inference-relay 1.6.13 → 1.6.14

package/dist/index-yGhLy31i.d.ts

@@ -1,270 +0,0 @@
-interface InferenceRelayConfig {
-    licenseKey: string;
-    endpoint?: string;
-    providers?: ProviderConfig[];
-    fallback?: boolean;
-    routing?: RoutingConfig;
-    concurrency?: ConcurrencyConfig;
-    warmPool?: number;
-    secureEnclave?: SecureEnclaveConfig;
-    telemetry?: boolean;
-    /**
-     * Maximum number of mid-stream failovers per call. After exhaustion, the
-     * original error is rethrown to the consumer. Default: 2.
-     */
-    maxMidStreamFailovers?: number;
-    /**
-     * Per-attempt delay budget (ms) for the gateway-bundle fetch retry loop.
-     * Defaults to `[0, 1000, 4000]` (3 attempts over ~5s). Env override:
-     * `IR_GATEWAY_RETRY_DELAYS_MS=0,5000,15000,60000`. Constructor option
-     * wins over env. Added in 1.6.2 to let ops ride out sustained
-     * Cloudflare/Convex 502s longer than the 5s default.
-     *
-     * Auto-mode (`import 'inference-relay/auto'`) reads the env var only —
-     * there's no constructor-options touchpoint in that flow.
-     */
-    gatewayRetryDelaysMs?: number[];
-    /**
-     * When true, the fallback cascade refuses to fall through from
-     * `claude-cli` to any paid provider (`anthropic-api`, `openai-api`).
-     * Failures bubble up as a clean `claude-cli:failed:*` chain rather
-     * than masking the real error behind a downstream 400/credit_balance
-     * response from a paid provider the user didn't intend to use.
-     *
-     * Also a billing safety net: with this on, a user who set
-     * `IR_LICENSE_KEY` but happens to have `ANTHROPIC_API_KEY` in env
-     * (e.g. for non-IR tooling) can't accidentally get charged on the
-     * paid API when claude-cli flakes. Added in 1.6.6.
-     */
-    subscriptionOnly?: boolean;
-    /**
-     * Maximum same-provider retries on `transient` failures (e.g., Bun
-     * crash on Windows under concurrent spawn — `CLI exited with code
-     * 3221226505`, illegal-instruction panics, empty-stdout closes).
-     * Defaults to 2 retries with exponential backoff (250ms, 1000ms).
-     * Set to 0 to disable. Added in 1.6.6.
-     */
-    maxSameProviderRetries?: number;
-}
-/**
- * Security state of the licence validation chain. Mirrors the state machine
- * documented in patent §5.3.4. SEC_DEGRADED enables Asynchronous Timing
- * Decorrelation (jitter) on outbound stream events.
- */
-type SecurityState = 'SEC_NOMINAL' | 'SEC_CACHE' | 'SEC_RECOVERY' | 'SEC_DEGRADED';
-interface ProviderConfig {
-    type: 'claude-cli' | 'kimi-cli' | 'anthropic-api' | 'openai-api' | 'ollama' | 'custom';
-    priority?: number;
-    apiKey?: string;
-    baseUrl?: string;
-    options?: Record<string, unknown>;
-}
-interface RelayResult {
-    content: ContentBlock[];
-    model: string;
-    usage: {
-        input_tokens: number;
-        output_tokens: number;
-        /**
-         * Prompt-cache tokens, when the provider reports them (Claude CLI, Anthropic API).
-         * `input_tokens` alone is the UNCACHED remainder: for a Claude CLI call it is often 3 to 10
-         * while tens of thousands of tokens ride in the cache fields. Cost accounting adds all three.
-         * Added 2026-09-10 (cache-usage-2026-09-10, ec8c30e); ported 2026-10-05.
-         */
-        cache_creation_input_tokens?: number;
-        cache_read_input_tokens?: number;
-    };
-    stop_reason: string;
-    provider: string;
-    costUsd: number;
-    userCostUsd?: number;
-    durationMs: number;
-    fallbackChain: string[];
-    headers?: Record<string, string>;
-}
-interface CreateParams {
-    model: string;
-    max_tokens: number;
-    messages: Message[];
-    system?: string;
-    temperature?: number;
-    tools?: unknown[];
-    stream?: boolean;
-    metadata?: Record<string, string>;
-}
-interface SegmentUsage {
-    provider: string;
-    inputTokens: number;
-    outputTokens: number;
-    costUsd: number;
-    userCostUsd: number;
-}
-interface AuditEvent {
-    timestamp: number;
-    provider: string;
-    model: string;
-    inputTokens: number;
-    outputTokens: number;
-    costUsd: number;
-    userCostUsd?: number;
-    durationMs: number;
-    stopReason: string;
-    fallbackChain: string[];
-    rateLimitHit: boolean;
-    securityState?: SecurityState;
-    metadata?: Record<string, string>;
-    promptContent: false;
-    completionContent: false;
-    /** True when the response was stitched across multiple providers. */
-    stitched?: boolean;
-    /** Per-segment usage when stitched; absent for single-segment calls. */
-    segments?: SegmentUsage[];
-    /** EngineEvent IDs correlated to this call (cryptographic lineage). */
-    lineage?: string[];
-}
-/**
- * Engine-internal event union. Emitted by FallbackEngine on a synchronous
- * Node EventEmitter so the TelemetryReporter can buffer them per-call and
- * flush a coherent AuditEvent at terminal stream_complete.
- *
- * Every variant carries id/callId/ts via EngineEventBase intersection so the
- * compiler enforces correlation fields on every emit.
- */
-interface EngineEventBase {
-    /** Unique per-event UUID. */
-    id: string;
-    /** Stable per-engine.stream() invocation UUID for correlation. */
-    callId: string;
-    /** Date.now() at emit time. */
-    ts: number;
-}
-type EngineEvent = (EngineEventBase & {
-    kind: 'segment_start';
-    provider: string;
-}) | (EngineEventBase & {
-    kind: 'segment_complete';
-    provider: string;
-    usage: SegmentUsage;
-}) | (EngineEventBase & {
-    kind: 'mid_stream_failover_attempt';
-    from: string;
-    bytesBuffered: number;
-    errorClass: string;
-    retriesLeft: number;
-}) | (EngineEventBase & {
-    kind: 'mid_stream_failover_success';
-    from: string;
-    to: string;
-}) | (EngineEventBase & {
-    kind: 'stream_complete';
-    stitched: boolean;
-    segments: SegmentUsage[];
-    provider: string;
-    model: string;
-    inputTokens: number;
-    outputTokens: number;
-    costUsd: number;
-    userCostUsd?: number;
-    durationMs: number;
-    stopReason: string;
-    fallbackChain: string[];
-    rateLimitHit: boolean;
-    securityState?: SecurityState;
-});
-interface ContentBlock {
-    type: 'text' | 'tool_use';
-    text?: string;
-    id?: string;
-    name?: string;
-    input?: unknown;
-}
-interface Message {
-    role: 'user' | 'assistant';
-    content: string | ContentBlock[];
-}
-interface RoutingConfig {
-    maxCostPerCall?: number;
-    preferLocal?: boolean;
-}
-interface ConcurrencyConfig {
-    maxParallel?: number;
-}
-/**
- * Configuration for the Hardware-Authorized Secure Enclave credential resolver.
- * Controls how the relay reads credentials from the platform's secure enclave.
