,L=Kt(O,Pe,R,q,j);if(L.action==="suppress"){"coalesceMode"in L&&(j=L.coalesceMode),L.clearedJustFailedOver&&(q=!1);continue}L.coalesceModeCleared&&(j=null),L.clearedJustFailedOver&&(q=!1);let A=L.event;if(Wt(R,A),A.type==="message_start"){if(E.segments.length>1)continue;B=A.message?.model||B,A.message?.usage?.input_tokens!=null&&(k.inputTokens=A.message.usage.input_tokens);}if(A.type==="message_delta"&&(A.delta?.stop_reason&&($=A.delta.stop_reason),A.usage)){Yt(k,h.type,A.usage),yield {...A,usage:void 0};continue}if(A.type==="message_stop")break;yield A;}if(R.content.length===0&&k.outputTokens===0)throw new Error("stream ended without producing any content");f.markHealthy(h.type),f.emitEngineEvent({kind:"segment_complete",callId:s,id:randomUUID(),ts:Date.now(),provider:h.type,usage:{...k}});let w=rt(E.segments);yield {type:"message_delta",delta:{stop_reason:$||"end_turn"},usage:w},yield {type:"message_stop"},f.emitEngineEvent({kind:"stream_complete",callId:s,id:randomUUID(),ts:Date.now(),stitched:E.segments.length>1,segments:E.segments.map(I=>({...I})),provider:E.segments.map(I=>I.provider).join("+"),model:B,inputTokens:w.input_tokens,outputTokens:w.output_tokens,costUsd:E.segments.reduce((I,O)=>I+O.costUsd,0),userCostUsd:E.segments.reduce((I,O)=>I+O.userCostUsd,0),durationMs:Date.now()-a,stopReason:$||"end_turn",fallbackChain:l,rateLimitHit:!1,securityState:r?.licenseState});return}catch(w){if(w?.name==="AbortError"||r?.signal?.aborted)throw w;let I=Ie(w);try{await g.return?.();}catch{}if(f.markUnhealthy(h.type,String(w)),f.emitEngineEvent({kind:"mid_stream_failover_attempt",callId:s,id:randomUUID(),ts:Date.now(),from:h.type,bytesBuffered:st(R),errorClass:I,retriesLeft:he}),l.push(`${h.type}:partial(${st(R)}b)`),he<=0)throw f.emitTerminalExhaustion(s,E,B,$,l,a,r?.licenseState),w;let O=jt(R.content),L=[...e.messages,{role:"assistant",content:O}],A=null,J=null,Le=h;for(;;){let x=f.nextHealthyProvider(o,Le,i,b);if(!x)break;Le=x,b.add(x);let vt={...e,messages:L,...x.type!=="anthropic-api"&&x.type!=="claude-cli"?{system:qt(e.system,"[SYSTEM]: CONTINUE_LAST_RESPONSE_WITHOUT_REPEATING")}:{}};try{J=await x.streamRaw(vt,{signal:r?.signal}),A=x;break}catch(xe){f.markUnhealthy(x.type,String(xe)),l.push(`${x.type}:failed_connect:${String(xe).slice(0,it)}`);}}if(!A||!J)throw f.emitTerminalExhaustion(s,E,B,$,l,a,r?.licenseState),w;l.push(`${A.type}:stitched`),f.emitEngineEvent({kind:"mid_stream_failover_success",callId:s,id:randomUUID(),ts:Date.now(),from:h.type,to:A.type}),Pe=Jt(R)+1,he--,h=A,g=J[Symbol.asyncIterator](),q=true,j=null,k=nt(h.type),E.segments.push(k),f.emitEngineEvent({kind:"segment_start",callId:s,id:randomUUID(),ts:Date.now(),provider:h.type});}}let y=u();return r?.licenseState==="SEC_DEGRADED"&&(y=zt(y)),z(y)}nextHealthyProvider(e,t,r,s){let o=e.indexOf(t);if(o===-1)return null;for(let i=o+1;i<e.length;i++){let a=e[i];if(!s.has(a)&&!(r&&!a.supportsMultimodal)&&a.supportsStreaming)return a}return null}emitEngineEvent(e){this.engineEvents.emitTyped(e);}emitTerminalExhaustion(e,t,r,s,o,i,a){let l=rt(t.segments);this.emitEngineEvent({kind:"stream_complete",callId:e,id:randomUUID(),ts:Date.now(),stitched:t.segments.length>1,segments:t.segments.map(p=>({...p})),provider:t.segments.map(p=>p.provider).join("+"),model:r,inputTokens:l.input_tokens,outputTokens:l.output_tokens,costUsd:t.segments.reduce((p,c)=>p+c.costUsd,0),userCostUsd:t.segments.reduce((p,c)=>p+c.userCostUsd,0),durationMs:Date.now()-i,stopReason:s||"error",fallbackChain:o,rateLimitHit:false,securityState:a});}};});function ot(n,e){if(e==="prefer-user"){let s=n.filter(i=>i.type==="claude-cli"),o=n.filter(i=>i.type!=="claude-cli");return [...s,...o]}if(e==="prefer-app"){let s=n.filter(i=>i.type==="claude-cli");return [...n.filter(i=>i.type!=="claude-cli"),...s]}let t=n.filter(s=>s.type===e),r=n.filter(s=>s.type!==e);return [...t,...r]}function at(n){return (e,t)=>n.maxCostPerCall!==void 0&&t.estimatedCost>n.maxCostPerCall?"prefer-user":n.preferLocal?"ollama":"prefer-user"}var lt=C(()=>{});function ct(n){if(Se[n])return Se[n];for(let[e,t]of Object.entries(Se))if(n.startsWith(e))return t;return [3,15]}var Se,ut=C(()=>{Se={"claude-opus-4":[15,75],"claude-sonnet-4":[3,15],"claude-haiku-4":[.8,4],"gpt-4o":[2.5,10],"gpt-4o-mini":[.15,.6],"gpt-4-turbo":[10,30],"gpt-3.5-turbo":[.5,1.5],llama:[0,0],mistral:[0,0],codellama:[0,0]};});function dt(n){let e=0;for(let a of n.messages)if(typeof a.content=="string")e+=a.content.length;else if(Array.isArray(a.content))for(let l of a.content)l.type==="text"&&l.text&&(e+=l.text.length);n.system&&(e+=n.system.length);let t=Math.ceil(e/4),r=Math.min(n.max_tokens||1024,4096),[s,o]=ct(n.model),i=(t*s+r*o)/1e6;return {estimatedInputTokens:t,estimatedCost:Math.round(i*1e4)/1e4,model:n.model,provider:""}}var pt=C(()=>{ut();});var le,mt=C(()=>{le=class{endpoint;enabled;keyId;lineageBuffer=new Map;unsubscribers=[];constructor(e){this.endpoint=e.endpoint||"https://api.inference-relay.com/v1/events",this.enabled=e.enabled??true,this.keyId=e.keyId||"";}subscribeToEngine(e){let t=s=>this.bufferLineageEvent(s),r=s=>{s.kind==="stream_complete"&&this.flushLineageForCall(s);};e.engineEvents.onTyped("segment_start",t),e.engineEvents.onTyped("segment_complete",t),e.engineEvents.onTyped("mid_stream_failover_attempt",t),e.engineEvents.onTyped("mid_stream_failover_success",t),e.engineEvents.onTyped("stream_complete",r),this.unsubscribers.push(()=>{e.engineEvents.offTyped("segment_start",t),e.engineEvents.offTyped("segment_complete",t),e.engineEvents.offTyped("mid_stream_failover_attempt",t),e.engineEvents.offTyped("mid_stream_failover_success",t),e.engineEvents.offTyped("stream_complete",r);});}unsubscribe(){for(let e of this.unsubscribers)e();this.unsubscribers=[],this.lineageBuffer.clear();}bufferLineageEvent(e){let t=this.lineageBuffer.get(e.callId)??[];t.push(e),this.lineageBuffer.set(e.callId,t);}flushLineageForCall(e){let t=this.lineageBuffer.get(e.callId)??[];this.lineageBuffer.delete(e.callId);let r={timestamp:e.ts,provider:e.provider,model:e.model,inputTokens:e.inputTokens,outputTokens:e.outputTokens,costUsd:e.costUsd,userCostUsd:e.userCostUsd,durationMs:e.durationMs,stopReason:e.stopReason,fallbackChain:e.fallbackChain,rateLimitHit:e.rateLimitHit,securityState:e.securityState,stitched:e.stitched,segments:e.segments,lineage:t.map(s=>s.id),promptContent:false,completionContent:false};this.report(r);}report(e){!this.enabled||!this.keyId||fetch(this.endpoint,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({keyId:this.keyId,provider:e.provider,model:e.model,inputTokens:e.inputTokens,outputTokens:e.outputTokens,costUsd:e.costUsd,...e.userCostUsd!==void 0?{userCostUsd:e.userCostUsd}:{},durationMs:e.durationMs,fallback:e.fallbackChain.length>1,timestamp:e.timestamp,...e.stitched?{stitched:true,segmentCount:e.segments?.length??0}:{},...e.lineage?{lineageIds:e.lineage}:{}})}).catch(()=>{});}};});async function ce(n,e){if(!n)return {valid:false,error:"License key required. Get one at inference-relay.com"};let t=U.get(n);if(t&&Date.now()-t.cachedAt<t.ttl)return t.result;let r=new AbortController,o=!U.has(n)?3e3:1e3,i=setTimeout(()=>r.abort(),o),a;try{let p=(e||Xt)+Qt;a=await(await fetch(p,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({key:n}),signal:r.signal})).json();}catch{return clearTimeout(i),t&&Date.now()-t.cachedAt<en?t.result:{valid:false,error:"License validation failed (network error). Visit inference-relay.com"}}clearTimeout(i);let l;if(a.jws)try{let{compactVerify:p,importSPKI:c}=await import('jose'),d=await c(tn,"RS256"),{payload:f}=await p(a.jws,d);l=JSON.parse(new TextDecoder().decode(f));}catch{return U.delete(n),{valid:false,error:"License validation signature invalid. Possible MITM \u2014 refusing to start."}}else if(l=a,l.valid)return U.delete(n),{valid:false,error:"License validation response not signed. Possible MITM."};return U.set(n,{result:l,cachedAt:Date.now(),ttl:l.valid?Zt:6e4}),l}var Xt,Qt,Zt,en,tn,U,ue,ht=C(()=>{ge();Ve();Ke();ze();Xe();we();Z();lt();pt();_e();mt();Xt="https://api.inference-relay.com",Qt="/v1/validate",Zt=1440*60*1e3,en=10080*60*1e3,tn=`-----BEGIN PUBLIC KEY-----
|
|
17
17
|
MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEAzk88Bt9LZSmuSd8ulKqy
|
|
18
18
|
EGNiPB+r8u1ZinBvMM4qX8dp4kzJEC/47L1RYrrxCphD05HToREjljbj+upH6rZH
|
|
19
19
|
t6D/5funYPi2TO1IOSC6BRVRwMqqTmJGVJcLcG/WF3gTE+2UQ3q5EppWJCEKcz6x
|
|
@@ -21,4 +21,4 @@ kfynZEzOcLikoQALIxEjxRfi6g83wNJS+WBwAJQ68s7N9FHE3rkLzEqv3FJRP+ZR
|
|
|
21
21
|
7lqwqIyuCLr9j9yc4XqF6VY9yooD2hyw/n1EVFejzjDgntWTqIkaCIXwh8lDzRrK
|
|
22
22
|
Rxn07gcI6ea2Eym9sXh4xZLFudXfAT2DFsH/XboDUuMKAV1YHc1NsIGApesWe2M+
|
|
23
23
|
UQIDAQAB
|
|
24
|
-
-----END PUBLIC KEY-----`,O=new Map;ce=class n{providers=[];config;fallbackEngine;telemetryReporter;routingFn=null;license=null;licensePromise=null;licenseValidatedAt=0;callsSinceValidation=0;manifestDecryptor=new V;capProbeTimer=null;capProbeTicksRemaining=0;messages;constructor(e){if(this.config=e,this.fallbackEngine=new ae,this.telemetryReporter=new oe({endpoint:e.endpoint?`${e.endpoint}/v1/events`:void 0,enabled:e.telemetry!==false,keyId:e.licenseKey}),this.telemetryReporter.subscribeToEngine(this.fallbackEngine),e.routing&&(this.routingFn=it(e.routing)),e.providers)for(let s of e.providers){let a=typeof s=="string"?{type:s}:s,i=this.createProvider(a);i&&(a.priority!==void 0&&(i.priority=a.priority),this.providers.push(i));}let t=100,r=3600*1e3;this.messages={create:async(s,a)=>{let i=this.license&&(this.callsSinceValidation>=t||Date.now()-this.licenseValidatedAt>=r);if(i&&(this.license=null),!this.license){this.licensePromise||(this.licensePromise=le(this.config.licenseKey,this.config.endpoint));try{this.license=await this.licensePromise;}finally{this.licensePromise=null;}if(this.licenseValidatedAt=Date.now(),this.callsSinceValidation=0,this.license.valid){if(this.license.capExceeded)throw this.startCapProbe(),new U({tier:this.license.tier||"solo",baseCap:this.license.baseCap??this.license.usageCap??null,topUps:this.license.topUpGrants??0,effectiveCap:this.license.usageCap??null,usedThisCycle:this.license.usageThisMonth??0});i||console.log(`[inference-relay] LICENSE AUTHORIZED [${(this.license.tier||"unknown").toUpperCase()} TIER] \u2014 READY.`);}else throw new Error(`[inference-relay] ${this.license.error||"License key invalid or expired. Visit inference-relay.com"}`)}return this.callsSinceValidation++,s.stream?this.streamMessage(s,a):this.executeMessage(s,a)}};}startCapProbe(){if(this.capProbeTimer!==null)return;let e=15e3,t=20,r=4;this.capProbeTicksRemaining=t;let s=async()=>{this.capProbeTicksRemaining-=1;let a=Math.ceil(this.capProbeTicksRemaining*e/6e4);this.capProbeTicksRemaining>0&&this.capProbeTicksRemaining%r===0&&console.log(`[inference-relay] cap exceeded \u2014 checking for top-up every 15s (${a}m remaining)`),this.config.licenseKey&&O.delete(this.config.licenseKey);try{let i=await this.probeValidate();if(i.valid&&i.capExceeded===!1){console.log("[inference-relay] cap cleared \u2014 probe exiting, inference resuming"),this.license=i,this.licenseValidatedAt=Date.now(),this.callsSinceValidation=0,this.stopCapProbe();return}}catch{}this.capProbeTicksRemaining<=0&&(console.log("[inference-relay] cap probe exhausted (5 min) \u2014 stopping"),this.stopCapProbe());};this.capProbeTimer=setInterval(s,e),typeof this.capProbeTimer.unref=="function"&&this.capProbeTimer.unref();}async probeValidate(){return le(this.config.licenseKey,this.config.endpoint)}stopCapProbe(){this.capProbeTimer!==null&&(clearInterval(this.capProbeTimer),this.capProbeTimer=null),this.capProbeTicksRemaining=0;}static async autoDetect(e){let t=[],r=await N();return r.available&&r.authenticated!==false&&t.push({type:"claude-cli",priority:1}),process.env.ANTHROPIC_API_KEY&&t.push({type:"anthropic-api",priority:2,apiKey:process.env.ANTHROPIC_API_KEY}),new n({licenseKey:e?.licenseKey||process.env.IR_LICENSE_KEY||"",providers:t,fallback:e?.fallback??true,...e})}dispose(){this.providers=[],this.config.licenseKey&&O.delete(this.config.licenseKey),this.license=null,this.licensePromise=null,this.callsSinceValidation=0,this.stopCapProbe(),this.telemetryReporter.unsubscribe(),this.manifestDecryptor.clearCache(),Me();}resolveProviders(e){if(!this.routingFn)return this.providers;let t=ct(e),r=this.routingFn(e,t);return st(this.providers,r)}async executeMessage(e,t){let r=this.resolveProviders(e);return this.fallbackEngine.execute(e,r,{signal:t?.signal,licenseState:this.manifestDecryptor.getSecurityState()})}async streamMessage(e,t){let r=this.resolveProviders(e);return this.fallbackEngine.stream(e,r,{signal:t?.signal,licenseState:this.manifestDecryptor.getSecurityState()})}createProvider(e){switch(e.type){case "claude-cli":return new Z(this.config.licenseKey);case "anthropic-api":return e.apiKey?new te({apiKey:e.apiKey,baseUrl:e.baseUrl}):null;case "openai-api":return e.apiKey?new ne({apiKey:e.apiKey,baseUrl:e.baseUrl}):null;case "ollama":return new re({baseUrl:e.baseUrl});default:return null}}};});var mt=C(()=>{Q();});var ft={};xe(ft,{runVerify:()=>sn});function S(n,e){let t=n.padEnd(13);console.log(`${v("[ir]")} ${t}${e}`);}function rn(n){if(n instanceof U)return {label:"CAP_EXCEEDED",hint:"Monthly cap reached. Buy a top-up at https://inference-relay.com/dashboard/billing",code:1};if(n instanceof F)return {label:"ALL_PROVIDERS_FAILED",hint:`Fallback chain exhausted: ${n.fallbackChain.join(" \u2192 ")}`,code:1};if(!(n instanceof Error))return {label:"FAILED",hint:String(n).slice(0,120),code:1};let e=n.message.toLowerCase(),t=n.name.toLowerCase();return e.includes("keychain_access_denied")||e.includes("keychain access denied")?{label:"KEYCHAIN_DENIED",hint:'macOS denied keychain access. Re-run and click "Always Allow" on the prompt.',code:1}:e.includes("rate limit")||e.includes("429")||t.includes("ratelimit")?{label:"RATE_LIMITED",hint:"Claude subscription is currently rate-limited. Wait a minute and re-run.",code:1}:e.includes("auth")||e.includes("401")||e.includes("403")||e.includes("unauthorized")||e.includes("forbidden")?{label:"AUTH_FAILED",hint:"Claude CLI OAuth token rejected. Run `claude` to refresh, then re-run verify.",code:1}:e.includes("timeout")||e.includes("timed out")||e.includes("aborted")||e.includes("etimedout")||e.includes("econnrefused")||e.includes("econnreset")||t==="aborterror"||t==="timeouterror"?{label:"TIMEOUT",hint:`No response within ${gt/1e3}s. Network issue, hung CLI, or keychain prompt waiting for input. Try \`curl -I api.anthropic.com\` to test connectivity.`,code:1}:e.includes("empty response")||e.includes("returned empty")?{label:"EMPTY_RESPONSE",hint:"CLI returned an empty body \u2014 possible truncation, model error, or stale CLI version.",code:1}:{label:"FAILED",hint:n.message.slice(0,120),code:1}}function ht(n){if(n.usageCap!==null&&n.usageCap!==void 0){let e=n.usageThisMonth??0,t=n.usageCap;if(t===0)console.log(v(` Usage: ${e.toLocaleString()} (unmetered)`));else {let r=Math.round(e/t*1e3)/10;console.log(v(` Usage: ${e.toLocaleString()} / ${t.toLocaleString()} (${r}%)`));}}}async function sn(n=process.argv.slice(2),e=process.env){let t=n.includes("--credentials-only"),r=e.IR_LICENSE_KEY;if(!r)return console.error(Y("[ir] IR_LICENSE_KEY environment variable is not set.")),console.error(" Get a key at https://inference-relay.com/pricing"),console.error(" Then: export IR_LICENSE_KEY=ir_live_xxxx"),2;if(!/^ir_(live|test)_[a-f0-9]{32}$/.test(r))return console.error(Y("[ir] IR_LICENSE_KEY does not look like a valid key.")),console.error(" Expected format: ir_live_<32 hex chars>"),console.error(` Got: ${r.slice(0,8)}... (${r.length} chars)`),2;let s;try{s=await le(r);}catch(E){return S("IDENTITY:",Y("FAILED")),console.error(` ${E.message||"unknown error"}`),1}if(!s.valid)return S("IDENTITY:",Y("REJECTED")),s.error&&console.error(` ${s.error}`),1;S("IDENTITY:",G("VALIDATED"));let a=(s.tier??"unknown").toUpperCase();S("TIER:",G(a));let i;try{i=await N();}catch{i=null;}if(!i||!i.available)return S("RELAY:",ue("INACTIVE")),console.log(v(" Claude CLI not detected. The relay will fall back")),console.log(v(" to direct API providers using your API keys.")),console.log(v(" For zero-cost routing, install Claude Code:")),console.log(v(" npm install -g @anthropic-ai/claude-code")),1;if(!i.authenticated)return S("RELAY:",ue("CLI_UNAUTHED")),console.log(v(" Claude CLI is installed but not authenticated.")),console.log(v(" Run `claude` once to sign in, then re-run verify.")),1;if(i.versionWarning&&console.log(ue(` Warning: ${i.versionWarning}`)),t)return S("RELAY:",G("READY")),console.log(v(" Credentials present. Round-trip skipped (--credentials-only).")),i.cliVersion&&console.log(v(` Claude CLI: v${i.cliVersion}`)),i.subscriptionType&&i.subscriptionType!=="unknown"&&console.log(v(` Tier: ${i.subscriptionType}`)),ht(s),0;let l=new ce({licenseKey:r,providers:[{type:"claude-cli",priority:1}],fallback:false,telemetry:false}),o=new AbortController,p=setTimeout(()=>o.abort(),gt),c=Date.now(),d;try{d=await l.messages.create(nn,{signal:o.signal});}catch(E){clearTimeout(p),l.dispose();let{label:u,hint:y}=rn(E);return S("RELAY:",Y(`DEGRADED \u2014 ${u}`)),console.log(v(` ${y}`)),i.cliVersion&&console.log(v(` Claude CLI: v${i.cliVersion}`)),1}clearTimeout(p),l.dispose();let h=Date.now()-c;return d.fallbackChain?.[0]?.startsWith("claude-cli:success")?(S("RELAY:",G("ACTIVE")),console.log(v(` Round-trip: ${h}ms via ${d.provider}`)),console.log(v(` Model echo: ${d.model}`)),console.log(v(` Tokens: ${d.usage.input_tokens} in / ${d.usage.output_tokens} out`)),i.cliVersion&&console.log(v(` Claude CLI: v${i.cliVersion}`)),console.log(v(` Multimodal: ${G("\u2713")} via claude-cli (subscription)`)),i.subscriptionType&&i.subscriptionType!=="unknown"&&console.log(v(` Tier: ${i.subscriptionType}`)),ht(s),0):(S("RELAY:",ue("UNEXPECTED")),console.log(v(` Native Gateway did not handle the call. Chain: ${d.fallbackChain?.join(" \u2192 ")}`)),1)}var tn,de,G,Y,ue,v,gt,nn,yt=C(()=>{pt();Ie();me();mt();tn=process.stdout.isTTY===true,de=(n,e)=>tn?`\x1B[${n}m${e}\x1B[0m`:e,G=n=>de("32",n),Y=n=>de("31",n),ue=n=>de("33",n),v=n=>de("2",n);gt=15e3,nn={model:"claude-haiku-4-5",max_tokens:1,messages:[{role:"user",content:"Reply with exactly the single character k and nothing else."}]};});var _t={verify:async()=>{let{runVerify:n}=await Promise.resolve().then(()=>(yt(),ft));return n()}};function Et(){console.error("Usage: inference-relay <command>"),console.error(""),console.error("Commands:"),console.error(" verify Smoke-test your installation (license + relay availability)"),console.error(""),console.error("Docs: https://inference-relay.com/docs");}var M=process.argv[2];(!M||M==="--help"||M==="-h")&&(Et(),process.exit(M?0:2));M in _t||(console.error(`Unknown command: ${M}`),console.error(""),Et(),process.exit(1));_t[M]().then(n=>process.exit(n)).catch(n=>{console.error(`[ir] failed to run command "${M}":`),console.error(n),process.exit(99);});
|
|
24
|
+
-----END PUBLIC KEY-----`,U=new Map;ue=class n{providers=[];config;fallbackEngine;telemetryReporter;routingFn=null;license=null;licensePromise=null;licenseValidatedAt=0;callsSinceValidation=0;manifestDecryptor=new G;capProbeTimer=null;capProbeTicksRemaining=0;messages;constructor(e){if(this.config=e,this.fallbackEngine=new ae,this.telemetryReporter=new le({endpoint:e.endpoint?`${e.endpoint}/v1/events`:void 0,enabled:e.telemetry!==false,keyId:e.licenseKey}),this.telemetryReporter.subscribeToEngine(this.fallbackEngine),e.routing&&(this.routingFn=at(e.routing)),e.providers)for(let s of e.providers){let o=typeof s=="string"?{type:s}:s,i=this.createProvider(o);i&&(o.priority!==void 0&&(i.priority=o.priority),this.providers.push(i));}let t=100,r=3600*1e3;this.messages={create:async(s,o)=>{let i=this.license&&(this.callsSinceValidation>=t||Date.now()-this.licenseValidatedAt>=r);if(i&&(this.license=null),!this.license){this.licensePromise||(this.licensePromise=ce(this.config.licenseKey,this.config.endpoint));try{this.license=await this.licensePromise;}finally{this.licensePromise=null;}if(this.licenseValidatedAt=Date.now(),this.callsSinceValidation=0,this.license.valid){if(this.license.capExceeded)throw this.startCapProbe(),new N({tier:this.license.tier||"solo",baseCap:this.license.baseCap??this.license.usageCap??null,topUps:this.license.topUpGrants??0,effectiveCap:this.license.usageCap??null,usedThisCycle:this.license.usageThisMonth??0});i||console.log(`[inference-relay] LICENSE AUTHORIZED [${(this.license.tier||"unknown").toUpperCase()} TIER] \u2014 READY.`);}else throw new Error(`[inference-relay] ${this.license.error||"License key invalid or expired. Visit inference-relay.com"}`)}return this.callsSinceValidation++,s.stream?this.streamMessage(s,o):this.executeMessage(s,o)}};}startCapProbe(){if(this.capProbeTimer!==null)return;let e=15e3,t=20,r=4;this.capProbeTicksRemaining=t;let s=async()=>{this.capProbeTicksRemaining-=1;let o=Math.ceil(this.capProbeTicksRemaining*e/6e4);this.capProbeTicksRemaining>0&&this.capProbeTicksRemaining%r===0&&console.log(`[inference-relay] cap exceeded \u2014 checking for top-up every 15s (${o}m remaining)`),this.config.licenseKey&&U.delete(this.config.licenseKey);try{let i=await this.probeValidate();if(i.valid&&i.capExceeded===!1){console.log("[inference-relay] cap cleared \u2014 probe exiting, inference resuming"),this.license=i,this.licenseValidatedAt=Date.now(),this.callsSinceValidation=0,this.stopCapProbe();return}}catch{}this.capProbeTicksRemaining<=0&&(console.log("[inference-relay] cap probe exhausted (5 min) \u2014 stopping"),this.stopCapProbe());};this.capProbeTimer=setInterval(s,e),typeof this.capProbeTimer.unref=="function"&&this.capProbeTimer.unref();}async probeValidate(){return ce(this.config.licenseKey,this.config.endpoint)}stopCapProbe(){this.capProbeTimer!==null&&(clearInterval(this.capProbeTimer),this.capProbeTimer=null),this.capProbeTicksRemaining=0;}static async autoDetect(e){let t=[],r=await F();return r.available&&r.authenticated!