@nxuss/lemma 0.5.2 → 0.5.4

package/dist/esm/cli/lemma-proxy.js

@@ -24,7 +24,7 @@ function getVersion() {
             catch { }
         }
     }
-    return '0.5.2';
+    return '0.5.3';
 }
 const VERSION = getVersion();
 import ComplexityRouter from '../proxy/ComplexityRouter';
@@ -78,13 +78,20 @@ async function readJson(file, fallback) {
     catch { }
     return fallback;
 }
+const writeQueues = new Map();
 async function writeJson(file, data) {
-    try {
-        const tempFile = file + '.tmp';
-        await fs.promises.writeFile(tempFile, JSON.stringify(data, null, 2));
-        await fs.promises.rename(tempFile, file);
-    }
-    catch { }
+    const chain = writeQueues.get(file) || Promise.resolve();
+    const writeOperation = async () => {
+        try {
+            const tempFile = `${file}.${crypto.randomUUID ? crypto.randomUUID() : Math.random().toString(36).substring(2, 15)}.tmp`;
+            await fs.promises.writeFile(tempFile, JSON.stringify(data, null, 2));
+            await fs.promises.rename(tempFile, file);
+        }
+        catch { }
+    };
+    const nextChain = chain.then(writeOperation).catch(() => { });
+    writeQueues.set(file, nextChain);
+    return nextChain;
 }
 function projectHash(name) {
     return crypto.createHash('sha1').update(name).digest('hex').slice(0, 12);
@@ -378,12 +385,21 @@ export function startBackgroundClipboardWatcher() {
     }
 }
 // ── License helpers ────────────────────────────────────────────────────────────
+let isProCached = null;
+let lastLicenseCheck = 0;
+const LICENSE_CACHE_TTL = 10000; // 10 seconds cache
 async function loadLicense() {
     return readJson(LICENSE_FILE, { isPro: false });
 }
 async function isPro() {
+    const now = Date.now();
+    if (isProCached !== null && (now - lastLicenseCheck) < LICENSE_CACHE_TTL) {
+        return isProCached;
+    }
     const l = await loadLicense();
-    return !!(l && l.isPro);
+    isProCached = !!(l && l.isPro);
+    lastLicenseCheck = now;
+    return isProCached;
 }
 async function validateKeyRemote(key) {
     return new Promise((resolve, reject) => {
@@ -446,16 +462,21 @@ async function recordStat(stats, project, fromCache, latencyMs, provider, tokens
     s.total++;
     s.totalLatency += latencyMs;
     fromCache ? s.hits++ : s.misses++;
-    s.totalTokensSaved += (fromCache ? (tokensSaved || 2000) : (tokensSaved || 0));
+    const tkSaved = fromCache ? (tokensSaved || 2000) : (tokensSaved || 0);
+    s.totalTokensSaved += tkSaved;
     if (!s.providers[provider])
         s.providers[provider] = { hits: 0, misses: 0 };
     fromCache ? s.providers[provider].hits++ : s.providers[provider].misses++;
+    if (fromCache && tkSaved > 0) {
+        // Automatically report cache hits directly to the server session ledger for absolute sync consistency
+        savingsLedger.recordTokens('cache', tkSaved);
+    }
     logEvent({
         type: fromCache ? 'cache:hit' : 'cache:miss',
         project,
         latency: latencyMs,
         provider,
-        tokens: fromCache ? (tokensSaved || 2000) : (tokensSaved || 0)
+        tokens: tkSaved
     });
     await writeJson(STATS_FILE, stats);
 }
@@ -476,7 +497,15 @@ function cacheGet(provider, prompt) {
     return e;
 }
 function cacheSet(provider, prompt, data) {
-    CACHE.set(cacheKey(provider, prompt), { input: prompt, data, createdAt: Date.now(), ttl: TTL_MS });
+    const MAX_CACHE_SIZE = 5000;
+    const k = cacheKey(provider, prompt);
+    if (!CACHE.has(k) && CACHE.size >= MAX_CACHE_SIZE) {
+        // Evict oldest cache entry (FIFO eviction)
+        const oldestKey = CACHE.keys().next().value;
+        if (oldestKey)
+            CACHE.delete(oldestKey);
+    }
+    CACHE.set(k, { input: prompt, data, createdAt: Date.now(), ttl: TTL_MS, provider });
 }
 // ── Semantic Search (Pro) ──────────────────────────────────────────────────────
 async function getEmbedding(text) {
@@ -521,7 +550,7 @@ async function semanticGet(provider, prompt) {
                 distance = distances[0][i];
             }
             const similarity = Math.max(0, 1 - distance);
-            const THRESHOLD = 0.7;
+            const THRESHOLD = 0.82; // raised from 0.7 to 0.82 to prevent false positives