- */
-interface SecureEnclaveConfig {
-    enabled?: boolean;
-    refreshTokens?: boolean;
-    allowPopup?: boolean;
-    fallbackToEnv?: boolean;
-}
-interface CostEstimate {
-    estimatedInputTokens: number;
-    estimatedCost: number;
-    model: string;
-    provider: string;
-}
-interface ProbeResult {
-    available: boolean;
-    authenticated: boolean;
-    subscriptionType?: string;
-    rateLimitTier?: string;
-    cliVersion?: string;
-    versionWarning?: string;
-}
-interface ProviderHealth {
-    status: 'healthy' | 'degraded' | 'unhealthy';
-    lastError?: string;
-    lastCheckAt: number;
-    errorCount: number;
-}
-interface RateLimitEvent {
-    provider: string;
-    utilization: number;
-    resetsAt: number;
-}
-/**
- * Credentials resolved from the Hardware-Authorized Secure Enclave.
- * Used by the Native Subscription Gateway provider.
- */
-interface SecureEnclaveCredentials {
-    accessToken: string;
-    refreshToken: string;
-    expiresAt: number;
-    subscriptionType: string;
-    rateLimitTier: string;
-    organizationUuid?: string;
-}
-/**
- * Structured error thrown when the license's effective cap (base cap plus
- * active top-up grants) has been exceeded for the current billing cycle.
- *
- * The library enters a "High-Frequency Probe" state on the first occurrence
- * of this error per instance: it re-validates against /v1/validate every
- * 15 seconds for up to 5 minutes, so a top-up purchased through the
- * dashboard resumes inference within seconds — not after the next normal
- * cache refresh.
- *
- * Consumers can catch this error and surface an in-app CTA to their end
- * users by inspecting the `resolutionUrl` and pack fields.
- */
-declare class RelayCapExceededError extends Error {
-    readonly code: "RELAY_CAP_EXCEEDED";
-    readonly tier: string;
-    readonly baseCap: number | null;
-    readonly topUps: number;
-    readonly effectiveCap: number | null;
-    readonly usedThisCycle: number;
-    readonly resolutionUrl: string;
-    constructor(opts: {
-        tier: string;
-        baseCap: number | null;
-        topUps: number;
-        effectiveCap: number | null;
-        usedThisCycle: number;
-        resolutionUrl?: string;
-    });
-}
-
-export { type AuditEvent as A, type CreateParams as C, type EngineEvent as E, type InferenceRelayConfig as I, type Message as M, type ProbeResult as P, type RelayResult as R, type SecurityState as S, type CostEstimate as a, type ProviderHealth as b, type RateLimitEvent as c, type RoutingConfig as d, type ConcurrencyConfig as e, type ContentBlock as f, type ProviderConfig as g, RelayCapExceededError as h, type SecureEnclaveConfig as i, type SecureEnclaveCredentials as j, type SegmentUsage as k };

package/dist/index.cjs

@@ -1,24 +0,0 @@
-'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. */
-var it=Object.defineProperty;var at=(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 ot=(n,e)=>()=>(n&&(e=n(n=0)),e);var lt=(n,e)=>{for(var t in e)it(n,t,{get:e[t],enumerable:true});};var Ke={};lt(Ke,{detectCLI:()=>V});function At(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 V(){if(!At("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&&!Rt.some(r=>e.startsWith(r.split(".").slice(0,2).join(".")))&&(t=`CLI version ${e} is not tested with this version of inference-relay. Streaming may be unstable.`),{available:true,authenticated:true,cliVersion:e||void 0,versionWarning:t}}var Rt,se=ot(()=>{Rt=["2.1.87"];});var K=class n extends Error{code="RELAY_CAP_EXCEEDED";tier;baseCap;topUps;effectiveCap;usedThisCycle;resolutionUrl;constructor(e){let t=e.resolutionUrl??"https://inference-relay.com/dashboard/billing?topup=1";super(`Cap reached (${e.usedThisCycle} / ${e.effectiveCap??"\u221E"} calls this cycle). Buy a top-up or upgrade at ${t}`),this.name="RelayCapExceededError",this.tier=e.tier,this.baseCap=e.baseCap,this.topUps=e.topUps,this.effectiveCap=e.effectiveCap,this.usedThisCycle=e.usedThisCycle,this.resolutionUrl=t,Object.setPrototypeOf(this,n.prototype);}};function M(n,e){let t=new Map,r=null,s=null,i=new Promise((p,f)=>{r=p,s=f;});i.catch(()=>{});let a=[],o=null,l="",d="";function c(p,...f){let u=t.get(p);u&&u.forEach(h=>h(...f));}async function*g(){try{for await(let p of n)if(yield p,p.type==="content_block_start")a[p.index]=p.content_block;else if(p.type==="content_block_delta"){let f=a[p.index];f&&p.delta?.type==="text_delta"&&(f.text=(f.text||"")+p.delta.text,c("text",p.delta.text));}else if(p.type==="message_start")l=p.message?.model||"",o=p.message?.usage;else if(p.type==="message_delta")d=p.delta?.stop_reason||"",p.usage&&(o={...o,...p.usage});else if(p.type==="message_stop"){let f={role:"assistant",model:l,content:a.filter(Boolean),usage:o,stop_reason:d};c("message",f),r&&(r(f),r=null),e&&e(f);}if(r){let p={role:"assistant",model:l,content:a.filter(Boolean),usage:o,stop_reason:d||"end_turn"};c("message",p),r(p),r=null,e&&e(p);}}catch(p){throw c("error",p),s&&(s(p),s=null),p}}let m=g(),y=false,_={[Symbol.asyncIterator](){if(y)throw new Error("[inference-relay] RelayStream can only be iterated once");return y=true,m},on(p,f){return t.has(p)||t.set(p,[]),t.get(p).push(f),_},finalMessage(){let p,f=new Promise((u,h)=>{p=setTimeout(()=>h(new Error("[inference-relay] finalMessage() timed out \u2014 no message_stop received after 5 minutes")),3e5);});return Promise.race([i,f]).finally(()=>clearTimeout(p))}};return _}var C=class{priority=100;async streamRaw(e,t){throw new Error(`[inference-relay] Provider '${this.type}' does not implement streamRaw(). Override either stream() or streamRaw() in the provider class.`)}async stream(e,t){return M(await this.streamRaw(e,t))}};var ve=`-----BEGIN PUBLIC KEY-----
-MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEAzk88Bt9LZSmuSd8ulKqy
-EGNiPB+r8u1ZinBvMM4qX8dp4kzJEC/47L1RYrrxCphD05HToREjljbj+upH6rZH
-t6D/5funYPi2TO1IOSC6BRVRwMqqTmJGVJcLcG/WF3gTE+2UQ3q5EppWJCEKcz6x
-kfynZEzOcLikoQALIxEjxRfi6g83wNJS+WBwAJQ68s7N9FHE3rkLzEqv3FJRP+ZR
-7lqwqIyuCLr9j9yc4XqF6VY9yooD2hyw/n1EVFejzjDgntWTqIkaCIXwh8lDzRrK
-Rxn07gcI6ea2Eym9sXh4xZLFudXfAT2DFsH/XboDUuMKAV1YHc1NsIGApesWe2M+
-UQIDAQAB