==false&&t.push({type:"claude-cli",priority:1}),process.env.ANTHROPIC_API_KEY&&t.push({type:"anthropic-api",priority:2,apiKey:process.env.ANTHROPIC_API_KEY}),new n({licenseKey:e?.licenseKey||process.env.IR_LICENSE_KEY||"",providers:t,fallback:e?.fallback??true,...e})}dispose(){this.providers=[],this.config.licenseKey&&U.delete(this.config.licenseKey),this.license=null,this.licensePromise=null,this.callsSinceValidation=0,this.stopCapProbe(),this.telemetryReporter.unsubscribe(),this.manifestDecryptor.clearCache(),Oe();}resolveProviders(e){if(!this.routingFn)return this.providers;let t=dt(e),r=this.routingFn(e,t);return ot(this.providers,r)}async executeMessage(e,t){let r=this.resolveProviders(e);return this.fallbackEngine.execute(e,r,{signal:t?.signal,licenseState:this.manifestDecryptor.getSecurityState()})}async streamMessage(e,t){let r=this.resolveProviders(e);return this.fallbackEngine.stream(e,r,{signal:t?.signal,licenseState:this.manifestDecryptor.getSecurityState()})}createProvider(e){switch(e.type){case "claude-cli":return new ee(this.config.licenseKey);case "anthropic-api":return e.apiKey?new ne({apiKey:e.apiKey,baseUrl:e.baseUrl}):null;case "openai-api":return e.apiKey?new re({apiKey:e.apiKey,baseUrl:e.baseUrl}):null;case "ollama":return new se({baseUrl:e.baseUrl});default:return null}}};});var gt=C(()=>{Z();});var yt={};De(yt,{runVerify:()=>ln});function T(n,e){let t=n.padEnd(13);console.log(`${_("[ir]")} ${t}${e}`);}function pe(n){if(n instanceof N)return {label:"CAP_EXCEEDED",hint:"Monthly cap reached. Buy a top-up at https://inference-relay.com/dashboard/billing",code:1};if(n instanceof H)return {label:"ALL_PROVIDERS_FAILED",hint:`Fallback chain exhausted: ${n.fallbackChain.join(" \u2192 ")}`,code:1};if(!(n instanceof Error))return {label:"FAILED",hint:String(n).slice(0,120),code:1};let e=n.message.toLowerCase(),t=n.name.toLowerCase();return e.includes("keychain_access_denied")||e.includes("keychain access denied")?{label:"KEYCHAIN_DENIED",hint:'macOS denied keychain access. Re-run and click "Always Allow" on the prompt.',code:1}:e.includes("rate limit")||e.includes("429")||t.includes("ratelimit")?{label:"RATE_LIMITED",hint:"Claude subscription is currently rate-limited. Wait a minute and re-run.",code:1}:e.includes("auth")||e.includes("401")||e.includes("403")||e.includes("unauthorized")||e.includes("forbidden")?{label:"AUTH_FAILED",hint:"Claude CLI OAuth token rejected. Run `claude` to refresh, then re-run verify.",code:1}:e.includes("timeout")||e.includes("timed out")||e.includes("aborted")||e.includes("etimedout")||e.includes("econnrefused")||e.includes("econnreset")||t==="aborterror"||t==="timeouterror"?{label:"TIMEOUT",hint:`No response within ${K/1e3}s. Network issue, hung CLI, or keychain prompt waiting for input. Try \`curl -I api.anthropic.com\` to test connectivity.`,code:1}:e.includes("empty response")||e.includes("returned empty")?{label:"EMPTY_RESPONSE",hint:"CLI returned an empty body \u2014 possible truncation, model error, or stale CLI version.",code:1}:{label:"FAILED",hint:n.message.slice(0,120),code:1}}function ft(n){if(n.usageCap!==null&&n.usageCap!==void 0){let e=n.usageThisMonth??0,t=n.usageCap;if(t===0)console.log(_(` Usage: ${e.toLocaleString()} (unmetered)`));else {let r=Math.round(e/t*1e3)/10;console.log(_(` Usage: ${e.toLocaleString()} / ${t.toLocaleString()} (${r}%)`));}}}async function ln(n=process.argv.slice(2),e=process.env){let t=n.includes("--credentials-only"),r=e.IR_LICENSE_KEY;if(!r)return console.error(D("[ir] IR_LICENSE_KEY environment variable is not set.")),console.error(" Get a key at https://inference-relay.com/pricing"),console.error(" Then: export IR_LICENSE_KEY=ir_live_xxxx"),2;if(!/^ir_(live|test)_[a-f0-9]{32}$/.test(r))return console.error(D("[ir] IR_LICENSE_KEY does not look like a valid key.")),console.error(" Expected format: ir_live_<32 hex chars>"),console.error(` Got: ${r.slice(0,8)}... (${r.length} chars)`),2;let s;try{s=await ce(r);}catch(b){return T("IDENTITY:",D("FAILED")),console.error(` ${b.message||"unknown error"}`),1}if(!s.valid)return T("IDENTITY:",D("REJECTED")),s.error&&console.error(` ${s.error}`),1;T("IDENTITY:",W("VALIDATED"));let o=(s.tier??"unknown").toUpperCase();T("TIER:",W(o));let i;try{i=await F();}catch{i=null;}if(!i||!i.available)return T("RELAY:",de("INACTIVE")),console.log(_(" Claude CLI not detected. The relay will fall back")),console.log(_(" to direct API providers using your API keys.")),console.log(_(" For zero-cost routing, install Claude Code:")),console.log(_(" npm install -g @anthropic-ai/claude-code")),1;if(!i.authenticated)return T("RELAY:",de("CLI_UNAUTHED")),console.log(_(" Claude CLI is installed but not authenticated.")),console.log(_(" Run `claude` once to sign in, then re-run verify.")),1;if(i.versionWarning&&console.log(de(` Warning: ${i.versionWarning}`)),t)return T("RELAY:",W("READY")),console.log(_(" Credentials present. Round-trip skipped (--credentials-only).")),i.cliVersion&&console.log(_(` Claude CLI: v${i.cliVersion}`)),i.subscriptionType&&i.subscriptionType!=="unknown"&&console.log(_(` Tier: ${i.subscriptionType}`)),ft(s),0;let a=new ue({licenseKey:r,providers:[{type:"claude-cli",priority:1}],fallback:false,telemetry:false}),l=new AbortController,p=setTimeout(()=>l.abort(),K),c=Date.now(),d;try{d=await a.messages.create(rn,{signal:l.signal});}catch(b){clearTimeout(p),a.dispose();let{label:u,hint:y}=pe(b);return T("RELAY:",D(`DEGRADED \u2014 ${u}`)),console.log(_(` ${y}`)),i.cliVersion&&console.log(_(` Claude CLI: v${i.cliVersion}`)),1}if(clearTimeout(p),process.platform==="win32"){let b=new AbortController,u=setTimeout(()=>b.abort(),K);try{await a.messages.create(sn,{signal:b.signal});}catch(E){clearTimeout(u),a.dispose();let{hint:k}=pe(E);return T("RELAY:",D("DEGRADED \u2014 LONG_PROMPT_FAILED")),console.log(_(` Long-system-prompt round-trip failed: ${k}`)),console.log(_(" This is the cmd.exe arg-length regression class.")),console.log(_(" Ensure inference-relay is at version >=1.5.3 (closes the bug).")),i.cliVersion&&console.log(_(` Claude CLI: v${i.cliVersion}`)),1}clearTimeout(u);let y=new AbortController,h=setTimeout(()=>y.abort(),K);try{await a.messages.create(on,{signal:y.signal});}catch(E){clearTimeout(h),a.dispose();let{hint:k}=pe(E);return T("RELAY:",D("DEGRADED \u2014 TOOLS_DISPATCH_FAILED")),console.log(_(` Tools round-trip failed: ${k}`)),console.log(_(" The claude-cli provider couldn't forward params.tools.")),console.log(_(" Ensure inference-relay is at version >=1.6.0 (closes the bug).")),i.cliVersion&&console.log(_(` Claude CLI: v${i.cliVersion}`)),1}clearTimeout(h);let g=new AbortController,R=setTimeout(()=>g.abort(),K);try{await a.messages.create(an,{signal:g.signal});}catch(E){clearTimeout(R),a.dispose();let{hint:k}=pe(E);return T("RELAY:",D("DEGRADED \u2014 MULTITURN_DISPATCH_FAILED")),console.log(_(` Multi-turn round-trip failed: ${k}`)),console.log(_(" The CLI rejected an assistant-role stream-json envelope.")),console.log(_(" Ensure inference-relay is at version >=1.6.0 (closes the bug),")),console.log(_(" OR ensure Claude CLI version is >=2.1.x (stream-json input floor).")),i.cliVersion&&console.log(_(` Claude CLI: v${i.cliVersion}`)),1}clearTimeout(R);}a.dispose();let f=Date.now()-c;return d.fallbackChain?.[0]?.startsWith("claude-cli:success")?(T("RELAY:",W("ACTIVE")),console.log(_(` Round-trip: ${f}ms via ${d.provider}`)),console.log(_(` Model echo: ${d.model}`)),console.log(_(` Tokens: ${d.usage.input_tokens} in / ${d.usage.output_tokens} out`)),i.cliVersion&&console.log(_(` Claude CLI: v${i.cliVersion}`)),console.log(_(` Multimodal: ${W("\u2713")} via claude-cli (subscription)`)),i.subscriptionType&&i.subscriptionType!=="unknown"&&console.log(_(` Tier: ${i.subscriptionType}`)),ft(s),0):(T("RELAY:",de("UNEXPECTED")),console.log(_(` Native Gateway did not handle the call. Chain: ${d.fallbackChain?.join(" \u2192 ")}`)),1)}var nn,me,W,D,de,_,K,rn,sn,on,an,_t=C(()=>{ht();we();ge();gt();nn=process.stdout.isTTY===true,me=(n,e)=>nn?`\x1B[${n}m${e}\x1B[0m`:e,W=n=>me("32",n),D=n=>me("31",n),de=n=>me("33",n),_=n=>me("2",n);K=15e3,rn={model:"claude-haiku-4-5",max_tokens:1,messages:[{role:"user",content:"Reply with exactly the single character k and nothing else."}]},sn={model:"claude-haiku-4-5",max_tokens:1,system:"Sentence. ".repeat(800),messages:[{role:"user",content:"Reply with exactly the single character k and nothing else."}]},on={model:"claude-haiku-4-5",max_tokens:50,messages:[{role:"user",content:"Reply with the single word: ok"}],tools:[{type:"web_search_20250305",name:"web_search",max_uses:1}]},an={model:"claude-haiku-4-5",max_tokens:10,messages:[{role:"user",content:"Reply with: hello"},{role:"assistant",content:[{type:"text",text:"hello"}]},{role:"user",content:"Reply with: ok"}]};});var Et={verify:async()=>{let{runVerify:n}=await Promise.resolve().then(()=>(_t(),yt));return n()}};function bt(){console.error("Usage: inference-relay <command>"),console.error(""),console.error("Commands:"),console.error(" verify Smoke-test your installation (license + relay availability)"),console.error(""),console.error("Docs: https://inference-relay.com/docs");}var M=process.argv[2];(!M||M==="--help"||M==="-h")&&(bt(),process.exit(M?0:2));M in Et||(console.error(`Unknown command: ${M}`),console.error(""),bt(),process.exit(1));Et[M]().then(n=>process.exit(n)).catch(n=>{console.error(`[ir] failed to run command "${M}":`),console.error(n),process.exit(99);});
|
package/dist/index.cjs
CHANGED
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
'use strict';var child_process=require('child_process'),fs=require('fs'),path=require('path'),os=require('os'),crypto$1=require('crypto'),vm=require('vm'),module$1=require('module'),events=require('events');/*! inference-relay (c) 2026 L2B II LLC. Patent pending. Proprietary. See LICENSE. */
|
|
2
|
-
var Je=Object.defineProperty;var ze=(n=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(n,{get:(e,t)=>(typeof require<"u"?require:e)[t]}):n)(function(n){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+n+'" is not supported')});var Xe=(n,e)=>()=>(n&&(e=n(n=0)),e);var Qe=(n,e)=>{for(var t in e)Je(n,t,{get:e[t],enumerable:true});};var Me={};Qe(Me,{detectCLI:()=>H});function dt(n){let e=process.env.PATH;if(!e)return null;let t=process.platform==="win32",r=t?";":":",s=t?(process.env.PATHEXT??".COM;.EXE;.BAT;.CMD").split(";").map(i=>i.toLowerCase()):[""];for(let i of e.split(r))if(i)for(let a of s){let o=path.join(i,n+a);if(fs.existsSync(o))return o}return null}async function H(){if(!dt("claude"))return {available:false,authenticated:false};let e=null;try{let r=child_process.execSync("claude --version",{encoding:"utf-8",timeout:5e3}).trim(),s=r.match(/(\d+\.\d+\.\d+)/);e=s?s[1]:r;}catch{}let t;return e&&!ut.some(r=>e.startsWith(r.split(".").slice(0,2).join(".")))&&(t=`CLI version ${e} is not tested with this version of inference-relay. Streaming may be unstable.`),{available:true,authenticated:true,cliVersion:e||void 0,versionWarning:t}}var ut,ee=Xe(()=>{ut=["2.1.87"];});var U=class n extends Error{code="RELAY_CAP_EXCEEDED";tier;baseCap;topUps;effectiveCap;usedThisCycle;resolutionUrl;constructor(e){let t=e.resolutionUrl??"https://inference-relay.com/dashboard/billing?topup=1";super(`Cap reached (${e.usedThisCycle} / ${e.effectiveCap??"\u221E"} calls this cycle). Buy a top-up or upgrade at ${t}`),this.name="RelayCapExceededError",this.tier=e.tier,this.baseCap=e.baseCap,this.topUps=e.topUps,this.effectiveCap=e.effectiveCap,this.usedThisCycle=e.usedThisCycle,this.resolutionUrl=t,Object.setPrototypeOf(this,n.prototype);}};function L(n,e){let t=new Map,r=null,s=null,i=new Promise((u,f)=>{r=u,s=f;});i.catch(()=>{});let a=[],o=null,l="",d="";function c(u,...f){let h=t.get(u);h&&h.forEach(g=>g(...f));}async function*p(){try{for await(let u of n)if(yield u,u.type==="content_block_start")a[u.index]=u.content_block;else if(u.type==="content_block_delta"){let f=a[u.index];f&&u.delta?.type==="text_delta"&&(f.text=(f.text||"")+u.delta.text,c("text",u.delta.text));}else if(u.type==="message_start")l=u.message?.model||"",o=u.message?.usage;else if(u.type==="message_delta")d=u.delta?.stop_reason||"",u.usage&&(o={...o,...u.usage});else if(u.type==="message_stop"){let f={role:"assistant",model:l,content:a.filter(Boolean),usage:o,stop_reason:d};c("message",f),r&&(r(f),r=null),e&&e(f);}if(r){let u={role:"assistant",model:l,content:a.filter(Boolean),usage:o,stop_reason:d||"end_turn"};c("message",u),r(u),r=null,e&&e(u);}}catch(u){throw c("error",u),s&&(s(u),s=null),u}}let y=p(),m=false,
|
|
2
|
+
var Je=Object.defineProperty;var ze=(n=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(n,{get:(e,t)=>(typeof require<"u"?require:e)[t]}):n)(function(n){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+n+'" is not supported')});var Xe=(n,e)=>()=>(n&&(e=n(n=0)),e);var Qe=(n,e)=>{for(var t in e)Je(n,t,{get:e[t],enumerable:true});};var Me={};Qe(Me,{detectCLI:()=>H});function dt(n){let e=process.env.PATH;if(!e)return null;let t=process.platform==="win32",r=t?";":":",s=t?(process.env.PATHEXT??".COM;.EXE;.BAT;.CMD").split(";").map(i=>i.toLowerCase()):[""];for(let i of e.split(r))if(i)for(let a of s){let o=path.join(i,n+a);if(fs.existsSync(o))return o}return null}async function H(){if(!dt("claude"))return {available:false,authenticated:false};let e=null;try{let r=child_process.execSync("claude --version",{encoding:"utf-8",timeout:5e3}).trim(),s=r.match(/(\d+\.\d+\.\d+)/);e=s?s[1]:r;}catch{}let t;return e&&!ut.some(r=>e.startsWith(r.split(".").slice(0,2).join(".")))&&(t=`CLI version ${e} is not tested with this version of inference-relay. Streaming may be unstable.`),{available:true,authenticated:true,cliVersion:e||void 0,versionWarning:t}}var ut,ee=Xe(()=>{ut=["2.1.87"];});var U=class n extends Error{code="RELAY_CAP_EXCEEDED";tier;baseCap;topUps;effectiveCap;usedThisCycle;resolutionUrl;constructor(e){let t=e.resolutionUrl??"https://inference-relay.com/dashboard/billing?topup=1";super(`Cap reached (${e.usedThisCycle} / ${e.effectiveCap??"\u221E"} calls this cycle). Buy a top-up or upgrade at ${t}`),this.name="RelayCapExceededError",this.tier=e.tier,this.baseCap=e.baseCap,this.topUps=e.topUps,this.effectiveCap=e.effectiveCap,this.usedThisCycle=e.usedThisCycle,this.resolutionUrl=t,Object.setPrototypeOf(this,n.prototype);}};function L(n,e){let t=new Map,r=null,s=null,i=new Promise((u,f)=>{r=u,s=f;});i.catch(()=>{});let a=[],o=null,l="",d="";function c(u,...f){let h=t.get(u);h&&h.forEach(g=>g(...f));}async function*p(){try{for await(let u of n)if(yield u,u.type==="content_block_start")a[u.index]=u.content_block;else if(u.type==="content_block_delta"){let f=a[u.index];f&&u.delta?.type==="text_delta"&&(f.text=(f.text||"")+u.delta.text,c("text",u.delta.text));}else if(u.type==="message_start")l=u.message?.model||"",o=u.message?.usage;else if(u.type==="message_delta")d=u.delta?.stop_reason||"",u.usage&&(o={...o,...u.usage});else if(u.type==="message_stop"){let f={role:"assistant",model:l,content:a.filter(Boolean),usage:o,stop_reason:d};c("message",f),r&&(r(f),r=null),e&&e(f);}if(r){let u={role:"assistant",model:l,content:a.filter(Boolean),usage:o,stop_reason:d||"end_turn"};c("message",u),r(u),r=null,e&&e(u);}}catch(u){throw c("error",u),s&&(s(u),s=null),u}}let y=p(),m=false,E={[Symbol.asyncIterator](){if(m)throw new Error("[inference-relay] RelayStream can only be iterated once");return m=true,y},on(u,f){return t.has(u)||t.set(u,[]),t.get(u).push(f),E},finalMessage(){let u,f=new Promise((h,g)=>{u=setTimeout(()=>g(new Error("[inference-relay] finalMessage() timed out \u2014 no message_stop received after 5 minutes")),3e5);});return Promise.race([i,f]).finally(()=>clearTimeout(u))}};return E}var R=class{priority=100;async streamRaw(e,t){throw new Error(`[inference-relay] Provider '${this.type}' does not implement streamRaw(). Override either stream() or streamRaw() in the provider class.`)}async stream(e,t){return L(await this.streamRaw(e,t))}};var me=`-----BEGIN PUBLIC KEY-----
|
|
3
3
|
MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEAzk88Bt9LZSmuSd8ulKqy
|
|
4
4
|
EGNiPB+r8u1ZinBvMM4qX8dp4kzJEC/47L1RYrrxCphD05HToREjljbj+upH6rZH
|
|
5
5
|
t6D/5funYPi2TO1IOSC6BRVRwMqqTmJGVJcLcG/WF3gTE+2UQ3q5EppWJCEKcz6x
|
|
@@ -7,10 +7,10 @@ kfynZEzOcLikoQALIxEjxRfi6g83wNJS+WBwAJQ68s7N9FHE3rkLzEqv3FJRP+ZR
|
|
|
7
7
|
7lqwqIyuCLr9j9yc4XqF6VY9yooD2hyw/n1EVFejzjDgntWTqIkaCIXwh8lDzRrK
|
|
8
8
|
Rxn07gcI6ea2Eym9sXh4xZLFudXfAT2DFsH/XboDUuMKAV1YHc1NsIGApesWe2M+
|
|
9
9
|
UQIDAQAB
|
|
10
|
-
-----END PUBLIC KEY-----`,pe={version:"1.0.0",cliVersion:"2.1.87",eventTypes:{system:"system",streamEvent:"stream_event",assistant:"assistant",result:"result"},fieldMappings:{sessionId:"session_id",model:"model",cliVersion:"claude_code_version",content:"content",usage:"usage",inputTokens:"input_tokens",outputTokens:"output_tokens",stopReason:"stop_reason",isError:"is_error",totalCostUsd:"total_cost_usd",durationMs:"duration_ms",error:"error",event:"event"},syntheticModel:"<synthetic>"},M=class{verificationFailures=0;cachedManifest=null;currentSecurityState="SEC_NOMINAL";async verifyManifest(e){try{let{compactVerify:t,importSPKI:r}=await import('jose'),s=await r(me,"RS256"),{payload:i}=await t(e,s),a=JSON.parse(new TextDecoder().decode(i));return this.verificationFailures=0,this.cachedManifest=a,this.currentSecurityState="SEC_NOMINAL",{manifest:a,securityState:"SEC_NOMINAL"}}catch{if(this.verificationFailures++,this.verificationFailures>=3){this.currentSecurityState="SEC_DEGRADED";let t=2e3+Math.random()*3e3;return await new Promise(r=>setTimeout(r,t)),console.warn("[inference-relay] Logic synchronization mismatch. Routing via local recovery."),{manifest:pe,securityState:"SEC_DEGRADED"}}return this.cachedManifest?(this.currentSecurityState="SEC_CACHE",{manifest:this.cachedManifest,securityState:"SEC_CACHE"}):(this.currentSecurityState="SEC_RECOVERY",{manifest:pe,securityState:"SEC_RECOVERY"})}}getSecurityState(){return this.currentSecurityState}clearCache(){this.cachedManifest=null,this.verificationFailures=0,this.currentSecurityState="SEC_NOMINAL";}},Ze=new M;function ke(){Ze.clearCache();}function et(){return pe}function tt(n,e){return e.split(".").reduce((t,r)=>{if(!(r==="__proto__"||r==="constructor"||r==="prototype"))return t&&t[r]},n)}var at=process.env.IR_GATEWAY_URL||"https://api.inference-relay.com";function ge(n){return path.join(os.homedir(),".inference-relay","gateway",n)}function Te(n){return path.join(ge(n),"bundle.js")}function xe(n){return path.join(ge(n),"meta.json")}var b=class extends Error{constructor(t,r,s){super(t);this.code=r;this.cause=s;this.name="GatewayLoaderError";}code;cause},F=new Map,B=class{static async load(e,t,r=at){let s=`${t}:${e.slice(-8)}`;if(F.has(s))return F.get(s);try{let p=await this.tryLoadFromDisk(t);if(p)return F.set(s,p),p}catch{}let{bytes:i,sha256:a,signature:o,libraryVersion:l}=await this.fetchFromAuthority(e,t,r);if(crypto$1.createHash("sha256").update(i).digest("hex")!==a)throw new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: HASH_MISMATCH)","GATEWAY_VERIFICATION_FAILED");await this.verifySignature(o,l,a);try{let p=ge(t);fs.mkdirSync(p,{recursive:!0}),fs.writeFileSync(Te(t),i),fs.writeFileSync(xe(t),JSON.stringify({sha256:a,signature:o,libraryVersion:l}));}catch{}let c=this.loadIntoVm(i.toString("utf-8"));return F.set(s,c),c}static async loadFromDisk(e){if(!fs.existsSync(e))throw new b(`[ir-system]: GATEWAY_LOAD_FAILED (REASON: BUNDLE_NOT_FOUND at ${e})`,"GATEWAY_LOAD_FAILED");let t=fs.readFileSync(e,"utf-8");return this.loadIntoVm(t)}static async tryLoadFromDisk(e){let t=Te(e),r=xe(e);if(!fs.existsSync(t)||!fs.existsSync(r))return null;let s=fs.readFileSync(t),i=JSON.parse(fs.readFileSync(r,"utf-8"));return crypto$1.createHash("sha256").update(s).digest("hex")!==i.sha256?null:(await this.verifySignature(i.signature,i.libraryVersion,i.sha256),this.loadIntoVm(s.toString("utf-8")))}static async fetchFromAuthority(e,t,r){let s,i=[0,1e3,4e3],a;for(let c of i){c>0&&await new Promise(p=>setTimeout(p,c));try{let p=await fetch(`${r}/v1/gateway/fetch`,{method:"POST",headers:{"Content-Type":"application/json",Authorization:`Bearer ${e}`},body:JSON.stringify({libraryVersion:t})});if(p.status===401||p.status===403)throw new b("[ir-system]: LICENSE_REJECTED (REASON: GATEWAY_FETCH_UNAUTHORIZED)","LICENSE_REJECTED");if(p.status>=500&&p.status<600){a=new b(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: HTTP_${p.status})`,"GATEWAY_FETCH_FAILED");continue}if(!p.ok)throw new b(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: HTTP_${p.status})`,"GATEWAY_FETCH_FAILED");s=await p.json(),a=null;break}catch(p){if(p instanceof b&&p.code==="LICENSE_REJECTED")throw p;a=p;}}if(a||!s)throw a instanceof b?a:new b("[ir-system]: GATEWAY_FETCH_FAILED (REASON: NETWORK_ERROR)","GATEWAY_FETCH_FAILED",a);let o;try{o=await fetch(s.bundleUrl);}catch(c){throw new b("[ir-system]: GATEWAY_FETCH_FAILED (REASON: BUNDLE_DOWNLOAD_FAILED)","GATEWAY_FETCH_FAILED",c)}if(!o.ok)throw new b(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: BUNDLE_HTTP_${o.status})`,"GATEWAY_FETCH_FAILED");let l=await o.arrayBuffer();return {bytes:Buffer.from(l),sha256:s.sha256,signature:s.signature,libraryVersion:s.libraryVersion,bundleVersion:s.bundleVersion}}static async verifySignature(e,t,r){try{let{compactVerify:s,importSPKI:i}=await import('jose'),a=await i(me,"RS256"),{payload:o}=await s(e,a),l=JSON.parse(new TextDecoder().decode(o));if(l.purpose!=="gateway-bundle")throw new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: PURPOSE_MISMATCH)","GATEWAY_VERIFICATION_FAILED");if(l.libraryVersion!==t)throw new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: LIBRARY_VERSION_MISMATCH)","GATEWAY_VERIFICATION_FAILED");if(l.sha256!