lynkr 7.2.5 → 8.0.1

package/src/routing/model-tiers.js

@@ -0,0 +1,365 @@
+/**
+ * Model Tier Selector
+ * Maps complexity scores to appropriate models per provider
+ * Uses config/model-tiers.json for tier preferences
+ */
+
+const fs = require('fs');
+const path = require('path');
+const logger = require('../logger');
+const config = require('../config');
+
+// Load tier config
+const TIER_CONFIG_PATH = path.join(__dirname, '../../config/model-tiers.json');
+
+// Tier definitions with complexity ranges
+const TIER_DEFINITIONS = {
+  SIMPLE: {
+    description: 'Greetings, simple Q&A, confirmations',
+    range: [0, 25],
+    priority: 1,
+  },
+  MEDIUM: {
+    description: 'Code reading, simple edits, research',
+    range: [26, 50],
+    priority: 2,
+  },
+  COMPLEX: {
+    description: 'Multi-file changes, debugging, architecture',
+    range: [51, 75],
+    priority: 3,
+  },
+  REASONING: {
+    description: 'Complex analysis, security audits, novel problems',
+    range: [76, 100],
+    priority: 4,
+  },
+};
+
+class ModelTierSelector {
+  constructor() {
+    this.tierConfig = null;
+    this.localProviders = {};
+    this.providerAliases = {};
+    this._loadConfig();
+  }
+
+  /**
+   * Load tier configuration from JSON file
+   */
+  _loadConfig() {
+    try {
+      if (fs.existsSync(TIER_CONFIG_PATH)) {
+        const data = JSON.parse(fs.readFileSync(TIER_CONFIG_PATH, 'utf8'));
+        this.tierConfig = data.tiers || {};
+        this.localProviders = data.localProviders || {};
+        this.providerAliases = data.providerAliases || {};
+        logger.debug({ tiers: Object.keys(this.tierConfig) }, '[ModelTiers] Config loaded');
+      } else {
+        logger.warn('[ModelTiers] Config file not found, using defaults');
+        this._loadDefaults();
+      }
+    } catch (err) {
+      logger.warn({ err: err.message }, '[ModelTiers] Config load failed, using defaults');
+      this._loadDefaults();
+    }
+  }
+
+  /**
+   * Load default tier config
+   */
+  _loadDefaults() {
+    this.tierConfig = {
+      SIMPLE: { preferred: { ollama: ['llama3.2'], openai: ['gpt-4o-mini'] } },
+      MEDIUM: { preferred: { openai: ['gpt-4o'], anthropic: ['claude-sonnet-4-20250514'] } },
+      COMPLEX: { preferred: { openai: ['o1-mini'], anthropic: ['claude-sonnet-4-20250514'] } },
+      REASONING: { preferred: { openai: ['o1'], anthropic: ['claude-opus-4-20250514'] } },
+    };
+    this.localProviders = {
+      ollama: { free: true, defaultTier: 'SIMPLE' },
+      llamacpp: { free: true, defaultTier: 'SIMPLE' },
+      lmstudio: { free: true, defaultTier: 'SIMPLE' },
+    };
+  }
+
+  /**
+   * Normalize provider name using aliases
+   */
+  _normalizeProvider(provider) {
+    if (!provider) return 'openai';
+    const lower = provider.toLowerCase();
+    return this.providerAliases[lower] || lower;
+  }
+
+  /**
+   * Get tier from complexity score
+   * @param {number} complexityScore - Score from 0-100
+   * @returns {string} Tier name (SIMPLE, MEDIUM, COMPLEX, REASONING)
+   */
+  getTier(complexityScore) {
+    const score = Math.max(0, Math.min(100, complexityScore || 0));
+
+    for (const [tier, def] of Object.entries(TIER_DEFINITIONS)) {
+      if (score >= def.range[0] && score <= def.range[1]) {
+        return tier;
+      }
+    }
+
+    return score > 75 ? 'REASONING' : 'SIMPLE';
+  }
+
+  /**
+   * Get tier definition
+   */
+  getTierDefinition(tier) {
