alvin-bot 4.5.1 → 4.7.0

package/docs/HANDBOOK.md

@@ -0,0 +1,856 @@
+# Alvin Bot Handbook
+
+> A complete, standalone reference for everything Alvin Bot can do — installation, architecture, providers, sub-agents, cron jobs, plugins, MCP, platforms, security, and troubleshooting. Written for users who want to understand the whole system, not just the quick-start.
+
+---
+
+## Table of contents
+
+1. [What Alvin Bot is](#1-what-alvin-bot-is)
+2. [Installation](#2-installation)
+3. [First run: the setup wizard](#3-first-run-the-setup-wizard)
+4. [Architecture at a glance](#4-architecture-at-a-glance)
+5. [AI providers](#5-ai-providers)
+6. [The chat session model](#6-the-chat-session-model)
+7. [Sub-agents](#7-sub-agents)
+8. [Cron jobs](#8-cron-jobs)
+9. [Commands reference](#9-commands-reference)
+10. [Platforms: Telegram, WhatsApp, Discord, Signal, Web, TUI](#10-platforms)
+11. [Plugins and MCP servers](#11-plugins-and-mcp-servers)
+12. [Skills](#12-skills)
+13. [Security](#13-security)
+14. [Running in production](#14-running-in-production)
+15. [Troubleshooting](#15-troubleshooting)
+16. [Upgrading](#16-upgrading)
+
+---
+
+## 1. What Alvin Bot is
+
+Alvin Bot is a self-hosted AI agent that lives where you chat. Instead of being a single-purpose chatbot, it connects a multi-model engine (Claude, OpenAI, Gemini, Groq, NVIDIA NIM, Ollama, OpenRouter, or any OpenAI-compatible endpoint) to every messaging platform you care about (Telegram, WhatsApp, Discord, Signal, Web UI, terminal), plus a set of superpowers: persistent memory, cron jobs, file access, skill files, MCP tool use, sub-agents for parallel work, and a comprehensive security audit.
+
+The design goal is **"the assistant that actually lives on your machine"**: it has real tool access (not a sandboxed chat), it remembers across sessions, it runs scheduled tasks, and you talk to it through whatever app you already have open.
+
+**Alvin Bot is the bot process.** Your identity — what it calls you, your preferences, long-running context — lives in `~/.alvin-bot/memory/` and is independent of the bot version.
+
+---
+
+## 2. Installation
+
+### 2.1 Quickest path — npm
+
+```bash
+npm install -g alvin-bot
+alvin-bot setup
+alvin-bot start
+```
+
+Requirements:
+- Node.js 18 or newer
+- A Telegram bot token (free from [@BotFather](https://t.me/BotFather))
+- Your Telegram user ID (from [@userinfobot](https://t.me/userinfobot))
+
+Everything else — an AI provider key — the setup wizard walks you through.
+
+### 2.2 From source
+
+```bash
+git clone https://github.com/alvbln/Alvin-Bot.git
+cd Alvin-Bot
+npm install
+npm run build
+node bin/cli.js setup
+node bin/cli.js start
+```
+
+This is the path if you want to track `main` or apply local patches.
+
+### 2.3 One-line install script
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/alvbln/Alvin-Bot/main/install.sh | bash
+```
+
+Downloads, builds, and runs the setup wizard in one shot. Linux and macOS only.
+
+### 2.4 Desktop app (macOS)
+
+Download the latest `.dmg` from [GitHub Releases](https://github.com/alvbln/Alvin-Bot/releases). The desktop app wraps the bot in an Electron shell with a system tray icon, auto-updater, and the embedded web UI.
+
+Current build: Apple Silicon (arm64). Windows and Linux desktop builds are on the roadmap.
+
+### 2.5 Docker
+
+```bash
+docker compose up -d
+```
+
+Use this for headless servers. See `docker-compose.yml` for the expected environment variables.
+
+### 2.6 The data directory
+
+Alvin Bot stores all user data under `~/.alvin-bot/` (overridable via the `ALVIN_DATA_DIR` environment variable):
+
+```
+~/.alvin-bot/
+├── .env                  # bot token, API keys, config
+├── memory/               # long-term memory (vector-indexed)
+├── sessions/             # session transcripts
+├── cron-jobs.json        # scheduled jobs
+├── sub-agents.json       # sub-agent config (maxParallel, visibility)
+├── delivery-queue.json   # pending outgoing messages with retry
+├── logs/                 # stdout/stderr if running under launchd
+├── custom-models.json    # user-added OpenAI-compatible providers
+├── users/                # multi-user approval + preferences
+└── whatsapp-groups.json  # WhatsApp group allowlist
+```
+
+The bot code itself lives wherever you installed it (`/usr/local/lib/node_modules/alvin-bot` for npm-global, or your git clone directory). The data directory is portable: copy it between machines and the bot comes up with the same memories, jobs, and settings.
