lynkr 9.7.2 → 9.9.0

package/src/server.js

@@ -37,13 +37,10 @@ const lazyLoader = require("./tools/lazy-loader");
 const { setLazyLoader } = require("./tools");
 const { waitForOllama } = require("./clients/ollama-startup");
 
-// Initialize MCP
 initialiseMcp();
 
-// Set up lazy tool loading
 setLazyLoader(lazyLoader);
 
-// Check if lazy loading is enabled (default: true)
 const LAZY_TOOLS_ENABLED = process.env.LAZY_TOOLS_ENABLED !== "false";
 
 if (LAZY_TOOLS_ENABLED) {
@@ -88,16 +85,13 @@ function createApp() {
   app.get('/dashboard/api/routing',  require('./dashboard/api').routing);
   app.get('/dashboard/api/logs',     require('./dashboard/api').logs);
 
-  // Initialize load shedder (log configuration)
   initializeLoadShedder();
 
-  // Load shedding (protect against overload)
   app.use(loadSheddingMiddleware);
 
   // Request logging (add request IDs, structured logs)
   app.use(requestLoggingMiddleware);
 
-  // Metrics collection
   app.use(metricsMiddleware);
 
   // Note: If using a tunnel (ngrok, Cloudflare Tunnel) and seeing BrotliDecompressionError,
@@ -107,9 +101,21 @@ function createApp() {
   app.use(sessionMiddleware);
   app.use(loggingMiddleware);
 
+  // Agent-framework endpoints: LangGraph/CrewAI/AutoGen talk to the
+  // OpenAI-compat surface, Claude Code to /v1/messages. Guards that protect
+  // against runaway agents must cover BOTH — mounting on /v1/messages only
+  // silently exempts every OpenAI-compat client (this bit us with client
+  // profiles already).
+  const AGENT_ENDPOINTS = ['/v1/messages', '/v1/chat/completions', '/v1/responses'];
+
+  // Loop guard — stateless runaway-agent circuit breaker
+  // (LYNKR_MAX_SESSION_TURNS / LYNKR_MAX_TOOL_TURNS; off unless set).
+  const { loopGuard } = require('./api/middleware/loop-guard');
+  for (const p of AGENT_ENDPOINTS) app.use(p, loopGuard);
+
   // Budget and rate limiting (can be disabled via config)
   if (config.budget?.enabled !== false) {
-    app.use('/v1/messages', budgetMiddleware);
+    for (const p of AGENT_ENDPOINTS) app.use(p, budgetMiddleware);
   }
 
   // Phase 6.1 — per-tenant routing policies (LYNKR-Tenant-Id header).
@@ -117,13 +123,11 @@ function createApp() {
   app.use('/v1/messages', tenantMiddleware);
 
   // Phase 6.2 — hierarchical budget enforcement (LYNKR_BUDGET_ENFORCER=false to disable).
-  app.use('/v1/messages', budgetEnforcer);
+  for (const p of AGENT_ENDPOINTS) app.use(p, budgetEnforcer);
 
-  // Health check endpoints
   app.get("/health/live", livenessCheck);
   app.get("/health/ready", readinessCheck);
 
-  // Metrics endpoints
   app.get("/metrics", (req, res) => {
     res.json(metrics.snapshot());
   });
@@ -176,10 +180,8 @@ function createApp() {
 
   app.use(router);
 
-  // Dashboard UI
   app.use('/dashboard', require('./dashboard/router'));
 
-  // Files API
   const filesRouter = require("./api/files-router");
   app.use("/v1", filesRouter);
 
@@ -193,8 +195,7 @@ function createApp() {
 }
 
 async function start() {
-  // Initialize Worker Thread Pool (if enabled)
-  // This pre-warms worker threads for CPU-intensive tasks
+  // Pre-warms worker threads for CPU-intensive tasks
   if (config.workerPool?.enabled !== false) {
     try {
       const poolOptions = {
@@ -245,17 +246,65 @@ async function start() {
     console.log(`Claude→Databricks proxy listening on http://localhost:${config.port}`);
   });
 
-  // Start session cleanup manager
+  // Start session cleanup manager. It also drives routing-side maintenance
+  // (telemetry retention + session-pin TTL) via its runCleanup tick — see
+  // src/sessions/cleanup.js.
   const { getSessionCleanupManager } = require("./sessions/cleanup");
   const sessionCleanup = getSessionCleanupManager();
   sessionCleanup.start();
 