             if (similarity >= THRESHOLD) {
                 return {
                     data: JSON.parse(metadata.response),
@@ -748,6 +777,16 @@ async function relayMissStream(res, body, provider, tier) {
         req.end();
     });
 }
+class WriteQueue {
+    constructor() {
+        this.queue = Promise.resolve();
+    }
+    async enqueue(fn) {
+        this.queue = this.queue.then(fn).catch(() => { });
+        return this.queue;
+    }
+}
+const usageWriteQueue = new WriteQueue();
 // ── Proxy Server class ─────────────────────────────────────────────────────────
 class LemmaServer {
     constructor(port, projectName, openBrowser = true) {
@@ -810,6 +849,18 @@ class LemmaServer {
                 const hitRate = stats.total > 0 ? stats.hits / stats.total : 0;
                 const avgLat = stats.total > 0 ? stats.totalLatency / stats.total : 0;
                 const secretsMasked = EVENT_LOG.filter(e => e.type === 'privacy:mask').length;
+                const ledgerSnap = savingsLedger.getSnapshot();
+                let activeAgents = 1; // Always include the Gateway itself
+                let totalAgents = 1;
+                // Query the stack orchestration API to merge active agents
+                try {
+                    const stackMetrics = await axios.get('http://127.0.0.1:8083/api/metrics', { timeout: 100 });
+                    if (stackMetrics.data) {
+                        activeAgents += stackMetrics.data.activeAgents || 0;
+                        totalAgents += stackMetrics.data.totalAgents || 0;
+                    }
+                }
+                catch { }
                 res.json({
                     timestamp: new Date().toISOString(),
                     totalRequests: stats.total,
@@ -817,30 +868,140 @@ class LemmaServer {
                     cacheMisses: stats.misses,
                     hitRate,
                     averageLatency: avgLat,
-                    tokensSaved: stats.totalTokensSaved,
-                    costSaved: (stats.totalTokensSaved / 1000) * 0.02,
+                    tokensSaved: ledgerSnap.total.tokensSaved,
+                    costSaved: ledgerSnap.total.costSaved,
                     secretsMasked: secretsMasked,
-                    activeAgents: 1,
-                    totalAgents: 1,
+                    activeAgents,
+                    totalAgents,
                     uptime: process.uptime(),
                     tier: (await isPro()) ? 'pro' : 'free'
                 });
             });
             this.app.get('/api/agents', async (req, res) => {
+                // Load statistics for the Gateway itself
                 const stats = await loadStats();
-                res.json({ agents: [
-                        { id: 'lemma-gateway', name: 'Lemma Gateway', status: 'active', capabilities: ['Privacy Firewall', 'Semantic Cache', 'Complexity Router'], tasksCompleted: stats[this.projectName]?.total || 0, lastSeen: new Date().toISOString(), errorCount: 0, cacheHitRate: 0.85 }
-                    ] });
+                const statsProj = stats[this.projectName] || { total: 0, hits: 0 };
+                const total = statsProj.total || 0;
+                const hits = statsProj.hits || 0;
+                const gatewayAgent = {
+                    id: 'lemma-gateway',
+                    name: 'Lemma Gateway',
+                    status: 'active',
+                    capabilities: ['Privacy Firewall', 'Semantic Cache', 'Complexity Router'],
+                    tasksCompleted: total,
+                    lastSeen: new Date().toISOString(),
+                    errorCount: 0,
+                    cacheHitRate: total > 0 ? hits / total : 0
+                };
+                let agentsList = [gatewayAgent];
+                // Query the stack orchestration API to merge dynamic websocket agents if running
+                try {
+                    const stackAgentsResponse = await axios.get('http://127.0.0.1:8083/api/agents', { timeout: 100 });
+                    if (stackAgentsResponse.data && Array.isArray(stackAgentsResponse.data.agents)) {
