New features and product updates.

Extension 38.3.8 is rolling out on the Chrome Web Store, and Chrome installs it on its own. It carries the extension side of this release: platform guidance from the first step, the new widget and dialog handling, sheet batch mode, next steps after a run, whole reads of background tabs, faster runs from AI apps, and the search and typing fixes. Extensions 38.3.3 to 38.3.7 shipped earlier this week. Nothing in it needs new permissions.

Secure logins. Save a site's username, password and optional 2FA key once, from the Logins page in Cloud or the side panel's Secure logins panel, where the site you are on is one click away. At that site's sign-in page, a Cloud run types the login, makes the 2FA code and keeps the session for later runs; the model never sees the password, and email or SMS codes, captchas and passkeys still come to you. Each site shows a plain status (Synced from Chrome, Kept from cloud sign-in, Signed out by the site), and saving a second login for the same account offers to update the existing one.

WordPress on any site. wp-admin, Elementor, WooCommerce and cPanel get the agent's guidance on every host, where what it learned on one site used to stay there. It builds an Elementor page as real widgets in one go, creates WooCommerce variations in one batch, and fills Rank Math, Yoast and ACF fields without fighting the editor: in our tests a 10-widget Elementor page took 53 seconds and under a credit. For WordPress.com and Jetpack sites, the agent can offer to connect the WordPress.com connector instead. In Cloud and in the extension.

Job portals, course sites, back offices and seller centres on any domain. Guidance now follows the platform, not the address: job applications on Workday, Greenhouse, Lever, SmartRecruiters, Ashby, iCIMS, Taleo, SuccessFactors and Dayforce (including forms embedded on an employer's careers page); Canvas, Brightspace, Blackboard and Moodle on school domains; n8n, GoHighLevel and Odoo on any host; Xero, Amazon Seller Central and Ads, Meta's business tools; and seller centres for Etsy, eBay, Walmart, Meesho, Shopee, TikTok Shop, Flipkart, Lazada and Trendyol. Replaying failed runs, the right guidance was there from the first step in 71 of 84 job and course runs (from 51) and 79 of 90 seller runs (from 37). Your own skills can now target a platform too, such as all Workday job sites. Naming a site in your prompt also brings its guidance from the first step in more languages, including Spanish, Italian, Portuguese, Vietnamese, Indonesian, and Hindi or Bengali typed in Latin letters: 70% of such prompts in our replays, from 47%.

Date pickers, typeahead lists, code editors, grids and search boxes. A date fills a calendar in one call, including month and year menus and both ends of a range; typeahead lists find options that load late; Monaco, CodeMirror, Ace and VS Code for the Web are written through the editor itself, so a long file goes in whole and undoes in one step; and scrolling a long virtual grid reads every row. Calendars now also open in tabs you are not looking at, and a picker or grid that only updates on screen says to bring its tab forward. A dropdown option already chosen is not clicked again in a background tab, so the click no longer lands on the next field. A search box that stands alone is submitted with Enter, and a box that already holds the query moves on to the next step instead of being retyped. In Cloud and in the extension; needs extension 38.3.8 in the extension.

Back-office runs stop tripping over the browser. We went through hundreds of failed runs from people managing stores, sites and admin panels and fixed the causes that were ours. A JavaScript alert, confirm or prompt is answered the way the task says instead of freezing the tab, including one a page shows while it is still loading (on https pages). A tab that stops responding is reopened fresh. An approval you gave covers the run, notices that leave the page usable no longer stop it, and cookie banners are declined. A form missing details is filled as far as it can go and asks you once, every change is read back from the live page before the run says done, and a request lost to a dropped connection is sent again. In Cloud and in the extension.

Next steps after a finished run. A completed answer now shows one-click follow-ups, and at most one offer to run the task again daily, on weekdays or weekly at the time you ran it. Offers appear only on a run you are watching, never in Private Mode or on scheduled runs, and you can snooze or turn off the schedule offer. In Cloud and in the extension.

Sheet batch mode. A task that works through a Google Sheet keeps each row's progress in an "rtrvr status" column, writes in blocks, and picks up where it stopped on the next run. It never rewrites a row marked done, skipped or failed, and a job that produces nothing for its first rows stops instead of stamping the whole sheet. In our test a 21-row job wrote all 21 rows in 37 seconds. Rows that found nothing are listed at the end and can be retried in one go, a list you ask to have in a new sheet is written there first and then worked row by row, and a website is checked for a contact page before a field is left blank. In Cloud and in the extension.

