Give every coding agent the smallest useful scope.

Broad repository access is convenient, but it turns every focused task into a potential system-wide change. A safer run begins with the objective and a declared implementation plan, then grants the files and commands required for that plan.

Kaplira can keep read, write, and execution permissions separate. An agent may inspect surrounding context without receiving permission to rewrite it, and write access can wait until the intended change is explicit.

• Readable context selected for the current objective
• Writable files bounded by the accepted change contract
• Locked blocks that the run must preserve
• Commands tied to validation rather than unrestricted shell use

Real implementation work sometimes reveals that another file or subsystem must change. The correct response is not to forbid every expansion. It is to make expansion explicit, review the new blast radius, and issue a new contract before editing continues.

This keeps legitimate discoveries possible without allowing a small fix to become an unreviewed refactor.

After the run, the declared files are compared with the files that actually changed. Tests and review receipts stay attached to the same scope, so a reviewer can distinguish a bounded outcome from a successful build that reached farther than intended.