Grok Bot's head of design runs weekly digests and dashboards as pages on a routine. Early users followed. Also: why your Bot's computer naps.

The Daily Bot

No. 014 · Thursday, Oct 22, 2026 · Read on the site

If you tried Grok Bot Pages yesterday, the next step is to hand the refresh to a routine: that's how one of the people credited for the launch says she runs hers. Also inside: OpenAI's new approval rule for routers and AI gateways, a Claude Code command for the new Claude API version, and why your Bot's computer freezes cron jobs overnight. Since this email was late yesterday, it also covers everything that went up on the site Wednesday afternoon.

In this issue: Pages on night one · What's new (4) · Since our last email (9) · How people use it (2) · Workflows (2) · Digest (16) · Known issues (6) · Coming up

Today's lead

What people built with Grok Bot Pages on night one (reported)

What happened: Nothing new shipped overnight; what's new is the first evidence of how Pages gets used. Jenny Wen, whose X bio reads "head of design @cursor_ai @bot," wrote at 5:07 PM ET Wednesday that she has "a few routines set up," naming "a weekly digest of team happenings, a dashboard of core product metrics." She added: "my bot makes pages for these and updates them on a weekly basis." Overnight, Carolina Dill says "My own copy updates every morning" of her numbers dashboard, and Cheesecake asked "Atlas, my chief of staff" for a dashboard with "A fresh version every weekday morning."

Who's affected: Grok Bot users on the latest mobile and desktop clients, where Pages is rolling out. It's one staff member's setup and two single-user reports, each a day old, and official docs still don't describe how pages refresh.

Why it matters: A chat reply scrolls away. A page stays put, has its own link, and can be rebuilt on a schedule. That turns recurring work (the weekly team digest, the metrics board, the morning decision list) into something you open instead of something you rebuild.

What to do:

1. Update Grok Bot. The rollout went to "the latest mobile and desktop clients," per @bot.
2. Ask for the page. Cheesecake's prompt, as posted: "Build me a dashboard of everything we're working on and the 3 to 5 decisions that need me next. Update it every morning."
3. Fix it in chat the way Dill did: ask for a graph instead of a table, labels, or a different order.
4. Put the refresh on a schedule.

Prompt (ours, untested):

Every Monday at 8 AM, rebuild my team digest page from this week's Slack threads and calendar, and keep the same layout.

Good to know: a scheduled refresh runs on the Bot's computer, which sleeps when idle and wakes for routines; staff's advice for overnight work is in today's digest. We haven't built a page ourselves yet.

What to watch: release notes for Pages, whether pages can refresh from connectors without a routine, and whether Tyler Howarth's call for examples turns into a gallery.

Read the full story

What's new

The Grok Bot changelog still tops out at 0.71.2 (Oct 14) as of 6:50 AM ET, with no Pages notes. (confirmed)

OpenAI will require approval for routers and AI gateways that serve its models (OpenAI · confirmed)

If you use OpenAI models through a router or AI gateway, ask that provider whether it has OpenAI's approval. From Nov 20, OpenAI's updated Usage Policies say a service "that enables third parties to directly access OpenAI models" must "obtain our prior review and approval," and OpenAI's post names "model routers or AI gateways." The policy adds: "This requirement does not apply solely because an application integrates OpenAI models to power its features or functionality." OpenAI hasn't published an approval process yet. (not legal advice) Read the full story

Claude Code 2.1.307 can move your code to the new Claude API version (Anthropic · confirmed)

If you have Claude API code to move to API version 2026-10-21, update Claude Code and run /claude-api api-version. The release reached npm at 4:47 PM ET Wednesday and also fixed "HTTP MCP tool calls being sent twice when the connection closed mid-call," per the CHANGELOG. @ClaudeDevs wrote that "2023-06-01 is not deprecated," so you can move when you're ready. Read the full story

A Hark user's four checks after it dropped one word (Hark · reported, one user's report)

Before you act on Hark's research, check the exact wording yourself. In field notes on Medium, Lakshmi narayana .U says Hark got a film festival shortlist right, then reported a subtitle rule contradiction that wasn't there: a rule read "non embedded subtitles only," with "non" at the end of a line, and the agent "omitted the word that sat at the line break." The author writes that it is "one run, three festivals"; the story has the author's four checks and a prompt you can paste. Read the full story

Anthropic previews Claude Duet, an agent app that messages you first (Anthropic · confirmed)

If you're on Claude Pro or Max in the US, Claude Duet's waitlist is still the way in. Anthropic calls it "a new app where Claude works as a proactive agent," and the Help Center says "Invitations go out in waves." Our guide covers how to join and what's still unclear. Read the full story

Since our last email

These went up on the site after yesterday's 2:39 PM ET email.