Background tabs read whole. Chrome draws nothing in a tab it is not showing, so feeds like YouTube, Reddit and LinkedIn used to read thin there. In the extension the tab now draws the moment the agent reads it, without coming to the front: first reads found 16 to 24 YouTube videos and 27 to 46 Reddit posts, and runs on Amazon, Indeed and Fiverr finished without pulling the window over your work. LinkedIn profiles read whole: a profile the agent opens in a background tab draws itself within about half a second, so the agent's next look has every section (a run listing a profile's sections answered in full in 24 seconds), and a run through six profiles in one background tab finished in 83 seconds. Pages that are already drawn, like Hacker News, are read at once, and lists that load as you scroll fill in without the tab coming forward: a Luma guest list read all 317 guests. In Cloud a tab is shown just until it draws, and four parallel YouTube reads went from 4 or 5 videos each to 22 to 30.

Research builds a list in one step, and search is faster. On paid plans the planner can hand a lead list or an enrichment job to one research run instead of searching and reading each site itself: on 10 sample tasks it took a median 26 seconds and 2.5 credits, with 90 of 100 rows checking out, where a planned run took about 200 seconds and 13.5 credits. It is billed at cost, uses your free monthly data-tool allowance first, needs Web Search on, and is off in Private Mode. News, site: and local searches answer in a median 0.43 seconds instead of 3, and local results carry phone numbers and websites. Live since October 27.

Annual plans and a free daily schedule. Paying yearly gives two months free: Starter $100, Pro $300, Team $1,000 and Scale $5,000 a year, and plan credits still refill every month. The free plan now includes one daily Cloud schedule and one in Chrome, run on Free Mode. Live since October 27.

Questions don't stall a run. A how-many or which-ones question marks a recommended choice, and the side panel takes it after 2 minutes without an answer and says so. A waiting run badges the extension icon and shows a Chrome notification, and how many is asked up front only before a live action. In Cloud, each step shows the question it asked and your answer, and files you attach while answering stay with the run. Live since October 27.

Better ways in, offered. Instead of clicking through Gmail, GitHub, Notion or Slack page by page, the agent can offer to connect the app. It can also offer to share a Google Sheet with it, ask you for a file, or ask you to sign in first. Saying no snoozes that offer for 14 days. Live since October 27.

Getting started is quicker. On a computer, the first-day welcome in Cloud now leads with the extension and a short walkthrough video, with Start in Cloud still one click away, and after your first clean Cloud run one card suggests running it in your own Chrome next time; on a phone, Cloud leads. Right after install, the page the extension opens signs it in as the account you already use on rtrvr.ai, or offers sign-in right there. In the side panel, the empty message box types out real example tasks, and Tab puts the one shown into the box.

Tailscale never switches IP without asking. If the tunnel cannot connect, the run stops before touching a site and offers Retry Tailscale, Use Retriever proxy for this run (at the normal proxy rate) or Cancel, and a retry keeps your prompt, files and login profile. Schedules retry once and skip that run unless you allowed the proxy on that schedule. Startup gives up in about 10 seconds, and each exit node is its own clear choice in the picker.

Chrome logins hold in Cloud. Cookie Sync now carries a site's partitioned cookies and browser storage from one consistent copy, and a login made in the cloud browser stays until you choose Sync from Chrome. Auto refresh checks social sites every 15 minutes and other sites every 30, a partly restored login carries on with a one-click resync, a login you removed never comes back, and Continue after Take control keeps the session you just fixed. Page reads mask password fields.

Cloud asks you to sign in instead of stopping short. After you accept Sign in now, or when a saved login needs a code only you have, the run pauses for Take control instead of ending as a partial result. With a password already saved the card offers Review login, and with several accounts on one site you pick one and Continue keeps your choice.

Schedules refuse what cannot run, and say why. A save that would fail at its first run, such as a page tool with no page to open, is refused with the reason on the editor tab that needs fixing, and tools a schedule cannot run are greyed out. A schedule whose tool broke later can still be renamed, retimed or paused. The schedule editor is reorganised into Task, Timing and Options, and switching views in Cloud no longer blanks the page. A Weekdays preset joins Daily, and switching a weekdays schedule to Daily now runs it every day instead of Monday to Friday.