------END PUBLIC KEY-----`,Ee={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>"},D=class{verificationFailures=0;cachedManifest=null;currentSecurityState="SEC_NOMINAL";async verifyManifest(e){try{let{compactVerify:t,importSPKI:r}=await import('jose'),s=await r(ve,"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:Ee,securityState:"SEC_DEGRADED"}}return this.cachedManifest?(this.currentSecurityState="SEC_CACHE",{manifest:this.cachedManifest,securityState:"SEC_CACHE"}):(this.currentSecurityState="SEC_RECOVERY",{manifest:Ee,securityState:"SEC_RECOVERY"})}}getSecurityState(){return this.currentSecurityState}clearCache(){this.cachedManifest=null,this.verificationFailures=0,this.currentSecurityState="SEC_NOMINAL";}},ct=new D;function De(){ct.clearCache();}function ut(){return Ee}function dt(n,e){return e.split(".").reduce((t,r)=>{if(!(r==="__proto__"||r==="constructor"||r==="prototype"))return t&&t[r]},n)}function S(n,...e){process.env.IR_DEBUG==="1"&&console.error(`[ir-debug:${n}]`,...e);}var gt=process.env.IR_GATEWAY_URL||"https://api.inference-relay.com";function Ce(n){return path.join(os.homedir(),".inference-relay","gateway",n)}function Fe(n){return path.join(Ce(n),"bundle.js")}function He(n){return path.join(Ce(n),"meta.json")}var E=class extends Error{constructor(t,r,s){super(t);this.code=r;this.cause=s;this.name="GatewayLoaderError";}code;cause},x=class extends E{constructor(e,t){super(e,"GATEWAY_FETCH_FAILED",t);}},G=new Map,ft=[0,1e3,4e3];function _t(n){let e=(r,s)=>{if(!Array.isArray(r)||r.length===0)return null;let i=[];for(let a of r){let o=typeof a=="string"?Number(a):a;if(typeof o!="number"||!Number.isFinite(o)||o<0)return S("loader.retry-config-invalid",{source:s,badValue:a}),null;i.push(o);}return i};if(n!==void 0){let r=e(n,"option");if(r)return r}let t=process.env.IR_GATEWAY_RETRY_DELAYS_MS;if(t){let r=e(t.split(",").map(s=>s.trim()),"env");if(r)return r}return ft}var N=class{static async load(e,t,r=gt,s){let i=`${t}:${e.slice(-8)}`;if(G.has(i))return G.get(i);try{let m=await this.tryLoadFromDisk(t);if(m)return S("loader.cache-hit",{libraryVersion:t}),G.set(i,m),m}catch{}S("loader.cache-miss",{libraryVersion:t});let{bytes:a,sha256:o,signature:l,libraryVersion:d}=await this.fetchFromAuthority(e,t,r,s);if(crypto$1.createHash("sha256").update(a).digest("hex")!==o)throw new E("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: HASH_MISMATCH)","GATEWAY_VERIFICATION_FAILED");await this.verifySignature(l,d,o);try{let m=Ce(t);fs.mkdirSync(m,{recursive:!0}),fs.writeFileSync(Fe(t),a),fs.writeFileSync(He(t),JSON.stringify({sha256:o,signature:l,libraryVersion:d}));}catch{}let g=this.loadIntoVm(a.toString("utf-8"));return G.set(i,g),g}static async loadFromDisk(e){if(!fs.existsSync(e))throw new E(`[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=Fe(e),r=He(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$1.createHash("sha256").update(s).digest("hex")!==i.sha256?null:(await this.verifySignature(i.signature,i.libraryVersion,i.sha256),this.loadIntoVm(s.toString("utf-8")))}static async fetchFromAuthority(e,t,r,s){let i,a=_t(s?.retryDelaysMs),o;for(let g=0;g<a.length;g++){let m=a[g];m>0&&await new Promise(y=>setTimeout(y,m)),S("loader.fetch",{url:`${r}/v1/gateway/fetch`,libraryVersion:t,attempt:g+1,delayMs:m});try{let y=await fetch(`${r}/v1/gateway/fetch`,{method:"POST",headers:{"Content-Type":"application/json",Authorization:`Bearer ${e}`},body:JSON.stringify({libraryVersion:t})});if(S("loader.fetch-result",{status:y.status,attempt:g+1}),y.status===401||y.status===403)throw new E("[ir-system]: LICENSE_REJECTED (REASON: GATEWAY_FETCH_UNAUTHORIZED)","LICENSE_REJECTED");if(y.status>=500&&y.status<600){o=new x(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: HTTP_${y.status})`);continue}if(!y.ok)throw new x(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: HTTP_${y.status})`);i=await y.json(),o=null;break}catch(y){if(y instanceof E&&y.code==="LICENSE_REJECTED")throw y;o=y;}}if(o||!i)throw o instanceof E?o:new x("[ir-system]: GATEWAY_FETCH_FAILED (REASON: NETWORK_ERROR)",o);let l;try{l=await fetch(i.bundleUrl);}catch(g){throw new x("[ir-system]: GATEWAY_FETCH_FAILED (REASON: BUNDLE_DOWNLOAD_FAILED)",g)}if(!l.ok)throw new x(`[ir-system]: GATEWAY_FETCH_FAILED (REASON: BUNDLE_HTTP_${l.status})`);let d=await l.arrayBuffer();return {bytes:Buffer.from(d),sha256:i.sha256,signature:i.signature,libraryVersion:i.libraryVersion,bundleVersion:i.bundleVersion}}static async verifySignature(e,t,r){try{let{compactVerify:s,importSPKI:i}=await import('jose'),a=await i(ve,"RS256"),{payload:o}=await s(e,a),l=JSON.parse(new TextDecoder().decode(o));if(l.purpose!=="gateway-bundle")throw new E("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: PURPOSE_MISMATCH)","GATEWAY_VERIFICATION_FAILED");if(l.libraryVersion!==t)throw new E("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: LIBRARY_VERSION_MISMATCH)","GATEWAY_VERIFICATION_FAILED");if(l.sha256!==r)throw new E("[ir-system]: GATEWAY_VERIFICATION_FAILED (REASON: SHA256_MISMATCH)","GATEWAY_VERIFICATION_FAILED")}catch(s){throw s instanceof E?s:new E("[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 E("[ir-system]: GATEWAY_LOAD_FAILED (REASON: MISSING_CLAUDE_CLI_EXPORT)","GATEWAY_LOAD_FAILED");return o}catch(t){throw t instanceof E?t:new E("[ir-system]: GATEWAY_LOAD_FAILED (REASON: VM_EXECUTION_ERROR)","GATEWAY_LOAD_FAILED",t)}}static clearCache(){G.clear();}};var Ge="1.6.13",ie=class extends C{type="claude-cli";supportsStreaming=true;supportsMultimodal=true;platformRequirement="desktop";licenseKey;loadOptions;loadedProvider=null;loadingPromise=null;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e,t){super(),this.licenseKey=e,this.loadOptions=t;}async ensureLoaded(){return this.loadedProvider?this.loadedProvider:this.loadingPromise?this.loadingPromise:(this.loadingPromise=(async()=>{try{let t=(await N.load(this.licenseKey,Ge,void 0,this.loadOptions)).ClaudeCLIProvider,r=new t;return this.loadedProvider=r,r}catch(e){throw this.loadingPromise=null,this.health.status="unhealthy",this.health.errorCount++,e instanceof E?this.health.lastError=e.message:e instanceof Error&&(this.health.lastError=e.message),e}})(),this.loadingPromise)}async warmup(){await this.ensureLoaded();}async probe(){let{detectCLI:e}=await Promise.resolve().then(()=>(se(),Ke));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());}},ae=class extends C{type="kimi-cli";supportsStreaming=true;supportsMultimodal=false;platformRequirement="desktop";licenseKey;loadOptions;loadedProvider=null;loadingPromise=null;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e,t){super(),this.licenseKey=e,this.loadOptions=t;}async ensureLoaded(){return this.loadedProvider?this.loadedProvider:this.loadingPromise?this.loadingPromise:(this.loadingPromise=(async()=>{try{let t=(await