==r)throw new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: SHA256_MISMATCH)","GATEWAY_VERIFICATION_FAILED")}catch(s){throw s instanceof b?s:new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: SIGNATURE_INVALID)","GATEWAY_VERIFICATION_FAILED",s)}}static loadIntoVm(e){try{let t=path.join(os.homedir(),".inference-relay","gateway","runtime.cjs"),r=path.dirname(t),s=module$1.createRequire(t),i={exports:{}};vm.compileFunction(e,["exports","require","module","__filename","__dirname"],{filename:"<gateway-bundle>"})(i.exports,s,i,t,r);let o=i.exports;if(!o.ClaudeCLIProvider)throw new b("[ir-system]: GATEWAY_LOAD_FAILED (REASON: MISSING_CLAUDE_CLI_EXPORT)","GATEWAY_LOAD_FAILED");return o}catch(t){throw t instanceof b?t:new b("[ir-system]: GATEWAY_LOAD_FAILED (REASON: VM_EXECUTION_ERROR)","GATEWAY_LOAD_FAILED",t)}}static clearCache(){F.clear();}};var pt="1.5.2",te=class extends R{type="claude-cli";supportsStreaming=true;supportsMultimodal=true;platformRequirement="desktop";licenseKey;loadedProvider=null;loadingPromise=null;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.licenseKey=e;}async ensureLoaded(){return this.loadedProvider?this.loadedProvider:this.loadingPromise?this.loadingPromise:(this.loadingPromise=(async()=>{try{let t=(await B.load(this.licenseKey,pt)).ClaudeCLIProvider,r=new t;return this.loadedProvider=r,r}catch(e){throw this.loadingPromise=null,this.health.status="unhealthy",this.health.errorCount++,e instanceof b?this.health.lastError=e.message:e instanceof Error&&(this.health.lastError=e.message),e}})(),this.loadingPromise)}async probe(){let{detectCLI:e}=await Promise.resolve().then(()=>(ee(),Me));return e()}async execute(e,t){return (await this.ensureLoaded()).execute(e,t)}async streamRaw(e,t){return (await this.ensureLoaded()).streamRaw(e,t)}estimateCost(e){if(this.loadedProvider)return this.loadedProvider.estimateCost(e);let r=JSON.stringify(e.messages).length/4+5e3,s=(r*3+1e3*15)/1e6;return {estimatedInputTokens:Math.round(r),estimatedCost:Math.round(s*1e4)/1e4,model:e.model,provider:"claude-cli"}}getHealth(){return this.loadedProvider?this.loadedProvider.getHealth():{...this.health}}reportRateLimit(e){this.loadedProvider?this.loadedProvider.reportRateLimit(e):(this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now());}};var G=[{canonical:"claude-haiku-4-5",apiId:"claude-haiku-4-5-20251001",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[.8,4]},{canonical:"claude-sonnet-4-6",apiId:"claude-sonnet-4-20250514",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[3,15]},{canonical:"claude-opus-4-6",apiId:"claude-opus-4-20250514",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[15,75]},{canonical:"gpt-4.1",apiId:"gpt-4.1",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[2,8]},{canonical:"gpt-4.1-mini",apiId:"gpt-4.1-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[.4,1.6]},{canonical:"gpt-4.1-nano",apiId:"gpt-4.1-nano",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[.1,.4]},{canonical:"gpt-4o",apiId:"gpt-4o",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:128e3},pricing:[2.5,10]},{canonical:"gpt-4o-mini",apiId:"gpt-4o-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:128e3},pricing:[.15,.6]},{canonical:"o4-mini",apiId:"o4-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[1.1,4.4]},{canonical:"o3",apiId:"o3",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[.4,1.6]},{canonical:"o3-mini",apiId:"o3-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[1.1,4.4]},{canonical:"o1",apiId:"o1",provider:"openai-api",capabilities:{streaming:false,tools:false,multimodal:false,contextWindow:2e5},pricing:[15,60]},{canonical:"llama3.1",apiId:"llama3.1",provider:"ollama",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:128e3}},{canonical:"mistral",apiId:"mistral",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:false,contextWindow:32e3}},{canonical:"mistral-nemo",apiId:"mistral-nemo",provider:"ollama",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:128e3}},{canonical:"llava",apiId:"llava",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:true,contextWindow:32e3}},{canonical:"qwen2:0.5b",apiId:"qwen2:0.5b",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:false,contextWindow:32e3}}],K=Object.fromEntries(G.filter(n=>n.canonical!==n.apiId).map(n=>[n.canonical,n.apiId]));K["claude-sonnet-4"]="claude-sonnet-4-20250514";K["claude-opus-4"]="claude-opus-4-20250514";K["claude-haiku-4"]="claude-haiku-4-5-20251001";function fe(n){return K[n]||n}function mt(n){return G.find(e=>e.canonical===n)}function re(n){return G.find(e=>e.canonical===n)?.provider}function ht(n){return G.filter(e=>e.provider===n)}var yt="https://api.anthropic.com/v1/messages",gt="2023-06-01",be={"claude-sonnet-4-5-20250514":[3,15],"claude-sonnet-4-20250514":[3,15],"claude-haiku-4-5-20250414":[.8,4],"claude-opus-4-20250514":[15,75]};function De(n){if(be[n])return be[n];for(let[e,t]of Object.entries(be))if(n.startsWith(e.split("-20")[0]))return t;return [3,15]}var V=class extends R{type="anthropic-api";supportsStreaming=true;supportsMultimodal=true;platformRequirement="any";apiKey;baseUrl;dangerousDirectBrowserAccess;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.apiKey=e.apiKey,this.baseUrl=e.baseUrl?.replace(/\/$/,"")||yt,this.dangerousDirectBrowserAccess=e.dangerousDirectBrowserAccess??false;}async probe(){try{let e=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify({model:"claude-haiku-4-5-20250414",max_tokens:1,messages:[{role:"user",content:"hi"}]})});return e.status===401?{available:!0,authenticated:!1}:e.ok||e.status===429?{available:!0,authenticated:!0,rateLimitTier:e.headers.get("anthropic-ratelimit-tier")??void 0}:e.status===529?{available:!0,authenticated:!0}:{available:!1,authenticated:!1}}catch{return {available:false,authenticated:false}}}async execute(e,t){let r=Date.now(),s=this.buildRequestBody(e,false),i=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(s),signal:t?.signal});if(!i.ok)throw this.handleErrorResponse(i),new Error("unreachable");let a=await i.json(),o=Date.now()-r;this.recordSuccess();let[l,d]=De(a.model||e.model),c=((a.usage?.input_tokens||0)*l+(a.usage?.output_tokens||0)*d)/1e6,p={},y=["request-id","x-request-id","anthropic-ratelimit-requests-limit","anthropic-ratelimit-requests-remaining","anthropic-ratelimit-requests-reset","anthropic-ratelimit-tokens-limit","anthropic-ratelimit-tokens-remaining","anthropic-ratelimit-tokens-reset","anthropic-ratelimit-tier"];for(let m of y){let _=i.headers.get(m);_&&(p[m]=_);}return {content:a.content||[],model:a.model||e.model,usage:{input_tokens:a.usage?.input_tokens||0,output_tokens:a.usage?.output_tokens||0},stop_reason:a.stop_reason||"end_turn",provider:"anthropic-api",costUsd:Math.round(c*1e6)/1e6,durationMs:o,fallbackChain:["anthropic-api:success"],headers:Object.keys(p).length>0?p:void 0}}async streamRaw(e,t){let r=this.buildRequestBody(e,true),s=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(r),signal:t?.signal});if(s.ok||this.handleErrorResponse(s),!s.body)throw new Error("ANTHROPIC_API_ERROR: Response body is null \u2014 streaming unavailable");let i=s.body.getReader(),a=new TextDecoder;return (async function*(){let o="",l=false;try{for(;;){let{done:d,value:c}=await i.read();if(d)break;o+=a.decode(c,{stream:!0});let p=o.split(`
|
|
11
|
-
`);o=p.pop()||"";let y=null;for(let m of p)if(m.startsWith("event: "))y=m.slice(7).trim();else if(m.startsWith("data: ")){let
|
|
12
|
-
`);l=u.pop()||"";for(let f of u){if(!f.startsWith("data: "))continue;let h=f.slice(6).trim();if(h==="[DONE]"){yield {type:"message_delta",delta:{stop_reason:y||"end_turn"},usage:p||{}},yield {type:"message_stop"},c=!0;return}try{let g=JSON.parse(h);g.usage&&(p={input_tokens:g.usage.prompt_tokens||0,output_tokens:g.usage.completion_tokens||0});let C=g.choices?.[0];if(!C)continue;let
|
|
13
|
-
`);l=m.pop()||"";for(let _ of m){let u=_.trim();if(u)try{let f=JSON.parse(u);if(f.message?.content&&(d||(d=!0,yield {type:"content_block_start",index:0,content_block:{type:"text",text:""}}),yield {type:"content_block_delta",index:0,delta:{type:"text_delta",text:f.message.content}}),f.done){d&&(yield {type:"content_block_stop",index:0}),yield {type:"message_delta",delta:{stop_reason:"end_turn"},usage:{input_tokens:f.prompt_eval_count||0,output_tokens:f.eval_count||0}},yield {type:"message_stop"},c=!0;return}}catch{}}}c||(yield {type:"message_stop"});}finally{try{i.releaseLock();}catch{}}})()}estimateCost(e){let t=Math.round(JSON.stringify(e.messages).length/4),r=e.system?Math.round(e.system.length/4):0;return {estimatedInputTokens:t+r,estimatedCost:0,model:e.model,provider:"ollama"}}getHealth(){return {...this.health}}reportRateLimit(e){this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now();}buildRequestBody(e,t){if(e.tools&&e.tools.length>0)throw new Error(`OLLAMA_TOOLS_NOT_SUPPORTED: The Ollama provider received ${e.tools.length} tool definition(s) in params.tools, but this provider does not implement function calling in v1.3.x. For tool-use workflows route to 'anthropic-api' or 'openai-api' instead. Tracking issue: 1.4.0 ships an Ollama tool-use shim for models that support native function calling (llama3.1, mistral-nemo, etc.).`);let r=Ct(e.messages,e.system),s={model:e.model,messages:r,stream:t};return e.temperature!==void 0&&(s.options={temperature:e.temperature}),s}handleErrorResponse(e){let t=e.status;throw this.health.errorCount++,this.health.lastCheckAt=Date.now(),t===404?(this.health.status="unhealthy",this.health.lastError="Model not found",new Error("OLLAMA_MODEL_NOT_FOUND: Requested model is not available locally (404)")):(this.health.status="unhealthy",this.health.lastError=`Request failed with status ${t}`,new Error(`OLLAMA_ERROR: Request failed with status ${t}`))}recordSuccess(){this.health.status="healthy",this.health.lastCheckAt=Date.now(),this.health.errorCount=0,this.health.lastError=void 0;}};var $=class extends events.EventEmitter{constructor(){super(),this.setMaxListeners(50);}emitTyped(e){return this.emit(e.kind,e)}onTyped(e,t){return this.on(e,t)}offTyped(e,t){return this.off(e,t)}};var We=200;function ne(n,e){if(e instanceof b)return `${n}:failed:${e.code}:${e.message}`;let t=e?.message;return `${n}:failed:${t?.slice(0,We)||"unknown"}`}var Fe=new Set(["anthropic-api","openai-api","ollama"]),q=class extends Error{code="ALL_PROVIDERS_FAILED";fallbackChain;constructor(e){super(`All providers failed. Chain: ${e.join(" \u2192 ")}`),this.name="RelayError",this.fallbackChain=e;}};function Be(n){return n.messages.some(e=>typeof e.content=="string"?false:Array.isArray(e.content)&&e.content.some(t=>t.type==="image"||t.type==="image_url"))}function _e(n){if(!(n instanceof Error))return "unknown";let e=n.code;if(typeof e=="string"){if(e==="LICENSE_REJECTED")return "auth_failed";if(e==="GATEWAY_FETCH_FAILED"||e==="GATEWAY_VERIFICATION_FAILED"||e==="GATEWAY_LOAD_FAILED")return "gateway_unavailable"}let t=n.message.toLowerCase(),r=n.name.toLowerCase();return t.startsWith("openai_model_translation_required:")||t.startsWith("ollama_tools_not_supported:")||t.startsWith("multimodal_cli_too_old:")?"provider_incompatible":t.includes("rate limit")||t.includes("429")||r.includes("ratelimit")?"rate_limited":t.includes("auth")||t.includes("401")||t.includes("403")||t.includes("unauthorized")||t.includes("forbidden")?"auth_failed":t.includes("timeout")||t.includes("timed out")||t.includes("aborted")||r==="aborterror"||r==="timeouterror"?"timeout":"unknown"}function He(n){return [...n].sort((e,t)=>e.priority-t.priority)}var j=class n{healthMap=new Map;engineEvents=new $;static UNHEALTHY_TTL_MS=300*1e3;getHealthStatus(e){let t=this.healthMap.get(e);return t?t.status==="unhealthy"&&Date.now()-t.updatedAt>=n.UNHEALTHY_TTL_MS?(this.healthMap.delete(e),"healthy"):t.status:"healthy"}markUnhealthy(e,t){this.healthMap.set(e,{status:"unhealthy",lastError:t,updatedAt:Date.now()});}markDegraded(e,t){this.healthMap.set(e,{status:"degraded",lastError:t,updatedAt:Date.now()});}markHealthy(e){this.healthMap.set(e,{status:"healthy",updatedAt:Date.now()});}resetHealth(){this.healthMap.clear();}async execute(e,t,r){let s=He(t),i=[],a=Be(e),o=crypto$1.randomUUID(),l=Date.now();for(let d of s){let c=d.type;if(this.getHealthStatus(c)==="unhealthy"){i.push(`${c}:skipped:unhealthy`);continue}if(a&&!d.supportsMultimodal){i.push(`${c}:skipped:no_multimodal`);continue}let y=re(e.model);if(y&&y!==c&&Fe.has(c)){i.push(`${c}:skipped:wrong_provider`);continue}try{let m=await d.execute(e,{signal:r?.signal});return this.markHealthy(c),i.push(`${c}:success`),this.emitEngineEvent({kind:"stream_complete",callId:o,id:crypto$1.randomUUID(),ts:Date.now(),stitched:!1,segments:[{provider:c,inputTokens:m.usage.input_tokens,outputTokens:m.usage.output_tokens,costUsd:m.costUsd,userCostUsd:m.userCostUsd??0}],provider:c,model:m.model,inputTokens:m.usage.input_tokens,outputTokens:m.usage.output_tokens,costUsd:m.costUsd,userCostUsd:m.userCostUsd,durationMs:m.durationMs??Date.now()-l,stopReason:m.stop_reason,fallbackChain:i,rateLimitHit:!1,securityState:r?.licenseState}),{...m,fallbackChain:i}}catch(m){if(m?.name==="AbortError"||r?.signal?.aborted)throw m;switch(_e(m)){case "rate_limited":d.reportRateLimit({provider:c,utilization:1,resetsAt:Date.now()+6e4}),this.markDegraded(c,"rate_limited"),i.push(`${c}:failed:rate_limited`);break;case "auth_failed":this.markUnhealthy(c,"auth_failed"),i.push(`${c}:failed:auth_failed`);break;case "timeout":i.push(`${c}:failed:timeout`);break;case "provider_incompatible":i.push(`${c}:incompatible`);break;case "gateway_unavailable":this.markUnhealthy(c,"gateway_unavailable"),i.push(ne(c,m));break;default:i.push(ne(c,m));break}}}throw new q(i)}async stream(e,t,r){let s=crypto$1.randomUUID(),i=He(t),a=Be(e),o=Date.now(),l=[],d=null,c=null;for(let h of i){let g=h.type;if(this.getHealthStatus(g)==="unhealthy"){l.push(`${g}:skipped:unhealthy`);continue}if(a&&!h.supportsMultimodal){l.push(`${g}:skipped:no_multimodal`);continue}if(!h.supportsStreaming){l.push(`${g}:skipped:no_streaming`);continue}let C=re(e.model);if(C&&C!==g&&Fe.has(g)){l.push(`${g}:skipped:wrong_provider`);continue}try{c=await h.streamRaw(e,{signal:r?.signal}),d=h,l.push(`${g}:success`);break}catch(E){if(E?.name==="AbortError"||r?.signal?.aborted)throw E;switch(_e(E)){case "rate_limited":h.reportRateLimit({provider:g,utilization:1,resetsAt:Date.now()+6e4}),this.markDegraded(g,"rate_limited"),l.push(`${g}:failed:rate_limited`);break;case "auth_failed":this.markUnhealthy(g,"auth_failed"),l.push(`${g}:failed:auth_failed`);break;case "timeout":l.push(`${g}:failed:timeout`);break;case "provider_incompatible":l.push(`${g}:incompatible`);break;case "gateway_unavailable":this.markUnhealthy(g,"gateway_unavailable"),l.push(ne(g,E));break;default:l.push(ne(g,E));break}}}if(!d||!c)throw new q(l);let y=this,m=r?.maxMidStreamFailovers??2,_=new Set([d]);async function*u(){let h=d,g=c[Symbol.asyncIterator](),C={role:"assistant",content:[]},E={segments:[]},I=Ge(h.type);E.segments.push(I);let Re=0,z=null,X=false,de=m,O="",N="";for(y.emitEngineEvent({kind:"segment_start",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),provider:h.type});;)try{for(;;){let A=await g.next();if(A.done)break;let T=A.value,k=Pt(T,Re,C,X,z);if(k.action==="suppress"){"coalesceMode"in k&&(z=k.coalesceMode),k.clearedJustFailedOver&&(X=!1);continue}k.coalesceModeCleared&&(z=null),k.clearedJustFailedOver&&(X=!1);let v=k.event;if(St(C,v),v.type==="message_start"){if(E.segments.length>1)continue;O=v.message?.model||O,v.message?.usage?.input_tokens!=null&&(I.inputTokens=v.message.usage.input_tokens);}if(v.type==="message_delta"&&(v.delta?.stop_reason&&(N=v.delta.stop_reason),v.usage)){Rt(I,h.type,v.usage),yield {...v,usage:void 0};continue}if(v.type==="message_stop")break;yield v;}if(C.content.length===0&&I.outputTokens===0)throw new Error("stream ended without producing any content");y.markHealthy(h.type),y.emitEngineEvent({kind:"segment_complete",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),provider:h.type,usage:{...I}});let S=Ke(E.segments);yield {type:"message_delta",delta:{stop_reason:N||"end_turn"},usage:S},yield {type:"message_stop"},y.emitEngineEvent({kind:"stream_complete",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),stitched:E.segments.length>1,segments:E.segments.map(A=>({...A})),provider:E.segments.map(A=>A.provider).join("+"),model:O,inputTokens:S.input_tokens,outputTokens:S.output_tokens,costUsd:E.segments.reduce((A,T)=>A+T.costUsd,0),userCostUsd:E.segments.reduce((A,T)=>A+T.userCostUsd,0),durationMs:Date.now()-o,stopReason:N||"end_turn",fallbackChain:l,rateLimitHit:!1,securityState:r?.licenseState});return}catch(S){if(S?.name==="AbortError"||r?.signal?.aborted)throw S;let A=_e(S);try{await g.return?.();}catch{}if(y.markUnhealthy(h.type,String(S)),y.emitEngineEvent({kind:"mid_stream_failover_attempt",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),from:h.type,bytesBuffered:Ve(C),errorClass:A,retriesLeft:de}),l.push(`${h.type}:partial(${Ve(C)}b)`),de<=0)throw y.emitTerminalExhaustion(s,E,O,N,l,o,r?.licenseState),S;let T=kt(C.content),k=[...e.messages,{role:"assistant",content:T}],v=null,Q=null,Se=h;for(;;){let w=y.nextHealthyProvider(i,Se,a,_);if(!w)break;Se=w,_.add(w);let je={...e,messages:k,...w.type!=="anthropic-api"&&w.type!=="claude-cli"?{system:wt(e.system,"[SYSTEM]: CONTINUE_LAST_RESPONSE_WITHOUT_REPEATING")}:{}};try{Q=await w.streamRaw(je,{signal:r?.signal}),v=w;break}catch(Pe){y.markUnhealthy(w.type,String(Pe)),l.push(`${w.type}:failed_connect:${String(Pe).slice(0,We)}`);}}if(!v||!Q)throw y.emitTerminalExhaustion(s,E,O,N,l,o,r?.licenseState),S;l.push(`${v.type}:stitched`),y.emitEngineEvent({kind:"mid_stream_failover_success",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),from:h.type,to:v.type}),Re=It(C)+1,de--,h=v,g=Q[Symbol.asyncIterator](),X=true,z=null,I=Ge(h.type),E.segments.push(I),y.emitEngineEvent({kind:"segment_start",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),provider:h.type});}}let f=u();return r?.licenseState==="SEC_DEGRADED"&&(f=Tt(f)),L(f)}nextHealthyProvider(e,t,r,s){let i=e.indexOf(t);if(i===-1)return null;for(let a=i+1;a<e.length;a++){let o=e[a];if(!s.has(o)&&!(r&&!o.supportsMultimodal)&&o.supportsStreaming)return o}return null}emitEngineEvent(e){this.engineEvents.emitTyped(e);}emitTerminalExhaustion(e,t,r,s,i,a,o){let l=Ke(t.segments);this.emitEngineEvent({kind:"stream_complete",callId:e,id:crypto$1.randomUUID(),ts:Date.now(),stitched:t.segments.length>1,segments:t.segments.map(d=>({...d})),provider:t.segments.map(d=>d.provider).join("+"),model:r,inputTokens:l.input_tokens,outputTokens:l.output_tokens,costUsd:t.segments.reduce((d,c)=>d+c.costUsd,0),userCostUsd:t.segments.reduce((d,c)=>d+c.userCostUsd,0),durationMs:Date.now()-a,stopReason:s||"error",fallbackChain:i,rateLimitHit:false,securityState:o});}};function Ge(n){return {provider:n,inputTokens:0,outputTokens:0,costUsd:0,userCostUsd:0}}function Rt(n,e,t){typeof t.input_tokens=="number"&&(n.inputTokens=t.input_tokens),typeof t.output_tokens=="number"&&(e==="anthropic-api"||e==="claude-cli"?n.outputTokens+=t.output_tokens:n.outputTokens=t.output_tokens);}function Ke(n){return {input_tokens:n[n.length-1]?.inputTokens??0,output_tokens:n.reduce((e,t)=>e+t.outputTokens,0)}}function St(n,e){if(e.type==="content_block_start"){let t=e.content_block;t?.type==="text"?n.content.push({type:"text",text:t.text||"",open:true}):t?.type==="tool_use"&&n.content.push({type:"tool_use",id:t.id,name:t.name,input:{},_partialJson:"",open:true});}else if(e.type==="content_block_delta"){let t=n.content[n.content.length-1];if(!t)return;t.type==="text"&&e.delta?.type==="text_delta"?t.text+=e.delta.text||"":t.type==="tool_use"&&e.delta?.type==="input_json_delta"&&(t._partialJson=(t._partialJson||"")+(e.delta.partial_json||""));}else if(e.type==="content_block_stop"){for(let t=n.content.length-1;t>=0;t--)if(n.content[t].open){if(n.content[t].open=false,n.content[t].type==="tool_use"){let r=n.content[t];try{r.input=JSON.parse(r._partialJson||"{}");}catch{}delete r._partialJson;}break}}}function Pt(n,e,t,r,s){if(s&&n.type==="content_block_delta"&&n.index===s.fromIndex)return {action:"yield",event:{...n,index:s.toIndex}};if(s&&n.type==="content_block_stop"&&n.index===s.fromIndex)return {action:"suppress",coalesceMode:null,coalesceModeCleared:true};if(r&&n.type==="content_block_start"&&n.content_block?.type==="text"){let a=t.content[t.content.length-1];if(a?.type==="text"&&a.open===true)return {action:"suppress",coalesceMode:{fromIndex:n.index+e,toIndex:t.content.length-1},clearedJustFailedOver:true}}let i=r&&n.type==="content_block_start";return typeof n.index=="number"&&e>0?{action:"yield",event:{...n,index:n.index+e},clearedJustFailedOver:i}:{action:"yield",event:n,clearedJustFailedOver:i}}function kt(n){let e=[];for(let t of n)t.type==="text"?t.text&&e.push({type:"text",text:t.text}):t.type==="tool_use"&&!t.open&&t.input&&Object.keys(t.input).length>0&&e.push({type:"tool_use",id:t.id,name:t.name,input:t.input});return e}function wt(n,e){return n?`${n}
|
|
10
|
+
