Grok Bot Release Notes
All Episodes
Grok Bot 0.78.1 Baseline and 0.79.3 Field Reports

Grok Bot 0.78.1 Baseline and 0.79.3 Field Reports

0:00|0:00

Explore the baseline features of Grok Bot version 0.78.1 alongside community findings from unannounced 0.79.2 and 0.79.3 desktop builds. Learn about new shopping templates, formatting updates, Windows local-exec drive limits, and essential troubleshooting workarounds.

Show Notes


Chapter 1

Official Baseline and What Moved Quietly

James Turner

The public Grok Bot changelog still says version 0.78.1, released October 27th. But desktop builds in the wild have quietly moved on to 0.79.2 and now 0.79.3, with no published notes. So this one is part changelog, part field report. Brought to you by Jellypod AI.

James Turner

Last time we covered custom cloud agent modes and early reports of 0.79.2 on Windows. Today I'm advancing that story. First, what's officially documented in 0.78.1, because that's the baseline you can actually rely on. Bots can start cloud agents in a custom mode you name, and keep replying to them in it. Shopping with a Bot now shows product cards with photos, rating, price, and store. You can connect Shop Pay in the Vault's Payments tab, and Buy now starts an order that you approve. After you order, the Bot tracks it and tells you when it ships or arrives.

James Turner

The templates are where I'd start. Try it, don't pay for it starts a free trial for you and cancels it, or reminds you, before it bills. Find unused subscriptions cancels plans you pick from the ones you still pay for, and Watch subscription prices checks renewal prices. And the Bot confirms with you before changing anything. So a practical move: test a trial SaaS tier with the first template, with your payment method set in the Vault, and let the reminder do the guarding.

James Turner

Some quieter 0.78.1 fixes matter too. Copy on a reply now keeps headings, lists, and links when you paste into Slack, Gmail, Docs, or Notion. Small, but you'll feel it daily. The Bot's browser types one character at a time and moves the mouse on every website it fills in. And when the Bot's computer has been refusing saves, Reset and Recover keep the current disk instead of restoring an older backup. That last one is a real safety improvement.

Chapter 2

Field Reports: Outages, Drives, and Limits

James Turner

Now the community side, and I'm pulling this from the Cursor forum, where support engineer Colin has been answering. On November 8th, from about 07:47 to 09:48 UTC plus 8, replies for a group of users failed before they reached the model. Even a plain hello got Bot failed to respond, on 0.79.3, on Windows and macOS. Colin said it was server side, not your Bot's computer. One user reported it fixed itself. And here's the lesson. Don't reflexively hit Reset or Recover. Several people tried resetting the computer, and it didn't help, because there was nothing to fix on their end.

James Turner

Second, Windows drives. Reports on 0.79.3 say Read, CopyFromBox, and CopyToBox are limited to C:\Users, while Shell is not. Reading a project on E: fails with, quote, Path is outside the allowed local-exec root and was refused. One user said this worked before and stopped around November 4th. A junction won't help, since the check uses the real path. So the practical advice: keep active repos inside your user profile folder if you use local execution on Windows.

James Turner

Third, one user with Execution on Local Computer set to Always allow said the Bot refused an authorized small edit and said coding has to go through a Cursor Cloud Agent. Their support bot couldn't say whether direct local edits are supported. So that's a report, not confirmed policy. Plan for that handoff, and watch for an official answer.

James Turner

Fourth, Auto-review. The error An error occured while classifying this action, please review manually, came from a size limit on the safety check. Colin said a save command carrying roughly 65,000 characters was rejected before any approval card appeared. His workaround: ask the Bot to write the analysis in several parts, each under 30,000 characters, then apply it. Nothing ran, so earlier results were untouched.

James Turner

Last, Linux. Colin confirmed local execution is supported on 0.79.2. But one user behind a proxy found the local execution daemon dials api2.cursor.sh directly and ignores HTTPS_PROXY, so with no DNS it failed with ENOTFOUND. Their workaround was a hosts entry plus a socat tunnel, and they've asked for proxy support. That's the week. Chunk your writes, mind your drives, and don't reset on an outage.