@relayplane/proxy 1.7.1 → 1.7.3

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) {
     const timestamp = new Date().toISOString();
     const status = success ? '✓' : '✗';
     const escalateTag = escalated ? ' [ESCALATED]' : '';
@@ -531,6 +534,8 @@ function logRequest(originalModel, targetModel, provider, latencyMs, success, mo
         costUsd: 0,
         taskType: taskType || 'general',
         complexity: complexity || 'simple',
+        agentFingerprint,
+        agentId,
     };
     requestHistory.push(entry);
     if (requestHistory.length > MAX_HISTORY) {
@@ -539,7 +544,7 @@ 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) {
+function updateLastHistoryEntry(tokensIn, tokensOut, costUsd, responseModel, cacheCreationTokens, cacheReadTokens, agentFingerprint, agentId, requestContent) {
     if (requestHistory.length > 0) {
         const last = requestHistory[requestHistory.length - 1];
         last.tokensIn = tokensIn;
@@ -548,8 +553,86 @@ function updateLastHistoryEntry(tokensIn, tokensOut, costUsd, responseModel) {
         if (responseModel) {
             last.responseModel = responseModel;
         }
+        if (cacheCreationTokens !== undefined)
+            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;
     }
 }
+/**
+ * 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: {},
@@ -580,6 +663,11 @@ 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())
@@ -773,6 +861,23 @@ function classifyComplexity(messages) {
         score += 1;
     if (andCount >= 5)
         score += 1;
+    // Calculate total tokens across ALL messages, not just last user message.
+    // For agent workloads (OpenClaw, aider, Claude Code) the last user message is
+    // often tiny while the real complexity lives in the 100K+ token context.
+    const allText = extractMessageText(messages);
+    const totalTokens = Math.ceil(allText.length / 4);
+    // Context size floor — use as a hard signal regardless of last-message score
+    if (totalTokens > 100000)
+        score += 5; // definitely complex
+    else if (totalTokens > 50000)
+        score += 3; // likely moderate+
+    else if (totalTokens > 20000)
+        score += 2;
+    // Message count signal — long conversations imply multi-step reasoning
+    if (messages.length > 50)
+        score += 2;
+    else if (messages.length > 20)
+        score += 1;
     if (score >= 4)
         return 'complex';
     if (score >= 2)
@@ -1505,11 +1610,13 @@ 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
+            // Pass through input token count from message_start (including cache tokens)
             const msgUsage = msg?.['usage'];
             if (msgUsage) {
                 baseChunk['usage'] = {
                     prompt_tokens: msgUsage['input_tokens'] ?? 0,
+                    cache_creation_tokens: msgUsage['cache_creation_input_tokens'] ?? 0,
+                    cache_read_tokens: msgUsage['cache_read_input_tokens'] ?? 0,
                 };
             }
             return `data: ${JSON.stringify(baseChunk)}\n\n`;
@@ -1749,7 +1856,7 @@ 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];
@@ -1917,7 +2024,7 @@ 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 };
             }
@@ -1969,6 +2076,7 @@ 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">
@@ -1979,9 +2087,11 @@ td{padding:8px 12px;border-bottom:1px solid #111318}
 </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>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>Cost</th><th>Latency</th><th>Status</th></tr></thead><tbody id="runs"></tbody></table></div>
+<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>
 <script>
 const $ = id => document.getElementById(id);
 function fmt(n,d=2){return typeof n==='number'?n.toFixed(d):'-'}
@@ -1989,12 +2099,13 @@ 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);
@@ -2021,9 +2132,26 @@ async function load(){
     ).join('')||'<tr><td colspan=4 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'}
-    $('runs').innerHTML=(runsR.runs||[]).map(r=>
