@relayplane/proxy 1.7.2 → 1.7.4

package/dist/standalone-proxy.js

@@ -56,6 +56,8 @@ Object.defineProperty(exports, "__esModule", { value: true });
 exports.SMART_ALIASES = exports.RELAYPLANE_ALIASES = exports.MODEL_MAPPING = exports.DEFAULT_ENDPOINTS = exports.proxyStatsCollector = void 0;
 exports.getAvailableModelNames = getAvailableModelNames;
 exports.resolveModelAlias = resolveModelAlias;
+exports.extractRequestContent = extractRequestContent;
+exports.extractResponseText = extractResponseText;
 exports.parseModelSuffix = parseModelSuffix;
 exports.classifyComplexity = classifyComplexity;
 exports.shouldEscalate = shouldEscalate;
@@ -76,6 +78,7 @@ const budget_js_1 = require("./budget.js");
 const anomaly_js_1 = require("./anomaly.js");
 const alerts_js_1 = require("./alerts.js");
 const downgrade_js_1 = require("./downgrade.js");
+const agent_tracker_js_1 = require("./agent-tracker.js");
 const version_status_js_1 = require("./utils/version-status.js");
 const PROXY_VERSION = (() => {
     try {
@@ -488,7 +491,7 @@ function shutdownHistory() {
     }
     flushHistoryBuffer();
 }
-function logRequest(originalModel, targetModel, provider, latencyMs, success, mode, escalated, taskType, complexity) {
+function logRequest(originalModel, targetModel, provider, latencyMs, success, mode, escalated, taskType, complexity, agentFingerprint, agentId, errorMessage, errorStatusCode) {
     const timestamp = new Date().toISOString();
     const status = success ? '✓' : '✗';
     const escalateTag = escalated ? ' [ESCALATED]' : '';
@@ -531,6 +534,10 @@ function logRequest(originalModel, targetModel, provider, latencyMs, success, mo
         costUsd: 0,
         taskType: taskType || 'general',
         complexity: complexity || 'simple',
+        agentFingerprint,
+        agentId,
+        error: errorMessage,
+        statusCode: errorStatusCode,
     };
     requestHistory.push(entry);
     if (requestHistory.length > MAX_HISTORY) {
@@ -539,7 +546,7 @@ function logRequest(originalModel, targetModel, provider, latencyMs, success, mo
     bufferHistoryEntry(entry);
 }
 /** Update the most recent history entry with token/cost info */
-function updateLastHistoryEntry(tokensIn, tokensOut, costUsd, responseModel, cacheCreationTokens, cacheReadTokens) {
+function updateLastHistoryEntry(tokensIn, tokensOut, costUsd, responseModel, cacheCreationTokens, cacheReadTokens, agentFingerprint, agentId, requestContent, errorMessage, errorStatusCode) {
     if (requestHistory.length > 0) {
         const last = requestHistory[requestHistory.length - 1];
         last.tokensIn = tokensIn;
@@ -552,8 +559,86 @@ function updateLastHistoryEntry(tokensIn, tokensOut, costUsd, responseModel, cac
             last.cacheCreationTokens = cacheCreationTokens;
         if (cacheReadTokens !== undefined)
             last.cacheReadTokens = cacheReadTokens;
+        if (agentFingerprint !== undefined)
+            last.agentFingerprint = agentFingerprint;
+        if (agentId !== undefined)
+            last.agentId = agentId;
+        if (requestContent)
+            last.requestContent = requestContent;
+        if (errorMessage !== undefined)
+            last.error = errorMessage;
+        if (errorStatusCode !== undefined)
+            last.statusCode = errorStatusCode;
     }
 }
+/**
+ * Extract request content for logging. Handles Anthropic and OpenAI formats.
+ */
+function extractRequestContent(body, isAnthropic) {
+    let systemPrompt = '';
+    let userMessage = '';
+    if (isAnthropic) {
+        if (typeof body.system === 'string') {
+            systemPrompt = body.system;
+        }
+        else if (Array.isArray(body.system)) {
+            systemPrompt = body.system
+                .map(p => p.type === 'text' ? (p.text ?? '') : (typeof p === 'string' ? String(p) : ''))
+                .join('');
+        }
+    }
+    else {
+        const sysmsgs = body.messages;
+        if (Array.isArray(sysmsgs)) {
+            for (const msg of sysmsgs) {
+                if (msg.role === 'system') {
+                    systemPrompt = typeof msg.content === 'string' ? msg.content : '';
+                    break;
+                }
+            }
+        }
+    }
+    const msgs = body.messages;
+    if (Array.isArray(msgs)) {
+        for (let i = msgs.length - 1; i >= 0; i--) {
+            if (msgs[i].role === 'user') {
+                const content = msgs[i].content;
+                if (typeof content === 'string') {
+                    userMessage = content;
+                }
+                else if (Array.isArray(content)) {
+                    userMessage = content
+                        .filter(p => p.type === 'text')
+                        .map(p => p.text ?? '')
+                        .join('');
+                }
+                break;
+            }
+        }
+    }
+    return {
+        systemPrompt: systemPrompt ? systemPrompt.slice(0, 200) : undefined,
+        userMessage: userMessage || undefined,
+    };
+}
+/**
+ * Extract assistant response text from response payload.
+ */
+function extractResponseText(responseData, isAnthropic) {
+    if (isAnthropic) {
+        const content = responseData.content;
+        if (Array.isArray(content)) {
+            return content.filter(p => p.type === 'text').map(p => p.text ?? '').join('');
+        }
+    }
+    else {
+        const choices = responseData.choices;
+        if (Array.isArray(choices) && choices[0]?.message?.content) {
+            return choices[0].message.content;
+        }
+    }
+    return '';
+}
 const DEFAULT_PROXY_CONFIG = {
     enabled: true,
     modelOverrides: {},
@@ -584,6 +669,11 @@ const DEFAULT_PROXY_CONFIG = {
         },
     },
 };
+/** Module-level ref to active proxy config (set during startProxy) */
+let _activeProxyConfig = {};
+function isContentLoggingEnabled() {
+    return _activeProxyConfig.dashboard?.showRequestContent !== false;
+}
 function getProxyConfigPath() {
     const customPath = process.env['RELAYPLANE_CONFIG_PATH'];
     if (customPath && customPath.trim())
@@ -835,19 +925,16 @@ function buildAnthropicHeadersWithAuth(ctx, apiKey, isMaxToken) {
         'Content-Type': 'application/json',
         'anthropic-version': ctx.versionHeader || '2023-06-01',
     };
-    // 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) {
+    // Auth: ALWAYS prefer incoming auth for passthrough (don't replace it)
+    // Incoming auth is from Claude Code/OpenClaw and is already the right token for the request
+    if (ctx.authHeader) {
+        // Incoming Authorization header takes priority - use it as-is
         headers['Authorization'] = ctx.authHeader;
     }
     else if (ctx.apiKeyHeader) {
-        // MAX/OAuth tokens (sk-ant-oat*) must use Authorization: Bearer, not x-api-key
+        // Incoming x-api-key header
         if (ctx.apiKeyHeader.startsWith('sk-ant-oat')) {
+            // MAX/OAuth tokens must use Authorization: Bearer, not x-api-key
             headers['Authorization'] = `Bearer ${ctx.apiKeyHeader}`;
         }
         else {
@@ -855,7 +942,7 @@ function buildAnthropicHeadersWithAuth(ctx, apiKey, isMaxToken) {
         }
     }
     else if (apiKey) {
-        // MAX tokens (OAuth) use Authorization: Bearer, API keys use x-api-key
+        // Fallback to configured API key (only if no incoming auth)
         if (isMaxToken || apiKey.startsWith('sk-ant-oat')) {
             headers['Authorization'] = `Bearer ${apiKey}`;
         }
@@ -867,6 +954,13 @@ function buildAnthropicHeadersWithAuth(ctx, apiKey, isMaxToken) {
     if (ctx.betaHeaders) {
         headers['anthropic-beta'] = ctx.betaHeaders;
     }
+    // Pass through OAuth identity headers (required by Anthropic for OAuth token validation)
+    if (ctx.userAgent) {
+        headers['user-agent'] = ctx.userAgent;
+    }
+    if (ctx.xApp) {
+        headers['x-app'] = ctx.xApp;
+    }
     return headers;
 }
 /**
@@ -907,6 +1001,13 @@ function buildAnthropicHeaders(ctx, envApiKey) {
     if (ctx.betaHeaders) {
         headers['anthropic-beta'] = ctx.betaHeaders;
     }
+    // Pass through OAuth identity headers (required by Anthropic for OAuth token validation)
+    if (ctx.userAgent) {
+        headers['user-agent'] = ctx.userAgent;
+    }
+    if (ctx.xApp) {
+        headers['x-app'] = ctx.xApp;
+    }
     return headers;
 }
 /**
@@ -1772,7 +1873,7 @@ function resolveExplicitModel(modelName) {
 function resolveConfigModel(modelName) {
     return resolveExplicitModel(modelName) ?? parsePreferredModel(modelName);
 }
-function extractResponseText(responseData) {
+function extractResponseTextAuto(responseData) {
     const openAiChoices = responseData['choices'];
     if (openAiChoices && openAiChoices.length > 0) {
         const first = openAiChoices[0];
@@ -1816,6 +1917,28 @@ function checkResponseModelMismatch(responseData, requestedModel, provider, log)
     }
     return responseModel;
 }
+/**
+ * Extract a human-readable error message from a provider error payload.