Grok Bot

• Grok Bot starts rolling out pages for reports, memos and dashboards (Grok Bot · feature · confirmed). The official account posted at 3:42 PM ET Wednesday: "Ask Bot for a page, then edit it together or have Bot keep it up to date." It's rolling out "to users today on the latest mobile and desktop clients," so update your app, and don't worry if you don't see it yet. Story

Anthropic

• Claude API version 2026-10-21 drops temperature; new SDKs switch by default (Anthropic · release · confirmed). Under the new version, a request that sends temperature, top_p or top_k gets a 400 error, "even at the parameter's default value," per the migration guide, and the new major SDK versions send it by default. Read the guide before you bump your SDK. Story
• Claude Code 2.1.304 to 2.1.306: resume warnings, wider auto mode setup (Anthropic · release · confirmed). On Pro or Max you now get a warning before resuming an inactive conversation would use a noticeable share of your 5-hour limit, and /auto-mode-setup runs on Bedrock, Vertex AI, Foundry and the Claude apps gateway, per the CHANGELOG. Update, Oct 22: stable, latest and next all point to 2.1.308, so stable users are past the 2.1.304 regression. Story
• Claude Security now rescans your code as it changes (Anthropic · release · confirmed). On Claude Enterprise, an owner can turn on a version that keeps scanning repositories as code lands, a public beta Anthropic announced Wednesday. The Help Center offers "Up to $25,000 in promotional credits for Claude Security usage in each of two consecutive months," ending Dec 31. Story
• Anthropic's free OSS Scanner accepts 114 projects, now encrypts every report (Anthropic · explainer · confirmed). Product manager Punit Shah posted that about 650 projects applied in the 11 days since launch and 114 were accepted. If you maintain an established, security-critical open source project, the story has the enrollment steps. Story

OpenAI

• OpenAI ships a Codex skill to check sites for PHP2Shell (OpenAI · release · confirmed). If you run PHP sites, upgrade to PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35, the patched releases CISA lists. To find sites still showing an old version, point Codex at the checker repo, which has phpcheck-url for a URL and phpcheck-server when you can log in. Story
• OpenAI API adds 12-hour prompt caching for context reused all day (OpenAI · feature · confirmed). You can now ask the API to keep a reused prefix cached for at least 12 hours instead of 30 minutes, per OpenAI's changelog. A 12-hour write costs "2.5× the uncached input-token rate," against 1.25× for 30 minutes, per the docs. Story
• ChatGPT can now lock chats behind a PIN (OpenAI · feature · confirmed). You can move a conversation out of your history, pinned chats and search, and open it only with a 4-digit account PIN or Face ID, Touch ID or device unlock on supported phones, per the help article. It is rolling out gradually, so you may not see Lock chat yet. Story

Hark

• Adcock says one Hark day can use a fifth of your month (Hark · explainer · reported). Hark founder Brett Adcock replied to a Pro² user: "Daily is not 1/30th of the monthly. We allow you to use up to 1/5th of the monthly each day." Our math: five full days can empty a month, so budget against the monthly bar, not the daily one. Story

How people are using it

• Thursday showcase: five recurring AI jobs people run. From a pizzeria's 6 AM books report in Claude Code to Grok Bot CRM reports, a team's CI fixes in GitHub Actions, a Hark usage panel and a ChatGPT dot as an assistant, each with a starter prompt. Credit: Chris Simone, Brad Smith, Akos, LarMedAI, Treb Gatte (self-reported; disclosures in the piece). Story
• A solo founder's weekday drafts, with selling kept human (Grok Bot). @bageventio, whose bio reads "Founder of BagEvent (since 2015) & MarketUP," wrote at 4:48 AM ET Tuesday: "Every weekday it drafts my X and LinkedIn posts, digs up event organizers worth talking to, and makes a 60 second industry news video every Monday." The line they draw: "Talking to customers is on me." In a reply, @KraJan_303 said merging plans "wiped out nine bots on one account." That's unconfirmed; in the similar X Subscription switch cases, staff say the Bots are safe but can land on a different account (our known issue). (self-reported, one user) Story

Workflows and Setup of the day

Setup of the day: a routine that clears spam texts and junk mail on a timer (Grok Bot · self-reported, two users)

Who runs it: @FoundAt5280 says an hourly routine now handles what used to be "no less then 30 text messages every day that were spam," and @nerdalator's Bot "runs every morning" and asks "who to unsubscribe from." Neither says how the Bot reaches their messages, and we found no official Grok Bot feature for reading texts.

