name: github-monitor description: Monitor one or more GitHub repositories and send low-noise alerts with configurable policy modes (major_only, balanced, verbose). Use when setting up recurring repo watch, release/security monitoring, PR merge tracking, and daily digest workflows via OpenClaw cron.

GitHub Monitor

Build one-cron multi-repo monitoring with deduplication, severity routing, and low-noise notifications.

Quick setup

Collect:

• repos: list of owner/repo
• policy_mode: major_only | balanced | verbose
• check_interval: recommended 2h
• timezone: e.g. America/New_York
• daily_digest_time: e.g. 21:00
• notify_target: user/channel id
• state_file: e.g. /root/.openclaw/workspace/memory/repo-watch/multi-repos.json

Use one isolated cron job to monitor all repos.

Data sources per repo

Check with GitHub public APIs each run:

1. Latest release
2. Recent commits (last 50, filtered by run window)
3. Recently merged PRs (filtered by run window)

Policy modes

1) major_only

• Notify only P0 events
• Skip P1/P2 notifications

P0 signals: - New release - PR keyword hit: breaking|migration|security|auth

更多技能請訪問小蔥技能站7w4.net。

2) balanced (default)

• P0/P1 immediate notify
• P2 daily digest

P1 keyword examples: - gateway|config schema|tooling|memory|cron|session|channel|provider|search|index|install|auth flow

3) verbose

• P0/P1 immediate
• P2 also immediate (or short-window batched)

Severity scoring recommendations

Do not rely on title keyword only; combine signals:

• Title/body keyword hit
• Label hit
• Changed paths weighting:
• higher weight: core/, gateway/, agent/, runtime/
• lower weight: docs/, tests/, examples/

Baseline + notification window (critical)

Track notification cursors in state to avoid backfilling old events:

• installed_at: skill first-run time
• last_checked_at: last successful scan time
• last_notified_at: last time user was notified

Rules:

1. On first install/run, initialize all three to now and do not send historical events.
2. Each run only evaluates events in (last_checked_at, now].
3. Notify only for events newer than last_notified_at.
4. After sending any alert/digest, set last_notified_at = now.
5. After successful scan, set last_checked_at = now.

Noise control

• Dedup fingerprint: repo + event_type + event_id/hash
• Silent when no new events
• Merge same-run alerts into one message grouped by repo
• Keep digest queue for P2 in state (pending_daily)

Reliability

• Retry transient API failures with exponential backoff
• If repeated failures, send one degradation alert
• Send one recovery alert when healthy again
• Persist state at end of successful run

Suggested OpenClaw cron behavior

• One job for all repos (easy scaling)
• Add/remove repos by editing list only
• Preserve one state file for dedup across restarts

Output mode (default: brief impact summary)

Default output must be concise and decision-oriented, not event-by-event long list.

Use this fixed structure:

1. 本視窗結論（是否有關鍵變化）
2. 對使用者可能影響（3-5 條，按優先順序）
3. 建議動作（最多 3 條）
4. 詳情入口：回覆“看詳情”再展開事件清單

Impact-first grouping (preferred):

• 配置/相容性風險（breaking、config schema、auth、gateway）
• 通知/渠道行為變化（telegram/slack/feishu/discord/channel）
• 會話與路由變化（session/dispatch/provider/tools）
• 穩定性/效能修復（timeout、race、retry）
• 文件/測試類（低優先）

Detailed mode (on-demand)

Only when the user explicitly asks “看詳情/展開/給清單”, output event-level details with links.

這個 GitHub 監控工具的設計思路不錯，支援多種告警策略，輸出格式也很清晰。但質量讓人擔心——壓縮包裡沒有程式碼檔案，只有配置檔案，根本無法驗證它能不能正常工作。文件寫得詳細但缺少實際使用示例，不知道效果如何。如果你用的是一個框架能自動理解 SKILL.md 來執行任務的平臺，那這個工具可能是好的設計；但如果需要自己寫程式碼實現，它就只是一個無法使用的設計草稿。