+
+---
+
+## 3. First run: the setup wizard
+
+Running `alvin-bot setup` walks you through four screens:
+
+1. **AI provider** — pick one of the supported providers. Free options first (Groq, Gemini, NVIDIA NIM, Ollama, Claude Max via the CLI). The wizard validates the API key with a test call before saving.
+2. **Telegram bot token** — paste the token from BotFather. The wizard calls `getMe` to confirm it's valid and shows you the bot's handle.
+3. **Authorized user ID** — paste your Telegram user ID. Only users listed in `ALLOWED_USERS` can chat with the bot; everyone else gets a silent deny.
+4. **Optional platforms** — Slack, WhatsApp, Discord, Signal. Each is configured independently; skip any you don't use.
+
+All values land in `~/.alvin-bot/.env` so you can edit them later.
+
+---
+
+## 4. Architecture at a glance
+
+```
+┌─────────────────────────────────────────────────────────────┐
+│ Platform layer                                              │
+│ Telegram · WhatsApp · Discord · Signal · Slack · TUI · Web  │
+└──────────────────────────┬──────────────────────────────────┘
+                           │
+                     platform-message
+                           │
+┌──────────────────────────▼──────────────────────────────────┐
+│ Handler layer                                               │
+│ handleMessage · handleVoice · handlePhoto · handleDocument  │
+└──────────────────────────┬──────────────────────────────────┘
+                           │
+                     Engine.query()
+                           │
+┌──────────────────────────▼──────────────────────────────────┐
+│ Provider registry                                           │
+│ Claude SDK · OpenAI-compat · Codex CLI · Ollama · custom... │
+│ Heartbeat monitor · fallback chain · retry on transient     │
+└──────────────────────────┬──────────────────────────────────┘
+                           │
+┌──────────────────────────▼──────────────────────────────────┐
+│ Services                                                    │
+│ Memory · Sub-agents · Cron · Skills · MCP · Plugins · Queue │
+└─────────────────────────────────────────────────────────────┘
+```
+
+Everything is wired together in `src/index.ts`. The process boots by:
+
+1. Creating `~/.alvin-bot/` and migrating any legacy data
+2. Validating the `.env` against the chosen provider
+3. Initializing the provider registry
+4. Loading plugins and MCP servers
+5. Creating the grammy bot instance and attaching the sub-agent delivery router
+6. Starting the web server, cron scheduler, delivery queue, and heartbeat
+7. Starting optional platform adapters
+8. Beginning Telegram polling
+
+A `shutdown()` function in the same file runs in reverse: cancel all running sub-agents (with Telegram notifications), stop the scheduler, drain the delivery queue, unload plugins, disconnect MCP, tear down provider lifecycles (Ollama, LM Studio), then exit cleanly.
+
+---
+
+## 5. AI providers
+
+Alvin Bot supports a multi-provider registry with a fallback chain. You pick a primary, optionally define fallbacks, and the bot picks the best available provider per request.
+
+### 5.1 Built-in providers
+
+| Registry key | Source | Auth | Cost | Notes |
+|---|---|---|---|---|
+| `claude-sdk` | Anthropic Claude Code SDK | OAuth (Claude Max) or API key | Max plan or per-token | Full tool use: Read, Write, Edit, Bash, Glob, Grep, WebSearch, WebFetch, Task. Needs the `claude` CLI installed. |
+| `codex-cli` | OpenAI Codex CLI | Local CLI | Per-token | Needs the `codex` CLI installed. |
+| `google` | Google Gemini 2.5 Flash | `GOOGLE_API_KEY` | Free tier available | Vision + embeddings. |
+| `ollama` | Local Ollama daemon | none | Free | Auto-detected. Bot can start/stop the Ollama daemon on demand. |
+
+### 5.2 Custom providers
+
+Any OpenAI-compatible endpoint (Groq, NVIDIA NIM, OpenRouter, LM Studio, vLLM, any self-hosted LLM server) goes into `~/.alvin-bot/custom-models.json`:
+
+```json
+[
+  {
+    "key": "groq",
+    "name": "Groq (Llama 3.3 70B)",
+    "model": "llama-3.3-70b-versatile",
+    "baseUrl": "https://api.groq.com/openai/v1",
+    "apiKeyEnv": "GROQ_API_KEY",
+    "supportsVision": false,
+    "supportsStreaming": true,
+    "maxTokens": 32768
+  }
+]
+```
+
+Set `PRIMARY_PROVIDER=groq` in `.env` and add `GROQ_API_KEY`. The setup wizard writes this for you for the common providers.