-      '<tr><td>'+fmtTime(r.started_at)+'</td><td>'+r.model+'</td><td class="col-tt"><span class="badge '+ttCls(r.taskType)+'">'+(r.taskType||'general').replace(/_/g,' ')+'</span></td><td class="col-cx"><span class="badge '+cxCls(r.complexity)+'">'+(r.complexity||'simple')+'</span></td><td>'+(r.tokensIn||0)+'</td><td>'+(r.tokensOut||0)+'</td><td>$'+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=9 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;')}
+    $('runs').innerHTML=(runsR.runs||[]).map((r,i)=>{
+      const row='<tr style="cursor:pointer" onclick="toggleDetail('+i+')" title="Click to expand"><td>'+fmtTime(r.started_at)+'</td><td>'+r.model+'</td><td class="col-tt"><span class="badge '+ttCls(r.taskType)+'">'+(r.taskType||'general').replace(/_/g,' ')+'</span></td><td class="col-cx"><span class="badge '+cxCls(r.complexity)+'">'+(r.complexity||'simple')+'</span></td><td>'+(r.tokensIn||0)+'</td><td>'+(r.tokensOut||0)+'</td><td class="col-cache" style="color:#60a5fa">'+(r.cacheCreationTokens||0)+'</td><td class="col-cache" style="color:#34d399">'+(r.cacheReadTokens||0)+'</td><td>$'+fmt(r.costUsd,4)+'</td><td>'+r.latencyMs+'ms</td><td><span class="badge '+(r.status==='success'?'ok':'err')+'">'+r.status+'</span></td></tr>';
+      const c=r.requestContent||{};
+      let detail='<tr id="run-detail-'+i+'" style="display:none"><td colspan="11" 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>';
+        detail+='<button onclick="event.stopPropagation();loadFullResponse(\''+r.id+'\','+i+')" id="full-btn-'+i+'" style="background:#1e293b;color:#e2e8f0;border:1px solid #334155;padding:6px 12px;border-radius:6px;cursor:pointer;font-size:.8rem">Show full response</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=11 style="color:#64748b">No runs yet</td></tr>';
+    const agents=(agentsR.agents||[]).sort((a,b)=>(b.totalCost||0)-(a.totalCost||0));
+    $('agents').innerHTML=agents.length?agents.map(a=>
+      '<tr><td><span class="agent-name" data-fp="'+a.fingerprint+'">'+a.name+'</span> <button class="rename-btn" onclick="renameAgent(\''+a.fingerprint+'\',\''+a.name.replace(/'/g,"\\'")+'\')">✏️</button></td><td>'+a.totalRequests+'</td><td>$'+fmt(a.totalCost,4)+'</td><td>'+fmtTime(a.lastSeen)+'</td><td style="font-size:.7rem;color:#64748b" title="'+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)+'%'):'';
@@ -2031,8 +2159,27 @@ 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();
+}
+function toggleDetail(i){var d=document.getElementById('run-detail-'+i);d.style.display=d.style.display==='none'?'table-row':'none'}
+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>
@@ -2129,8 +2276,10 @@ 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);
@@ -2139,6 +2288,7 @@ 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 {
@@ -2516,7 +2666,9 @@ 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-4-6', r.tokensIn, r.tokensOut);
+                    // Savings should reflect routing decisions only — pass same cache tokens to baseline
+                    // so the cache discount doesn't get counted as "savings from routing"
+                    const origCost = (0, telemetry_js_1.estimateCost)('claude-opus-4-6', r.tokensIn, r.tokensOut, r.cacheCreationTokens || undefined, r.cacheReadTokens || undefined);
                     const perRunSavings = Math.max(0, origCost - r.costUsd);
                     return {
                         id: r.id,
@@ -2536,8 +2688,16 @@ async function startProxy(config = {}) {
                         latencyMs: r.latencyMs,
                         tokensIn: r.tokensIn,
                         tokensOut: r.tokensOut,
+                        cacheCreationTokens: r.cacheCreationTokens ?? 0,
+                        cacheReadTokens: r.cacheReadTokens ?? 0,
                         savings: Math.round(perRunSavings * 10000) / 10000,
                         escalated: r.escalated,
+                        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' });
@@ -2553,7 +2713,9 @@ async function startProxy(config = {}) {
                 let totalSavedAmount = 0;
                 const byDayMap = new Map();
                 for (const r of requestHistory) {
-                    const origCost = (0, telemetry_js_1.estimateCost)(OPUS_BASELINE, r.tokensIn, r.tokensOut);
+                    // 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;
