@relayplane/proxy 1.5.44 → 1.5.46

package/README.md

@@ -5,6 +5,13 @@
 
 An open-source LLM proxy that sits between your AI agents and providers. Tracks every request, shows where the money goes, and offers configurable task-aware routing — all running locally.
 
+**Key features:**
+- 📊 Per-request cost tracking across 11+ providers
+- 💰 **Cache-aware cost tracking** — accurately tracks Anthropic prompt caching with cache read savings, creation costs, and true per-request costs including cached tokens
+- 🔀 Configurable task-aware routing (complexity-based, cascade, model overrides)
+- 🛡️ Circuit breaker architecture — if the proxy fails, your agent doesn't notice
+- 📈 Local dashboard with cost breakdown, savings analysis, and provider health
+
 ## Quick Start
 
 ```bash

package/dist/standalone-proxy.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"standalone-proxy.d.ts","sourceRoot":"","sources":["../src/standalone-proxy.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAEH,OAAO,KAAK,IAAI,MAAM,WAAW,CAAC;AAKlC,OAAO,KAAK,EAAE,QAAQ,EAAY,MAAM,kBAAkB,CAAC;AAG3D,OAAO,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAU5C,2DAA2D;AAC3D,eAAO,MAAM,mBAAmB,gBAAuB,CAAC;AAExD;;GAEG;AACH,MAAM,WAAW,gBAAgB;IAC/B,OAAO,EAAE,MAAM,CAAC;IAChB,SAAS,EAAE,MAAM,CAAC;CACnB;AAED;;GAEG;AACH,eAAO,MAAM,iBAAiB,EAAE,MAAM,CAAC,MAAM,EAAE,gBAAgB,CA6C9D,CAAC;AAEF;;GAEG;AACH,eAAO,MAAM,aAAa,EAAE,MAAM,CAAC,MAAM,EAAE;IAAE,QAAQ,EAAE,QAAQ,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,CAc/E,CAAC;AAEF;;;GAGG;AACH,eAAO,MAAM,kBAAkB,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAGrD,CAAC;AAEF;;;GAGG;AACH,eAAO,MAAM,aAAa,EAAE,MAAM,CAAC,MAAM,EAAE;IAAE,QAAQ,EAAE,QAAQ,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,CAQ/E,CAAC;AAsCF;;GAEG;AACH,wBAAgB,sBAAsB,IAAI,MAAM,EAAE,CAWjD;AAED;;;GAGG;AACH,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,CAMvD;AAkBD,KAAK,aAAa,GAAG,MAAM,GAAG,MAAM,GAAG,SAAS,CAAC;AAEjD,UAAU,WAAW;IACnB,SAAS,EAAE,MAAM,CAAC;IAClB,MAAM,EAAE,aAAa,GAAG,IAAI,CAAC;CAC9B;AAcD,UAAU,aAAa;IACrB,OAAO,EAAE,OAAO,CAAC;IACjB,MAAM,EAAE,MAAM,EAAE,CAAC;IACjB,UAAU,EAAE,aAAa,GAAG,SAAS,GAAG,OAAO,CAAC;IAChD,cAAc,EAAE,MAAM,CAAC;CACxB;AAmBD,KAAK,UAAU,GAAG,QAAQ,GAAG,UAAU,GAAG,SAAS,CAAC;AA6EpD;;GAEG;AACH,MAAM,WAAW,WAAW;IAC1B,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,OAAO,CAAC,EAAE,OAAO,CAAC;IAClB;;;;;OAKG;IACH,aAAa,CAAC,EAAE,aAAa,GAAG,KAAK,GAAG,MAAM,CAAC;CAChD;AAuZD,wBAAgB,gBAAgB,CAAC,KAAK,EAAE,MAAM,GAAG,WAAW,CAe3D;AAuDD,wBAAgB,kBAAkB,CAAC,QAAQ,EAAE,KAAK,CAAC;IAAE,IAAI,CAAC,EAAE,MAAM,CAAC;IAAC,OAAO,CAAC,EAAE,OAAO,CAAA;CAAE,CAAC,GAAG,UAAU,CAuCpG;AAED,wBAAgB,cAAc,CAAC,YAAY,EAAE,MAAM,EAAE,OAAO,EAAE,aAAa,CAAC,YAAY,CAAC,GAAG,OAAO,CAIlG;AAuoDD;;GAEG;AACH,wBAAsB,UAAU,CAAC,MAAM,GAAE,WAAgB,GAAG,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CAu5C/E"}