-----END PUBLIC KEY-----`,pe={version:"1.0.0",cliVersion:"2.1.87",eventTypes:{system:"system",streamEvent:"stream_event",assistant:"assistant",result:"result"},fieldMappings:{sessionId:"session_id",model:"model",cliVersion:"claude_code_version",content:"content",usage:"usage",inputTokens:"input_tokens",outputTokens:"output_tokens",stopReason:"stop_reason",isError:"is_error",totalCostUsd:"total_cost_usd",durationMs:"duration_ms",error:"error",event:"event"},syntheticModel:"<synthetic>"},M=class{verificationFailures=0;cachedManifest=null;currentSecurityState="SEC_NOMINAL";async verifyManifest(e){try{let{compactVerify:t,importSPKI:r}=await import('jose'),s=await r(me,"RS256"),{payload:i}=await t(e,s),a=JSON.parse(new TextDecoder().decode(i));return this.verificationFailures=0,this.cachedManifest=a,this.currentSecurityState="SEC_NOMINAL",{manifest:a,securityState:"SEC_NOMINAL"}}catch{if(this.verificationFailures++,this.verificationFailures>=3){this.currentSecurityState="SEC_DEGRADED";let t=2e3+Math.random()*3e3;return await new Promise(r=>setTimeout(r,t)),console.warn("[inference-relay] Logic synchronization mismatch. Routing via local recovery."),{manifest:pe,securityState:"SEC_DEGRADED"}}return this.cachedManifest?(this.currentSecurityState="SEC_CACHE",{manifest:this.cachedManifest,securityState:"SEC_CACHE"}):(this.currentSecurityState="SEC_RECOVERY",{manifest:pe,securityState:"SEC_RECOVERY"})}}getSecurityState(){return this.currentSecurityState}clearCache(){this.cachedManifest=null,this.verificationFailures=0,this.currentSecurityState="SEC_NOMINAL";}},Ze=new M;function ke(){Ze.clearCache();}function et(){return pe}function tt(n,e){return e.split(".").reduce((t,r)=>{if(!(r==="__proto__"||r==="constructor"||r==="prototype"))return t&&t[r]},n)}var at=process.env.IR_GATEWAY_URL||"https://api.inference-relay.com";function ge(n){return path.join(os.homedir(),".inference-relay","gateway",n)}function Te(n){return path.join(ge(n),"bundle.js")}function xe(n){return path.join(ge(n),"meta.json")}var b=class extends Error{constructor(t,r,s){super(t);this.code=r;this.cause=s;this.name="GatewayLoaderError";}code;cause},F=new Map,B=class{static async load(e,t,r=at){let s=`${t}:${e.slice(-8)}`;if(F.has(s))return F.get(s);try{let p=await this.tryLoadFromDisk(t);if(p)return F.set(s,p),p}catch{}let{bytes:i,sha256:a,signature:o,libraryVersion:l}=await this.fetchFromAuthority(e,t,r);if(crypto$1.createHash("sha256").update(i).digest("hex")!==a)throw new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: HASH_MISMATCH)","GATEWAY_VERIFICATION_FAILED");await this.verifySignature(o,l,a);try{let p=ge(t);fs.mkdirSync(p,{recursive:!0}),fs.writeFileSync(Te(t),i),fs.writeFileSync(xe(t),JSON.stringify({sha256:a,signature:o,libraryVersion:l}));}catch{}let c=this.loadIntoVm(i.toString("utf-8"));return F.set(s,c),c}static async loadFromDisk(e){if(!fs.existsSync(e))throw new b(`[ir-system]: GATEWAY_LOAD_FAILED (REASON: BUNDLE_NOT_FOUND at ${e})`,"GATEWAY_LOAD_FAILED");let t=fs.readFileSync(e,"utf-8");return this.loadIntoVm(t)}static async tryLoadFromDisk(e){let t=Te(e),r=xe(e);if(!fs.existsSync(t)||!fs.existsSync(r))return null;let s=fs.readFileSync(t),i=JSON.parse(fs.readFileSync(r,"utf-8"));return crypto$1.createHash("sha256").update(s).digest("hex")!==i.sha256?null:(await this.verifySignature(i.signature,i.libraryVersion,i.sha256),this.loadIntoVm(s.toString("utf-8")))}static async fetchFromAuthority(e,t,r){let s,i=[0,1e3,4e3],a;for(let c of i){c>0&&await new Promise(p=>setTimeout(p,c));try{let p=await fetch(`${r}/v1/gateway/fetch`,{method:"POST",headers:{"Content-Type":"application/json",Authorization:`Bearer ${e}`},body:JSON.stringify({libraryVersion:t})});if(p.status===401||p.status===403)throw new b("[ir-system]: LICENSE_REJECTED (REASON: GATEWAY_FETCH_UNAUTHORIZED)","LICENSE_REJECTED");if(p.status>=500&&p.status<600){a=new b(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: HTTP_${p.status})`,"GATEWAY_FETCH_FAILED");continue}if(!p.ok)throw new b(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: HTTP_${p.status})`,"GATEWAY_FETCH_FAILED");s=await p.json(),a=null;break}catch(p){if(p instanceof b&&p.code==="LICENSE_REJECTED")throw p;a=p;}}if(a||!s)throw a instanceof b?a:new b("[ir-system]: GATEWAY_FETCH_FAILED (REASON: NETWORK_ERROR)","GATEWAY_FETCH_FAILED",a);let o;try{o=await fetch(s.bundleUrl);}catch(c){throw new b("[ir-system]: GATEWAY_FETCH_FAILED (REASON: BUNDLE_DOWNLOAD_FAILED)","GATEWAY_FETCH_FAILED",c)}if(!o.ok)throw new b(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: BUNDLE_HTTP_${o.status})`,"GATEWAY_FETCH_FAILED");let l=await o.arrayBuffer();return {bytes:Buffer.from(l),sha256:s.sha256,signature:s.signature,libraryVersion:s.libraryVersion,bundleVersion:s.bundleVersion}}static async verifySignature(e,t,r){try{let{compactVerify:s,importSPKI:i}=await import('jose'),a=await i(me,"RS256"),{payload:o}=await s(e,a),l=JSON.parse(new TextDecoder().decode(o));if(l.purpose!=="gateway-bundle")throw new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: PURPOSE_MISMATCH)","GATEWAY_VERIFICATION_FAILED");if(l.libraryVersion!==t)throw new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: LIBRARY_VERSION_MISMATCH)","GATEWAY_VERIFICATION_FAILED");if(l.sha256!==r)throw new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: SHA256_MISMATCH)","GATEWAY_VERIFICATION_FAILED")}catch(s){throw s instanceof b?s:new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: SIGNATURE_INVALID)","GATEWAY_VERIFICATION_FAILED",s)}}static loadIntoVm(e){try{let t=path.join(os.homedir(),".inference-relay","gateway","runtime.cjs"),r=path.dirname(t),s=module$1.createRequire(t),i={exports:{}};vm.compileFunction(e,["exports","require","module","__filename","__dirname"],{filename:"<gateway-bundle>"})(i.exports,s,i,t,r);let o=i.exports;if(!o.ClaudeCLIProvider)throw new b("[ir-system]: GATEWAY_LOAD_FAILED (REASON: MISSING_CLAUDE_CLI_EXPORT)","GATEWAY_LOAD_FAILED");return o}catch(t){throw t instanceof b?t:new b("[ir-system]: GATEWAY_LOAD_FAILED (REASON: VM_EXECUTION_ERROR)","GATEWAY_LOAD_FAILED",t)}}static clearCache(){F.clear();}};var pt="1.6.0",te=class extends R{type="claude-cli";supportsStreaming=true;supportsMultimodal=true;platformRequirement="desktop";licenseKey;loadedProvider=null;loadingPromise=null;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.licenseKey=e;}async ensureLoaded(){return this.loadedProvider?this.loadedProvider:this.loadingPromise?this.loadingPromise:(this.loadingPromise=(async()=>{try{let t=(await B.load(this.licenseKey,pt)).ClaudeCLIProvider,r=new t;return this.loadedProvider=r,r}catch(e){throw this.loadingPromise=null,this.health.status="unhealthy",this.health.errorCount++,e instanceof b?this.health.lastError=e.message:e instanceof Error&&(this.health.lastError=e.message),e}})(),this.loadingPromise)}async probe(){let{detectCLI:e}=await Promise.resolve().then(()=>(ee(),Me));return e()}async execute(e,t){return (await this.ensureLoaded()).execute(e,t)}async streamRaw(e,t){return (await this.ensureLoaded()).streamRaw(e,t)}estimateCost(e){if(this.loadedProvider)return this.loadedProvider.estimateCost(e);let r=JSON.stringify(e.messages).length/4+5e3,s=(r*3+1e3*15)/1e6;return {estimatedInputTokens:Math.round(r),estimatedCost:Math.round(s*1e4)/1e4,model:e.model,provider:"claude-cli"}}getHealth(){return this.loadedProvider?this.loadedProvider.getHealth():{...this.health}}reportRateLimit(e){this.loadedProvider?this.loadedProvider.reportRateLimit(e):(this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now());}};var G=[{canonical:"claude-haiku-4-5",apiId:"claude-haiku-4-5-20251001",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[.8,4]},{canonical:"claude-sonnet-4-6",apiId:"claude-sonnet-4-20250514",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[3,15]},{canonical:"claude-opus-4-6",apiId:"claude-opus-4-20250514",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[15,75]},{canonical:"gpt-4.1",apiId:"gpt-4.1",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[2,8]},{canonical:"gpt-4.1-mini",apiId:"gpt-4.1-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[.4,1.6]},{canonical:"gpt-4.1-nano",apiId:"gpt-4.1-nano",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[.1,.4]},{canonical:"gpt-4o",apiId:"gpt-4o",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:128e3},pricing:[2.5,10]},{canonical:"gpt-4o-mini",apiId:"gpt-4o-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:128e3},pricing:[.15,.6]},{canonical:"o4-mini",apiId:"o4-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[1.1,4.4]},{canonical:"o3",apiId:"o3",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[.4,1.6]},{canonical:"o3-mini",apiId:"o3-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[1.1,4.4]},{canonical:"o1",apiId:"o1",provider:"openai-api",capabilities:{streaming:false,tools:false,multimodal:false,contextWindow:2e5},pricing:[15,60]},{canonical:"llama3.1",apiId:"llama3.1",provider:"ollama",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:128e3}},{canonical:"mistral",apiId:"mistral",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:false,contextWindow:32e3}},{canonical:"mistral-nemo",apiId:"mistral-nemo",provider:"ollama",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:128e3}},{canonical:"llava",apiId:"llava",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:true,contextWindow:32e3}},{canonical:"qwen2:0.5b",apiId:"qwen2:0.5b",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:false,contextWindow:32e3}}],K=Object.fromEntries(G.filter(n=>n.canonical!==n.apiId).map(n=>[n.canonical,n.apiId]));K["claude-sonnet-4"]="claude-sonnet-4-20250514";K["claude-opus-4"]="claude-opus-4-20250514";K["claude-haiku-4"]="claude-haiku-4-5-20251001";function fe(n){return K[n]||n}function mt(n){return G.find(e=>e.canonical===n)}function re(n){return G.find(e=>e.canonical===n)?.provider}function ht(n){return G.filter(e=>e.provider===n)}var yt="https://api.anthropic.com/v1/messages",gt="2023-06-01",be={"claude-sonnet-4-5-20250514":[3,15],"claude-sonnet-4-20250514":[3,15],"claude-haiku-4-5-20250414":[.8,4],"claude-opus-4-20250514":[15,75]};function De(n){if(be[n])return be[n];for(let[e,t]of Object.entries(be))if(n.startsWith(e.split("-20")[0]))return t;return [3,15]}var V=class extends R{type="anthropic-api";supportsStreaming=true;supportsMultimodal=true;platformRequirement="any";apiKey;baseUrl;dangerousDirectBrowserAccess;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.apiKey=e.apiKey,this.baseUrl=e.baseUrl?.replace(/\/$/,"")||yt,this.dangerousDirectBrowserAccess=e.dangerousDirectBrowserAccess??false;}async probe(){try{let e=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify({model:"claude-haiku-4-5-20250414",max_tokens:1,messages:[{role:"user",content:"hi"}]})});return e.status===401?{available:!0,authenticated:!1}:e.ok||e.status===429?{available:!0,authenticated:!0,rateLimitTier:e.headers.get("anthropic-ratelimit-tier")??void 0}:e.status===529?{available:!0,authenticated:!0}:{available:!1,authenticated:!1}}catch{return {available:false,authenticated:false}}}async execute(e,t){let r=Date.now(),s=this.buildRequestBody(e,false),i=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(s),signal:t?.signal});if(!i.ok)throw this.handleErrorResponse(i),new Error("unreachable");let a=await i.json(),o=Date.now()-r;this.recordSuccess();let[l,d]=De(a.model||e.model),c=((a.usage?.input_tokens||0)*l+(a.usage?.output_tokens||0)*d)/1e6,p={},y=["request-id","x-request-id","anthropic-ratelimit-requests-limit","anthropic-ratelimit-requests-remaining","anthropic-ratelimit-requests-reset","anthropic-ratelimit-tokens-limit","anthropic-ratelimit-tokens-remaining","anthropic-ratelimit-tokens-reset","anthropic-ratelimit-tier"];for(let m of y){let E=i.headers.get(m);E&&(p[m]=E);}return {content:a.content||[],model:a.model||e.model,usage:{input_tokens:a.usage?.input_tokens||0,output_tokens:a.usage?.output_tokens||0},stop_reason:a.stop_reason||"end_turn",provider:"anthropic-api",costUsd:Math.round(c*1e6)/1e6,durationMs:o,fallbackChain:["anthropic-api:success"],headers:Object.keys(p).length>0?p:void 0}}async streamRaw(e,t){let r=this.buildRequestBody(e,true),s=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(r),signal:t?.signal});if(s.ok||this.handleErrorResponse(s),!s.body)throw new Error("ANTHROPIC_API_ERROR: Response body is null \u2014 streaming unavailable");let i=s.body.getReader(),a=new TextDecoder;return (async function*(){let o="",l=false;try{for(;;){let{done:d,value:c}=await i.read();if(d)break;o+=a.decode(c,{stream:!0});let p=o.split(`
|
|
11
|
+
`);o=p.pop()||"";let y=null;for(let m of p)if(m.startsWith("event: "))y=m.slice(7).trim();else if(m.startsWith("data: ")){let E=m.slice(6);if(y&&E)try{let u=JSON.parse(E);u.type==="message_stop"&&(l=!0),yield u;}catch{console.warn("[inference-relay] Skipped malformed SSE data:",E.substring(0,100));}y=null;}else m.trim()===""&&(y=null);}l||(yield {type:"message_stop"});}finally{try{i.releaseLock();}catch{}}})()}estimateCost(e){let t=Math.round(JSON.stringify(e.messages).length/4),r=e.system?Math.round(e.system.length/4):0,s=t+r,[i,a]=De(e.model),o=(s*i+1e3*a)/1e6;return {estimatedInputTokens:s,estimatedCost:Math.round(o*1e4)/1e4,model:e.model,provider:"anthropic-api"}}getHealth(){return {...this.health}}reportRateLimit(e){this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now();}buildHeaders(){let e={"content-type":"application/json","x-api-key":this.apiKey,"anthropic-version":gt};return this.dangerousDirectBrowserAccess&&(e["anthropic-dangerous-direct-browser-access"]="true"),e}buildRequestBody(e,t){let r={model:fe(e.model),max_tokens:e.max_tokens,messages:e.messages};return t&&(r.stream=true),e.system&&(r.system=e.system),e.temperature!==void 0&&(r.temperature=e.temperature),e.tools&&e.tools.length>0&&(r.tools=e.tools),e.metadata&&(r.metadata=e.metadata),r}handleErrorResponse(e){let t=e.status;if(t===401)throw this.health.status="unhealthy",this.health.lastError="Authentication failed",this.health.errorCount++,this.health.lastCheckAt=Date.now(),new Error("ANTHROPIC_AUTH_ERROR: Invalid API key (401)");if(t===429){let r=e.headers.get("retry-after"),s=r?parseInt(r,10)*1e3:6e4;this.health.status="degraded",this.health.lastError=`Rate limited \u2014 retry after ${s}ms`,this.health.lastCheckAt=Date.now();let i=new Error(`ANTHROPIC_RATE_LIMIT: Too many requests (429). Retry after ${s}ms`);throw i.retryAfterMs=s,i}throw t===529?(this.health.status="degraded",this.health.lastError="API overloaded",this.health.lastCheckAt=Date.now(),new Error("ANTHROPIC_OVERLOADED: API is temporarily overloaded (529)")):(this.health.errorCount++,this.health.lastCheckAt=Date.now(),new Error(`ANTHROPIC_API_ERROR: Request failed with status ${t}`))}recordSuccess(){this.health.status="healthy",this.health.lastCheckAt=Date.now(),this.health.errorCount=0,this.health.lastError=void 0;}};var ft="https://api.openai.com/v1/chat/completions",Oe=["gpt-","o1","o3","o4","chatgpt-","text-","davinci","babbage","ft:"],_e={"gpt-4.1":[2,8],"gpt-4.1-mini":[.4,1.6],"gpt-4.1-nano":[.1,.4],"gpt-4o":[2.5,10],"gpt-4o-mini":[.15,.6],"o4-mini":[1.1,4.4],o3:[.4,1.6],"o3-mini":[1.1,4.4],o1:[15,60],"o1-mini":[3,12],"gpt-4-turbo":[10,30],"gpt-4":[30,60],"gpt-3.5-turbo":[.5,1.5]};function Ne(n){if(_e[n])return _e[n];for(let[e,t]of Object.entries(_e))if(n.startsWith(e))return t;return [2.5,10]}function bt(n){return typeof n=="string"?n:n.filter(e=>e.type==="text"&&e.text).map(e=>e.text).join("")}function _t(n,e){let t=[];e&&t.push({role:"system",content:e});for(let r of n){let s=bt(r.content);if(r.role==="assistant"&&Array.isArray(r.content)){let i=r.content.filter(a=>a.type==="tool_use");if(i.length>0){t.push({role:"assistant",content:s||"",tool_calls:i.map(a=>({id:a.id||`call_${Math.random().toString(36).slice(2)}`,type:"function",function:{name:a.name,arguments:typeof a.input=="string"?a.input:JSON.stringify(a.input)}}))});continue}}t.push({role:r.role,content:s});}return t}function Ue(n){switch(n){case "stop":return "end_turn";case "length":return "max_tokens";case "tool_calls":case "function_call":return "tool_use";default:return "end_turn"}}var W=class extends R{type="openai-api";supportsStreaming=true;supportsMultimodal=true;platformRequirement="any";apiKey;baseUrl;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.apiKey=e.apiKey,this.baseUrl=e.baseUrl?.replace(/\/$/,"")||ft;}async probe(){try{let e=this.baseUrl.replace(/\/chat\/completions$/,"/models"),t=await fetch(e,{method:"GET",headers:this.buildHeaders()});return t.status===401?{available:!0,authenticated:!1}:t.ok||t.status===429?{available:!0,authenticated:!0}:{available:!1,authenticated:!1}}catch{return {available:false,authenticated:false}}}async execute(e,t){let r=Date.now(),s=this.buildRequestBody(e,false),i=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(s),signal:t?.signal});if(!i.ok)throw this.handleErrorResponse(i),new Error("unreachable");let a=await i.json(),o=Date.now()-r;this.recordSuccess();let l=a.choices?.[0],d=[];if(l?.message?.content&&d.push({type:"text",text:l.message.content}),l?.message?.tool_calls)for(let h of l.message.tool_calls)d.push({type:"tool_use",id:h.id,name:h.function?.name,input:h.function?.arguments?JSON.parse(h.function.arguments):{}});let c=a.usage?.prompt_tokens||0,p=a.usage?.completion_tokens||0,[y,m]=Ne(a.model||e.model),E=(c*y+p*m)/1e6,u={},f=["x-request-id","x-ratelimit-limit-requests","x-ratelimit-remaining-requests","x-ratelimit-limit-tokens","x-ratelimit-remaining-tokens","x-ratelimit-reset-requests","x-ratelimit-reset-tokens"];for(let h of f){let g=i.headers.get(h);g&&(u[h]=g);}return {content:d,model:a.model||e.model,usage:{input_tokens:c,output_tokens:p},stop_reason:Ue(l?.finish_reason),provider:"openai-api",costUsd:Math.round(E*1e6)/1e6,durationMs:o,fallbackChain:["openai-api:success"],headers:Object.keys(u).length>0?u:void 0}}async streamRaw(e,t){let r=this.buildRequestBody(e,true),s=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(r),signal:t?.signal});if(s.ok||this.handleErrorResponse(s),!s.body)throw new Error("OPENAI_API_ERROR: Response body is null \u2014 streaming unavailable");let i=s.body.getReader(),a=new TextDecoder,o=e.model;return (async function*(){let l="",d=false,c=false,p,y;yield {type:"message_start",message:{model:o,usage:{input_tokens:0,output_tokens:0}}};try{for(;;){let{done:m,value:E}=await i.read();if(m)break;l+=a.decode(E,{stream:!0});let u=l.split(`
|
|
12
|
+
`);l=u.pop()||"";for(let f of u){if(!f.startsWith("data: "))continue;let h=f.slice(6).trim();if(h==="[DONE]"){yield {type:"message_delta",delta:{stop_reason:y||"end_turn"},usage:p||{}},yield {type:"message_stop"},c=!0;return}try{let g=JSON.parse(h);g.usage&&(p={input_tokens:g.usage.prompt_tokens||0,output_tokens:g.usage.completion_tokens||0});let C=g.choices?.[0];if(!C)continue;let _=C.delta;_?.content&&!d&&(d=!0,yield {type:"content_block_start",index:0,content_block:{type:"text",text:""}}),_?.content&&(yield {type:"content_block_delta",index:0,delta:{type:"text_delta",text:_.content}}),C.finish_reason&&(d&&(yield {type:"content_block_stop",index:0}),y=Ue(C.finish_reason));}catch{}}}c||(yield {type:"message_delta",delta:{stop_reason:y||"end_turn"},usage:p||{}},yield {type:"message_stop"});}finally{try{i.releaseLock();}catch{}}})()}estimateCost(e){let t=Math.round(JSON.stringify(e.messages).length/4),r=e.system?Math.round(e.system.length/4):0,s=t+r,[i,a]=Ne(e.model),o=(s*i+1e3*a)/1e6;return {estimatedInputTokens:s,estimatedCost:Math.round(o*1e4)/1e4,model:e.model,provider:"openai-api"}}getHealth(){return {...this.health}}reportRateLimit(e){this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now();}buildHeaders(){return {"Content-Type":"application/json",Authorization:`Bearer ${this.apiKey}`}}buildRequestBody(e,t){let r=e.model.toLowerCase();if(!Oe.some(o=>r.startsWith(o)))throw new Error(`OPENAI_MODEL_TRANSLATION_REQUIRED: The OpenAI provider received the model name '${e.model}', which does not match any known OpenAI model prefix (${Oe.join(", ")}). The relay does NOT auto-translate model names across providers in v1.3.x. Either (a) pass an OpenAI model name like 'gpt-4o-mini' when routing to this provider, or (b) route non-OpenAI models through a provider that understands them (e.g. 'anthropic-api' for Claude models).`);let i=_t(e.messages,e.system),a={model:e.model,max_tokens:e.max_tokens,messages:i};return t&&(a.stream=true,a.stream_options={include_usage:true}),e.temperature!==void 0&&(a.temperature=e.temperature),e.tools&&e.tools.length>0&&(a.tools=e.tools),a}handleErrorResponse(e){let t=e.status;if(t===401)throw this.health.status="unhealthy",this.health.lastError="Authentication failed",this.health.errorCount++,this.health.lastCheckAt=Date.now(),new Error("OPENAI_AUTH_ERROR: Invalid API key (401)");if(t===429){let r=e.headers.get("retry-after"),s=r?parseInt(r,10)*1e3:6e4;this.health.status="degraded",this.health.lastError=`Rate limited \u2014 retry after ${s}ms`,this.health.lastCheckAt=Date.now();let i=new Error(`OPENAI_RATE_LIMIT: Too many requests (429). Retry after ${s}ms`);throw i.retryAfterMs=s,i}throw this.health.errorCount++,this.health.lastCheckAt=Date.now(),new Error(`OPENAI_API_ERROR: Request failed with status ${t}`)}recordSuccess(){this.health.status="healthy",this.health.lastCheckAt=Date.now(),this.health.errorCount=0,this.health.lastError=void 0;}};var Et="http://localhost:11434";function vt(n){return typeof n=="string"?n:n.filter(e=>e.type==="text"&&e.text).map(e=>e.text).join("")}function Ct(n,e){let t=[];e&&t.push({role:"system",content:e});for(let r of n)t.push({role:r.role,content:vt(r.content)});return t}var Y=class extends R{type="ollama";supportsStreaming=true;supportsMultimodal=false;platformRequirement="desktop";baseUrl;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.baseUrl=e?.baseUrl?.replace(/\/$/,"")||Et;}async probe(){try{return (await fetch(`${this.baseUrl}/api/tags`,{method:"GET"})).ok?{available:!0,authenticated:!0}:{available:!1,authenticated:!1}}catch{return {available:false,authenticated:false}}}async execute(e,t){let r=Date.now(),s=this.buildRequestBody(e,false),i=await fetch(`${this.baseUrl}/api/chat`,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify(s),signal:t?.signal});if(!i.ok)throw this.handleErrorResponse(i),new Error("unreachable");let a=await i.json(),o=Date.now()-r;this.recordSuccess();let l=[];a.message?.content&&l.push({type:"text",text:a.message.content});let d=a.prompt_eval_count||0,c=a.eval_count||0;return {content:l,model:a.model||e.model,usage:{input_tokens:d,output_tokens:c},stop_reason:(a.done,"end_turn"),provider:"ollama",costUsd:0,durationMs:o,fallbackChain:["ollama:success"]}}async streamRaw(e,t){let r=this.buildRequestBody(e,true),s=await fetch(`${this.baseUrl}/api/chat`,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify(r),signal:t?.signal});if(s.ok||this.handleErrorResponse(s),!s.body)throw new Error("OLLAMA_ERROR: Response body is null \u2014 streaming unavailable");let i=s.body.getReader(),a=new TextDecoder,o=e.model;return (async function*(){let l="",d=false,c=false;yield {type:"message_start",message:{model:o,usage:{input_tokens:0,output_tokens:0}}};try{for(;;){let{done:p,value:y}=await i.read();if(p)break;l+=a.decode(y,{stream:!0});let m=l.split(`
|
|
13
|
+