Answers that match the ask. A top-10 request with a field list returns ten rows, not one (8 of 8 replays, from 1 of 5). The agent ranks candidates before acting on the best one, a task that says apply or pick any password carries that permission to the step that fills the form, and extraction keeps only each item's own values, so a directory's own email and phone stop filling every company row. A result that prints undefined gets one more look, and one failed search no longer fails the whole step.

Clicks, keys and page reads behave. Reading a page no longer leaves it scrolled to the bottom (on X a single read had dropped the first 9 posts). A loose match never clicks Dismiss inside a card the plan meant to open, Space toggles checkboxes and chips, Enter opens a link in the same tab, a late pop-up gets its own tab, and a drag that changed nothing retries as a mouse drag. LinkedIn member lists read whole, the first read sees past a cookie dialog, and a confirm the agent cancelled is accepted on the retry.

Long runs answer faster and cost less. Each planning pass starts from the run's progress so far instead of recounting finished items, the pass after you answer a question reuses its cached history, and resuming over MCP no longer re-sends the whole run. A lead-list run no longer pauses before free lookups or holds its scoring while slow profile scrapes catch up. Page reads drop repeated icons and duplicate long names and never carry page script text or an icon's own stylesheet, about 26% fewer tokens on the largest pages.

For AI apps and the API: check_results pages a long run's steps (one reply went from 560,095 to 37,799 characters) and a resume keeps the run's output schema; the MCP schedule tool edits, pauses and resumes Cloud schedules with no browser connected; cloud_agent takes continue_from and dataInputs, and treats a reused trajectoryId as already started (409 over REST); and a $V: variable named without a value is reported instead of typed, while extract and crawl now accept variables. A run sent to your browser starts and answers about 3 seconds sooner: it waits for its tabs to load instead of a fixed 2 seconds and closes them after the answer, so outside the model a Hacker News top 10 went from 4.9 to 1.8 seconds.

Smaller fixes. GPT-6.1 Terra takes GPT-6.1 Sol's place in the model menus at half the price, and a saved Sol keeps working. readFile reads every Excel column, row and sheet once per run (48.5 seconds became 0.6). A dropped connection right after sending no longer shows "Couldn't reach" for a run that started, and Cancel no longer leaves a run on Running. A finished run ends with its own closing note instead of the model's last reasoning, and Cloud plays its success sound and confetti only for a clean success, never for a failed, cancelled or partial run. A Free Mode run in Cloud checks the daily allowance at each step, as the extension does. Cloud and scheduled Sheets runs check Drive access with a fresh token before asking you to reconnect, a file the model provider cannot fetch no longer fails the step, element labels never include page script text, and site guidance is refreshed for LinkedIn, Pearson MyLab, Zillow, MercadoLibre and Reddit search.

Tailscale Tunnels are in beta. Install Tailscale on a computer that stays on, make it an exit node, then open Routing in Cloud, choose Add a proxy or Tailscale tunnel, then Tailscale, and press Sign in with Tailscale. The tunnel is saved as Home via Tailscale: pick it for one chat or make it your default, and a schedule made from a run keeps its route. Sites see your connection's IP, and the browser's time zone and language match your exit node.

Safe by design. The cloud browser cannot open your router, printer, NAS or any other device on your network or tailnet; only public sites go through the tunnel. If your exit node is offline, the run stops before it starts and says why, and it never falls back to a data-center route. WebRTC stays inside the tunnel, so it cannot reveal the cloud server's address. Traffic uses your own connection, so it costs no proxy credits, and an OAuth client or auth key you connect with is encrypted and never shown again.

Beta limits. A tunnel you signed in with runs one task at a time; connect with a Tailscale OAuth client for parallel runs. The exit node has to stay on and awake, pages load at your upload speed, and the rtrvr device's key expires after 180 days unless you turn off key expiry for it in Tailscale.

Browser profiles. A profile is a named set of saved logins with its own exit location, such as one per client. A run on a profile reads only that profile's logins and starts signed out everywhere else, and a sign-in made during the run is saved back to it. Default keeps working as before: the logins Cookie Sync uploads plus the sign-ins saved during cloud runs. Switch, create, rename and delete profiles from the Logins chip under the composer, and set your account default there. Free on every plan.

