aiden-runtime 3.16.2 → 3.18.0

package/dist/core/agentLoop.js

@@ -66,6 +66,7 @@ const entityGraph_1 = require("./entityGraph");
 const learningMemory_1 = require("./learningMemory");
 const conversationMemory_1 = require("./conversationMemory");
 const router_1 = require("../providers/router");
+const modelRegistry_1 = require("./modelRegistry");
 const index_1 = require("../providers/index");
 const knowledgeBase_1 = require("./knowledgeBase");
 const skillTeacher_1 = require("./skillTeacher");
@@ -80,6 +81,7 @@ const semanticMemory_1 = require("./semanticMemory");
 const sessionMemory_1 = require("./sessionMemory");
 const goalTracker_1 = require("./goalTracker");
 const hooks_1 = require("./hooks");
+const pluginLoader_1 = require("./pluginLoader");
 const instinctSystem_1 = require("./instinctSystem");
 const workflowTracker_1 = require("./workflowTracker");
 const parallelExecutor_1 = require("./parallelExecutor");
@@ -427,6 +429,7 @@ function inferPhasesFromSteps(steps) {
         clipboard_read: 'execution', clipboard_write: 'execution',
         window_list: 'execution', window_focus: 'execution',
         app_launch: 'execution', app_close: 'execution',
+        system_volume: 'execution',
         watch_folder: 'execution', watch_folder_list: 'execution',
     };
     const phaseNames = {
@@ -717,7 +720,7 @@ async function planWithLLM(message, history, apiKey, model, provider, memoryCont
         'screenshot', 'screen_read', 'vision_loop', 'wait',
         'code_interpreter_python', 'code_interpreter_node',
         'clipboard_read', 'clipboard_write', 'window_list', 'window_focus',
-        'app_launch', 'app_close',
+        'app_launch', 'app_close', 'system_volume',
         'watch_folder', 'watch_folder_list',
         'send_file_local', 'receive_file_local',
         'get_briefing',
@@ -953,11 +956,13 @@ If requires_execution is false:
 
 NOTE: "goals" is only required when 2+ distinct intents are present. Single-goal messages may omit it.
 
-THE 'respond' TOOL — use this for ALL conversational messages:
-- 'respond' is ALWAYS a valid plan. When no external tool is needed, plan a single respond step.
+THE 'respond' TOOL — use this for conversational messages ONLY:
 - respond: { "message": "your answer text here" }
 - Use respond for: greetings, capability questions, simple facts from training data, clarifying questions, short answers.
 - Example: user says "hi" → { "goal": "hi", "requires_execution": true, "plan": [{ "step": 1, "tool": "respond", "input": { "message": "Hi! What can I help you with today?" } }] }
+- NEVER use respond as the only step when the user expects a physical action on their machine.
+  "open chrome", "close spotify", "increase volume" -> these REQUIRE the actual tool (see SYSTEM ACTION GATE).
+  Responding with "Done. I've opened Chrome." without calling app_launch is lying — never do it.
 
 ACTION GATE RULES — apply BEFORE creating any plan:
 1. CAPABILITY GATE: If message is "Can you do X?" / "Can you X?" / "Are you able to X?" → plan respond with answer
@@ -970,6 +975,28 @@ ACTION GATE RULES — apply BEFORE creating any plan:
 4. NEVER create comparison tables, reports, or verdicts unless user explicitly asked for them
 5. NEVER mention Pega, BlueWinston, Gaude Digital, or any third-party product by name
 
+SYSTEM ACTION GATE — CRITICAL, apply before any system-action request:
+ANY request to open, launch, start, close, kill, quit an app OR change system volume
+MUST use the corresponding tool. Using 'respond' alone is WRONG for these.
+
+Exact mappings (use these, do not improvise):
+- "open chrome" / "launch chrome" / "open Google Chrome"  → app_launch { app_name: "chrome" }
+- "close chrome" / "kill chrome" / "quit chrome"          → app_close  { app_name: "chrome" }
+- "open spotify" / "launch spotify"                       → app_launch { app_name: "spotify" }
+- "close spotify"                                         → app_close  { app_name: "spotify" }
+- "open discord" / "open VS Code" / "open notepad"        → app_launch { app_name: "<name>" }
+- "increase volume" / "volume up 20" / "turn up volume"   → system_volume { volume: 20 }
+- "decrease volume" / "volume down 10"                    → system_volume { volume: 10, direction: "down" }
+- "mute" / "mute sound"                                   → system_volume { mute: true }
+- "unmute"                                                → system_volume { unmute: true }
+- "open file explorer"                                    → app_launch { app_name: "explorer" }
+
+WRONG (never do this for the above requests):
+  { "tool": "respond", "input": { "message": "Done. I've opened Chrome." } }  <- FAKE, LYING
+
+CORRECT:
+  { "tool": "app_launch", "input": { "app_name": "chrome" } }  <- actually opens Chrome
+
 ## SKILL DISCOVERY
 