+    return TIER_DEFINITIONS[tier] || TIER_DEFINITIONS.MEDIUM;
+  }
+
+  /**
+   * Get tier priority (1-4)
+   */
+  getTierPriority(tier) {
+    return TIER_DEFINITIONS[tier]?.priority || 2;
+  }
+
+  /**
+   * Compare two tiers, returns positive if tier1 > tier2
+   */
+  compareTiers(tier1, tier2) {
+    return this.getTierPriority(tier1) - this.getTierPriority(tier2);
+  }
+
+  /**
+   * Get preferred models for a tier and provider
+   * @param {string} tier - Tier name
+   * @param {string} provider - Provider name
+   * @returns {string[]} Array of model names
+   */
+  getPreferredModels(tier, provider) {
+    const normalizedProvider = this._normalizeProvider(provider);
+    return this.tierConfig[tier]?.preferred?.[normalizedProvider] || [];
+  }
+
+  /**
+   * Select model for tier from TIER_* env var (mandatory)
+   * @param {string} tier - Tier name (SIMPLE, MEDIUM, COMPLEX, REASONING)
+   * @param {string} _unused - Deprecated parameter
+   * @returns {Object} { model, provider, source, tier }
+   */
+  selectModel(tier, _unused = null) {
+    const tierConfig = config.modelTiers?.[tier];
+    if (!tierConfig) {
+      throw new Error(`TIER_${tier} not configured. Set TIER_${tier}=provider:model in .env`);
+    }
+
+    const parsed = this._parseTierConfig(tierConfig);
+    if (!parsed) {
+      throw new Error(`Invalid TIER_${tier} format. Expected provider:model, got: ${tierConfig}`);
+    }
+
+    return {
+      model: parsed.model,
+      provider: parsed.provider,
+      source: 'env_tier',
+      tier,
+    };
+  }
+
+  /**
+   * Parse tier config string (format: provider:model)
+   * Examples: "ollama:llama3.2", "azure-openai:gpt-5.2-chat", "openai:gpt-4o"
+   */
+  _parseTierConfig(configStr) {
+    if (!configStr || typeof configStr !== 'string') return null;
+
+    const colonIndex = configStr.indexOf(':');
+    if (colonIndex === -1) {
+      // No colon - treat as model name, use default provider
+      return {
+        provider: config.modelProvider?.type || 'openai',
+        model: configStr.trim(),
+      };
+    }
+
+    const provider = configStr.substring(0, colonIndex).trim().toLowerCase();
+    const model = configStr.substring(colonIndex + 1).trim();
+
+    if (!provider || !model) return null;
+
+    return { provider, model };
+  }
+
+  /**
+   * Get the model configured for a provider from .env
+   */
+  _getProviderModel(provider) {
+    switch (provider) {
+      case 'azure-openai':
+      case 'azureopenai':
+        return config.azureOpenAI?.deployment || null;
+      case 'openai':
+        return config.openai?.model || null;
+      case 'ollama':
+        return config.ollama?.model || null;
+      case 'openrouter':
+        return config.openrouter?.model || null;
+      case 'llamacpp':
+        return config.llamacpp?.model || null;
+      case 'lmstudio':
+        return config.lmstudio?.model || null;
+      case 'bedrock':
+        return config.bedrock?.modelId || null;
+      case 'zai':
+        return config.zai?.model || null;
+      case 'moonshot':
+        return config.moonshot?.model || null;
+      case 'vertex':
+        return config.vertex?.model || null;
+      case 'databricks':
+        return config.modelProvider?.defaultModel || null;
+      default:
+        return null;
+    }
+  }
+
+  /**
+   * Get provider for a specific tier (from env or fallback)
+   */
+  getProviderForTier(tier) {