Bring your logins in. Import a Playwright storage state, a cookies.txt file or a Cookie header into Default or into a profile, from the Logins menu or over the API. A profile's cookie values are encrypted with that profile's own key, and the API lists names only: values are never returned and there is no export.

For AI apps and the API: manage profiles at /v1/profiles with the same API key as /agent and /scrape, and pass profile on /agent or on the MCP cloud_agent tool to run on one.

Extension 38.2.3 is rolling out on the Chrome Web Store, and Chrome installs it on its own. It carries the extension side of this release: the new Continue rule, Partial and abandoned run endings, honest step counts, answers on more page-read runs, faster Stop, whole page reads when a script reads one tab several times at once, and a stricter tab filter for AI apps. Nothing in it needs new permissions.

Fewer Continue pauses. A run now pauses for Continue after 4 rounds in a row with no measurable progress, where it used to pause after 2, and the 40-round ceiling stops only a run that has stopped making progress. A run that never makes progress still pauses after its 5 free rounds, as before. Replaying a week of runs with the real pause logic, Continue clicks fell 54% and the share of tasks that needed a click went from 9.2% to 4.2%; after a first Continue, most runs had gone on to finish. Every pause now also records why it fired. In Cloud and in the extension; needs extension 38.2.3 in the extension.

Every run says how it ended. A run that kept what it had and stopped short ends as Partial, and the side panel says Partial result instead of Error. A run whose side panel was closed, whose extension restarted, or that another task replaced is now recorded as abandoned instead of sending no ending at all; on October 23 that was about 1 in 5 extension runs. A Cloud, API or MCP run that stopped short is recorded as partial too, and the schedules panel shows a partial run as its own state, not a failure. Needs extension 38.2.3 in the extension.

Counts you can trust. Every browser step now ends with what its own calls did, such as "refresh_page ran 4 times; refused, not run: 1", whether it finished, stopped short or was stopped by the repeat guard, and the planner counts from that instead of the step's own summary. Asked to reload a page 8 times, a run that split the work across two steps used to report 4; it now reports the 8 that ran. In Cloud and in the extension; needs extension 38.2.3 in the extension.

A run that only read a page ends with an answer on more pages. The check that spots a raw page read now recognises pages whose only element ids are on links, such as an author's page on a quotes site or most article pages, so a run that returns only what it read is held for one review that writes the answer, instead of returning the page. A one-pass Cloud run gets the same review. In Cloud and in the extension; needs extension 38.2.3 in the extension.

Stop takes seconds while a script waits. A run stopped inside rtrvr.waitFor or rtrvr.waitForUrl ends at the next check, where it used to take about 40 seconds, and a stopped Cloud run's card shows its final credits. In Cloud and in the extension; needs extension 38.2.3 in the extension.

A resumed Cloud run shows its whole cost. After you answer a question a run asked, its card, live credits, end line and stored record add the phases before the pause; before, a run that cost 4.08 credits could show 2.31. Billing was always correct.

Files saved to rtrvr Files come back with their link. A Cloud run that saves a file to rtrvr Files gets the file's id and link in its result and gives you that, and it never opens rtrvr.ai's own app in the cloud browser to look for it; one run had waited 5 minutes on a sign-in page there. A plan told to save to rtrvr Files no longer saves to Google Drive.

Page reads hold up. In the extension, a plan script's own page read now waits for an app such as Google Maps to draw, as Cloud's does. Several reads of one tab at the same moment each get the whole page, where most used to come back broken, and a new tab is read once it has left its blank starting page. In Cloud and in the extension; needs extension 38.2.3 in the extension.

A click that does nothing says so. When a click had no effect because the element wants another gesture, the hint now starts with "your click did nothing: this target expects a drag" (or a double click, a choice from a list, a scroll) before the call to make, so the agent stops retrying the click. In Cloud and in the extension; needs extension 38.2.3 in the extension.

Free Mode on shared networks. Accounts signed in with Google no longer count toward or get refused by the per-network account limit, so students and coworkers on one school or office network each keep Free Mode; on the website this needs the browser to pass Free Mode's device check. When the network is the reason for a refusal, the message now says network instead of device. Live since October 23.

AI apps see only the tabs they ask for. Over MCP, get_browser_tabs with a domain filter returns only matching tabs; it used to add your active tab even when it did not match. Needs extension 38.2.3.