N.load(this.licenseKey,Ge,void 0,this.loadOptions)).KimiCLIProvider;if(!t)throw new E("[ir-system]: GATEWAY_LOAD_FAILED (REASON: STALE_BUNDLE_NO_KIMI_CLI)","GATEWAY_LOAD_FAILED");let r=new t;return this.loadedProvider=r,r}catch(e){throw this.loadingPromise=null,this.health.status="unhealthy",this.health.errorCount++,this.health.lastError=e instanceof Error?e.message:String(e),e}})(),this.loadingPromise)}async warmup(){await this.ensureLoaded();}async probe(){let{existsSync:e}=await import('fs'),{homedir:t}=await import('os'),{join:r}=await import('path'),s=process.env.KIMI_CODE_HOME||r(t(),".kimi-code"),i=process.env.IR_KIMI_BIN||r(s,"bin","kimi"),a=e(i),o=e(r(s,"credentials"))||e(r(s,"oauth"));return {available:a,authenticated:o,subscriptionType:o?"kimi-code":void 0}}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 t=JSON.stringify(e.messages).length/4;return {estimatedInputTokens:Math.round(t+2e3),estimatedCost:0,model:e.model,provider:"kimi-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 Y=[{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}}],$=Object.fromEntries(Y.filter(n=>n.canonical!==n.apiId).map(n=>[n.canonical,n.apiId]));$["claude-sonnet-4"]="claude-sonnet-4-20250514";$["claude-opus-4"]="claude-opus-4-20250514";$["claude-haiku-4"]="claude-haiku-4-5-20251001";function Se(n){return $[n]||n}function Ct(n){return Y.find(e=>e.canonical===n)}function oe(n){return Y.find(e=>e.canonical===n)?.provider}function St(n){return Y.filter(e=>e.provider===n)}var wt="https://api.anthropic.com/v1/messages",Pt="2023-06-01",we={"claude-sonnet-4-5-20250514":[3,15],"claude-sonnet-4-20250514":[3,15],"claude-haiku-4-5-20250414":[.8,4],"claude-opus-4-20250514":[15,75]};function Ve(n){if(we[n])return we[n];for(let[e,t]of Object.entries(we))if(n.startsWith(e.split("-20")[0]))return t;return [3,15]}var W=class extends C{type="anthropic-api";supportsStreaming=true;supportsMultimodal=true;platformRequirement="any";apiKey;baseUrl;dangerousDirectBrowserAccess;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.apiKey=e.apiKey,this.baseUrl=e.baseUrl?.replace(/\/$/,"")||wt,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]=Ve(a.model||e.model),c=((a.usage?.input_tokens||0)*l+(a.usage?.output_tokens||0)*d)/1e6,g={},m=["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 y of m){let _=i.headers.get(y);_&&(g[y]=_);}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(g).length>0?g: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 g=o.split(`
-`);o=g.pop()||"";let m=null;for(let y of g)if(y.startsWith("event: "))m=y.slice(7).trim();else if(y.startsWith("data: ")){let _=y.slice(6);if(m&&_)try{let p=JSON.parse(_);p.type==="message_stop"&&(l=!0),yield p;}catch{console.warn("[inference-relay] Skipped malformed SSE data:",_.substring(0,100));}m=null;}else y.trim()===""&&(m=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]=Ve(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":Pt};return this.dangerousDirectBrowserAccess&&(e["anthropic-dangerous-direct-browser-access"]="true"),e}buildRequestBody(e,t){let r={model:Se(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 kt="https://api.openai.com/v1/chat/completions",Ye=["gpt-","o1","o3","o4","chatgpt-","text-","davinci","babbage","ft:"],Pe={"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 $e(n){if(Pe[n])return Pe[n];for(let[e,t]of Object.entries(Pe))if(n.startsWith(e))return t;return [2.5,10]}function It(n){return typeof n=="string"?n:n.filter(e=>e.type==="text"&&e.text).map(e=>e.text).join("")}function Tt(n,e){let t=[];e&&t.push({role:"system",content:e});for(let r of n){let s=It(r.content);if(r.role==="assistant"&&Array.isArray(r.content)){let i=r.content.filter(a=>a.type==="tool_use");if(i.length>0){t.push({role:"assistant",content:s||"",tool_calls:i.map(a=>({id:a.id||`call_${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 We(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 q=class extends C{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(/\/$/,"")||kt;}async probe(){try{let e=this.baseUrl.replace(/\/chat\/completions$/,"/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 u of l.message.tool_calls)d.push({type:"tool_use",id:u.id,name:u.function?.name,input:u.function?.arguments?JSON.parse(u.function.arguments):{}});let c=a.usage?.prompt_tokens||0,g=a.usage?.completion_tokens||0,[m,y]=$e(a.model||e.model),_=(c*m+g*y)/1e6,p={},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 u of f){let h=i.headers.get(u);h&&(p[u]=h);}return {content:d,model:a.model||e.model,usage:{input_tokens:c,output_tokens:g},stop_reason:We(l?.finish_reason),provider:"openai-api",costUsd:Math.round(_*1e6)/1e6,durationMs:o,fallbackChain:["openai-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("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,g,m;yield {type:"message_start",message:{model:o,usage:{input_tokens:0,output_tokens:0}}};try{for(;;){let{done:y,value:_}=await i.read();if(y)break;l+=a.decode(_,{stream:!0});let p=l.split(`
-`);l=p.pop()||"";for(let f of p){if(!f.startsWith("data: "))continue;let u=f.slice(6).trim();if(u==="[DONE]"){yield {type:"message_delta",delta:{stop_reason:m||"end_turn"},usage:g||{}},yield {type:"message_stop"},c=!0;return}try{let h=JSON.parse(u);h.usage&&(g={input_tokens:h.usage.prompt_tokens||0,output_tokens:h.usage.completion_tokens||0});let R=h.choices?.[0];if(!R)continue;let b=R.delta;b?.content&&!d&&(d=!0,yield {type:"content_block_start",index:0,content_block:{type:"text",text:""}}),b?.content&&(yield {type:"content_block_delta",index:0,delta:{type:"text_delta",text:b.content}}),R.finish_reason&&(d&&(yield {type:"content_block_stop",index:0}),m=We(R.finish_reason));}catch{}}}c||(yield {type:"message_delta",delta:{stop_reason:m||"end_turn"},usage:g||{}},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]=$e(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 ${this.apiKey}`}}buildRequestBody(e,t){let r=e.model.toLowerCase();if(!Ye.some(o=>r.startsWith(o)))throw new Error(`OPENAI_MODEL_TRANSLATION_REQUIRED: The OpenAI provider received the model name '${e.model}', which does not match any known OpenAI model prefix (${Ye.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=Tt(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 ${t}`)}recordSuccess(){this.health.status="healthy",this.health.lastCheckAt=Date.now(),this.health.errorCount=0,this.health.lastError=void 0;}};var xt="http://localhost:11434";function Lt(n){return typeof n=="string"?n:n.filter(e=>e.type==="text"&&e.text).map(e=>e.text).join("")}function Ot(n,e){let t=[];e&&t.push({role:"system",content:e});for(let r of n)t.push({role:r.role,content:Lt(r.content)});return t}var j=class extends C{type="ollama";supportsStreaming=true;supportsMultimodal=false;platformRequirement="desktop";baseUrl;health={status:"healthy",lastCheckAt:Date.now(),errorCount:0};constructor(e){super(),this.baseUrl=e?.baseUrl?.replace(/\/$/,"")||xt;}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:g,value:m}=await i.read();if(g)break;l+=a.decode(m,{stream:!0});let y=l.split(`