+{"version":3,"file":"standalone-proxy.d.ts","sourceRoot":"","sources":["../src/standalone-proxy.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAEH,OAAO,KAAK,IAAI,MAAM,WAAW,CAAC;AAKlC,OAAO,KAAK,EAAE,QAAQ,EAAY,MAAM,kBAAkB,CAAC;AAG3D,OAAO,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAU5C,2DAA2D;AAC3D,eAAO,MAAM,mBAAmB,gBAAuB,CAAC;AAExD;;GAEG;AACH,MAAM,WAAW,gBAAgB;IAC/B,OAAO,EAAE,MAAM,CAAC;IAChB,SAAS,EAAE,MAAM,CAAC;CACnB;AAED;;GAEG;AACH,eAAO,MAAM,iBAAiB,EAAE,MAAM,CAAC,MAAM,EAAE,gBAAgB,CA6C9D,CAAC;AAEF;;GAEG;AACH,eAAO,MAAM,aAAa,EAAE,MAAM,CAAC,MAAM,EAAE;IAAE,QAAQ,EAAE,QAAQ,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,CAc/E,CAAC;AAEF;;;GAGG;AACH,eAAO,MAAM,kBAAkB,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAGrD,CAAC;AAEF;;;GAGG;AACH,eAAO,MAAM,aAAa,EAAE,MAAM,CAAC,MAAM,EAAE;IAAE,QAAQ,EAAE,QAAQ,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,CAQ/E,CAAC;AAiCF;;GAEG;AACH,wBAAgB,sBAAsB,IAAI,MAAM,EAAE,CAWjD;AAED;;;GAGG;AACH,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,CAMvD;AAkBD,KAAK,aAAa,GAAG,MAAM,GAAG,MAAM,GAAG,SAAS,CAAC;AAEjD,UAAU,WAAW;IACnB,SAAS,EAAE,MAAM,CAAC;IAClB,MAAM,EAAE,aAAa,GAAG,IAAI,CAAC;CAC9B;AAcD,UAAU,aAAa;IACrB,OAAO,EAAE,OAAO,CAAC;IACjB,MAAM,EAAE,MAAM,EAAE,CAAC;IACjB,UAAU,EAAE,aAAa,GAAG,SAAS,GAAG,OAAO,CAAC;IAChD,cAAc,EAAE,MAAM,CAAC;CACxB;AAmBD,KAAK,UAAU,GAAG,QAAQ,GAAG,UAAU,GAAG,SAAS,CAAC;AA6EpD;;GAEG;AACH,MAAM,WAAW,WAAW;IAC1B,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,OAAO,CAAC,EAAE,OAAO,CAAC;IAClB;;;;;OAKG;IACH,aAAa,CAAC,EAAE,aAAa,GAAG,KAAK,GAAG,MAAM,CAAC;CAChD;AAuZD,wBAAgB,gBAAgB,CAAC,KAAK,EAAE,MAAM,GAAG,WAAW,CAe3D;AAuDD,wBAAgB,kBAAkB,CAAC,QAAQ,EAAE,KAAK,CAAC;IAAE,IAAI,CAAC,EAAE,MAAM,CAAC;IAAC,OAAO,CAAC,EAAE,OAAO,CAAA;CAAE,CAAC,GAAG,UAAU,CAuCpG;AAED,wBAAgB,cAAc,CAAC,YAAY,EAAE,MAAM,EAAE,OAAO,EAAE,aAAa,CAAC,YAAY,CAAC,GAAG,OAAO,CAIlG;AAulDD;;GAEG;AACH,wBAAsB,UAAU,CAAC,MAAM,GAAE,WAAgB,GAAG,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CA2xC/E"}

package/dist/standalone-proxy.js

@@ -133,14 +133,14 @@ exports.DEFAULT_ENDPOINTS = {
  */
 exports.MODEL_MAPPING = {
     // Anthropic models (using correct API model IDs)
-    'claude-opus-4-5': { provider: 'anthropic', model: 'claude-opus-4-20250514' },
-    'claude-sonnet-4': { provider: 'anthropic', model: 'claude-sonnet-4-20250514' },
-    'claude-3-5-sonnet': { provider: 'anthropic', model: 'claude-3-5-sonnet-20241022' },
-    'claude-3-5-haiku': { provider: 'anthropic', model: 'claude-3-5-haiku-20241022' },
-    'claude-haiku-4-5': { provider: 'anthropic', model: 'claude-haiku-4-5-20250514' },
-    haiku: { provider: 'anthropic', model: 'claude-haiku-4-5-20250514' },
-    sonnet: { provider: 'anthropic', model: 'claude-sonnet-4-20250514' },
-    opus: { provider: 'anthropic', model: 'claude-opus-4-20250514' },
+    'claude-opus-4-5': { provider: 'anthropic', model: 'claude-opus-4-6' },
+    'claude-sonnet-4': { provider: 'anthropic', model: 'claude-sonnet-4-6' },
+    'claude-3-5-sonnet': { provider: 'anthropic', model: 'claude-3-5-sonnet-latest' },
+    'claude-3-5-haiku': { provider: 'anthropic', model: 'claude-haiku-4-5' },
+    'claude-haiku-4-5': { provider: 'anthropic', model: 'claude-haiku-4-5' },
+    haiku: { provider: 'anthropic', model: 'claude-haiku-4-5' },
+    sonnet: { provider: 'anthropic', model: 'claude-sonnet-4-6' },
+    opus: { provider: 'anthropic', model: 'claude-opus-4-6' },
     // OpenAI models
     'gpt-4o': { provider: 'openai', model: 'gpt-4o' },
     'gpt-4o-mini': { provider: 'openai', model: 'gpt-4o-mini' },
@@ -171,10 +171,10 @@ exports.SMART_ALIASES = {
  * Send a telemetry event to the cloud (anonymous or authenticated).
  * Non-blocking — errors are silently swallowed.