 Before planning any multi-step task (>=2 tools), call lookup_skill with the user's message as the query.
@@ -986,7 +1013,14 @@ TIER 1 (USE FIRST): lookup_skill, respond, web_search, fetch_page, fetch_url, de
 TIER 2 (USE SECOND): file_write, file_read, file_list, shell_exec, run_powershell, run_python, run_node, code_interpreter_python, code_interpreter_node, git_status, git_commit, git_push, clipboard_read, clipboard_write, spawn_subagent, swarm
   → Use when you need to read/write files, run scripts, or run git commands
 
-TIER 3 (USE THIRD): open_browser, browser_click, browser_type, browser_extract, browser_screenshot, window_list, window_focus, app_launch, app_close
+TIER 3a — SYSTEM ACTIONS (use whenever user asks for OS-level actions):
+  app_launch, app_close, system_volume, window_focus, window_list
+  → USE IMMEDIATELY when user asks to open/close/launch/kill an app, change volume, or focus a window
+  → Do NOT substitute with respond — the user wants the ACTION to happen, not acknowledgment
+  → Do NOT use shell_exec as a substitute; app_launch/app_close are the correct tools
+
+TIER 3b — BROWSER UI (use when task requires interacting with a website UI):
+  open_browser, browser_click, browser_type, browser_extract, browser_screenshot
   → ONLY when task requires interacting with a website UI
   → NEVER use browser when an API tool can do the same job
   → For other selectors always pass selector: "<css selector>", never guess at element text.
@@ -1168,11 +1202,19 @@ Output ONLY valid JSON, nothing else:`;
                 e.message?.includes('429') ||
                 e.message?.includes('rate') ||
                 e.message?.includes('aborted')) {
-                try {
-                    (0, router_1.markRateLimited)(curApiName);
-                    console.log(`[Planner] Marked ${curApiName} as rate limited — will rotate away`);
+                // Try next model within the same provider before marking whole entry rate-limited
+                const nextModel = (0, modelRegistry_1.getNextModelOnFailure)(curProvider, curModel);
+                if (nextModel) {
+                    console.log(`[Planner] Model ${curModel} failed — trying next model ${nextModel} on same provider (${curApiName})`);
+                    curModel = nextModel;
+                }
+                else {
+                    try {
+                        (0, router_1.markRateLimited)(curApiName);
+                        console.log(`[Planner] Marked ${curApiName} as rate limited — will rotate away`);
+                    }
+                    catch { }
                 }
-                catch { }
             }
         }
         // Wait before next attempt — helps with rate-limit recovery
@@ -1366,7 +1408,7 @@ const VALID_TOOLS = [
     'screenshot', 'screen_read', 'vision_loop', 'wait',
     'code_interpreter_python', 'code_interpreter_node',
     'clipboard_read', 'clipboard_write', 'window_list', 'window_focus',
-    'app_launch', 'app_close',
+    'app_launch', 'app_close', 'system_volume',
     'watch_folder', 'watch_folder_list',
     'send_file_local', 'receive_file_local',
     'clarify', 'todo', 'cronjob', 'vision_analyze',
@@ -1667,11 +1709,25 @@ const NO_RETRY_TOOLS = new Set([
 async function executeToolWithRetry(tool, input, maxRetries = 2) {
     const retryable = !NO_RETRY_TOOLS.has(tool);
     const effectiveMax = retryable ? maxRetries : 0;
+    // ── Plugin preTool hooks ──────────────────────────────────────
+    let effectiveInput = input;
+    for (const hook of pluginLoader_1.pluginHooks.preTool) {
+        try {
+            const r = await hook(tool, effectiveInput);
+            if (r.skip)
+                return { success: true, output: '[skipped by plugin]', skippedByPlugin: true };
+            if (r.input)
+                effectiveInput = r.input;
+        }
+        catch (e) {
+            console.warn(`[PluginHook] preTool error for ${tool}:`, e.message);
+        }
+    }
     for (let attempt = 0; attempt <= effectiveMax; attempt++) {
         try {
-            const result = await (0, toolRegistry_1.executeTool)(tool, input);
+            const result = await (0, toolRegistry_1.executeTool)(tool, effectiveInput);