@@ -2633,6 +2795,63 @@ 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' });
@@ -2713,6 +2932,14 @@ 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) {
@@ -3151,7 +3378,7 @@ async function startProxy(config = {}) {
                                 model: targetModel || requestedModel,
                                 tokensIn: nativeUsage?.input_tokens ?? 0,
                                 tokensOut: nativeUsage?.output_tokens ?? 0,
-                                costUsd: (0, telemetry_js_1.estimateCost)(targetModel || requestedModel, nativeUsage?.input_tokens ?? 0, nativeUsage?.output_tokens ?? 0),
+                                costUsd: (0, telemetry_js_1.estimateCost)(targetModel || requestedModel, nativeUsage?.input_tokens ?? 0, nativeUsage?.output_tokens ?? 0, nativeUsage?.cache_creation_input_tokens || undefined, nativeUsage?.cache_read_input_tokens || undefined),
                                 taskType,
                             });
                             log(`Cache STORE for ${targetModel || requestedModel} (hash: ${cacheHash.slice(0, 8)})`);
@@ -3174,7 +3401,22 @@ 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);
+                // 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) {
@@ -3183,6 +3425,10 @@ async function startProxy(config = {}) {
                         prompt: promptText.slice(0, 500),
                         taskType,
                         model: `${targetProvider}:${targetModel || requestedModel}`,
+                    })
+                        .then((runResult) => {
+                        // Backfill token/cost data — relay.run() has no adapters so records NULLs
+                        relay.patchRunTokens(runResult.runId, nativeTokIn, nativeTokOut, nativeCostUsd);
                     })
                         .catch(() => { });
                     sendCloudTelemetry(taskType, targetModel || requestedModel, nativeTokIn, nativeTokOut, durationMs, true, undefined, originalModel ?? undefined, nativeCacheCreation || undefined, nativeCacheRead || undefined);
@@ -3268,6 +3514,14 @@ 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;
@@ -3568,7 +3822,7 @@ 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) {
@@ -3605,8 +3859,12 @@ async function startProxy(config = {}) {
                     const cascadeUsage = responseData?.usage;
                     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, cascadeTokensIn, cascadeTokensOut);
-                    updateLastHistoryEntry(cascadeTokensIn, cascadeTokensOut, cascadeCost, chatCascadeRespModel);
+                    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, chatAgentFingerprint, chatExplicitAgentId);
+                    if (chatAgentFingerprint && chatAgentFingerprint !== 'unknown')
+                        (0, agent_tracker_js_1.updateAgentCost)(chatAgentFingerprint, cascadeCost);
                     if (recordTelemetry) {
                         try {
                             const runResult = await relay.run({
@@ -3614,6 +3872,8 @@ async function startProxy(config = {}) {
                                 taskType,
                                 model: `${cascadeResult.provider}:${cascadeResult.model}`,
                             });
+                            // Backfill token/cost data — relay.run() has no adapters so records NULLs
+                            relay.patchRunTokens(runResult.runId, cascadeTokensIn, cascadeTokensOut, cascadeCost);
                             responseData['_relayplane'] = {
                                 runId: runResult.runId,
                                 routedTo: `${cascadeResult.provider}/${cascadeResult.model}`,
@@ -3628,7 +3888,7 @@ async function startProxy(config = {}) {
                         catch (err) {
                             log(`Failed to record run: ${err}`);
                         }
-                        sendCloudTelemetry(taskType, cascadeResult.model, cascadeTokensIn, cascadeTokensOut, durationMs, true, undefined, originalRequestedModel ?? undefined);
+                        sendCloudTelemetry(taskType, cascadeResult.model, cascadeTokensIn, cascadeTokensOut, durationMs, true, undefined, originalRequestedModel ?? undefined, cascadeCacheCreation, cascadeCacheRead);