  */
-function sendCloudTelemetry(taskType, model, tokensIn, tokensOut, latencyMs, success, costUsd, requestedModel, cacheCreationTokens, cacheReadTokens) {
+function sendCloudTelemetry(taskType, model, tokensIn, tokensOut, latencyMs, success, costUsd, requestedModel) {
     try {
-        const cost = costUsd ?? (0, telemetry_js_1.estimateCost)(model, tokensIn, tokensOut, cacheCreationTokens, cacheReadTokens);
-        const event = {
+        const cost = costUsd ?? (0, telemetry_js_1.estimateCost)(model, tokensIn, tokensOut);
+        (0, telemetry_js_1.recordTelemetry)({
             task_type: taskType,
             model,
             tokens_in: tokensIn,
@@ -183,12 +183,7 @@ function sendCloudTelemetry(taskType, model, tokensIn, tokensOut, latencyMs, suc
             success,
             cost_usd: cost,
             requested_model: requestedModel,
-        };
-        if (cacheCreationTokens)
-            event.cache_creation_tokens = cacheCreationTokens;
-        if (cacheReadTokens)
-            event.cache_read_tokens = cacheReadTokens;
-        (0, telemetry_js_1.recordTelemetry)(event);
+        });
     }
     catch {
         // Telemetry should never break the proxy
@@ -494,18 +489,18 @@ const DEFAULT_PROXY_CONFIG = {
         cascade: {
             enabled: true,
             models: [
-                'claude-3-5-haiku-20241022',
-                'claude-sonnet-4-20250514',
-                'claude-opus-4-20250514',
+                'claude-haiku-4-5',
+                'claude-sonnet-4-6',
+                'claude-opus-4-6',
             ],
             escalateOn: 'uncertainty',
             maxEscalations: 1,
         },
         complexity: {
             enabled: true,
-            simple: 'claude-3-5-haiku-20241022',
-            moderate: 'claude-sonnet-4-20250514',
-            complex: 'claude-opus-4-20250514',
+            simple: 'claude-haiku-4-5',
+            moderate: 'claude-sonnet-4-6',
+            complex: 'claude-opus-4-6',
         },
     },
     reliability: {
@@ -746,24 +741,29 @@ function getAuthForModel(model, authConfig, envApiKey) {
  * MAX tokens (sk-ant-oat*) use Authorization: Bearer header
  * API keys (sk-ant-api*) use x-api-key header
  */
-function buildAnthropicHeadersWithAuth(ctx, apiKey, isMaxToken, isRerouted) {
+function buildAnthropicHeadersWithAuth(ctx, apiKey, isMaxToken) {
     const headers = {
         'Content-Type': 'application/json',
         'anthropic-version': ctx.versionHeader || '2023-06-01',
     };
-    // Detect if incoming auth is OAuth
-    const incomingIsOAuth = ctx.apiKeyHeader?.startsWith('sk-ant-oat') || ctx.authHeader?.includes('sk-ant-oat');
-    const apiKeyIsRegular = apiKey && apiKey.startsWith('sk-ant-api');
-    // When rerouted (auto mode changed the model) and incoming is OAuth,
-    // prefer the regular API key — OAuth doesn't work for all models (e.g. Haiku)
-    if (isRerouted && incomingIsOAuth && apiKeyIsRegular) {
+    // Auth: prefer incoming auth for passthrough, but OAuth doesn't work for all models (e.g. Haiku)
+    // When we have a regular API key AND incoming auth is OAuth, prefer the API key for rerouted requests
+    // because OAuth may not be supported on the target model. The API key works for ALL models.
+    const incomingIsOAuth = !!(ctx.apiKeyHeader?.startsWith('sk-ant-oat') || ctx.authHeader?.includes('sk-ant-oat'));
+    if (incomingIsOAuth && apiKey && !apiKey.startsWith('sk-ant-oat')) {
         headers['x-api-key'] = apiKey;
     }
     else if (ctx.authHeader) {
         headers['Authorization'] = ctx.authHeader;
     }
     else if (ctx.apiKeyHeader) {
-        headers['x-api-key'] = ctx.apiKeyHeader;
+        // MAX/OAuth tokens (sk-ant-oat*) must use Authorization: Bearer, not x-api-key
+        if (ctx.apiKeyHeader.startsWith('sk-ant-oat')) {
+            headers['Authorization'] = `Bearer ${ctx.apiKeyHeader}`;
+        }
+        else {
+            headers['x-api-key'] = ctx.apiKeyHeader;
+        }
     }
     else if (apiKey) {
         // MAX tokens (OAuth) use Authorization: Bearer, API keys use x-api-key
@@ -850,8 +850,8 @@ async function forwardToAnthropicStream(request, targetModel, ctx, envApiKey) {
  * Forward native Anthropic /v1/messages request (passthrough with routing)
  * Used for Claude Code direct integration
  */
-async function forwardNativeAnthropicRequest(body, ctx, envApiKey, isMaxToken, isRerouted) {
-    const headers = buildAnthropicHeadersWithAuth(ctx, envApiKey, isMaxToken, isRerouted);
+async function forwardNativeAnthropicRequest(body, ctx, envApiKey, isMaxToken) {
+    const headers = buildAnthropicHeadersWithAuth(ctx, envApiKey, isMaxToken);
     const response = await fetch('https://api.anthropic.com/v1/messages', {
         method: 'POST',
         headers,
@@ -1412,11 +1412,9 @@ function convertAnthropicResponse(anthropicData) {
             },
         ],
         usage: {
-            prompt_tokens: (anthropicData.usage?.input_tokens ?? 0) + (anthropicData.usage?.cache_creation_input_tokens ?? 0) + (anthropicData.usage?.cache_read_input_tokens ?? 0),
+            prompt_tokens: anthropicData.usage?.input_tokens ?? 0,
             completion_tokens: anthropicData.usage?.output_tokens ?? 0,
-            total_tokens: (anthropicData.usage?.input_tokens ?? 0) + (anthropicData.usage?.cache_creation_input_tokens ?? 0) + (anthropicData.usage?.cache_read_input_tokens ?? 0) + (anthropicData.usage?.output_tokens ?? 0),
-            cache_creation_input_tokens: anthropicData.usage?.cache_creation_input_tokens ?? 0,
-            cache_read_input_tokens: anthropicData.usage?.cache_read_input_tokens ?? 0,
+            total_tokens: (anthropicData.usage?.input_tokens ?? 0) + (anthropicData.usage?.output_tokens ?? 0),
         },
     };
 }