Cost and risk: an hourly routine runs about 24 times a day, and big backlogs can burn through usage: on a 35,000-email pile, @FoundAt5280 wrote the Bot "ran out of tokens overnight and only deleted 8,000." Start with move-only, so a real message can't vanish before you notice.

1. Pick the inbox and the timer. Texts every hour for @FoundAt5280, mail every morning for @nerdalator.
2. Write the exceptions down. @FoundAt5280 targets "all the poltical stuff" but says "I had it exclude my representatives."
3. Let the Bot ask before it unsubscribes. @nerdalator's Bot asks "who to unsubscribe from" instead of deciding alone.
4. Park before you delete. @nerdalator started "with creating folder and label named 'groktobedeleted' so I can review what is moving there before ultimate deletion."

Prompt (ours, untested):

Every morning at 7, look through my inbox from the last 24 hours. Move anything that is clearly spam, political fundraising or a newsletter I never open into a label called "to-delete". Never touch mail from these senders: [your list]. Do not delete or unsubscribe from anything yourself. Send me a short list: which senders you moved, and which ones you think I should unsubscribe from. Wait for my yes before unsubscribing.

Read the full story

A 45-minute morning walk call with a ChatGPT dot (OpenAI · self-reported, one user)

Riley Brown, a creator whose bio reads "Building an @agentnative_ company," wrote at 9:35 AM ET Wednesday: "Every morning I call my dot and put my AirPods in and I walk have a conversation for about 45 minutes." Riley answers email, writes instructions for editors and hands longer jobs to Codex, and described the same call on Oct 1. Dots need ChatGPT Pro or Business Premium, per OpenAI's Help Center. Read the full story

The digest

Grok Bot

• Your Bot's computer sleeps when nothing is using it, and cron jobs inside it freeze too. Follow-up to our Oct 11 story. Cursor staff member Colin explained at 7:42 AM ET Oct 21 that programs inside the computer, such as a sync, a watchdog, an rclone mount and cron, "don’t count as activity, so they freeze along with it," and "There’s no way to keep the computer running 24/7 today." His advice: "Move overnight work into scheduled Bots rather than cron jobs or daemons inside the computer," make long jobs catch up on wake, and note that "A command a Bot starts in its own shell should keep the computer awake for a while (up to about 30 minutes)." He says there's "no need to Reset, Recover or Update." The poster replied that it "helps me rethink how I’m using my Grok Bot setup," and a second user says their Bot's computer also sleeps overnight and asked for a way to keep it awake. (Grok Bot; confirmed by staff)
• A CI routine on one pull request won't fire if Cursor cloud agents pushed the commits. Colin listed four setups where a PR-scoped CI routine never fires: a PR "opened from a fork," CI that "reports results as commit statuses," a commit in "more than one open pull request," and commits "pushed by the bot through the Cursor GitHub App rather than by you." That last one "is an issue the team is already tracking," he wrote at 3:48 AM ET. Staff workarounds in the same reply: a routine on "PR pushed" that waits for the checks, or "A short scheduled routine that checks the PR’s checks and deletes itself once they finish." After the poster confirmed it was their case, Colin added at 4:44 AM ET that he'd added the report "so the Grok Bot routine side is covered too." (Grok Bot routines and Cursor cloud agents; known issue)

Cursor

• Cursor Projects can't use your local MCP servers directly. Colin wrote at 4:43 AM ET that "the Project’s coordinator always runs as a Cloud Agent on Cursor’s servers," so it doesn't load ~/.cursor/mcp.json. Agents it starts on your machine can. His steps: make sure Run on includes your machine, then ask for a local agent by name, for example "Start a local agent on my machine to pull last week’s failed Airflow runs using the airflow MCP tools." (Cursor; staff explanation)
• Bugbot and security subagents failing with a model-not-found error: a fix is rolling out. Colin said at 4:39 AM ET: "A fix is already done and should reach you automatically, without reinstalling." Until it reaches you, ask the main chat to do the review "yourself, without using a subagent." (Cursor; staff says rolling out)
• The Explore subagent model setting only covers Explore subagents. Colin explained at 3:58 AM ET that general-purpose subagents use "the same model as your chat," and the main agent can name a model, a choice that "takes priority over the setting." To steer it, ask in the prompt, for example "launch any subagents with Grok 4.7." (Cursor; staff explanation)
• Cursor windows crash with "oom" at about 4 GB, regardless of RAM. Colin said at 3:45 AM ET it's "a known issue we’re already tracking," worst in long agent chats. His tips: start a new chat regularly, turn off Continue Interrupted Agents under Cursor Settings > General > Startup, and update, since "Newer versions no longer keep full copies of very large files in the chat’s edit history." (Cursor; known issue)

