v1.5.153

Features

• auth login --bwu logs in with your Bitwarden vault. On a login page it asks bitwarden-use (0.7.0+) which vault logins match the site. It sees them masked, most recently used first, and uses the only one or the one named with --item; several are listed for you to choose from. The values go through bwu run: you confirm once with Touch ID, chrome-use runs itself again with them in that process's environment, and the result lists the steps filled, never the values. Default steps are username, password and Enter. It waits for a password field that only appears after the username was sent, and fills the TOTP on a code page, including one opened separately and sites that submit the code themselves. An item's _autotype field (rofi-rbw syntax, e.g. username:enter:delay:password:enter) sets the steps instead. --no-submit fills only. A secret command's secrets list is scrubbed from its response entry by entry. (#391)
• addinitscript <js> / addinitscript --file <path> adds a script that runs before the page's own in every page the session loads from now on, and prints its handle for removeinitscript. It had only been reachable through --init-script at launch. (#390)
• help and help <command> work as commands. (#390)

Changed

• Output is shaped by how agents actually call chrome-use. Over 215 real agent sessions, 74% of calls piped output through tail/head and 24% discarded stderr, so an error printed only to stderr, or one whose last line was boilerplate, never reached the model. Now:
  • an error also goes to stdout when stderr goes to /dev/null, and its last line is the next step;
  • the --launch test-profile notice prints only on the call that launches, not on every command;
  • plain-mode eval prints a string result as text, byte for byte, instead of a JSON-quoted literal, so a JSON.stringify(...) result is parsed once. A string that would read as another type (empty, or a bare number/true/false/null) keeps its quotes; --json is unchanged. (#390)
• "session unresponsive" names one successor session (foo → foo-2 → foo-3) and says to keep it. It used to say "use a different --session name", and one agent went through 98 names, each leaving a daemon behind. (#390)
• Commands agents guess point to the real one: js → eval, requests → network requests, har → network har start, logs → console, links → snapshot -i -f link. Edit distance had suggested is for js. <command> --help without a help page of its own prints the lines of the full help that mention it, not all 568; tabs --help shows the tab page. (#390)
• core/SKILL.md has a "before you write eval" table: in real sessions eval was 28% of all calls, mostly reading text, clicking by text and reading geometry, which get text, click "text=…" and get box already do. (#390)

Bug Fixes

• Tests no longer write to your friction log. About three quarters of the 4000 lines in ~/.chrome-use/friction.jsonl were test fixtures. Its categories now follow the error codes --json reports, plus page_fetch_failed and policy. Run chrome-use friction --clear once to drop the old test records. (#390)

Contributors

• @leeguooooo