IceCube Heartbeat

🧊 IceCube Heartbeat — Proactive agent maintenance during idle periods. Self-check, memory governance, improvement detection, and evolution triggers. Keeps y...

Security Scan

The skill's instructions ask the agent to access email/calendar, ClawHub, git repos, and install other skills or push code, but the package declares no credentials or install steps — the operational demands and privileges don't match what the skill requests.

Purpose & Capability

The declared purpose (agent maintenance during idle periods) is plausible, but the instructions require access to email, calendar, ClawHub (skill store), git remotes, cron/system status and multiple local memory files. The skill declares no required env vars, credentials, or config paths to enable those actions, which is inconsistent — any agent performing these tasks would need credentials or platform APIs explicitly documented.

Instruction Scope

The SKILL.md instructs the agent to read and modify many files (unclosed_work.yaml, mistake_log.md, success_patterns.md, MEMORY.md, heartbeat-state.json), check cron jobs and agent sessions, check ClawHub, install new skills, run git status/commit/push, and proactively notify users (email/calendar). Those are broad side-effects and external interactions beyond a simple heartbeat reply and are not scoped or limited in the document (no constraints on which repos/accounts/endpoints).

Install Mechanism

There is no install spec and no code files, which reduces the static attack surface. However, because the instructions explicitly tell an agent to install skills from ClawHub and to commit/push code, the runtime behavior could still lead to code changes or new skill installs even though nothing is installed by the package itself.

Credentials

The skill lists no required environment variables or credentials yet expects access to email, calendar, ClawHub, and git remotes. That mismatch is concerning: these capabilities normally require service tokens, OAuth credentials, or repo keys. The SKILL.md does not document what credentials will be used or how the agent should obtain permission, so it's unclear what the agent will actually do at runtime or which accounts it would affect.

Persistence & Privilege

always:false (good), but the skill explicitly instructs the agent to install new skills, modify MEMORY.md, write heartbeat-state.json, and commit/push changes. Combined with the platform-default ability for the agent to invoke skills autonomously, these instructions give the skill potential to alter system state and other skills without declaring that privilege — this broad modification capability is a material risk unless you run in a tightly sandboxed environment.

What to consider before installing

This SKILL.md tells the agent to read/write many local files, check email/calendar/cron, install other skills, and commit/push code — but it declares no credentials or install steps. Before installing: (1) don't enable this on a production agent with access to your email, git, or external accounts; test in an isolated agent/sandbox first; (2) ask the author to document exactly which accounts/endpoints will be used and what credentials are required; (3) require explicit user approval before installing skills or pushing code; (4) back up repositories and important files and enforce least privilege (separate service accounts or read-only access where possible). If you cannot confirm or constrain these behaviors, treat the skill as risky and proceed cautiously.

The pulse that keeps your agent alive.

When your agent receives periodic "heartbeat" polls, don't just reply "HEARTBEAT_OK". Use that time to maintain, improve, and evolve.

Why IceCube Heartbeat?

Problem: Most agents sit idle between user requests. No learning. No improvement. No self-maintenance.

Solution: Heartbeat turns idle time into productive evolution cycles.

Result: An agent that gets better every day, without you doing anything.

What Heartbeat Does

1. System Health Check

• Scan unclosed_work.yaml for pending tasks
• Check cron job status (recent failures?)
• Verify agent sessions are healthy
• Alert if anything needs attention

2. Memory Governance

• Scan mistake_log.md for new errors
• Scan success_patterns.md for new wins
• Evaluate if memory consolidation needed
• Promote important daily entries to MEMORY.md

3. Self-Improvement Detection

• Repeated fallback patterns → flag for optimization
• Repeated rollback patterns → flag for root cause analysis
• Task stagnation (>24h no progress) → alert
• User repeated corrections → update rules

4. Evolution Triggers

• Weekly: Check ClawHub for new relevant skills
• Weekly: Assess pending API configs can be activated
• Daily: Generate brief status report

