Omp Operator — DevOps skill for Claude Code

omp-operator

A Kubernetes operator that runs [omp](https://github.com/can1357/oh-my-pi) (oh-my-pi) agents on webhooks and schedules, and as cloud agents that you hand a local omp conversation to.

You describe an agent as a custom resource: its model, instructions, skills, workspace, credentials, MCP servers and the actions it may take. You describe what wakes it up as a trigger: an Alertmanager, Grafana or generic JSON webhook, or a cron schedule. When a trigger fires, the operator starts one pod per session that drives `omp --mode rpc`, lets the agent investigate, and records a verdict, a report and a transcript. The agent can act only through actions you define, and the operator reports the result to ntfy, Slack, Discord or a webhook and shows it in a read-only dashboard.

A cloud agent works the other way round. In your local omp, `/cloud ` ships the repository (with your uncommitted changes), the conversation, an allowlist of secrets and your omp setup (settings, extensions, skills, `AGENTS.md`, memory) to an agent on the operator. It works asynchronously with the same omp, the same paths and every built-in tool, and returns its work as a branch `omp-cloud/`. See [cloud agents](docs/cloud.md).

v0.1 was the core platform and v0.2 adds cloud agents. GitHub mode (a robomp-style PR agent), attaching to a running cloud agent, and launching or chatting with agents from the dashboard are planned for later releases.

What you get

• Five CRDs in group omp.techdufus.com/v1alpha1: OmpAgent, OmpTrigger, OmpSession, OmpAction and OmpThread.
• Webhook triggers with adapters for Alertmanager, Grafana and arbitrary JSON, bearer, basic or HMAC authentication, CEL filters, grouping into episodes, dedupe, and per-group, per-agent and non-critical daily budgets.
• Schedules with 5-field cron expressions and IANA time zones.
• Sessions that run up to limits.maxConcurrentSessions at a time per agent (default 1), with priori