`);l=m.pop()||"";for(let E of m){let u=E.trim();if(u)try{let f=JSON.parse(u);if(f.message?.content&&(d||(d=!0,yield {type:"content_block_start",index:0,content_block:{type:"text",text:""}}),yield {type:"content_block_delta",index:0,delta:{type:"text_delta",text:f.message.content}}),f.done){d&&(yield {type:"content_block_stop",index:0}),yield {type:"message_delta",delta:{stop_reason:"end_turn"},usage:{input_tokens:f.prompt_eval_count||0,output_tokens:f.eval_count||0}},yield {type:"message_stop"},c=!0;return}}catch{}}}c||(yield {type:"message_stop"});}finally{try{i.releaseLock();}catch{}}})()}estimateCost(e){let t=Math.round(JSON.stringify(e.messages).length/4),r=e.system?Math.round(e.system.length/4):0;return {estimatedInputTokens:t+r,estimatedCost:0,model:e.model,provider:"ollama"}}getHealth(){return {...this.health}}reportRateLimit(e){this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now();}buildRequestBody(e,t){if(e.tools&&e.tools.length>0)throw new Error(`OLLAMA_TOOLS_NOT_SUPPORTED: The Ollama provider received ${e.tools.length} tool definition(s) in params.tools, but this provider does not implement function calling in v1.3.x. For tool-use workflows route to 'anthropic-api' or 'openai-api' instead. Tracking issue: 1.4.0 ships an Ollama tool-use shim for models that support native function calling (llama3.1, mistral-nemo, etc.).`);let r=Ct(e.messages,e.system),s={model:e.model,messages:r,stream:t};return e.temperature!==void 0&&(s.options={temperature:e.temperature}),s}handleErrorResponse(e){let t=e.status;throw this.health.errorCount++,this.health.lastCheckAt=Date.now(),t===404?(this.health.status="unhealthy",this.health.lastError="Model not found",new Error("OLLAMA_MODEL_NOT_FOUND: Requested model is not available locally (404)")):(this.health.status="unhealthy",this.health.lastError=`Request failed with status ${t}`,new Error(`OLLAMA_ERROR: Request failed with status ${t}`))}recordSuccess(){this.health.status="healthy",this.health.lastCheckAt=Date.now(),this.health.errorCount=0,this.health.lastError=void 0;}};var $=class extends events.EventEmitter{constructor(){super(),this.setMaxListeners(50);}emitTyped(e){return this.emit(e.kind,e)}onTyped(e,t){return this.on(e,t)}offTyped(e,t){return this.off(e,t)}};var We=200;function ne(n,e){if(e instanceof b)return `${n}:failed:${e.code}:${e.message}`;let t=e?.message;return `${n}:failed:${t?.slice(0,We)||"unknown"}`}var Fe=new Set(["anthropic-api","openai-api","ollama"]),q=class extends Error{code="ALL_PROVIDERS_FAILED";fallbackChain;constructor(e){super(`All providers failed. Chain: ${e.join(" \u2192 ")}`),this.name="RelayError",this.fallbackChain=e;}};function Be(n){return n.messages.some(e=>typeof e.content=="string"?false:Array.isArray(e.content)&&e.content.some(t=>t.type==="image"||t.type==="image_url"))}function Ee(n){if(!(n instanceof Error))return "unknown";let e=n.code;if(typeof e=="string"){if(e==="LICENSE_REJECTED")return "auth_failed";if(e==="GATEWAY_FETCH_FAILED"||e==="GATEWAY_VERIFICATION_FAILED"||e==="GATEWAY_LOAD_FAILED")return "gateway_unavailable"}let t=n.message.toLowerCase(),r=n.name.toLowerCase();return t.startsWith("openai_model_translation_required:")||t.startsWith("ollama_tools_not_supported:")||t.startsWith("multimodal_cli_too_old:")||t.startsWith("stream_json_cli_too_old:")?"provider_incompatible":t.includes("rate limit")||t.includes("429")||r.includes("ratelimit")?"rate_limited":t.includes("auth")||t.includes("401")||t.includes("403")||t.includes("unauthorized")||t.includes("forbidden")?"auth_failed":t.includes("timeout")||t.includes("timed out")||t.includes("aborted")||r==="aborterror"||r==="timeouterror"?"timeout":"unknown"}function He(n){return [...n].sort((e,t)=>e.priority-t.priority)}var j=class n{healthMap=new Map;engineEvents=new $;static UNHEALTHY_TTL_MS=300*1e3;getHealthStatus(e){let t=this.healthMap.get(e);return t?t.status==="unhealthy"&&Date.now()-t.updatedAt>=n.UNHEALTHY_TTL_MS?(this.healthMap.delete(e),"healthy"):t.status:"healthy"}markUnhealthy(e,t){this.healthMap.set(e,{status:"unhealthy",lastError:t,updatedAt:Date.now()});}markDegraded(e,t){this.healthMap.set(e,{status:"degraded",lastError:t,updatedAt:Date.now()});}markHealthy(e){this.healthMap.set(e,{status:"healthy",updatedAt:Date.now()});}resetHealth(){this.healthMap.clear();}async execute(e,t,r){let s=He(t),i=[],a=Be(e),o=crypto$1.randomUUID(),l=Date.now();for(let d of s){let c=d.type;if(this.getHealthStatus(c)==="unhealthy"){i.push(`${c}:skipped:unhealthy`);continue}if(a&&!d.supportsMultimodal){i.push(`${c}:skipped:no_multimodal`);continue}let y=re(e.model);if(y&&y!==c&&Fe.has(c)){i.push(`${c}:skipped:wrong_provider`);continue}try{let m=await d.execute(e,{signal:r?.signal});return this.markHealthy(c),i.push(`${c}:success`),this.emitEngineEvent({kind:"stream_complete",callId:o,id:crypto$1.randomUUID(),ts:Date.now(),stitched:!1,segments:[{provider:c,inputTokens:m.usage.input_tokens,outputTokens:m.usage.output_tokens,costUsd:m.costUsd,userCostUsd:m.userCostUsd??0}],provider:c,model:m.model,inputTokens:m.usage.input_tokens,outputTokens:m.usage.output_tokens,costUsd:m.costUsd,userCostUsd:m.userCostUsd,durationMs:m.durationMs??Date.now()-l,stopReason:m.stop_reason,fallbackChain:i,rateLimitHit:!1,securityState:r?.licenseState}),{...m,fallbackChain:i}}catch(m){if(m?.name==="AbortError"||r?.signal?.aborted)throw m;switch(Ee(m)){case "rate_limited":d.reportRateLimit({provider:c,utilization:1,resetsAt:Date.now()+6e4}),this.markDegraded(c,"rate_limited"),i.push(`${c}:failed:rate_limited`);break;case "auth_failed":this.markUnhealthy(c,"auth_failed"),i.push(`${c}:failed:auth_failed`);break;case "timeout":i.push(`${c}:failed:timeout`);break;case "provider_incompatible":i.push(`${c}:incompatible`);break;case "gateway_unavailable":this.markUnhealthy(c,"gateway_unavailable"),i.push(ne(c,m));break;default:i.push(ne(c,m));break}}}throw new q(i)}async stream(e,t,r){let s=crypto$1.randomUUID(),i=He(t),a=Be(e),o=Date.now(),l=[],d=null,c=null;for(let h of i){let g=h.type;if(this.getHealthStatus(g)==="unhealthy"){l.push(`${g}:skipped:unhealthy`);continue}if(a&&!h.supportsMultimodal){l.push(`${g}:skipped:no_multimodal`);continue}if(!h.supportsStreaming){l.push(`${g}:skipped:no_streaming`);continue}let C=re(e.model);if(C&&C!==g&&Fe.has(g)){l.push(`${g}:skipped:wrong_provider`);continue}try{c=await h.streamRaw(e,{signal:r?.signal}),d=h,l.push(`${g}:success`);break}catch(_){if(_?.name==="AbortError"||r?.signal?.aborted)throw _;switch(Ee(_)){case "rate_limited":h.reportRateLimit({provider:g,utilization:1,resetsAt:Date.now()+6e4}),this.markDegraded(g,"rate_limited"),l.push(`${g}:failed:rate_limited`);break;case "auth_failed":this.markUnhealthy(g,"auth_failed"),l.push(`${g}:failed:auth_failed`);break;case "timeout":l.push(`${g}:failed:timeout`);break;case "provider_incompatible":l.push(`${g}:incompatible`);break;case "gateway_unavailable":this.markUnhealthy(g,"gateway_unavailable"),l.push(ne(g,_));break;default:l.push(ne(g,_));break}}}if(!d||!c)throw new q(l);let y=this,m=r?.maxMidStreamFailovers??2,E=new Set([d]);async function*u(){let h=d,g=c[Symbol.asyncIterator](),C={role:"assistant",content:[]},_={segments:[]},I=Ge(h.type);_.segments.push(I);let Re=0,z=null,X=false,de=m,O="",N="";for(y.emitEngineEvent({kind:"segment_start",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),provider:h.type});;)try{for(;;){let A=await g.next();if(A.done)break;let T=A.value,k=Pt(T,Re,C,X,z);if(k.action==="suppress"){"coalesceMode"in k&&(z=k.coalesceMode),k.clearedJustFailedOver&&(X=!1);continue}k.coalesceModeCleared&&(z=null),k.clearedJustFailedOver&&(X=!1);let v=k.event;if(St(C,v),v.type==="message_start"){if(_.segments.length>1)continue;O=v.message?.model||O,v.message?.usage?.input_tokens!=null&&(I.inputTokens=v.message.usage.input_tokens);}if(v.type==="message_delta"&&(v.delta?.stop_reason&&(N=v.delta.stop_reason),v.usage)){Rt(I,h.type,v.usage),yield {...v,usage:void 0};continue}if(v.type==="message_stop")break;yield v;}if(C.content.length===0&&I.outputTokens===0)throw new Error("stream ended without producing any content");y.markHealthy(h.type),y.emitEngineEvent({kind:"segment_complete",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),provider:h.type,usage:{...I}});let S=Ke(_.segments);yield {type:"message_delta",delta:{stop_reason:N||"end_turn"},usage:S},yield {type:"message_stop"},y.emitEngineEvent({kind:"stream_complete",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),stitched:_.segments.length>1,segments:_.segments.map(A=>({...A})),provider:_.segments.map(A=>A.provider).join("+"),model:O,inputTokens:S.input_tokens,outputTokens:S.output_tokens,costUsd:_.segments.reduce((A,T)=>A+T.costUsd,0),userCostUsd:_.segments.reduce((A,T)=>A+T.userCostUsd,0),durationMs:Date.now()-o,stopReason:N||"end_turn",fallbackChain:l,rateLimitHit:!1,securityState:r?.licenseState});return}catch(S){if(S?.name==="AbortError"||r?.signal?.aborted)throw S;let A=Ee(S);try{await g.return?.();}catch{}if(y.markUnhealthy(h.type,String(S)),y.emitEngineEvent({kind:"mid_stream_failover_attempt",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),from:h.type,bytesBuffered:Ve(C),errorClass:A,retriesLeft:de}),l.push(`${h.type}:partial(${Ve(C)}b)`),de<=0)throw y.emitTerminalExhaustion(s,_,O,N,l,o,r?.licenseState),S;let T=kt(C.content),k=[...e.messages,{role:"assistant",content:T}],v=null,Q=null,Se=h;for(;;){let w=y.nextHealthyProvider(i,Se,a,E);if(!w)break;Se=w,E.add(w);let je={...e,messages:k,...w.type!=="anthropic-api"&&w.type!=="claude-cli"?{system:wt(e.system,"[SYSTEM]: CONTINUE_LAST_RESPONSE_WITHOUT_REPEATING")}:{}};try{Q=await w.streamRaw(je,{signal:r?.signal}),v=w;break}catch(Pe){y.markUnhealthy(w.type,String(Pe)),l.push(`${w.type}:failed_connect:${String(Pe).slice(0,We)}`);}}if(!v||!Q)throw y.emitTerminalExhaustion(s,_,O,N,l,o,r?.licenseState),S;l.push(`${v.type}:stitched`),y.emitEngineEvent({kind:"mid_stream_failover_success",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),from:h.type,to:v.type}),Re=It(C)+1,de--,h=v,g=Q[Symbol.asyncIterator](),X=true,z=null,I=Ge(h.type),_.segments.push(I),y.emitEngineEvent({kind:"segment_start",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),provider:h.type});}}let f=u();return r?.licenseState==="SEC_DEGRADED"&&(f=Tt(f)),L(f)}nextHealthyProvider(e,t,r,s){let i=e.indexOf(t);if(i===-1)return null;for(let a=i+1;a<e.length;a++){let o=e[a];if(!s.has(o)&&!(r&&!o.supportsMultimodal)&&o.supportsStreaming)return o}return null}emitEngineEvent(e){this.engineEvents.emitTyped(e);}emitTerminalExhaustion(e,t,r,s,i,a,o){let l=Ke(t.segments);this.emitEngineEvent({kind:"stream_complete",callId:e,id:crypto$1.randomUUID(),ts:Date.now(),stitched:t.segments.length>1,segments:t.segments.map(d=>({...d})),provider:t.segments.map(d=>d.provider).join("+"),model:r,inputTokens:l.input_tokens,outputTokens:l.output_tokens,costUsd:t.segments.reduce((d,c)=>d+c.costUsd,0),userCostUsd:t.segments.reduce((d,c)=>d+c.userCostUsd,0),durationMs:Date.now()-a,stopReason:s||"error",fallbackChain:i,rateLimitHit:false,securityState:o});}};function Ge(n){return {provider:n,inputTokens:0,outputTokens:0,costUsd:0,userCostUsd:0}}function Rt(n,e,t){typeof t.input_tokens=="number"&&(n.inputTokens=t.input_tokens),typeof t.output_tokens=="number"&&(e==="anthropic-api"||e==="claude-cli"?n.outputTokens+=t.output_tokens:n.outputTokens=t.output_tokens);}function Ke(n){return {input_tokens:n[n.length-1]?.inputTokens??0,output_tokens:n.reduce((e,t)=>e+t.outputTokens,0)}}function St(n,e){if(e.type==="content_block_start"){let t=e.content_block;t?.type==="text"?n.content.push({type:"text",text:t.text||"",open:true}):t?.type==="tool_use"&&n.content.push({type:"tool_use",id:t.id,name:t.name,input:{},_partialJson:"",open:true});}else if(e.type==="content_block_delta"){let t=n.content[n.content.length-1];if(!t)return;t.type==="text"&&e.delta?.type==="text_delta"?t.text+=e.delta.text||"":t.type==="tool_use"&&e.delta?.type==="input_json_delta"&&(t._partialJson=(t._partialJson||"")+(e.delta.partial_json||""));}else if(e.type==="content_block_stop"){for(let t=n.content.length-1;t>=0;t--)if(n.content[t].open){if(n.content[t].open=false,n.content[t].type==="tool_use"){let r=n.content[t];try{r.input=JSON.parse(r._partialJson||"{}");}catch{}delete r._partialJson;}break}}}function Pt(n,e,t,r,s){if(s&&n.type==="content_block_delta"&&n.index===s.fromIndex)return {action:"yield",event:{...n,index:s.toIndex}};if(s&&n.type==="content_block_stop"&&n.index===s.fromIndex)return {action:"suppress",coalesceMode:null,coalesceModeCleared:true};if(r&&n.type==="content_block_start"&&n.content_block?.type==="text"){let a=t.content[t.content.length-1];if(a?.type==="text"&&a.open===true)return {action:"suppress",coalesceMode:{fromIndex:n.index+e,toIndex:t.content.length-1},clearedJustFailedOver:true}}let i=r&&n.type==="content_block_start";return typeof n.index=="number"&&e>0?{action:"yield",event:{...n,index:n.index+e},clearedJustFailedOver:i}:{action:"yield",event:n,clearedJustFailedOver:i}}function kt(n){let e=[];for(let t of n)t.type==="text"?t.text&&e.push({type:"text",text:t.text}):t.type==="tool_use"&&!t.open&&t.input&&Object.keys(t.input).length>0&&e.push({type:"tool_use",id:t.id,name:t.name,input:t.input});return e}function wt(n,e){return n?`${n}
|
|
14
14
|
|
|
15
15
|
${e}`:e}function Ve(n){let e=0;for(let t of n.content)t.type==="text"?e+=t.text.length:t.type==="tool_use"&&(e+=(t._partialJson||"").length);return e}function It(n){return Math.max(0,n.content.length-1)}async function*Tt(n){for await(let e of n)await new Promise(t=>setTimeout(t,50+Math.random()*50)),yield e;}ee();function se(n,e){if(e==="prefer-user"){let s=n.filter(a=>a.type==="claude-cli"),i=n.filter(a=>a.type!=="claude-cli");return [...s,...i]}if(e==="prefer-app"){let s=n.filter(a=>a.type==="claude-cli");return [...n.filter(a=>a.type!=="claude-cli"),...s]}let t=n.filter(s=>s.type===e),r=n.filter(s=>s.type!==e);return [...t,...r]}function ie(n){return (e,t)=>n.maxCostPerCall!==void 0&&t.estimatedCost>n.maxCostPerCall?"prefer-user":n.preferLocal?"ollama":"prefer-user"}var ve={"claude-opus-4":[15,75],"claude-sonnet-4":[3,15],"claude-haiku-4":[.8,4],"gpt-4o":[2.5,10],"gpt-4o-mini":[.15,.6],"gpt-4-turbo":[10,30],"gpt-3.5-turbo":[.5,1.5],llama:[0,0],mistral:[0,0],codellama:[0,0]};function ae(n){if(ve[n])return ve[n];for(let[e,t]of Object.entries(ve))if(n.startsWith(e))return t;return [3,15]}function oe(n){let e=0;for(let o of n.messages)if(typeof o.content=="string")e+=o.content.length;else if(Array.isArray(o.content))for(let l of o.content)l.type==="text"&&l.text&&(e+=l.text.length);n.system&&(e+=n.system.length);let t=Math.ceil(e/4),r=Math.min(n.max_tokens||1024,4096),[s,i]=ae(n.model),a=(t*s+r*i)/1e6;return {estimatedInputTokens:t,estimatedCost:Math.round(a*1e4)/1e4,model:n.model,provider:""}}var D=class{endpoint;enabled;keyId;lineageBuffer=new Map;unsubscribers=[];constructor(e){this.endpoint=e.endpoint||"https://api.inference-relay.com/v1/events",this.enabled=e.enabled??true,this.keyId=e.keyId||"";}subscribeToEngine(e){let t=s=>this.bufferLineageEvent(s),r=s=>{s.kind==="stream_complete"&&this.flushLineageForCall(s);};e.engineEvents.onTyped("segment_start",t),e.engineEvents.onTyped("segment_complete",t),e.engineEvents.onTyped("mid_stream_failover_attempt",t),e.engineEvents.onTyped("mid_stream_failover_success",t),e.engineEvents.onTyped("stream_complete",r),this.unsubscribers.push(()=>{e.engineEvents.offTyped("segment_start",t),e.engineEvents.offTyped("segment_complete",t),e.engineEvents.offTyped("mid_stream_failover_attempt",t),e.engineEvents.offTyped("mid_stream_failover_success",t),e.engineEvents.offTyped("stream_complete",r);});}unsubscribe(){for(let e of this.unsubscribers)e();this.unsubscribers=[],this.lineageBuffer.clear();}bufferLineageEvent(e){let t=this.lineageBuffer.get(e.callId)??[];t.push(e),this.lineageBuffer.set(e.callId,t);}flushLineageForCall(e){let t=this.lineageBuffer.get(e.callId)??[];this.lineageBuffer.delete(e.callId);let r={timestamp:e.ts,provider:e.provider,model:e.model,inputTokens:e.inputTokens,outputTokens:e.outputTokens,costUsd:e.costUsd,userCostUsd:e.userCostUsd,durationMs:e.durationMs,stopReason:e.stopReason,fallbackChain:e.fallbackChain,rateLimitHit:e.rateLimitHit,securityState:e.securityState,stitched:e.stitched,segments:e.segments,lineage:t.map(s=>s.id),promptContent:false,completionContent:false};this.report(r);}report(e){!this.enabled||!this.keyId||fetch(this.endpoint,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({keyId:this.keyId,provider:e.provider,model:e.model,inputTokens:e.inputTokens,outputTokens:e.outputTokens,costUsd:e.costUsd,...e.userCostUsd!==void 0?{userCostUsd:e.userCostUsd}:{},durationMs:e.durationMs,fallback:e.fallbackChain.length>1,timestamp:e.timestamp,...e.stitched?{stitched:true,segmentCount:e.segments?.length??0}:{},...e.lineage?{lineageIds:e.lineage}:{}})}).catch(()=>{});}};var xt="https://api.inference-relay.com",Lt="/v1/validate",Mt=1440*60*1e3,Dt=10080*60*1e3,Ot=`-----BEGIN PUBLIC KEY-----
|
|
16
16
|
MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEAzk88Bt9LZSmuSd8ulKqy
|
package/dist/index.js
CHANGED
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
import {execSync}from'child_process';import {mkdirSync,writeFileSync,existsSync,readFileSync}from'fs';import {join,dirname}from'path';import {homedir}from'os';import {createHash,randomUUID}from'crypto';import {compileFunction}from'vm';import {createRequire}from'module';import {EventEmitter}from'events';/*! inference-relay (c) 2026 L2B II LLC. Patent pending. Proprietary. See LICENSE. */
|
|
2
|
-
var Je=Object.defineProperty;var ze=(n=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(n,{get:(e,t)=>(typeof require<"u"?require:e)[t]}):n)(function(n){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+n+'" is not supported')});var Xe=(n,e)=>()=>(n&&(e=n(n=0)),e);var Qe=(n,e)=>{for(var t in e)Je(n,t,{get:e[t],enumerable:true});};var Me={};Qe(Me,{detectCLI:()=>H});function dt(n){let e=process.env.PATH;if(!e)return null;let t=process.platform==="win32",r=t?";":":",s=t?(process.env.PATHEXT??".COM;.EXE;.BAT;.CMD").split(";").map(i=>i.toLowerCase()):[""];for(let i of e.split(r))if(i)for(let a of s){let o=join(i,n+a);if(existsSync(o))return o}return null}async function H(){if(!dt("claude"))return {available:false,authenticated:false};let e=null;try{let r=execSync("claude --version",{encoding:"utf-8",timeout:5e3}).trim(),s=r.match(/(\d+\.\d+\.\d+)/);e=s?s[1]:r;}catch{}let t;return e&&!ut.some(r=>e.startsWith(r.split(".").slice(0,2).join(".")))&&(t=`CLI version ${e} is not tested with this version of inference-relay. Streaming may be unstable.`),{available:true,authenticated:true,cliVersion:e||void 0,versionWarning:t}}var ut,ee=Xe(()=>{ut=["2.1.87"];});var U=class n extends Error{code="RELAY_CAP_EXCEEDED";tier;baseCap;topUps;effectiveCap;usedThisCycle;resolutionUrl;constructor(e){let t=e.resolutionUrl??"https://inference-relay.com/dashboard/billing?topup=1";super(`Cap reached (${e.usedThisCycle} / ${e.effectiveCap??"\u221E"} calls this cycle). Buy a top-up or upgrade at ${t}`),this.name="RelayCapExceededError",this.tier=e.tier,this.baseCap=e.baseCap,this.topUps=e.topUps,this.effectiveCap=e.effectiveCap,this.usedThisCycle=e.usedThisCycle,this.resolutionUrl=t,Object.setPrototypeOf(this,n.prototype);}};function L(n,e){let t=new Map,r=null,s=null,i=new Promise((u,f)=>{r=u,s=f;});i.catch(()=>{});let a=[],o=null,l="",d="";function c(u,...f){let h=t.get(u);h&&h.forEach(g=>g(...f));}async function*p(){try{for await(let u of n)if(yield u,u.type==="content_block_start")a[u.index]=u.content_block;else if(u.type==="content_block_delta"){let f=a[u.index];f&&u.delta?.type==="text_delta"&&(f.text=(f.text||"")+u.delta.text,c("text",u.delta.text));}else if(u.type==="message_start")l=u.message?.model||"",o=u.message?.usage;else if(u.type==="message_delta")d=u.delta?.stop_reason||"",u.usage&&(o={...o,...u.usage});else if(u.type==="message_stop"){let f={role:"assistant",model:l,content:a.filter(Boolean),usage:o,stop_reason:d};c("message",f),r&&(r(f),r=null),e&&e(f);}if(r){let u={role:"assistant",model:l,content:a.filter(Boolean),usage:o,stop_reason:d||"end_turn"};c("message",u),r(u),r=null,e&&e(u);}}catch(u){throw c("error",u),s&&(s(u),s=null),u}}let y=p(),m=false,
|
|
2
|
+
var Je=Object.defineProperty;var ze=(n=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(n,{get:(e,t)=>(typeof require<"u"?require:e)[t]}):n)(function(n){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+n+'" is not supported')});var Xe=(n,e)=>()=>(n&&(e=n(n=0)),e);var Qe=(n,e)=>{for(var t in e)Je(n,t,{get:e[t],enumerable:true});};var Me={};Qe(Me,{detectCLI:()=>H});function dt(n){let e=process.env.PATH;if(!e)return null;let t=process.platform==="win32",r=t?";":":",s=t?(process.env.PATHEXT??".COM;.EXE;.BAT;.CMD").split(";").map(i=>i.toLowerCase()):[""];for(let i of e.split(r))if(i)for(let a of s){let o=join(i,n+a);if(existsSync(o))return o}return null}async function H(){if(!dt("claude"))return {available:false,authenticated:false};let e=null;try{let r=execSync("claude --version",{encoding:"utf-8",timeout:5e3}).trim(),s=r.match(/(\d+\.\d+\.\d+)/);e=s?s[1]:r;}catch{}let t;return e&&!ut.some(r=>e.startsWith(r.split(".").slice(0,2).join(".")))&&(t=`CLI version ${e} is not tested with this version of inference-relay. Streaming may be unstable.`),{available:true,authenticated:true,cliVersion:e||void 0,versionWarning:t}}var ut,ee=Xe(()=>{ut=["2.1.87"];});var U=class n extends Error{code="RELAY_CAP_EXCEEDED";tier;baseCap;topUps;effectiveCap;usedThisCycle;resolutionUrl;constructor(e){let t=e.resolutionUrl??"https://inference-relay.com/dashboard/billing?topup=1";super(`Cap reached (${e.usedThisCycle} / ${e.effectiveCap??"\u221E"} calls this cycle). Buy a top-up or upgrade at ${t}`),this.name="RelayCapExceededError",this.tier=e.tier,this.baseCap=e.baseCap,this.topUps=e.topUps,this.effectiveCap=e.effectiveCap,this.usedThisCycle=e.usedThisCycle,this.resolutionUrl=t,Object.setPrototypeOf(this,n.prototype);}};function L(n,e){let t=new Map,r=null,s=null,i=new Promise((u,f)=>{r=u,s=f;});i.catch(()=>{});let a=[],o=null,l="",d="";function c(u,...f){let h=t.get(u);h&&h.forEach(g=>g(...f));}async function*p(){try{for await(let u of n)if(yield u,u.type==="content_block_start")a[u.index]=u.content_block;else if(u.type==="content_block_delta"){let f=a[u.index];f&&u.delta?.type==="text_delta"&&(f.text=(f.text||"")+u.delta.text,c("text",u.delta.text));}else if(u.type==="message_start")l=u.message?.model||"",o=u.message?.usage;else if(u.type==="message_delta")d=u.delta?.stop_reason||"",u.usage&&(o={...o,...u.usage});else if(u.type==="message_stop"){let f={role:"assistant",model:l,content:a.filter(Boolean),usage:o,stop_reason:d};c("message",f),r&&(r(f),r=null),e&&e(f);}if(r){let u={role:"assistant",model:l,content:a.filter(Boolean),usage:o,stop_reason:d||"end_turn"};c("message",u),r(u),r=null,e&&e(u);}}catch(u){throw c("error",u),s&&(s(u),s=null),u}}let y=p(),m=false,E={[Symbol.asyncIterator](){if(m)throw new Error("[inference-relay] RelayStream can only be iterated once");return m=true,y},on(u,f){return t.has(u)||t.set(u,[]),t.get(u).push(f),E},finalMessage(){let u,f=new Promise((h,g)=>{u=setTimeout(()=>g(new Error("[inference-relay] finalMessage() timed out \u2014 no message_stop received after 5 minutes")),3e5);});return Promise.race([i,f]).finally(()=>clearTimeout(u))}};return E}var R=class{priority=100;async streamRaw(e,t){throw new Error(`[inference-relay] Provider '${this.type}' does not implement streamRaw(). Override either stream() or streamRaw() in the provider class.`)}async stream(e,t){return L(await this.streamRaw(e,t))}};var me=`-----BEGIN PUBLIC KEY-----
|
|
3
3
|
MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEAzk88Bt9LZSmuSd8ulKqy
|
|
4
4
|
EGNiPB+r8u1ZinBvMM4qX8dp4kzJEC/47L1RYrrxCphD05HToREjljbj+upH6rZH
|
|
5
5
|
t6D/5funYPi2TO1IOSC6BRVRwMqqTmJGVJcLcG/WF3gTE+2UQ3q5EppWJCEKcz6x
|
|
@@ -7,10 +7,10 @@ kfynZEzOcLikoQALIxEjxRfi6g83wNJS+WBwAJQ68s7N9FHE3rkLzEqv3FJRP+ZR
|
|
|
7
7
|