5. Proactive Outreach (Optional)

• Important email arrived → notify
• Calendar event soon (<2h) → remind
• Weather relevant → mention if user might go out

Setup

1. Create HEARTBEAT.md

# HEARTBEAT.md

## 自检任务（每次 heartbeat 执行）

### 1. 系统健康检查
- [ ] 检查 unclosed_work.yaml 中的待办事项
- [ ] 检查 cron 任务状态
- [ ] 检查 agent sessions 是否有异常

### 2. 记忆治理
- [ ] 扫描 mistake_log 是否有新错误
- [ ] 扫描 success_patterns 是否有新模式
- [ ] 评估是否需要 memory consolidation

### 3. 自我改进触发条件检测
- [ ] 是否有 repeated fallback
- [ ] 是否有 repeated rollback
- [ ] 是否有 task stagnation
- [ ] 是否有用户重复纠正

### 4. 技能扩展检查（每周一次）
- [ ] ClawHub 是否有新技能适合当前需求
- [ ] pending API 配置技能是否可以激活

2. Configure Heartbeat Interval

In OpenClaw config:

{
  "agents": {
    "heartbeat": {
      "intervalMs": 1800000,  // 30 minutes
      "prompt": "Read HEARTBEAT.md. Follow it strictly. Reply HEARTBEAT_OK if nothing needs attention."
    }
  }
}

3. Create Supporting Files

mistake_log.md:

# Mistake Log

## Format
- Date: YYYY-MM-DD HH:MM
- Mistake: Description
- Impact: What went wrong
- Fix: How to prevent recurrence

success_patterns.md:

# Success Patterns

## Format
- Date: YYYY-MM-DD HH:MM
- Pattern: Description
- Result: What worked well
- Applicability: When to use this pattern

Heartbeat Response Rules

When to Reply HEARTBEAT_OK

• Late night (23:00-08:00) unless urgent
• Nothing new since last check
• Checked <30 minutes ago
• User clearly busy

When to Reach Out

• Important email arrived
• Calendar event in <2h
• Something interesting found

• 8h since last contact

When to Stay Silent

• Routine checks with no findings
• User hasn't responded to previous proactive messages
• Late night quiet hours

Example Heartbeat Cycle

08:00 Morning heartbeat:

1. Check unclosed_work.yaml → 3 pending items
2. Check mistake_log → 1 new entry (fallback loop)
3. Check calendar → meeting at 10:00
4. Action: Remind user about meeting, flag fallback issue for investigation
5. Reply: "Meeting at 10:00. Found a fallback pattern to fix later."

14:00 Afternoon heartbeat:

1. Check unclosed_work → same 3 items, no progress on TASK-A
2. Flag: task stagnation >24h
3. Check ClawHub → new skill matching current project
4. Reply: "TASK-A stalled >24h. New skill available for [project]."

23:00 Evening heartbeat:

1. All checks pass
2. Late night → stay quiet
3. Reply: HEARTBEAT_OK

Anti-Patterns

❌ Don't:

• Reply HEARTBEAT_OK when there's unfinished work
• Ignore stale tasks
• Skip memory governance
• Never promote learnings to long-term memory

✅ Do:

• Check before you OK
• Act on findings immediately
• Keep heartbeat-state.json for tracking
• Batch checks to reduce API calls

Proactive Work Without Asking

These are safe to do during heartbeat:

• Read and organize memory files
• Check on projects (git status)
• Update documentation
• Commit and push your own changes
• Review and update MEMORY.md
• Install new skills from ClawHub
• Run self-improvement routines

Integration with Other IceCube Skills

icecube-memory: Heartbeat triggers memory distillation weekly icecube-ops: Heartbeat checks system health status icecube-evolution: Heartbeat drives continuous improvement loops

Tracking State

heartbeat-state.json:

{
  "lastChecks": {
    "email": 1703275200,
    "calendar": 1703260800,
    "weather": null,
    "clawhub": 1703260800
  },
  "lastPromotion": "2026-03-17",
  "heartbeatCount": 42
}

License

MIT — Use freely.

---

Idle time is evolution time.