Staff say All screens in use can mean one Bot's screen never started

Grok BotWhat's new#

Update, Oct 13, 10:44 AM ET: Five staff-confirmed known issues landed on the forum after the morning issue. 1Password reconnect can stick on "previous credential delivery must be reconciled" (app-memory lock plus rate limits). Grok Bot GitHub page reads can show 0 stars and an empty README (use the API or the Bot browser). A Windows Cursor IDE agent report says a quoting bug let rmdir /s /q hit a drive root; staff say Auto-review should have asked first. Synced personal skills are not available inside Cursor automation runs. New permission-based user API keys fail CLI/SDK agent runs server-side. Workarounds are in the linked entries. (confirmed)

Update, Oct 13: Two more staff-confirmed known issues. Saving a Grok Bot routine checks your Cursor plan, so on Start or with only a SuperGrok link the save is refused (routines need a Cursor plan with Automations). A platform-wide limit on automation starts was saturated for several hours today and has cleared; PR, Slack and webhook runs it blocked were dropped, not retried (how to find and re-fire them). (confirmed)

Update, Oct 13: Anthropic's Frontier Red Team says small AI models may be months away from being able to self-replicate on cheap smart devices like off-brand Android TV boxes, and it ran a 26B model across 15 of them (what it found and how to lock down your devices). (confirmed)

Update, Oct 13: Gemini 4 Argon is in Cursor, and Cursor says it matches Opus 5.5 on CursorBench at half the token price; you can't choose it in Grok Bot (pricing, effort levels and the Start-plan catch). (confirmed)

Update, Oct 13: Staff say Grok Bot routines were held back again for a few hours Tuesday morning US time and have run since about noon ET; missed runs won't be replayed (what to check and re-run). (confirmed)

Update, Oct 13, 2:30 PM ET: Cursor says Origin now supports over 20 integrations, and its App Marketplace is open to public submissions from any developer-tool team (what changed and how to install or publish an app). (confirmed)

Update, Oct 13, 5:13 PM ET: Cursor's status page confirms Grok Bot routines started by webhooks, Slack, git, email and Linear failed from 1:52 to 4:29 PM ET, and events from that window won't be retried (what to resend). (confirmed)

Update, Oct 13, 8:40 PM ET: Staff say a planned server move around 4:30 PM ET rebuilt Grok Bot computers; home folders and saved logins survived, system-wide apps and the browser profile did not (what to reinstall). Separately, staff confirmed bots made before an account merge into The X Subscription can stick on "Working on it"; don't reset or delete them (what to do). (confirmed)

1Password reconnect stuck on must be reconciled (confirmed)#

What happened: Colin said the error is an in-app lock after a failed final connect step, often from 1Password rate-limiting a new service account, and each retry creates another "Grok Bot agents" service account. ([[entry:1password-reconnect-must-be-reconciled]])

Who is affected: Mac users reconnecting 1Password after a disconnect or partial sync.

What to do now:

1. Wait about a day; do not keep clicking Connect.
2. Delete leftover "Grok Bot agents" service accounts; keep the shared vault well under 2,000 items.
3. Cmd-Q quit Grok Bot, reopen, connect once.

GitHub repo pages read as 0 stars (confirmed)#

What happened: Colin reproduced empty star, fork, README, and file-list reads from GitHub HTML pages. ([[entry:github-repo-page-reads-empty]])

Who is affected: Anyone asking a Bot to judge a repo from a github.com link.

What to do now: Ask the Bot to curl https://api.github.com/repos/OWNER/REPO or open the repo in its browser.

Cursor IDE rmdir hit a drive root (confirmed staff acknowledgment)#

What happened: deanrie said Auto-review should have stopped a recursive delete after a PowerShell-to-cmd quoting bug on Windows with Grok 4.7, and that this was not caused by user settings. Single report. ([[entry:cursor-rmdir-drive-root-auto-review]])

Who is affected: Cursor IDE agent users on Windows running destructive shell commands with Auto-review.

What to do now: Set Agents > Run Mode to Ask Every Time. Prefer scoped deletes. Keep work on a remote.

Automations miss synced personal skills (confirmed)#

What happened: Colin corrected an earlier reply: synced personal skills are not available inside automation runs even with "Run as you." ([[entry:automations-synced-skills-unavailable]])

What to do now: Put the skill in .cursor/skills/ in the repo, or paste the instructions into the automation.

New permission-based API keys fail CLI agent runs (confirmed)#

What happened: Colin said new permission-based keys fail agent runs on the server while sign-in and model lists work. ([[entry:permission-api-keys-cli-sdk-fail]])

What to do now: Use an older key, or cursor-agent login without CURSOR_API_KEY.

Update, Oct 13: The changelog now lists 0.70.2, dated Oct 11, covering 0.69.0 through 0.70.1. It names Bot email addresses, X search and in-chat apps, but not the unfocused-window fix staff described on the forum. Cursor's status page is All Systems Operational after a short overnight incident. (confirmed)

All screens in use can mean one Bot's screen failed to start (confirmed)#

What happened: On Oct 13 at 4:14 AM ET, Colin said a shared-computer "All screens on the shared computer are in use" error "isn't a capacity problem" when other Bots on that computer still have screens. The message "is misleading in this case." Staff are tracking it as a known issue. ([[entry:screens-in-use-misleading]])

Who is affected: Anyone running several Bots that need browser screens on one Bot computer, especially overnight unattended browser jobs.

What to do now:

1. See whether other Bots on the same computer can still screenshot.
2. Retry once, then try Update Grok Bot's Computer if it keeps failing.
3. If it returns, file times and timezone in Bug Reports rather than Resetting first.