+ * Handles Anthropic ({ error: { type, message } }) and OpenAI ({ error: { message } }) formats.
+ */
+function extractProviderErrorMessage(payload, statusCode) {
+    const err = payload['error'];
+    if (typeof err === 'string')
+        return err;
+    if (err && typeof err === 'object') {
+        const errType = err['type'];
+        const errMsg = err['message'];
+        if (errType && errMsg)
+            return `${errType}: ${errMsg}`;
+        if (errMsg)
+            return errMsg;
+        if (errType)
+            return errType;
+    }
+    if (statusCode)
+        return `HTTP ${statusCode}`;
+    return 'Unknown error';
+}
 class ProviderResponseError extends Error {
     status;
     payload;
@@ -1841,6 +1964,8 @@ function extractRequestContext(req) {
         betaHeaders: req.headers['anthropic-beta'],
         versionHeader: req.headers['anthropic-version'],
         apiKeyHeader: req.headers['x-api-key'],
+        userAgent: req.headers['user-agent'],
+        xApp: req.headers['x-app'],
     };
 }
 const MAX_BODY_SIZE = 10 * 1024 * 1024; // 10MB max request body
@@ -1940,7 +2065,7 @@ async function cascadeRequest(config, makeRequest, log) {
         const isLastModel = i === config.models.length - 1;
         try {
             const { responseData, provider, model: resolvedModel } = await makeRequest(model);
-            const text = extractResponseText(responseData);
+            const text = extractResponseTextAuto(responseData);
             if (isLastModel || escalations >= config.maxEscalations) {
                 return { responseData, provider, model: resolvedModel, escalations };
             }
@@ -1969,7 +2094,7 @@ async function cascadeRequest(config, makeRequest, log) {
 function getDashboardHTML() {
     return `<!DOCTYPE html><html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>RelayPlane Dashboard</title>
 <style>
-*{margin:0;padding:0;box-sizing:border-box}body{background:#0a0b0d;color:#e2e8f0;font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',sans-serif;padding:20px;max-width:1200px;margin:0 auto}
+*{margin:0;padding:0;box-sizing:border-box}body{background:#0a0b0d;color:#e2e8f0;font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',sans-serif;padding:20px;max-width:1600px;margin:0 auto}
 a{color:#34d399}h1{font-size:1.5rem;font-weight:600}
 .header{display:flex;justify-content:space-between;align-items:center;padding:16px 0;border-bottom:1px solid #1e293b;margin-bottom:24px}
 .header .meta{font-size:.8rem;color:#64748b}
@@ -1977,13 +2102,18 @@ a{color:#34d399}h1{font-size:1.5rem;font-weight:600}
 .card{background:#111318;border:1px solid #1e293b;border-radius:12px;padding:20px}
 .card .label{font-size:.75rem;color:#64748b;text-transform:uppercase;letter-spacing:.05em;margin-bottom:6px}
 .card .value{font-size:1.75rem;font-weight:700}.green{color:#34d399}
+.tooltip-wrap{position:relative;display:inline-block}
+.tooltip-wrap .tooltip-box{visibility:hidden;opacity:0;background:#1e293b;color:#e2e8f0;font-size:.8rem;font-weight:400;text-transform:none;letter-spacing:0;line-height:1.5;border:1px solid #334155;border-radius:8px;padding:10px 14px;position:absolute;top:calc(100% + 8px);left:50%;transform:translateX(-50%);width:280px;z-index:999;pointer-events:none;transition:opacity .15s;box-shadow:0 4px 16px rgba(0,0,0,.4)}
+.tooltip-wrap .tooltip-box::after{content:'';position:absolute;bottom:100%;left:50%;transform:translateX(-50%);border:6px solid transparent;border-bottom-color:#334155}
+.tooltip-wrap:hover .tooltip-box{visibility:visible;opacity:1}
+.info-icon{cursor:help;color:#64748b;font-size:.75rem;vertical-align:middle;margin-left:4px}
 table{width:100%;border-collapse:collapse;font-size:.85rem}
 th{text-align:left;color:#64748b;font-weight:500;padding:8px 12px;border-bottom:1px solid #1e293b;font-size:.75rem;text-transform:uppercase;letter-spacing:.04em}
 td{padding:8px 12px;border-bottom:1px solid #111318}
 .section{margin-bottom:32px}.section h2{font-size:1rem;font-weight:600;margin-bottom:12px;color:#94a3b8}
 .dot{display:inline-block;width:8px;height:8px;border-radius:50%;margin-right:6px}.dot.up{background:#34d399}.dot.warn{background:#fbbf24}.dot.down{background:#ef4444}
 .badge{display:inline-block;padding:2px 8px;border-radius:6px;font-size:.75rem;font-weight:500}
-.badge.ok{background:#052e1633;color:#34d399}.badge.err{background:#2d0a0a;color:#ef4444}
+.badge.ok{background:#052e1633;color:#34d399}.badge.err{background:#2d0a0a;color:#ef4444}.badge.err-auth{background:#2d0a0a;color:#ef4444}.badge.err-rate{background:#2d2a0a;color:#fbbf24}.badge.err-timeout{background:#2d1a0a;color:#fb923c}
 .badge.tt-code{background:#1e3a5f;color:#60a5fa}.badge.tt-analysis{background:#3b1f6e;color:#a78bfa}.badge.tt-summarization{background:#1a3a2a;color:#6ee7b7}.badge.tt-qa{background:#3a2f1e;color:#fbbf24}.badge.tt-general{background:#1e293b;color:#94a3b8}
 .badge.cx-simple{background:#052e1633;color:#34d399}.badge.cx-moderate{background:#2d2a0a;color:#fbbf24}.badge.cx-complex{background:#2d0a0a;color:#ef4444}
 .vstat{display:inline-flex;align-items:center;gap:6px;margin-left:8px;padding:1px 8px;border-radius:999px;border:1px solid #334155;font-size:.72rem}
@@ -1992,19 +2122,22 @@ td{padding:8px 12px;border-bottom:1px solid #111318}
 .vstat.unavailable{color:#a3a3a3;border-color:#52525b66;background:#18181b66}
 @media(max-width:768px){.col-tt,.col-cx{display:none}}
 .prov{display:flex;gap:16px;flex-wrap:wrap}.prov-item{display:flex;align-items:center;font-size:.85rem;background:#111318;padding:8px 14px;border-radius:8px;border:1px solid #1e293b}
+.rename-btn{background:none;border:none;cursor:pointer;font-size:.75rem;opacity:.5;padding:2px}.rename-btn:hover{opacity:1}
 </style></head><body>
 <div class="header"><div><h1>⚡ RelayPlane Dashboard</h1></div><div class="meta"><a href="/dashboard/config">Config</a> · <span id="ver"></span><span id="vstat" class="vstat unavailable">Unable to check</span> · up <span id="uptime"></span> · refreshes every 5s</div></div>
 <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</div><div class="value green" id="savings">—</div></div>