-`);l=y.pop()||"";for(let _ of y){let p=_.trim();if(p)try{let f=JSON.parse(p);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=Ot(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 ${t}`))}recordSuccess(){this.health.status="healthy",this.health.lastCheckAt=Date.now(),this.health.errorCount=0,this.health.lastError=void 0;}};var J=class extends events.EventEmitter{constructor(){super(),this.setMaxListeners(50);}emitTyped(e){return this.emit(e.kind,e)}onTyped(e,t){return this.on(e,t)}offTyped(e,t){return this.off(e,t)}};var et=200;function le(n,e){if(e instanceof E)return `${n}:failed:${e.code}:${e.message}`;let t=e?.message;return `${n}:failed:${t?.slice(0,et)||"unknown"}`}function qe(n){let e=n.filter(t=>t.includes(":failed:"));return e.length===0?false:e.every(t=>t.endsWith(":failed:request_too_large"))}var je=new Set(["anthropic-api","openai-api","ollama"]),U=class extends Error{code="ALL_PROVIDERS_FAILED";fallbackChain;constructor(e){super(`All providers failed. Chain: ${e.join(" \u2192 ")}`),this.name="RelayError",this.fallbackChain=e;}},z=class extends U{constructor(e){super(e);}};function Je(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 ce(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.includes("prompt is too long")||t.includes("prompt too long")||t.includes("prompt_too_large")||t.includes("input length")||t.includes("input is too long")||t.includes("context_length_exceeded")||t.includes("context length exceeded")||t.includes("request too large")||t.includes("tokens exceed")||t.includes("exceeds the maximum")||t.includes("too many tokens")?"request_too_large":t.includes("code 3221226505")||t.includes("cli exited with code 3 ")||t.includes("cli stream exited with code 3 ")||t.includes("bun has crashed")||t.includes("illegal instruction")||t.includes("stream ended without producing any content")||t.includes("cli returned empty response")||t.includes("cli closed stdout without sending any response")||t.includes("kimi_stall_silent")?"transient":t.includes("credit balance is too low")||t.includes("credit balance too low")?"auth_failed":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 ze(n){return [...n].sort((e,t)=>e.priority-t.priority)}var X=class n{healthMap=new Map;engineEvents=new J;static UNHEALTHY_TTL_DEFAULT_MS=300*1e3;static UNHEALTHY_TTL_AUTH_MS=30*1e3;getHealthStatus(e){let t=this.healthMap.get(e);if(!t)return "healthy";let r=t.ttlMs??n.UNHEALTHY_TTL_DEFAULT_MS;if(t.status==="unhealthy"&&Date.now()-t.updatedAt>=r){let s=Date.now()-t.updatedAt;return this.healthMap.delete(e),S("engine.recovered",{providerType:e,ageMs:s,ttlMs:r}),"healthy"}return t.status}markUnhealthy(e,t,r){this.healthMap.set(e,{status:"unhealthy",lastError:t,updatedAt:Date.now(),ttlMs:r}),S("engine.unhealthy",{providerType:e,reason:t,ttlMs:r});}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),i=[],a=Je(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(`${c}:skipped:no_multimodal`);continue}let m=oe(e.model);if(m&&m!==c&&je.has(c)){i.push(`${c}:skipped:wrong_provider`);continue}let y=r?.maxSameProviderRetries??2,_=0,p=null,f=false;for(;!f;)try{let u=await d.execute(e,{signal:r?.signal});return f=!0,this.markHealthy(c),_>0&&i.push(`${c}:retry_success:${_}`),i.push(`${c}:success`),this.emitEngineEvent({kind:"stream_complete",callId:o,id:crypto$1.randomUUID(),ts:Date.now(),stitched:!1,segments:[{provider:c,inputTokens:u.usage.input_tokens,outputTokens:u.usage.output_tokens,costUsd:u.costUsd,userCostUsd:u.userCostUsd??0}],provider:c,model:u.model,inputTokens:u.usage.input_tokens,outputTokens:u.usage.output_tokens,costUsd:u.costUsd,userCostUsd:u.userCostUsd,durationMs:u.durationMs??Date.now()-l,stopReason:u.stop_reason,fallbackChain:i,rateLimitHit:!1,securityState:r?.licenseState}),{...u,fallbackChain:i}}catch(u){if(u?.name==="AbortError"||r?.signal?.aborted)throw u;if(ce(u)==="transient"&&_<y){_++;let R=250*Math.pow(2,_-1);S("engine.transient-retry",{id:c,attempt:_,backoffMs:R,msg:String(u?.message||"").slice(0,120)}),i.push(`${c}:retry:${_}`),await new Promise(b=>setTimeout(b,R));continue}p=u;break}if(!f){let u=p;switch(ce(u)){case "rate_limited":d.reportRateLimit({provider:c,utilization:1,resetsAt:Date.now()+6e4}),this.markDegraded(c,"rate_limited"),i.push(`${c}:failed:rate_limited`);break;case "auth_failed":this.markUnhealthy(c,"auth_failed",n.UNHEALTHY_TTL_AUTH_MS),i.push(`${c}:failed:auth_failed`);break;case "timeout":i.push(`${c}:failed:timeout`);break;case "provider_incompatible":i.push(`${c}:incompatible`);break;case "gateway_unavailable":this.markUnhealthy(c,"gateway_unavailable"),i.push(le(c,u));break;case "request_too_large":i.push(`${c}:failed:request_too_large`);break;case "transient":this.markDegraded(c,"transient_exhausted"),i.push(`${c}:failed:transient_exhausted`);break;default:i.push(le(c,u));break}if(r?.subscriptionOnly&&c==="claude-cli"){i.push("subscription_only:cascade_halted");break}}}throw S("engine.cascade-exhausted",{chain:i,mode:"execute"}),qe(i)?new z(i):new U(i)}async stream(e,t,r){let s=crypto$1.randomUUID(),i=ze(t),a=Je(e),o=Date.now(),l=[],d=null,c=null;for(let u of i){let h=u.type;if(this.getHealthStatus(h)==="unhealthy"){l.push(`${h}:skipped:unhealthy`);continue}if(a&&!u.supportsMultimodal){l.push(`${h}:skipped:no_multimodal`);continue}if(!u.supportsStreaming){l.push(`${h}:skipped:no_streaming`);continue}let R=oe(e.model);if(R&&R!