pVisor
pVisor runs an existing Agent command inside a controlled execution environment. It gives each Run its own workspace boundary, records the controls that were actually installed, and lets you review filesystem changes before they reach the project.
In Persisting, pVisor runs one Agent and reviews its changes. You can use it without pChronicle.
By the end of the first walkthrough: the Agent has stopped; its changes remain
in a Run-owned staging directory; you deliberately write them into the project
or discard them. Your first pvisor review shows the record of controls that
were actually installed.
pVisor does not replace the Agent's reasoning loop. You can keep using Agent CLIs, scripts, and frameworks you already have.
Run, review, decide
From a project directory:
pvisor run --stage ./runs/task-001 -- codex
pvisor review last
pvisor apply last --path src
With --stage ./runs/task-001, the Agent writes to a staged view of the project. After the Run,
you can apply all changes, accept selected paths in several batches, or discard
the stage:
pvisor apply last --all
# or
pvisor drop last
The exact filesystem and network boundary depends on the platform and chosen executor. pVisor records the effective controls so that a Run is not described as more isolated than it was.
Choose your next step
Start with Run your first Agent if this is your first session. It ends with a reviewed stage and gives you the vocabulary used by the rest of the documentation.
When you already know what you need, follow the matching path:
- Keep or discard changes: Review and apply
- Compare host, container, and VM: Execution layouts
- Constrain network access: Network policy
- Publish trajectory events: Capture trajectories
- Look up exact flags: CLI reference
pVisor's local run-review-apply loop works on its own. pChronicle is optional: use it when you want to retain and query trajectory Datasets after a Run.
A useful reading order
- Run your first Agent to see the complete success loop.
- Review and apply when you need finer control over changes.
- Execution layouts when provider boundaries affect your decision.
- Capabilities and evidence when you need to interpret a Run Bundle.
- CLI reference only when you need an exact flag or output field.