lynkr 9.0.1 → 9.1.2

package/src/clients/xml-tool-extractor.js

@@ -0,0 +1,307 @@
+/**
+ * Universal Tool Call Extractor
+ *
+ * Extracts tool calls embedded as raw text (XML, JSON, custom tokens)
+ * from LLM output.  Covers Minimax, Hermes/Qwen, Qwen3-Coder, GLM,
+ * Llama, Mistral, DeepSeek, GPT-OSS, and generic formats.
+ *
+ * @module clients/xml-tool-extractor
+ */
+
+const logger = require("../logger");
+
+let callCounter = 0;
+function nextId() {
+  return `call_extracted_${Date.now()}_${callCounter++}`;
+}
+
+function tryParseJSON(str) {
+  try { return JSON.parse(str.trim()); } catch { return null; }
+}
+
+// ── Individual extractors ────────────────────────────────────────────
+
+/** 1. Minimax: <invoke name="X"><parameter name="K">V</parameter></invoke> */
+function extractMinimax(text) {
+  const calls = [];
+  const re = /<invoke\s+name="([^"]+)">([\s\S]*?)<\/invoke>/g;
+  const paramRe = /<parameter\s+name="([^"]+)">([\s\S]*?)<\/parameter>/g;
+  let m;
+  while ((m = re.exec(text)) !== null) {
+    const name = m[1];
+    const body = m[2];
+    const args = {};
+    let pm;
+    while ((pm = paramRe.exec(body)) !== null) {
+      let val = pm[2].trim();
+      const parsed = tryParseJSON(val);
+      args[pm[1]] = parsed !== null ? parsed : val;
+    }
+    paramRe.lastIndex = 0;
+    calls.push({ name, arguments: args, _match: m[0] });
+  }
+  // Also strip wrapper tags
+  let cleaned = text;
+  if (calls.length > 0) {
+    for (const c of calls) cleaned = cleaned.replace(c._match, "");
+    cleaned = cleaned.replace(/<\/?minimax:tool_call>/g, "");
+  }
+  return { calls, cleaned };
+}
+
+/** 2. GLM: <tool_call>func_name <arg_key>k</arg_key> <arg_value>v</arg_value></tool_call> */
+function extractGLM(text) {
+  const calls = [];
+  const re = /<tool_call>([\s\S]*?)<\/tool_call>/g;
+  let m;
+  while ((m = re.exec(text)) !== null) {
+    const body = m[1].trim();
+    // Check if it's GLM style (has <arg_key> tags)
+    if (!body.includes("<arg_key>")) continue;
+    const nameMatch = body.match(/^(\S+)/);
+    if (!nameMatch) continue;
+    const name = nameMatch[1];
+    const args = {};
+    const kvRe = /<arg_key>([\s\S]*?)<\/arg_key>\s*<arg_value>([\s\S]*?)<\/arg_value>/g;
+    let kv;
+    while ((kv = kvRe.exec(body)) !== null) {
+      let val = kv[2].trim();
+      const parsed = tryParseJSON(val);
+      args[kv[1].trim()] = parsed !== null ? parsed : val;
+    }
+    calls.push({ name, arguments: args, _match: m[0] });
+  }
+  let cleaned = text;
+  for (const c of calls) cleaned = cleaned.replace(c._match, "");
+  return { calls, cleaned };
+}
+
+/** 3. Hermes/Qwen JSON: <tool_call>{"name":"X","arguments":{...}}</tool_call> */
+function extractHermesQwen(text) {
+  const calls = [];
+  const re = /<tool_call>\s*([\s\S]*?)\s*<\/tool_call>/g;
+  let m;
+  while ((m = re.exec(text)) !== null) {
+    const body = m[1].trim();
+    // Skip GLM format (handled above)
+    if (body.includes("<arg_key>")) continue;
+    // Skip Qwen3-Coder XML format
+    if (body.includes("<tool_name>")) continue;
+    const json = tryParseJSON(body);
+    if (json && (json.name || json.function)) {
+      const name = json.name || json.function?.name || json.function || "unknown";
+      const args = json.arguments || json.parameters || json.params || {};
+      calls.push({ name, arguments: typeof args === "string" ? tryParseJSON(args) || {} : args, _match: m[0] });
+    }
+  }
+  let cleaned = text;
+  for (const c of calls) cleaned = cleaned.replace(c._match, "");
+  return { calls, cleaned };
+}
+
+/** 4. Qwen3-Coder: <tool_call><tool_name>X</tool_name><parameter name="K">V</parameter></tool_call> */
+function extractQwenCoder(text) {
+  const calls = [];
