AI Orchestrator — DeepSeek via Puppeteer

Launch DeepSeek AI via Puppeteer browser automation with CDP interceptor for full API responses, persistent daemon for fast startup, health checks, graceful...

Security Scan

The skill mostly does what it says (Puppeteer-based DeepSeek orchestration) but includes several intrusive local operations and persistent artifacts (process-killing, rm -rf on temp dirs, PM2 restarts, persistent Chrome profile and diagnostic logs) that are plausible for this purpose but warrant caution before installing.

Purpose & Capability

Name/description (Puppeteer + CDP + persistent daemon for DeepSeek) match the code and SKILL.md: ask-puppeteer.js, deepseek-daemon.js, healthcheck and diagnostics implement exactly that functionality. Requests (no env vars) align with a web-automation skill; requiring PM2 and npm is expected for a Node daemon.

Instruction Scope

Runtime instructions and code perform a wide set of local actions beyond simply opening pages: they create persistent Chrome user-data (.profile) that will contain cookies/session, write diagnostics (.diagnostics/, logs), read/write endpoint and session files, forcibly kill Chrome processes for a profile, remove lock files and prune /dev/shm and /tmp Chrome entries, and run shell commands (pm2 restart etc.). While coherent for managing a persistent browser daemon, these actions touch system processes, temp directories and local data and will store user prompts/trace data on disk — this is broader scope than a stateless API client and may expose sensitive prompts or credentials if the browser is logged in.

Install Mechanism

No formal install spec in registry but SKILL.md instructs running 'npm install' and using PM2. Code is provided (package.json and package-lock.json) so npm will install dependencies locally. This is a moderate-risk install pattern (npm packages are common); the package files are bundled here so there is no external arbitrary URL download, but running npm install and PM2 will install/enable persistent processes on the host.

Credentials

The skill declares no required environment variables or credentials, which matches the code. However it persists a Chromium user-data dir under the skill (.profile) and expects you to authenticate manually — that directory will contain cookies/session tokens for DeepSeek once you log in. Diagnostic traces and metrics (written to .diagnostics/) may include prompt metadata. No external credential exfiltration was observed, but local persistence of session cookies means the skill can make authenticated requests on behalf of whoever logs in.

Persistence & Privilege

The skill creates a persistent daemon (PM2-managed deepseek-daemon) and writes a browser websocket endpoint file (.daemon-ws-endpoint) and profile data (.profile). Healthcheck code can auto-restart the daemon via PM2 and will call 'pm2 restart' or other shell commands. It does not set always:true, but it requests persistent presence via PM2 and modifies local system state (files, processes). The ability to kill Chrome processes and execute restarts is powerful — justified for daemon management but increases blast radius if misused.

What to consider before installing

This skill implements a persistent Puppeteer-based daemon for interacting with DeepSeek and is largely coherent with its description — but it performs several intrusive local operations you should review and accept before installing: - Persistent browser profile: the daemon uses a local Chrome user-data dir (.profile) that will store login cookies and session state; if you log into DeepSeek in that profile, the skill can act using that account. Consider using an isolated account or container if that is a concern. - Diagnostic traces and logs: the skill writes trace/metrics files to .diagnostics/ and may include prompt metadata; do not send sensitive secrets in prompts unless you are comfortable with local logs retaining them. - Process and filesystem operations: code may kill Chrome processes tied to the profile, remove Chrome-related files in /dev/shm and /tmp, and delete/rotate local files. These are intrusive but explained as remediation for locked profiles; run as an unprivileged user and avoid running as root. - PM2 control: healthcheck and daemon scripts call pm2 (restart/start/stop). If you do not want a persistent background process, avoid running the PM2 steps and run the script in ephemeral mode (without --daemon / PM2). - Review code: because the skill runs shell commands (child_process.execSync) and manipulates local files, inspect ask-puppeteer.js, daemon-healthcheck.js and deepseek-daemon.js yourself. If you plan to run this on a sensitive system, run it inside a sandbox/container or a separate user account. In summary: the behavior is plausible for a persistent Puppeteer orchestrator, not obviously exfiltrative, but it carries non-trivial local privileges and persistence. If you want to proceed, run it in an isolated environment, avoid using your main DeepSeek account or any account containing sensitive data, and audit/limit PM2 usage and log retention.

AI Orchestrator

Production-ready DeepSeek AI access via Puppeteer + CDP interceptor.

Quick Start

# Simple
ask-deepseek.sh "What is HTTP?"

# Session (keeps context)
ask-deepseek.sh "Explain OAuth2" --session work
ask-deepseek.sh "What about OpenID Connect?" --session work
ask-deepseek.sh --session work --end-session

# With daemon (faster, ~35ms startup vs ~15s cold)
ask-deepseek.sh "Question" --daemon

# Pipe / heredoc support
cat code.py | ask-deepseek.sh "Find bugs"
ask-deepseek.sh <<'EOF'
Long multi-line prompt
EOF

All Flags

Flag	Purpose
--session NAME	Persistent context across requests
--daemon	Use running Chrome daemon (fast startup ~35ms)
--search	Enable DeepSeek web search
--think	Enable DeepThink mode
--new-chat	Start new chat within session
--end-session	Close session
--visible	Open visible browser (for auth/CAPTCHA)
--wait	Wait for manual auth (with --visible)
--close	Force close browser after request
--dry-run	Test auth + composer without sending prompt
--debug	Verbose debug output
-h, --help	Show help

Daemon Setup

# First time
cd ~/.openclaw/workspace/skills/ai-orchestrator
npm install

# Start daemon
pm2 start deepseek-daemon.js --name deepseek-daemon --no-autorestart
pm2 save

# Auto-start on boot
pm2 startup

# Status check
pm2 status deepseek-daemon
pm2 logs deepseek-daemon --lines 20 --nostream

DeepSeek Free Tier Limits

• Per response: ~13,000 characters (verified 2026-04-03)
• Button "Continue" may appear — auto-clicked up to 30 times
• For longer answers: use --session with chained questions
• Rate limit: 5 seconds between requests

Workaround for Long Answers (>13k chars)

# Start session
ask-deepseek.sh "Write a comprehensive Python backend guide covering data structures and OOP" --session guide

# Continue in same conversation
ask-deepseek.sh "Now cover async programming, databases, and REST APIs" --session guide

# End when done
ask-deepseek.sh --session guide --end-session

Configuration (.deepseek.json)

Override any default by creating/editing .deepseek.json:

{
  "idleTimeout": 20000,
  "maxContinueRounds": 15,
  "logToFile": true,
  "logPath": ".logs/deepseek.log"
}

Diagnostics

Each request generates trace + metrics files in .diagnostics/:

• trace-<id>.jsonl — phase-by-phase JSONL log
• metrics-<id>.json — memory, timing, counts summary

Troubleshooting → REFERENCE.md

See REFERENCE.md for:

• Architecture overview
• Complete troubleshooting with solutions
• Diagnostic system details
• Performance benchmarks
• Common issues (CAPTCHA, rate limits, selector changes)