+                        const stackAgents = stackAgentsResponse.data.agents.filter((a) => a.id !== 'lemma-gateway');
+                        agentsList = [...agentsList, ...stackAgents];
+                    }
+                }
+                catch { }
+                res.json({ agents: agentsList });
             });
-            this.app.get('/api/events', (req, res) => {
+            this.app.get('/api/events', async (req, res) => {
                 const page = parseInt(req.query.page) || 1;
                 const limit = parseInt(req.query.limit) || 50;
                 const start = (page - 1) * limit;
-                const events = EVENT_LOG.slice().reverse().slice(start, start + limit);
-                res.json({ events, pagination: { total: EVENT_LOG.length, page, limit } });
+                // Get local proxy events (normalized schema matching what dashboard expects)
+                const localEvents = EVENT_LOG.map(e => ({
+                    id: e.id,
+                    timestamp: new Date(e.timestamp).toISOString(),
+                    type: e.type, // e.g., 'cache:hit' or 'cache:miss' which client normalizes to 'cache_hit' or 'cache_miss'
+                    latency: e.latency,
+                    tokensSaved: e.tokens,
+                    costSaved: e.tokens ? (e.tokens / 1000) * 0.002 : 0,
+                    metadata: {
+                        provider: e.provider,
+                        project: e.project
+                    }
+                }));
+                let mergedEvents = [...localEvents];
+                // Query the stack orchestration API if running to merge stack events
+                try {
+                    const stackEventsResponse = await axios.get(`http://127.0.0.1:8083/api/events?page=1&limit=100`, { timeout: 100 });
+                    if (stackEventsResponse.data && Array.isArray(stackEventsResponse.data.events)) {
+                        mergedEvents = [...mergedEvents, ...stackEventsResponse.data.events];
+                    }
+                }
+                catch { }
+                // Sort merged events chronologically (newest first)
+                mergedEvents.sort((a, b) => {
+                    const tA = new Date(a.timestamp).getTime();
+                    const tB = new Date(b.timestamp).getTime();
+                    return tB - tA;
+                });
+                const total = mergedEvents.length;
+                const paginatedEvents = mergedEvents.slice(start, start + limit);
+                res.json({
+                    events: paginatedEvents,
+                    pagination: {
+                        total,
+                        page,
+                        limit,
+                        hasMore: start + limit < total
+                    }
+                });
             });
-            this.app.get('/api/cache-stats', (req, res) => {
-                res.json({ topQueries: [] });
+            this.app.get('/api/cache-stats', async (req, res) => {
+                let topQueries = [];
+                let totalCached = CACHE.size;
+                // Try to query ChromaDB for semantic cache count
+                try {
+                    const collection = await chroma.getOrCreateCollection({
+                        name: 'lemma-cache',
+                        embeddingFunction: dummyEmbeddingFunction,
+                        metadata: { "hnsw:space": "cosine" }
+                    });
+                    const count = await collection.count();
+                    totalCached += count;
+                }
+                catch { }
+                // Build top queries from in-memory CACHE
+                for (const [key, entry] of CACHE.entries()) {
+                    const hitCount = EVENT_LOG.filter(e => e.type === 'cache:hit' && e.provider === entry.provider).length || 1;
+                    const tokens = 2000;
+                    topQueries.push({
+                        queryHash: key.substring(0, 16),
+                        query: entry.input,
+                        hitCount,
+                        totalTokens: tokens * hitCount,
+                        costSaved: (tokens * hitCount / 1000) * 0.002,
