Skip to content

Open-source AI browser agent for Chrome and Firefox (monorepo) 🧠

Changelog: cut 42.2.0

Emre Sokullu0de7125bd02f
diff --git a/CHANGELOG.md b/CHANGELOG.md
index a3669d619..31bc94f6c 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -6,6 +6,8 @@ This changelog was generated from the repository Git history and release tags. V
## [Unreleased]
+## [42.2.0] - 2026-10-20
+
### Added
- Mode picker in the composer's "!" menu: Auto stays the default, and you can choose Ask, Act or Dev there to keep that mode for the rest of the chat when Auto picks wrong. The "!" button then shows the mode (for example "⚠ Act"), and a new chat starts in Auto again. The Hindi and Bengali names for Act, which read as "law", are corrected.
- Automatic mode selection: WebBrain now picks Ask, Act or Dev for each message, and the Ask, Act and Dev buttons are gone. It stays read-only unless your message asks it to do something in the browser. With a decision model configured (WebBrain Cloud includes one), the choice is made while the page is being read, so a message waits at most a quarter of a second longer. Without one, or when the decision model is unsure, the message starts in Ask; if it needs clicking, typing or opening pages, the model hands it over to Act, quoting your words. That hand-off counts only when your words, read again without the page, also ask for an action; the same message then restarts in Act with the usual plan review and permission prompts. Page content can never switch a message into Act, and right-click and selection prompts stay in Ask. Start a message with /ask, /act or /dev, or press Ctrl+Shift+A, X or D, to choose the mode for that message yourself. Text you selected on the page and the standalone window always stay in Ask.
@@ -27,6 +29,10 @@ This changelog was generated from the repository Git history and release tags. V
- Finishing a form you asked WebBrain not to submit no longer takes most of a minute. The completion check removed every line that mentioned a password, so "fill out this form with a random password … don't submit" reached its judge as "[credential omitted]", and so did the approved plan and the agent's summary. Seeing only an unsubmitted signup form, the judge answered "pending" again and again; in the trace it took four attempts and about 50 seconds, ending only when a slow screenshot check ran out of time. The active model's check, which already holds these texts in the conversation, now receives every line of them, with only addresses shortened and bearer tokens and entered passwords hidden, as in all judge evidence. A decision model on this device, or WebBrain Cloud's own while WebBrain Cloud is also your provider, receives that same copy; any other decision model keeps whole-line redaction, and when that leaves it without the task, it is no longer asked while the active model can judge. A password WebBrain typed into a password field during the run is now hidden in the text judges receive, such as the agent's summary or a field a "show password" control revealed, unless the typing tool could not tell what kind of field it was; other entered values stay visible, so a judge can still catch a mistyped one. The judges are told that a task asking not to submit is complete once the staged values match, that a hidden value neither proves nor disproves an entry, and, for each recorded entry, whether its field read back what was entered, whether it is a credential field and which field it targeted; a credential field's line in the page evidence keeps its role and reference. Staged field edits are judged from the accessibility tree first: a screenshot shows passwords only as dots, and the focused field's blinking caret made each one stale before its verdict returned.
- Filling several known form fields now takes one `browser_batch` instead of one model turn per field. The Act prompts still said to fill one field at a time and to click each field before typing, and the planner named only `set_field`; they now send already observed fields, including those in the latest page observation, through one batch with one `set_field` per field. A verified password entry continues the batch like any other text field, so a login's username, password and button can run in one batch. The completion check now also counts the actions inside a batch, which it did not see before.
- Logging in with one `browser_batch` (username, password, Log in) no longer stops at the button with "The browser changed during action preparation" after you approve the submit card. The page monitor carried a verified entry into the batch's checks for every text field except a password, so the button's form no longer matched what it had observed, and the click waited for another model turn. A password field now counts like any other text field; its typed value is checked inside the page monitor and never leaves it. When a click is refused because its binding changed, the trace note now names the reason. Chrome and Firefox.
+- Ask mode can no longer change a site through its API. The API mutations setting, on by default, and `/allow-api` let `fetch_url` send POST, PUT, PATCH and DELETE from an Ask run, so with Automatic mode starting every message in Ask, a model could log in, post or delete through an API before any mode choice, plan review or card. An Ask run now refuses a write-method request whatever the setting says, and the block follows the run's own mode, so a side panel opened during a run cannot lift it by changing the conversation's mode; Act and Dev keep the setting as before.
+- A form submission you declined is no longer asked about again in the same turn. Answering "Do not submit" and then retrying the same submission β€” the same form, button and values β€” brought the same card back every time, which the agent read as more attempts to push it through. A submission is now remembered by a fingerprint of everything it would send: every value the form would submit, passwords, hidden fields and file names included, hashed inside the frame so nothing leaves the page, together with the form and button it goes through and where that button sends it. A retry of that same submission in the same turn is refused without a card, and the agent is told you already declined it; any changed value, a new turn, a correction, or your own answer on a later card asks again, and an answer that timed out still counts as a decline. Controls that only run page script, image buttons (whose click position is part of what they send) and submissions with no readable form have no such identity and ask on their own card every time. `set_field({submit})`, which submits without a button, is identified by its form's own action and method.
+- A form field's own placeholder text is no longer submitted as your value. In live QA on 42.1.2 the agent typed an email field's placeholder, "[email protected]", and asked to submit that instead of asking you for the address. A submission whose changed field still holds its placeholder β€” a value none of your messages or card answers gave β€” is now refused without a card, and the agent is told to ask you for it. Only a whole answer you gave on one of the active task's cards supplies such a value; the request's own words cannot, since no text match can tell "use [email protected]" from "do not use [email protected]" in any language. The check runs first among the pre-submit checks and runs whether or not you ask to confirm submissions, so a scheduled run or a disabled confirmation no longer lets an example through; it covers every control however many the form has (an edited field could previously hide behind 5,000 irrelevant inputs), fields outside any form, closed shadow-root forms in Chrome, and fields WebBrain itself typed into β€” an append whose result was never reported keeps a receipt bound to its field, rebuilt from the task's history after the background worker restarts. Passwords are never listed, and disabled controls, which are not sent with the form, are skipped.
+- A click or typing aimed at the second or later match of a selector inside an iframe no longer fails its submit check. `iframe_click` and `iframe_type` act on the nth match of their selector in a frame, but the submit probe read the first: in live QA on 42.1.2 a request for a button at index 1 asked for approval six times, and each click then failed with "Element not found in any matching iframe"; with another button at that index the card would have described one control while WebBrain clicked another. The probe now reads the same indexed match, and none when the frame has no such match.
## [40.8.0] - 2026-10-16