Stop eyeballing your agent's permissions. Prove them.

Description and use

OpenShell’s prover automatically identifies concrete policy violations, revealing hidden leaks that manual review would miss.

Problem

Reading permission files by eye misses subtle policy combinations that allow unintended access.

Outcome

The prover quickly identifies exact requests or filesystem writes that violate a boundary, revealing leaks that manual review would miss.

Workflow

• Write a boundary policy for the parent agent.
• Generate a candidate policy for a child agent.
• Run openshell-prover with the boundary and candidate files.
• Interpret the prover’s output: if exit code 0, policy is within boundary; otherwise, examine the counterexample or unsupported message.
• Adjust the candidate policy accordingly and re‑run.

Evidence

The sandbox blocked the agent's write to a GitHub repo it was not allowed to touch.

Source

The prover says within_boundary and exits 0.

Source

result: exceeds_boundary counterexample: network binary=- ancestor_binary=- binary_identity_required=false host=api.github.com:443 protocol=rest method=POST path=/repos/acme/app/issues

Source

When the prover cannot answer, it says so and exits 3, instead of passing.

Source