+
+### 5.3 The fallback chain
+
+`.env` can define a fallback order:
+
+```
+PRIMARY_PROVIDER=claude-sdk
+FALLBACK_PROVIDERS=google,groq
+```
+
+On every query, the registry tries the primary first. If it's unavailable (auth fails, rate-limited, service down), it cascades through the fallback list in order. The registry also learns — a provider that fails twice in a row gets temporarily deprioritized until the heartbeat monitor confirms it's healthy again.
+
+### 5.4 The heartbeat monitor
+
+Every 5 minutes, the heartbeat monitor pings each configured provider with a tiny probe. Providers that fail twice in a row are marked unhealthy and get skipped until they recover. You see the current state in `/status` and in the web UI's Dashboard.
+
+### 5.5 Model switching from chat
+
+`/model` — pick a different provider as your current primary.
+`/fallback` — rearrange the fallback chain from Telegram.
+`/effort low|medium|high|max` — adjust reasoning depth (where supported).
+
+Switches are session-scoped by default; persist via `/model save` or the web UI.
+
+---
+
+## 6. The chat session model
+
+Alvin Bot separates **sessions** from **users**. Every platform identity maps to a session, and sessions carry:
+
+- Current provider and fallback chain
+- Conversation history (bounded — older messages age out)
+- Active skills and personality flags
+- Working directory (from `/dir`)
+- Locale (`/language`)
+
+Session isolation is configurable via the `SESSION_MODE` env var:
+
+- `per-user` (default) — one shared session per authorized user across all their platform identities. `/new` on Telegram wipes the history for the TUI too.
+- `per-channel` — each platform gets its own session. `/new` on Telegram leaves the TUI untouched.
+- `per-channel-peer` — each platform × user combo gets its own session. Useful when you share a bot between family members.
+
+Memory (the long-term knowledge base at `~/.alvin-bot/memory/`) is **always** per-user, regardless of session mode.
+
+### 6.1 Long-running sessions
+
+Alvin Bot's defining feature is that **he can think for as long as he needs**. There is no absolute timeout on a query — the only cap is an adaptive "stuck" detector that fires if there's no progress (no text, no tool calls) for `ALVIN_STUCK_TIMEOUT_MINUTES` (default: 10). A legitimate 40-minute reasoning run with continuous tool use never gets killed.
+
+This is deliberate and different from most AI chat wrappers. It means Alvin can work overnight on a research task and deliver the answer when you wake up.
+
+### 6.2 Mid-stream abort and resume
+
+If a provider drops the connection mid-stream (network blip, provider restart), the registry silently retries the **same** provider once. If visible text was already streamed to the user, mid-stream failover is disabled — we don't fail over to a different provider mid-answer because that would jumble the response.
+
+---
+
+## 7. Sub-agents
+
+**Sub-agents are parallel isolated workers that run in the background.** Use them when you have multiple independent tasks, when you need to keep a long-running task from blocking your main chat, or when you want to dispatch work from a cron job.
+
+The system distinguishes three sources:
+
+| Source | Origin | Delivery |
+|---|---|---|
+| `implicit` | Main Claude calls the SDK `Task` tool | Parent stream (no separate delivery) |
+| `user` | You ask the bot to spawn a sub-agent | Banner + final to your chat |
+| `cron` | Scheduled job uses `ai-query` | Banner + final to `chatId` in the cron job's target |
+
+### 7.1 Spawning
+
+From chat, the easiest way is to ask Main Claude to spawn one:
+
+> *spawne einen sub-agent der mir in einem Wort 'pong' zurückgibt*
+
+Claude invokes the SDK `Task` tool, which runs as `source: "implicit"`. The result streams back into the parent message.
+
+From a cron job of type `ai-query`, the scheduler calls `spawnSubAgent({ source: "cron" })` internally. The cron job's `target.chatId` becomes the `parentChatId` so the delivery lands in the right chat.
+
+### 7.2 The `/subagents` command
+
+| Command | Effect |
+|---|---|
+| `/subagents` | Show status: max parallel, visibility mode, running list |
+| `/subagents max <n>` | Set max parallel (0 = auto = min(cpu, 16)) |
+| `/subagents visibility auto\|banner\|silent` | Default delivery mode for new spawns |
+| `/subagents list` | Tree view with depth indent and source badges |
+| `/subagents cancel <name\|id>` | Cancel a running sub-agent |
+| `/subagents result <name\|id>` | Show a completed sub-agent's stored result |
+
+### 7.3 Name-first addressing
+
+Sub-agents are addressable by name, not just UUID. If you spawn three agents called `code-review` in quick succession, they appear as `code-review`, `code-review#2`, `code-review#3` in `/subagents list`. The name-resolver handles collisions automatically.
+
+`/subagents cancel code-review#2` cancels the second one specifically. `/subagents cancel code-review` on an unambiguous base name cancels that one. On an ambiguous base name, the bot replies with a disambiguation list.