-  // Setup graceful shutdown
+  // WS5.7 — auto-calibration scheduler.
+  //
+  // Recomputes `data/calibrated-thresholds.json` from live telemetry every
+  // 24 h and hot-reloads the model-tier selector so the update takes
+  // effect without a restart. First run is jittered 30–90 min into the
+  // process lifetime so a cluster of proxies won't all recalibrate at the
+  // same moment. When telemetry is sparse (<100 rows in-window), the
+  // calibration step no-ops itself via `runCalibration`'s
+  // `insufficient_samples` skip path. Unconditionally armed — the pre-B
+  // `LYNKR_AUTO_CALIBRATE` env flag was removed because the "off" state
+  // was never useful in prod (calibration is idempotent and self-gating).
+  {
+    const { runCalibration } = require('./routing/calibration');
+    const { reloadCalibratedThresholds } = require('./routing/model-tiers');
+    const degradation = require('./routing/degradation');
+    const DAY_MS = 24 * 60 * 60 * 1000;
+    const FIRST_RUN_MIN_MS = 30 * 60 * 1000;
+    const FIRST_RUN_JITTER_MS = 60 * 60 * 1000;
+    const firstDelay = FIRST_RUN_MIN_MS + Math.floor(Math.random() * FIRST_RUN_JITTER_MS);
+
+    const runOnce = () => {
+      try {
+        const result = runCalibration({ days: 7 });
+        if (result.skipped) {
+          logger.info({ reason: result.reason, count: result.count }, '[Calibration] Skipped');
+          return;
+        }
+        const before = reloadCalibratedThresholds();
+        logger.info({
+          calibratedAt: result.calibratedAt,
+          sampleCount: result.sampleCount,
+          ranges: result.ranges,
+          previousRanges: before,
+        }, '[Calibration] Ranges refreshed');
+      } catch (err) {
+        degradation.record('calibration', err);
+      }
+    };
+
+    const firstTimer = setTimeout(() => {
+      runOnce();
+      const interval = setInterval(runOnce, DAY_MS);
+      interval.unref();
+    }, firstDelay);
+    firstTimer.unref();
+    logger.info({ firstDelayMs: firstDelay, intervalMs: DAY_MS }, '[Calibration] Scheduler armed');
+  }
+
   const shutdownManager = getShutdownManager();
   shutdownManager.registerServer(server);
   shutdownManager.setupSignalHandlers();
 
-  // Register Headroom shutdown callback
   if (config.headroom?.enabled) {
     shutdownManager.onShutdown(async () => {
       logger.info("Stopping Headroom sidecar on shutdown");
@@ -263,7 +312,6 @@ async function start() {
     });
   }
 
-  // Register Worker Pool shutdown callback
   if (config.workerPool?.enabled !== false && isWorkerPoolReady()) {
     shutdownManager.onShutdown(async () => {
       logger.info("Stopping worker thread pool on shutdown");
@@ -272,7 +320,6 @@ async function start() {
     });
   }
 
-  // Register Codex process shutdown callback
   shutdownManager.onShutdown(async () => {
     try {
       const { getCodexProcess } = require("./clients/codex-process");
@@ -283,7 +330,6 @@ async function start() {
     } catch { /* ignore if codex never started */ }
   });
 
-  // Initialize hot reload config watcher
   if (config.hotReload?.enabled !== false) {
     const watcher = initConfigWatcher({
       paths: [".env"],
@@ -300,7 +346,6 @@ async function start() {
       }
     });
 
-    // Stop watcher on shutdown
     shutdownManager.onShutdown(() => {
       const w = getConfigWatcher();
       if (w) w.stop();

package/src/sessions/cleanup.js

@@ -32,6 +32,23 @@ class SessionCleanupManager {
     } catch (error) {
       logger.error({ error }, "Session cleanup failed");
     }
+
+    // Routing-side maintenance piggybacks on the same tick — both live in
+    // .lynkr/telemetry.db, both are unbounded without a scheduled prune,
+    // and neither had a caller before this hook. WS5 auto-calibration will
+    // hang off this same manager.
+    try {
+      const telemetry = require("../routing/telemetry");
+      const affinityStore = require("../routing/affinity-store");
+      const pinTtlMs = Number(process.env.LYNKR_STICKY_TTL_MS) || 6 * 60 * 60 * 1000;
+      const telemetryDeleted = telemetry.cleanup();
+      const pinsDeleted = affinityStore.cleanup(pinTtlMs);
+      if (telemetryDeleted || pinsDeleted) {
+        logger.debug({ telemetryDeleted, pinsDeleted }, "Routing telemetry/pin cleanup completed");
+      }
+    } catch (err) {
+      logger.debug({ err: err.message }, "Routing cleanup tick failed (non-fatal)");
+    }
   }
 
   stop() {

package/src/tools/index.js

@@ -1,6 +1,5 @@
 const logger = require("../logger");
 const { truncateToolOutput } = require("./truncate");
-const { isGPTProvider, formatToolResultForGPT } = require("../clients/gpt-utils");
 
 const registry = new Map();
 const registryLowercase = new Map();
@@ -258,19 +257,7 @@ async function executeToolCall(call, context = {}) {
     );
     const formatted = normalizeHandlerResult(result);
 
