qoobeeHermes

把静态 HTML 站点部署到服务器对外访问的标准 SOP。覆盖 Caddy(Docker,:80/:443) 与 nginx(宿主,独立端口) 两条路径、端口约定、80 端口冲突坑、部署目录、reload 与 curl 验证。任何"部署/上线/开放访问一个 HTML 页面或静态站点"的请求都必须先读这里，不要自己瞎配。

Diagnose RL agent behavioral issues by analyzing reward structure. Use when the agent exhibits unwanted behavior (random skills, wandering, avoiding combat, etc.) to identify root causes in the reward function.

Author/validate/export Google's DESIGN.md token spec files.

Configure, troubleshoot, and operate the Feishu (Lark) gateway integration — group chats, @mention gating, permissions, message routing, and diagnostics.

Scrape training metrics from Tencent AI Arena (tencentarena.com) monitoring dashboard. Covers JWT auth, x-kaiwu-auth signature, API endpoints, and metric extraction.

Write and maintain Kaiwu_2026 答辩材料 (defense/presentation docs). Covers technical writing for innovation points, updating the main summary doc, and git workflow for doc commits.

Use when working with the Kaiwu platform monitoring system — fetching training metrics, plotting charts, running anomaly checks, or updating monitor config. Covers the monitoring/ directory: fetch.py, check.py, plot.py, config.yaml.

Use when the user asks about training status, or when a cron alert fires. Fetches training monitoring data (including eval/matchup win rates and skill usage) from Tencent Arena API, runs anomaly detection, generates charts, keeps a per-run long-term record, and sends code-based analysis with prioritized recommendations via Feishu.

Use when initializing or rebuilding the wiki index. Scans the entire Kaiwu_2026 repository, extracts code structure via Python AST parsing, summarizes docs and experiment logs, and generates wiki/ index files.

从 GitHub 拉取最新代码，处理 stash/冲突，同步后通知飞书群聊并触发 wiki 更新。

Use when answering questions about the Kaiwu_2026 repository. Searches wiki/, memory, and git log to provide informed answers about code structure, experiments, competition stages, and RL design decisions.

Diagnose RL agent behavioral issues by analyzing reward structure. Use when the agent exhibits unwanted behavior (random skills, wandering, avoiding combat, etc.) to identify root causes in the reward function.

Use after a git commit to incrementally update wiki/ based on the diff. Triggered by post-commit hook automatically. Updates affected wiki files and Hermes memory.

Sync betting data from agents/*/round-N/bets.md into the HTML file's AI_ROUNDS array. Parses each model's bets.md, extracts bet entries, deploys to Caddy.

Automatically fetch match results from ESPN API and update the HTML display page. Covers result fetching, bet settlement, HTML patching, and Caddy deployment.

Sync world-cup repo from GitHub, update nginx-served HTML, and reload. Used when pulling latest changes or deploying updates.