v2.3.1: A Quiet Patch With Four Real Fixes

v2.3.1 is a small patch release. No new features, four quiet fixes that you will feel even if you do not notice them.

Inbound mail delivery for a few edge cases

A small number of customers were having inbound emails to their agent silently bounce when the sender was on certain corporate mail relays. The cause was an SMTP handshake quirk: the way our mail receiver checked whether the sender was "talking too fast" was too strict, and a handful of legitimate corporate setups were getting filtered out by an anti-spam heuristic.

The check has been relaxed in this release. Legitimate senders that pipeline their commands aggressively now go through cleanly, and the regular slow-talker path still works exactly as before. If you had a customer or contact whose emails to your agent were not arriving, this is most likely why.

Multiple tabs and devices, without surprises

Opening the same employee in two tabs (or on your laptop and your phone in parallel) is a perfectly normal thing to do, and the chat already broadcasts every message to every connected client. What was not handled well was the moment one of those tabs did something destructive: clearing the session, switching the AI provider, that kind of thing. The other tabs kept showing the old view, kept accepting input, and the next message you sent could end up landing in the wrong place.

In v2.3.1, the moment any tab triggers a destructive action, every other tab and every other device gets a clear blocking message explaining what happened and a button back to the Dashboard. No more half-stale conversations.

Greetings on every kind of employee

Claude employees have always opened a fresh chat session with a personalised "salut, on continuait sur X la dernière fois?" greeting, drawn from recent memory. Employees running on Codex, OpenCode, or a custom LLM did not get the same treatment. They sat in silence until the user typed first.

Now every provider opens with the same warm context-aware greeting. The experience is consistent regardless of which model is under the hood.

French punctuation no longer confuses OpenCode

A small bug specific to OpenCode-backed employees: any sentence containing an exclamation mark in the middle ("Genere-moi une fonction Python !", "C'est top !") was being misread by the input parser as a shell command instead of a normal prompt. Mostly invisible to English-speaking users, mildly maddening for everyone writing in French.

The fix wraps user input so the parser ignores the misleading character. Legitimate shell escapes at the start of input still work; everything else just passes through as plain text the way it always should have.

Why this kind of release matters

Patch releases like this are usually the ones where the most measurable improvement to daily user experience happens. None of these were headline features. All four were the kind of thing that quietly degraded the product for a subset of users and is now off the list.

Hard refresh to pick up the new version tag.

Want to test the most advanced AI employees? Try it here: https://Geta.Team