7lqwqIyuCLr9j9yc4XqF6VY9yooD2hyw/n1EVFejzjDgntWTqIkaCIXwh8lDzRrK
|
|
8
8
|
Rxn07gcI6ea2Eym9sXh4xZLFudXfAT2DFsH/XboDUuMKAV1YHc1NsIGApesWe2M+
|
|
9
9
|
UQIDAQAB
|
|
10
|
-
-----END PUBLIC KEY-----`,pe={version:"1.0.0",cliVersion:"2.1.87",eventTypes:{system:"system",streamEvent:"stream_event",assistant:"assistant",result:"result"},fieldMappings:{sessionId:"session_id",model:"model",cliVersion:"claude_code_version",content:"content",usage:"usage",inputTokens:"input_tokens",outputTokens:"output_tokens",stopReason:"stop_reason",isError:"is_error",totalCostUsd:"total_cost_usd",durationMs:"duration_ms",error:"error",event:"event"},syntheticModel:"<synthetic>"},M=class{verificationFailures=0;cachedManifest=null;currentSecurityState="SEC_NOMINAL";async verifyManifest(e){try{let{compactVerify:t,importSPKI:r}=await import('jose'),s=await r(me,"RS256"),{payload:i}=await t(e,s),a=JSON.parse(new TextDecoder().decode(i));return this.verificationFailures=0,this.cachedManifest=a,this.currentSecurityState="SEC_NOMINAL",{manifest:a,securityState:"SEC_NOMINAL"}}catch{if(this.verificationFailures++,this.verificationFailures>=3){this.currentSecurityState="SEC_DEGRADED";let t=2e3+Math.random()*3e3;return await new Promise(r=>setTimeout(r,t)),console.warn("[inference-relay] Logic synchronization mismatch. Routing via local recovery."),{manifest:pe,securityState:"SEC_DEGRADED"}}return this.cachedManifest?(this.currentSecurityState="SEC_CACHE",{manifest:this.cachedManifest,securityState:"SEC_CACHE"}):(this.currentSecurityState="SEC_RECOVERY",{manifest:pe,securityState:"SEC_RECOVERY"})}}getSecurityState(){return this.currentSecurityState}clearCache(){this.cachedManifest=null,this.verificationFailures=0,this.currentSecurityState="SEC_NOMINAL";}},Ze=new M;function ke(){Ze.clearCache();}function et(){return pe}function tt(n,e){return e.split(".").reduce((t,r)=>{if(!(r==="__proto__"||r==="constructor"||r==="prototype"))return t&&t[r]},n)}var at=process.env.IR_GATEWAY_URL||"https://api.inference-relay.com";function ge(n){return join(homedir(),".inference-relay","gateway",n)}function Te(n){return join(ge(n),"bundle.js")}function xe(n){return join(ge(n),"meta.json")}var b=class extends Error{constructor(t,r,s){super(t);this.code=r;this.cause=s;this.name="GatewayLoaderError";}code;cause},F=new Map,B=class{static async load(e,t,r=at){let s=`${t}:${e.slice(-8)}`;if(F.has(s))return F.get(s);try{let p=await this.tryLoadFromDisk(t);if(p)return F.set(s,p),p}catch{}let{bytes:i,sha256:a,signature:o,libraryVersion:l}=await this.fetchFromAuthority(e,t,r);if(createHash("sha256").update(i).digest("hex")!==a)throw new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: HASH_MISMATCH)","GATEWAY_VERIFICATION_FAILED");await this.verifySignature(o,l,a);try{let p=ge(t);mkdirSync(p,{recursive:!0}),writeFileSync(Te(t),i),writeFileSync(xe(t),JSON.stringify({sha256:a,signature:o,libraryVersion:l}));}catch{}let c=this.loadIntoVm(i.toString("utf-8"));return F.set(s,c),c}static async loadFromDisk(e){if(!existsSync(e))throw new b(`[ir-system]: GATEWAY_LOAD_FAILED (REASON: BUNDLE_NOT_FOUND at ${e})`,"GATEWAY_LOAD_FAILED");let t=readFileSync(e,"utf-8");return this.loadIntoVm(t)}static async tryLoadFromDisk(e){let t=Te(e),r=xe(e);if(!existsSync(t)||!existsSync(r))return null;let s=readFileSync(t),i=JSON.parse(readFileSync(r,"utf-8"));return createHash("sha256").update(s).digest("hex")!==i.sha256?null:(await this.verifySignature(i.signature,i.libraryVersion,i.sha256),this.loadIntoVm(s.toString("utf-8")))}static async fetchFromAuthority(e,t,r){let s,i=[0,1e3,4e3],a;for(let c of i){c>0&&await new Promise(p=>setTimeout(p,c));try{let p=await fetch(`${r}/v1/gateway/fetch`,{method:"POST",headers:{"Content-Type":"application/json",Authorization:`Bearer ${e}`},body:JSON.stringify({libraryVersion:t})});if(p.status===401||p.status===403)throw new b("[ir-system]: LICENSE_REJECTED (REASON: GATEWAY_FETCH_UNAUTHORIZED)","LICENSE_REJECTED");if(p.status>=500&&p.status<600){a=new b(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: HTTP_${p.status})`,"GATEWAY_FETCH_FAILED");continue}if(!p.ok)throw new b(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: HTTP_${p.status})`,"GATEWAY_FETCH_FAILED");s=await p.json(),a=null;break}catch(p){if(p instanceof b&&p.code==="LICENSE_REJECTED")throw p;a=p;}}if(a||!s)throw a instanceof b?a:new b("[ir-system]: GATEWAY_FETCH_FAILED (REASON: NETWORK_ERROR)","GATEWAY_FETCH_FAILED",a);let o;try{o=await fetch(s.bundleUrl);}catch(c){throw new b("[ir-system]: GATEWAY_FETCH_FAILED (REASON: BUNDLE_DOWNLOAD_FAILED)","GATEWAY_FETCH_FAILED",c)}if(!o.ok)throw new b(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: BUNDLE_HTTP_${o.status})`,"GATEWAY_FETCH_FAILED");let l=await o.arrayBuffer();return {bytes:Buffer.from(l),sha256:s.sha256,signature:s.signature,libraryVersion:s.libraryVersion,bundleVersion:s.bundleVersion}}static async verifySignature(e,t,r){try{let{compactVerify:s,importSPKI:i}=await import('jose'),a=await i(me,"RS256"),{payload:o}=await s(e,a),l=JSON.parse(new TextDecoder().decode(o));if(l.purpose!=="gateway-bundle")throw new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: PURPOSE_MISMATCH)","GATEWAY_VERIFICATION_FAILED");if(l.libraryVersion!==t)throw new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: LIBRARY_VERSION_MISMATCH)","GATEWAY_VERIFICATION_FAILED");if(l.sha256!==r)throw new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: SHA256_MISMATCH)","GATEWAY_VERIFICATION_FAILED")}catch(s){throw s instanceof b?s:new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: SIGNATURE_INVALID)","GATEWAY_VERIFICATION_FAILED",s)}}static loadIntoVm(e){try{let t=join(homedir(),".inference-relay","gateway","runtime.cjs"),r=dirname(t),s=createRequire(t),i={exports:{}};compileFunction(e,["exports","require","module","__filename","__dirname"],{filename:"<gateway-bundle>"})(i.exports,s,i,t,r);let o=i.exports;if(!o.ClaudeCLIProvider)throw new b("[ir-system]: GATEWAY_LOAD_FAILED (REASON: MISSING_CLAUDE_CLI_EXPORT)","GATEWAY_LOAD_FAILED");return o}catch(t){throw t instanceof b?t:new b("[ir-system]: GATEWAY_LOAD_FAILED (REASON: VM_EXECUTION_ERROR)","GATEWAY_LOAD_FAILED",t)}}static clearCache(){F.clear();}};var pt="1.5.2",te=class extends R{type="claude-cli";supportsStreaming=true;supportsMultimodal=true;platformRequirement="desktop";licenseKey;loadedProvider=null;loadingPromise=null;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.licenseKey=e;}async ensureLoaded(){return this.loadedProvider?this.loadedProvider:this.loadingPromise?this.loadingPromise:(this.loadingPromise=(async()=>{try{let t=(await B.load(this.licenseKey,pt)).ClaudeCLIProvider,r=new t;return this.loadedProvider=r,r}catch(e){throw this.loadingPromise=null,this.health.status="unhealthy",this.health.errorCount++,e instanceof b?this.health.lastError=e.message:e instanceof Error&&(this.health.lastError=e.message),e}})(),this.loadingPromise)}async probe(){let{detectCLI:e}=await Promise.resolve().then(()=>(ee(),Me));return e()}async execute(e,t){return (await this.ensureLoaded()).execute(e,t)}async streamRaw(e,t){return (await this.ensureLoaded()).streamRaw(e,t)}estimateCost(e){if(this.loadedProvider)return this.loadedProvider.estimateCost(e);let r=JSON.stringify(e.messages).length/4+5e3,s=(r*3+1e3*15)/1e6;return {estimatedInputTokens:Math.round(r),estimatedCost:Math.round(s*1e4)/1e4,model:e.model,provider:"claude-cli"}}getHealth(){return this.loadedProvider?this.loadedProvider.getHealth():{...this.health}}reportRateLimit(e){this.loadedProvider?this.loadedProvider.reportRateLimit(e):(this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now());}};var G=[{canonical:"claude-haiku-4-5",apiId:"claude-haiku-4-5-20251001",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[.8,4]},{canonical:"claude-sonnet-4-6",apiId:"claude-sonnet-4-20250514",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[3,15]},{canonical:"claude-opus-4-6",apiId:"claude-opus-4-20250514",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[15,75]},{canonical:"gpt-4.1",apiId:"gpt-4.1",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[2,8]},{canonical:"gpt-4.1-mini",apiId:"gpt-4.1-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[.4,1.6]},{canonical:"gpt-4.1-nano",apiId:"gpt-4.1-nano",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[.1,.4]},{canonical:"gpt-4o",apiId:"gpt-4o",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:128e3},pricing:[2.5,10]},{canonical:"gpt-4o-mini",apiId:"gpt-4o-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:128e3},pricing:[.15,.6]},{canonical:"o4-mini",apiId:"o4-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[1.1,4.4]},{canonical:"o3",apiId:"o3",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[.4,1.6]},{canonical:"o3-mini",apiId:"o3-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[1.1,4.4]},{canonical:"o1",apiId:"o1",provider:"openai-api",capabilities:{streaming:false,tools:false,multimodal:false,contextWindow:2e5},pricing:[15,60]},{canonical:"llama3.1",apiId:"llama3.1",provider:"ollama",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:128e3}},{canonical:"mistral",apiId:"mistral",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:false,contextWindow:32e3}},{canonical:"mistral-nemo",apiId:"mistral-nemo",provider:"ollama",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:128e3}},{canonical:"llava",apiId:"llava",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:true,contextWindow:32e3}},{canonical:"qwen2:0.5b",apiId:"qwen2:0.5b",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:false,contextWindow:32e3}}],K=Object.fromEntries(G.filter(n=>n.canonical!==n.apiId).map(n=>[n.canonical,n.apiId]));K["claude-sonnet-4"]="claude-sonnet-4-20250514";K["claude-opus-4"]="claude-opus-4-20250514";K["claude-haiku-4"]="claude-haiku-4-5-20251001";function fe(n){return K[n]||n}function mt(n){return G.find(e=>e.canonical===n)}function re(n){return G.find(e=>e.canonical===n)?.provider}function ht(n){return G.filter(e=>e.provider===n)}var yt="https://api.anthropic.com/v1/messages",gt="2023-06-01",be={"claude-sonnet-4-5-20250514":[3,15],"claude-sonnet-4-20250514":[3,15],"claude-haiku-4-5-20250414":[.8,4],"claude-opus-4-20250514":[15,75]};function De(n){if(be[n])return be[n];for(let[e,t]of Object.entries(be))if(n.startsWith(e.split("-20")[0]))return t;return [3,15]}var V=class extends R{type="anthropic-api";supportsStreaming=true;supportsMultimodal=true;platformRequirement="any";apiKey;baseUrl;dangerousDirectBrowserAccess;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.apiKey=e.apiKey,this.baseUrl=e.baseUrl?.replace(/\/$/,"")||yt,this.dangerousDirectBrowserAccess=e.dangerousDirectBrowserAccess??false;}async probe(){try{let e=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify({model:"claude-haiku-4-5-20250414",max_tokens:1,messages:[{role:"user",content:"hi"}]})});return e.status===401?{available:!0,authenticated:!1}:e.ok||e.status===429?{available:!0,authenticated:!0,rateLimitTier:e.headers.get("anthropic-ratelimit-tier")??void 0}:e.status===529?{available:!0,authenticated:!0}:{available:!1,authenticated:!1}}catch{return {available:false,authenticated:false}}}async execute(e,t){let r=Date.now(),s=this.buildRequestBody(e,false),i=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(s),signal:t?.signal});if(!i.ok)throw this.handleErrorResponse(i),new Error("unreachable");let a=await i.json(),o=Date.now()-r;this.recordSuccess();let[l,d]=De(a.model||e.model),c=((a.usage?.input_tokens||0)*l+(a.usage?.output_tokens||0)*d)/1e6,p={},y=["request-id","x-request-id","anthropic-ratelimit-requests-limit","anthropic-ratelimit-requests-remaining","anthropic-ratelimit-requests-reset","anthropic-ratelimit-tokens-limit","anthropic-ratelimit-tokens-remaining","anthropic-ratelimit-tokens-reset","anthropic-ratelimit-tier"];for(let m of y){let _=i.headers.get(m);_&&(p[m]=_);}return {content:a.content||[],model:a.model||e.model,usage:{input_tokens:a.usage?.input_tokens||0,output_tokens:a.usage?.output_tokens||0},stop_reason:a.stop_reason||"end_turn",provider:"anthropic-api",costUsd:Math.round(c*1e6)/1e6,durationMs:o,fallbackChain:["anthropic-api:success"],headers:Object.keys(p).length>0?p:void 0}}async streamRaw(e,t){let r=this.buildRequestBody(e,true),s=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(r),signal:t?.signal});if(s.ok||this.handleErrorResponse(s),!s.body)throw new Error("ANTHROPIC_API_ERROR: Response body is null \u2014 streaming unavailable");let i=s.body.getReader(),a=new TextDecoder;return (async function*(){let o="",l=false;try{for(;;){let{done:d,value:c}=await i.read();if(d)break;o+=a.decode(c,{stream:!0});let p=o.split(`
|
|
11
|
-
`);o=p.pop()||"";let y=null;for(let m of p)if(m.startsWith("event: "))y=m.slice(7).trim();else if(m.startsWith("data: ")){let
|
|
12
|
-
`);l=u.pop()||"";for(let f of u){if(!f.startsWith("data: "))continue;let h=f.slice(6).trim();if(h==="[DONE]"){yield {type:"message_delta",delta:{stop_reason:y||"end_turn"},usage:p||{}},yield {type:"message_stop"},c=!0;return}try{let g=JSON.parse(h);g.usage&&(p={input_tokens:g.usage.prompt_tokens||0,output_tokens:g.usage.completion_tokens||0});let C=g.choices?.[0];if(!C)continue;let
|
|
13
|
-
`);l=m.pop()||"";for(let _ of m){let u=_.trim();if(u)try{let f=JSON.parse(u);if(f.message?.content&&(d||(d=!0,yield {type:"content_block_start",index:0,content_block:{type:"text",text:""}}),yield {type:"content_block_delta",index:0,delta:{type:"text_delta",text:f.message.content}}),f.done){d&&(yield {type:"content_block_stop",index:0}),yield {type:"message_delta",delta:{stop_reason:"end_turn"},usage:{input_tokens:f.prompt_eval_count||0,output_tokens:f.eval_count||0}},yield {type:"message_stop"},c=!0;return}}catch{}}}c||(yield {type:"message_stop"});}finally{try{i.releaseLock();}catch{}}})()}estimateCost(e){let t=Math.round(JSON.stringify(e.messages).length/4),r=e.system?Math.round(e.system.length/4):0;return {estimatedInputTokens:t+r,estimatedCost:0,model:e.model,provider:"ollama"}}getHealth(){return {...this.health}}reportRateLimit(e){this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now();}buildRequestBody(e,t){if(e.tools&&e.tools.length>0)throw new Error(`OLLAMA_TOOLS_NOT_SUPPORTED: The Ollama provider received ${e.tools.length} tool definition(s) in params.tools, but this provider does not implement function calling in v1.3.x. For tool-use workflows route to 'anthropic-api' or 'openai-api' instead. Tracking issue: 1.4.0 ships an Ollama tool-use shim for models that support native function calling (llama3.1, mistral-nemo, etc.).`);let r=Ct(e.messages,e.system),s={model:e.model,messages:r,stream:t};return e.temperature!==void 0&&(s.options={temperature:e.temperature}),s}handleErrorResponse(e){let t=e.status;throw this.health.errorCount++,this.health.lastCheckAt=Date.now(),t===404?(this.health.status="unhealthy",this.health.lastError="Model not found",new Error("OLLAMA_MODEL_NOT_FOUND: Requested model is not available locally (404)")):(this.health.status="unhealthy",this.health.lastError=`Request failed with status ${t}`,new Error(`OLLAMA_ERROR: Request failed with status ${t}`))}recordSuccess(){this.health.status="healthy",this.health.lastCheckAt=Date.now(),this.health.errorCount=0,this.health.lastError=void 0;}};var $=class extends EventEmitter{constructor(){super(),this.setMaxListeners(50);}emitTyped(e){return this.emit(e.kind,e)}onTyped(e,t){return this.on(e,t)}offTyped(e,t){return this.off(e,t)}};var We=200;function ne(n,e){if(e instanceof b)return `${n}:failed:${e.code}:${e.message}`;let t=e?.message;return `${n}:failed:${t?.slice(0,We)||"unknown"}`}var Fe=new Set(["anthropic-api","openai-api","ollama"]),q=class extends Error{code="ALL_PROVIDERS_FAILED";fallbackChain;constructor(e){super(`All providers failed. Chain: ${e.join(" \u2192 ")}`),this.name="RelayError",this.fallbackChain=e;}};function Be(n){return n.messages.some(e=>typeof e.content=="string"?false:Array.isArray(e.content)&&e.content.some(t=>t.type==="image"||t.type==="image_url"))}function _e(n){if(!(n instanceof Error))return "unknown";let e=n.code;if(typeof e=="string"){if(e==="LICENSE_REJECTED")return "auth_failed";if(e==="GATEWAY_FETCH_FAILED"||e==="GATEWAY_VERIFICATION_FAILED"||e==="GATEWAY_LOAD_FAILED")return "gateway_unavailable"}let t=n.message.toLowerCase(),r=n.name.toLowerCase();return t.startsWith("openai_model_translation_required:")||t.startsWith("ollama_tools_not_supported:")||t.startsWith("multimodal_cli_too_old:")?"provider_incompatible":t.includes("rate limit")||t.includes("429")||r.includes("ratelimit")?"rate_limited":t.includes("auth")||t.includes("401")||t.includes("403")||t.includes("unauthorized")||t.includes("forbidden")?"auth_failed":t.includes("timeout")||t.includes("timed out")||t.includes("aborted")||r==="aborterror"||r==="timeouterror"?"timeout":"unknown"}function He(n){return [...n].sort((e,t)=>e.priority-t.priority)}var j=class n{healthMap=new Map;engineEvents=new $;static UNHEALTHY_TTL_MS=300*1e3;getHealthStatus(e){let t=this.healthMap.get(e);return t?t.status==="unhealthy"&&Date.now()-t.updatedAt>=n.UNHEALTHY_TTL_MS?(this.healthMap.delete(e),"healthy"):t.status:"healthy"}markUnhealthy(e,t){this.healthMap.set(e,{status:"unhealthy",lastError:t,updatedAt:Date.now()});}markDegraded(e,t){this.healthMap.set(e,{status:"degraded",lastError:t,updatedAt:Date.now()});}markHealthy(e){this.healthMap.set(e,{status:"healthy",updatedAt:Date.now()});}resetHealth(){this.healthMap.clear();}async execute(e,t,r){let s=He(t),i=[],a=Be(e),o=randomUUID(),l=Date.now();for(let d of s){let c=d.type;if(this.getHealthStatus(c)==="unhealthy"){i.push(`${c}:skipped:unhealthy`);continue}if(a&&!d.supportsMultimodal){i.push(`${c}:skipped:no_multimodal`);continue}let y=re(e.model);if(y&&y!==c&&Fe.has(c)){i.push(`${c}:skipped:wrong_provider`);continue}try{let m=await d.execute(e,{signal:r?.signal});return this.markHealthy(c),i.push(`${c}:success`),this.emitEngineEvent({kind:"stream_complete",callId:o,id:randomUUID(),ts:Date.now(),stitched:!1,segments:[{provider:c,inputTokens:m.usage.input_tokens,outputTokens:m.usage.output_tokens,costUsd:m.costUsd,userCostUsd:m.userCostUsd??0}],provider:c,model:m.model,inputTokens:m.usage.input_tokens,outputTokens:m.usage.output_tokens,costUsd:m.costUsd,userCostUsd:m.userCostUsd,durationMs:m.durationMs??Date.now()-l,stopReason:m.stop_reason,fallbackChain:i,rateLimitHit:!1,securityState:r?.licenseState}),{...m,fallbackChain:i}}catch(m){if(m?.name==="AbortError"||r?.signal?.aborted)throw m;switch(_e(m)){case "rate_limited":d.reportRateLimit({provider:c,utilization:1,resetsAt:Date.now()+6e4}),this.markDegraded(c,"rate_limited"),i.push(`${c}:failed:rate_limited`);break;case "auth_failed":this.markUnhealthy(c,"auth_failed"),i.push(`${c}:failed:auth_failed`);break;case "timeout":i.push(`${c}:failed:timeout`);break;case "provider_incompatible":i.push(`${c}:incompatible`);break;case "gateway_unavailable":this.markUnhealthy(c,"gateway_unavailable"),i.push(ne(c,m));break;default:i.push(ne(c,m));break}}}throw new q(i)}async stream(e,t,r){let s=randomUUID(),i=He(t),a=Be(e),o=Date.now(),l=[],d=null,c=null;for(let h of i){let g=h.type;if(this.getHealthStatus(g)==="unhealthy"){l.push(`${g}:skipped:unhealthy`);continue}if(a&&!h.supportsMultimodal){l.push(`${g}:skipped:no_multimodal`);continue}if(!h.supportsStreaming){l.push(`${g}:skipped:no_streaming`);continue}let C=re(e.model);if(C&&C!==g&&Fe.has(g)){l.push(`${g}:skipped:wrong_provider`);continue}try{c=await h.streamRaw(e,{signal:r?.signal}),d=h,l.push(`${g}:success`);break}catch(E){if(E?.name==="AbortError"||r?.signal?.aborted)throw E;switch(_e(E)){case "rate_limited":h.reportRateLimit({provider:g,utilization:1,resetsAt:Date.now()+6e4}),this.markDegraded(g,"rate_limited"),l.push(`${g}:failed:rate_limited`);break;case "auth_failed":this.markUnhealthy(g,"auth_failed"),l.push(`${g}:failed:auth_failed`);break;case "timeout":l.push(`${g}:failed:timeout`);break;case "provider_incompatible":l.push(`${g}:incompatible`);break;case "gateway_unavailable":this.markUnhealthy(g,"gateway_unavailable"),l.push(ne(g,E));break;default:l.push(ne(g,E));break}}}if(!d||!c)throw new q(l);let y=this,m=r?.maxMidStreamFailovers??2,_=new Set([d]);async function*u(){let h=d,g=c[Symbol.asyncIterator](),C={role:"assistant",content:[]},E={segments:[]},I=Ge(h.type);E.segments.push(I);let Re=0,z=null,X=false,de=m,O="",N="";for(y.emitEngineEvent({kind:"segment_start",callId:s,id:randomUUID(),ts:Date.now(),provider:h.type});;)try{for(;;){let A=await g.next();if(A.done)break;let T=A.value,k=Pt(T,Re,C,X,z);if(k.action==="suppress"){"coalesceMode"in k&&(z=k.coalesceMode),k.clearedJustFailedOver&&(X=!1);continue}k.coalesceModeCleared&&(z=null),k.clearedJustFailedOver&&(X=!1);let v=k.event;if(St(C,v),v.type==="message_start"){if(E.segments.length>1)continue;O=v.message?.model||O,v.message?.usage?.input_tokens!=null&&(I.inputTokens=v.message.usage.input_tokens);}if(v.type==="message_delta"&&(v.delta?.stop_reason&&(N=v.delta.stop_reason),v.usage)){Rt(I,h.type,v.usage),yield {...v,usage:void 0};continue}if(v.type==="message_stop")break;yield v;}if(C.content.length===0&&I.outputTokens===0)throw new Error("stream ended without producing any content");y.markHealthy(h.type),y.emitEngineEvent({kind:"segment_complete",callId:s,id:randomUUID(),ts:Date.now(),provider:h.type,usage:{...I}});let S=Ke(E.segments);yield {type:"message_delta",delta:{stop_reason:N||"end_turn"},usage:S},yield {type:"message_stop"},y.emitEngineEvent({kind:"stream_complete",callId:s,id:randomUUID(),ts:Date.now(),stitched:E.segments.length>1,segments:E.segments.map(A=>({...A})),provider:E.segments.map(A=>A.provider).join("+"),model:O,inputTokens:S.input_tokens,outputTokens:S.output_tokens,costUsd:E.segments.reduce((A,T)=>A+T.costUsd,0),userCostUsd:E.segments.reduce((A,T)=>A+T.userCostUsd,0),durationMs:Date.now()-o,stopReason:N||"end_turn",fallbackChain:l,rateLimitHit:!1,securityState:r?.licenseState});return}catch(S){if(S?.name==="AbortError"||r?.signal?.aborted)throw S;let A=_e(S);try{await g.return?.();}catch{}if(y.markUnhealthy(h.type,String(S)),y.emitEngineEvent({kind:"mid_stream_failover_attempt",callId:s,id:randomUUID(),ts:Date.now(),from:h.type,bytesBuffered:Ve(C),errorClass:A,retriesLeft:de}),l.push(`${h.type}:partial(${Ve(C)}b)`),de<=0)throw y.emitTerminalExhaustion(s,E,O,N,l,o,r?.licenseState),S;let T=kt(C.content),k=[...e.messages,{role:"assistant",content:T}],v=null,Q=null,Se=h;for(;;){let w=y.nextHealthyProvider(i,Se,a,_);if(!w)break;Se=w,_.add(w);let je={...e,messages:k,...w.type!