-    // Apply tool output truncation for token efficiency
-    let truncatedContent = truncateToolOutput(normalisedCall.name, formatted.content);
-
-    // GPT-specific formatting temporarily disabled for testing
-    // const isGPT = context?.provider && isGPTProvider(context.provider);
-    // if (isGPT) {
-    //   truncatedContent = formatToolResultForGPT(
-    //     normalisedCall.name,
-    //     truncatedContent,
-    //     normalisedCall.arguments
-    //   );
-    // }
-    const isGPT = false; // Disabled for testing
+    const truncatedContent = truncateToolOutput(normalisedCall.name, formatted.content);
 
     return {
       id: normalisedCall.id,
@@ -284,7 +271,6 @@ async function executeToolCall(call, context = {}) {
         truncated: truncatedContent !== formatted.content,
         originalLength: formatted.content?.length,
         truncatedLength: truncatedContent?.length,
-        gptFormatted: isGPT,
       },
     };
   } catch (err) {

package/src/tools/smart-selection.js

@@ -122,6 +122,14 @@ function estimateToolTokens(tools) {
 /**
  * Apply conservative stripping to the tool list.
  */
+function recordStrippingSavings(before, after) {
+  if (after >= before) return;
+  try {
+    const telemetry = require('../routing/telemetry');
+    telemetry.recordSavings('tool_stripping', (before - after) * 175);
+  } catch { /* telemetry is best-effort */ }
+}
+
 function selectToolsSmartly(tools, classification, options = {}) {
   if (!Array.isArray(tools) || tools.length === 0) return tools;
 
@@ -130,6 +138,7 @@ function selectToolsSmartly(tools, classification, options = {}) {
 
   // Greeting: strip everything
   if (classification.type === 'conversational') {
+    recordStrippingSavings(tools.length, 0);
     return [];
   }
 
@@ -159,6 +168,7 @@ function selectToolsSmartly(tools, classification, options = {}) {
     selected = selected.slice(0, 10);
   }
 
+  recordStrippingSavings(tools.length, selected.length);
   return selected;
 }
 

package/src/tools/web-client.js

@@ -1,4 +1,4 @@
-const { Agent, setGlobalDispatcher } = require("undici");
+const { Agent, fetch: undiciFetch, setGlobalDispatcher } = require("undici");
 const logger = require("../logger");
 
 /**
@@ -42,10 +42,10 @@ function createWebAgent() {
 const webAgent = createWebAgent();
 
 /**
- * Fetch with the optimized agent
+ * Fetch with the optimized agent (uses undici.fetch for dispatcher support)
  */
 async function fetchWithAgent(url, options = {}) {
-  return fetch(url, {
+  return undiciFetch(url, {
     ...options,
     dispatcher: webAgent,
   });

package/.eslintrc.cjs

@@ -1,12 +0,0 @@
-module.exports = {
-  env: {
-    node: true,
-    es2021: true,
-  },
-  extends: ["eslint:recommended"],
-  parserOptions: {
-    ecmaVersion: "latest",
-    sourceType: "script",
-  },
-  rules: {},
-};