+
+### 7.4 Depth cap (F2)
+
+Nested spawning is allowed up to depth 2:
+
+- depth 0 = root (spawned by main thread)
+- depth 1 = spawned by a depth-0 agent
+- depth 2 = spawned by a depth-1 agent
+- depth 3 = rejected
+
+This allows the "scatter-gather" pattern (main → orchestrator → 10 workers) without runaway recursion.
+
+### 7.5 Visibility modes
+
+- `auto` (default) — source-based routing: implicit stays in the parent stream, user and cron get a banner+final delivery.
+- `banner` — always send a banner+final, even for implicit spawns.
+- `silent` — never send. The result is still stored in the activeAgents map for 30 minutes and pullable via `/subagents result <name>`.
+- **`live`** — stream incremental updates into a single Telegram message as the agent thinks. Only applies to `source: "user"` spawns with a `parentChatId`. The live message is plain text (so half-formed markdown during streaming can't break the edit), updates are throttled to 800 ms between edits, and a separate banner message is posted at the end so you get a completion notification. If the bot API doesn't support `editMessageText` or the live setup fails, we fall through to `banner` mode automatically.
+
+### 7.6 Inheritance
+
+Sub-agents inherit from the spawning context:
+
+- **Working directory** — the parent's `cwd` (from `/dir`). Opt out with `inheritCwd: false`.
+- **CLAUDE.md** — via the Claude SDK's `settingSources: ["project"]`. Loads automatically if the cwd is inside a project with a CLAUDE.md.
+- **Model and tools** — inherited via the provider registry.
+- **Conversation history** — **not inherited.** Sub-agents receive only their own prompt. This forces clean, self-describing spawn requests.
+
+### 7.7 Bounded priority queue
+
+When the running pool hits `maxParallel`, new spawn requests land in a bounded queue instead of being rejected immediately.
+
+- **Default cap:** 20 slots. Configure via `/subagents queue <n>` (clamped to 0–200).
+- **Disable:** `/subagents queue 0` — restores the old reject-on-full behavior.
+- **Priority order on drain:** `user > cron > implicit`. Within each priority class, FIFO.
+- **`/subagents list`** shows queued entries with a `#N` suffix indicating their position.
+- **Cancel a queued entry** with `/subagents cancel <name>` — it's removed from the queue without ever starting.
+
+Reject is only triggered when the pool **and** the queue are both full. The reject message is priority-aware and names who's holding the slots.
+
+### 7.8 Stats
+
+`/subagents stats` shows a summary of the last 24 hours of sub-agent runs:
+
+- Total runs + total tokens + total wall time
+- Runs per source (user / cron / implicit)
+- Runs per status (completed / cancelled / timeout / error)
+
+The backing data is an append-only JSON ring buffer at `~/.alvin-bot/subagent-stats.json`. Entries older than 24 hours are pruned automatically. A hard cap of 5000 entries protects against runaway growth on very busy bots.
+
+### 7.9 Shutdown notifications
+
+When you restart the bot (SIGTERM), any still-running sub-agents get a cancellation delivery before the process exits:
+
+> ⚠️ `<name>` cancelled · `<duration>` · 0 in / 0 out
+> ⚠️ Agent wurde durch Bot-Restart unterbrochen. Bitte neu triggern.
+
+The shutdown phase is capped at 5 seconds total so a hanging Telegram send can't block the restart.
+
+---
+
+## 8. Cron jobs
+
+Scheduled tasks that run regardless of whether you're chatting with the bot. Use them for daily news digests, weekly reports, periodic health checks, or any recurring automation.
+
+### 8.1 Job types
+
+| Type | Payload | Execution |
+|---|---|---|
+| `reminder` | `text` | Sends the text to the target chat at the scheduled time |
+| `message` | `text` | Same as reminder |
+| `shell` | `command` | Runs the shell command, sends the output |
+| `http` | `url`, `method`, `headers`, `body` | Fires an HTTP request, sends the response |
+| `ai-query` | `prompt` | Spawns a sub-agent with `source: "cron"`, result delivered via I3 |
+
+### 8.2 Scheduling syntax
+
+Supports two formats:
+
+- **Interval** — `30s`, `5m`, `1h`, `6h`, `1d`
+- **Cron** — standard 5-field (`MIN HOUR DAY MONTH WEEKDAY`, 0 = Sunday)
+
+### 8.3 Managing jobs
+
+From Telegram:
+
+```
+/cron              — list all jobs
+/cron add          — create a new one (interactive)
+/cron delete <id>  — remove a job
+/cron toggle <id>  — enable/disable
+/cron run <id>     — trigger a job immediately, outside its schedule
+```
+
+From the command line (useful for scripts):
+
+```
+node scripts/cron-manage.js add \
+  --name "Daily email summary" \
+  --type ai-query \
+  --schedule "0 8 * * *" \
+  --prompt "Check my inbox and summarize unread messages" \
+  --chatId <your-telegram-id>
+
+node scripts/cron-manage.js list
+node scripts/cron-manage.js delete --id <job-id>
+```
+
+The scheduler polls `~/.alvin-bot/cron-jobs.json` every 30 seconds, so hand-edited jobs come into effect at the next tick without a bot restart.
