v2.7.19: Codex Gets a Model Picker, Guardrails Get Friendlier, and Voice Prompts Get Room to Breathe

This one is about employees answering more reliably, and Codex finally catching up with everyone else.

Codex employees get a model picker

Until now, the model and reasoning picker in the chat only showed up for Claude and custom-LLM employees. If you hired a Codex employee, you were stuck with whatever it started on.

That is fixed. Codex employees get the same picker, with the full model list and reasoning levels from none all the way to xhigh. Better still, switching happens mid-conversation. No restart, no lost thread. Change the model in the middle of a task and the next turn simply runs on the new one, with your whole conversation intact.

Model and reasoning level move independently, so you can keep a fast model on high effort, or a big model on low, whatever suits the work.

One small tip: if you want to know which model an employee is running, look at the picker rather than asking the employee. It genuinely cannot see that information about itself, so it will make a polite guess and it may well be wrong.

Codex employees now read their profile properly

Related, and more important than it sounds. Codex employees were not reliably picking up their CLAUDE.md, which is where their personality, instructions and working context live. An employee without its profile is technically functional and completely generic.

They read it properly now. If your Codex employee has felt oddly impersonal lately, this is why, and it should feel like itself again after a respawn.

Guardrails got noticeably friendlier

Two changes here, both of which you will feel immediately.

First, harmless messages are no longer misread. A cheerful ca va? could occasionally come back refused, which is a frustrating thing to explain to anyone. The safety classifier now receives conversation context in a way that keeps it firmly in the role of reviewer rather than participant, and casual chat sails through the way it always should have.

Second, and this is the bigger shift: if the safety check itself cannot run, your message now goes through. Previously a check that timed out or failed technically was treated as a refusal, so you could be blocked for something nobody had actually judged. Only a genuine block verdict blocks now. Technical failures are still recorded and shown in the admin panel as an amber "Check failed, let through" note, so nothing is hidden, but they no longer cost you a message.

And getting blocked no longer freezes your chat. You used to need a refresh, sometimes a full session clear. Now the conversation carries on normally with your next message.

Custom-LLM employees are talking again

A missing folder permission was stopping custom-LLM employees from reaching their model on some instances, which showed up as silence rather than an error. The folder is mounted correctly now and those employees answer again.

Voice prompts have room to breathe

If your employee handles calls, its custom call instructions were capped at 4,000 characters when written through the skill, even though the dashboard happily accepted far more. Anyone with a serious call script, and clinics in particular run very long ones, hit that wall.

The ceiling is now 100,000 characters, matching what the interface already allowed. There is still a ceiling on purpose: this text goes into every single call, so a very long prompt costs a little setup time on each one. But 100,000 is real room.

Also in this release

• The context gauge no longer occasionally shows an impossible number and pins itself at 100%
• Discord bots can actually be disconnected now, including ones stuck in an error state
• Connectors now lists the browser extension and the desktop app, so when your employee tells you to find them there, they are there
• Admins can disconnect a Google or Microsoft mail account from any employee, which is the fix for that forgotten mailbox quietly waking an agent up all day
• Switching an employee's engine now clears the old session properly, so it stays switched

Update and enjoy. If you run Codex employees, go and try switching models mid-conversation, it is a nicer feeling than it sounds.