+    const tierConfig = config.modelTiers?.[tier];
+    if (tierConfig) {
+      const parsed = this._parseTierConfig(tierConfig);
+      if (parsed) return parsed.provider;
+    }
+    return config.modelProvider?.type || 'openai';
+  }
+
+  /**
+   * Get fallback model if provider can't handle requested tier
+   */
+  _getFallbackModel(requestedTier, provider) {
+    const tierOrder = ['REASONING', 'COMPLEX', 'MEDIUM', 'SIMPLE'];
+    const startIndex = tierOrder.indexOf(requestedTier);
+
+    // Try lower tiers
+    for (let i = startIndex + 1; i < tierOrder.length; i++) {
+      const fallbackTier = tierOrder[i];
+      const models = this.getPreferredModels(fallbackTier, provider);
+
+      if (models.length > 0) {
+        logger.debug({
+          from: requestedTier,
+          to: fallbackTier,
+          provider,
+          model: models[0],
+        }, '[ModelTiers] Downgrading tier');
+
+        return { model: models[0], tier: fallbackTier };
+      }
+    }
+
+    return null;
+  }
+
+  /**
+   * Check if provider can handle a specific tier
+   */
+  canHandleTier(provider, tier) {
+    const normalizedProvider = this._normalizeProvider(provider);
+    const models = this.getPreferredModels(tier, normalizedProvider);
+    return models.length > 0;
+  }
+
+  /**
+   * Check if provider is local/free
+   */
+  isLocalProvider(provider) {
+    const normalizedProvider = this._normalizeProvider(provider);
+    return this.localProviders[normalizedProvider]?.free === true;
+  }
+
+  /**
+   * Get all providers that can handle a tier
+   */
+  getProvidersForTier(tier) {
+    const tierConfig = this.tierConfig[tier];
+    if (!tierConfig?.preferred) return [];
+    return Object.keys(tierConfig.preferred);
+  }
+
+  /**
+   * Get all tiers a provider can handle
+   */
+  getTiersForProvider(provider) {
+    const normalizedProvider = this._normalizeProvider(provider);
+    const tiers = [];
+
+    for (const tier of Object.keys(TIER_DEFINITIONS)) {
+      if (this.canHandleTier(normalizedProvider, tier)) {
+        tiers.push(tier);
+      }
+    }
+
+    return tiers;
+  }
+
+  /**
+   * Get tier stats for metrics endpoint
+   */
+  getTierStats() {
+    const stats = {
+      tiers: {},
+      providers: {},
+    };
+
+    for (const [tier, def] of Object.entries(TIER_DEFINITIONS)) {
+      const providers = this.getProvidersForTier(tier);
+      stats.tiers[tier] = {
+        ...def,
+        providerCount: providers.length,
+        providers: providers,
+      };
+    }
+
+    // Count models per provider
+    const allProviders = new Set();
+    for (const tierConfig of Object.values(this.tierConfig)) {
+      if (tierConfig.preferred) {
+        Object.keys(tierConfig.preferred).forEach(p => allProviders.add(p));
+      }
+    }
+
+    for (const provider of allProviders) {
+      stats.providers[provider] = {
+        tiers: this.getTiersForProvider(provider),
+        isLocal: this.isLocalProvider(provider),
+      };
+    }
+
+    return stats;
+  }
+
+  /**
+   * Reload configuration (for hot reload)
+   */
+  reload() {
+    this._loadConfig();
+    logger.info('[ModelTiers] Configuration reloaded');
+  }
+}
+
+// Singleton instance
+let instance = null;
+
+function getModelTierSelector() {
+  if (!instance) {
+    instance = new ModelTierSelector();
+  }
+  return instance;
+}
+
+module.exports = {
+  ModelTierSelector,
+  getModelTierSelector,
+  TIER_DEFINITIONS,
+};

