Cancel, then ask the agent "What is your | OpenAI | Page 1

cerulean notch Apr 27, 2026, 2:00 PM

#

Model interrupted to submit steer instructions.

› What is your current sandbox, network, and writable_roots settings?

• Sandbox: workspace-write

Network: restricted

Writable roots:

/home/jon/.codex/memories
/home/jon/climex
/tmp

Approval policy: never, so I cannot request elevated permissions.

trail palm Apr 27, 2026, 2:04 PM

#

OK, there are two likely things here. You have network access restricted, and the agent said something about trading market mechanics. They're probably doing API calls or something that requires network. If you wanna enable network you do this

[sandbox_workspace_write]
network_access = true

Secondly, I would set approval_policy = "on-request", that way if they hit real sandbox issues they have a chance to escalate, otherwise they will either refuse to continue due to restrictions, or they will work tirelessly to find workarounds.

cerulean notch Apr 27, 2026, 2:05 PM

#

I'll try that, thank you!

#

Any suggestions to how I can diagnose and fix this myself? I like to learn how to fish 🙂

trail palm Apr 27, 2026, 2:15 PM

#

Hmmm, are you using the guardian feature or something? Usually elicitations include the command or tool being ran, but it's just describing the tool...

In any case here's the official docs: https://developers.openai.com/codex/config-advanced#approval-policies-and-sandbox-modes

And what I like to do is clone https://github.com/openai/codex.git, checkout the tag matching the locally installed version, then use gpt-5.4-mini and ask it questions about the codebase.

cerulean notch Apr 27, 2026, 2:15 PM

#

maybe I should just use --yolo ?

#

gotcha I'll check out those docs, ty

#

this is my user-level config

#

trail palm Apr 27, 2026, 2:19 PM

#

Yeah, that's a really solid and sensible setup.

I would definitely also ask the agent why they keep requesting elicitation. Like what commands or tools are they trying to use that require approval. I don't recommend --yolo, the sandbox can be tuned to be safe while also giving them more agency and it's generally better to follow that route than give em full access

cerulean notch Apr 27, 2026, 2:19 PM

#

got it. thank you!

#

#

it's taking me a bit to remember that I can literally just ask the AI about the issue i'm seeing with the AI itself

trail palm Apr 27, 2026, 2:28 PM

#

Haha ok, that happens a lot 😛

I think when the MCP tool required elicitation, it fell into the habit of asking constantly even when there was no sandbox concern anymore. This happens to me all the time!

cerulean notch Apr 27, 2026, 2:32 PM

#

Dang. still not fixed. It's freaking whack-a-mole lol

trail palm Apr 27, 2026, 2:35 PM

#

cerulean notch Dang. still not fixed. It's freaking whack-a-mole lol

At this point I would ask them specifically what command and why, not just agree to "ask less often". Say you need to know precisely why so you can tune the sandbox to allow these things.

Then I would clone the openai/codex repo and have gpt-5.4-mini sweep through based on what that agent says to find the right config combo, mode, instructions to make the agent work more autonomously

#Cancel, then ask the agent "What is your