                         meshCapture(cascadeResult.model, cascadeResult.provider, taskType, cascadeTokensIn, cascadeTokensOut, cascadeCost, durationMs, true);
                     }
                     const chatCascadeRpHeaders = buildRelayPlaneResponseHeaders(cascadeResult.model, originalRequestedModel ?? 'unknown', complexity, cascadeResult.provider, 'cascade');
@@ -3649,7 +3909,7 @@ 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);
             }
         }
     });
@@ -3792,7 +4052,7 @@ 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) {
@@ -3845,9 +4105,11 @@ async function handleStreamingRequest(res, request, targetProvider, targetModel,
         'Connection': 'keep-alive',
         ...streamRpHeaders,
     });
-    // Track token usage from streaming events
+    // Track token usage from streaming events (including Anthropic prompt cache tokens)
     let streamTokensIn = 0;
     let streamTokensOut = 0;
+    let streamCacheCreation = 0;
+    let streamCacheRead = 0;
     const shouldCacheStream = !!(cacheHash && !cacheBypass);
     const rawChunks = [];
     try {
@@ -3859,7 +4121,8 @@ async function handleStreamingRequest(res, request, targetProvider, targetModel,
                     res.write(chunk);
                     if (shouldCacheStream)
                         rawChunks.push(chunk);
-                    // Parse OpenAI-format chunks for usage (emitted at end of stream)
+                    // Parse OpenAI-format chunks for usage — the converter embeds
+                    // cache_creation_tokens and cache_read_tokens from message_start.
                     try {
                         const lines = chunk.split('\n');
                         for (const line of lines) {
@@ -3868,6 +4131,8 @@ 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_tokens ?? streamCacheCreation;
+                                    streamCacheRead = evt.usage.cache_read_tokens ?? streamCacheRead;
                                 }
                             }
                         }
@@ -3927,13 +4192,13 @@ async function handleStreamingRequest(res, request, targetProvider, targetModel,
         const streamPayload = JSON.stringify({
             _relayplaneStreamCache: true,
             ssePayload: rawChunks.join(''),
-            usage: { input_tokens: streamTokensIn, output_tokens: streamTokensOut, prompt_tokens: streamTokensIn, completion_tokens: streamTokensOut },
+            usage: { input_tokens: streamTokensIn, output_tokens: streamTokensOut, prompt_tokens: streamTokensIn, completion_tokens: streamTokensOut, cache_creation_input_tokens: streamCacheCreation, cache_read_input_tokens: streamCacheRead },
         });
         responseCache.set(cacheHash, streamPayload, {
             model: targetModel,
             tokensIn: streamTokensIn,
             tokensOut: streamTokensOut,
-            costUsd: (0, telemetry_js_1.estimateCost)(targetModel, streamTokensIn, streamTokensOut),
+            costUsd: (0, telemetry_js_1.estimateCost)(targetModel, streamTokensIn, streamTokensOut, streamCacheCreation || undefined, streamCacheRead || undefined),
             taskType,
         });
         log(`Cache STORE (stream) for chat/completions ${targetModel} (hash: ${cacheHash.slice(0, 8)})`);
@@ -3944,9 +4209,11 @@ async function handleStreamingRequest(res, request, targetProvider, targetModel,
     const durationMs = Date.now() - startTime;
     // Always log the request for stats/telemetry tracking
     logRequest(request.model ?? 'unknown', targetModel, targetProvider, durationMs, true, routingMode, undefined, taskType, complexity);
-    // Update token/cost info on the history entry
-    const streamCost = (0, telemetry_js_1.estimateCost)(targetModel, streamTokensIn, streamTokensOut);
-    updateLastHistoryEntry(streamTokensIn, streamTokensOut, streamCost);
+    // Update token/cost info on the history entry (with cache token discount)
+    const streamCost = (0, telemetry_js_1.estimateCost)(targetModel, streamTokensIn, streamTokensOut, streamCacheCreation || undefined, streamCacheRead || undefined);
+    updateLastHistoryEntry(streamTokensIn, streamTokensOut, streamCost, undefined, streamCacheCreation || undefined, streamCacheRead || undefined, agentFingerprint, agentId);
