Hello OpenAI team,
I use ChatGPT Work and Codex for macOS automation and software development. A typical task starts with a read-only audit, moves to a small approved change, and ends with tests and a record I can inspect later.
My highest priority is approval tied to the exact action. Before a deletion, installation, or external transfer, show the target, command or app action, destination, and scope of the grant. Ask for fresh approval if those details change. Let me revoke a grant before the next action, and show which restrictions each tool can enforce.
I’d also like a local activity log with a readable view and a JSONL export. Record attempted, approved, denied, and completed actions, who or what approved them, the affected resources, and the result. Redact sensitive values by default.
For file edits, add checkpoints with a diff preview and selective restore that preserves later manual edits or surfaces conflicts. After an interruption, verify whether an external action already completed before retrying it. Attach the commands, exit codes, and skipped checks to the task so I can verify the result.
A practical acceptance test would use a scratch project with one writable folder and one restricted folder. Check how each enabled tool handles the restricted path, change the target after a one-time approval, revoke access during the task, and restore an agent edit while preserving a later manual edit. The activity log should show what happened at each step.
Thank you for considering this.