#sandbox = "read-only" and

1 messages · Page 1 of 1 (latest)

pine peak
#

How those two settings related to command approval? As I understand sandbox is related to filesystem and network access so I don't think it controls the approvals for commands, right? Also I can't find the trust_level config in the config reference page. There is only a approval_policy setting.

#

From the docs: approval_policy

When to ask for command approval:

- untrusted: only known-safe read-only commands auto-run; others prompt

- on-request: model decides when to ask (default)

- never: never prompt (risky)

#

So no option for always ask?

trim cosmos
#

Oh I gotcha, I got it backwards. approval_policy = "untrusted" would be what you're looking for. Then you would use

https://developers.openai.com/codex/rules

Rules. By default I think read-only commands like ls or rg are allowed, but you can write rules that require approval for all commands to harden the approval_policy beyond default

pine peak
#

Thanks. I didn't look into the rules section of the docs. In other agents I've used, rules refer to agent instructions, not to configs like it looks to be in codex.

#

Do you happen to know where can I find the list of commands that are deemed safe to run by codex, when using the "untrusted" setting?