             if (result.success) {
-                const quality = validateResultQuality(tool, input, result.output || result);
+                const quality = validateResultQuality(tool, effectiveInput, result.output || result);
                 if (!quality.valid) {
                     console.log(`[Quality] ${tool} returned but quality check failed: ${quality.reason}`);
                     if (attempt < effectiveMax) {
@@ -1683,7 +1739,19 @@ async function executeToolWithRetry(tool, input, maxRetries = 2) {
                     console.log(`[Quality] ${tool} — accepting low-quality result after ${effectiveMax} retries`);
                     appendLesson(`${tool} produced low-quality output (${quality.reason}) after ${effectiveMax} retries — consider alternative approach for this tool.`);
                 }
-                return result;
+                // ── Plugin postTool hooks ─────────────────────────────
+                let finalResult = result;
+                for (const hook of pluginLoader_1.pluginHooks.postTool) {
+                    try {
+                        const r = await hook(tool, effectiveInput, finalResult);
+                        if (r.result)
+                            finalResult = r.result;
+                    }
+                    catch (e) {
+                        console.warn(`[PluginHook] postTool error for ${tool}:`, e.message);
+                    }
+                }
+                return finalResult;
             }
             if (attempt < effectiveMax) {
                 const delay = Math.min(1000 * Math.pow(2, attempt), 5000);
@@ -1720,7 +1788,7 @@ const SEQUENTIAL_ONLY = new Set([
     'open_browser', 'browser_click', 'browser_type', 'browser_extract',
     'mouse_move', 'mouse_click', 'keyboard_type', 'keyboard_press',
     'screenshot', 'screen_read', 'vision_loop', 'notify', 'wait',
-    'clipboard_write', 'window_focus', 'app_launch', 'app_close',
+    'clipboard_write', 'window_focus', 'app_launch', 'app_close', 'system_volume',
     'watch_folder',
 ]);
 function buildDependencyGroups(steps) {
@@ -1791,6 +1859,7 @@ async function executePlan(plan, onStep, onPhaseChange, existingState, replanApi
         clipboard_read: 'execution', clipboard_write: 'execution',
         window_list: 'execution', window_focus: 'execution',
         app_launch: 'execution', app_close: 'execution',
+        system_volume: 'execution',
         watch_folder: 'execution', watch_folder_list: 'execution',
     };
     let lastCapability = '';
@@ -2198,6 +2267,18 @@ function responderSystem(userName, date) {
     return (0, aidenPersonality_1.AIDEN_RESPONDER_SYSTEM)(userName, date);
 }
 async function respondWithResults(originalMessage, plan, results, history, userName, apiKey, model, providerName, onToken, sessionId, goals) {
+    // ── CommandGate / PermissionGate short-circuit ───────────────
+    // If ANY tool was blocked with an approval gate, stream the
+    // approval question directly — never let the LLM hallucinate "Done".
+    const gatedResult = results.find(r => !r.success && r.error &&
+        (r.error.startsWith('CommandGate:') || r.error.startsWith('PermissionGate:')));
+    if (gatedResult) {
+        const blocked = gatedResult.error
+            .replace(/^(CommandGate|PermissionGate):\s*/i, '')
+            .replace(/:\s*$/, '');
+        onToken(`I need your approval before I can do that.\n\n**Blocked action:** ${blocked}\n\nReply **yes** to confirm, or tell me what you'd like instead.`);
+        return;
+    }
     const date = new Date().toLocaleDateString('en-US', {
         weekday: 'long', month: 'long', day: 'numeric', year: 'numeric',
     });

package/dist/core/channels/email.js

@@ -134,7 +134,7 @@ class EmailAdapter {
                     host: this.imapHost,
                     port: this.imapPort,
                     tls: true,
-                    tlsOptions: { rejectUnauthorized: false },
+                    tlsOptions: { rejectUnauthorized: false }, // user-configured IMAP server may use self-signed cert
                     user: this.imapUser,
                     password: this.imapPassword,
                     authTimeout: 5000,

