agent-ops-plus

Built on affaan-m/ECC by @affaan-m (254,867 stars, MIT). All credit for the original idea to them. This fork improves and repackages it; upstream license preserved in UPSTREAM_LICENSE.

A Claude Code skill for running long-lived agents with safe limits, clear logs, careful updates, and fast recovery.

Built for teams that run agents in the cloud or as always-on services. It fits into your current Claude Code flow and works with systemd, PM2, containers, and release checks.

💡 Why

Long-lived agents can get stuck, retry bad work, leak data in logs, or spend too much.

Updates can also fail while tasks are still running.

This skill gives Claude Code a clear set of rules for task limits, safe retries, shutdowns, restarts, updates, rollback, and incident response.

It is a single skill file with no added packages.

🚀 Install

Run one command:

mkdir -p ~/.claude/skills/enterprise-agent-ops && curl -fsSL https://raw.githubusercontent.com/lukehmw/agent-ops-plus/main/skill/SKILL.md -o ~/.claude/skills/enterprise-agent-ops/SKILL.md

Restart Claude Code after install.

🛠️ Usage

Ask Claude Code to plan or review an always-on agent:

Use enterprise-agent-ops to review my support agent service.

It runs with systemd. Each task may call paid tools and write to a state store.
Create safe task limits, retry rules, a kill switch, a staged update plan,
restart steps, rollback steps, and an incident checklist.

Expected output:

Task controls
- Give each task a clear ID.
- Set time, retry, work, and cost limits.
- Do not retry bad input or other errors that cannot improve.
- Use task IDs to stop the same risky action from running twice.

Service controls
- Stop new work before shutdown.
- Check saved task state after restart.
- Block risky work if the audit log or state store fails.
- Use a kill switch when harm, leaks, loops, or open-ended spend may grow.

Update plan
- Test the fixed release.
- Start with a small share of work.
- Watch failures, cost, and task time.
- Roll back to the last good release when a set limit is crossed.

🔄 What we changed vs upstream

• Rewrites the Japanese description and jargon-heavy sections in clearer, simpler English.
• Adds explicit ECC attribution while preserving the original origin: ECC metadata.
• Expands baseline safeguards with log redaction, backoff, non-retryable errors, cost/work limits, and tested restart/rollback procedures.
• Adds detailed task lifecycle, shutdown/recovery, staged rollout, state-store failure, and idempotency rules.
• Greatly broadens incident guidance with concrete edge cases, escalation triggers, audit requirements, and a systemd example.

📄 License

MIT licensed. The upstream MIT license and credit are preserved in UPSTREAM_LICENSE.