+  <div class="card"><div class="label">Routing Savings <span class="tooltip-wrap"><span class="info-icon">ⓘ</span><span class="tooltip-box" id="savings-tooltip">Loading...</span></span></div><div class="value green" id="savings">—</div><div id="savings-detail" style="font-size:.75rem;color:#64748b;margin-top:4px">—</div></div>
   <div class="card"><div class="label">Avg Latency</div><div class="value" id="avgLat">—</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>
+<table><thead><tr><th>Provider</th><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>Agent Cost Breakdown</h2>
+<table><thead><tr><th>Agent</th><th>Requests</th><th>Total Cost</th><th>Last Active</th><th></th></tr></thead><tbody id="agents"></tbody></table></div>
 <div class="section"><h2>Provider Status</h2><div class="prov" id="providers"></div></div>
 <div class="section"><h2>Recent Runs</h2>
-<table><thead><tr><th>Time</th><th>Model</th><th class="col-tt">Task Type</th><th class="col-cx">Complexity</th><th>Tokens In</th><th>Tokens Out</th><th class="col-cache">Cache Create</th><th class="col-cache">Cache Read</th><th>Cost</th><th>Latency</th><th>Status</th></tr></thead><tbody id="runs"></tbody></table></div>
+<table><thead><tr><th>Time</th><th>Agent</th><th>Model</th><th class="col-tt">Task Type</th><th class="col-cx">Complexity</th><th>Tokens In</th><th>Tokens Out</th><th class="col-cache">Cache Create</th><th class="col-cache">Cache Read</th><th>Cost</th><th>Latency</th><th>Status</th></tr></thead><tbody id="runs"></tbody></table></div>
 <script>
 const $ = id => document.getElementById(id);
 function fmt(n,d=2){return typeof n==='number'?n.toFixed(d):'-'}
@@ -2012,12 +2145,13 @@ function fmtTime(s){const d=new Date(s);return d.toLocaleTimeString()}
 function dur(s){const h=Math.floor(s/3600),m=Math.floor(s%3600/60);return h?h+'h '+m+'m':m+'m'}
 async function load(){
   try{
-    const [health,stats,runsR,sav,provH]=await Promise.all([
+    const [health,stats,runsR,sav,provH,agentsR]=await Promise.all([
       fetch('/health').then(r=>r.json()),
       fetch('/v1/telemetry/stats').then(r=>r.json()),
       fetch('/v1/telemetry/runs?limit=20').then(r=>r.json()),
       fetch('/v1/telemetry/savings').then(r=>r.json()),
-      fetch('/v1/telemetry/health').then(r=>r.json())
+      fetch('/v1/telemetry/health').then(r=>r.json()),
+      fetch('/api/agents').then(r=>r.json()).catch(()=>({agents:[]}))
     ]);
     $('ver').textContent='v'+health.version;
     $('uptime').textContent=dur(health.uptime);
@@ -2037,16 +2171,65 @@ async function load(){
     const total=stats.summary?.totalEvents||0;
     $('totalReq').textContent=total;
     $('totalCost').textContent='$'+fmt(stats.summary?.totalCostUsd??0,4);
-    $('savings').textContent=(sav.percentage??0)+'%';
+    const savAmt=sav.savedAmount??sav.savings??0;
+    const cacheSav=sav.cacheSavings??0;
+    const routeSav=sav.routingSavings??0;
+    const actual=sav.actualCost??0;
+    const hasAnthropic=sav.hasAnthropicCalls!==false;
+    const baseline=sav.potentialSavings??sav.total??0;
+    // Headline = routing savings % (RelayPlane's actual contribution)
+    const routeBaseline=baseline>0?baseline:1;
+    const routePct=hasAnthropic?Math.round((routeSav/routeBaseline)*100):0;
+    const totalPct=sav.percentage??0;
+    $('savings').textContent='$'+fmt(routeSav,2);
+    // Secondary: show total % including cache as context
+    if(hasAnthropic){
+      $('savings-detail').innerHTML='<span style="color:#60a5fa">routing savings</span> · <span style="color:#64748b" title="Includes Anthropic prompt cache hits which happen regardless of routing">'+totalPct+'% total incl. cache</span>';
+    } else {
+      $('savings-detail').innerHTML='<span style="color:#a78bfa">$'+fmt(cacheSav,2)+' cache</span> · <span style="color:#64748b">'+totalPct+'% total</span>';
+    }
+    const tipEl=$('savings-tooltip');
+    if(tipEl){
+      let tip='<strong>How savings are calculated</strong><br><br>';
+      if(hasAnthropic){
+        tip+='<span style="color:#60a5fa">🔀 Routing savings: $'+fmt(routeSav,2)+'</span><br><small>Requests routed to cheaper models (e.g. Sonnet) vs always using Opus. RelayPlane contribution.</small><br><br>';
+        tip+='<span style="color:#a78bfa">💾 Cache savings: $'+fmt(cacheSav,2)+'</span><br><small>Anthropic prompt cache hits (10× cheaper reads). This would happen without RelayPlane too.</small><br><br>';
+      } else {
+        tip+='<span style="color:#a78bfa">💾 Cache savings: $'+fmt(cacheSav,2)+'</span><br><small>Provider cache hits. Happens automatically, not specific to RelayPlane.</small><br><br>';
+      }
+      tip+='💳 Actual cost: <b>$'+fmt(actual,2)+'</b><br>✅ Total saved: <b>$'+fmt(savAmt,2)+'</b>';
+      tipEl.innerHTML=tip;
+    }
     $('avgLat').textContent=(stats.summary?.avgLatencyMs??0)+'ms';
     $('models').innerHTML=(stats.byModel||[]).map(m=>
-      '<tr><td>'+m.model+'</td><td>'+m.count+'</td><td>$'+fmt(m.costUsd,4)+'</td><td>'+fmt(total>0?m.count/total*100:0,1)+'%</td></tr>'
-    ).join('')||'<tr><td colspan=4 style="color:#64748b">No data yet</td></tr>';
+      '<tr><td style="color:#94a3b8;font-size:.85rem">'+(m.provider||'—')+'</td><td>'+m.model+'</td><td>'+m.count+'</td><td>$'+fmt(m.costUsd,4)+'</td><td>'+fmt(total>0?m.count/total*100:0,1)+'%</td></tr>'
+    ).join('')||'<tr><td colspan=5 style="color:#64748b">No data yet</td></tr>';
     function ttCls(t){const m={code_generation:'tt-code',analysis:'tt-analysis',summarization:'tt-summarization',question_answering:'tt-qa'};return m[t]||'tt-general'}
     function cxCls(c){const m={simple:'cx-simple',moderate:'cx-moderate',complex:'cx-complex'};return m[c]||'cx-simple'}
-    $('runs').innerHTML=(runsR.runs||[]).map(r=>
-      '<tr><td>'+fmtTime(r.started_at)+'</td><td>'+r.model+'</td><td class="col-tt"><span class="badge '+ttCls(r.taskType)+'">'+(r.taskType||'general').replace(/_/g,' ')+'</span></td><td class="col-cx"><span class="badge '+cxCls(r.complexity)+'">'+(r.complexity||'simple')+'</span></td><td>'+(r.tokensIn||0)+'</td><td>'+(r.tokensOut||0)+'</td><td class="col-cache" style="color:#60a5fa">'+(r.cacheCreationTokens||0)+'</td><td class="col-cache" style="color:#34d399">'+(r.cacheReadTokens||0)+'</td><td>$'+fmt(r.costUsd,4)+'</td><td>'+r.latencyMs+'ms</td><td><span class="badge '+(r.status==='success'?'ok':'err')+'">'+r.status+'</span></td></tr>'