+                        lastHit: new Date(entry.createdAt).toISOString()
+                    });
+                }
+                // Try to query the stack orchestration API if running to append stack cache stats
+                try {
+                    const stackCache = await axios.get('http://127.0.0.1:8083/api/cache-stats', { timeout: 100 });
+                    if (stackCache.data) {
+                        totalCached += stackCache.data.totalCached || 0;
+                        if (Array.isArray(stackCache.data.topQueries)) {
+                            const stackTop = stackCache.data.topQueries.filter((q) => !topQueries.some(t => t.query === q.query));
+                            topQueries = [...topQueries, ...stackTop];
+                        }
+                    }
+                }
+                catch { }
+                // Sort by hitCount desc
+                topQueries.sort((a, b) => b.hitCount - a.hitCount);
+                // Latency percentiles from EVENT_LOG
+                const latencies = EVENT_LOG.filter(e => e.latency).map(e => e.latency).sort((a, b) => a - b);
+                const latencyPercentiles = {
+                    p50: latencies[Math.floor(latencies.length * 0.5)] || 0,
+                    p95: latencies[Math.floor(latencies.length * 0.95)] || 0,
+                    p99: latencies[Math.floor(latencies.length * 0.99)] || 0,
+                };
+                res.json({
+                    topQueries: topQueries.slice(0, 10),
+                    latencyPercentiles,
+                    totalCached
+                });
             });
             this.app.get('/api/search', async (req, res) => {
                 const query = req.query.q;
@@ -877,24 +1038,57 @@ class LemmaServer {
                     res.status(500).json({ error: e.message });
                 }
             });
-            this.app.post('/api/cache/clear', (req, res) => {
-                const count = CACHE.size;
-                CACHE.clear();
-                EVENT_LOG.length = 0; // Clear live event log
-                savingsLedger.reset(); // Reset savings counters
-                // Reset and persist statistics for the current project
-                if (this.stats[this.projectName]) {
-                    this.stats[this.projectName] = {
-                        total: 0,
-                        hits: 0,
-                        misses: 0,
-                        totalLatency: 0,
-                        totalTokensSaved: 0,
-                        providers: {}
-                    };
-                    writeJson(STATS_FILE, this.stats).catch(() => { });
+            this.app.post('/api/cache/clear', async (req, res) => {
+                const { provider } = req.body || {};
+                let clearedCount = 0;
+                if (!provider) {
+                    // Global clear
+                    clearedCount = CACHE.size;
+                    CACHE.clear();
+                    EVENT_LOG.length = 0; // Clear live event log
+                    savingsLedger.reset(); // Reset savings counters
+                    // Reset and persist statistics for the current project
+                    if (this.stats[this.projectName]) {
+                        this.stats[this.projectName] = {
+                            total: 0,
+                            hits: 0,
+                            misses: 0,
+                            totalLatency: 0,
+                            totalTokensSaved: 0,
+                            providers: {}
+                        };
+                        writeJson(STATS_FILE, this.stats).catch(() => { });
+                    }
+                    // Clear Chroma DB collection
+                    try {
+                        await chroma.deleteCollection({ name: 'lemma-cache' }).catch(() => { });
+                    }
+                    catch (e) {
+                        console.error(`[CacheClear] Failed to clear Chroma collection: ${e.message}`);
+                    }
                 }
-                res.json({ success: true, clearedEntries: count });
+                else {
+                    // Granular clear
+                    for (const [k, e] of CACHE.entries()) {