Anthropic

• API platforms can now see when Anthropic restricts one of their end customers. If you send user profiles with your requests, each profile now shows whether Anthropic has an open restriction or ban on that end customer. The release notes for Oct 21 say Anthropic added the enforcement_status and enforcement_actions fields "so you can see whether Anthropic has restricted or banned one of your end customers," plus a new enforcement_status[] filter for listing profiles. The user profiles guide says to send these requests "with no user-profiles- value in the anthropic-beta header, because user-profiles-2026-09-04 and earlier values leave the fields out," and to "Treat an enforcement_status you don't recognize as restricted." On Claude Platform on AWS, "enforcement_status is always unrestricted." At 5:56 AM ET Thursday, the List User Profiles API reference didn't show the new filter yet, so work from the guide. (confirmed)
• Two short Claude incidents, both resolved. Anthropic's status page logged "Degraded performance for Claude Fable 5" from 3:49 PM to 4:29 PM ET Wednesday, opening with "elevated errors on requests to Claude Fable 5," against claude.ai, the Claude API, Claude Code and Claude Cowork. A second incident, "Elevated errors affecting Connectors" on claude.ai and Claude Cowork, opened at 11:50 PM ET Wednesday. Anthropic deployed a fix at 3:28 AM ET and marked it resolved at 4:07 AM ET Thursday. Both were rated minor. Our read: if a connector failed overnight in a scheduled task, run it again. (confirmed)
• Update, Oct 22: Claude Code's stable channel is past the 2.1.304 bug. Our Oct 21 story warned that stable still pointed to 2.1.304, the release with a provider and gateway regression. When we checked npm at 5:56 AM ET Thursday, stable, latest and next all pointed to 2.1.308, which adds only "Bug fixes and reliability improvements," per the CHANGELOG. We've updated the story. (confirmed)

OpenAI

• Codex for OSS grants move up to Pro 200. Charlie Marsh, whose X bio reads "@OpenAI. Building Ruff, uv, ty, and other high-performance Python tools with the @astral_sh team," posted at 1:37 PM ET Oct 21: "we’re upgrading all Codex for OSS grants to Pro 200!" He said there will "still be the 10,000 grants in this round, now at the higher tier," that past recipients should see the upgrade "automatically," and that "Applications are still open." @OpenAIDevs reposted it. The September round offered "$100 Pro plans," per his Sep 15 post. Disclosure: Marsh works at OpenAI. (confirmed, staff)
• API admins can limit which service tiers a project uses. Per OpenAI's API changelog for Oct 21: "You can now choose which service tiers projects can use in Platform settings. Organization settings can restrict service tiers for all projects." Our read: it's a way to keep a test project off a tier such as Ultrafast. The entry links OpenAI's service tier errors page for what a blocked request returns. (confirmed)
• Tibo: "it’s reset day today." OpenAI's Tibo Sottiaux, whose X bio reads "Codex & ChatGPT @OpenAI," posted at 2:55 AM ET Oct 22, as Day 17 of his daily posts: "Bet you thought we forgot, but no, it’s reset day today." That's his claim; he didn't say which usage limits reset or for whom. In the same post he urged readers to "get Daybreak Blue access, it only takes 15 mins." Daybreak Blue is one of two access tiers "for approved defenders," per the API changelog. (staff claim)

Known issues on OpenAI's status page, Oct 21 to 22 (times ET): - Oct 21, 10:18 PM to Oct 22, 4:50 AM: Elevated ChatGPT conversation and Responses API errors. "This is impacting users in India, Japan, and Korea," per the status page. Mitigation applied 10:42 PM. Resolved. - Oct 21, 9:42 AM to 3:34 PM: Elevated FedRAMP API error rates for Astra and Sol. Resolved at 3:34 PM; yesterday's issue had it still monitoring at 2:50 PM, and we've updated that line.

