// the find
TNT-Likely/PanWatch
PanWatch — AI stock monitoring for A-shares, HK & US markets, powered by TradingAgents. Portfolio insights, real-time alerts & automated reports.｜盯盘侠：覆盖 A股/港股/美股的 AI 盯盘、持仓分析、实时提醒与自动报告。
A self-hosted dashboard for watching A-share/HK/US stocks that wires in TauricResearch's TradingAgents for multi-agent deep analysis — bull/bear debate, risk review, PM decision — on top of portfolio tracking, price alerts, and paper trading. Aimed at individual traders who want LLM-assisted research on their own holdings without handing data to a SaaS, and who are comfortable running their own OpenAI-compatible or Ollama backend.
The TradingAgents integration is a real architecture, not a thin GPT wrapper — it runs analyst roles through a debate and risk-review pipeline and keeps the full reasoning trail, which is more honest than a single-shot 'buy/sell' prompt. Built-in observability (trace_id per agent run, agent_runs table, node-level cost/progress) works with zero external services, and optional OTel export to Jaeger/Tempo/Langfuse is a genuine no-op when unused, not a half-wired feature. Market-calendar handling across three exchanges — correct local trade dates, half-days, DST transitions for US paper-trading fills, blocking execution on unpublished dates instead of silently running — shows someone actually hit these edge cases in production rather than assuming UTC everywhere. Notification coverage (Telegram, WeCom, DingTalk, Feishu, Bark, webhooks) is broader than the typical English-first alerting tool, which matters a lot for its actual user base.
Auth is single-user by design (AUTI_USERNAME/PASSWORD env vars plus a JWT) — fine for a personal appliance, useless if you wanted to share it with family or a small group. The annual market-calendar data is bundled and manually maintained; the README admits this outright, and a missed update means alerts and scheduled agents quietly stop firing on days that should be trading days. First container startup downloads a full headless Chromium via Playwright just for screenshot generation, which is a heavy, network-dependent cold start for something billed as a lightweight self-hosted service — you have to know to set PLAYWRIGHT_SKIP_BROWSER_INSTALL if you don't want that. Cost and runtime of the deep-analysis pipeline are explicitly 'depends on the model you pick,' with no built-in budget cap or per-run cost ceiling, so someone running scheduled pre-market/intraday/closing agents across a real portfolio can rack up API spend without much warning.