==h&&je.has(h)){l.push(`${h}:skipped:wrong_provider`);continue}try{c=await u.streamRaw(e,{signal:r?.signal}),d=u,l.push(`${h}:success`);break}catch(b){if(b?.name==="AbortError"||r?.signal?.aborted)throw b;switch(ce(b)){case "rate_limited":u.reportRateLimit({provider:h,utilization:1,resetsAt:Date.now()+6e4}),this.markDegraded(h,"rate_limited"),l.push(`${h}:failed:rate_limited`);break;case "auth_failed":this.markUnhealthy(h,"auth_failed",n.UNHEALTHY_TTL_AUTH_MS),l.push(`${h}:failed:auth_failed`);break;case "timeout":l.push(`${h}:failed:timeout`);break;case "provider_incompatible":l.push(`${h}:incompatible`);break;case "gateway_unavailable":this.markUnhealthy(h,"gateway_unavailable"),l.push(le(h,b));break;case "request_too_large":l.push(`${h}:failed:request_too_large`);break;case "transient":this.markDegraded(h,"transient_exhausted"),l.push(`${h}:failed:transient_exhausted`);break;default:l.push(le(h,b));break}if(r?.subscriptionOnly&&h==="claude-cli"){l.push("subscription_only:cascade_halted");break}}}if(!d||!c)throw S("engine.cascade-exhausted",{chain:l,mode:"stream"}),qe(l)?new z(l):new U(l);let m=this,y=r?.maxMidStreamFailovers??2,_=new Set([d]);async function*p(){let u=d,h=c[Symbol.asyncIterator](),R={role:"assistant",content:[]},b={segments:[]},T=Xe(u.type);b.segments.push(T);let Le=0,ee=null,te=false,be=y,H="",B="";for(m.emitEngineEvent({kind:"segment_start",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),provider:u.type});;)try{for(;;){let A=await h.next();if(A.done)break;let L=A.value,k=Ut(L,Le,R,te,ee);if(k.action==="suppress"){"coalesceMode"in k&&(ee=k.coalesceMode),k.clearedJustFailedOver&&(te=!1);continue}k.coalesceModeCleared&&(ee=null),k.clearedJustFailedOver&&(te=!1);let v=k.event;if(Nt(R,v),v.type==="message_start"){if(b.segments.length>1)continue;H=v.message?.model||H,v.message?.usage?.input_tokens!=null&&(T.inputTokens=v.message.usage.input_tokens);}if(v.type==="message_delta"&&(v.delta?.stop_reason&&(B=v.delta.stop_reason),v.usage)){Dt(T,u.type,v.usage),yield {...v,usage:void 0};continue}if(v.type==="message_stop")break;yield v;}if(R.content.length===0&&T.outputTokens===0)throw new Error("stream ended without producing any content");m.markHealthy(u.type),m.emitEngineEvent({kind:"segment_complete",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),provider:u.type,usage:{...T}});let w=Qe(b.segments);yield {type:"message_delta",delta:{stop_reason:B||"end_turn"},usage:w},yield {type:"message_stop"},m.emitEngineEvent({kind:"stream_complete",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),stitched:b.segments.length>1,segments:b.segments.map(A=>({...A})),provider:b.segments.map(A=>A.provider).join("+"),model:H,inputTokens:w.input_tokens,outputTokens:w.output_tokens,costUsd:b.segments.reduce((A,L)=>A+L.costUsd,0),userCostUsd:b.segments.reduce((A,L)=>A+L.userCostUsd,0),durationMs:Date.now()-o,stopReason:B||"end_turn",fallbackChain:l,rateLimitHit:!1,securityState:r?.licenseState});return}catch(w){if(w?.name==="AbortError"||r?.signal?.aborted)throw w;let A=ce(w);try{await h.return?.();}catch{}if(m.markUnhealthy(u.type,String(w)),m.emitEngineEvent({kind:"mid_stream_failover_attempt",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),from:u.type,bytesBuffered:Ze(R),errorClass:A,retriesLeft:be}),l.push(`${u.type}:partial(${Ze(R)}b)`),be<=0)throw m.emitTerminalExhaustion(s,b,H,B,l,o,r?.licenseState),w;let L=Ft(R.content),k=[...e.messages,{role:"assistant",content:L}],v=null,re=null,Oe=u;for(;;){let I=m.nextHealthyProvider(i,Oe,a,_);if(!I)break;Oe=I,_.add(I);let st={...e,messages:k,...I.type!=="anthropic-api"&&I.type!=="claude-cli"?{system:Ht(e.system,"[SYSTEM]: CONTINUE_LAST_RESPONSE_WITHOUT_REPEATING")}:{}};try{re=await I.streamRaw(st,{signal:r?.signal}),v=I;break}catch(Me){m.markUnhealthy(I.type,String(Me)),l.push(`${I.type}:failed_connect:${String(Me).slice(0,et)}`);}}if(!v||!re)throw m.emitTerminalExhaustion(s,b,H,B,l,o,r?.licenseState),w;l.push(`${v.type}:stitched`),m.emitEngineEvent({kind:"mid_stream_failover_success",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),from:u.type,to:v.type}),Le=Bt(R)+1,be--,u=v,h=re[Symbol.asyncIterator](),te=true,ee=null,T=Xe(u.type),b.segments.push(T),m.emitEngineEvent({kind:"segment_start",callId:s,id:crypto$1.randomUUID(),ts:Date.now(),provider:u.type});}}let f=p();return r?.licenseState==="SEC_DEGRADED"&&(f=Kt(f)),M(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=Qe(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 Xe(n){return {provider:n,inputTokens:0,outputTokens:0,costUsd:0,userCostUsd:0}}function Dt(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 Qe(n){return {input_tokens:n[n.length-1]?.inputTokens??0,output_tokens:n.reduce((e,t)=>e+t.outputTokens,0)}}function Nt(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 Ut(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 Ft(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 Ht(n,e){return n?`${n}
-
-${e}`:e}function Ze(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 Bt(n){return Math.max(0,n.content.length-1)}async function*Kt(n){for await(let e of n)await new Promise(t=>setTimeout(t,50+Math.random()*50)),yield e;}se();function ue(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 de(n){return (e,t)=>n.maxCostPerCall!==void 0&&t.estimatedCost>n.maxCostPerCall?"prefer-user":n.preferLocal?"ollama":"prefer-user"}var pe={"claude-fable-5":[10,50],"claude-mythos-5":[10,50],"claude-opus-5-5":[4,20],"claude-opus-5":[5,25],"claude-opus-4-8":[5,25],"claude-opus-4-7":[5,25],"claude-opus-4-6":[5,25],"claude-opus-4-5":[5,25],"claude-opus-4":[15,75],"claude-sonnet-5":[2,10],"claude-sonnet-4":[3,15],"claude-haiku-5-5":[.1,.5],"claude-haiku-4-5":[1,5],"claude-3-5-haiku":[.8,4],"claude-3-haiku":[.25,1.25],"kimi-k3":[3,15],"kimi-k2.7-code":[.95,4],"kimi-for-coding":[.95,4],"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 me(n){if(pe[n])return pe[n];let e=null;for(let t of Object.keys(pe))n.startsWith(t)&&(e===null||t.length>e.length)&&(e=t);return e!==null?pe[e]:[3,15]}function he(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]=me(n.model),a=(t*s+r*i)/1e6;return {estimatedInputTokens:t,estimatedCost:Math.round(a*1e4)/1e4,model:n.model,provider:""}}var F=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 Q="1.6.13",ye=Object.freeze({streamKeepsHistory:true,streamToolBlocks:true,streamBareToolNames:true,stopsWhenCallerGoes:true,stopsAtFirstCap:true});function ke(){return {name:"inference-relay",version:Q,features:ye}}var Gt="https://api.inference-relay.com",Vt="/v1/validate",Yt=1440*60*1e3,$t=10080*60*1e3,Wt=`-----BEGIN PUBLIC KEY-----
-MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEAzk88Bt9LZSmuSd8ulKqy
-EGNiPB+r8u1ZinBvMM4qX8dp4kzJEC/47L1RYrrxCphD05HToREjljbj+upH6rZH
-t6D/5funYPi2TO1IOSC6BRVRwMqqTmJGVJcLcG/WF3gTE+2UQ3q5EppWJCEKcz6x
-kfynZEzOcLikoQALIxEjxRfi6g83wNJS+WBwAJQ68s7N9FHE3rkLzEqv3FJRP+ZR
-7lqwqIyuCLr9j9yc4XqF6VY9yooD2hyw/n1EVFejzjDgntWTqIkaCIXwh8lDzRrK
-Rxn07gcI6ea2Eym9sXh4xZLFudXfAT2DFsH/XboDUuMKAV1YHc1NsIGApesWe2M+
-UQIDAQAB