+  const re = /<tool_call>\s*([\s\S]*?)\s*<\/tool_call>/g;
+  let m;
+  while ((m = re.exec(text)) !== null) {
+    const body = m[1].trim();
+    if (!body.includes("<tool_name>")) continue;
+    const nameMatch = body.match(/<tool_name>([\s\S]*?)<\/tool_name>/);
+    if (!nameMatch) continue;
+    const name = nameMatch[1].trim();
+    const args = {};
+    const paramRe = /<parameter\s+name="([^"]+)">([\s\S]*?)<\/parameter>/g;
+    let pm;
+    while ((pm = paramRe.exec(body)) !== null) {
+      let val = pm[2].trim();
+      const parsed = tryParseJSON(val);
+      args[pm[1]] = parsed !== null ? parsed : val;
+    }
+    calls.push({ name, arguments: args, _match: m[0] });
+  }
+  let cleaned = text;
+  for (const c of calls) cleaned = cleaned.replace(c._match, "");
+  return { calls, cleaned };
+}
+
+/** 5. DeepSeek: <｜tool▁call▁begin｜>function<｜tool▁sep｜>name\n```json\n{...}\n```\n<｜tool▁call▁end｜> */
+function extractDeepSeek(text) {
+  const calls = [];
+  // Match both Unicode and ASCII approximations
+  const re = /(?:<｜tool▁call▁begin｜>|<\|tool_call_begin\|>|<\|tool_call_start\|>)\s*(?:function)?\s*(?:<｜tool▁sep｜>|<\|tool_sep\|>)?\s*(\S+)\s*```(?:json)?\s*([\s\S]*?)```\s*(?:<｜tool▁call▁end｜>|<\|tool_call_end\|>)/g;
+  let m;
+  while ((m = re.exec(text)) !== null) {
+    const name = m[1].trim();
+    const json = tryParseJSON(m[2]);
+    if (name) {
+      calls.push({ name, arguments: json || {}, _match: m[0] });
+    }
+  }
+  let cleaned = text;
+  for (const c of calls) cleaned = cleaned.replace(c._match, "");
+  return { calls, cleaned };
+}
+
+/** 6. Mistral: [TOOL_CALLS] [{"name":"X","arguments":{...}}] */
+function extractMistral(text) {
+  const calls = [];
+  const re = /\[TOOL_CALLS\]\s*(\[[\s\S]*?\])/g;
+  let m;
+  while ((m = re.exec(text)) !== null) {
+    const arr = tryParseJSON(m[1]);
+    if (Array.isArray(arr)) {
+      for (const item of arr) {
+        if (item.name || item.function) {
+          const name = item.name || item.function?.name || "unknown";
+          const args = item.arguments || item.parameters || {};
+          calls.push({ name, arguments: typeof args === "string" ? tryParseJSON(args) || {} : args, _match: m[0] });
+        }
+      }
+    }
+  }
+  let cleaned = text;
+  for (const c of calls) cleaned = cleaned.replace(c._match, "");
+  return { calls, cleaned };
+}
+
+/** 7. Llama python_tag: <|python_tag|>{"name":"X","arguments":{...}} */
+function extractLlamaPythonTag(text) {
+  const calls = [];
+  const re = /<\|python_tag\|>\s*(\{[\s\S]*?\})(?:<\|eom_id\|>|<\|eot_id\|>|\s*$)/g;
+  let m;
+  while ((m = re.exec(text)) !== null) {
+    const json = tryParseJSON(m[1]);
+    if (json && (json.name || json.function)) {
+      const name = json.name || json.function || "unknown";
+      const args = json.arguments || json.parameters || {};
+      calls.push({ name, arguments: typeof args === "string" ? tryParseJSON(args) || {} : args, _match: m[0] });
+    }
+  }
+  let cleaned = text;
+  for (const c of calls) cleaned = cleaned.replace(c._match, "");
+  return { calls, cleaned };
+}
+
+/** 8. GPT-OSS Harmony: <|call|>name(key=value, ...) */
+function extractGptOss(text) {
+  const calls = [];
+  const re = /<\|call\|>\s*(\w+)\(([^)]*)\)/g;
+  let m;
+  while ((m = re.exec(text)) !== null) {
+    const name = m[1];
+    const argsStr = m[2].trim();
+    const args = {};
+    if (argsStr) {
+      // Parse key=value pairs
+      const kvRe = /(\w+)\s*=\s*(?:"([^"]*)"|'([^']*)'|(\S+))/g;
+      let kv;
+      while ((kv = kvRe.exec(argsStr)) !== null) {
+        args[kv[1]] = kv[2] ?? kv[3] ?? kv[4];
+      }
+    }
+    calls.push({ name, arguments: args, _match: m[0] });
+  }
+  let cleaned = text;