package/dist/core/modelRegistry.js

@@ -0,0 +1,261 @@
+"use strict";
+/**
+ * core/modelRegistry.ts
+ * Curated list of best free/cheap models per provider.
+ * Updated manually — not auto-discovered (keeps things simple and predictable).
+ *
+ * Usage:
+ *   getDefaultModel('groq')              → 'llama-3.3-70b-versatile'
+ *   getNextModelOnFailure('groq', 'llama-3.3-70b-versatile') → 'llama-3.1-70b-versatile'
+ *   getRegistryEntry('groq', 'llama-3.3-70b-versatile')      → ModelConfig | undefined
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.MODEL_REGISTRY = void 0;
+exports.getDefaultModel = getDefaultModel;
+exports.getNextModelOnFailure = getNextModelOnFailure;
+exports.getRegistryEntry = getRegistryEntry;
+exports.getModelsForProvider = getModelsForProvider;
+/**
+ * Ordered by preference — first entry is the default.
+ * Free models come before paid unless quality difference is large.
+ * Env var override: set ${PROVIDER_UPPER}_MODEL to force a specific model.
+ *   e.g. GROQ_MODEL=mixtral-8x7b-32768 overrides groq default
+ */
+exports.MODEL_REGISTRY = {
+    groq: [
+        {
+            id: 'llama-3.3-70b-versatile',
+            contextWindow: 128000,
+            pricing: 'free',
+            quality: 'high',
+            speed: 'fast',
+            notes: 'Primary — fastest + highest quality free tier',
+        },
+        {
+            id: 'llama-3.1-70b-versatile',
+            contextWindow: 128000,
+            pricing: 'free',
+            quality: 'high',
+            speed: 'fast',
+            notes: 'Fallback when 3.3 is rate-limited',
+        },
+        {
+            id: 'llama3-70b-8192',
+            contextWindow: 8192,
+            pricing: 'free',
+            quality: 'high',
+            speed: 'fast',
+            notes: 'Smaller context but very reliable',
+        },
+        {
+            id: 'mixtral-8x7b-32768',
+            contextWindow: 32768,
+            pricing: 'free',
+            quality: 'medium',
+            speed: 'fast',
+            notes: 'Good for structured JSON tasks',
+        },
+        {
+            id: 'gemma2-9b-it',
+            contextWindow: 8192,
+            pricing: 'free',
+            quality: 'medium',
+            speed: 'fast',
+            notes: 'Light fallback',
+        },
+    ],
+    openrouter: [
+        {
+            id: 'meta-llama/llama-3.3-70b-instruct:free',
+            contextWindow: 131072,
+            pricing: 'free',
+            quality: 'high',
+            speed: 'medium',
+            notes: 'Best free model on OpenRouter',
+        },
+        {
+            id: 'meta-llama/llama-3.1-70b-instruct:free',
+            contextWindow: 131072,
+            pricing: 'free',
+            quality: 'high',
+            speed: 'medium',
+            notes: 'Reliable free fallback',
+        },
+        {
+            id: 'mistralai/mistral-7b-instruct:free',
+            contextWindow: 32768,
+            pricing: 'free',
+            quality: 'medium',
+            speed: 'fast',
+            notes: 'Fast small model for simple tasks',
+        },
+        {
+            id: 'google/gemma-2-9b-it:free',
+            contextWindow: 8192,
+            pricing: 'free',
+            quality: 'medium',
+            speed: 'fast',
+            notes: 'Emergency fallback',
+        },
+    ],
+    together: [
+        {
+            id: 'meta-llama/llama-3.1-405b-instruct',
+            contextWindow: 130000,
+            pricing: 'paid',
+            quality: 'high',
+            speed: 'medium',
+            notes: '405B — highest quality, use sparingly ($5 credit)',
+        },
+        {
+            id: 'meta-llama/llama-3.3-70b-instruct-turbo',
+            contextWindow: 131072,
+            pricing: 'paid',
+            quality: 'high',
+            speed: 'fast',
+            notes: 'Faster cheaper Together option',
+        },
+        {
+            id: 'meta-llama/llama-3.1-70b-instruct-turbo',
+            contextWindow: 131072,
+            pricing: 'paid',
+            quality: 'high',
+            speed: 'fast',
+            notes: 'Fallback paid',
+        },
+    ],
+    nvidia: [
+        {
+            id: 'nvidia/llama-3.3-nemotron-super-49b-v1',
