qwen-orchestrator（Qwen 编排器）

Qwen Orchestrator v1.5.8

What this is: Browser automation that talks to Qwen Chat via Puppeteer. Default runtime policy: before sending a prompt, qwen-orchestrator must switch the Qwen model selector to Qwen3.6-Max-Preview and then switch the Qwen mode selector to thinking mode when those selectors are available. 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 qwen-orchestrator commands from the skill root: ~/.openclaw/workspace/skills/qwen-orchestrator. If you are not already there, run:

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

Quick Start

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

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

# With daemon (fast startup ~35ms vs ~15s cold)
ask-qwen.sh "Question" --daemon

# With web search
ask-qwen.sh "Latest news about AI" --search

# Pipe content
cat code.py | ask-qwen.sh "Find bugs"
ask-qwen.sh --stdin < code.py
ask-qwen.sh <<'EOF'
Long multi-line prompt here
EOF

Before you start

• Fastest normal path: prefer --daemon
• Safe readiness check: run ask-qwen.sh --dry-run --daemon
• Need context across calls: add --session NAME
• Need browser automation / scraping / clicking / forms: use agent-browser, not qwen-orchestrator

When to use qwen-orchestrator

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

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

Follow-up continuity invariant: after restoring an existing session chat, the next prompt send must remain bound to that same chat unless --new-chat was explicitly requested. Before follow-up submit, qwen-orchestrator should prefer the cheapest safe continuity check first: navigate to the saved chatUrl only if the current tab is not already bound to that chat, wait for the composer on that exact chat, and require hydrated follow-up chat state rather than a shell-only page before submitting. Hydration readiness now means visible prior history, not just URL + composer presence. If the restored page is still shell-only, qwen-orchestrator must use exact-URL hard rebind rather than sidebar title activation, because duplicate chat titles can bind to the wrong conversation. Submit should stay bound to the composer/send control for that chat, not a page-global Enter fallback unless no safer path exists. If submit still jumps to /c/new-chat or a different chat id, treat it as a continuity failure; do not silently accept or persist the new chat as the session state.

All Flags

Health check

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

# 2. Dry-run path works?
ask-qwen.sh --dry-run --daemon || echo "Qwen dry-run failed"

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

Failure Recovery (mandatory routing)

Symptom → exact command

Failure Recovery (English quick table)

Daemon Setup

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

Exit Codes

Minimal Contract & Known Limitations

What you can rely on

• Exit codes:

- 0 → success, response printed to stdout - 1 → fatal config/arg error or generic unrecoverable runtime failure - 2 → Qwen/daemon unavailable; restart or repair daemon connectivity before retrying - 3 → follow-up continuity failure; the stored chat binding was not preserved safely

• Session continuity: with --session NAME + --daemon, qwen-orchestrator keeps the saved chatUrl unless --new-chat was explicitly requested.
• Daemon behavior: warm daemon calls are the normal fast path; local mode must not race against the same Chromium profile while qwen-daemon is active.

Non-fatal warnings

These warnings may still return exit code 0 if a response was produced:

• thinking-mode selector could not be opened or confirmed
• web-search toggle could not be confirmed even though the request continued

If the task depends on deep reasoning or web search specifically, treat these warnings as a signal to retry with --visible or manually verify mode state.

Known limitations

• Selector drift: Qwen UI changes can break the thinking/search selectors before the rest of the flow fails.
• Long answers: use --session and continuation prompts for large outputs.
• Oversized prompts: Qwen web UI can silently refuse an oversized single-shot submit before any API request is emitted. qwen-orchestrator now validates whether submit really left the page, retries with adaptive shortening when needed, and fails explicitly with QWEN_SUBMIT_BLOCKED_OVERSIZE_PROMPT if delivery never succeeds.
• Continuity failures: if exit code 3 occurs, end the affected session and start it fresh instead of silently trusting the new chat.
• Hydration lag: restored follow-up chats may briefly present the composer before prior messages hydrate; qwen-orchestrator now waits for hydrated chat state before follow-up submit, but severe upstream UI changes can still require visible inspection.

Recovery shortcuts

# Exit 2: Qwen/daemon unavailable
pm2 restart qwen-daemon && sleep 8
bash ask-qwen.sh --dry-run --daemon

# Exit 3: continuity failure
bash ask-qwen.sh --session NAME --end-session
bash ask-qwen.sh "your prompt" --session NAME --daemon

Session file trust boundary

Session state in .sessions/<name>.json is trusted internal state.

• Do not manually edit or swap session files.
• Continuity guarantees apply only while session files are managed exclusively by ask-qwen.sh.
• If a session file looks stale or suspicious, end that session and start a fresh one instead of trying to repair the JSON by hand.

Configuration (.qwen.json)

{
  "browserLaunchTimeout": 30000,
  "answerTimeout": 600000,
  "composerTimeout": 10000,
  "navigationTimeout": 30000,
  "idleTimeout": 15000,
  "heartbeatInterval": 15000,
  "domErrorIdleMs": 25000,
  "rateLimitMs": 5000,
  "maxContinueRounds": 30,
  "preferredModel": "Qwen3.6-Max-Preview",
  "logToFile": true,
  "logPath": ".logs/qwen.log"
}

• rateLimitMs: client-side delay between prompt sends. Helps avoid accidental rapid-fire requests against Qwen. Set to 0 to disable.
• preferredModel: model name that qwen-orchestrator should force before each prompt. Default: Qwen3.6-Max-Preview.

Minimal Decision Rules

• If you need the fastest normal run → --daemon
• If you need context across prompts → --session NAME
• If you need a fresh Qwen thread inside an existing session namespace → --new-chat
• If you need to close stored context → --end-session
• Before sending a prompt → force the Qwen model selector to Qwen3.6-Max-Preview (or .qwen.json.preferredModel) if the selector exists; warn if it cannot be confirmed
• Before sending a prompt → force the Qwen mode selector to thinking mode if the selector exists; warn if it cannot be confirmed
• If you enable search → verify the Web Search toggle really became active; warn if it cannot be confirmed
• Auth repair → stop daemon → --visible --wait --dry-run
• Health check → --dry-run
• Long answers → --session + chunk
• If qwen-daemon is running, prefer --daemon; local mode currently must not run concurrently against the same Chromium profile
• Browser automation / scraping / form filling → use agent-browser, not qwen-orchestrator
• If a follow-up session behaves oddly, prefer reusing the same --session NAME; do not silently switch to a different open Qwen tab
• Warning-only mode/search selector issues may still return success if the response was produced; treat the warning text as diagnostic signal, not as a guaranteed hard failure
• Need deeper runtime details or repair commands → read REFERENCE.md. Purpose: debug failures.