Execution boundaries
Evolver primarily produces protocol-bound prompts, candidate assets, and event records. Source edits, command execution, and external writes are normally performed by the agent application consuming its output, a validator, or an explicitly enabled automation path.
In this section
- Review mode: add human review before applying a proposal or approving an asset.
- Hard caps and rollback: use Git, file limits, and forbidden paths to bound impact.
- Validation commands: run only command forms accepted by the active verifier.
Three levels to inspect
| Level | Question |
|---|---|
| Proposal | Is the output plain text, a diff proposal, or an action directive interpreted by an agent application? |
| Execution | Which process will actually write files, run commands, or access the network? |
| Validation | Which test, build, state read, or human check proves the result? |
Do not describe “Evolver produced a proposal” as “the system completed the change.” Continuous loop mode is also not a live assistant in the current conversation; it primarily supports background maintenance, Worker, Validator, and solidification workflows.
EvoX Docs · Security · Execution boundaries