-    ).join('')||'<tr><td colspan=11 style="color:#64748b">No runs yet</td></tr>';
+    function esc(s){if(!s)return'';return s.replace(/&/g,'&amp;').replace(/</g,'&lt;').replace(/>/g,'&gt;')}
+    const agents=(agentsR.agents||[]).sort((a,b)=>(b.totalCost||0)-(a.totalCost||0));
+    $('runs').innerHTML=(runsR.runs||[]).map((r,i)=>{
+      function errBadge(r){if(r.status==='success')return '<span class="badge ok">success</span>';var cls='err';var label=r.error||'error';if(r.statusCode===401||r.statusCode===403||(r.error&&/auth/i.test(r.error)))cls='err-auth';else if(r.statusCode===429||(r.error&&/rate.?limit/i.test(r.error)))cls='err-rate';else if(r.error&&/timeout/i.test(r.error))cls='err-timeout';return '<span class="badge '+cls+'" title="'+esc(r.error||'')+' (HTTP '+( r.statusCode||'?')+')">'+(r.statusCode?r.statusCode+' ':'')+ (label.length>40?label.slice(0,40)+'…':label)+'</span>';}
+      const agentName=agents.find(a=>a.fingerprint===r.agentFingerprint)?.name||(r.agentId||'—');
+      const row='<tr style="cursor:pointer" onclick="toggleDetail('+i+')"><td><span id="arrow-'+i+'" style="color:#64748b;font-size:.7rem;margin-right:6px">▶</span>'+fmtTime(r.started_at)+'</td><td style="font-size:.85rem">'+esc(agentName)+'</td><td>'+r.model+'</td><td class="col-tt"><span class="badge '+ttCls(r.taskType)+'">'+(r.taskType||'general').replace(/_/g,' ')+'</span></td><td class="col-cx"><span class="badge '+cxCls(r.complexity)+'">'+(r.complexity||'simple')+'</span></td><td>'+(r.tokensIn||0)+'</td><td>'+(r.tokensOut||0)+'</td><td class="col-cache" style="color:#60a5fa">'+(r.cacheCreationTokens||0)+'</td><td class="col-cache" style="color:#34d399">'+(r.cacheReadTokens||0)+'</td><td>$'+fmt(r.costUsd,4)+'</td><td>'+r.latencyMs+'ms</td><td>'+errBadge(r)+'</td></tr>';
+      const c=r.requestContent||{};
+      let detail='<tr id="run-detail-'+i+'" style="display:none"><td colspan="12" style="padding:16px;background:#111217;border-bottom:1px solid #1e293b">';
+      if(c.systemPrompt||c.userMessage||c.responsePreview){
+        if(c.systemPrompt) detail+='<div style="color:#64748b;font-size:.85rem;margin-bottom:10px;font-style:italic"><strong style="color:#94a3b8">System:</strong> '+esc(c.systemPrompt)+'</div>';
+        if(c.userMessage) detail+='<div style="background:#1a1c23;border:1px solid #1e293b;border-radius:8px;padding:12px;margin-bottom:10px"><strong style="color:#94a3b8;font-size:.8rem">User Message</strong><div style="margin-top:6px;white-space:pre-wrap">'+esc(c.userMessage)+'</div></div>';
+        if(c.responsePreview) detail+='<div style="background:#1a1c23;border:1px solid #1e293b;border-radius:8px;padding:12px;margin-bottom:10px"><strong style="color:#94a3b8;font-size:.8rem">Response Preview</strong><div style="margin-top:6px;white-space:pre-wrap">'+esc(c.responsePreview)+'</div></div>';
+        const btnAttrs='id="full-btn-'+i+'" style="background:#1e293b;color:#e2e8f0;border:1px solid #334155;padding:6px 12px;border-radius:6px;font-size:.8rem"';
+        detail+=(r.tokensOut>0?'<button onclick="event.stopPropagation();loadFullResponse(&quot;'+r.id+'&quot;,'+i+')" '+btnAttrs+'>Show full response</button>':'<button disabled '+btnAttrs+' style="opacity:.4;cursor:default">Response not available (streaming)</button>')+'<pre id="full-resp-'+i+'" style="display:none;white-space:pre-wrap;margin-top:10px;background:#0d0e11;border:1px solid #1e293b;border-radius:8px;padding:12px;max-height:400px;overflow:auto;font-size:.8rem"></pre>';
+      } else {
+        detail+='<span style="color:#64748b">No content captured for this request</span>';
+      }
+      detail+='</td></tr>';
+      return row+detail;
+    }).join('')||'<tr><td colspan=12 style="color:#64748b">No runs yet</td></tr>';
+    restoreExpanded();
+    $('agents').innerHTML=agents.length?agents.map(a=>
+      '<tr><td><span class="agent-name" data-fp="'+a.fingerprint+'">'+esc(a.name)+'</span> <button class="rename-btn" onclick="renameAgent(&quot;'+a.fingerprint+'&quot;,&quot;'+a.name.replace(/"/g,'')+'&quot;)">✏️</button></td><td>'+a.totalRequests+'</td><td>$'+fmt(a.totalCost,4)+'</td><td>'+fmtTime(a.lastSeen)+'</td><td style="font-size:.7rem;color:#64748b" title="'+esc(a.systemPromptPreview||'')+'">'+a.fingerprint+'</td></tr>'
+    ).join(''):'<tr><td colspan=5 style="color:#64748b">No agents detected yet</td></tr>';
     $('providers').innerHTML=(provH.providers||[]).map(p=>{
       const dotClass = p.status==='healthy'?'up':(p.status==='degraded'?'warn':'down');
       const rate = p.successRate!==undefined?(' '+Math.round(p.successRate*100)+'%'):'';
@@ -2054,13 +2237,34 @@ async function load(){
     }).join('');
   }catch(e){console.error(e)}
 }
+async function renameAgent(fp,currentName){
+  const name=prompt('Rename agent:',currentName);
+  if(!name||name===currentName)return;
+  await fetch('/api/agents/rename',{method:'POST',headers:{'Content-Type':'application/json'},body:JSON.stringify({fingerprint:fp,name:name})});
+  load();
+}
+const expandedRows=new Set();
+function toggleDetail(i){var d=document.getElementById('run-detail-'+i);var arrow=document.getElementById('arrow-'+i);if(d.style.display==='none'){d.style.display='table-row';expandedRows.add(i);if(arrow)arrow.textContent='▼'}else{d.style.display='none';expandedRows.delete(i);if(arrow)arrow.textContent='▶'}}
+function restoreExpanded(){expandedRows.forEach(i=>{var d=document.getElementById('run-detail-'+i);var arrow=document.getElementById('arrow-'+i);if(d)d.style.display='table-row';if(arrow)arrow.textContent='▼'})}
+async function loadFullResponse(runId,i){
+  const btn=document.getElementById('full-btn-'+i);
+  const pre=document.getElementById('full-resp-'+i);
+  if(pre.style.display!=='none'){pre.style.display='none';btn.textContent='Show full response';return}
+  btn.textContent='Loading...';
+  try{