+                        if (e.provider === provider) {
+                            CACHE.delete(k);
+                            clearedCount++;
+                        }
+                    }
+                    // Selective delete from Chroma DB collection using provider metadata filter
+                    try {
+                        const collection = await chroma.getOrCreateCollection({
+                            name: 'lemma-cache',
+                            embeddingFunction: dummyEmbeddingFunction,
+                            metadata: { "hnsw:space": "cosine" }
+                        });
+                        await collection.delete({ where: { provider } });
+                    }
+                    catch (e) {
+                        console.error(`[CacheClear] Failed to selectively delete provider ${provider} from Chroma: ${e.message}`);
+                    }
+                }
+                res.json({ success: true, clearedEntries: clearedCount });
             });
             // Savings breakdown by source for the dashboard
             this.app.get('/api/savings-breakdown', (req, res) => {
@@ -938,89 +1132,35 @@ class LemmaServer {
             logEvent({ type: 'privacy:mask', count: Object.keys(tokenMap).length, project: this.projectName });
         }
         const originalModel = req.body.model || 'gpt-4o';
-        const routingDecision = complexityRouter.evaluate(prompt, originalModel);
-        const model = routingDecision.model;
-        if (req.body.model)
-            req.body.model = model;
-        // Record complexity routing savings when model is downgraded
-        if (routingDecision.complexity === 'low' && model !== originalModel) {
-            // Estimated token savings: cheap model costs ~5x less, so avg prompt of ~800 tokens saves ~640 tokens
-            const estimatedPromptTokens = Math.floor(prompt.length / 4);
-            const routingSavings = Math.floor(estimatedPromptTokens * 0.8); // conservative 80% cost reduction
-            savingsLedger.recordTokens('complexityRouting', routingSavings);
-        }
-        // Apply the Agent Multiplexer to local-verify the prompt in parallel
-        const enrichedPrompt = await agentMultiplexer.multiplexPrompt(prompt, model);
-        // Resolve conversation session key and retrieve/create the fileHistory map
-        let fileHistory;
-        if (req.body.messages && req.body.messages.length > 0) {
-            try {
-                const firstMsg = req.body.messages[0];
-                const content = typeof firstMsg.content === 'string' ? firstMsg.content : JSON.stringify(firstMsg.content);
-                const sessionKey = crypto.createHash('sha1').update(content).digest('hex');
-                if (!this.sessions.has(sessionKey)) {
-                    this.sessions.set(sessionKey, { fileHistory: new Map() });
-                }
-                fileHistory = this.sessions.get(sessionKey).fileHistory;
-            }
-            catch { }
-        }
-        // Compress long TS/JS codebase file blocks in the enriched prompt using the Context Squeezer AST Tree-Shaker!
-        // The query context is the user's primary prompt to ensure we keep relevant methods!
-        const squeezeResult = squeezePrompt(enrichedPrompt, prompt, fileHistory);
-        const squeezedPrompt = squeezeResult.squeezed;
-        const compressionSavings = squeezeResult.originalSize - squeezeResult.squeezedSize;
-        if (compressionSavings > 0) {
-            savingsLedger.record('contextSqueeze', squeezeResult.originalSize, squeezeResult.squeezedSize);
-            console.log(`⚡ [Squeezer] Compressed prompt from ${squeezeResult.originalSize} to ${squeezeResult.squeezedSize} chars (Saved ${(squeezeResult.compressionRatio * 100).toFixed(1)}% context tokens!)`);
-        }
-        if (provider === 'openai' || provider === 'anthropic') {
-            if (req.body.messages && req.body.messages.length > 0) {
-                injectSqueezedContent(req.body.messages[req.body.messages.length - 1], squeezedPrompt);
-                // Apply conversational history pruning to save tokens on historical messages!