+  for (const c of calls) cleaned = cleaned.replace(c._match, "");
+  return { calls, cleaned };
+}
+
+/** 9. Generic <function_call>{...}</function_call> */
+function extractGenericFunctionCall(text) {
+  const calls = [];
+  const re = /<function_call>\s*([\s\S]*?)\s*<\/function_call>/g;
+  let m;
+  while ((m = re.exec(text)) !== null) {
+    const json = tryParseJSON(m[1]);
+    if (json && (json.name || json.function)) {
+      const name = json.name || json.function?.name || json.function || "unknown";
+      const args = json.arguments || json.parameters || {};
+      calls.push({ name, arguments: typeof args === "string" ? tryParseJSON(args) || {} : args, _match: m[0] });
+    }
+  }
+  let cleaned = text;
+  for (const c of calls) cleaned = cleaned.replace(c._match, "");
+  return { calls, cleaned };
+}
+
+/** 10. Raw JSON fallback: text starts with {"name":"...","arguments":{...}} */
+function extractRawJSON(text) {
+  const trimmed = text.trim();
+  if (!trimmed.startsWith("{")) return { calls: [], cleaned: text };
+  const json = tryParseJSON(trimmed);
+  if (!json || typeof json !== "object") return { calls: [], cleaned: text };
+  if (!json.name && !json.function) return { calls: [], cleaned: text };
+  const name = json.name || json.function?.name || json.function || "unknown";
+  const args = json.arguments || json.parameters || {};
+  return {
+    calls: [{ name, arguments: typeof args === "string" ? tryParseJSON(args) || {} : args, _match: trimmed }],
+    cleaned: "",
+  };
+}
+
+// ── Main entry point ─────────────────────────────────────────────────
+
+const EXTRACTORS = [
+  { name: "minimax", fn: extractMinimax },
+  { name: "glm", fn: extractGLM },
+  { name: "hermes_qwen", fn: extractHermesQwen },
+  { name: "qwen_coder", fn: extractQwenCoder },
+  { name: "deepseek", fn: extractDeepSeek },
+  { name: "mistral", fn: extractMistral },
+  { name: "llama_python", fn: extractLlamaPythonTag },
+  { name: "gpt_oss", fn: extractGptOss },
+  { name: "generic_function_call", fn: extractGenericFunctionCall },
+  { name: "raw_json", fn: extractRawJSON },
+];
+
+/**
+ * Extract tool calls from model text output.
+ * Tries all known patterns (most specific → most generic).
+ * Returns on first extractor that finds tool calls.
+ *
+ * @param {string} text - Raw model text content
+ * @returns {{ toolCalls: Array, cleanedText: string|null }}
+ */
+function extractToolCallsFromText(text) {
+  if (!text || typeof text !== "string") {
+    return { toolCalls: [], cleanedText: text };
+  }
+
+  for (const { name: extractorName, fn } of EXTRACTORS) {
+    try {
+      const { calls, cleaned } = fn(text);
+      if (calls.length > 0) {
+        const toolCalls = calls.map((c) => ({
+          id: nextId(),
+          type: "function",
+          function: {
+            name: c.name,
+            arguments: typeof c.arguments === "string" ? c.arguments : JSON.stringify(c.arguments),
+          },
+        }));
+
+        // Clean up stray tokens
+        let cleanedText = (cleaned || "")
+          .replace(/<\|eom_id\|>/g, "")
+          .replace(/<\|eot_id\|>/g, "")
+          .replace(/<\|end\|>/g, "")
+          .trim() || null;
+
+        logger.info({
+          extractor: extractorName,
+          toolCount: toolCalls.length,
+          tools: toolCalls.map((t) => t.function.name),
+        }, "Extracted tool calls from model text output");
+
+        return { toolCalls, cleanedText };
+      }
+    } catch (err) {
+      logger.debug({ extractor: extractorName, error: err.message }, "Tool call extractor failed, trying next");
+    }
+  }
+
+  return { toolCalls: [], cleanedText: text };
+}
+
+module.exports = { extractToolCallsFromText };