Hark

• One user reports a sign-in that opened an empty new account. Samuel Voltolini (@Wololand) posted at 2:44 AM ET Thursday that after being logged out, signing back in with a phone number, "the same method used during setup", gave this result: "it shows as a new account, all projects and chats and information gone." Voltolini wrote: "I had the most costly plan when this happened." As of 6:10 AM ET we found no reply from Hark, and status.hark.com listed every component, including Authentication Services, as operational. It is one report so far. (user report)
• iOS rating slips as the app stays on 2.4.0. Apple's App Store listing still shows version 2.4.0 from Oct 16 ("New interactions. Several bug fixes. And more refinements to Hark."). Its average is now 3.74 stars from 337 ratings in Apple's lookup data (the listing rounds it to 3.7), down from 3.85 from 300 when we checked at 2:34 PM ET Wednesday (App Store, 6:09 AM ET Thursday). Google Play still shows its Oct 20 update and 100K+ downloads. (store listings)

Known issues and workarounds

The newest open items from Known issues, plus one touched in today's issue:

• Background scripts freeze when the Bot's computer sleeps (Grok Bot, confirmed; updated Oct 22): move overnight work into scheduled Bots instead of cron jobs or daemons inside the computer. Details
• Bots vanished after the X Subscription switch (Grok Bot, confirmed, tracked): don't reset, delete or pay again; sign out everywhere and back in with the same method, then email [email protected]. Details
• Bots show no usage after a half-finished X Subscription switch (Grok Bot, confirmed, tracked): don't buy a second plan or retry the switch; email [email protected] with every account email and a screenshot. Details
• /review-bugbot bills as Composer 2.5 Fast (Cursor, confirmed): review in the main chat with no subagents, or use your own subagent set to [fast=false]. Details
• Gmail connector rewrites links into Google redirects (Grok Bot, confirmed, fix in progress): add or re-paste links yourself in Gmail before sending. Details
• Old Grok app messages keep reaching some Bots (Grok Bot, confirmed): delete old Grok Tasks that point at your Bots; don't just archive them. Details

Coming up

From this week's stories. Everything here is announced, hinted or scheduled, not shipped.

On the calendar

• Nov 20: OpenAI's approval rule for routers and AI gateways takes effect, per the updated Usage Policies (story). (scheduled)
• Dec 31: Claude Security's promotional credits end, per the Help Center (story). (scheduled)

Confirmed, no date

• A fix for Bugbot's model-not-found error. Colin said it "should reach you automatically, without reinstalling." (staff says rolling out)
• A fix for PR-scoped CI routines on bot-pushed commits. Colin said it "is an issue the team is already tracking." No date.

Early signals (unverified)

• Pages release notes. Watch for notes or a help page, and whether pages can refresh from connectors without a routine. (open question)
• OpenAI's approval process. Watch for an approval process or approved-provider list from OpenAI before Nov 20. (open question)

Daily Bot Check

1. How does Jenny Wen keep her pages current? (a) She edits them by hand every week (b) Routines make the pages and update them weekly (c) Pages refresh from connectors on their own
2. Staff's advice for overnight work on the Bot's computer? (a) Run it as a cron job inside the computer (b) Keep your Mac awake all night (c) Move it into scheduled Bots (routines)
3. What did Cursor staff say about Bugbot's model-not-found error? (a) Reinstall Cursor to fix it (b) A fix is done and reaching users automatically (c) There is no fix planned

Answers: 1 (b), 2 (c), 3 (b).

Every story in this email

• What people built with Grok Bot Pages on night one
• OpenAI will require approval for routers and AI gateways that serve its models
• Claude Code 2.1.307 can move your code to the new Claude API version
• A Hark user's four checks after it dropped one word
• Anthropic previews Claude Duet, an agent app that messages you first
• Grok Bot starts rolling out pages for reports, memos and dashboards
• Claude API version 2026-10-21 drops temperature; new SDKs switch by default
• Claude Code 2.1.304 to 2.1.306: resume warnings, wider auto mode setup
• Claude Security now rescans your code as it changes
• Anthropic's free OSS Scanner accepts 114 projects, now encrypts every report
• OpenAI ships a Codex skill to check sites for PHP2Shell
• OpenAI API adds 12-hour prompt caching for context reused all day
• ChatGPT can now lock chats behind a PIN
• Adcock says one Hark day can use a fifth of your month
• Five recurring AI jobs people actually run, with prompts to copy
• A routine that clears spam texts and junk mail on a timer
• A 45-minute morning walk call that clears email and plans the day
• Background scripts freeze when the Bot's computer sleeps
• Full issue, Oct 22

You're getting this because you subscribed at thedailybot.vercel.app. Reply with corrections or tips; a person reads every reply.

How we label: confirmed means we checked a primary source or a staff post; reported means a credible claim we haven't verified; "Our read" is our own analysis.