v2.7.10: Your Employees Read Images in a Second Now, and the Context Window You Pick Is the One You Get

Yesterday we told you the built-in vision fallback got sharper. Today it got fast. v2.7.10 swaps the local vision model for Google Gemini, called server-side, and the result is the kind of speed jump you feel immediately: reading an image drops from 13 to 33 seconds down to about one or two. It also fixes a quieter but costlier bug, the custom context window you pick in settings is now actually the one your employees use.

Reading images went from seconds to instant

A quick reminder of what this feature is. When an employee runs on a model that cannot see images itself, it leans on a vision fallback to read the picture for it: the text in a screenshot, the icon in a UI, the contents of a photo you sent. Until now that fallback ran on a local model, and while it was reliable, it was slow. A single image could take anywhere from 13 to 33 seconds, which is a long time to wait for an employee to simply look at something.

This release moves that work to Gemini, running server-side with the platform's key, and the numbers tell the story:

That is not a tweak, it is a different experience. An employee that used to visibly stall while it puzzled over a screenshot now glances at it and keeps moving. For anything where an image is part of the task, reading a receipt, checking a chart, describing a photo, the whole interaction feels responsive instead of stuck.

Two things worth knowing about how it works. Your employees did not change at all: they still call the same internal endpoint with their own employee token, and the vision skill itself is untouched, so nothing you built on top needs updating. And the key is never exposed. The Gemini call happens on the server with the provided key, and the employee only ever sees its own token.

There is a nice side effect too. Moving this off a bundled local model removed roughly three and a half gigabytes of model files from the install, along with the runtime that supported them. Lighter footprint, faster reads, same simple interface.

The context window you pick is the one you get

The other fix in this release is small to describe and large in impact. If you run custom models and set your context window in Admin, Settings, that number is supposed to be the amount your employees can hold in mind before the conversation gets compacted. It was not being honored. Pick 512K and the system would still compact at 256K, exactly half, no matter what you chose.

The cause was a plumbing mismatch: the setting was being saved in one place and read from another, so the value your employees actually used quietly fell back to a fixed default and ignored your choice entirely. If you had raised your context window and wondered why longer conversations kept getting trimmed early, this is why, and it is fixed.

• Set it, and it sticks. The context window you choose in settings is now the one your employees actually run with.
• Picked up on the next session. Employees already working keep their current size until they restart, so nothing gets cut off mid-task. Clear a session and it comes back on the size you set.

If you run custom models, it is worth taking a moment to confirm your context window is where you want it, then clearing a session to pick it up. You may find your employees can suddenly hold a lot more of a long conversation than they were letting on.

The pattern here

Two changes, one theme. A vision fallback that reads in a second instead of half a minute, and a context setting that finally means what it says. Neither is a flashy new capability. Both are the difference between a tool that quietly underperforms and one that does exactly what you told it to, as fast as it should. We would rather ship that than another headline.

v2.7.10 is live now. Send an employee an image and watch how quickly it reads, and if you run custom models, go check that context window.