ai-orchestrator（AI 协调器）

AI Orchestrator v2.0.11

What this is: Browser automation that talks to DeepSeek via Puppeteer. What this is NOT: A general AI router or multi-model orchestrator. It's one browser → one LLM.

Execution

Working directory rule: Unless a command here uses an absolute path, run ai-orchestrator commands from the skill root: ~/.openclaw/workspace/skills/ai-orchestrator. If you are not already there, run:

cd ~/.openclaw/workspace/skills/ai-orchestrator

Relative paths such as .daemon-ws-endpoint, .profile/, .deepseek.json, and scripts/setup-daemon.sh assume this directory.

Quick Start

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

# Session (keeps context between questions)
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 (fast startup ~35ms vs ~15s cold)
ask-deepseek.sh "Question" --daemon

# Pipe code or long prompts
cat code.py | ask-deepseek.sh "Find bugs"
ask-deepseek.sh --stdin < code.py
ask-deepseek.sh <<'EOF'
Long multi-line prompt here
EOF

When to use ai-orchestrator

Rule: Need LLM reasoning without API key → ai-orchestrator. Need browser tasks (scrape, click, fill forms) → use agent-browser skill instead.

Default site mode: ai-orchestrator must keep DeepSeek in Expert mode whenever the site exposes the Instant / Expert switch. Treat Expert as the canonical default. Do not fall back to Instant unless a future implementation explicitly adds an override and the user asks for it.

Session continuity: in daemon mode, follow-up requests using --session NAME must restore the exact saved DeepSeek chatUrl for that session unless --new-chat was explicitly requested. Reusing an arbitrary open DeepSeek tab is incorrect because it breaks multi-round review continuity.

All Flags

Health check

# 1. Daemon running?
pm2 status | grep -q "deepseek.*online" || echo "Daemon down"

# 2. Dry-run path works?
ask-deepseek.sh --dry-run --daemon || echo "DeepSeek dry-run failed"  # Checks auth + composer through the daemon path

# 3. Real response path works?
ask-deepseek.sh "OK" --daemon | grep -q "OK" || echo "DeepSeek not responding"

If any step fails, do not improvise. Use the troubleshooting ladder below.

Failure Recovery (mandatory routing)

First failure rule — do not skip

When any command fails:

1. STOP — do not try a random second fix yet
2. CAPTURE — read the exact error you saw
3. MATCH — find the closest symptom row below
4. RUN — execute that row's exact command
5. VERIFY — retry with the specified verification command

Symptom → exact command

If 3 fixes failed already

Do not keep guessing. Reset to a known-good baseline:

cd ~/.openclaw/workspace/skills/ai-orchestrator
pm2 stop deepseek-daemon
pkill -f puppeteer || true
rm -f .daemon-ws-endpoint .profile/Singleton*
bash scripts/setup-daemon.sh
bash ask-deepseek.sh "Say OK" --daemon

If that still fails, escalate to manual auth/inspection with --visible --wait.

Manual Auth with --wait

--wait blocks until you complete auth in the visible browser or interrupt the command yourself. There is no automatic timeout documented here. Use Ctrl+C if you need to abort.

Daemon Setup

One-liner:

cd ~/.openclaw/workspace/skills/ai-orchestrator && bash scripts/setup-daemon.sh

Or manual:

bash scripts/setup-daemon.sh  # Or run step-by-step below
npm install   # First time only
pm2 start deepseek-daemon.js --name deepseek-daemon --no-autorestart
pm2 save
pm2 startup   # Auto-start on boot

Exit Codes

Note: Partial success (truncated response) still returns 0. Use "Continue" button handling.

Known Limitations (v2.0.11)

• Exact-string prompts: DeepSeek may prepend explanation text before the requested suffix. For machine-sensitive checks, prefer structured JSON prompts when possible, or validate that the returned payload still contains the expected terminal token instead of assuming exact literal obedience.
• Oversized prompts: if a giant single-shot prompt does not actually leave the page, ai-orchestrator now validates submit outcome, retries with adaptive shortening, and fails explicitly with DEEPSEEK_SUBMIT_BLOCKED_OVERSIZE_PROMPT if delivery never succeeds.
• Publish hygiene: local runtime state must not be published. Before ClawHub publish, verify that .clawhubignore excludes local profile, diagnostics, logs, sessions, and daemon state files.

DeepSeek Free Tier Limits

• Per response: ~13,000 characters (verified 2026-04-03)
• Button "Continue" auto-clicked up to 30 times
• Rate limit: 5 seconds between requests
• Session limit: Use --session for long conversations instead of single huge prompts

Workaround for Long Answers (>13k chars)

ask-deepseek.sh "Write a Python backend guide" --session guide
ask-deepseek.sh "Now cover async and databases" --session guide  # Continue
ask-deepseek.sh --session guide --end-session  # Done

Logging

After important orchestrator runs, append to memory/episodic/YYYY-MM-DD.md:

## ai-orchestrator
- Task: code review app.js (tool: deepseek, duration: 12s, status: success)
- Failure: CAPTCHA on deepseek (fallback: manual via --visible)

Not every request needs logging. Log decisions and failures, not routine questions.

What NOT to do

• ❌ Use for web scraping or browser automation (wrong tool — use agent-browser skill)
• ❌ Bypass Dual Thinking for complex decisions (orchestrator executes, doesn't think)
• ❌ Ignore failure logs (they tell you what's broken)
• ❌ Run without --daemon unless debugging (cold start = 15s vs 35ms)
• ❌ Guess random fixes when a known failure case already matches
• ❌ Open a visible browser first if the problem is only a dead daemon — restart daemon first
• ❌ Keep reusing one broken long prompt after free-tier cutoff — switch to --session chunks instead

Configuration (.deepseek.json)

{
  "idleTimeout": 15000,
  "heartbeatInterval": 15000,
  "domErrorIdleMs": 25000,
  "maxContinueRounds": 30,
  "logToFile": true,
  "logPath": ".logs/deepseek.log"
}

Minimal Contract for Weak Runtimes

• Exit codes: 0 = success, 1 = config/runtime error, 2 = DeepSeek unreachable or daemon unavailable, 3 = blocked or rate-limited. Never treat 2 or 3 as success.
• Mode enforcement: when DeepSeek shows Instant / Expert, ai-orchestrator must use Expert unless the user explicitly requests otherwise via a future supported override.
• Session restore: --session NAME must restore the exact saved chatUrl unless --new-chat was explicitly requested.
• Publish hygiene: before publishing to ClawHub, ensure .clawhubignore exists and excludes local runtime state.

Minimal Decision Rules for Weak Models

When choosing what to do next, use these defaults:

• Need fastest normal run → use --daemon
• Need context across multiple prompts on one topic → use --session
• Need a brand new DeepSeek chat for that topic → add --new-chat
• Need to close the topic chat → use --end-session
• Need auth/login repair → stop daemon first, then --visible --wait --dry-run
• Need a health check without sending a real prompt → use --dry-run
• Need a long answer → use --session and split it into chunks
• Need browser automation on websites → stop and use agent-browser, not ai-orchestrator`
• Need failure recovery → do not guess; use the symptom table above
• Before every real prompt send, keep the DeepSeek site in Expert mode when that switch exists
• If --think is used, enable the site DeepThink toggle before sending the prompt

Troubleshooting → REFERENCE.md

See REFERENCE.md for:

• Architecture overview (Puppeteer + CDP interceptor)
• Complete troubleshooting guide
• Diagnostic system (trace + metrics + summary JSON + file log)
• Performance benchmarks
• Common issues: CAPTCHA, rate limits, selector changes, visible-browser vs daemon behavior