package/benchmark-configs/litellm_config.yaml

@@ -1,86 +0,0 @@
-# ─── LiteLLM Benchmark Config ─────────────────────────────────────────────────
-# Multi-provider tier routing via LiteLLM Complexity Router.
-#
-# Start: litellm --config benchmark-configs/litellm_config.yaml --port 8082
-#
-# Required env vars:
-#   AZURE_OPENAI_API_KEY
-#   AZURE_OPENAI_ENDPOINT   (https://YOUR-RESOURCE.openai.azure.com)
-#   MOONSHOT_API_KEY
-#   (Ollama needs no key — running locally on :11434)
-#
-# Tier mapping (matches Lynkr benchmark config):
-#   SIMPLE    → ollama:minimax-m2.5:cloud
-#   MEDIUM    → ollama:minimax-m2.5:cloud
-#   COMPLEX   → moonshot:moonshot-v1-auto
-#   REASONING → azure-openai:gpt-5.2-chat
-
-model_list:
-
-  # ── SIMPLE + MEDIUM → Ollama minimax-m2.5:cloud ───────────────────────────
-  # Note: the model tag is "minimax-m2.5:cloud" — the colon is part of the
-  # Ollama model name, NOT a provider separator here.
-  - model_name: smart-router
-    litellm_params:
-      model: "ollama/minimax-m2.5:cloud"
-      api_base: http://localhost:11434
-  - model_name: smart-router
-    litellm_params:
-      model: "ollama/minimax-m2.5:cloud"
-      api_base: http://localhost:11434
-
-  # ── COMPLEX → Moonshot moonshot-v1-auto (matches Lynkr TIER_COMPLEX) ────────
-  - model_name: smart-router
-    litellm_params:
-      model: openai/moonshot-v1-auto
-      api_base: https://api.moonshot.ai/v1
-      api_key: os.environ/MOONSHOT_API_KEY
-
-  # ── REASONING → Azure OpenAI gpt-5.2-chat ─────────────────────────────────
-  - model_name: smart-router
-    litellm_params:
-      model: azure/gpt-5.2-chat
-      api_base: os.environ/AZURE_OPENAI_ENDPOINT
-      api_key: os.environ/AZURE_OPENAI_API_KEY
-      api_version: "2024-12-01-preview"
-
-  # ── Direct aliases (for targeted calls outside the benchmark) ─────────────
-  - model_name: ollama-minimax
-    litellm_params:
-      model: "ollama/minimax-m2.5:cloud"
-      api_base: http://localhost:11434
-
-  - model_name: moonshot-kimi-k2
-    litellm_params:
-      model: openai/moonshot-v1-auto
-      api_base: https://api.moonshot.ai/v1
-      api_key: os.environ/MOONSHOT_API_KEY
-
-  - model_name: azure-gpt5
-    litellm_params:
-      model: azure/gpt-5.2-chat
-      api_base: os.environ/AZURE_OPENAI_ENDPOINT
-      api_key: os.environ/AZURE_OPENAI_API_KEY
-      api_version: "2024-12-01-preview"
-
-router_settings:
-  routing_strategy: cost-based-routing
-  # Fallback: if smart-router fails on one deployment, try the next
-  fallbacks:
-    - smart-router:
-        - ollama-minimax
-        - moonshot-kimi-k2
-        - azure-gpt5
-  num_retries: 2
-  timeout: 90
-
-litellm_settings:
-  drop_params: true
-  use_responses_api: false
-  return_response_headers: true
-  success_callback: []
-  failure_callback: []
-
-general_settings:
-  master_key: sk-1234        # change this
-  port: 8082

package/benchmark-configs/lynkr.env

@@ -1,48 +0,0 @@
-# ─── Lynkr Benchmark Config ───────────────────────────────────────────────────
-# Multi-provider tier routing: Ollama → Moonshot → Azure OpenAI
-# Copy to .env and fill in your credentials.
-
-PORT=8081
-
-# ── Ollama (local, free) ───────────────────────────────────────────────────────
-OLLAMA_ENDPOINT=http://localhost:11434
-OLLAMA_MODEL=qwen2.5-coder:7b
-OLLAMA_TIMEOUT_MS=120000
-OLLAMA_EMBEDDINGS_MODEL=nomic-embed-text
-OLLAMA_EMBEDDINGS_ENDPOINT=http://localhost:11434/api/embeddings
-
-# ── Azure OpenAI ───────────────────────────────────────────────────────────────
-AZURE_OPENAI_ENDPOINT=https://YOUR-RESOURCE.openai.azure.com
-AZURE_OPENAI_API_KEY=your-azure-openai-key
-AZURE_OPENAI_DEPLOYMENT=gpt-4o
-AZURE_OPENAI_API_VERSION=2024-08-01-preview
-
-# ── Moonshot (Kimi) ────────────────────────────────────────────────────────────
-MOONSHOT_API_KEY=your-moonshot-api-key
-MOONSHOT_ENDPOINT=https://api.moonshot.ai/v1/chat/completions
-MOONSHOT_MODEL=kimi-k2-turbo-preview
-
-# ── Primary provider (Lynkr uses this when no tier matches) ───────────────────
-# Set to whichever you want as the default fallback
-MODEL_PROVIDER=azure-openai
-
-# ── Tier Routing ───────────────────────────────────────────────────────────────
-# SIMPLE   → Ollama local (free)
-# MEDIUM   → Moonshot Kimi (cheap, fast)
-# COMPLEX  → Azure OpenAI GPT-4o (powerful)
-# REASONING→ Azure OpenAI o3-mini (best reasoning)
-TIER_SIMPLE=ollama:minimax-m2.5:cloud
-TIER_MEDIUM=ollama:minimax-m2.5:cloud
-TIER_COMPLEX=moonshot:kimi-k2.6
-TIER_REASONING=azure-openai:gpt-5.2-chat
-
-# ── Token Optimisations (these are what LiteLLM/Portkey don't have) ────────────
-SMART_TOOL_SELECTION=true
-PROMPT_CACHE_ENABLED=true
-SEMANTIC_CACHE_ENABLED=true
-SEMANTIC_CACHE_THRESHOLD=0.95
-HISTORY_COMPRESSION_ENABLED=true
-TOOL_INJECTION_ENABLED=false
-
-# ── Optional: make routing decisions visible in responses ──────────────────────
-LYNKR_VISIBLE_ROUTING=true

