Review and apply Agent changes¶
An Agent can work freely inside a staged workspace without receiving automatic permission to modify the base project. After the Run, you decide which effects cross that boundary.
Run with a manual stage¶
The short form is:
For explicit paths and commit behavior:
pvisor run \
--overlayfs-base "$PWD" \
--overlayfs-stage /tmp/my-agent-stage \
--overlayfs-commit manual \
-- codex
Use a separate stage for every concurrent Run.
Review before accepting¶
review summarizes the Run Bundle, file effects, network evidence, and safety
warnings. inspect runs a read-only inspection command against the staged view.
Apply only selected files¶
Selections can be path-based or pattern-based:
A filtered apply consumes only the selected dependency-closed batch. Opaque directories and hard-link groups are not split when doing so would produce an invalid result.
Apply more than once¶
Applying a subset does not close the stage. Review what remains, then apply another subset:
Successful batches are recorded in apply-ledger.json. You can stop at any
point and discard the remaining changes:
Continue from an accepted point¶
Use a checkpoint before starting another line of work:
The new Run receives its own identity and stage while preserving lineage to the checkpoint.
What this boundary does not cover¶
Selective apply governs staged filesystem effects. It does not undo a message, payment, deployment, direct database write, or external API mutation. Those effects require admission before execution, provider-specific containment when available, and durable evidence afterward.
Next: control network access or capture the Run.