@@ -1439,16 +1437,11 @@ function convertAnthropicStreamEvent(eventType, eventData, messageId, model, too
             const msg = eventData['message'];
             baseChunk.id = msg?.['id'] || messageId;
             choice.delta = { role: 'assistant', content: '' };
-            // Pass through input token count from message_start (including cache tokens)
+            // Pass through input token count from message_start
             const msgUsage = msg?.['usage'];
             if (msgUsage) {
-                const cacheCreation = msgUsage['cache_creation_input_tokens'] ?? 0;
-                const cacheRead = msgUsage['cache_read_input_tokens'] ?? 0;
-                const inputTokens = msgUsage['input_tokens'] ?? 0;
                 baseChunk['usage'] = {
-                    prompt_tokens: inputTokens + cacheCreation + cacheRead,
-                    cache_creation_input_tokens: cacheCreation,
-                    cache_read_input_tokens: cacheRead,
+                    prompt_tokens: msgUsage['input_tokens'] ?? 0,
                 };
             }
             return `data: ${JSON.stringify(baseChunk)}\n\n`;
@@ -1819,7 +1812,7 @@ function getCascadeConfig(config) {
     const c = config.routing?.cascade;
     return {
         enabled: c?.enabled ?? true,
-        models: c?.models ?? ['claude-3-5-haiku-20241022', 'claude-sonnet-4-20250514', 'claude-opus-4-20250514'],
+        models: c?.models ?? ['claude-haiku-4-5', 'claude-sonnet-4-6', 'claude-opus-4-6'],
         escalateOn: c?.escalateOn ?? 'uncertainty',
         maxEscalations: c?.maxEscalations ?? 1,
     };
@@ -1836,18 +1829,18 @@ function getCooldownConfig(config) {
 function getCostModel(config) {
     return (config.routing?.complexity?.simple ||
         config.routing?.cascade?.models?.[0] ||
-        'claude-3-5-haiku-20241022');
+        'claude-haiku-4-5');
 }
 function getFastModel(config) {
     return (config.routing?.complexity?.simple ||
         config.routing?.cascade?.models?.[0] ||
-        'claude-3-5-haiku-20241022');
+        'claude-haiku-4-5');
 }
 function getQualityModel(config) {
     return (config.routing?.complexity?.complex ||
         config.routing?.cascade?.models?.[config.routing?.cascade?.models?.length ? config.routing.cascade.models.length - 1 : 0] ||
         process.env['RELAYPLANE_QUALITY_MODEL'] ||
-        'claude-sonnet-4-20250514');
+        'claude-sonnet-4-6');
 }
 async function cascadeRequest(config, makeRequest, log) {
     let escalations = 0;
@@ -1909,18 +1902,9 @@ td{padding:8px 12px;border-bottom:1px solid #111318}
 <div class="cards">
   <div class="card"><div class="label">Total Requests</div><div class="value" id="totalReq">—</div></div>
   <div class="card"><div class="label">Total Cost</div><div class="value" id="totalCost">—</div></div>
-  <div class="card"><div class="label">Savings (vs Opus)</div><div class="value green" id="savings">—</div></div>
+  <div class="card"><div class="label">Savings</div><div class="value green" id="savings">—</div></div>
   <div class="card"><div class="label">Avg Latency</div><div class="value" id="avgLat">—</div></div>
 </div>
-<div class="section"><h2>Auth & Routing</h2>
-<div id="routingDetails" style="display:flex;gap:16px;flex-wrap:wrap;margin-bottom:12px">
-  <div class="prov-item"><span class="dot" id="authDot"></span> <strong>Auth:</strong>&nbsp;<span id="authInfo">—</span></div>
-  <div class="prov-item"><strong>Routing:</strong>&nbsp;<span id="routingMode">—</span></div>
-  <div class="prov-item"><strong>Simple→</strong>&nbsp;<span id="routeSimple">—</span></div>
-  <div class="prov-item"><strong>Moderate→</strong>&nbsp;<span id="routeModerate">—</span></div>
-  <div class="prov-item"><strong>Complex→</strong>&nbsp;<span id="routeComplex">—</span></div>
-</div>
-</div>
 <div class="section"><h2>Model Breakdown</h2>
 <table><thead><tr><th>Model</th><th>Requests</th><th>Cost</th><th>% of Total</th></tr></thead><tbody id="models"></tbody></table></div>
 <div class="section"><h2>Provider Status</h2><div class="prov" id="providers"></div></div>
@@ -1941,39 +1925,6 @@ async function load(){
       fetch('/v1/telemetry/health').then(r=>r.json())
     ]);
     $('ver').textContent='v'+health.version;
-    const authDot=$('authDot'),authInfo=$('authInfo');
-    if(health.auth){
-      if(health.auth.anthropicApiKey){authDot.className='dot up';authInfo.textContent='API key ('+health.auth.anthropicApiKeyPrefix+')';}
-      else{authDot.className='dot warn';authInfo.textContent='OAuth only (no API key)';}
-    }
-    if(health.routing){
-      const mode=health.routing.mode||'passthrough';
-      $('routingMode').textContent=mode;
-      const routingSection=document.getElementById('routingDetails');
-      const hasApiKey=health.auth&&health.auth.anthropicApiKey;
-      if(mode==='passthrough'){
-        if(routingSection)routingSection.innerHTML='<div class="prov-item">Routing: passthrough → model from incoming requests</div>';
-      }else{
-        if(health.routing.complexity){
-          const cx=health.routing.complexity;
-          const authLabel=function(model){
-            if(hasApiKey)return '<span style="color:#34d399">● API key</span>';
-            const isHaiku=model&&model.toLowerCase().includes('haiku');
-            if(isHaiku)return '<span style="color:#ef4444">⚠️ OAuth - may fail</span>';
-            return '<span style="color:#fbbf24">● OAuth</span>';
-          };
-          $('routeSimple').innerHTML=(cx.simple||'—')+' <small>'+authLabel(cx.simple)+'</small>';
-          $('routeModerate').innerHTML=(cx.moderate||'—')+' <small>'+authLabel(cx.moderate)+'</small>';
-          $('routeComplex').innerHTML=(cx.complex||'—')+' <small>'+authLabel(cx.complex)+'</small>';
-          if(!hasApiKey&&cx.simple&&cx.simple.toLowerCase().includes('haiku')){
