メニュー
Start and manage the full heartbeat ecosystem for agent-studio. Registers all 7 heartbeat loops plus auto-reschedule task via CronCreate to keep the agent ecosystem healthy, indexed, informed, and connected.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/oimiragieo/agent-studio --skill heartbeatコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
Agent frontmatter enhancements — disallowedTools, mcpServers scoping, fork_eligible field
Context management — pre-compact persistence, threshold alignment, microcompact/circuit breaker detection
Hook enhancements — updatedInput for bash safety prefixes, suppressOutput for verbose blocks, denial-based routing feedback
SKILL.md を表示中
SOC 職業分類に基づく
| name | heartbeat |
| version | 1.0.0 |
| description | Start and manage the full heartbeat ecosystem for agent-studio. Registers all 7 heartbeat loops plus auto-reschedule task via CronCreate to keep the agent ecosystem healthy, indexed, informed, and connected. |
| category | infrastructure |
| trigger | when user wants to start heartbeat monitoring, recurring health checks, the full heartbeat ecosystem, scheduled maintenance, or autonomous reflection loops |
| tools | ["CronCreate","CronList","CronDelete","Read","Bash","TaskCreate","TaskUpdate","TaskList","MemoryRecord"] |
| dependencies | ["scheduled-tasks"] |
| tags | ["heartbeat","cron","monitoring","health","ecosystem","telegram","reflection","evolution"] |
| model | haiku |
| invoked_by | both |
| user_invocable | true |
| error_handling | graceful |
| verified | true |
| created_by | direct (retroactive attribution) |
| compliance_status | legacy-direct-creation |
| source | builtin |
| trust_score | 100 |
| provenance_sha | 7d69437c7bdc3180 |
The Heartbeat Ecosystem keeps agent-studio healthy, indexed, and informed by running 7 background loops via Claude Code's cron scheduler (CronCreate/CronList/CronDelete).
Key constraint — session-scoped: All loops die when the terminal closes. Use Loop 0 (auto-reschedule) to prevent silent 3-day expiry, and re-run Quick-Start commands after each session restart.
Run these /loop commands in your Claude Code session:
/loop 2h Heartbeat tick: Spawn heartbeat-orchestrator subagent via Task({ task_id: 'hb-reflection-' + Date.now(), subagent_type: 'heartbeat-orchestrator' }) to run reflection-check.cjs. Reply HEARTBEAT_OK after spawning.
/loop 4h Heartbeat tick: Spawn heartbeat-orchestrator subagent via Task({ task_id: 'hb-indexing-' + Date.now(), subagent_type: 'heartbeat-orchestrator' }) to check mtime of .claude/context/code-index/metadata.json and reindex if older than 4 hours or missing. Reply HEARTBEAT_OK after spawning.
/loop 30m Heartbeat tick: Spawn heartbeat-orchestrator subagent via Task({ task_id: 'hb-drain-' + Date.now(), subagent_type: 'heartbeat-orchestrator' }) to run context-drain.cjs and report stdout. Reply HEARTBEAT_OK after spawning.
/loop 24h Heartbeat tick: Spawn heartbeat-orchestrator subagent via Task({ task_id: 'hb-evolution-' + Date.now(), subagent_type: 'heartbeat-orchestrator' }) to run evolution-check.cjs. Reply HEARTBEAT_OK after spawning.
Prevents silent 3-day expiry. CRITICAL ORDER: Always CronCreate new tasks BEFORE CronDelete old ones.
CronCreate({
schedule: '0 0 */2 * *',
task: 'Self-maintenance: CronList() to inventory active tasks. Identify missing heartbeat loops from the expected set (reflection-2h, evolution-24h, briefing-8am, indexing-4h, drain-30m, telegram-5m, research-7am, reschedule-2d). Recreate any missing tasks using schedules in .claude/context/plans/heartbeat-ecosystem-design-2026-03-07.md. Report recreated task IDs.',
});
Extracts patterns from session transcripts before they are lost.