package/src/server.js

@@ -78,18 +78,8 @@ function createApp() {
   // Metrics collection
   app.use(metricsMiddleware);
 
-  // Enable compression for all responses (gzip/deflate)
-  app.use(compression({
-    level: 6, // Balanced compression level
-    threshold: 1024, // Only compress responses > 1KB
-    filter: (req, res) => {
-      // Don't compress event streams
-      if (res.getHeader('Content-Type') === 'text/event-stream') {
-        return false;
-      }
-      return compression.filter(req, res);
-    }
-  }));
+  // Note: If using a tunnel (ngrok, Cloudflare Tunnel) and seeing BrotliDecompressionError,
+  // start ngrok with: ngrok http 8081 --request-header-remove "Accept-Encoding"
 
   app.use(express.json({ limit: config.server.jsonLimit }));
   app.use(sessionMiddleware);
@@ -201,9 +191,9 @@ async function start() {
 
   const app = createApp();
 
-  // Wait for Ollama if it's the configured provider or preferred for routing
+  // Wait for Ollama if it's the configured provider or referenced in tier config
   const provider = config.modelProvider?.type?.toLowerCase();
-  if (provider === "ollama" || config.modelProvider?.preferOllama) {
+  if (provider === "ollama" || config.tiersReferenceOllama()) {
     await waitForOllama();
   }
 

package/src/sessions/cleanup.js

@@ -4,9 +4,9 @@ const { cleanupOldSessions, cleanupOldHistory } = require("./store");
 class SessionCleanupManager {
   constructor(options = {}) {
     this.enabled = options.enabled !== false;
-    this.intervalMs = options.intervalMs || 3600000; // 1 hour
-    this.sessionMaxAgeMs = options.sessionMaxAgeMs || 7 * 24 * 60 * 60 * 1000; // 7 days
-    this.historyMaxAgeMs = options.historyMaxAgeMs || 30 * 24 * 60 * 60 * 1000; // 30 days
+    this.intervalMs = options.intervalMs || 300000; // 5 minutes (was 1 hour)
+    this.sessionMaxAgeMs = options.sessionMaxAgeMs || 24 * 60 * 60 * 1000; // 1 day (was 7 days)
+    this.historyMaxAgeMs = options.historyMaxAgeMs || 7 * 24 * 60 * 60 * 1000; // 7 days (was 30 days)
     this.timer = null;
   }
 

package/src/sessions/record.js

@@ -1,5 +1,8 @@
 const { appendSessionTurn } = require("./store");
 
+// Cap in-memory history to prevent unbounded growth during long tool loops
+const MAX_IN_MEMORY_HISTORY = 100;
+
 function ensureSessionShape(session) {
   if (!session) return null;
   if (!Array.isArray(session.history)) {
@@ -19,7 +22,13 @@ function appendTurnToSession(session, entry) {
   target.history.push(turn);
   target.updatedAt = turn.timestamp;
 
-  if (target.id) {
+  // Trim in-memory history if it exceeds the cap
+  if (target.history.length > MAX_IN_MEMORY_HISTORY) {
+    target.history = target.history.slice(-MAX_IN_MEMORY_HISTORY);
+  }
+
+  // Skip DB write for ephemeral sessions (auto-generated, no client session ID)
+  if (target.id && !target._ephemeral) {
     appendSessionTurn(target.id, turn, target.metadata ?? {});
   }
 

package/src/sessions/store.js

@@ -4,11 +4,15 @@ const logger = require("../logger");
 const selectSessionStmt = db.prepare(
   "SELECT id, created_at, updated_at, metadata FROM sessions WHERE id = ?",
 );
+// Limit history to last 50 entries to prevent unbounded memory growth.
+// Older entries remain in DB for auditing but aren't loaded into memory.
+const MAX_HISTORY_ROWS = 50;
 const selectHistoryStmt = db.prepare(
   `SELECT role, type, status, content, metadata, timestamp
    FROM session_history
    WHERE session_id = ?
-   ORDER BY timestamp ASC, id ASC`,
+   ORDER BY timestamp DESC, id DESC
+   LIMIT ${MAX_HISTORY_ROWS}`,
 );
 const insertSessionStmt = db.prepare(
   "INSERT INTO sessions (id, created_at, updated_at, metadata) VALUES (@id, @created_at, @updated_at, @metadata)",
@@ -75,7 +79,8 @@ function getSession(sessionId) {
   if (!sessionId) return null;
   const sessionRow = selectSessionStmt.get(sessionId);
   if (!sessionRow) return null;
-  const historyRows = selectHistoryStmt.all(sessionId);
+  // Query returns rows in DESC order (for LIMIT to grab newest), reverse to ASC
+  const historyRows = selectHistoryStmt.all(sessionId).reverse();
   return toSession(sessionRow, historyRows);
 }
 

package/src/tools/agent-task.js

@@ -2,6 +2,50 @@ const { registerTool } = require(".");
 const { spawnAgent, autoSelectAgent } = require("../agents");
 const logger = require("../logger");
 