Correction check on the 0.70.2 idle fix (confirmed staff conflict)#

Update, Oct 13: Our Oct 12 fix quoted mohitjain that 0.70.2 keeps the local helper running when the window is unfocused. Later the same day at 4:49 PM ET, deanrie still described the Mac as away after about 10 minutes out of focus or when the screen is locked, and said the helper "doesn't start again automatically when the bot wakes up for a scheduled job." A user asked whether 0.70.2 fixed that thread. No staff confirmation there yet. We added Update notes to that story and reopened the known issue for re-test. Screen lock and scheduled-job restart stay open. 0.70.2 is now in the changelog, but its notes don't mention this fix, and whether it fixes this issue is now disputed. (confirmed)

Overnight Cursor status incident hit Grok Bot (confirmed)#

From 8:22 PM to 11:07 PM ET on Oct 12, status.cursor.com logged "Degraded performance and errors across Cursor services," including the IDE, Cloud Agents, Automations, Grok Bot, and Origin. Grok Bot was marked recovered at about 9:49 PM ET. The incident resolved at 11:07 PM ET. This morning the page shows All Systems Operational. (confirmed)

The X Subscription follow-up (confirmed)#

The Oct 12 unify story still holds. New this morning: bundled-subscription-terms (Last Updated Oct 7, 2026) name X Subscription Corp, 800 W Cesar Chavez St., Austin TX 78701, as merchant of record, and Anysphere as seller for Cursor Top-Ups. Absolute pool sizes are still unpublished. Forum thread 174450 filled with questions on third-party model billing after a switch. At 3:43 AM ET Colin answered that included third-party usage stays with your existing Cursor plan for as long as you keep it, while on The X Subscription third-party models are "billed separately at API list prices through top-ups," with an optional spending cap. At 4:07 AM ET he added that "All 3rd party model usage on the X Subscription is from top-ups." Colin also pointed account-specific issues to support@x.ai. (confirmed)

PersonalAgentBench (confirmed)#

micro1's PersonalAgentBench now ranks six personal agents. Grok Bot is tied fourth at 65% task and trusted completion, behind OpenAI's Dots and Hark. The full table and Brett Adcock's reaction are in [[entry:hark-personalagentbench-second]].

Beyond Grok Bot#

OpenAI's GPT-6.1 Sol was degraded in Codex and ChatGPT Work overnight, and Sol's default context in Codex and Work is now 500k. Anthropic shipped dynamic workflows for Managed Agents and says new distillation defenses cut harvested data by about 90 to 99 percent.

Grok BotHow people are using it#

• Setup of the day: Check every draft invoice was actually sent. Sunnyvale's morning pass marks the tracker Sent only after QuickBooks shows a real send (a ChatGPT stack, per their notes).

Grok BotThe digest#

Screens and overnight work. Devon_Lybbert's Oct 12 report of overnight browser stalls with a stuck "all screens in use" error got a same-morning staff confirmation that the message can be wrong when other Bots still work. See the standalone known-issue. Related older screen-lease reports remain in forum topic 173022.

Idle helper still needs re-test on 0.70.2. mohitjain's early Oct 12 posts named an unfocused-window fix. deanrie's 4:49 PM ET reply the same day still described away after ~10 minutes unfocused, screen lock, helper stop, and no restart for scheduled jobs. Bradley_Street asked if 0.70.2 fixed it; xufeng said they are on 0.70.2 and will report back. 0.70.2 is a forum-named build, not in the changelog, and whether it fixes this is disputed. We added Update notes to our Oct 12 story and reopened the known issue.

X Subscription reaction. Colin answered the biggest question overnight: on The X Subscription, third-party models in Cursor are paid from top-ups at API list prices, not from the shared pool, while existing Cursor plans keep their included third-party usage. He also said nothing happens automatically to existing Cursor Pro plans, and that users who don't migrate will be moved later "with advance notice." SuperGrok Heavy holders who keep Heavy and don't combine accounts keep their linked Cursor Ultra for now; combining turns Heavy into Ultra on The X Subscription (post 48). Premium ($30) is the comparable plan to Cursor Pro, and Pro+ ($60) has no direct equivalent (post 52). Absolute numbers for the Plus-multiple pools are still not on x.ai/pricing. Legal terms confirm Austin merchant-of-record and Anysphere Top-Ups.

Changelog now lists 0.70.2 (Oct 11). Linux users may still need the staff-posted 0.70.2 .deb if apt has not caught up. Ubuntu 22.04 local-exec C++ mismatch and routine inbox-tool gaps stay open as published known issues.

Workflow: a Managed Agents morning brief. Anthropic's Oct 8 reference automation runs a scheduled Claude Managed Agents session that reads bookmarked Slack channels and GitHub PRs and posts one dated brief to Slack. It stores a per-source bookmark and only advances it after Slack confirms the post, so late or early runs don't miss or repeat items. Staff authors Lance Martin and CJ Avilla say the pattern is "frequently used" inside Anthropic, and the post suggests a per-run budget of about three to five times a normal run's cost. We have not run it ourselves; it needs a Slack app, a read-only GitHub token and the Managed Agents beta.

Late known issues (staff, Oct 13). 1Password reconnect can stick on an in-app "must be reconciled" lock after rate limits (174582). GitHub HTML page reads can show 0 stars and an empty README; use the API or the Bot browser (174504). A Windows Cursor IDE report of rmdir at a drive root got a staff note that Auto-review should have asked first (174587). Synced personal skills do not load in automation runs (174578). New permission-based API keys fail CLI/SDK agent runs (174583). Routine saves refused Automations are not available with your plan on Start + SuperGrok are a plan check, staff confirmed (174600; entry).