+
+### 8.4 Delivery
+
+ai-query jobs route through the sub-agent delivery router (see §7). Other job types (reminder, shell, http, message) use the legacy delivery-queue path with retry and exponential backoff. Both paths respect the job's `target.platform` and `target.chatId`.
+
+---
+
+## 9. Commands reference
+
+All commands are triggered from any platform that supports commands (Telegram, Discord).
+
+### 9.1 Core
+
+| Command | Purpose |
+|---|---|
+| `/start` | Show the bot intro |
+| `/help` | List all commands |
+| `/new` | Start a fresh session (wipes current history) |
+| `/status` | Current status: provider, model, session age, token usage |
+| `/cancel` | Cancel the currently running request |
+| `/dir <path>` | Change the bot's working directory |
+
+### 9.2 AI control
+
+| Command | Purpose |
+|---|---|
+| `/model` | Switch provider |
+| `/model save` | Persist the current provider as default |
+| `/fallback` | Rearrange fallback chain |
+| `/effort low\|medium\|high\|max` | Set reasoning depth |
+| `/voice on\|off` | Toggle voice replies (ElevenLabs or Edge TTS) |
+| `/language en\|de\|es\|fr` | Switch session locale |
+
+### 9.3 Memory and history
+
+| Command | Purpose |
+|---|---|
+| `/remember <text>` | Add something to long-term memory |
+| `/recall <query>` | Semantic search across memories |
+| `/export` | Export the current conversation to JSON |
+
+### 9.4 Tools
+
+| Command | Purpose |
+|---|---|
+| `/web <query>` | Quick web search |
+| `/imagine <prompt>` | Generate an image (Gemini Nano Banana / Google Imagen) |
+| `/remind 30m <text>` | One-shot reminder |
+
+### 9.5 Bot management
+
+| Command | Purpose |
+|---|---|
+| `/cron` | Manage scheduled jobs |
+| `/subagents` | Show sub-agent status |
+| `/subagents max <n>` | Set max parallel (0 = auto) |
+| `/subagents queue <n>` | Set bounded-queue cap (0 = disabled) |
+| `/subagents visibility <auto\|banner\|silent\|live>` | Delivery mode |
+| `/subagents list` | List all (queued + running + recent) |
+| `/subagents cancel <name\|id>` | Cancel one |
+| `/subagents result <name\|id>` | Show a completed result |
+| `/subagents stats` | Last 24h run stats (by source + status) |
+| `/webui` | Open web UI URL |
+| `/setup` | Re-run the setup wizard flow from chat |
+| `/restart` | Restart the bot process |
+| `/update` | Pull latest, rebuild, restart |
+| `/autoupdate on\|off\|status` | Configure auto-update |
+
+### 9.6 CLI commands (from a terminal)
+
+| Command | Purpose |
+|---|---|
+| `alvin-bot setup` | Initial setup wizard |
+| `alvin-bot start` | Start in background (pm2) |
+| `alvin-bot start -f` | Start in foreground (for debugging) |
+| `alvin-bot stop` | Stop the bot |
+| `alvin-bot tui` | Terminal chat UI |
+| `alvin-bot doctor` | Health check and diagnostics |
+| `alvin-bot audit` | Security audit |
+| `alvin-bot search <query>` | Search assets, memories, and skills |
+| `alvin-bot update` | Git pull, npm install, rebuild |
+| `alvin-bot launchd install` | **macOS only** — install as LaunchAgent (see §14.1) |
+| `alvin-bot launchd uninstall` | Remove the LaunchAgent |
+| `alvin-bot launchd status` | Show LaunchAgent state + recent logs |
+| `alvin-bot version` | Print version |
+
+---
+
+## 10. Platforms
+
+### 10.1 Telegram
+
+The primary platform. Full support: streaming messages, inline keyboards, voice transcription (via OpenAI Whisper or Edge TTS), photo understanding, document ingestion, sticker replies, command autocomplete.
+
+Setup: `BOT_TOKEN` from BotFather + your user ID in `ALLOWED_USERS`.