=="anthropic-api"&&w.type!=="claude-cli"?{system:wt(e.system,"[SYSTEM]: CONTINUE_LAST_RESPONSE_WITHOUT_REPEATING")}:{}};try{Q=await w.streamRaw(je,{signal:r?.signal}),v=w;break}catch(Pe){y.markUnhealthy(w.type,String(Pe)),l.push(`${w.type}:failed_connect:${String(Pe).slice(0,We)}`);}}if(!v||!Q)throw y.emitTerminalExhaustion(s,E,O,N,l,o,r?.licenseState),S;l.push(`${v.type}:stitched`),y.emitEngineEvent({kind:"mid_stream_failover_success",callId:s,id:randomUUID(),ts:Date.now(),from:h.type,to:v.type}),Re=It(C)+1,de--,h=v,g=Q[Symbol.asyncIterator](),X=true,z=null,I=Ge(h.type),E.segments.push(I),y.emitEngineEvent({kind:"segment_start",callId:s,id:randomUUID(),ts:Date.now(),provider:h.type});}}let f=u();return r?.licenseState==="SEC_DEGRADED"&&(f=Tt(f)),L(f)}nextHealthyProvider(e,t,r,s){let i=e.indexOf(t);if(i===-1)return null;for(let a=i+1;a<e.length;a++){let o=e[a];if(!s.has(o)&&!(r&&!o.supportsMultimodal)&&o.supportsStreaming)return o}return null}emitEngineEvent(e){this.engineEvents.emitTyped(e);}emitTerminalExhaustion(e,t,r,s,i,a,o){let l=Ke(t.segments);this.emitEngineEvent({kind:"stream_complete",callId:e,id:randomUUID(),ts:Date.now(),stitched:t.segments.length>1,segments:t.segments.map(d=>({...d})),provider:t.segments.map(d=>d.provider).join("+"),model:r,inputTokens:l.input_tokens,outputTokens:l.output_tokens,costUsd:t.segments.reduce((d,c)=>d+c.costUsd,0),userCostUsd:t.segments.reduce((d,c)=>d+c.userCostUsd,0),durationMs:Date.now()-a,stopReason:s||"error",fallbackChain:i,rateLimitHit:false,securityState:o});}};function Ge(n){return {provider:n,inputTokens:0,outputTokens:0,costUsd:0,userCostUsd:0}}function Rt(n,e,t){typeof t.input_tokens=="number"&&(n.inputTokens=t.input_tokens),typeof t.output_tokens=="number"&&(e==="anthropic-api"||e==="claude-cli"?n.outputTokens+=t.output_tokens:n.outputTokens=t.output_tokens);}function Ke(n){return {input_tokens:n[n.length-1]?.inputTokens??0,output_tokens:n.reduce((e,t)=>e+t.outputTokens,0)}}function St(n,e){if(e.type==="content_block_start"){let t=e.content_block;t?.type==="text"?n.content.push({type:"text",text:t.text||"",open:true}):t?.type==="tool_use"&&n.content.push({type:"tool_use",id:t.id,name:t.name,input:{},_partialJson:"",open:true});}else if(e.type==="content_block_delta"){let t=n.content[n.content.length-1];if(!t)return;t.type==="text"&&e.delta?.type==="text_delta"?t.text+=e.delta.text||"":t.type==="tool_use"&&e.delta?.type==="input_json_delta"&&(t._partialJson=(t._partialJson||"")+(e.delta.partial_json||""));}else if(e.type==="content_block_stop"){for(let t=n.content.length-1;t>=0;t--)if(n.content[t].open){if(n.content[t].open=false,n.content[t].type==="tool_use"){let r=n.content[t];try{r.input=JSON.parse(r._partialJson||"{}");}catch{}delete r._partialJson;}break}}}function Pt(n,e,t,r,s){if(s&&n.type==="content_block_delta"&&n.index===s.fromIndex)return {action:"yield",event:{...n,index:s.toIndex}};if(s&&n.type==="content_block_stop"&&n.index===s.fromIndex)return {action:"suppress",coalesceMode:null,coalesceModeCleared:true};if(r&&n.type==="content_block_start"&&n.content_block?.type==="text"){let a=t.content[t.content.length-1];if(a?.type==="text"&&a.open===true)return {action:"suppress",coalesceMode:{fromIndex:n.index+e,toIndex:t.content.length-1},clearedJustFailedOver:true}}let i=r&&n.type==="content_block_start";return typeof n.index=="number"&&e>0?{action:"yield",event:{...n,index:n.index+e},clearedJustFailedOver:i}:{action:"yield",event:n,clearedJustFailedOver:i}}function kt(n){let e=[];for(let t of n)t.type==="text"?t.text&&e.push({type:"text",text:t.text}):t.type==="tool_use"&&!t.open&&t.input&&Object.keys(t.input).length>0&&e.push({type:"tool_use",id:t.id,name:t.name,input:t.input});return e}function wt(n,e){return n?`${n}
|
|
10
|
+
-----END PUBLIC KEY-----`,pe={version:"1.0.0",cliVersion:"2.1.87",eventTypes:{system:"system",streamEvent:"stream_event",assistant:"assistant",result:"result"},fieldMappings:{sessionId:"session_id",model:"model",cliVersion:"claude_code_version",content:"content",usage:"usage",inputTokens:"input_tokens",outputTokens:"output_tokens",stopReason:"stop_reason",isError:"is_error",totalCostUsd:"total_cost_usd",durationMs:"duration_ms",error:"error",event:"event"},syntheticModel:"<synthetic>"},M=class{verificationFailures=0;cachedManifest=null;currentSecurityState="SEC_NOMINAL";async verifyManifest(e){try{let{compactVerify:t,importSPKI:r}=await import('jose'),s=await r(me,"RS256"),{payload:i}=await t(e,s),a=JSON.parse(new TextDecoder().decode(i));return this.verificationFailures=0,this.cachedManifest=a,this.currentSecurityState="SEC_NOMINAL",{manifest:a,securityState:"SEC_NOMINAL"}}catch{if(this.verificationFailures++,this.verificationFailures>=3){this.currentSecurityState="SEC_DEGRADED";let t=2e3+Math.random()*3e3;return await new Promise(r=>setTimeout(r,t)),console.warn("[inference-relay] Logic synchronization mismatch. Routing via local recovery."),{manifest:pe,securityState:"SEC_DEGRADED"}}return this.cachedManifest?(this.currentSecurityState="SEC_CACHE",{manifest:this.cachedManifest,securityState:"SEC_CACHE"}):(this.currentSecurityState="SEC_RECOVERY",{manifest:pe,securityState:"SEC_RECOVERY"})}}getSecurityState(){return this.currentSecurityState}clearCache(){this.cachedManifest=null,this.verificationFailures=0,this.currentSecurityState="SEC_NOMINAL";}},Ze=new M;function ke(){Ze.clearCache();}function et(){return pe}function tt(n,e){return e.split(".").reduce((t,r)=>{if(!(r==="__proto__"||r==="constructor"||r==="prototype"))return t&&t[r]},n)}var at=process.env.IR_GATEWAY_URL||"https://api.inference-relay.com";function ge(n){return join(homedir(),".inference-relay","gateway",n)}function Te(n){return join(ge(n),"bundle.js")}function xe(n){return join(ge(n),"meta.json")}var b=class extends Error{constructor(t,r,s){super(t);this.code=r;this.cause=s;this.name="GatewayLoaderError";}code;cause},F=new Map,B=class{static async load(e,t,r=at){let s=`${t}:${e.slice(-8)}`;if(F.has(s))return F.get(s);try{let p=await this.tryLoadFromDisk(t);if(p)return F.set(s,p),p}catch{}let{bytes:i,sha256:a,signature:o,libraryVersion:l}=await this.fetchFromAuthority(e,t,r);if(createHash("sha256").update(i).digest("hex")!==a)throw new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: HASH_MISMATCH)","GATEWAY_VERIFICATION_FAILED");await this.verifySignature(o,l,a);try{let p=ge(t);mkdirSync(p,{recursive:!0}),writeFileSync(Te(t),i),writeFileSync(xe(t),JSON.stringify({sha256:a,signature:o,libraryVersion:l}));}catch{}let c=this.loadIntoVm(i.toString("utf-8"));return F.set(s,c),c}static async loadFromDisk(e){if(!existsSync(e))throw new b(`[ir-system]: GATEWAY_LOAD_FAILED (REASON: BUNDLE_NOT_FOUND at ${e})`,"GATEWAY_LOAD_FAILED");let t=readFileSync(e,"utf-8");return this.loadIntoVm(t)}static async tryLoadFromDisk(e){let t=Te(e),r=xe(e);if(!existsSync(t)||!existsSync(r))return null;let s=readFileSync(t),i=JSON.parse(readFileSync(r,"utf-8"));return createHash("sha256").update(s).digest("hex")!==i.sha256?null:(await this.verifySignature(i.signature,i.libraryVersion,i.sha256),this.loadIntoVm(s.toString("utf-8")))}static async fetchFromAuthority(e,t,r){let s,i=[0,1e3,4e3],a;for(let c of i){c>0&&await new Promise(p=>setTimeout(p,c));try{let p=await fetch(`${r}/v1/gateway/fetch`,{method:"POST",headers:{"Content-Type":"application/json",Authorization:`Bearer ${e}`},body:JSON.stringify({libraryVersion:t})});if(p.status===401||p.status===403)throw new b("[ir-system]: LICENSE_REJECTED (REASON: GATEWAY_FETCH_UNAUTHORIZED)","LICENSE_REJECTED");if(p.status>=500&&p.status<600){a=new b(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: HTTP_${p.status})`,"GATEWAY_FETCH_FAILED");continue}if(!p.ok)throw new b(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: HTTP_${p.status})`,"GATEWAY_FETCH_FAILED");s=await p.json(),a=null;break}catch(p){if(p instanceof b&&p.code==="LICENSE_REJECTED")throw p;a=p;}}if(a||!s)throw a instanceof b?a:new b("[ir-system]: GATEWAY_FETCH_FAILED (REASON: NETWORK_ERROR)","GATEWAY_FETCH_FAILED",a);let o;try{o=await fetch(s.bundleUrl);}catch(c){throw new b("[ir-system]: GATEWAY_FETCH_FAILED (REASON: BUNDLE_DOWNLOAD_FAILED)","GATEWAY_FETCH_FAILED",c)}if(!o.ok)throw new b(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: BUNDLE_HTTP_${o.status})`,"GATEWAY_FETCH_FAILED");let l=await o.arrayBuffer();return {bytes:Buffer.from(l),sha256:s.sha256,signature:s.signature,libraryVersion:s.libraryVersion,bundleVersion:s.bundleVersion}}static async verifySignature(e,t,r){try{let{compactVerify:s,importSPKI:i}=await import('jose'),a=await i(me,"RS256"),{payload:o}=await s(e,a),l=JSON.parse(new TextDecoder().decode(o));if(l.purpose!=="gateway-bundle")throw new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: PURPOSE_MISMATCH)","GATEWAY_VERIFICATION_FAILED");if(l.libraryVersion!==t)throw new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: LIBRARY_VERSION_MISMATCH)","GATEWAY_VERIFICATION_FAILED");if(l.sha256!==r)throw new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: SHA256_MISMATCH)","GATEWAY_VERIFICATION_FAILED")}catch(s){throw s instanceof b?s:new b("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: SIGNATURE_INVALID)","GATEWAY_VERIFICATION_FAILED",s)}}static loadIntoVm(e){try{let t=join(homedir(),".inference-relay","gateway","runtime.cjs"),r=dirname(t),s=createRequire(t),i={exports:{}};compileFunction(e,["exports","require","module","__filename","__dirname"],{filename:"<gateway-bundle>"})(i.exports,s,i,t,r);let o=i.exports;if(!o.ClaudeCLIProvider)throw new b("[ir-system]: GATEWAY_LOAD_FAILED (REASON: MISSING_CLAUDE_CLI_EXPORT)","GATEWAY_LOAD_FAILED");return o}catch(t){throw t instanceof b?t:new b("[ir-system]: GATEWAY_LOAD_FAILED (REASON: VM_EXECUTION_ERROR)","GATEWAY_LOAD_FAILED",t)}}static clearCache(){F.clear();}};var pt="1.6.0",te=class extends R{type="claude-cli";supportsStreaming=true;supportsMultimodal=true;platformRequirement="desktop";licenseKey;loadedProvider=null;loadingPromise=null;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.licenseKey=e;}async ensureLoaded(){return this.loadedProvider?this.loadedProvider:this.loadingPromise?this.loadingPromise:(this.loadingPromise=(async()=>{try{let t=(await B.load(this.licenseKey,pt)).ClaudeCLIProvider,r=new t;return this.loadedProvider=r,r}catch(e){throw this.loadingPromise=null,this.health.status="unhealthy",this.health.errorCount++,e instanceof b?this.health.lastError=e.message:e instanceof Error&&(this.health.lastError=e.message),e}})(),this.loadingPromise)}async probe(){let{detectCLI:e}=await Promise.resolve().then(()=>(ee(),Me));return e()}async execute(e,t){return (await this.ensureLoaded()).execute(e,t)}async streamRaw(e,t){return (await this.ensureLoaded()).streamRaw(e,t)}estimateCost(e){if(this.loadedProvider)return this.loadedProvider.estimateCost(e);let r=JSON.stringify(e.messages).length/4+5e3,s=(r*3+1e3*15)/1e6;return {estimatedInputTokens:Math.round(r),estimatedCost:Math.round(s*1e4)/1e4,model:e.model,provider:"claude-cli"}}getHealth(){return this.loadedProvider?this.loadedProvider.getHealth():{...this.health}}reportRateLimit(e){this.loadedProvider?this.loadedProvider.reportRateLimit(e):(this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now());}};var G=[{canonical:"claude-haiku-4-5",apiId:"claude-haiku-4-5-20251001",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[.8,4]},{canonical:"claude-sonnet-4-6",apiId:"claude-sonnet-4-20250514",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[3,15]},{canonical:"claude-opus-4-6",apiId:"claude-opus-4-20250514",provider:"anthropic-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:2e5},pricing:[15,75]},{canonical:"gpt-4.1",apiId:"gpt-4.1",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[2,8]},{canonical:"gpt-4.1-mini",apiId:"gpt-4.1-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[.4,1.6]},{canonical:"gpt-4.1-nano",apiId:"gpt-4.1-nano",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:1e6},pricing:[.1,.4]},{canonical:"gpt-4o",apiId:"gpt-4o",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:128e3},pricing:[2.5,10]},{canonical:"gpt-4o-mini",apiId:"gpt-4o-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:true,contextWindow:128e3},pricing:[.15,.6]},{canonical:"o4-mini",apiId:"o4-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[1.1,4.4]},{canonical:"o3",apiId:"o3",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[.4,1.6]},{canonical:"o3-mini",apiId:"o3-mini",provider:"openai-api",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:2e5},pricing:[1.1,4.4]},{canonical:"o1",apiId:"o1",provider:"openai-api",capabilities:{streaming:false,tools:false,multimodal:false,contextWindow:2e5},pricing:[15,60]},{canonical:"llama3.1",apiId:"llama3.1",provider:"ollama",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:128e3}},{canonical:"mistral",apiId:"mistral",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:false,contextWindow:32e3}},{canonical:"mistral-nemo",apiId:"mistral-nemo",provider:"ollama",capabilities:{streaming:true,tools:true,multimodal:false,contextWindow:128e3}},{canonical:"llava",apiId:"llava",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:true,contextWindow:32e3}},{canonical:"qwen2:0.5b",apiId:"qwen2:0.5b",provider:"ollama",capabilities:{streaming:true,tools:false,multimodal:false,contextWindow:32e3}}],K=Object.fromEntries(G.filter(n=>n.canonical!==n.apiId).map(n=>[n.canonical,n.apiId]));K["claude-sonnet-4"]="claude-sonnet-4-20250514";K["claude-opus-4"]="claude-opus-4-20250514";K["claude-haiku-4"]="claude-haiku-4-5-20251001";function fe(n){return K[n]||n}function mt(n){return G.find(e=>e.canonical===n)}function re(n){return G.find(e=>e.canonical===n)?.provider}function ht(n){return G.filter(e=>e.provider===n)}var yt="https://api.anthropic.com/v1/messages",gt="2023-06-01",be={"claude-sonnet-4-5-20250514":[3,15],"claude-sonnet-4-20250514":[3,15],"claude-haiku-4-5-20250414":[.8,4],"claude-opus-4-20250514":[15,75]};function De(n){if(be[n])return be[n];for(let[e,t]of Object.entries(be))if(n.startsWith(e.split("-20")[0]))return t;return [3,15]}var V=class extends R{type="anthropic-api";supportsStreaming=true;supportsMultimodal=true;platformRequirement="any";apiKey;baseUrl;dangerousDirectBrowserAccess;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.apiKey=e.apiKey,this.baseUrl=e.baseUrl?.replace(/\/$/,"")||yt,this.dangerousDirectBrowserAccess=e.dangerousDirectBrowserAccess??false;}async probe(){try{let e=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify({model:"claude-haiku-4-5-20250414",max_tokens:1,messages:[{role:"user",content:"hi"}]})});return e.status===401?{available:!0,authenticated:!1}:e.ok||e.status===429?{available:!0,authenticated:!0,rateLimitTier:e.headers.get("anthropic-ratelimit-tier")??void 0}:e.status===529?{available:!0,authenticated:!0}:{available:!1,authenticated:!1}}catch{return {available:false,authenticated:false}}}async execute(e,t){let r=Date.now(),s=this.buildRequestBody(e,false),i=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(s),signal:t?.signal});if(!i.ok)throw this.handleErrorResponse(i),new Error("unreachable");let a=await i.json(),o=Date.now()-r;this.recordSuccess();let[l,d]=De(a.model||e.model),c=((a.usage?.input_tokens||0)*l+(a.usage?.output_tokens||0)*d)/1e6,p={},y=["request-id","x-request-id","anthropic-ratelimit-requests-limit","anthropic-ratelimit-requests-remaining","anthropic-ratelimit-requests-reset","anthropic-ratelimit-tokens-limit","anthropic-ratelimit-tokens-remaining","anthropic-ratelimit-tokens-reset","anthropic-ratelimit-tier"];for(let m of y){let E=i.headers.get(m);E&&(p[m]=E);}return {content:a.content||[],model:a.model||e.model,usage:{input_tokens:a.usage?.input_tokens||0,output_tokens:a.usage?.output_tokens||0},stop_reason:a.stop_reason||"end_turn",provider:"anthropic-api",costUsd:Math.round(c*1e6)/1e6,durationMs:o,fallbackChain:["anthropic-api:success"],headers:Object.keys(p).length>0?p:void 0}}async streamRaw(e,t){let r=this.buildRequestBody(e,true),s=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(r),signal:t?.signal});if(s.ok||this.handleErrorResponse(s),!s.body)throw new Error("ANTHROPIC_API_ERROR: Response body is null \u2014 streaming unavailable");let i=s.body.getReader(),a=new TextDecoder;return (async function*(){let o="",l=false;try{for(;;){let{done:d,value:c}=await i.read();if(d)break;o+=a.decode(c,{stream:!0});let p=o.split(`
|
|
11
|
+
`);o=p.pop()||"";let y=null;for(let m of p)if(m.startsWith("event: "))y=m.slice(7).trim();else if(m.startsWith("data: ")){let E=m.slice(6);if(y&&E)try{let u=JSON.parse(E);u.type==="message_stop"&&(l=!0),yield u;}catch{console.warn("[inference-relay] Skipped malformed SSE data:",E.substring(0,100));}y=null;}else m.trim()===""&&(y=null);}l||(yield {type:"message_stop"});}finally{try{i.releaseLock();}catch{}}})()}estimateCost(e){let t=Math.round(JSON.stringify(e.messages).length/4),r=e.system?Math.round(e.system.length/4):0,s=t+r,[i,a]=De(e.model),o=(s*i+1e3*a)/1e6;return {estimatedInputTokens:s,estimatedCost:Math.round(o*1e4)/1e4,model:e.model,provider:"anthropic-api"}}getHealth(){return {...this.health}}reportRateLimit(e){this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now();}buildHeaders(){let e={"content-type":"application/json","x-api-key":this.apiKey,"anthropic-version":gt};return this.dangerousDirectBrowserAccess&&(e["anthropic-dangerous-direct-browser-access"]="true"),e}buildRequestBody(e,t){let r={model:fe(e.model),max_tokens:e.max_tokens,messages:e.messages};return t&&(r.stream=true),e.system&&(r.system=e.system),e.temperature!==void 0&&(r.temperature=e.temperature),e.tools&&e.tools.length>0&&(r.tools=e.tools),e.metadata&&(r.metadata=e.metadata),r}handleErrorResponse(e){let t=e.status;if(t===401)throw this.health.status="unhealthy",this.health.lastError="Authentication failed",this.health.errorCount++,this.health.lastCheckAt=Date.now(),new Error("ANTHROPIC_AUTH_ERROR: Invalid API key (401)");if(t===429){let r=e.headers.get("retry-after"),s=r?parseInt(r,10)*1e3:6e4;this.health.status="degraded",this.health.lastError=`Rate limited \u2014 retry after ${s}ms`,this.health.lastCheckAt=Date.now();let i=new Error(`ANTHROPIC_RATE_LIMIT: Too many requests (429). Retry after ${s}ms`);throw i.retryAfterMs=s,i}throw t===529?(this.health.status="degraded",this.health.lastError="API overloaded",this.health.lastCheckAt=Date.now(),new Error("ANTHROPIC_OVERLOADED: API is temporarily overloaded (529)")):(this.health.errorCount++,this.health.lastCheckAt=Date.now(),new Error(`ANTHROPIC_API_ERROR: Request failed with status ${t}`))}recordSuccess(){this.health.status="healthy",this.health.lastCheckAt=Date.now(),this.health.errorCount=0,this.health.lastError=void 0;}};var ft="https://api.openai.com/v1/chat/completions",Oe=["gpt-","o1","o3","o4","chatgpt-","text-","davinci","babbage","ft:"],_e={"gpt-4.1":[2,8],"gpt-4.1-mini":[.4,1.6],"gpt-4.1-nano":[.1,.4],"gpt-4o":[2.5,10],"gpt-4o-mini":[.15,.6],"o4-mini":[1.1,4.4],o3:[.4,1.6],"o3-mini":[1.1,4.4],o1:[15,60],"o1-mini":[3,12],"gpt-4-turbo":[10,30],"gpt-4":[30,60],"gpt-3.5-turbo":[.5,1.5]};function Ne(n){if(_e[n])return _e[n];for(let[e,t]of Object.entries(_e))if(n.startsWith(e))return t;return [2.5,10]}function bt(n){return typeof n=="string"?n:n.filter(e=>e.type==="text"&&e.text).map(e=>e.text).join("")}function _t(n,e){let t=[];e&&t.push({role:"system",content:e});for(let r of n){let s=bt(r.content);if(r.role==="assistant"&&Array.isArray(r.content)){let i=r.content.filter(a=>a.type==="tool_use");if(i.length>0){t.push({role:"assistant",content:s||"",tool_calls:i.map(a=>({id:a.id||`call_${Math.random().toString(36).slice(2)}`,type:"function",function:{name:a.name,arguments:typeof a.input=="string"?a.input:JSON.stringify(a.input)}}))});continue}}t.push({role:r.role,content:s});}return t}function Ue(n){switch(n){case "stop":return "end_turn";case "length":return "max_tokens";case "tool_calls":case "function_call":return "tool_use";default:return "end_turn"}}var W=class extends R{type="openai-api";supportsStreaming=true;supportsMultimodal=true;platformRequirement="any";apiKey;baseUrl;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.apiKey=e.apiKey,this.baseUrl=e.baseUrl?.replace(/\/$/,"")||ft;}async probe(){try{let e=this.baseUrl.replace(/\/chat\/completions$/,"/models"),t=await fetch(e,{method:"GET",headers:this.buildHeaders()});return t.status===401?{available:!0,authenticated:!1}:t.ok||t.status===429?{available:!0,authenticated:!0}:{available:!1,authenticated:!1}}catch{return {available:false,authenticated:false}}}async execute(e,t){let r=Date.now(),s=this.buildRequestBody(e,false),i=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(s),signal:t?.signal});if(!i.ok)throw this.handleErrorResponse(i),new Error("unreachable");let a=await i.json(),o=Date.now()-r;this.recordSuccess();let l=a.choices?.[0],d=[];if(l?.message?.content&&d.push({type:"text",text:l.message.content}),l?.message?.tool_calls)for(let h of l.message.tool_calls)d.push({type:"tool_use",id:h.id,name:h.function?.name,input:h.function?.arguments?JSON.parse(h.function.arguments):{}});let c=a.usage?.prompt_tokens||0,p=a.usage?.completion_tokens||0,[y,m]=Ne(a.model||e.model),E=(c*y+p*m)/1e6,u={},f=["x-request-id","x-ratelimit-limit-requests","x-ratelimit-remaining-requests","x-ratelimit-limit-tokens","x-ratelimit-remaining-tokens","x-ratelimit-reset-requests","x-ratelimit-reset-tokens"];for(let h of f){let g=i.headers.get(h);g&&(u[h]=g);}return {content:d,model:a.model||e.model,usage:{input_tokens:c,output_tokens:p},stop_reason:Ue(l?.finish_reason),provider:"openai-api",costUsd:Math.round(E*1e6)/1e6,durationMs:o,fallbackChain:["openai-api:success"],headers:Object.keys(u).length>0?u:void 0}}async streamRaw(e,t){let r=this.buildRequestBody(e,true),s=await fetch(this.baseUrl,{method:"POST",headers:this.buildHeaders(),body:JSON.stringify(r),signal:t?.signal});if(s.ok||this.handleErrorResponse(s),!s.body)throw new Error("OPENAI_API_ERROR: Response body is null \u2014 streaming unavailable");let i=s.body.getReader(),a=new TextDecoder,o=e.model;return (async function*(){let l="",d=false,c=false,p,y;yield {type:"message_start",message:{model:o,usage:{input_tokens:0,output_tokens:0}}};try{for(;;){let{done:m,value:E}=await i.read();if(m)break;l+=a.decode(E,{stream:!0});let u=l.split(`
|
|
12
|
+