-                const pruneResult = pruneHistoryMessages(req.body.messages);
-                req.body.messages = pruneResult.messages;
-                historyTokensSaved = pruneResult.tokensSavedEstimate + Math.floor(compressionSavings / 4);
-                if (pruneResult.tokensSavedEstimate > 0) {
-                    savingsLedger.recordTokens('historyPrune', pruneResult.tokensSavedEstimate);
-                }
-            }
-        }
-        else if (provider === 'gemini') {
-            if (req.body.contents && req.body.contents.length > 0) {
-                const lastContent = req.body.contents[req.body.contents.length - 1];
-                if (lastContent.parts && lastContent.parts.length > 0) {
-                    lastContent.parts[lastContent.parts.length - 1].text = squeezedPrompt;
-                }
-            }
-        }
+        // ── 1. Cache-First Hot Path ──
+        // Perform exact cache check immediately (resolves in <1ms)
         const hit = cacheGet(provider, prompt);
         if (hit) {
-            const cacheTokensSaved = 2000 + historyTokensSaved;
-            savingsLedger.recordTokens('cache', cacheTokensSaved);
+            const cacheTokensSaved = 2000;
             await recordStat(this.stats, this.projectName, true, Date.now() - t0, provider, cacheTokensSaved);
             const unmaskedData = semanticScrubber.unmask(hit.data, tokenMap);
             if (isStream)
-                return simulateHitStream(res, unmaskedData, provider, model, tier, 1.0);
+                return simulateHitStream(res, unmaskedData, provider, originalModel, tier, 1.0);
             res.setHeader('X-Lemma-Cache', 'HIT');
             res.setHeader('X-Lemma-Similarity', '1.000');
             res.setHeader('X-Lemma-Tier', tier);
             return res.json(unmaskedData);
         }
+        // Perform semantic cache check (resolves in ~15ms)
         if (pro) {
             const semHit = await semanticGet(provider, prompt);
             if (semHit) {
                 if (semHit.similarity >= 0.90) {
-                    await recordStat(this.stats, this.projectName, true, Date.now() - t0, provider, 2000 + historyTokensSaved);
+                    await recordStat(this.stats, this.projectName, true, Date.now() - t0, provider, 2000);
                     const unmaskedData = semanticScrubber.unmask(semHit.data, tokenMap);
                     if (isStream)
-                        return simulateHitStream(res, unmaskedData, provider, model, tier, semHit.similarity);
+                        return simulateHitStream(res, unmaskedData, provider, originalModel, tier, semHit.similarity);
                     res.setHeader('X-Lemma-Cache', 'HIT');
                     res.setHeader('X-Lemma-Similarity', semHit.similarity.toFixed(3));
                     res.setHeader('X-Lemma-Tier', tier);
                     return res.json(unmaskedData);
                 }
-                else if (semHit.similarity >= 0.70) {
+                else if (semHit.similarity >= 0.82) { // raised from 0.70 to 0.82 to prevent false positives
                     console.log(`💡 [CARS] Near Cache Hit (Similarity: ${(semHit.similarity * 100).toFixed(1)}%). Synthesizing answer locally via Ollama...`);
                     let cachedContent = '';
                     if (provider === 'openai') {
@@ -1058,10 +1198,10 @@ Adjusted Answer:`;
                                 else if (provider === 'anthropic') {
                                     synthesizedData.content[0].text = adjustedText;
                                 }
-                                await recordStat(this.stats, this.projectName, true, Date.now() - t0, provider, 2000 + historyTokensSaved);
+                                await recordStat(this.stats, this.projectName, true, Date.now() - t0, provider, 2000);
                                 const unmaskedData = semanticScrubber.unmask(synthesizedData, tokenMap);
                                 if (isStream) {
-                                    return simulateHitStream(res, unmaskedData, provider, model, tier, semHit.similarity);
+                                    return simulateHitStream(res, unmaskedData, provider, originalModel, tier, semHit.similarity);
                                 }
                                 res.setHeader('X-Lemma-Cache', 'HIT-CARS');
                                 res.setHeader('X-Lemma-Similarity', semHit.similarity.toFixed(3));
@@ -1078,22 +1218,86 @@ Adjusted Answer:`;