For AI apps and the API: a page the cloud browser's network route refuses, or a page that cannot be read, answers 422 instead of a server error; a tool that ran in your browser and failed answers 424 device_tool_failed over the Direct API, while a browser that does not respond keeps its 503; a call refused for a zero balance carries a Retry-After that grows while the key keeps retrying; and a prewarm that gives way to a real run answers 409.

Smaller fixes. A pause at a run's credit limit no longer says a result came back with it when none did, and the tab list you resume with is current. Run records keep every step even when a script fires many calls at once. Background-tab notes reach the planner only when a page really drew less there. Cloud no longer warns about too many listeners on long sessions.

Extension 38.2.2 is live on the Chrome Web Store, and Chrome installs it on its own. It carries the extension side of this release: formatted typing into rich editors, Google Sheets access that stays granted and the Grant access notice, Private Mode in the model menu, Tools menu runs on Free Mode, clearer step records for the planner, and steps that no longer wait for the local record of the last one. Nothing in it needs new permissions.

The planner stops re-deciding the same questions. Reviewing 134 extension planner traces and 60 Cloud ones, we found much of the heaviest reasoning going to questions the instructions left open on every pass: whether earlier decisions still held, where a question to you belongs, when a script may finish the task, how much one act should take on, how sure a match must be. In Cloud, re-reading and re-deciding earlier passes alone took 23% of GPT-6 Luna's reasoning on its heaviest passes. Each question now has one rule: an earlier step's decisions stand unless a new result contradicts them; an approval you gave covers later steps of the same kind on the same target; a blocker only you can clear gets one question, and the task then finishes with what it has; the agent keeps going while a source it has not tried can still add what you asked for; and the bar for a match is set once and kept. The agent also judges up to about 50 items it can already read itself, in the same step, instead of sending them to another model call. In Cloud and in the extension.

Formatted text reaches rich editors. When the agent types into a rich-text editor such as Wix's blog editor, ProseMirror, Tiptap or Quill, it can paste formatted text: headings, bold, italic, links and lists arrive as formatting instead of plain paragraphs, and the step says which kinds of formatting landed. A Google Doc can be read through Drive and pasted with its formatting. Before this, a Cloud schedule posting a Google Doc to a Wix blog lost every heading, bold run and link on all 17 of its runs; in our tests a pasted Doc kept them on Wix's own editor. Plain typing into these editors now replaces their content instead of landing at an old cursor position. In Cloud and in the extension; needs extension 38.2.2 in the extension.

Google Sheets stop asking for access on every run. When Chrome itself was not signed in to Google, or in Edge or Brave, the extension's Google access came from a one-hour token that was never renewed, so the access prompt and then the sheet picker reopened on every run. It now renews quietly, counts access you already granted, and checks a sheet with the Google account the run reads it with. When Google still refuses a sheet, the side panel, the Cloud run card and the error message offer Grant access, which opens the Google picker on that one sheet. Needs extension 38.2.2 in the extension.

Schedules check the sheets they name. A schedule now checks every Google Sheet its task names, when you save it and before each run, and saving lists any sheet it cannot open with a Grant access link. After 3 runs in a row refused on the same sheet, the schedule pauses and emails you a one-click fix. Before, it ended partial every day without telling anyone: one daily schedule did that from October 15 to 21. In Cloud.

A model answer cut off mid-stream is sent once more. When a provider's response stream broke partway, the step ended with The model response was interrupted and the run stopped: 19 Cloud runs from October 7 to 21, one of them a scheduled article that never got its Google Doc. Now the call is sent once more, to the same provider when its answer had started or to the next provider of the same model, never to another model, and the cut attempt bills only what the provider reports. In Cloud and in the extension.

Cloud reads apps after they draw. Every page the Cloud agent reads, not only a run's first, now waits up to 5 seconds for a single-page app such as Google Maps or LinkedIn to draw, once per page and inside the read's own time budget, and so does a plan script's own getPageTree read. Before, an app opened partway through a run was read as its empty shell, so the agent spent a step waiting or an extraction came back empty. In the first 8 hours after the update, a Cloud run's first page read came back empty 7% of the time, against 14 to 15% before. In Cloud; the extension already waited this way.

