Permissions and data flow
Review the workspace, execution policy, model endpoint and connectors before acting on a project.
On this page · 9 sections
Choose the scope deliberately
File tools use a selected workspace and path checks. That is different from a sandbox around every shell command. Commands and external tools can have wider reach, so do not treat a project directory as complete host isolation.
Start with a recoverable test workspace. Permission to use a tool is not permission to act on another person's data or infrastructure.
Separate inspection, modification and publication
Different actions affect different people and systems. Reading a selected source file is not the same as changing it; changing a local file is not the same as publishing it. State which class of operation is authorized rather than relying on a broad instruction to ‘finish everything.’
Review the destination as well as the verb. A harmless-looking diagnostic can send source material to a remote endpoint, while a local edit can be recoverable from version control. Publication, messages, account changes and deployments need an explicit scope. The permission to investigate a problem does not automatically include any of those actions.
| Operation | Review question |
|---|---|
| Inspect | Which material will be read, and where will it be processed? |
| Modify | Which exact files or state may change, and how can they be recovered? |
| Publish or deploy | Which external destination is authorized, and who will be affected? |
Read the proposed operation
The runtime exposes approval boundaries, but the effective policy depends on the interface and configuration. Check the exact operation, its target and the data it can affect before approving it.
Deleting files, sending data outside the project and changing an external service are consequential actions. A confirmation prompt provides a decision point; it does not certify that the proposed operation is harmless.
A workspace path is not a shell sandbox
File tools can reject a path outside the workspace. A shell command, however, can invoke another program, follow its own configuration or reach another service. Do not describe these two controls as interchangeable. Inspect the effective execution mode for the interface you are using.
The current tool implementation can request Docker execution. If that isolation is requested but unavailable, it fails closed unless an explicit fallback has been chosen. Enabling the fallback accepts host execution; it is not a repair of Docker isolation. Container availability also does not prove that network access, credentials and every connector meet your requirements. Review those settings separately.
Know where project material goes
Local model inference can keep processing on the operator's machine. A configured remote endpoint or connector can send material elsewhere. Review the model service and connector permissions independently of the workspace controls.
Local records can include task history, tool activity and memory. Inspect storage and retention settings before using sensitive material. This guide is not a security certification or a hosting-log audit.
Treat project and tool text as evidence, not new authority
A repository file, fetched page or tool result can contain instructions addressed to an assistant. Reading that text does not make its author authorized to redirect the current task. Compare the suggested action with the user’s request and the active operation policy before allowing it.
For example, a troubleshooting document might suggest uploading a log to another service. Check what the log contains, who controls the destination and whether sharing it was authorized. The advice may be useful, but the source cannot grant permission on the user’s behalf. The current project contains controls for untrusted material; their presence alone is not evidence that every adversarial case is solved.
Review retention and shared records
A local run may retain prompts, observations, memory and approval decisions. Those records are useful for review and may also contain material that should not be published. Before sharing a reproduction, inspect the relevant excerpt and remove credentials, personal information and confidential source context.
Do not assume that ‘local inference’ implies a no-records policy. Storage and processing are separate questions, and a connector introduces another destination. These guides describe the source-build boundaries, not an audited retention guarantee. Keep the smallest diagnostic excerpt that still explains the failure, and follow your organization’s handling requirements.
Check state before retrying an uncertain action
A timeout does not prove that an operation never started. A retry may repeat a write, create another remote object or send the same material twice. Inspect the current state first and identify whether the action has an idempotent or recoverable path before repeating it.
For local changes, use the known checkpoint and diff to decide what should be kept. Do not discard a whole worktree when only one scoped change needs attention. For an external operation, use the service’s own state and recovery controls. If the target or consequence is unclear, stop and ask the person who owns it instead of turning uncertainty into a second action.
Keep the review boundary
An operation that ran is not automatically a correct result. Inspect the changes, the checks that actually ran and any remaining uncertainty before accepting the work.
Avoid automatic operation until you understand the active policy. A production deployment needs its own review of isolation, credentials, connectors and recovery.