------END PUBLIC KEY-----`,O=new Map,Ie=new Set,qt=n=>!!n&&typeof n.aborted=="boolean"&&typeof n.addEventListener=="function";function jt(n,e){let t=n?.signal;if(t===void 0)return {params:n,options:e};let{signal:r,...s}=n,i=s;if(!qt(t))return {params:i,options:e};let a=e?.signal;if(!a||a===t)return {params:i,options:{...e,signal:t}};let o=new AbortController,l=()=>{a.removeEventListener("abort",d),t.removeEventListener("abort",d);},d=()=>{l(),o.abort();};return a.aborted||t.aborted?o.abort():(a.addEventListener("abort",d,{once:true}),t.addEventListener("abort",d,{once:true})),{params:i,options:{...e,signal:o.signal},release:l}}async function Te(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 s=!O.has(n)?1e4:3e3,i=(e||Gt)+Vt,a=async()=>{let d=new AbortController,c=setTimeout(()=>d.abort(),s);try{return await(await fetch(i,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({key:n}),signal:d.signal})).json()}finally{clearTimeout(c);}},o;try{o=await a();}catch{try{if(t)throw new Error("cached answer stands in");o=await a();}catch{return t&&Date.now()-t.cachedAt<$t?t.result:{valid:false,error:"License validation failed (network error). Visit inference-relay.com"}}}let l;if(o.jws)try{let{compactVerify:d,importSPKI:c}=await import('jose'),g=await c(Wt,"RS256"),{payload:m}=await d(o.jws,g);l=JSON.parse(new TextDecoder().decode(m));}catch{return O.delete(n),{valid:false,error:"License validation signature invalid. Possible MITM \u2014 refusing to start."}}else if(l=o,l.valid)return O.delete(n),{valid:false,error:"License validation response not signed. Possible MITM."};return O.set(n,{result:l,cachedAt:Date.now(),ttl:l.valid?Yt:6e4}),l}var xe=class n{static version=Q;get version(){return Q}get features(){return ye}hello(){return ke()}providers=[];config;fallbackEngine;telemetryReporter;routingFn=null;license=null;licensePromise=null;licenseValidatedAt=0;callsSinceValidation=0;manifestDecryptor=new D;capProbeTimer=null;capProbeTicksRemaining=0;messages;constructor(e){if(this.config=e,this.fallbackEngine=new X,this.telemetryReporter=new F({endpoint:e.endpoint?`${e.endpoint}/v1/events`:void 0,enabled:e.telemetry!==false,keyId:e.licenseKey}),this.telemetryReporter.subscribeToEngine(this.fallbackEngine),e.routing&&(this.routingFn=de(e.routing)),e.providers)for(let s of e.providers){let i=typeof s=="string"?{type:s}:s,a=this.createProvider(i);a&&(i.priority!==void 0&&(a.priority=i.priority),this.providers.push(a));}let t=100,r=3600*1e3;this.messages={create:async(s,i)=>{let{params:a,options:o,release:l}=jt(s,i),d=false;try{let c=this.license&&(this.callsSinceValidation>=t||Date.now()-this.licenseValidatedAt>=r);if(c&&(this.license=null),!this.license){this.licensePromise||(this.licensePromise=Te(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 K({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});!c&&!Ie.has(this.config.licenseKey)&&(Ie.add(this.config.licenseKey),console.log(`[inference-relay] LICENSE AUTHORIZED [${(this.license.tier||"unknown").toUpperCase()} TIER] \u2014 READY.`));}else throw new Error(`[inference-relay] ${this.license.error||"License key invalid or expired. Visit inference-relay.com"}`)}if(this.callsSinceValidation++,a.stream){let g=await this.streamMessage(a,o);return d=!0,g}return await this.executeMessage(a,o)}finally{d||l?.();}}};}startCapProbe(){if(this.capProbeTimer!==null)return;let e=15e3,t=20,r=4;this.capProbeTicksRemaining=t;let s=async()=>{this.capProbeTicksRemaining-=1;let i=Math.ceil(this.capProbeTicksRemaining*e/6e4);this.capProbeTicksRemaining>0&&this.capProbeTicksRemaining%r===0&&console.log(`[inference-relay] cap exceeded \u2014 checking for top-up every 15s (${i}m remaining)`),this.config.licenseKey&&O.delete(this.config.licenseKey);try{let a=await this.probeValidate();if(a.valid&&a.capExceeded===!1){console.log("[inference-relay] cap cleared \u2014 probe exiting, inference resuming"),this.license=a,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 Te(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 V();r.available&&r.authenticated!==false&&t.push({type:"claude-cli",priority:1});let s=e?.subscriptionOnly===true||process.env.IR_SUBSCRIPTION_ONLY==="1";return process.env.ANTHROPIC_API_KEY&&!s&&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,subscriptionOnly:s,...e})}async warmup(){let e=this.providers.find(t=>t.type==="claude-cli");e&&"warmup"in e&&typeof e.warmup=="function"&&await e.warmup();}resetHealth(){this.fallbackEngine.resetHealth();}markHealthy(e){this.fallbackEngine.markHealthy(e);}dispose(){this.providers=[],this.config.licenseKey&&(O.delete(this.config.licenseKey),Ie.delete(this.config.licenseKey)),this.license=null,this.licensePromise=null,this.callsSinceValidation=0,this.stopCapProbe(),this.telemetryReporter.unsubscribe(),this.manifestDecryptor.clearCache(),De();}resolveProviders(e){if(!this.routingFn)return this.providers;let t=he(e),r=this.routingFn(e,t);return ue(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(),subscriptionOnly:this.config.subscriptionOnly,maxSameProviderRetries:this.config.maxSameProviderRetries})}async streamMessage(e,t){let r=this.resolveProviders(e);return this.fallbackEngine.stream(e,r,{signal:t?.signal,licenseState:this.manifestDecryptor.getSecurityState(),subscriptionOnly:this.config.subscriptionOnly,maxSameProviderRetries:this.config.maxSameProviderRetries})}createProvider(e){switch(e.type){case "claude-cli":return new ie(this.config.licenseKey,{retryDelaysMs:this.config.gatewayRetryDelaysMs});case "kimi-cli":return new ae(this.config.licenseKey,{retryDelaysMs:this.config.gatewayRetryDelaysMs});case "anthropic-api":return e.apiKey?new W({apiKey:e.apiKey,baseUrl:e.baseUrl}):null;case "openai-api":return e.apiKey?new q({apiKey:e.apiKey,baseUrl:e.baseUrl}):null;case "ollama":return new j({baseUrl:e.baseUrl});default:return null}}};function Jt(n){if(!n)return 0;if(typeof n=="string")return Math.ceil(n.length/4);if(!Array.isArray(n))return 0;let e=0;for(let t of n){if(!t||typeof t!="object")continue;let r=t;if(r.type==="image"||r.type==="image_url"){e+=1500;continue}if(typeof r.text=="string"){e+=Math.ceil(r.text.length/4);continue}try{e+=Math.ceil(JSON.stringify(t).length/4);}catch{e+=50;}}return e}function zt(n){return 8+Jt(n.content)}function Xt(n){let e=50;typeof n.system=="string"&&(e+=Math.ceil(n.system.length/4));for(let t of n.messages)e+=zt(t);return e}se();var Qt="https://api.anthropic.com/v1/messages",Zt="2023-06-01";async function tt(n){try{let e=await fetch(Qt,{method:"POST",headers:{"content-type":"application/json","x-api-key":n,"anthropic-version":Zt},body:JSON.stringify({model:"claude-haiku-4-5-20250414",max_tokens:1,messages:[{role:"user",content:"hi"}]})});return e.status===401?