-            const warn=document.createElement('div');
-            warn.className='prov-item';warn.style.borderColor='#ef4444';warn.style.color='#ef4444';
-            warn.innerHTML='⚠️ Haiku requires ANTHROPIC_API_KEY — OAuth not supported';
-            if(routingSection)routingSection.appendChild(warn);
-          }
-        }
-      }
-    }
     $('uptime').textContent=dur(health.uptime);
     const total=stats.summary?.totalEvents||0;
     $('totalReq').textContent=total;
@@ -2102,60 +2053,6 @@ async function startProxy(config = {}) {
     process.on('SIGTERM', handleShutdown);
     const configPath = getProxyConfigPath();
     let proxyConfig = await loadProxyConfig(configPath, log);
-    // Auto-config on startup: detect available auth and set optimal routing
-    const configExists = await fs.promises.access(configPath).then(() => true).catch(() => false);
-    if (!configExists || proxyConfig.routing?.mode === 'auto') {
-        const envAnthropicKey = process.env['ANTHROPIC_API_KEY'];
-        const hasRegularApiKey = !!envAnthropicKey && envAnthropicKey.startsWith('sk-ant-api');
-        if (hasRegularApiKey) {
-            // Full 3-tier routing with API key
-            log('Auto-config: ANTHROPIC_API_KEY detected — enabling 3-tier routing (haiku/sonnet/opus)');
-            if (!configExists) {
-                const autoConfig = {
-                    enabled: true,
-                    modelOverrides: {},
-                    routing: {
-                        mode: 'auto',
-                        cascade: { enabled: false, models: [], escalateOn: 'uncertainty', maxEscalations: 1 },
-                        complexity: {
-                            enabled: true,
-                            simple: 'claude-haiku-4-5',
-                            moderate: 'claude-sonnet-4-6',
-                            complex: 'claude-opus-4-6',
-                        },
-                    },
-                    reliability: proxyConfig.reliability,
-                };
-                await saveProxyConfig(configPath, autoConfig);
-                proxyConfig = await loadProxyConfig(configPath, log);
-                log('Auto-config: wrote config to ' + configPath);
-            }
-        }
-        else {
-            // No regular API key — OAuth only, skip Haiku
-            console.warn('[relayplane] ⚠️  No ANTHROPIC_API_KEY set — Haiku routing disabled (OAuth not supported). Set ANTHROPIC_API_KEY to enable 3-tier routing.');
-            if (!configExists) {
-                const autoConfig = {
-                    enabled: true,
-                    modelOverrides: {},
-                    routing: {
-                        mode: 'auto',
-                        cascade: { enabled: false, models: [], escalateOn: 'uncertainty', maxEscalations: 1 },
-                        complexity: {
-                            enabled: true,
-                            simple: 'claude-sonnet-4-6',
-                            moderate: 'claude-sonnet-4-6',
-                            complex: 'claude-opus-4-6',
-                        },
-                    },
-                    reliability: proxyConfig.reliability,
-                };
-                await saveProxyConfig(configPath, autoConfig);
-                proxyConfig = await loadProxyConfig(configPath, log);
-                log('Auto-config: wrote OAuth-safe config to ' + configPath + ' (no Haiku)');
-            }
-        }
-    }
     const cooldownManager = new CooldownManager(getCooldownConfig(proxyConfig));
     let configWatcher = null;
     let configReloadTimer = null;
@@ -2171,8 +2068,6 @@ async function startProxy(config = {}) {
             reloadConfig().catch(() => { });
         }, 50);
     };
-    let credentialsWatcher = null;
-    const credentialsPath = path.join(path.dirname(configPath), 'credentials.json');
     const startConfigWatcher = () => {
         if (configWatcher)
             return;
@@ -2184,45 +2079,7 @@ async function startProxy(config = {}) {
             log(`Config watch error: ${error.message}`);
         }
     };
-    const startCredentialsWatcher = () => {
-        if (credentialsWatcher)
-            return;
-        try {
-            // Watch credentials.json so login triggers a reload without proxy restart
-            if (fs.existsSync(credentialsPath)) {
-                credentialsWatcher = fs.watch(credentialsPath, () => {
-                    log('Credentials changed — reloading config');
-                    scheduleConfigReload();
-                });
-            }
-            else {
-                // Watch the directory for credentials.json creation
-                const dir = path.dirname(credentialsPath);
-                const dirWatcher = fs.watch(dir, (_, filename) => {
-                    if (filename === 'credentials.json') {
-                        log('Credentials file created — reloading config');
-                        scheduleConfigReload();
-                        // Now watch the file directly
-                        dirWatcher.close();
-                        try {
-                            credentialsWatcher = fs.watch(credentialsPath, () => {
-                                log('Credentials changed — reloading config');
-                                scheduleConfigReload();
-                            });
-                        }
-                        catch { }
-                    }
-                });
-                credentialsWatcher = dirWatcher;
-            }
-        }
-        catch (err) {
-            const error = err;
-            log(`Credentials watch error: ${error.message}`);
-        }
-    };
     startConfigWatcher();
-    startCredentialsWatcher();
     // Initialize RelayPlane