+/**
+ * Extract text from Anthropic content blocks format
+ * Handles: [{"type":"text","text":"..."}] -> "..."
+ */
+function extractTextFromContentBlocks(content) {
+  if (typeof content !== 'string') {
+    return content;
+  }
+
+  const trimmed = content.trim();
+  if (!trimmed.startsWith('[')) {
+    return content;
+  }
+
+  try {
+    const parsed = JSON.parse(trimmed);
+    if (!Array.isArray(parsed)) {
+      return content;
+    }
+
+    // Extract text from content blocks
+    const textParts = parsed
+      .filter(block => block && typeof block === 'object')
+      .map(block => {
+        if (block.type === 'text' && typeof block.text === 'string') {
+          return block.text;
+        }
+        if (typeof block.text === 'string') {
+          return block.text;
+        }
+        return null;
+      })
+      .filter(text => text !== null);
+
+    if (textParts.length > 0) {
+      return textParts.join('\n\n');
+    }
+
+    return content;
+  } catch {
+    return content;
+  }
+}
+
 function registerAgentTaskTool() {
   registerTool(
     "Task",
@@ -49,10 +93,13 @@ function registerAgentTaskTool() {
         });
 
         if (result.success) {
+          // Extract text from Anthropic content blocks if present
+          const cleanContent = extractTextFromContentBlocks(result.result);
+
           return {
             ok: true,
             status: 200,
-            content: result.result,
+            content: cleanContent,
             metadata: {
               agentType: subagentType,
               agentId: result.stats.agentId,

package/src/tools/index.js

@@ -1,5 +1,6 @@
 const logger = require("../logger");
 const { truncateToolOutput } = require("./truncate");
+const { isGPTProvider, formatToolResultForGPT } = require("../clients/gpt-utils");
 
 const registry = new Map();
 const registryLowercase = new Map();
@@ -29,6 +30,10 @@ const TOOL_ALIASES = {
   WebSearch: "web_search",
   web_fetch: "web_fetch",
   webfetch: "web_fetch",
+  web_agent: "web_agent",
+  webagent: "web_agent",
+  WebAgent: "web_agent",
+  tinyfish: "web_agent",
   task: "fs_write",
   write: "fs_write",
   filewrite: "fs_write",
@@ -254,7 +259,18 @@ async function executeToolCall(call, context = {}) {
     const formatted = normalizeHandlerResult(result);
 
     // Apply tool output truncation for token efficiency
-    const truncatedContent = truncateToolOutput(normalisedCall.name, formatted.content);
+    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
 
     return {
       id: normalisedCall.id,
@@ -267,7 +283,8 @@ async function executeToolCall(call, context = {}) {
         registered: true,
         truncated: truncatedContent !== formatted.content,
         originalLength: formatted.content?.length,
-        truncatedLength: truncatedContent?.length
+        truncatedLength: truncatedContent?.length,
+        gptFormatted: isGPT,
       },
     };
   } catch (err) {

package/src/tools/lazy-loader.js

@@ -57,6 +57,11 @@ const TOOL_CATEGORIES = {
     loader: () => require('./tasks').registerTaskTools,
     priority: 2,
   },
+  tinyfish: {
+    keywords: ['tinyfish', 'web_agent', 'automate', 'scrape', 'extract', 'crawl', 'browser'],
+    loader: () => require('./tinyfish').registerTinyFishTools,
+    priority: 2,
+  },
   tests: {
     keywords: ['test', 'jest', 'mocha', 'pytest', 'unittest', 'spec', 'coverage', 'assert'],
     loader: () => require('./tests').registerTestTools,
@@ -277,6 +282,8 @@ function loadCategoryForTool(toolName) {
     'workspace_mcp_servers': 'mcp',
 
     // Agent task
+    // TinyFish (web agent)
+    'web_agent': 'tinyfish',
     'agent_task': 'agentTask',
   };