Example scenarios¶
These examples begin after a reviewed deployment is already running. Each one opens with the outcome and the shortest safe path, then points to the exact policy and runtime files needed for review.
| Scenario | Use |
|---|---|
| Praxis token quotas | Inspect providers, preview, apply, and verify shared rolling-day model-token allowances |
| Streamlit before and after OpenShell | Show credential-free managed inference and a confined filesystem in one short demo |
| OpenCode PR-candidate queue | Turn public GitHub issues or operator prompts into reviewable patch bundles without giving the agent GitHub write access |
| Review as a service | Delegate pull-request review in a restricted sandbox that can read public repositories but cannot push or comment |
| API key theft | Replay the reported key-stealing agent: the key is findable, and every path to spend it at another provider dies at connect() |
| Document exfiltration | A malicious downloaded skill tries five legitimate-looking ways to ship a raw document to an attacker bucket; the boundary stops each |
| Model escalation | The same agent image with and without the boundary: one flag spends the key on a premium model in the open; behind the gateway the catalog refuses it before credentials |
Each example names its prerequisites and the evidence it does not claim. Treat a successful demo as a workflow result, not a substitute for the deployment's separate qualification tests.