+
+### 10.2 WhatsApp
+
+Via WhatsApp Web using `whatsapp-web.js`. Two modes:
+
+- **Self-chat** — treat your own WhatsApp number as a personal assistant notepad
+- **Group whitelist** — the bot only responds in explicitly allowed groups, with per-contact access control inside those groups
+
+Group messages go through an approval flow: the owner gets a Telegram notification (or WhatsApp DM fallback) before the bot responds, so there's no surprise chatter from the bot in a group.
+
+Setup: scan the QR code printed on bot startup. Session cookies persist in `~/.alvin-bot/.wwebjs_auth/`.
+
+### 10.3 Discord
+
+Bot with mention detection, slash commands, voice channel support. Uses the standard Discord bot token flow.
+
+Setup: `DISCORD_TOKEN` in `.env`, invite the bot to your server with the correct intents (Message Content + Guild Messages).
+
+### 10.4 Signal
+
+Via `signal-cli` REST API. Less feature-rich than the others because of Signal's tight UI model, but text, images, and voice messages work.
+
+### 10.5 Web UI
+
+A full dashboard at `http://localhost:3100` (configurable via `WEB_PORT`). Password-protected via `WEB_PASSWORD`. Features:
+
+- Live chat with streaming
+- Dashboard: provider status, heartbeat, resource usage
+- Memory manager: browse and edit long-term memories
+- File browser: navigate and edit files the bot has access to
+- Cron job editor
+- Settings and env editor
+- Plugin and MCP server manager
+- Web terminal (in-browser shell)
+- Security audit report
+
+### 10.6 TUI
+
+`alvin-bot tui` starts a terminal chat with ANSI colors and streaming. Useful on remote servers without a browser.
+
+### 10.7 Slack
+
+Uses `@slack/bolt`. Supports DMs and channel mentions.
+
+---
+
+## 11. Plugins and MCP servers
+
+### 11.1 Plugins
+
+JavaScript modules under `plugins/` that register new Telegram commands and extend the bot's capabilities. Each plugin exports:
+
+- `name`
+- `version`
+- `commands` (array of `{ name, description, handler }`)
+- `init(ctx)` — called at startup, gets access to the bot instance
+- `unload()` — called at shutdown
+
+Built-in plugins:
+
+| Plugin | Purpose |
+|---|---|
+| `calendar` | Google Calendar integration |
+| `email` | Gmail / IMAP ingest |
+| `finance` | Bank account queries |
+| `notes` | Apple Notes / Obsidian sync |
+| `smarthome` | Home Assistant control |
+| `weather` | Weather forecasts |
+
+User plugins go in `~/.alvin-bot/plugins/` and hot-reload on save.
+
+### 11.2 MCP servers
+
+Model Context Protocol servers extend the bot with new tools. Configure them in `~/.mcp.json` or `~/.alvin-bot/mcp.json`:
+
+```json
+{
+  "mcpServers": {
+    "filesystem": {
+      "command": "npx",
+      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path"]
+    }
+  }
+}
+```
+
+The bot auto-discovers MCP tools at startup and makes them available to the AI provider (Claude SDK supports them natively; other providers get the tools as descriptions in the system prompt).
+
+---
+
+## 12. Skills
+
+Skills are Markdown files under `skills/` that teach the AI how to do specific tasks. Each skill is a `SKILL.md` file with:
+
+- Title and description
+- When to use it (triggers)
+- Step-by-step procedure
+- Code templates and examples
+
+Skills auto-activate based on message context. Built-in skills include:
+
+- **apple-notes** — create and search Apple Notes
+- **browser-automation** — control Chrome via Playwright
+- **code-project** — scaffold, refactor, and test code projects
+- **data-analysis** — pandas/jupyter-style data exploration
+- **document-creation** — generate PDFs and Word docs
+- **email-summary** — triage and summarize email
+- **github** — issues, PRs, releases
+- **summarize** — generic summarization with citations
+- **system-administration** — shell tasks on Linux/macOS
+- **weather** — forecast queries
+- **web-research** — structured web research with sources
+
+User skills go in `~/.alvin-bot/skills/` and are hot-reloaded.
+
+---
+
+## 13. Security
+
+Security is taken seriously because this bot has real system access.
+
+### 13.1 Authentication
+
+Only users listed in `ALLOWED_USERS` can chat with the Telegram bot. The `authMiddleware` silently drops everyone else.
+
+For WhatsApp groups, each group must be explicitly whitelisted and the owner approves each message before the bot responds.
+
+For the web UI, access is gated by `WEB_PASSWORD`.
+
+### 13.2 Execution sandboxing
+
+The `EXEC_SECURITY` env var controls how shell commands are gated:
+
+- `full` — no restrictions (use only on trusted machines)
+- `allowlist` (default) — only whitelisted binaries from `~/.alvin-bot/exec-allowlist.txt` can run
+- `deny` — shell execution completely disabled
+
+### 13.3 Security audit
+
+`alvin-bot audit` runs a series of health checks:
+
+- Are secrets in `.env` readable only by you?