`);l=u.pop()||"";for(let f of u){if(!f.startsWith("data: "))continue;let h=f.slice(6).trim();if(h==="[DONE]"){yield {type:"message_delta",delta:{stop_reason:y||"end_turn"},usage:p||{}},yield {type:"message_stop"},c=!0;return}try{let g=JSON.parse(h);g.usage&&(p={input_tokens:g.usage.prompt_tokens||0,output_tokens:g.usage.completion_tokens||0});let C=g.choices?.[0];if(!C)continue;let _=C.delta;_?.content&&!d&&(d=!0,yield {type:"content_block_start",index:0,content_block:{type:"text",text:""}}),_?.content&&(yield {type:"content_block_delta",index:0,delta:{type:"text_delta",text:_.content}}),C.finish_reason&&(d&&(yield {type:"content_block_stop",index:0}),y=Ue(C.finish_reason));}catch{}}}c||(yield {type:"message_delta",delta:{stop_reason:y||"end_turn"},usage:p||{}},yield {type:"message_stop"});}finally{try{i.releaseLock();}catch{}}})()}estimateCost(e){let t=Math.round(JSON.stringify(e.messages).length/4),r=e.system?Math.round(e.system.length/4):0,s=t+r,[i,a]=Ne(e.model),o=(s*i+1e3*a)/1e6;return {estimatedInputTokens:s,estimatedCost:Math.round(o*1e4)/1e4,model:e.model,provider:"openai-api"}}getHealth(){return {...this.health}}reportRateLimit(e){this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now();}buildHeaders(){return {"Content-Type":"application/json",Authorization:`Bearer ${this.apiKey}`}}buildRequestBody(e,t){let r=e.model.toLowerCase();if(!Oe.some(o=>r.startsWith(o)))throw new Error(`OPENAI_MODEL_TRANSLATION_REQUIRED: The OpenAI provider received the model name '${e.model}', which does not match any known OpenAI model prefix (${Oe.join(", ")}). The relay does NOT auto-translate model names across providers in v1.3.x. Either (a) pass an OpenAI model name like 'gpt-4o-mini' when routing to this provider, or (b) route non-OpenAI models through a provider that understands them (e.g. 'anthropic-api' for Claude models).`);let i=_t(e.messages,e.system),a={model:e.model,max_tokens:e.max_tokens,messages:i};return t&&(a.stream=true,a.stream_options={include_usage:true}),e.temperature!==void 0&&(a.temperature=e.temperature),e.tools&&e.tools.length>0&&(a.tools=e.tools),a}handleErrorResponse(e){let t=e.status;if(t===401)throw this.health.status="unhealthy",this.health.lastError="Authentication failed",this.health.errorCount++,this.health.lastCheckAt=Date.now(),new Error("OPENAI_AUTH_ERROR: Invalid API key (401)");if(t===429){let r=e.headers.get("retry-after"),s=r?parseInt(r,10)*1e3:6e4;this.health.status="degraded",this.health.lastError=`Rate limited \u2014 retry after ${s}ms`,this.health.lastCheckAt=Date.now();let i=new Error(`OPENAI_RATE_LIMIT: Too many requests (429). Retry after ${s}ms`);throw i.retryAfterMs=s,i}throw this.health.errorCount++,this.health.lastCheckAt=Date.now(),new Error(`OPENAI_API_ERROR: Request failed with status ${t}`)}recordSuccess(){this.health.status="healthy",this.health.lastCheckAt=Date.now(),this.health.errorCount=0,this.health.lastError=void 0;}};var Et="http://localhost:11434";function vt(n){return typeof n=="string"?n:n.filter(e=>e.type==="text"&&e.text).map(e=>e.text).join("")}function Ct(n,e){let t=[];e&&t.push({role:"system",content:e});for(let r of n)t.push({role:r.role,content:vt(r.content)});return t}var Y=class extends R{type="ollama";supportsStreaming=true;supportsMultimodal=false;platformRequirement="desktop";baseUrl;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.baseUrl=e?.baseUrl?.replace(/\/$/,"")||Et;}async probe(){try{return (await fetch(`${this.baseUrl}/api/tags`,{method:"GET"})).ok?{available:!0,authenticated:!0}:{available:!1,authenticated:!1}}catch{return {available:false,authenticated:false}}}async execute(e,t){let r=Date.now(),s=this.buildRequestBody(e,false),i=await fetch(`${this.baseUrl}/api/chat`,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify(s),signal:t?.signal});if(!i.ok)throw this.handleErrorResponse(i),new Error("unreachable");let a=await i.json(),o=Date.now()-r;this.recordSuccess();let l=[];a.message?.content&&l.push({type:"text",text:a.message.content});let d=a.prompt_eval_count||0,c=a.eval_count||0;return {content:l,model:a.model||e.model,usage:{input_tokens:d,output_tokens:c},stop_reason:(a.done,"end_turn"),provider:"ollama",costUsd:0,durationMs:o,fallbackChain:["ollama:success"]}}async streamRaw(e,t){let r=this.buildRequestBody(e,true),s=await fetch(`${this.baseUrl}/api/chat`,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify(r),signal:t?.signal});if(s.ok||this.handleErrorResponse(s),!s.body)throw new Error("OLLAMA_ERROR: Response body is null \u2014 streaming unavailable");let i=s.body.getReader(),a=new TextDecoder,o=e.model;return (async function*(){let l="",d=false,c=false;yield {type:"message_start",message:{model:o,usage:{input_tokens:0,output_tokens:0}}};try{for(;;){let{done:p,value:y}=await i.read();if(p)break;l+=a.decode(y,{stream:!0});let m=l.split(`
|
|
13
|
+
`);l=m.pop()||"";for(let E of m){let u=E.trim();if(u)try{let f=JSON.parse(u);if(f.message?.content&&(d||(d=!0,yield {type:"content_block_start",index:0,content_block:{type:"text",text:""}}),yield {type:"content_block_delta",index:0,delta:{type:"text_delta",text:f.message.content}}),f.done){d&&(yield {type:"content_block_stop",index:0}),yield {type:"message_delta",delta:{stop_reason:"end_turn"},usage:{input_tokens:f.prompt_eval_count||0,output_tokens:f.eval_count||0}},yield {type:"message_stop"},c=!0;return}}catch{}}}c||(yield {type:"message_stop"});}finally{try{i.releaseLock();}catch{}}})()}estimateCost(e){let t=Math.round(JSON.stringify(e.messages).length/4),r=e.system?Math.round(e.system.length/4):0;return {estimatedInputTokens:t+r,estimatedCost:0,model:e.model,provider:"ollama"}}getHealth(){return {...this.health}}reportRateLimit(e){this.health.status="degraded",this.health.lastError=`Rate limited: ${e.utilization*100}% utilization`,this.health.lastCheckAt=Date.now();}buildRequestBody(e,t){if(e.tools&&e.tools.length>0)throw new Error(`OLLAMA_TOOLS_NOT_SUPPORTED: The Ollama provider received ${e.tools.length} tool definition(s) in params.tools, but this provider does not implement function calling in v1.3.x. For tool-use workflows route to 'anthropic-api' or 'openai-api' instead. Tracking issue: 1.4.0 ships an Ollama tool-use shim for models that support native function calling (llama3.1, mistral-nemo, etc.).`);let r=Ct(e.messages,e.system),s={model:e.model,messages:r,stream:t};return e.temperature!==void 0&&(s.options={temperature:e.temperature}),s}handleErrorResponse(e){let t=e.status;throw this.health.errorCount++,this.health.lastCheckAt=Date.now(),t===404?(this.health.status="unhealthy",this.health.lastError="Model not found",new Error("OLLAMA_MODEL_NOT_FOUND: Requested model is not available locally (404)")):(this.health.status="unhealthy",this.health.lastError=`Request failed with status ${t}`,new Error(`OLLAMA_ERROR: Request failed with status ${t}`))}recordSuccess(){this.health.status="healthy",this.health.lastCheckAt=Date.now(),this.health.errorCount=0,this.health.lastError=void 0;}};var $=class extends EventEmitter{constructor(){super(),this.setMaxListeners(50);}emitTyped(e){return this.emit(e.kind,e)}onTyped(e,t){return this.on(e,t)}offTyped(e,t){return this.off(e,t)}};var We=200;function ne(n,e){if(e instanceof b)return `${n}:failed:${e.code}:${e.message}`;let t=e?.message;return `${n}:failed:${t?.slice(0,We)||"unknown"}`}var Fe=new Set(["anthropic-api","openai-api","ollama"]),q=class extends Error{code="ALL_PROVIDERS_FAILED";fallbackChain;constructor(e){super(`All providers failed. Chain: ${e.join(" \u2192 ")}`),this.name="RelayError",this.fallbackChain=e;}};function Be(n){return n.messages.some(e=>typeof e.content=="string"?false:Array.isArray(e.content)&&e.content.some(t=>t.type==="image"||t.type==="image_url"))}function Ee(n){if(!(n instanceof Error))return "unknown";let e=n.code;if(typeof e=="string"){if(e==="LICENSE_REJECTED")return "auth_failed";if(e==="GATEWAY_FETCH_FAILED"||e==="GATEWAY_VERIFICATION_FAILED"||e==="GATEWAY_LOAD_FAILED")return "gateway_unavailable"}let t=n.message.toLowerCase(),r=n.name.toLowerCase();return t.startsWith("openai_model_translation_required:")||t.startsWith("ollama_tools_not_supported:")||t.startsWith("multimodal_cli_too_old:")||t.startsWith("stream_json_cli_too_old:")?"provider_incompatible":t.includes("rate limit")||t.includes("429")||r.includes("ratelimit")?"rate_limited":t.includes("auth")||t.includes("401")||t.includes("403")||t.includes("unauthorized")||t.includes("forbidden")?"auth_failed":t.includes("timeout")||t.includes("timed out")||t.includes("aborted")||r==="aborterror"||r==="timeouterror"?"timeout":"unknown"}function He(n){return [...n].sort((e,t)=>e.priority-t.priority)}var j=class n{healthMap=new Map;engineEvents=new $;static UNHEALTHY_TTL_MS=300*1e3;getHealthStatus(e){let t=this.healthMap.get(e);return t?t.status==="unhealthy"&&Date.now()-t.updatedAt>=n.UNHEALTHY_TTL_MS?(this.healthMap.delete(e),"healthy"):t.status:"healthy"}markUnhealthy(e,t){this.healthMap.set(e,{status:"unhealthy",lastError:t,updatedAt:Date.now()});}markDegraded(e,t){this.healthMap.set(e,{status:"degraded",lastError:t,updatedAt:Date.now()});}markHealthy(e){this.healthMap.set(e,{status:"healthy",updatedAt:Date.now()});}resetHealth(){this.healthMap.clear();}async execute(e,t,r){let s=He(t),i=[],a=Be(e),o=randomUUID(),l=Date.now();for(let d of s){let c=d.type;if(this.getHealthStatus(c)==="unhealthy"){i.push(`${c}:skipped:unhealthy`);continue}if(a&&!d.supportsMultimodal){i.push(`${c}:skipped:no_multimodal`);continue}let y=re(e.model);if(y&&y!==c&&Fe.has(c)){i.push(`${c}:skipped:wrong_provider`);continue}try{let m=await d.execute(e,{signal:r?.signal});return this.markHealthy(c),i.push(`${c}:success`),this.emitEngineEvent({kind:"stream_complete",callId:o,id:randomUUID(),ts:Date.now(),stitched:!1,segments:[{provider:c,inputTokens:m.usage.input_tokens,outputTokens:m.usage.output_tokens,costUsd:m.costUsd,userCostUsd:m.userCostUsd??0}],provider:c,model:m.model,inputTokens:m.usage.input_tokens,outputTokens:m.usage.output_tokens,costUsd:m.costUsd,userCostUsd:m.userCostUsd,durationMs:m.durationMs??Date.now()-l,stopReason:m.stop_reason,fallbackChain:i,rateLimitHit:!1,securityState:r?.licenseState}),{...m,fallbackChain:i}}catch(m){if(m?.name==="AbortError"||r?.signal?.aborted)throw m;switch(Ee(m)){case "rate_limited":d.reportRateLimit({provider:c,utilization:1,resetsAt:Date.now()+6e4}),this.markDegraded(c,"rate_limited"),i.push(`${c}:failed:rate_limited`);break;case "auth_failed":this.markUnhealthy(c,"auth_failed"),i.push(`${c}:failed:auth_failed`);break;case "timeout":i.push(`${c}:failed:timeout`);break;case "provider_incompatible":i.push(`${c}:incompatible`);break;case "gateway_unavailable":this.markUnhealthy(c,"gateway_unavailable"),i.push(ne(c,m));break;default:i.push(ne(c,m));break}}}throw new q(i)}async stream(e,t,r){let s=randomUUID(),i=He(t),a=Be(e),o=Date.now(),l=[],d=null,c=null;for(let h of i){let g=h.type;if(this.getHealthStatus(g)==="unhealthy"){l.push(`${g}:skipped:unhealthy`);continue}if(a&&!h.supportsMultimodal){l.push(`${g}:skipped:no_multimodal`);continue}if(!h.supportsStreaming){l.push(`${g}:skipped:no_streaming`);continue}let C=re(e.model);if(C&&C!==g&&Fe.has(g)){l.push(`${g}:skipped:wrong_provider`);continue}try{c=await h.streamRaw(e,{signal:r?.signal}),d=h,l.push(`${g}:success`);break}catch(_){if(_?.name==="AbortError"||r?.signal?.aborted)throw _;switch(Ee(_)){case "rate_limited":h.reportRateLimit({provider:g,utilization:1,resetsAt:Date.now()+6e4}),this.markDegraded(g,"rate_limited"),l.push(`${g}:failed:rate_limited`);break;case "auth_failed":this.markUnhealthy(g,"auth_failed"),l.push(`${g}:failed:auth_failed`);break;case "timeout":l.push(`${g}:failed:timeout`);break;case "provider_incompatible":l.push(`${g}:incompatible`);break;case "gateway_unavailable":this.markUnhealthy(g,"gateway_unavailable"),l.push(ne(g,_));break;default:l.push(ne(g,_));break}}}if(!d||!c)throw new q(l);let y=this,m=r?.maxMidStreamFailovers??2,E=new Set([d]);async function*u(){let h=d,g=c[Symbol.asyncIterator](),C={role:"assistant",content:[]},_={segments:[]},I=Ge(h.type);_.segments.push(I);let Re=0,z=null,X=false,de=m,O="",N="";for(y.emitEngineEvent({kind:"segment_start",callId:s,id:randomUUID(),ts:Date.now(),provider:h.type});;)try{for(;;){let A=await g.next();if(A.done)break;let T=A.value,k=Pt(T,Re,C,X,z);if(k.action==="suppress"){"coalesceMode"in k&&(z=k.coalesceMode),k.clearedJustFailedOver&&(X=!1);continue}k.coalesceModeCleared&&(z=null),k.clearedJustFailedOver&&(X=!1);let v=k.event;if(St(C,v),v.type==="message_start"){if(_.segments.length>1)continue;O=v.message?.model||O,v.message?.usage?.input_tokens!=null&&(I.inputTokens=v.message.usage.input_tokens);}if(v.type==="message_delta"&&(v.delta?.stop_reason&&(N=v.delta.stop_reason),v.usage)){Rt(I,h.type,v.usage),yield {...v,usage:void 0};continue}if(v.type==="message_stop")break;yield v;}if(C.content.length===0&&I.outputTokens===0)throw new Error("stream ended without producing any content");y.markHealthy(h.type),y.emitEngineEvent({kind:"segment_complete",callId:s,id:randomUUID(),ts:Date.now(),provider:h.type,usage:{...I}});let S=Ke(_.segments);yield {type:"message_delta",delta:{stop_reason:N||"end_turn"},usage:S},yield {type:"message_stop"},y.emitEngineEvent({kind:"stream_complete",callId:s,id:randomUUID(),ts:Date.now(),stitched:_.segments.length>1,segments:_.segments.map(A=>({...A})),provider:_.segments.map(A=>A.provider).join("+"),model:O,inputTokens:S.input_tokens,outputTokens:S.output_tokens,costUsd:_.segments.reduce((A,T)=>A+T.costUsd,0),userCostUsd:_.segments.reduce((A,T)=>A+T.userCostUsd,0),durationMs:Date.now()-o,stopReason:N||"end_turn",fallbackChain:l,rateLimitHit:!1,securityState:r?.licenseState});return}catch(S){if(S?.name==="AbortError"||r?.signal?.aborted)throw S;let A=Ee(S);try{await g.return?.();}catch{}if(y.markUnhealthy(h.type,String(S)),y.emitEngineEvent({kind:"mid_stream_failover_attempt",callId:s,id:randomUUID(),ts:Date.now(),from:h.type,bytesBuffered:Ve(C),errorClass:A,retriesLeft:de}),l.push(`${h.type}:partial(${Ve(C)}b)`),de<=0)throw y.emitTerminalExhaustion(s,_,O,N,l,o,r?.licenseState),S;let T=kt(C.content),k=[...e.messages,{role:"assistant",content:T}],v=null,Q=null,Se=h;for(;;){let w=y.nextHealthyProvider(i,Se,a,E);if(!w)break;Se=w,E.add(w);let je={...e,messages:k,...w.type!=="anthropic-api"&&w.type!=="claude-cli"?{system:wt(e.system,"[SYSTEM]: CONTINUE_LAST_RESPONSE_WITHOUT_REPEATING")}:{}};try{Q=await w.streamRaw(je,{signal:r?.signal}),v=w;break}catch(Pe){y.markUnhealthy(w.type,String(Pe)),l.push(`${w.type}:failed_connect:${String(Pe).slice(0,We)}`);}}if(!v||!Q)throw y.emitTerminalExhaustion(s,_,O,N,l,o,r?.licenseState),S;l.push(`${v.type}:stitched`),y.emitEngineEvent({kind:"mid_stream_failover_success",callId:s,id:randomUUID(),ts:Date.now(),from:h.type,to:v.type}),Re=It(C)+1,de--,h=v,g=Q[Symbol.asyncIterator](),X=true,z=null,I=Ge(h.type),_.segments.push(I),y.emitEngineEvent({kind:"segment_start",callId:s,id:randomUUID(),ts:Date.now(),provider:h.type});}}let f=u();return r?.licenseState==="SEC_DEGRADED"&&(f=Tt(f)),L(f)}nextHealthyProvider(e,t,r,s){let i=e.indexOf(t);if(i===-1)return null;for(let a=i+1;a<e.length;a++){let o=e[a];if(!s.has(o)&&!(r&&!o.supportsMultimodal)&&o.supportsStreaming)return o}return null}emitEngineEvent(e){this.engineEvents.emitTyped(e);}emitTerminalExhaustion(e,t,r,s,i,a,o){let l=Ke(t.segments);this.emitEngineEvent({kind:"stream_complete",callId:e,id:randomUUID(),ts:Date.now(),stitched:t.segments.length>1,segments:t.segments.map(d=>({...d})),provider:t.segments.map(d=>d.provider).join("+"),model:r,inputTokens:l.input_tokens,outputTokens:l.output_tokens,costUsd:t.segments.reduce((d,c)=>d+c.costUsd,0),userCostUsd:t.segments.reduce((d,c)=>d+c.userCostUsd,0),durationMs:Date.now()-a,stopReason:s||"error",fallbackChain:i,rateLimitHit:false,securityState:o});}};function Ge(n){return {provider:n,inputTokens:0,outputTokens:0,costUsd:0,userCostUsd:0}}function Rt(n,e,t){typeof t.input_tokens=="number"&&(n.inputTokens=t.input_tokens),typeof t.output_tokens=="number"&&(e==="anthropic-api"||e==="claude-cli"?n.outputTokens+=t.output_tokens:n.outputTokens=t.output_tokens);}function Ke(n){return {input_tokens:n[n.length-1]?.inputTokens??0,output_tokens:n.reduce((e,t)=>e+t.outputTokens,0)}}function St(n,e){if(e.type==="content_block_start"){let t=e.content_block;t?.type==="text"?n.content.push({type:"text",text:t.text||"",open:true}):t?.type==="tool_use"&&n.content.push({type:"tool_use",id:t.id,name:t.name,input:{},_partialJson:"",open:true});}else if(e.type==="content_block_delta"){let t=n.content[n.content.length-1];if(!t)return;t.type==="text"&&e.delta?.type==="text_delta"?t.text+=e.delta.text||"":t.type==="tool_use"&&e.delta?.type==="input_json_delta"&&(t._partialJson=(t._partialJson||"")+(e.delta.partial_json||""));}else if(e.type==="content_block_stop"){for(let t=n.content.length-1;t>=0;t--)if(n.content[t].open){if(n.content[t].open=false,n.content[t].type==="tool_use"){let r=n.content[t];try{r.input=JSON.parse(r._partialJson||"{}");}catch{}delete r._partialJson;}break}}}function Pt(n,e,t,r,s){if(s&&n.type==="content_block_delta"&&n.index===s.fromIndex)return {action:"yield",event:{...n,index:s.toIndex}};if(s&&n.type==="content_block_stop"&&n.index===s.fromIndex)return {action:"suppress",coalesceMode:null,coalesceModeCleared:true};if(r&&n.type==="content_block_start"&&n.content_block?.type==="text"){let a=t.content[t.content.length-1];if(a?.type==="text"&&a.open===true)return {action:"suppress",coalesceMode:{fromIndex:n.index+e,toIndex:t.content.length-1},clearedJustFailedOver:true}}let i=r&&n.type==="content_block_start";return typeof n.index=="number"&&e>0?{action:"yield",event:{...n,index:n.index+e},clearedJustFailedOver:i}:{action:"yield",event:n,clearedJustFailedOver:i}}function kt(n){let e=[];for(let t of n)t.type==="text"?t.text&&e.push({type:"text",text:t.text}):t.type==="tool_use"&&!t.open&&t.input&&Object.keys(t.input).length>0&&e.push({type:"tool_use",id:t.id,name:t.name,input:t.input});return e}function wt(n,e){return n?`${n}
|
|
14
14
|
|
|
15
15
|
${e}`:e}function Ve(n){let e=0;for(let t of n.content)t.type==="text"?e+=t.text.length:t.type==="tool_use"&&(e+=(t._partialJson||"").length);return e}function It(n){return Math.max(0,n.content.length-1)}async function*Tt(n){for await(let e of n)await new Promise(t=>setTimeout(t,50+Math.random()*50)),yield e;}ee();function se(n,e){if(e==="prefer-user"){let s=n.filter(a=>a.type==="claude-cli"),i=n.filter(a=>a.type!=="claude-cli");return [...s,...i]}if(e==="prefer-app"){let s=n.filter(a=>a.type==="claude-cli");return [...n.filter(a=>a.type!=="claude-cli"),...s]}let t=n.filter(s=>s.type===e),r=n.filter(s=>s.type!==e);return [...t,...r]}function ie(n){return (e,t)=>n.maxCostPerCall!==void 0&&t.estimatedCost>n.maxCostPerCall?"prefer-user":n.preferLocal?"ollama":"prefer-user"}var ve={"claude-opus-4":[15,75],"claude-sonnet-4":[3,15],"claude-haiku-4":[.8,4],"gpt-4o":[2.5,10],"gpt-4o-mini":[.15,.6],"gpt-4-turbo":[10,30],"gpt-3.5-turbo":[.5,1.5],llama:[0,0],mistral:[0,0],codellama:[0,0]};function ae(n){if(ve[n])return ve[n];for(let[e,t]of Object.entries(ve))if(n.startsWith(e))return t;return [3,15]}function oe(n){let e=0;for(let o of n.messages)if(typeof o.content=="string")e+=o.content.length;else if(Array.isArray(o.content))for(let l of o.content)l.type==="text"&&l.text&&(e+=l.text.length);n.system&&(e+=n.system.length);let t=Math.ceil(e/4),r=Math.min(n.max_tokens||1024,4096),[s,i]=ae(n.model),a=(t*s+r*i)/1e6;return {estimatedInputTokens:t,estimatedCost:Math.round(a*1e4)/1e4,model:n.model,provider:""}}var D=class{endpoint;enabled;keyId;lineageBuffer=new Map;unsubscribers=[];constructor(e){this.endpoint=e.endpoint||"https://api.inference-relay.com/v1/events",this.enabled=e.enabled??true,this.keyId=e.keyId||"";}subscribeToEngine(e){let t=s=>this.bufferLineageEvent(s),r=s=>{s.kind==="stream_complete"&&this.flushLineageForCall(s);};e.engineEvents.onTyped("segment_start",t),e.engineEvents.onTyped("segment_complete",t),e.engineEvents.onTyped("mid_stream_failover_attempt",t),e.engineEvents.onTyped("mid_stream_failover_success",t),e.engineEvents.onTyped("stream_complete",r),this.unsubscribers.push(()=>{e.engineEvents.offTyped("segment_start",t),e.engineEvents.offTyped("segment_complete",t),e.engineEvents.offTyped("mid_stream_failover_attempt",t),e.engineEvents.offTyped("mid_stream_failover_success",t),e.engineEvents.offTyped("stream_complete",r);});}unsubscribe(){for(let e of this.unsubscribers)e();this.unsubscribers=[],this.lineageBuffer.clear();}bufferLineageEvent(e){let t=this.lineageBuffer.get(e.callId)??[];t.push(e),this.lineageBuffer.set(e.callId,t);}flushLineageForCall(e){let t=this.lineageBuffer.get(e.callId)??[];this.lineageBuffer.delete(e.callId);let r={timestamp:e.ts,provider:e.provider,model:e.model,inputTokens:e.inputTokens,outputTokens:e.outputTokens,costUsd:e.costUsd,userCostUsd:e.userCostUsd,durationMs:e.durationMs,stopReason:e.stopReason,fallbackChain:e.fallbackChain,rateLimitHit:e.rateLimitHit,securityState:e.securityState,stitched:e.stitched,segments:e.segments,lineage:t.map(s=>s.id),promptContent:false,completionContent:false};this.report(r);}report(e){!this.enabled||!this.keyId||fetch(this.endpoint,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({keyId:this.keyId,provider:e.provider,model:e.model,inputTokens:e.inputTokens,outputTokens:e.outputTokens,costUsd:e.costUsd,...e.userCostUsd!==void 0?{userCostUsd:e.userCostUsd}:{},durationMs:e.durationMs,fallback:e.fallbackChain.length>1,timestamp:e.timestamp,...e.stitched?{stitched:true,segmentCount:e.segments?.length??0}:{},...e.lineage?{lineageIds:e.lineage}:{}})}).catch(()=>{});}};var xt="https://api.inference-relay.com",Lt="/v1/validate",Mt=1440*60*1e3,Dt=10080*60*1e3,Ot=`-----BEGIN PUBLIC KEY-----
|
|
16
16
|
MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEAzk88Bt9LZSmuSd8ulKqy
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "inference-relay",
|
|
3
|
-
"version": "1.
|
|
3
|
+
"version": "1.6.0",
|
|
4
4
|
"private": false,
|
|
5
5
|
"description": "Route AI inference to the user's existing subscription",
|
|
6
6
|
"type": "module",
|
|
@@ -59,7 +59,7 @@
|
|
|
59
59
|
"lint": "eslint src/",
|
|
60
60
|
"typecheck": "tsc --noEmit",
|
|
61
61
|
"clean": "rm -rf dist",
|
|
62
|
-
"prepublishOnly": "echo '[publish] 1.
|
|
62
|
+
"prepublishOnly": "echo '[publish] 1.6.0 release — guard intentionally relaxed for this publish' && exit 0"
|
|
63
63
|
},
|
|
64
64
|
"devDependencies": {
|
|
65
65
|
"@types/node": "^22.0.0",
|