+    const data=await fetch('/api/runs/'+runId).then(r=>r.json());
+    const full=data.requestContent&&data.requestContent.fullResponse;
+    if(full){pre.textContent=full;pre.style.display='block';btn.textContent='Hide full response'}
+    else{btn.textContent='No full response available'}
+  }catch{btn.textContent='Error loading response'}
+}
 load();setInterval(load,5000);
-</script></body></html>`;
+</script><footer style="text-align:center;padding:20px 0;color:#475569;font-size:.75rem;border-top:1px solid #1e293b;margin-top:20px">🔒 Request content stays on your machine. Never sent to cloud.</footer></body></html>`;
 }
 function getConfigDashboardHTML() {
     return `<!DOCTYPE html><html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>RelayPlane Config</title>
 <style>
-*{margin:0;padding:0;box-sizing:border-box}body{background:#0a0b0d;color:#e2e8f0;font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',sans-serif;padding:20px;max-width:1200px;margin:0 auto}
+*{margin:0;padding:0;box-sizing:border-box}body{background:#0a0b0d;color:#e2e8f0;font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',sans-serif;padding:20px;max-width:1600px;margin:0 auto}
 a{color:#34d399}h1{font-size:1.5rem;font-weight:600}
 .header{display:flex;justify-content:space-between;align-items:center;padding:16px 0;border-bottom:1px solid #1e293b;margin-bottom:24px}
 .header .meta{font-size:.8rem;color:#64748b}
@@ -2152,8 +2356,10 @@ async function startProxy(config = {}) {
     };
     // Load persistent history from disk
     loadHistoryFromDisk();
+    (0, agent_tracker_js_1.loadAgentRegistry)();
     // Flush history on shutdown
     const handleShutdown = () => {
+        (0, agent_tracker_js_1.flushAgentRegistry)();
         meshHandle.stop();
         shutdownHistory();
         process.exit(0);
@@ -2162,6 +2368,7 @@ async function startProxy(config = {}) {
     process.on('SIGTERM', handleShutdown);
     const configPath = getProxyConfigPath();
     let proxyConfig = await loadProxyConfig(configPath, log);
+    _activeProxyConfig = proxyConfig;
     const cooldownManager = new CooldownManager(getCooldownConfig(proxyConfig));
     // === Startup config validation (Task 4) ===
     try {
@@ -2500,11 +2707,11 @@ async function startProxy(config = {}) {
                 const days = parseInt(params.get('days') || '7', 10);
                 const cutoff = Date.now() - days * 86400000;
                 const recent = requestHistory.filter(r => new Date(r.timestamp).getTime() >= cutoff);
-                // Model breakdown
+                // Model breakdown (keyed by provider/model for disambiguation)
                 const modelMap = new Map();
                 for (const r of recent) {
-                    const key = r.targetModel;
-                    const cur = modelMap.get(key) || { count: 0, cost: 0 };
+                    const key = `${r.provider || 'unknown'}/${r.targetModel}`;
+                    const cur = modelMap.get(key) || { count: 0, cost: 0, provider: r.provider || 'unknown', model: r.targetModel };
                     cur.count++;
                     cur.cost += r.costUsd;
                     modelMap.set(key, cur);
@@ -2527,7 +2734,7 @@ async function startProxy(config = {}) {
                         avgLatencyMs: recent.length ? Math.round(totalLatency / recent.length) : 0,
                         successRate: recent.length ? recent.filter(r => r.success).length / recent.length : 0,
                     },
-                    byModel: Array.from(modelMap.entries()).map(([model, v]) => ({ model, count: v.count, costUsd: v.cost, savings: 0 })),
+                    byModel: Array.from(modelMap.entries()).map(([, v]) => ({ model: v.model, provider: v.provider, count: v.count, costUsd: v.cost, savings: 0 })),
                     dailyCosts: Array.from(dailyMap.entries()).map(([date, v]) => ({ date, costUsd: v.cost, requests: v.requests })),
                 };
                 res.writeHead(200, { 'Content-Type': 'application/json' });
@@ -2565,6 +2772,16 @@ async function startProxy(config = {}) {
                         cacheReadTokens: r.cacheReadTokens ?? 0,
                         savings: Math.round(perRunSavings * 10000) / 10000,
                         escalated: r.escalated,
+                        error: r.error ?? null,
+                        statusCode: r.statusCode ?? null,
+                        agentFingerprint: r.agentFingerprint ?? null,
+                        agentId: r.agentId ?? null,
+                        requestContent: r.requestContent ? {
+                            systemPrompt: r.requestContent.systemPrompt,
+                            userMessage: r.requestContent.userMessage,
+                            responsePreview: r.requestContent.responsePreview,
+                            // fullResponse excluded from list endpoint to keep payloads small
+                        } : undefined,
                     };
                 });
                 res.writeHead(200, { 'Content-Type': 'application/json' });
@@ -2572,26 +2789,34 @@ async function startProxy(config = {}) {
                 return;
             }
             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-4-6';
-                let totalOriginalCost = 0;
+                // Routing savings: cost at same model with no cache vs actual cost
+                // Cache savings: what cache hits saved vs paying full input price
+                // Baseline: each request at full input price (no cache, no routing)
                 let totalActualCost = 0;
-                let totalSavedAmount = 0;
+                let totalCacheSavings = 0; // savings from cache hits (Anthropic feature)
+                let totalRoutingSavings = 0; // savings from routing to cheaper model
+                let hasAnthropicCalls = false;
                 const byDayMap = new Map();
                 for (const r of requestHistory) {
-                    // Pass same cache tokens to baseline so savings only reflect routing decisions,
-                    // not prompt-cache discounts (those happen regardless of which model is chosen).