package/src/cluster.js

@@ -0,0 +1,82 @@
+/**
+ * Cluster Mode — Multi-Core Scaling
+ *
+ * Forks one worker per CPU core. Each worker runs a full Lynkr
+ * instance with its own Express server, event loop, and connection pool.
+ *
+ * Enable: CLUSTER_ENABLED=true (default: false for dev, recommended for prod)
+ * Workers: CLUSTER_WORKERS=auto (default) or a number
+ *
+ * Architecture:
+ *   Primary process → forks N workers → each worker calls start()
+ *   Primary handles: signal forwarding, worker respawning, health monitoring
+ *   Workers handle: HTTP requests, LLM proxying, tool execution
+ *
+ * Shared state considerations:
+ *   - SQLite: WAL mode supports concurrent readers across processes
+ *   - In-memory caches (prompt, circuit breaker): per-worker (not shared)
+ *   - Rate limiting: per-worker (sessions are sticky via round-robin)
+ *
+ * @module cluster
+ */
+
+const cluster = require('node:cluster');
+const os = require('node:os');
+
+const WORKER_COUNT = (() => {
+  const env = process.env.CLUSTER_WORKERS;
+  if (!env || env === 'auto') return Math.max(os.cpus().length - 1, 1);
+  const n = parseInt(env, 10);
+  return Number.isNaN(n) || n < 1 ? Math.max(os.cpus().length - 1, 1) : n;
+})();
+
+function startCluster() {
+  if (cluster.isPrimary) {
+    console.log(`[cluster] Primary ${process.pid} starting ${WORKER_COUNT} workers`);
+
+    // Fork workers
+    for (let i = 0; i < WORKER_COUNT; i++) {
+      cluster.fork();
+    }
+
+    // Respawn crashed workers
+    cluster.on('exit', (worker, code, signal) => {
+      if (signal) {
+        console.log(`[cluster] Worker ${worker.process.pid} killed by signal ${signal}`);
+      } else if (code !== 0) {
+        console.log(`[cluster] Worker ${worker.process.pid} exited with code ${code}, respawning...`);
+        cluster.fork();
+      } else {
+        console.log(`[cluster] Worker ${worker.process.pid} exited cleanly`);
+      }
+    });
+
+    // Forward SIGTERM/SIGINT to all workers for graceful shutdown
+    const shutdown = (sig) => {
+      console.log(`[cluster] Primary received ${sig}, shutting down workers...`);
+      for (const id in cluster.workers) {
+        cluster.workers[id].process.kill(sig);
+      }
+      // Give workers 10s to drain, then force exit
+      setTimeout(() => {
+        console.log('[cluster] Force exit after 10s drain timeout');
+        process.exit(0);
+      }, 10000).unref();
+    };
+
+    process.on('SIGTERM', () => shutdown('SIGTERM'));
+    process.on('SIGINT', () => shutdown('SIGINT'));
+
+    // Log worker status
+    cluster.on('online', (worker) => {
+      console.log(`[cluster] Worker ${worker.process.pid} online`);
+    });
+
+  } else {
+    // Worker process — start the normal Lynkr server
+    const { start } = require('./server');
+    start();
+  }
+}
+
+module.exports = { startCluster, WORKER_COUNT };