Earlier results stay readable. `rtrvr.priorResult(id)` now numbers results in the order a script makes its calls, so an id no longer points at a different result from one run to the next, a replayed shortcut reads the results its earlier scripts stored, and a resumed, retried or continued task never reads a newer result under an old id. A replayed shortcut in Cloud also hands each script the variables the script before it saved, and a pass that failed the first time is not replayed: its repair runs instead. In Cloud and in the extension; needs extension 38.2.2 in the extension.

The planner's record of each step is clearer. A call that answered with a failure is now marked failed, with its error code, instead of done; a step that saved no data for the next one says so; and the record leads with the progress the script reported and which of its calls failed. When the agent's own finish is held for one review of the sheet it wrote, it is told why and how to finish. In Cloud and in the extension; the extension's record needs 38.2.2. In the extension, a tab the agent reads in the background also says why it counts as background: another tab is active in its window, or the window is minimized, covered or not focused.

Steps start sooner in the extension. Before each call to rtrvr, the extension waited for its local record of the step to be saved; that record is now saved while the next step runs, still in order and with nothing dropped. Needs extension 38.2.2.

Private Mode moves into the model menu of the side panel, right under Free Mode. Turning it on turns Free Mode off and says so. Needs extension 38.2.2.

On the free plan, a Tools menu tool that uses AI runs on Free Mode with its own sponsored card, shown in the tool's window while it runs. Before, such a tool failed with Free Mode is unavailable unless a task was running beside it. Needs extension 38.2.2.

The agent submits only when you ask. A task to fill in a form, type text or draft a message now stops before Send, Submit or Post and leaves that click to you; before, a draft-only task could click Send. In Cloud and in the extension.

An action that keeps getting the same result is stopped. When the agent is about to make the same call a 5th time and every earlier try came back the same, or keeps switching between two calls that undo each other, that call is not run: the agent is told why, such as the same page loading again or the same error, and what to try instead, and asking for it once more ends that browser step as incomplete. A background tab is named as the cause only when its page really drew less there. In Cloud and in the extension; needs extension 38.2.2 in the extension.

Browser shortcuts are refused with the tool that does the job. A key the agent presses lands in the page, so Ctrl+R, F5, Ctrl+T or Alt+Left never reached the browser, yet the step came back as a success: asked to refresh a page six times, the agent counted presses that reloaded nothing. Now such a press fails and names the tool to use, such as goto_url with action reload, and Backspace, Delete, the scrolling keys and Ctrl+A do what a real keystroke does. In Cloud and in the extension; needs extension 38.2.2 in the extension.

The agent reads text in ProseMirror editors, Tiptap among them. Reading the page made these editors redraw their content, so the agent saw an empty text box and called the editor blank. Reading now leaves the editor's content untouched, and its headings, text and links show up. In Cloud and in the extension; needs extension 38.2.2 in the extension.

A run that only read a page leads with its answer. When a plan script reads a page and the agent answers from what it read, the run's result is that answer, over the API and MCP too, instead of the page's whole tree and every link on it. The Cloud run card and the Result tab put the answer first and keep the page data one click away, and a partial run names the page that failed. In Cloud and in the extension; needs extension 38.2.2 in the extension.

Extraction-only tasks finish when they are done. The extension held a finished task for one more review of a Google Sheet whenever it had extracted data, even data kept only in memory, so 46 of 206 held finishes in one week went back for a review with no sheet to check. Now only a sheet the task wrote or read holds it, and the review shows an extraction's header row as its header. Needs extension 38.2.2.

Smaller fixes. An answer never carries a link placeholder the page could not fill: a field meant to hold a URL from the page, when the page had none, is left empty and named in the answer. Over MCP, a free account's saved tool and webhook trigger get the paid-plan notice instead of starting a Cloud run. Rover's planner never rejects a visitor's complete answer for a missing receipt, and visitors no longer see an internal retry error. In the Cloud composer, the Cookies, Tools and Models pickers can use the window's full height, so the Cookies site list shows about three times as many sites and Tools about twice as many rows. A page title cut mid-emoji no longer loses that step's page record in Cloud. Every Cloud charge made through an API key, MCP, a webhook, a schedule or WhatsApp is now recorded against its run. Cloud and its live view run on Node.js 22 with the latest security fixes.

Site guidance gets its monthly review: 51 site skills refreshed from a month of runs and recordings, and 16 new ones, among them coursework platforms.