package/benchmark-configs/portkey-config.json

@@ -1,60 +0,0 @@
-{
-  "_comment": "Portkey Gateway Config — multi-provider conditional routing",
-  "_note": "Portkey has NO automatic complexity detection. This config uses max_tokens as a proxy for complexity. For real tier routing pass x-portkey-metadata: { 'tier': 'simple|medium|complex|reasoning' } from your client.",
-
-  "strategy": {
-    "mode": "conditional"
-  },
-
-  "conditions": [
-    {
-      "_comment": "SIMPLE — short requests (max_tokens <= 256) → Ollama",
-      "condition": {
-        "query.max_tokens": { "$lte": 256 }
-      },
-      "target": {
-        "provider": "ollama",
-        "customHost": "http://localhost:11434",
-        "override_params": {
-          "model": "qwen2.5-coder:7b"
-        }
-      }
-    },
-    {
-      "_comment": "MEDIUM — metadata tier=medium → Moonshot",
-      "condition": {
-        "metadata.tier": { "$eq": "medium" }
-      },
-      "target": {
-        "provider": "openai",
-        "apiKey": "{{MOONSHOT_API_KEY}}",
-        "baseURL": "https://api.moonshot.ai/v1",
-        "override_params": {
-          "model": "moonshot-v1-8k"
-        }
-      }
-    },
-    {
-      "_comment": "REASONING — metadata tier=reasoning → Azure OpenAI o3-mini",
-      "condition": {
-        "metadata.tier": { "$eq": "reasoning" }
-      },
-      "target": {
-        "provider": "azure-openai",
-        "apiKey": "{{AZURE_OPENAI_API_KEY}}",
-        "resourceName": "YOUR-RESOURCE",
-        "deploymentId": "o3-mini",
-        "apiVersion": "2024-12-01-preview"
-      }
-    }
-  ],
-
-  "default": {
-    "_comment": "COMPLEX — everything else → Azure OpenAI GPT-4o",
-    "provider": "azure-openai",
-    "apiKey": "{{AZURE_OPENAI_API_KEY}}",
-    "resourceName": "YOUR-RESOURCE",
-    "deploymentId": "gpt-4o",
-    "apiVersion": "2024-08-01-preview"
-  }
-}

package/benchmark-configs/portkey-docker.sh

@@ -1,23 +0,0 @@
-#!/bin/bash
-# Run Portkey local AI Gateway with provider credentials injected
-
-docker run -d \
-  --name portkey-gateway \
-  -p 8083:8787 \
-  -e AZURE_OPENAI_API_KEY="${AZURE_OPENAI_API_KEY}" \
-  -e MOONSHOT_API_KEY="${MOONSHOT_API_KEY}" \
-  portkeyai/gateway:latest
-
-echo "Portkey gateway running on http://localhost:8083"
-echo ""
-echo "To use Azure OpenAI directly (no tier routing):"
-echo "  curl http://localhost:8083/v1/chat/completions \\"
-echo "    -H 'x-portkey-provider: azure-openai' \\"
-echo "    -H 'x-portkey-api-key: \$PORTKEY_API_KEY' \\"
-echo "    -H 'x-portkey-azure-resource-name: YOUR-RESOURCE' \\"
-echo "    -H 'x-portkey-azure-deployment-id: gpt-4o' \\"
-echo "    -H 'x-portkey-azure-api-version: 2024-08-01-preview' \\"
-echo "    -d '{\"model\": \"gpt-4o\", \"messages\": [...]}'"
-echo ""
-echo "To use conditional routing config, pass:"
-echo "  -H 'x-portkey-config: <base64-encoded portkey-config.json>'"