             const cloudHit = await cloudSync.get(prompt);
             if (cloudHit) {
                 cacheSet(provider, prompt, cloudHit.data);
-                await recordStat(this.stats, this.projectName, true, Date.now() - t0, provider, 2000 + historyTokensSaved);
+                await recordStat(this.stats, this.projectName, true, Date.now() - t0, provider, 2000);
                 const unmaskedData = semanticScrubber.unmask(cloudHit.data, tokenMap);
                 if (isStream)
-                    return simulateHitStream(res, unmaskedData, provider, model, tier, cloudHit.similarity || 0.95);
+                    return simulateHitStream(res, unmaskedData, provider, originalModel, tier, cloudHit.similarity || 0.95);
                 res.setHeader('X-Lemma-Cache', 'HIT');
                 res.setHeader('X-Lemma-Similarity', (cloudHit.similarity || 0.95).toFixed(3));
                 res.setHeader('X-Lemma-Tier', tier);
                 return res.json(unmaskedData);
             }
         }
+        // ── 2. Cache Miss Path ──
+        const routingDecision = complexityRouter.evaluate(prompt, originalModel);
+        const model = routingDecision.model;
+        if (req.body.model)
+            req.body.model = model;
+        // Record complexity routing savings when model is downgraded
+        if (routingDecision.complexity === 'low' && model !== originalModel) {
+            const estimatedPromptTokens = Math.floor(prompt.length / 4);
+            const routingSavings = Math.floor(estimatedPromptTokens * 0.8);
+            savingsLedger.recordTokens('complexityRouting', routingSavings);
+        }
+        // Apply the Agent Multiplexer to local-verify the prompt in parallel
+        const enrichedPrompt = await agentMultiplexer.multiplexPrompt(prompt, model);
+        // Resolve conversation session key and retrieve/create the fileHistory map
+        let fileHistory;
+        if (req.body.messages && req.body.messages.length > 0) {
+            try {
+                const firstMsg = req.body.messages[0];
+                const content = typeof firstMsg.content === 'string' ? firstMsg.content : JSON.stringify(firstMsg.content);
+                const sessionKey = crypto.createHash('sha1').update(content).digest('hex');
+                if (!this.sessions.has(sessionKey)) {
+                    // Enforce 500 active sessions cap to prevent memory leak (FIFO eviction)
+                    const MAX_SESSIONS = 500;
+                    if (this.sessions.size >= MAX_SESSIONS) {
+                        const oldestKey = this.sessions.keys().next().value;
+                        if (oldestKey)
+                            this.sessions.delete(oldestKey);
+                    }
+                    this.sessions.set(sessionKey, { fileHistory: new Map() });
+                }
+                fileHistory = this.sessions.get(sessionKey).fileHistory;
+            }
+            catch { }
+        }
+        // Compress long TS/JS codebase file blocks in the enriched prompt
+        const squeezeResult = squeezePrompt(enrichedPrompt, prompt, fileHistory);
+        const squeezedPrompt = squeezeResult.squeezed;
+        const compressionSavings = squeezeResult.originalSize - squeezeResult.squeezedSize;
+        if (compressionSavings > 0) {
+            savingsLedger.record('contextSqueeze', squeezeResult.originalSize, squeezeResult.squeezedSize);
+            console.log(`⚡ [Squeezer] Compressed prompt from ${squeezeResult.originalSize} to ${squeezeResult.squeezedSize} chars (Saved ${(squeezeResult.compressionRatio * 100).toFixed(1)}% context tokens!)`);
+        }
+        if (provider === 'openai' || provider === 'anthropic') {
+            if (req.body.messages && req.body.messages.length > 0) {
+                injectSqueezedContent(req.body.messages[req.body.messages.length - 1], squeezedPrompt);
+                // Apply conversational history pruning
+                const pruneResult = pruneHistoryMessages(req.body.messages);
+                req.body.messages = pruneResult.messages;
+                historyTokensSaved = pruneResult.tokensSavedEstimate + Math.floor(compressionSavings / 4);
+                if (pruneResult.tokensSavedEstimate > 0) {
+                    savingsLedger.recordTokens('historyPrune', pruneResult.tokensSavedEstimate);
+                }
+            }
+        }
+        else if (provider === 'gemini') {
+            if (req.body.contents && req.body.contents.length > 0) {
+                const lastContent = req.body.contents[req.body.contents.length - 1];