Extension 38.2.1 is live on the Chrome Web Store, and Chrome installs it on its own. It carries the extension side of this release: background-tab clicks, Retry with note continuing a run, fewer pauses after Continue, public document links, math fields, Private Mode files kept in your browser, and smaller uploads. Nothing in it needs new permissions.

The agent keeps working until the task is done. When a planning step ends with a reply that only announces what it will do next, such as "I'll now open each profile", the agent is sent back to do it instead of the run ending there; across 1,266 runs that ended on a written reply, the check caught 35, and 28 of the 30 we read had stopped right after a promise. A long task's 45-minute stretch is no longer read as its end, and the agent is told to serve the goal behind your words, so a task that asks for a button that is not on the page opens the menu that holds it. In Cloud and in the extension.

Fewer pauses once you press Continue. After you answer a pause with Continue, rounds that make no visible progress no longer pause the run again, and a round that adds 3 or more new lines to the page counts as progress; a repeated script, the run's time limit and its credit limit still pause it. Replaying the Continue clicks of 451 runs from October 5 to 18, pauses fell by 42% and runs that paused 3 or more times went from 113 to 1. In Cloud and in the extension; needs extension 38.2.1 in the extension.

Retry continues, it does not restart. Retry with note on a run that stopped or failed carries on from the progress that run had saved instead of planning from scratch: in Cloud from a checkpoint the run leaves behind, and in the extension for up to 2 hours. A plan that works through a list records which item comes next, so continuing picks up there, and a follow-up in the same chat hears what a stopped, failed or paused run finished. Stop gives a running script up to 40 seconds to save the sheet rows it was writing, and asking a run to stop and hand over what it has gets you the deliverable built from what it collected so far. A Cloud follow-up that names no page starts on the page the previous run ended on, with a note in the chat saying so. Needs extension 38.2.1 for Retry with note in the extension.

Clicks land in background tabs. Chrome draws nothing in a tab you are not looking at, so a click whose effect shows on the next frame read as no effect, and the agent clicked again, closing the menu it had just opened: 185 of 357 such clicks from October 18 to 20 were in background tabs. Now, while the Chrome window is on screen, the agent shows that tab for a moment without taking focus from your window, checks the first click there, and puts your previous tab back. In our test, a menu that took 2 clicks and 2 seconds and ended closed now opens with 1 click, confirmed in 0.3 seconds. Needs extension 38.2.1.

More of the page is clickable, and a dead click says what to try. Buttons a page wires up after it loads, through jQuery or as plain elements that only show a pointer cursor, are now click targets: on a Bubble app's page, 47 of 47 such buttons, against none before. When a click does nothing because the element wants another gesture, such as a drag to rank, a double click, a custom dropdown or a picker wheel, the agent is told which one. The agent also sees which field has focus when its tab is in the background or you are in another window, where 230 of 474 focus steps looked like nothing had happened. After three actions in a row that leave the page exactly as it was, the agent gets a note on what it tried and where to look next; it is a note, never a stop. In Cloud and in the extension; the extension half needs 38.2.1.

Results reach your inbox. A schedule with email turned on now sends the run's answer, its rows as a table and links to the Sheets, Docs, Slides and pages it wrote, instead of only saying it finished, and Slack and Discord get a summary too. In a Cloud run, ask to have the results emailed and close the tab: when the run finishes, they arrive at your account's verified email address, once per run. A scheduled run sends only its schedule's email, so asking in the prompt as well does not send two.

Today means your today. Every Cloud run you start on the website now carries your browser's time zone, and a scheduled run carries its schedule's, so the agent reads today, yesterday and the last 24 hours on your clock instead of the cloud browser's. Before, a run started in Melbourne read today as the day before.

Scheduled runs longer than 5 minutes are tracked to the end. Since October 16, every scheduled run that worked past 5 minutes lost contact with the scheduler at that mark, 5 of 5, while the run itself carried on, so its status could be wrong. A schedule whose run could not start because a lookup on our side failed now says it did not run and used no credits, tries again, and no longer counts toward pausing the schedule. In Cloud.

A run that reaches its credit limit ends as Partial. It keeps what it did, says Stopped at this run's credit limit with the number of credits, where it was shown as Task Cancelled by User. Your own Stop still reads as cancelled. Over the API, such a run's webhook is rtrvr.execution.failed with status partial instead of rtrvr.execution.cancelled. In Cloud.