{valid:!1,error:"Invalid API key (401)"}:e.ok?{valid:!0}:e.status===429?{valid:!0,error:"Rate limited (429)"}:e.status===529?{valid:!0,error:"API overloaded (529)"}:{valid:!1,error:`Unexpected status ${e.status}`}}catch(e){return {valid:false,error:`Network error: ${e.message}`}}}var ge=class{filePath;safeStorage=null;memoryFallback=new Map;constructor(e){let t=e?.storagePath||path.join(os.homedir(),".inference-relay");fs.existsSync(t)||fs.mkdirSync(t,{recursive:true}),this.filePath=path.join(t,"credentials.enc");try{let r=at("electron");r?.safeStorage?.isEncryptionAvailable?.()&&(this.safeStorage=r.safeStorage);}catch{}}readDisk(){if(!fs.existsSync(this.filePath))return {};try{let e=fs.readFileSync(this.filePath,"utf-8");return JSON.parse(e)}catch{return {}}}writeDisk(e){fs.writeFileSync(this.filePath,JSON.stringify(e,null,2),{mode:384});}async get(e){if(!this.safeStorage)return this.memoryFallback.get(e)??null;let r=this.readDisk()[e];if(!r)return null;try{let s=Buffer.from(r.data,"base64");return this.safeStorage.decryptString(s)}catch{return null}}async set(e,t){if(!this.safeStorage){this.memoryFallback.set(e,t);return}let r=this.readDisk(),s=this.safeStorage.encryptString(t);r[e]={data:s.toString("base64")},this.writeDisk(r);}async delete(e){if(!this.safeStorage){this.memoryFallback.delete(e);return}let t=this.readDisk();e in t&&(delete t[e],this.writeDisk(t));}async has(e){if(!this.safeStorage)return this.memoryFallback.has(e);let t=this.readDisk();return e in t}};var Z="AES-GCM";var fe=class{dbName="inference-relay-credentials";storeName="keys";cryptoKey=null;passphrase;initPromise=null;constructor(e){this.passphrase=e?.passphrase,e?.dbName&&(this.dbName=e.dbName);}async ensureReady(){if(!this.cryptoKey)return this.initPromise?this.initPromise:(this.initPromise=this.init(),this.initPromise)}async init(){if(await this.openDB(),this.passphrase)this.cryptoKey=await this.deriveKey(this.passphrase,this.getOrCreateSessionSalt());else {let e=typeof sessionStorage<"u"?sessionStorage.getItem("ir-cred-key"):null;if(e){let t=Uint8Array.from(atob(e),r=>r.charCodeAt(0));this.cryptoKey=await crypto.subtle.importKey("raw",t,Z,false,["encrypt","decrypt"]);}else {this.cryptoKey=await crypto.subtle.generateKey({name:Z,length:256},true,["encrypt","decrypt"]);let t=await crypto.subtle.exportKey("raw",this.cryptoKey),r=btoa(String.fromCharCode(...new Uint8Array(t)));typeof sessionStorage<"u"&&sessionStorage.setItem("ir-cred-key",r);}}}openDB(){return new Promise((e,t)=>{let r=indexedDB.open(this.dbName,1);r.onupgradeneeded=()=>{let s=r.result;s.objectStoreNames.contains(this.storeName)||s.createObjectStore(this.storeName);},r.onsuccess=()=>e(r.result),r.onerror=()=>t(r.error);})}async idbGet(e){let t=await this.openDB();return new Promise((r,s)=>{let o=t.transaction(this.storeName,"readonly").objectStore(this.storeName).get(e);o.onsuccess=()=>r(o.result),o.onerror=()=>s(o.error);})}async idbPut(e,t){let r=await this.openDB();return new Promise((s,i)=>{let l=r.transaction(this.storeName,"readwrite").objectStore(this.storeName).put(t,e);l.onsuccess=()=>s(),l.onerror=()=>i(l.error);})}async idbDelete(e){let t=await this.openDB();return new Promise((r,s)=>{let o=t.transaction(this.storeName,"readwrite").objectStore(this.storeName).delete(e);o.onsuccess=()=>r(),o.onerror=()=>s(o.error);})}async idbHas(e){let t=await this.openDB();return new Promise((r,s)=>{let o=t.transaction(this.storeName,"readonly").objectStore(this.storeName).count(e);o.onsuccess=()=>r(o.result>0),o.onerror=()=>s(o.error);})}getOrCreateSessionSalt(){let e="ir-cred-salt";if(typeof sessionStorage<"u"){let r=sessionStorage.getItem(e);if(r)return Uint8Array.from(atob(r),s=>s.charCodeAt(0))}let t=crypto.getRandomValues(new Uint8Array(16));return typeof sessionStorage<"u"&&sessionStorage.setItem(e,btoa(String.fromCharCode(...t))),t}async deriveKey(e,t){let r=new TextEncoder,s=await crypto.subtle.importKey("raw",r.encode(e),"PBKDF2",false,["deriveKey"]);return crypto.subtle.deriveKey({name:"PBKDF2",salt:t.buffer,iterations:1e5,hash:"SHA-256"},s,{name:Z,length:256},false,["encrypt","decrypt"])}async encrypt(e){let t=crypto.getRandomValues(new Uint8Array(12)),r=this.getOrCreateSessionSalt(),s=new TextEncoder().encode(e),i=await crypto.subtle.encrypt({name:Z,iv:t},this.cryptoKey,s);return {iv:btoa(String.fromCharCode(...t)),ciphertext:btoa(String.fromCharCode(...new Uint8Array(i))),salt:btoa(String.fromCharCode(...r))}}async decrypt(e){let t=Uint8Array.from(atob(e.iv),i=>i.charCodeAt(0)),r=Uint8Array.from(atob(e.ciphertext),i=>i.charCodeAt(0)),s=await crypto.subtle.decrypt({name:Z,iv:t},this.cryptoKey,r);return new TextDecoder().decode(s)}async get(e){await this.ensureReady();let t=await this.idbGet(e);if(!t)return null;try{return await this.decrypt(t)}catch{return null}}async set(e,t){await this.ensureReady();let r=await this.encrypt(t);await this.idbPut(e,r);}async delete(e){await this.ensureReady(),await this.idbDelete(e);}async has(e){return await this.ensureReady(),this.idbHas(e)}};var _e=class{store=new Map;async get(e){return this.store.get(e)??null}async set(e,t){this.store.set(e,t);}async delete(e){this.store.delete(e);}async has(e){return this.store.has(e)}};
-exports.AnthropicAPIProvider=W;exports.BaseProvider=C;exports.ElectronCredentialStore=ge;exports.EngineEventEmitter=J;exports.FallbackEngine=X;exports.GatewayLoader=N;exports.GatewayLoaderError=E;exports.GatewayUnavailableError=x;exports.IndexedDBCredentialStore=fe;exports.InferenceRelay=xe;exports.MODEL_ALIASES=$;exports.MODEL_REGISTRY=Y;exports.ManifestDecryptor=D;exports.MemoryCredentialStore=_e;exports.OllamaProvider=j;exports.OpenAIAPIProvider=q;exports.RELAY_FEATURES=ye;exports.RELAY_VERSION=Q;exports.RelayCapExceededError=K;exports.RelayError=U;exports.RequestTooLargeError=z;exports.TelemetryReporter=F;exports.applyRouting=ue;exports.createRelayStream=M;exports.createRoutingFunction=de;exports.detectCLI=V;exports.estimateCost=he;exports.estimateInputTokens=Xt;exports.findModel=Ct;exports.getLastKnownGoodManifest=ut;exports.getModelPricing=me;exports.getValue=dt;exports.modelsForProvider=St;exports.normalizeModel=Se;exports.providerForModel=oe;exports.relayHello=ke;exports.validateAnthropicKey=tt;exports.validateLicense=Te;