     const relay = new core_1.RelayPlane({ dbPath: config.dbPath });
     // Startup migration: clear default routing rules so complexity config takes priority
@@ -2253,10 +2110,6 @@ async function startProxy(config = {}) {
         if (req.method === 'GET' && (pathname === '/health' || pathname === '/healthz')) {
             const uptimeMs = Date.now() - globalStats.startedAt;
             res.writeHead(200, { 'Content-Type': 'application/json' });
-            const anthropicEnvKeySet = !!process.env['ANTHROPIC_API_KEY'];
-            const anthropicEnvKeyPrefix = anthropicEnvKeySet ? process.env['ANTHROPIC_API_KEY'].slice(0, 12) + '...' : null;
-            const routingMode = proxyConfig.routing?.mode || 'passthrough';
-            const complexityConfig = proxyConfig.routing?.complexity;
             res.end(JSON.stringify({
                 status: 'ok',
                 version: PROXY_VERSION,
@@ -2266,15 +2119,6 @@ async function startProxy(config = {}) {
                 successRate: globalStats.totalRequests > 0
                     ? parseFloat(((globalStats.successfulRequests / globalStats.totalRequests) * 100).toFixed(1))
                     : null,
-                auth: {
-                    anthropicApiKey: anthropicEnvKeySet,
-                    anthropicApiKeyPrefix: anthropicEnvKeyPrefix,
-                    note: anthropicEnvKeySet ? 'API key available for models that don\'t support OAuth' : 'No API key — OAuth passthrough only',
-                },
-                routing: {
-                    mode: routingMode,
-                    complexity: complexityConfig,
-                },
                 stats: {
                     totalRequests: globalStats.totalRequests,
                     successfulRequests: globalStats.successfulRequests,
@@ -2400,7 +2244,7 @@ async function startProxy(config = {}) {
                 const offset = parseInt(params.get('offset') || '0', 10);
                 const sorted = [...requestHistory].reverse();
                 const runs = sorted.slice(offset, offset + limit).map(r => {
-                    const origCost = (0, telemetry_js_1.estimateCost)('claude-opus-4-20250514', r.tokensIn, r.tokensOut);
+                    const origCost = (0, telemetry_js_1.estimateCost)('claude-opus-4-6', r.tokensIn, r.tokensOut);
                     const perRunSavings = Math.max(0, origCost - r.costUsd);
                     return {
                         id: r.id,
@@ -2431,7 +2275,7 @@ async function startProxy(config = {}) {
             if (req.method === 'GET' && telemetryPath === 'savings') {
                 // Savings = cost if everything ran on Opus - actual cost
                 // Always compare against Opus as the baseline
-                const OPUS_BASELINE = 'claude-opus-4-20250514';
+                const OPUS_BASELINE = 'claude-opus-4-6';
                 let totalOriginalCost = 0;
                 let totalActualCost = 0;
                 let totalSavedAmount = 0;
@@ -2556,7 +2400,8 @@ async function startProxy(config = {}) {
             useAnthropicEnvKey = undefined; // Only use incoming auth
         }
         else {
-            // 'auto': Always keep env key available — needed for OAuth→API key fallback on rerouted requests
+            // 'auto': Use incoming auth if present, fallback to env
+            // ALWAYS keep env key available — OAuth (sk-ant-oat) doesn't work for all models (e.g. Haiku)
             useAnthropicEnvKey = anthropicEnvKey;
         }
         // === Native Anthropic /v1/messages endpoint (for Claude Code) ===
@@ -2795,8 +2640,7 @@ async function startProxy(config = {}) {
                         if (modelAuth.isMax) {
                             log(`Using MAX token for ${resolved.model}`);
                         }
-                        const isCascadeRerouted = resolved.model !== originalModel;
-                        const providerResponse = await forwardNativeAnthropicRequest(attemptBody, ctx, modelAuth.apiKey, modelAuth.isMax, isCascadeRerouted);
+                        const providerResponse = await forwardNativeAnthropicRequest(attemptBody, ctx, modelAuth.apiKey, modelAuth.isMax);
                         const responseData = (await providerResponse.json());
                         if (!providerResponse.ok) {
                             if (proxyConfig.reliability?.cooldowns?.enabled) {
@@ -2823,12 +2667,7 @@ async function startProxy(config = {}) {
                     if (modelAuth.isMax) {
                         log(`Using MAX token for ${finalModel}`);
                     }
-                    // isRerouted: true when auto-routing changed the model from what the user requested
-                    const isRerouted = routingMode !== 'passthrough' && finalModel !== originalModel;
-                    if (isRerouted) {
-                        log(`Rerouted: ${originalModel} → ${finalModel} (auth fallback enabled)`);
-                    }
-                    const providerResponse = await forwardNativeAnthropicRequest({ ...requestBody, model: finalModel }, ctx, modelAuth.apiKey, modelAuth.isMax, isRerouted);
+                    const providerResponse = await forwardNativeAnthropicRequest({ ...requestBody, model: finalModel }, ctx, modelAuth.apiKey, modelAuth.isMax);
                     if (!providerResponse.ok) {
                         const errorPayload = (await providerResponse.json());
                         if (proxyConfig.reliability?.cooldowns?.enabled) {
@@ -2854,8 +2693,6 @@ async function startProxy(config = {}) {
                         const reader = providerResponse.body?.getReader();
                         let streamTokensIn = 0;
                         let streamTokensOut = 0;
-                        let streamCacheCreation = 0;
-                        let streamCacheRead = 0;