package/src/config/index.js

@@ -638,6 +638,9 @@ var config = {
     fallbackProvider,
   },
   toolExecutionMode,
+  toolResultCompression: {
+    enabled: true,
+  },
   server: {
     jsonLimit: process.env.REQUEST_JSON_LIMIT ?? "1gb",
   },
@@ -917,6 +920,13 @@ var config = {
     weightedScoring: true,
     costOptimization: true,
     agenticDetection: true,
+    // Embed an interaction block in the response body so the user can
+    // see *why* a particular tier/provider was chosen.
+    visibleInteraction: process.env.LYNKR_VISIBLE_ROUTING === 'true',
+    // Run user-supplied preflight commands before invoking the model.
+    // If all exit 0, short-circuit the request with zero LLM cost.
+    preflightEnabled: process.env.LYNKR_PREFLIGHT_ENABLED === 'true',
+    preflightTimeoutMs: Number(process.env.LYNKR_PREFLIGHT_TIMEOUT_MS) || 120000,
   },
 
   // Model Tier Configuration (REQUIRED)
@@ -929,6 +939,12 @@ var config = {
     REASONING: process.env.TIER_REASONING?.trim() || null,
   },
 
+  // Cluster mode (multi-core scaling for 50+ concurrent users)
+  cluster: {
+    enabled: process.env.CLUSTER_ENABLED === 'true',
+    workers: process.env.CLUSTER_WORKERS || 'auto',
+  },
+
   // Graphify knowledge graph integration (structural analysis)
   codeGraph: {
     enabled: process.env.CODE_GRAPH_ENABLED === 'true',

package/src/context/distill.js

@@ -71,9 +71,19 @@ function jaccardSimilarity(setA, setB) {
   return union === 0 ? 0 : intersection / union;
 }
 
+// Try to load native Rust implementation (3.7x faster for 100+ line blocks)
+let nativeSimilarity = null;
+try {
+  const native = require('../../native');
+  if (native.available && native.structuralSimilarity) {
+    nativeSimilarity = native.structuralSimilarity;
+  }
+} catch { /* native module not available — use JS */ }
+
 /**
  * Compute structural similarity between two text blocks.
  * Uses normalized line signatures + Jaccard index.
+ * Delegates to Rust native when available (3.7x faster).
  *
  * @param {string} a - First text
  * @param {string} b - Second text
@@ -83,6 +93,11 @@ function structuralSimilarity(a, b) {
   if (!a && !b) return 1;
   if (!a || !b) return 0;
 
+  // Use Rust for large inputs where the speedup offsets Napi boundary cost
+  if (nativeSimilarity && a.length + b.length > 500) {
+    return nativeSimilarity(a, b);
+  }
+
   const sigA = extractSignature(a);
   const sigB = extractSignature(b);