-                    const origCost = (0, telemetry_js_1.estimateCost)(OPUS_BASELINE, r.tokensIn, r.tokensOut, r.cacheCreationTokens || undefined, r.cacheReadTokens || undefined);
                     const actualCost = r.costUsd;
-                    const saved = Math.max(0, origCost - actualCost);
-                    totalOriginalCost += origCost;
                     totalActualCost += actualCost;
-                    totalSavedAmount += saved;
+                    // Cache savings: full input price vs what was paid with cache
+                    const fullInputCost = (0, telemetry_js_1.estimateCost)(r.targetModel, r.tokensIn + (r.cacheCreationTokens || 0) + (r.cacheReadTokens || 0), r.tokensOut);
+                    const cachedCost = r.costUsd;
+                    const cacheSaved = Math.max(0, fullInputCost - cachedCost);
+                    totalCacheSavings += cacheSaved;
+                    // Routing savings: what would this request cost at full Opus price (no cache)
+                    // vs what the routed model cost (no cache). Only meaningful for Anthropic.
+                    if (r.provider === 'anthropic') {
+                        hasAnthropicCalls = true;
+                        const opusCost = (0, telemetry_js_1.estimateCost)('claude-opus-4-6', r.tokensIn, r.tokensOut);
+                        const modelCost = (0, telemetry_js_1.estimateCost)(r.targetModel, r.tokensIn, r.tokensOut);
+                        const routingSaved = Math.max(0, opusCost - modelCost);
+                        totalRoutingSavings += routingSaved;
+                    }
                     const date = r.timestamp.slice(0, 10);
                     const day = byDayMap.get(date) || { savedAmount: 0, originalCost: 0, actualCost: 0 };
-                    day.savedAmount += saved;
-                    day.originalCost += origCost;
+                    day.savedAmount += Math.max(0, totalCacheSavings + totalRoutingSavings);
                     day.actualCost += actualCost;
                     byDayMap.set(date, day);
                 }
@@ -2603,16 +2828,19 @@ async function startProxy(config = {}) {
                     originalCost: Math.round(v.originalCost * 10000) / 10000,
                     actualCost: Math.round(v.actualCost * 10000) / 10000,
                 }));
+                const totalSaved = totalCacheSavings + totalRoutingSavings;
+                const baseline = totalActualCost + totalSaved;
                 res.writeHead(200, { 'Content-Type': 'application/json' });
                 res.end(JSON.stringify({
-                    total: Math.round(totalOriginalCost * 10000) / 10000,
                     actualCost: Math.round(totalActualCost * 10000) / 10000,
-                    savings: Math.round(totalSavedAmount * 10000) / 10000,
-                    savedAmount: Math.round(totalSavedAmount * 10000) / 10000,
-                    potentialSavings: Math.round(totalOriginalCost * 10000) / 10000,
-                    percentage: totalOriginalCost > 0
-                        ? Math.round((totalSavedAmount / totalOriginalCost) * 100)
-                        : 0,
+                    savedAmount: Math.round(totalSaved * 10000) / 10000,
+                    savings: Math.round(totalSaved * 10000) / 10000,
+                    cacheSavings: Math.round(totalCacheSavings * 10000) / 10000,
+                    routingSavings: Math.round(totalRoutingSavings * 10000) / 10000,
+                    hasAnthropicCalls,
+                    potentialSavings: Math.round(baseline * 10000) / 10000,
+                    total: Math.round(baseline * 10000) / 10000,
+                    percentage: baseline > 0 ? Math.round((totalSaved / baseline) * 100) : 0,
                     byDay,
                 }));
                 return;
@@ -2662,6 +2890,63 @@ async function startProxy(config = {}) {
             res.end(JSON.stringify({ error: 'Not found' }));
             return;
         }
+        // === Agent tracking API ===
+        // === /api/runs/:id — full request/response content for a single run ===
+        const runsIdMatch = pathname.match(/^\/api\/runs\/(.+)$/);
+        if (req.method === 'GET' && runsIdMatch) {
+            const runId = runsIdMatch[1];
+            const run = requestHistory.find(r => r.id === runId);
+            if (!run) {
+                res.writeHead(404, { 'Content-Type': 'application/json' });
+                res.end(JSON.stringify({ error: 'Run not found' }));
+                return;
+            }
+            res.writeHead(200, { 'Content-Type': 'application/json' });
+            res.end(JSON.stringify({
+                id: run.id,
+                model: run.targetModel,
+                provider: run.provider,
+                timestamp: run.timestamp,
+                tokensIn: run.tokensIn,
+                tokensOut: run.tokensOut,
+                costUsd: run.costUsd,
+                latencyMs: run.latencyMs,
+                success: run.success,
+                requestContent: run.requestContent,
+            }));
+            return;
+        }
+        if (req.method === 'GET' && pathname === '/api/agents') {
+            const summaries = (0, agent_tracker_js_1.getAgentSummaries)(requestHistory);
+            res.writeHead(200, { 'Content-Type': 'application/json' });
+            res.end(JSON.stringify({ agents: summaries }));
+            return;
+        }
+        if (req.method === 'POST' && pathname === '/api/agents/rename') {
+            try {
+                const body = await readJsonBody(req);
+                const fingerprint = body['fingerprint'];
+                const name = body['name'];
+                if (!fingerprint || !name) {
+                    res.writeHead(400, { 'Content-Type': 'application/json' });
+                    res.end(JSON.stringify({ error: 'Missing fingerprint or name' }));
+                    return;
+                }
+                const ok = (0, agent_tracker_js_1.renameAgent)(fingerprint, name);
+                if (!ok) {
+                    res.writeHead(404, { 'Content-Type': 'application/json' });
+                    res.end(JSON.stringify({ error: 'Agent not found' }));
+                    return;
+                }
+                res.writeHead(200, { 'Content-Type': 'application/json' });
+                res.end(JSON.stringify({ ok: true }));
+            }
+            catch {
+                res.writeHead(400, { 'Content-Type': 'application/json' });
+                res.end(JSON.stringify({ error: 'Invalid JSON' }));
+            }
+            return;
+        }
         // === Dashboard ===
         if (req.method === 'GET' && (pathname === '/' || pathname === '/dashboard')) {
             res.writeHead(200, { 'Content-Type': 'text/html' });
@@ -2742,6 +3027,14 @@ async function startProxy(config = {}) {
                 res.end(JSON.stringify({ error: 'Invalid JSON' }));
                 return;
             }
+            // Extract agent fingerprint and explicit agent ID
+            const nativeSystemPrompt = (0, agent_tracker_js_1.extractSystemPromptFromBody)(requestBody);
+            const nativeExplicitAgentId = getHeaderValue(req, 'x-relayplane-agent') || undefined;
+            let nativeAgentFingerprint;
+            if (nativeSystemPrompt) {
+                const agentResult = (0, agent_tracker_js_1.trackAgent)(nativeSystemPrompt, 0, nativeExplicitAgentId);
+                nativeAgentFingerprint = agentResult.fingerprint;
+            }
             const originalModel = requestBody['model'];
             let requestedModel = headerModelOverride ?? originalModel ?? '';
             if (headerModelOverride) {
@@ -3076,7 +3369,8 @@ async function startProxy(config = {}) {
                             cooldownManager.recordFailure(targetProvider, JSON.stringify(errorPayload));
