Shadow Stack Widget — SKILL.md

Electron + React + Vite виджет для автоматизации macOS/M1 окружения.

Быстрый старт

cd ~/shadow-stack-widget

# GUI режим (виджет)
npm run start

# Headless режим (CLI)
npm run headless -- --step=0
npm run headless -- --all
npm run headless -- --all --debug

---

Режимы работы

🖥️ GUI Mode

npm run start

• Electron окно-виджет
• Интерактивный Autopilot
• Логи в реальном времени
• Ollama status indicator

💻 Headless Mode

npm run headless -- [опции]

Опция	Описание
--step=N	Выполнить шаг N
--all	Выполнить все шаги
--debug	Debug логи
-d	То же что --debug

Примеры:

npm run headless -- --step=0        # System audit
npm run headless -- --step=1         # Homebrew/Node
npm run headless -- --all            # Все шаги
npm run headless -- --all --debug    # С debug логами

---

Shadow Steps (0-7)

ID	Название	Описание
0	Разведка и аудит	Системный аудит: чип, память, диск, софт
1	Homebrew, Node, Python	Установка node@22, python@3.12
2	Ollama + модели	llama3.2, phi3, mistral, qwen2.5
3	OpenClaw Agent	opencode-ai CLI
4	NeMo Agent Toolkit	nemotoolkit
5	OpenCode SDK	@opencode-ai/sdk
6	Vercel AI SDK v4+	ai + @ai-sdk/openai
7	Playwright MCP	@playwright/mcp

---

Debug логи

Включить:

npm run headless -- --all --debug

Лог-файл:

tail -f /tmp/shadow-widget.log

Custom log path:

SHADOW_LOG=/path/to/log npm run headless -- --all

Что логируется:

• Startup + args
• IPC вызовы (execute-bash, set-always-on-top)
• Comet/Ollama запросы
• Ошибки и fallback-и

---

Environment Variables

VITE_DEV_SERVER_URL=http://127.0.0.1:5175
COMET_ENDPOINT=http://localhost:8080/api/fix
COMET_API_KEY=your-key
SHADOW_LOG=/tmp/shadow-widget.log

---

IPC API (preload.cjs)

window.electronAPI.runCommand(cmd); // Bash команда
window.electronAPI.setAlwaysOnTop(flag); // Пин окна
window.electronAPI.setFullscreen(flag); // Полноэкранный
window.electronAPI.sendErrorToComet(payload); // Ошибка → Ollama
window.electronAPI.getPlatform(); // Platform info

---

Ollama Integration

Port: 11434

Models: llama3.2, phi3, mistral, qwen2.5:3b, llava

Check status:

ollama list
curl http://localhost:11434/api/tags

Restart:

brew services restart ollama

---

Troubleshooting

Electron SIGTERM

npm run kill-port && npm run start

Port 5175 занят

npm run kill-port

Ollama не отвечает

brew services restart ollama
ollama pull llama3.2

Debug mode

npm run headless -- --step=0 --debug
tail -f /tmp/shadow-widget.log

---

Файлы проекта

~/shadow-stack-widget/
├── main.cjs           # Electron main + headless
├── preload.cjs        # IPC bridge
├── vite.config.js     # Port 5175
├── postcss.config.js  # Tailwind v4
├── package.json       # "main": "main.cjs"
├── AGENTS.md          # OpenCode rules (8 фаз)
├── SKILL.md           # This file
└── src/
    ├── App.jsx        # React widget
    ├── index.css
    └── main.jsx

---

Stack

• Electron 41 + Node 22
• Vite 8 + React 19 + Tailwind CSS v4
• Ollama (local LLM)
• OpenCode AI CLI

---

Auto-Router v1.0

Routing Rules

routing_rules:
  command_routes:
    /deploy:   { route: vercel-deploy, provider: vercel, model: vercel-cli }
    /premium:  { route: claude-premium, provider: anthropic, model: claude-sonnet-4 }
    /deep:     { route: perplexity, provider: perplexity, model: sonar }
    /grok:     { route: grok, provider: xai, model: grok-2 }
    /kimi:     { route: kimi, provider: moonshot, model: moonshot-v1-8k }

  text_routes:
    short: { condition: "text.length < 80", route: ollama-short, model: qwen2.5-coder:3b }
    long:  { condition: "text.length >= 80", route: ollama-analysis, model: qwen2.5:7b }

  default: { route: openrouter-free, model: qwen/qwen-2.5-coder-32b-instruct }

Providers Priority

providers_priority:
  1_local_free:    [ ollama ]           # $0, fastest, RAM only
  2_proxy_free:    [ antigravity ]      # $0, Gemini via OpenCode proxy
  3_cloud_free:    [ openrouter ]       # $0, API key required
  4_subscription:  [ perplexity ]       # monthly sub
  5_api_key:       [ kimi ]             # API key, per-token
  6_paid:          [ anthropic, openai ] # per-token, /premium only

Fallback Chain

fallback_chain:
  on_error: [ 429, 5xx, timeout ]
  retry: { max: 3, backoff: exponential, delays: [1s, 2s, 4s] }
  cascade:
    - ollama qwen2.5-coder:3b          # 1. try local first
    - antigravity gemini-2.0-flash      # 2. Gemini via OpenCode proxy ($0)
    - openrouter qwen-2.5-32b          # 3. free cloud tier
    - kimi moonshot-v1-8k              # 4. API key provider
    - claude claude-sonnet-4            # 5. paid, last resort
  quota_threshold: 90%                  # switch to next when 90% quota used

Usage Monitoring

usage_tracking:
  storage: in-memory (resets hourly)
  limits:
    ollama: unlimited
    antigravity: 50/hr
    openrouter: 100/hr
    perplexity: 30/hr
    kimi: 50/hr
    claude: 20/hr
  endpoints:
    GET /api/auto-router/usage    # JSON quota stats
    /usage (Telegram)             # visual bars

State Machine

IDLE → ROUTING → EXECUTING → DONE
                      ↓
                    ERROR → Fallback Cascade → DONE | ERROR

Telegram Commands

Command	Action	Cost
/help	Show command list	$0
/status	Query Ralph Loop orchestrator	$0
/deploy	Vercel production deploy	$0
/premium	Claude Sonnet API	$$
/deep	Perplexity search	sub
/grok	Grok AI	sub
/kimi	Kimi (Moonshot)	key
/reset	Reset session	$0
/test-router	Test routing (shows route info)	$0*
/usage	Provider quota stats (visual)	$0
(text)	Auto-route by length	$0

* /test-router still calls the LLM, but shows routing metadata

Meta-Escalation

meta_escalation:
  trigger: all_providers_failed OR manual /escalate command
  chain:
    tier_1: Perplexity API (search + context)
    tier_2: GPT-4o via OpenRouter (reasoning)
    tier_3: Telegram human (async, WAITING_FOR_HUMAN)
  endpoint: POST /api/meta-escalate { "problem": "..." }
  telegram: /escalate <description>
  statuses: [ resolved, waiting_for_human, all_failed ]

API Endpoints

# Auto-Router (Express, port 3001)
curl -X POST http://localhost:3001/api/auto-router \
  -H 'Content-Type: application/json' \
  -d '{"text": "hello"}'

# Meta-Escalation
curl -X POST http://localhost:3001/api/meta-escalate \
  -H 'Content-Type: application/json' \
  -d '{"problem": "All providers failed for query X"}'

# Usage Monitoring
curl http://localhost:3001/api/auto-router/usage

# Telegram Webhook (Next.js)
# Set via: https://api.telegram.org/bot$TOKEN/setWebhook?url=$VERCEL_URL/api/telegram-webhook