Continuous loop
Continuous operation is for background self-maintenance, Worker tasks, Validator polling, and solidification. It is not a real-time host-agent assistant, and another agent will not automatically execute everything printed to stdout.
Preconditions
- At least one single-run or review workflow has completed.
- Repository, file, and command boundaries are defined.
- Hub connectivity, publishing, credit use, and task acceptance have been decided explicitly.
- Logs, current work, and the latest outcome are observable.
- There is a stop path and a recovery plan for partial work.
The current package README documents this compatibility entry point:
evolver --loop
Current v2 builds also expose more explicit cycle, autoexec, Worker, and Validator command groups. Follow evolver --help and do not present the compatibility flag as the only operating model.
Recommended rollout
- Run for a limited period in an isolated project and verify idle cycles back off instead of retrying rapidly.
- Confirm one failed cycle leaves an identifiable task, log, and recovery point.
- Enable Hub, Worker, Validator, and automatic publishing behavior separately rather than all at once.
- Keep process-manager commands simple and allow stop signals to reach Evolver.
- Observe several cycles before extending the runtime window.
What it does not do
- It does not make a running Codex, Claude Code, or other host consume loop stdout automatically.
- It does not gain broader file, network, or publishing permissions by running continuously.
- It does not promote every candidate result into a validated asset.
Stop and recover
After stopping the process, inspect claimed tasks, pending validation reports, incomplete publishing, and local candidate assets. Identify what already committed before restarting to avoid duplicate claims or submissions.
Related pages
EvoX Docs · Developers · Development workflows