+- Is the data directory mode `0700`?
+- Are there any API keys leaked into version-controlled files?
+- Is the bot exposed on a public port?
+- Are the allowed users list and exec allowlist intact?
+
+Each check reports `PASS` / `WARN` / `FAIL` with a fix suggestion.
+
+### 13.4 Checkpoints during long runs
+
+Alvin's Claude SDK provider injects a `[CHECKPOINT]` reminder into the prompt every 5 turns once certain thresholds are crossed (10+ messages or 15+ tool calls in a session). The reminder tells Claude to write a checkpoint to its memory file before processing the next request. This protects against context compaction losing progress on long multi-hour sessions.
+
+### 13.5 Git safety
+
+The bot's own code is protected against accidental destruction:
+
+- Never changes user passwords in projects
+- Never touches auth functions without explicit instruction
+- Never runs `git push --force` on main/master
+- Never bypasses pre-commit hooks (`--no-verify`)
+- Never amends commits that might be lost
+
+---
+
+## 14. Running in production
+
+### 14.1 macOS: `launchd` (recommended)
+
+On macOS with the Claude Code SDK, you **must** run the bot as a `launchd` user agent, not via pm2. Claude Code stores the OAuth token for the Max subscription in the macOS Keychain under the service `"Claude Code-credentials"`, and only processes running inside the user's GUI login session can access that Keychain without a manual unlock.
+
+Pm2 starts its daemon in a detached launchd context without Keychain access. Result: the bot sees `Not logged in` on every Claude SDK call. Launchd user agents, on the other hand, run inside the GUI session and inherit the unlocked Keychain automatically.
+
+```bash
+alvin-bot launchd install
+```
+
+This writes `~/Library/LaunchAgents/com.alvinbot.app.plist` with:
+
+- `RunAtLoad true` — starts on login
+- `KeepAlive Crashed=true` — auto-restart on crash
+- `ThrottleInterval 10` — prevents restart loops
+- `PATH` covering Apple Silicon and Intel Homebrew plus `~/.local/bin`
+- stdout → `~/.alvin-bot/logs/alvin-bot.out.log`
+- stderr → `~/.alvin-bot/logs/alvin-bot.err.log`
+
+Management:
+
+```bash
+alvin-bot launchd status       # plist state, PID, last log lines
+alvin-bot launchd uninstall    # unload and remove the plist
+
+# Restart:
+launchctl kickstart -k gui/$UID/com.alvinbot.app
+```
+
+If you'd previously set up the bot via pm2, uninstall the pm2 process first (`pm2 delete alvin-bot; pm2 kill`) before running `launchd install`.
+
+### 14.2 Linux / Windows: pm2
+
+The `alvin-bot start` command wraps pm2:
+
+```bash
+alvin-bot start       # start in background, auto-restart on crash
+alvin-bot start -f    # start in foreground (for debugging)
+alvin-bot stop        # stop the bot
+pm2 logs alvin-bot    # live logs
+```
+
+For systemd users, you can write a user unit pointing at `node dist/index.js` — see the example in `docs/systemd.md` (if included in your release).
+
+### 14.3 Auto-update
+
+`/autoupdate on` tells the bot to check for updates daily. When an update is found, it runs `git pull`, `npm install`, `npm run build`, and restarts itself. The auto-update flag lives at `~/.alvin-bot/auto-update.flag`.