                         }
                         const durationMs = Date.now() - startTime;
-                        logRequest(originalModel ?? 'unknown', targetModel || requestedModel, targetProvider, durationMs, false, routingMode, undefined, taskType, complexity);
+                        const errMsg = extractProviderErrorMessage(errorPayload, providerResponse.status);
+                        logRequest(originalModel ?? 'unknown', targetModel || requestedModel, targetProvider, durationMs, false, routingMode, undefined, taskType, complexity, undefined, undefined, errMsg, providerResponse.status);
                         res.writeHead(providerResponse.status, { 'Content-Type': 'application/json' });
                         res.end(JSON.stringify(errorPayload));
                         return;
@@ -3203,7 +3497,22 @@ async function startProxy(config = {}) {
                 const nativeTokIn = nativeBaseTokIn + nativeCacheCreation + nativeCacheRead;
                 // Cost calculation expects inputTokens to include cache tokens when cache params are provided
                 const nativeCostUsd = (0, telemetry_js_1.estimateCost)(targetModel || requestedModel, nativeTokIn, nativeTokOut, nativeCacheCreation || undefined, nativeCacheRead || undefined);
-                updateLastHistoryEntry(nativeTokIn, nativeTokOut, nativeCostUsd, undefined, nativeCacheCreation || undefined, nativeCacheRead || undefined);
+                // Build request content if logging enabled
+                let nativeContentData;
+                if (isContentLoggingEnabled()) {
+                    const extracted = extractRequestContent(requestBody, true);
+                    const responseText = nativeResponseData ? extractResponseText(nativeResponseData, true) : '';
+                    nativeContentData = {
+                        ...extracted,
+                        responsePreview: responseText ? responseText.slice(0, 500) : undefined,
+                        fullResponse: responseText || undefined,
+                    };
+                }
+                updateLastHistoryEntry(nativeTokIn, nativeTokOut, nativeCostUsd, undefined, nativeCacheCreation || undefined, nativeCacheRead || undefined, nativeAgentFingerprint, nativeExplicitAgentId, nativeContentData);
+                // Update agent cost now that we know the actual cost
+                if (nativeAgentFingerprint && nativeAgentFingerprint !== 'unknown') {
+                    (0, agent_tracker_js_1.updateAgentCost)(nativeAgentFingerprint, nativeCostUsd);
+                }
                 // ── Post-request: budget spend + anomaly detection ──
                 postRequestRecord(targetModel || requestedModel, nativeTokIn, nativeTokOut, nativeCostUsd);
                 if (recordTelemetry) {
@@ -3224,7 +3533,17 @@ async function startProxy(config = {}) {
             }
             catch (err) {
                 const durationMs = Date.now() - startTime;
-                logRequest(originalModel ?? 'unknown', targetModel || requestedModel, targetProvider, durationMs, false, routingMode, undefined, taskType, complexity);
+                let catchErrMsg;
+                let catchErrStatus;
+                if (err instanceof ProviderResponseError) {
+                    catchErrMsg = extractProviderErrorMessage(err.payload, err.status);
+                    catchErrStatus = err.status;
+                }
+                else {
+                    catchErrMsg = err instanceof Error ? err.message : String(err);
+                    catchErrStatus = 500;
+                }
+                logRequest(originalModel ?? 'unknown', targetModel || requestedModel, targetProvider, durationMs, false, routingMode, undefined, taskType, complexity, undefined, undefined, catchErrMsg, catchErrStatus);
                 if (err instanceof ProviderResponseError) {
                     res.writeHead(err.status, { 'Content-Type': 'application/json' });
                     res.end(JSON.stringify(err.payload));
@@ -3301,6 +3620,14 @@ async function startProxy(config = {}) {
             return;
         }
         const isStreaming = request.stream === true;
+        // Extract agent fingerprint for chat/completions
+        const chatSystemPrompt = (0, agent_tracker_js_1.extractSystemPromptFromBody)(request);
+        const chatExplicitAgentId = getHeaderValue(req, 'x-relayplane-agent') || undefined;
+        let chatAgentFingerprint;
+        if (chatSystemPrompt) {
+            const agentResult = (0, agent_tracker_js_1.trackAgent)(chatSystemPrompt, 0, chatExplicitAgentId);
+            chatAgentFingerprint = agentResult.fingerprint;
+        }
         // ── Response Cache: check for cached response (chat/completions) ──
         const chatCacheBypass = responseCache.shouldBypass(request);
         let chatCacheHash;
@@ -3601,7 +3928,7 @@ async function startProxy(config = {}) {
         const startTime = Date.now();
         // Handle streaming vs non-streaming
         if (isStreaming) {
-            await handleStreamingRequest(res, request, targetProvider, targetModel, apiKey, ctx, relay, promptText, taskType, confidence, useCascade ? 'cascade' : routingMode, recordTelemetry, startTime, log, cooldownManager, cooldownsEnabled, complexity, chatCacheHash, chatCacheBypass);
+            await handleStreamingRequest(res, request, targetProvider, targetModel, apiKey, ctx, relay, promptText, taskType, confidence, useCascade ? 'cascade' : routingMode, recordTelemetry, startTime, log, cooldownManager, cooldownsEnabled, complexity, chatCacheHash, chatCacheBypass, chatAgentFingerprint, chatExplicitAgentId);
         }
         else {
             if (useCascade && cascadeConfig) {
@@ -3641,7 +3968,9 @@ async function startProxy(config = {}) {
                     const cascadeCacheCreation = cascadeUsage?.cache_creation_input_tokens || undefined;
                     const cascadeCacheRead = cascadeUsage?.cache_read_input_tokens || undefined;
                     const cascadeCost = (0, telemetry_js_1.estimateCost)(cascadeResult.model, cascadeTokensIn, cascadeTokensOut, cascadeCacheCreation, cascadeCacheRead);
-                    updateLastHistoryEntry(cascadeTokensIn, cascadeTokensOut, cascadeCost, chatCascadeRespModel, cascadeCacheCreation, cascadeCacheRead);
+                    updateLastHistoryEntry(cascadeTokensIn, cascadeTokensOut, cascadeCost, chatCascadeRespModel, cascadeCacheCreation, cascadeCacheRead, chatAgentFingerprint, chatExplicitAgentId);
+                    if (chatAgentFingerprint && chatAgentFingerprint !== 'unknown')
+                        (0, agent_tracker_js_1.updateAgentCost)(chatAgentFingerprint, cascadeCost);
                     if (recordTelemetry) {
                         try {
                             const runResult = await relay.run({
@@ -3674,7 +4003,17 @@ async function startProxy(config = {}) {
                 }
                 catch (err) {
                     const durationMs = Date.now() - startTime;
-                    logRequest(originalRequestedModel ?? 'unknown', targetModel || 'unknown', targetProvider, durationMs, false, 'cascade', undefined, taskType, complexity);
+                    let cascadeErrMsg;
+                    let cascadeErrStatus;
+                    if (err instanceof ProviderResponseError) {
+                        cascadeErrMsg = extractProviderErrorMessage(err.payload, err.status);
+                        cascadeErrStatus = err.status;
+                    }
+                    else {
+                        cascadeErrMsg = err instanceof Error ? err.message : String(err);
+                        cascadeErrStatus = 500;
+                    }
+                    logRequest(originalRequestedModel ?? 'unknown', targetModel || 'unknown', targetProvider, durationMs, false, 'cascade', undefined, taskType, complexity, undefined, undefined, cascadeErrMsg, cascadeErrStatus);