+                if (lastContent.parts && lastContent.parts.length > 0) {
+                    lastContent.parts[lastContent.parts.length - 1].text = squeezedPrompt;
+                }
+            }
+        }
         if (isStream) {
             await recordStat(this.stats, this.projectName, false, Date.now() - t0, provider, historyTokensSaved);
             if (!pro) {
-                const u = await loadUsage();
-                u.count++;
-                await writeJson(USAGE_FILE, u);
+                await usageWriteQueue.enqueue(async () => {
+                    const u = await loadUsage();
+                    u.count++;
+                    await writeJson(USAGE_FILE, u);
+                });
             }
             return relayMissStream(res, req.body, provider, tier);
         }
@@ -1105,9 +1309,11 @@ Adjusted Answer:`;
                 await cloudSync.set(prompt, data);
             }
             if (!pro) {
-                const u = await loadUsage();
-                u.count++;
-                await writeJson(USAGE_FILE, u);
+                await usageWriteQueue.enqueue(async () => {
+                    const u = await loadUsage();
+                    u.count++;
+                    await writeJson(USAGE_FILE, u);
+                });
             }
             await recordStat(this.stats, this.projectName, false, Date.now() - t0, provider, historyTokensSaved);
             return res.json(semanticScrubber.unmask(data, tokenMap));
@@ -1118,6 +1324,43 @@ Adjusted Answer:`;
     }
     async callUpstream(body, provider) {
         let apiKey = provider === 'openai' ? process.env.OPENAI_API_KEY : (provider === 'anthropic' ? process.env.ANTHROPIC_API_KEY : process.env.GEMINI_API_KEY);
+        // If the API key is an OpenAI key (sk-proj-) but requested for Anthropic or Gemini,
+        // multiplex it internally to OpenAI gpt-4o-mini to allow live, cheap simulated provider routing.
+        const isMockedProvider = apiKey && apiKey.startsWith('sk-') && provider !== 'openai';
+        if (isMockedProvider) {
+            console.log(`\n\x1b[36m🔀 [Gateway Router] Multiplexing B2B ${provider.toUpperCase()} request to OpenAI using key: ...${apiKey.substring(8, 14)}...\x1b[0m`);
+            const payload = {
+                model: 'gpt-4o-mini',
+                messages: body.messages || [{ role: 'user', content: extractPrompt(body, provider) || 'Hello' }],
+                stream: false // keep stream false for simulation simplicity
+            };
+            const baseURL = 'https://api.openai.com/v1/chat/completions';
+            const resp = await axios.post(baseURL, payload, {
+                headers: {
+                    Authorization: `Bearer ${apiKey}`,
+                    'Content-Type': 'application/json'
+                }
+            });
+            if (provider === 'anthropic') {
+                const text = resp.data.choices?.[0]?.message?.content || '';
+                return {
+                    id: `msg-${Date.now()}`,
+                    type: 'message',
+                    role: 'assistant',
+                    content: [{ type: 'text', text }],
+                    model: body.model || 'claude-3-5-sonnet'
+                };
+            }
+            else if (provider === 'gemini') {
+                const text = resp.data.choices?.[0]?.message?.content || '';
+                return {
+                    candidates: [{
+                            content: { parts: [{ text }] },
+                            finishReason: 'STOP'
+                        }]
+                };
+            }
+        }
         if (!apiKey) {
             console.log(`\n\x1b[33m⚠️  No API Key found for ${provider.toUpperCase()}. Falling back to local Ollama (Zero-Cost Mode)!\x1b[0m`);
             // If the requested model is a cloud model (e.g. gpt-4o, claude), map it to a local model like llama3
@@ -1144,6 +1387,15 @@ Adjusted Answer:`;
     }
     async start() {
         this.stats = await loadStats();
+        // Start active cache eviction sweep every hour
+        setInterval(() => {
+            const now = Date.now();
+            for (const [k, e] of CACHE.entries()) {
+                if (e.ttl > 0 && now - e.createdAt > e.ttl) {
+                    CACHE.delete(k);
+                }
+            }
+        }, 3600000).unref();
         const server = http.createServer(this.app);
         server.on('error', (e) => {
             if (e.code === 'EADDRINUSE') {