+
+### 14.4 Environment variables reference
+
+Common env vars (all optional unless marked):
+
+| Variable | Default | Purpose |
+|---|---|---|
+| `BOT_TOKEN` | — | **Required**. Telegram bot token. |
+| `ALLOWED_USERS` | — | **Required**. Comma-separated Telegram user IDs. |
+| `PRIMARY_PROVIDER` | `claude-sdk` | Registry key of the default provider. |
+| `FALLBACK_PROVIDERS` | — | Comma-separated fallback chain. |
+| `GOOGLE_API_KEY` | — | Required if using `google` as provider. |
+| `GROQ_API_KEY` | — | For custom Groq provider. |
+| `OPENAI_API_KEY` | — | For custom OpenAI provider. |
+| `OPENROUTER_API_KEY` | — | For custom OpenRouter provider. |
+| `NVIDIA_API_KEY` | — | For custom NVIDIA NIM provider. |
+| `ELEVENLABS_API_KEY` | — | Voice synthesis (falls back to Edge TTS). |
+| `WEB_PORT` | `3100` | Web UI port. |
+| `WEB_PASSWORD` | — | Web UI access password. |
+| `SESSION_MODE` | `per-user` | Session isolation mode. |
+| `ALVIN_STUCK_TIMEOUT_MINUTES` | `10` | Adaptive stuck-detector window. |
+| `MAX_SUBAGENTS` | `0` (auto) | Max parallel sub-agents. |
+| `SUBAGENT_TIMEOUT` | `300000` | Per-agent timeout in ms. |
+| `COMPACTION_THRESHOLD` | `80000` | Tokens before context compaction triggers. |
+| `EXEC_SECURITY` | `allowlist` | Shell execution sandbox mode. |
+| `ALVIN_DATA_DIR` | `~/.alvin-bot` | Data directory path. |
+
+---
+
+## 15. Troubleshooting
+
+### 15.1 Bot starts but doesn't respond
+
+1. `alvin-bot doctor` — runs 10+ health checks
+2. Check the log: `pm2 logs alvin-bot` (or `~/.alvin-bot/logs/alvin-bot.err.log` under launchd)
+3. Verify your provider: `/status` in chat shows the active provider and its health
+
+### 15.2 Claude SDK says "Not logged in"
+
+You're hitting the Keychain problem documented in §14.1. Three paths:
+
+- **Preferred**: switch to `alvin-bot launchd install` on macOS
+- **Temporary**: `security unlock-keychain -p '<your-login-password>' ~/Library/Keychains/login.keychain-db`, then restart the bot in the same terminal session
+- **Fallback**: run interactively: `claude` (opens the interactive CLI), then `/login`, then re-run the bot
+
+### 15.3 Sub-agent limit reached
+
+Either increase the cap:
+
+```
+/subagents max 8
+```
+
+Or cancel a stuck agent:
+
+```
+/subagents list
+/subagents cancel <name>
+```
+
+### 15.4 Cron job not firing
+
+- Is it `enabled: true`? `/cron` or `cat ~/.alvin-bot/cron-jobs.json`
+- Is the schedule valid? Try `/cron run <id>` to trigger manually
+- Check the bot log for `[cron]` lines
+
+### 15.5 Web UI shows "Not authorized"
+
+`WEB_PASSWORD` isn't set or doesn't match. Add `WEB_PASSWORD=<pick-a-password>` to `.env` and restart.
+
+### 15.6 Sub-agent results shown twice
+
+Happens if you upgrade from a version before 4.6.0 and a mid-stream shutdown race fired. Fixed in 4.6.0 — see CHANGELOG §9.2.
+
+### 15.7 Upgrading from 4.5.x
+
+- Run `npm install -g alvin-bot@latest`
+- Run `alvin-bot update` from chat (pulls, rebuilds, restarts)
+- On macOS, consider switching from pm2 to `alvin-bot launchd install` for the Keychain fix
+
+---
+
+## 16. Upgrading
+
+### 16.1 Data directory compatibility
+
+Alvin Bot follows semver for the **data directory format**. Minor version bumps never break `~/.alvin-bot/`; major bumps might introduce migration scripts that run automatically on first startup.
+
+### 16.2 From 4.5.x to 4.6.0
+
+- Sub-agents: new fields in `sub-agents.json` (`visibility`). Old files auto-upgrade — the loader treats missing fields as defaults.
+- Cron jobs: no schema change.
+- Memory: no schema change.
+- `.env`: no new required variables. `MAX_SUBAGENTS` and `SUBAGENT_TIMEOUT` are optional.
+
+### 16.3 From 4.6.x to 4.7.0
+
+- Sub-agents: new fields in `sub-agents.json` (`queueCap`, defaults to 20). Old files auto-upgrade.
+- New file `~/.alvin-bot/subagent-stats.json` — auto-created when the first sub-agent finishes.
+- `start`/`stop` now auto-detect the LaunchAgent on macOS. No migration needed; if you previously installed the LaunchAgent in 4.6.0, `alvin-bot start` now correctly reloads it instead of spawning a parallel pm2 process.
+- No new required `.env` variables.
+
+### 16.3 From git
+
+```bash
+cd ~/path/to/alvin-bot
+git pull
+npm install
+npm run build
+
+# If running under launchd:
+launchctl kickstart -k gui/$UID/com.alvinbot.app
+
+# If running under pm2:
+alvin-bot start
+
+# If running in foreground:
+# Stop with Ctrl+C, then alvin-bot start -f
+```
+
+### 16.4 From npm
+
+```bash
+npm install -g alvin-bot@latest
+alvin-bot stop
+alvin-bot start
+```
+
+### 16.5 Rollback
+
+```bash
+npm install -g alvin-bot@4.5.1
+```
+
+Old data in `~/.alvin-bot/` continues to work — nothing is deleted on upgrade.
+
+---
+
+## Feedback and bug reports
+
+- GitHub Issues: <https://github.com/alvbln/Alvin-Bot/issues>
+- Pull requests welcome
+- Maintained by @alvbln