                     if (err instanceof ProviderResponseError) {
                         res.writeHead(err.status, { 'Content-Type': 'application/json' });
                         res.end(JSON.stringify(err.payload));
@@ -3686,7 +4025,7 @@ async function startProxy(config = {}) {
                 }
             }
             else {
-                await handleNonStreamingRequest(res, request, targetProvider, targetModel, apiKey, ctx, relay, promptText, taskType, confidence, routingMode, recordTelemetry, startTime, log, cooldownManager, cooldownsEnabled, complexity);
+                await handleNonStreamingRequest(res, request, targetProvider, targetModel, apiKey, ctx, relay, promptText, taskType, confidence, routingMode, recordTelemetry, startTime, log, cooldownManager, cooldownsEnabled, complexity, chatAgentFingerprint, chatExplicitAgentId);
             }
         }
     });
@@ -3829,7 +4168,7 @@ async function executeNonStreamingProviderRequest(request, targetProvider, targe
     }
     return { responseData, ok: true, status: 200 };
 }
-async function handleStreamingRequest(res, request, targetProvider, targetModel, apiKey, ctx, relay, promptText, taskType, confidence, routingMode, recordTelemetry, startTime, log, cooldownManager, cooldownsEnabled, complexity = 'simple', cacheHash, cacheBypass) {
+async function handleStreamingRequest(res, request, targetProvider, targetModel, apiKey, ctx, relay, promptText, taskType, confidence, routingMode, recordTelemetry, startTime, log, cooldownManager, cooldownsEnabled, complexity = 'simple', cacheHash, cacheBypass, agentFingerprint, agentId) {
     let providerResponse;
     try {
         switch (targetProvider) {
@@ -3857,7 +4196,8 @@ async function handleStreamingRequest(res, request, targetProvider, targetModel,
                 cooldownManager.recordFailure(targetProvider, JSON.stringify(errorData));
             }
             const durationMs = Date.now() - startTime;
-            logRequest(request.model ?? 'unknown', targetModel, targetProvider, durationMs, false, routingMode, undefined, taskType, complexity);
+            const streamErrMsg = extractProviderErrorMessage(errorData, providerResponse.status);
+            logRequest(request.model ?? 'unknown', targetModel, targetProvider, durationMs, false, routingMode, undefined, taskType, complexity, undefined, undefined, streamErrMsg, providerResponse.status);
             res.writeHead(providerResponse.status, { 'Content-Type': 'application/json' });
             res.end(JSON.stringify(errorData));
             return;
@@ -3869,7 +4209,7 @@ async function handleStreamingRequest(res, request, targetProvider, targetModel,
             cooldownManager.recordFailure(targetProvider, errorMsg);
         }
         const durationMs = Date.now() - startTime;
-        logRequest(request.model ?? 'unknown', targetModel, targetProvider, durationMs, false, routingMode, undefined, taskType, complexity);
+        logRequest(request.model ?? 'unknown', targetModel, targetProvider, durationMs, false, routingMode, undefined, taskType, complexity, undefined, undefined, errorMsg, 500);
         res.writeHead(500, { 'Content-Type': 'application/json' });
         res.end(JSON.stringify({ error: `Provider error: ${errorMsg}` }));
         return;
@@ -3988,7 +4328,9 @@ async function handleStreamingRequest(res, request, targetProvider, targetModel,
     logRequest(request.model ?? 'unknown', targetModel, targetProvider, durationMs, true, routingMode, undefined, taskType, complexity);
     // Update token/cost info on the history entry (with cache token discount)
     const streamCost = (0, telemetry_js_1.estimateCost)(targetModel, streamTokensIn, streamTokensOut, streamCacheCreation || undefined, streamCacheRead || undefined);
-    updateLastHistoryEntry(streamTokensIn, streamTokensOut, streamCost, undefined, streamCacheCreation || undefined, streamCacheRead || undefined);
+    updateLastHistoryEntry(streamTokensIn, streamTokensOut, streamCost, undefined, streamCacheCreation || undefined, streamCacheRead || undefined, agentFingerprint, agentId);
+    if (agentFingerprint && agentFingerprint !== 'unknown')
+        (0, agent_tracker_js_1.updateAgentCost)(agentFingerprint, streamCost);
     // ── Post-request: budget spend + anomaly detection ──
     try {
         (0, budget_js_1.getBudgetManager)().recordSpend(streamCost, targetModel);
@@ -4024,7 +4366,7 @@ async function handleStreamingRequest(res, request, targetProvider, targetModel,
 /**
  * Handle non-streaming request
  */
-async function handleNonStreamingRequest(res, request, targetProvider, targetModel, apiKey, ctx, relay, promptText, taskType, confidence, routingMode, recordTelemetry, startTime, log, cooldownManager, cooldownsEnabled, complexity = 'simple') {
+async function handleNonStreamingRequest(res, request, targetProvider, targetModel, apiKey, ctx, relay, promptText, taskType, confidence, routingMode, recordTelemetry, startTime, log, cooldownManager, cooldownsEnabled, complexity = 'simple', agentFingerprint, agentId) {
     let responseData;
     try {
         const result = await executeNonStreamingProviderRequest(request, targetProvider, targetModel, apiKey, ctx);
@@ -4034,7 +4376,8 @@ async function handleNonStreamingRequest(res, request, targetProvider, targetMod
                 cooldownManager.recordFailure(targetProvider, JSON.stringify(responseData));
             }
             const durationMs = Date.now() - startTime;
-            logRequest(request.model ?? 'unknown', targetModel, targetProvider, durationMs, false, routingMode, undefined, taskType, complexity);
+            const nsErrMsg = extractProviderErrorMessage(responseData, result.status);
+            logRequest(request.model ?? 'unknown', targetModel, targetProvider, durationMs, false, routingMode, undefined, taskType, complexity, undefined, undefined, nsErrMsg, result.status);
             res.writeHead(result.status, { 'Content-Type': 'application/json' });
             res.end(JSON.stringify(responseData));
             return;
@@ -4046,7 +4389,7 @@ async function handleNonStreamingRequest(res, request, targetProvider, targetMod
             cooldownManager.recordFailure(targetProvider, errorMsg);
         }
         const durationMs = Date.now() - startTime;
-        logRequest(request.model ?? 'unknown', targetModel, targetProvider, durationMs, false, routingMode, undefined, taskType, complexity);
+        logRequest(request.model ?? 'unknown', targetModel, targetProvider, durationMs, false, routingMode, undefined, taskType, complexity, undefined, undefined, errorMsg, 500);
         res.writeHead(500, { 'Content-Type': 'application/json' });
         res.end(JSON.stringify({ error: `Provider error: ${errorMsg}` }));
         return;
@@ -4066,7 +4409,9 @@ async function handleNonStreamingRequest(res, request, targetProvider, targetMod
     const cacheCreationTokens = usage?.cache_creation_input_tokens ?? 0;
     const cacheReadTokens = usage?.cache_read_input_tokens ?? 0;
     const cost = (0, telemetry_js_1.estimateCost)(targetModel, tokensIn, tokensOut, cacheCreationTokens || undefined, cacheReadTokens || undefined);
-    updateLastHistoryEntry(tokensIn, tokensOut, cost, nonStreamRespModel, cacheCreationTokens || undefined, cacheReadTokens || undefined);
+    updateLastHistoryEntry(tokensIn, tokensOut, cost, nonStreamRespModel, cacheCreationTokens || undefined, cacheReadTokens || undefined, agentFingerprint, agentId);
+    if (agentFingerprint && agentFingerprint !== 'unknown')
+        (0, agent_tracker_js_1.updateAgentCost)(agentFingerprint, cost);
     // ── Post-request: budget spend + anomaly detection ──
     try {
         (0, budget_js_1.getBudgetManager)().recordSpend(cost, targetModel);