v2.3.6: See How Full Your Conversation Is, Compact It in One Click, and Tabs That Fix Themselves

Every chat session with an AI employee has a working memory, and until today you had no idea how full it was. v2.3.6 fixes that with a live gauge right in the chat bar, plus a one-click way to free up space. Around it: Google connections can no longer end up assigned to nobody, tabs recover on their own after a restart, and five small everyday annoyances are gone.

You can now see how full a conversation is

Long conversations fill up. When they get too full, the platform compresses the older parts to make room, a process called compaction. That has always happened behind the scenes; you just never knew when it was coming, and on some setups it triggered at the wrong moment or not at all.

v2.3.6 makes the whole thing visible. A discreet gauge now sits next to the model picker in the chat input bar, showing in real time how much of the session's capacity is used. It works for every kind of employee, whether it runs on Claude, Codex, or a custom model.

The gauge changes color as the session fills: amber at 50%, red at 85%. From amber onward, a small compact button appears next to it. One click and the session compresses itself, same effect as the /compact command, no typing needed. The gauge refreshes after every response from your employee, so it is always current.

Pick the right context size for your custom models

If you run employees on custom or self-hosted models, you can now tell the platform how big their working memory actually is. A new AI Context Window setting in Admin Settings offers four sizes, from 128k up to 1M tokens, with a sensible default.

This matters more than it sounds. Custom models used to report their usage as a permanent 0%, because the platform had no way of knowing their real capacity. Worse, in some configurations the compression math went wrong in a spectacular way: we caught a session live that was compacting itself in an endless loop, producing about a hundred compression summaries in fifteen minutes. With the window size set correctly, the percentages are true and compaction triggers exactly when it should.

The setting applies to each employee the next time its session is cleared.

Google connections can no longer be orphaned

Connecting a Google account used to have a silent failure mode: you finished the OAuth dance, the connection was created, and then nothing asked you which employee should use it. The connection sat there assigned to nobody, and the employee you meant to empower kept answering "I don't have access to Gmail".

Now the moment you come back from Google, a dialog opens and asks you to assign the new connection to an employee. It will not let you wander off without deciding, and if older unassigned connections are lying around from before, it catches those too, one by one. The only time it lets you close without assigning is when your instance has no employees yet, for obvious reasons.

Tabs heal themselves after a restart

After a backend restart or a session clear, open chat tabs used to get stuck on "Reconnecting..." forever. The platform was actually telling them "give me a minute, I am in cooldown", but the tabs treated that as a permanent goodbye. The only cure was closing and reopening every tab.

Tabs now understand the cooldown for what it is: temporary. They wait it out and reconnect on their own a couple of seconds after it ends. Leave the tab open, go get a coffee, it will be connected when you return.

Five irritants, gone

• The hire wizard position trap. You could reach the final step of hiring without having entered a position, at which point hiring failed and there was no Back button. The wizard now refuses to advance until the position is filled, and the final step has a Back button regardless.
• Empty diffs on custom-model employees. When an OpenCode-based employee edited a file, the red/green change view sometimes rendered completely empty. The changes were real; the viewer just did not understand that runtime's format. It does now.
• The stuck "Compacting..." banner. On some employees, the compacting notice stayed on screen long after compaction had finished. It now dismisses itself properly.
• Cleaner skill outputs. A protective component printed two log lines at the top of nearly every command an agent ran, which polluted outputs and occasionally confused the agents themselves. Those lines now go to a log file where they belong.
• Fresh-session accuracy. Brand-new sessions could briefly report nonsense usage numbers because of empty transcript files. The gauge now skips those.

Picking it up

Hard refresh (Ctrl+Shift+R) to load the new version. The context window setting takes effect per employee at the next session clear.

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