+            contextWindow: 131072,
+            pricing: 'free',
+            quality: 'high',
+            speed: 'medium',
+            notes: 'NVIDIA NIM — high quality free inference',
+        },
+        {
+            id: 'meta/llama-3.3-70b-instruct',
+            contextWindow: 131072,
+            pricing: 'free',
+            quality: 'high',
+            speed: 'medium',
+            notes: 'NVIDIA-hosted Llama fallback',
+        },
+        {
+            id: 'mistralai/mixtral-8x7b-instruct-v0.1',
+            contextWindow: 32768,
+            pricing: 'free',
+            quality: 'medium',
+            speed: 'fast',
+            notes: 'Lightweight NVIDIA fallback',
+        },
+    ],
+    gemini: [
+        {
+            id: 'gemini-2.5-flash',
+            contextWindow: 1000000,
+            pricing: 'free',
+            quality: 'high',
+            speed: 'fast',
+            notes: '1M context, thinking model, best free Gemini',
+        },
+        {
+            id: 'gemini-2.0-flash',
+            contextWindow: 1000000,
+            pricing: 'free',
+            quality: 'high',
+            speed: 'fast',
+            notes: 'Stable previous gen, good fallback',
+        },
+        {
+            id: 'gemini-1.5-flash',
+            contextWindow: 1000000,
+            pricing: 'free',
+            quality: 'medium',
+            speed: 'fast',
+            notes: 'Conservative fallback if 2.x rate-limited',
+        },
+        {
+            id: 'gemini-1.5-flash-8b',
+            contextWindow: 1000000,
+            pricing: 'free',
+            quality: 'low',
+            speed: 'fast',
+            notes: 'Emergency fallback — smallest Gemini',
+        },
+    ],
+    ollama: [
+        {
+            id: 'gemma4:e4b',
+            contextWindow: 8192,
+            pricing: 'free',
+            quality: 'medium',
+            speed: 'medium',
+            notes: 'Local default — requires GTX 1060 VRAM',
+        },
+        {
+            id: 'qwen2.5-coder:7b',
+            contextWindow: 32768,
+            pricing: 'free',
+            quality: 'medium',
+            speed: 'medium',
+            notes: 'Local coder model',
+        },
+        {
+            id: 'llama3.2:latest',
+            contextWindow: 128000,
+            pricing: 'free',
+            quality: 'medium',
+            speed: 'fast',
+            notes: 'Local fast model',
+        },
+    ],
+};
+/**
+ * Returns the default model ID for a provider.
+ * Env var ${PROVIDER_UPPER}_MODEL overrides the registry default.
+ *
+ * e.g. GROQ_MODEL=mixtral-8x7b-32768 → uses that instead
+ */
+function getDefaultModel(provider) {
+    const envKey = `${provider.toUpperCase()}_MODEL`;
+    const envOverride = process.env[envKey];
+    if (envOverride)
+        return envOverride;
+    const models = exports.MODEL_REGISTRY[provider.toLowerCase()];
+    if (!models || models.length === 0)
+        return '';
+    return models[0].id;
+}
+/**
+ * Returns the next model to try after currentModel fails (rate-limited / error).
+ * Returns null if currentModel is already the last in the list — caller should
+ * then mark the whole provider entry rate-limited and rotate to next provider.
+ */
+function getNextModelOnFailure(provider, currentModel) {
+    const models = exports.MODEL_REGISTRY[provider.toLowerCase()];
+    if (!models || models.length === 0)
+        return null;
+    const idx = models.findIndex(m => m.id === currentModel);
+    if (idx === -1 || idx >= models.length - 1)
+        return null;
+    return models[idx + 1].id;
+}
+/**
+ * Returns the ModelConfig for a specific provider + model id.
+ */
+function getRegistryEntry(provider, modelId) {
+    const models = exports.MODEL_REGISTRY[provider.toLowerCase()];
+    if (!models)
+        return undefined;
+    return models.find(m => m.id === modelId);
+}
+/**
+ * Returns all models for a provider, optionally filtered by pricing tier.
+ */
+function getModelsForProvider(provider, filter) {
+    const models = exports.MODEL_REGISTRY[provider.toLowerCase()] ?? [];
+    if (!filter)
+        return models;
+    return models.filter(m => !filter.pricing || m.pricing === filter.pricing);
+}