                         if (reader) {
                             const decoder = new TextDecoder();
                             let sseBuffer = '';
@@ -2878,12 +2715,9 @@ async function startProxy(config = {}) {
                                                 if (evt.type === 'message_delta' && evt.usage) {
                                                     streamTokensOut = evt.usage.output_tokens ?? streamTokensOut;
                                                 }
-                                                // Anthropic: message_start has usage.input_tokens + cache token fields
+                                                // Anthropic: message_start has usage.input_tokens
                                                 if (evt.type === 'message_start' && evt.message?.usage) {
-                                                    const u = evt.message.usage;
-                                                    streamCacheCreation = u.cache_creation_input_tokens ?? 0;
-                                                    streamCacheRead = u.cache_read_input_tokens ?? 0;
-                                                    streamTokensIn = (u.input_tokens ?? 0) + streamCacheCreation + streamCacheRead;
+                                                    streamTokensIn = evt.message.usage.input_tokens ?? streamTokensIn;
                                                 }
                                                 // OpenAI format: choices with usage
                                                 if (evt.usage) {
@@ -2903,7 +2737,7 @@ async function startProxy(config = {}) {
                             }
                         }
                         // Store streaming token counts so telemetry can use them
-                        nativeResponseData = { usage: { input_tokens: streamTokensIn, output_tokens: streamTokensOut, cache_creation_input_tokens: streamCacheCreation, cache_read_input_tokens: streamCacheRead } };
+                        nativeResponseData = { usage: { input_tokens: streamTokensIn, output_tokens: streamTokensOut } };
                         res.end();
                     }
                     else {
@@ -2920,12 +2754,9 @@ async function startProxy(config = {}) {
                 // nativeResponseData holds response JSON for non-streaming, or { usage: { input_tokens, output_tokens } }
                 // synthesised from SSE events for streaming
                 const nativeUsageData = nativeResponseData?.usage;
-                const nativeCacheCreation = nativeUsageData?.cache_creation_input_tokens ?? 0;
-                const nativeCacheRead = nativeUsageData?.cache_read_input_tokens ?? 0;
-                const nativeRawIn = nativeUsageData?.input_tokens ?? nativeUsageData?.prompt_tokens ?? 0;
-                const nativeTokIn = nativeRawIn + nativeCacheCreation + nativeCacheRead;
+                const nativeTokIn = nativeUsageData?.input_tokens ?? nativeUsageData?.prompt_tokens ?? 0;
                 const nativeTokOut = nativeUsageData?.output_tokens ?? nativeUsageData?.completion_tokens ?? 0;
-                updateLastHistoryEntry(nativeTokIn, nativeTokOut, (0, telemetry_js_1.estimateCost)(targetModel || requestedModel, nativeRawIn, nativeTokOut, nativeCacheCreation, nativeCacheRead));
+                updateLastHistoryEntry(nativeTokIn, nativeTokOut, (0, telemetry_js_1.estimateCost)(targetModel || requestedModel, nativeTokIn, nativeTokOut));
                 if (recordTelemetry) {
                     relay
                         .run({
@@ -2934,7 +2765,7 @@ async function startProxy(config = {}) {
                         model: `${targetProvider}:${targetModel || requestedModel}`,
                     })
                         .catch(() => { });
-                    sendCloudTelemetry(taskType, targetModel || requestedModel, nativeTokIn, nativeTokOut, durationMs, true, undefined, originalModel ?? undefined, nativeCacheCreation, nativeCacheRead);
+                    sendCloudTelemetry(taskType, targetModel || requestedModel, nativeTokIn, nativeTokOut, durationMs, true, undefined, originalModel ?? undefined);
                 }
             }
             catch (err) {
@@ -3272,12 +3103,9 @@ async function startProxy(config = {}) {
                     // Log cascade request for stats tracking
                     logRequest(originalRequestedModel ?? 'unknown', cascadeResult.model, cascadeResult.provider, durationMs, true, 'cascade', cascadeResult.escalations > 0, taskType, complexity);
                     const cascadeUsage = responseData?.usage;
-                    const cascadeCacheCreation = cascadeUsage?.cache_creation_input_tokens ?? 0;
-                    const cascadeCacheRead = cascadeUsage?.cache_read_input_tokens ?? 0;
-                    const cascadeRawIn = cascadeUsage?.input_tokens ?? cascadeUsage?.prompt_tokens ?? 0;
-                    const cascadeTokensIn = cascadeRawIn + cascadeCacheCreation + cascadeCacheRead;
+                    const cascadeTokensIn = cascadeUsage?.input_tokens ?? cascadeUsage?.prompt_tokens ?? 0;
                     const cascadeTokensOut = cascadeUsage?.output_tokens ?? cascadeUsage?.completion_tokens ?? 0;
-                    const cascadeCost = (0, telemetry_js_1.estimateCost)(cascadeResult.model, cascadeRawIn, cascadeTokensOut, cascadeCacheCreation, cascadeCacheRead);
+                    const cascadeCost = (0, telemetry_js_1.estimateCost)(cascadeResult.model, cascadeTokensIn, cascadeTokensOut);
                     updateLastHistoryEntry(cascadeTokensIn, cascadeTokensOut, cascadeCost, chatCascadeRespModel);
                     if (recordTelemetry) {
                         try {
@@ -3300,7 +3128,7 @@ async function startProxy(config = {}) {
                         catch (err) {