CronCreate({
schedule: '0 */2 * * *',
task: "Heartbeat tick: Spawn heartbeat-orchestrator subagent via Task({ task_id: 'hb-reflection-' + Date.now(), subagent_type: 'heartbeat-orchestrator' }) to execute this tick. The orchestrator must: (1) call TaskUpdate(in_progress), (2) run `node .claude/tools/cli/reflection-check.cjs`, (3) if stdout is HEARTBEAT_OK, call TaskUpdate(completed) and exit. Do NOT run the script directly in the router session. Reply HEARTBEAT_OK immediately after spawning.",
});
Applies accumulated learnings to improve agent definitions.
CronCreate({
schedule: '0 3 * * *',
task: "Heartbeat tick: Spawn heartbeat-orchestrator subagent via Task({ task_id: 'hb-evolution-' + Date.now(), subagent_type: 'heartbeat-orchestrator' }) to execute this tick. The orchestrator must: (1) call TaskUpdate(in_progress), (2) run `node .claude/tools/cli/evolution-check.cjs`, (3) if stdout is HEARTBEAT_OK, call TaskUpdate(completed) and exit. Do NOT run the script directly in the router session. Reply HEARTBEAT_OK immediately after spawning.",
});
Summarizes overnight state and suggests priority work.
CronCreate({
schedule: '0 8 * * 1-5',
task: 'Morning briefing: Spawn researcher via Task() to read issues.md, learnings.md, git log, and generate morning briefing report. Do NOT wait for sub-agent. Reply HEARTBEAT_OK and exit.',
});
Keeps hybrid search index fresh.
CronCreate({
schedule: '0 */4 * * *',
// Fixed 2026-04-21: was bm25-index.json which never exists — real artifact is metadata.json
task: "Heartbeat tick: Spawn heartbeat-orchestrator subagent via Task({ task_id: 'hb-indexing-' + Date.now(), subagent_type: 'heartbeat-orchestrator' }) to execute this tick. The orchestrator must: (1) call TaskUpdate(in_progress), (2) check mtime of .claude/context/code-index/metadata.json via Bash; if older than 4 hours or missing, run `pnpm code:index:reindex`, (3) call TaskUpdate(completed) and exit. Do NOT run the script directly in the router session. Reply HEARTBEAT_OK immediately after spawning.",
});
Detects pipeline idle state — warns user, does NOT auto-clear.
CronCreate({
schedule: '*/30 * * * *',
task: "Heartbeat tick: Spawn heartbeat-orchestrator subagent via Task({ task_id: 'hb-drain-' + Date.now(), subagent_type: 'heartbeat-orchestrator' }) to execute this tick. The orchestrator must: (1) call TaskUpdate(in_progress), (2) run `node .claude/tools/cli/context-drain.cjs`, (3) reply with the exact stdout output, then call TaskUpdate(completed) and exit. Do NOT run the script directly in the router session. Reply HEARTBEAT_OK immediately after spawning.",
});
Polls Telegram Bot API for user messages and routes to agents.
Configuration required:
TELEGRAM_BOT_TOKEN=your_token in .env.env is configuredCronCreate({
schedule: '*/5 * * * *',
task: "Heartbeat tick: Spawn heartbeat-orchestrator subagent via Task({ task_id: 'hb-telegram-' + Date.now(), subagent_type: 'heartbeat-orchestrator' }) to execute this tick. The orchestrator must: (1) call TaskUpdate(in_progress), (2) run `node .claude/tools/cli/telegram-poll.cjs`, (3) if stdout is HEARTBEAT_OK, call TaskUpdate(completed) and exit. Do NOT run the script directly in the router session. Reply HEARTBEAT_OK immediately after spawning.",
});
Discord note: Discord uses webhooks (push-based) rather than polling. Send messages via Discord webhook URL, but receiving requires a persistent process — recommend Telegram for bidirectional communication.