Public documents read from their link. readFile, searchFile and file inputs in plan scripts now take a public link to a PDF or other document: in the extension your browser downloads it into the run, up to 25 MB, and in Cloud the cloud browser does. A fillable PDF form gives the values typed into its fields under their printed labels instead of an empty form, and a scanned page is read with OCR. Before, a link answered No attached file matches, and some government sites refuse downloads from anything but a browser. In Cloud and in the extension; needs extension 38.2.1 in the extension.

Sheets runs only change the rows they wrote. A task that works through a list adds each row as it reads that item, so a run that stops early keeps every row it finished; in Cloud, a run about to end without writing to a sheet you made required gets one more pass to write first; and fixes and follow-up passes may change only the rows this run wrote, never your older rows. In Cloud and in the extension.

Private Mode keeps your files in your browser. With extension 38.2.1, screenshots, generated images, PDFs, webpages and filled-in forms come back inline and nothing is stored on our side. On every extension version, a message sent after you stop a run no longer fails before it starts because a file was missing.

Inkling and Inkling Small bill 99% off Tinker's list price: $0.01 per million input tokens, $0.0017 cached and $0.0405 output for Inkling, and $0.003, $0.0006 and $0.012 for Inkling Small. The model pickers in the side panel and the Cloud composer show a 99% off chip, with the list price on hover.

Models and search. DeepSeek Flash on Merge now runs DeepSeek V4.1 Flash 1017, the newer build, at the same price. When a provider's content filter cuts a reply short, rtrvr treats it as that provider refusing and tries the next provider once, and on your own key the error says so; before, about 160 steps from October 7 to 16 ended as a success with a cut-off answer. Web search results from Liner now respect site: and -site: in your query, and a list left empty moves on to the next provider.

Math fields and drawn surfaces. Typing into a math editor such as Desmos or DeltaMath goes in key by key and counts only when the equation on screen changes, keys report what they typed, and punctuation keys type the right character. Taps and drags reach shapes drawn in SVG, such as a music editor's note grid. In Cloud and in the extension; needs extension 38.2.1 in the extension.

Crawls in Cloud get the crawl fixes the extension got on October 16: they keep paging up to the number of pages you set, reaching it counts as complete, each page's extraction reads only its own page, and a failed page is reported as an error instead of ending the crawl as if the listing were over.

On the free plan, Cloud API calls made with an API key, the MCP Cloud tools (cloud_agent and cloud_scrape, and agent or scrape when they run in Cloud), WhatsApp tasks and the extension agent's Cloud steps now need a paid plan. Free Mode covers the tasks you start in the extension or the Cloud dashboard, and the browser tools your own AI drives over MCP stay free on every plan. An API key stops working when its owner's account is deleted or disabled.

Smaller fixes. Sign-in checks on Bubble sites no longer take a visitor id for a login or miss a live one, in Cloud, in the extension and on the website. A wildcard host such as *.example.com stays text in the Cloud composer instead of becoming a start URL, and Cloud answers TARGET_UNREACHABLE for one instead of failing. Plan Timeout in the extension takes whole minutes from 1 to 60 and says it limits one plan script, not the whole task. A Cloud run that stops to ask you something no longer fails with a server error when its browser is briefly stuck: it asks, and picks up on the run's starting page. A Jev step that never answers ends at its own time limit, unbilled, instead of holding the run for 15 minutes. The extension compresses what it sends to rtrvr, so a page read is several times smaller on the wire. A model's thinking tags no longer show in the side panel or in Cloud.

For AI apps and plan scripts: the context field of an MCP planner call now reaches the agent, where runs in your browser ignored it; a Cloud tool called with too few credits answers MCP error -32007 or REST 402 with the credit numbers behind the refusal; upload_file answers 422 FILE_FETCH_FAILED with the host and reason when its source URL cannot be fetched, instead of 500; and a later plan script can read an earlier act, extract or crawl result in full with rtrvr.priorResult(id).

Simulations v2: approved results on all 113 tasks, with re-recorded reference runs. Claude Opus 5.5 solved 92.9% at $17.79 an episode at list price, GPT-6.1 Sol 90.3% at $3.54, and DeepSeek V4 Flash at Max reasoning 62.4% at $1.38.