                             log(`Failed to record run: ${err}`);
                         }
-                        sendCloudTelemetry(taskType, cascadeResult.model, cascadeTokensIn, cascadeTokensOut, durationMs, true, undefined, originalRequestedModel ?? undefined, cascadeCacheCreation, cascadeCacheRead);
+                        sendCloudTelemetry(taskType, cascadeResult.model, cascadeTokensIn, cascadeTokensOut, durationMs, true, undefined, originalRequestedModel ?? undefined);
                     }
                     const chatCascadeRpHeaders = buildRelayPlaneResponseHeaders(cascadeResult.model, originalRequestedModel ?? 'unknown', complexity, cascadeResult.provider, 'cascade');
                     res.writeHead(200, { 'Content-Type': 'application/json', ...chatCascadeRpHeaders });
@@ -3449,8 +3277,6 @@ async function handleStreamingRequest(res, request, targetProvider, targetModel,
     // Track token usage from streaming events
     let streamTokensIn = 0;
     let streamTokensOut = 0;
-    let streamCacheCreation = 0;
-    let streamCacheRead = 0;
     try {
         // Stream the response based on provider format
         switch (targetProvider) {
@@ -3467,8 +3293,6 @@ async function handleStreamingRequest(res, request, targetProvider, targetModel,
                                 if (evt.usage) {
                                     streamTokensIn = evt.usage.prompt_tokens ?? streamTokensIn;
                                     streamTokensOut = evt.usage.completion_tokens ?? streamTokensOut;
-                                    streamCacheCreation = evt.usage.cache_creation_input_tokens ?? streamCacheCreation;
-                                    streamCacheRead = evt.usage.cache_read_input_tokens ?? streamCacheRead;
                                 }
                             }
                         }
@@ -3525,9 +3349,7 @@ async function handleStreamingRequest(res, request, targetProvider, targetModel,
     // Always log the request for stats/telemetry tracking
     logRequest(request.model ?? 'unknown', targetModel, targetProvider, durationMs, true, routingMode, undefined, taskType, complexity);
     // Update token/cost info on the history entry
-    // For cost calculation with cache breakdown, pass raw input (total minus cache) separately
-    const streamRawIn = streamCacheCreation || streamCacheRead ? streamTokensIn - streamCacheCreation - streamCacheRead : streamTokensIn;
-    const streamCost = (0, telemetry_js_1.estimateCost)(targetModel, streamRawIn, streamTokensOut, streamCacheCreation || undefined, streamCacheRead || undefined);
+    const streamCost = (0, telemetry_js_1.estimateCost)(targetModel, streamTokensIn, streamTokensOut);
     updateLastHistoryEntry(streamTokensIn, streamTokensOut, streamCost);
     if (recordTelemetry) {
         // Record the run (non-blocking)
@@ -3543,7 +3365,7 @@ async function handleStreamingRequest(res, request, targetProvider, targetModel,
             .catch((err) => {
             log(`Failed to record run: ${err}`);
         });
-        sendCloudTelemetry(taskType, targetModel, streamTokensIn, streamTokensOut, durationMs, true, undefined, request.model ?? undefined, streamCacheCreation || undefined, streamCacheRead || undefined);
+        sendCloudTelemetry(taskType, targetModel, streamTokensIn, streamTokensOut, durationMs, true, undefined, request.model ?? undefined);
     }
     res.end();
 }
@@ -3587,12 +3409,9 @@ async function handleNonStreamingRequest(res, request, targetProvider, targetMod
     logRequest(request.model ?? 'unknown', targetModel, targetProvider, durationMs, true, routingMode, undefined, taskType, complexity);
     // Update token/cost info
     const usage = responseData?.usage;
-    const cacheCreation = usage?.cache_creation_input_tokens ?? 0;
-    const cacheRead = usage?.cache_read_input_tokens ?? 0;
-    const rawIn = usage?.input_tokens ?? usage?.prompt_tokens ?? 0;
-    const tokensIn = rawIn + cacheCreation + cacheRead;
+    const tokensIn = usage?.input_tokens ?? usage?.prompt_tokens ?? 0;
     const tokensOut = usage?.output_tokens ?? usage?.completion_tokens ?? 0;
-    const cost = (0, telemetry_js_1.estimateCost)(targetModel, rawIn, tokensOut, cacheCreation || undefined, cacheRead || undefined);
+    const cost = (0, telemetry_js_1.estimateCost)(targetModel, tokensIn, tokensOut);
     updateLastHistoryEntry(tokensIn, tokensOut, cost, nonStreamRespModel);
     if (recordTelemetry) {
         // Record the run in RelayPlane
@@ -3617,13 +3436,10 @@ async function handleNonStreamingRequest(res, request, targetProvider, targetMod
             log(`Failed to record run: ${err}`);
         }
         // Extract token counts from response if available (Anthropic/OpenAI format)
-        const usage2 = responseData?.usage;
-        const cc2 = usage2?.cache_creation_input_tokens ?? 0;
-        const cr2 = usage2?.cache_read_input_tokens ?? 0;
-        const rawIn2 = usage2?.input_tokens ?? usage2?.prompt_tokens ?? 0;
-        const tokensIn2 = rawIn2 + cc2 + cr2;
-        const tokensOut2 = usage2?.output_tokens ?? usage2?.completion_tokens ?? 0;
-        sendCloudTelemetry(taskType, targetModel, tokensIn2, tokensOut2, durationMs, true, undefined, undefined, cc2 || undefined, cr2 || undefined);
+        const usage = responseData?.usage;
+        const tokensIn = usage?.input_tokens ?? usage?.prompt_tokens ?? 0;
+        const tokensOut = usage?.output_tokens ?? usage?.completion_tokens ?? 0;
+        sendCloudTelemetry(taskType, targetModel, tokensIn, tokensOut, durationMs, true);
     }
     // Send response with RelayPlane routing headers
     const nonStreamRpHeaders = buildRelayPlaneResponseHeaders(targetModel, request.model ?? 'unknown', complexity, targetProvider, routingMode);