Surfaces relevant academic papers and web news. Delegates to the dedicated monitor skills.
CronCreate({
schedule: '0 7 * * *',
task: 'Research digest: Spawn researcher via Task() to invoke arxiv-monitor and exa-monitor skills. Do NOT wait for sub-agent. Reply HEARTBEAT_OK and exit.',
});
Dedicated skills: arxiv-monitor (every 6h standalone) | exa-monitor (every 4h standalone)
Config: ARXIV_KEYWORDS and EXA_MONITOR_TOPICS env vars (see .env.example)
To activate all 8 tasks programmatically:
// Invoke via Skill({ skill: 'heartbeat' }) — this skill registers all loops
// Loop 0: Auto-reschedule (must be first — keeps everything alive)
CronCreate({ schedule: '0 0 */2 * *', task: '...' }); // see Loop 0 above
// Loop 1: Reflection
CronCreate({ schedule: '0 */2 * * *', task: '...' }); // see Loop 1 above
// Loop 2: Evolution
CronCreate({ schedule: '0 3 * * *', task: '...' }); // see Loop 2 above
// Loop 3: Morning Briefing
CronCreate({ schedule: '0 8 * * 1-5', task: '...' }); // see Loop 3 above
// Loop 4: Indexing
CronCreate({ schedule: '0 */4 * * *', task: '...' }); // see Loop 4 above
// Loop 5: Drain Check
CronCreate({ schedule: '*/30 * * * *', task: '...' }); // see Loop 5 above
// Loop 6: Telegram Polling
CronCreate({ schedule: '*/5 * * * *', task: '...' }); // see Loop 6 above
// Loop 7: Research Digest
CronCreate({ schedule: '0 7 * * *', task: '...' }); // see Loop 7 above
// Verify all registered
CronList();
CronList(); // Returns all active tasks with IDs, schedules, next fire time
CronDelete({ id: 'abc12345' }); // Use ID from CronList()
// Get all IDs via CronList(), then delete each
const tasks = await CronList();
for (const task of tasks) {
CronDelete({ id: task.id });
}
minute hour day-of-month month day-of-week
*/5 * * * * = Every 5 minutes
0 */2 * * * = Every 2 hours
0 8 * * 1-5 = Weekdays at 8am
0 */4 * * * = Every 4 hours
*/30 * * * * = Every 30 minutes
0 0 */2 * * = Every 2 days at midnight
Supported: wildcards (*), single values (5), steps (*/15), ranges (1-5), comma lists (1,15,30).
NOT supported: L, W, ?, named aliases (MON, JAN).
| Risk | Mitigation |
|---|---|
| session-scoped: loops die on terminal close | Re-run Quick-Start commands on each session restart |
| 3-day auto-expiry: tasks silently self-delete | Loop 0 auto-reschedule runs every 2 days |
| 50-task cap: max concurrent scheduled tasks | 8 loops = 16% of cap; leaves 84% for ad-hoc tasks |
| No catch-up: missed fires are NOT replayed | Design loops to be idempotent (safe to skip) |
| Jitter: up to 15min delay on recurring tasks | Expected; design checks to tolerate timing variance |
| Telegram token not configured | Loop 6 gracefully skips if TELEGRAM_BOT_TOKEN unset |
memory-rotator.cjs for automatic file rotationagent-health.json for degradation signals.claude/context/memory/research-digest.mdTaskList() same as the router drain gatescheduled-tasks — low-level cron patterns (this skill builds on it)arxiv-monitor — dedicated ArXiv paper monitor (ARXIV_KEYWORDS, 6h interval)exa-monitor — dedicated Exa web search monitor (EXA_MONITOR_TOPICS, 4h interval)telegram-polling — Telegram Bot API polling (TELEGRAM_BOT_TOKEN required)task-management-protocol — task tracking patternsmemory-search — semantic memory queries used by loop prompts