Roadmap and limits¶
| Phase | What | State |
|---|---|---|
| 0. Record | Alert schema, append-only event log, OpenObserve pull adapter, webhooks, source staleness, heartbeat | ✅ built |
| 1. Observe | Correlation into incidents, flap and maintenance flags, change calendar, triage in shadow mode, daily digest with feedback | ✅ built |
| 2. Investigate | Read-only investigation by a strong model, analysis by notification, postmortem drafts | ✅ built |
| 3. Approve and act | Action catalog, signed Go / No-go, executor, verification, rollback, runbooks | ✅ built, not yet proven in use |
| 4. UI and MCP | Console (board, incident, log, ledger, actions, autonomy, postmortems), read-only MCP server, /brief for a daily brief |
✅ built |
| 5. Earned autonomy | Promotion proposals, auto, Undo, automatic demotion |
✅ built, not yet proven in use |
| 6. Release | Docs, threat model, a Prometheus and Docker quick start | ✅ written; the "stranger in under an hour" test is still to be run |
"Built" means written and covered by tests. The exit criteria of the
design are about use (ten approved fixes
without a manual cleanup, a month of auto with no bad action): those take a server and time.
Limits as built¶
- Actions are for one Docker host, the Coolify instance and the deployment repository you configure. There is none on another host, on the network or on Kubernetes.
config_changechanges a value that already exists in a YAML file. It cannot add a key, edit another format, or open a pull request anywhere but GitHub (elsewhere the branch is pushed and the notification says so).- A step is verified by the container's state, its health check, Coolify's own status and an optional URL. The alert's own query is not re-run by the executor; the incident resolving is what confirms the fix.
coolify_restartis written from Coolify's published API description and has not been run against an instance. The trace export has been tested against a collector double, not against a real OTLP endpoint.- The quick start was run end to end without a model key: the investigation step of it was not.
- Replay judges a triage verdict by itself; for an investigation it shows the old and new root cause and leaves the judgment to you.
- There is no retention period for tool results in the event log.
- Of the notification channels, ntfy, Gotify, e-mail and the webhook were run against real servers; the others were tested against each service's documented request, not the live service. A tap cannot be answered from inside a chat: a button opens the executor's page.
- Of the sources, the OpenObserve adapter was written against a real instance. Prometheus,
Alertmanager's API, Grafana, Loki, VictoriaLogs, Gatus, Healthchecks, Netdata, Zabbix and
Beszel are written from each service's documentation and tested against the payloads and
answers it documents, not the live service.
paeon source <name>is the first thing to run. - A Prometheus without Alertmanager has no way in yet: its alerts are not polled. Netdata Cloud's notifications are not read. For logs in Loki and VictoriaLogs the model gives words to look for, not a query.
- The console cannot approve. Tier 2 is approved by merging on the forge; nothing riskier exists.