+    if (agentFingerprint && agentFingerprint !== 'unknown')
+        (0, agent_tracker_js_1.updateAgentCost)(agentFingerprint, streamCost);
     // ── Post-request: budget spend + anomaly detection ──
     try {
         (0, budget_js_1.getBudgetManager)().recordSpend(streamCost, targetModel);
@@ -3967,12 +4234,14 @@ async function handleStreamingRequest(res, request, targetProvider, targetModel,
             model: `${targetProvider}:${targetModel}`,
         })
             .then((runResult) => {
+            // Backfill token/cost data — relay.run() has no adapters so records NULLs
+            relay.patchRunTokens(runResult.runId, streamTokensIn, streamTokensOut, streamCost);
             log(`Completed streaming in ${durationMs}ms, runId: ${runResult.runId}`);
         })
             .catch((err) => {
             log(`Failed to record run: ${err}`);
         });
-        sendCloudTelemetry(taskType, targetModel, streamTokensIn, streamTokensOut, durationMs, true, undefined, request.model ?? undefined);
+        sendCloudTelemetry(taskType, targetModel, streamTokensIn, streamTokensOut, durationMs, true, undefined, request.model ?? undefined, streamCacheCreation || undefined, streamCacheRead || undefined);
         meshCapture(targetModel, targetProvider, taskType, streamTokensIn, streamTokensOut, streamCost, durationMs, true);
     }
     res.end();
@@ -3980,7 +4249,7 @@ 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);
@@ -4015,12 +4284,16 @@ async function handleNonStreamingRequest(res, request, targetProvider, targetMod
     const nonStreamRespModel = checkResponseModelMismatch(responseData, targetModel, targetProvider, log);
     // Log the successful request
     logRequest(request.model ?? 'unknown', targetModel, targetProvider, durationMs, true, routingMode, undefined, taskType, complexity);
-    // Update token/cost info
+    // Update token/cost info (including Anthropic prompt cache tokens)
     const usage = responseData?.usage;
     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, tokensIn, tokensOut);
-    updateLastHistoryEntry(tokensIn, tokensOut, cost, nonStreamRespModel);
+    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, 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);
@@ -4040,6 +4313,8 @@ async function handleNonStreamingRequest(res, request, targetProvider, targetMod
                 taskType,
                 model: `${targetProvider}:${targetModel}`,
             });
+            // Backfill token/cost data — relay.run() has no adapters so records NULLs
+            relay.patchRunTokens(runResult.runId, tokensIn, tokensOut, cost);
             // Add routing metadata to response
             responseData['_relayplane'] = {
                 runId: runResult.runId,
@@ -4054,12 +4329,14 @@ async function handleNonStreamingRequest(res, request, targetProvider, targetMod
         catch (err) {
             log(`Failed to record run: ${err}`);
         }
-        // Extract token counts from response if available (Anthropic/OpenAI format)
-        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);
-        meshCapture(targetModel, targetProvider, taskType, tokensIn, tokensOut, cost, durationMs, true);
+        // Extract token counts from response if available (Anthropic/OpenAI format, including cache)
+        const innerUsage = responseData?.usage;
+        const innerTokIn = innerUsage?.input_tokens ?? innerUsage?.prompt_tokens ?? 0;
+        const innerTokOut = innerUsage?.output_tokens ?? innerUsage?.completion_tokens ?? 0;
+        const innerCacheCreation = innerUsage?.cache_creation_input_tokens ?? 0;
+        const innerCacheRead = innerUsage?.cache_read_input_tokens ?? 0;
+        sendCloudTelemetry(taskType, targetModel, innerTokIn, innerTokOut, durationMs, true, undefined, undefined, innerCacheCreation || undefined, innerCacheRead || undefined);
+        meshCapture(targetModel, targetProvider, taskType, innerTokIn, innerTokOut, cost, durationMs, true);
     }
     // ── Cache: store non-streaming chat/completions response ──
     const chatRespCache = (0, response_cache_js_1.getResponseCache)();