Skip to content

turno-cli 0.4.0

Latest

Choose a tag to compare

@github-actions github-actions released this 19 Nov 14:17
· 11 commits to main since this release
e1267a5

What's new in turno 0.4.0: https://turno.wimo.dev/docs/releases/v0-4-0

Minor Changes

  • 50facc9: Record and replay. browser_record saves an agent's actions on an identity as a named flow, each
    element kept four ways (Playwright's locator, role and name, text, CSS path) and typed text as
    parameters; passwords, one-time codes and values marked sensitive are never saved, and a vault
    sign-in is kept as browser_login with the site name. browser_replay runs a flow with
    parameters (keychain:<item> reads one from the secret store) and, when a step no longer fits,
    stops with a report the agent can continue from; browser_flows lists them. Opt-in per identity
    (record in identity.json, turno browser record <id>): every session as a flow, Playwright
    traces and a screenshot per action under a size cap, with the trace stopped around sign-ins.
    New commands turno browser flows|replay|record|traces; traces --open uses the trace viewer
    bundled with patchright-core.
    With "sharing": "tabs" a recording holds only the session that started it and replay runs in
    that session's tabs; traces and screenshots are refused in attach mode and in read mode.

  • e94a153: Natural language actions and captcha solving for browser identities, both not recommended and off by default.

    • turno identity set browser:<id> --act lists browser_act (an instruction carried out one stock browser action at a time, every step shown), browser_extract (data off the page as JSON) and browser_observe (proposed calls, not run). The model is your client's, through MCP sampling, or a chat completions API you name with --act-provider chat-completions --act-endpoint <url> --act-model <name> --act-key keychain:<item>. The model sees the instruction and a snapshot of the current tab with vault values and password, one-time code and card fields hidden; it can only click, type, select, press keys, hover, drag, wait and go back on that tab; steps, tokens and time are capped per call (--act-max-steps, --act-max-tokens, --act-timeout). A recorded flow keeps the stock steps it chose, never the instruction or the page.
    • --captcha-provider task-api --captcha-endpoint <url> --captcha-key keychain:<item> lists browser_solve_captcha, which sends a reCAPTCHA v2, hCaptcha or Turnstile widget and the page address to that solver service and puts the token into the page. The service sees the page, and solving a site's captchas by machine may break its terms: it is meant for your own accounts. Every attempt is a line in the identity's reputation log, and turno browser reputation shows a solved column once a solver answered. Read mode refuses it.
    • Keys are only keychain: references read from the secret store at use time, refused as literal values without being echoed, and hidden from every result.
    • identity.json format 8 (act, captcha); turno-browser API 1.7, which turno-cli now requires.
  • dcc2b97: Two opt-in engines for browser identities and servers, both off by default.

    Read mode runs an identity (or one server) on Lightpanda instead of Chrome, for reading pages
    with a fraction of the memory: turno browser set <id> --engine lightpanda, or
    turno-browser-mcp --engine lightpanda. turno never downloads Lightpanda; it runs the
    lightpanda on your PATH (or at TURNO_LIGHTPANDA_PATH), on a private Unix socket with
    Lightpanda's telemetry off. The tools that work there run; the others (screenshots, dialogs,
    drag and drop, uploads, browser_login) are refused with the reason before they reach the
    browser. An identity in read mode keeps its own cookie jar, apart from its Chrome profile.

    Attach mode connects to the Chrome you already have open, so the agent works in your signed-in
    browser: turno browser set <id> --attach auto (after you turn on remote debugging at
    chrome://inspect/#remote-debugging, Chrome 144 or newer), a Chrome user data folder, or a
    loopback endpoint; turno-browser-mcp --attach <target> for one server. Chrome asks you to allow
    each connection. turno works in a tab of its own, closes only the tabs it opened (on
    browser_close, at the end of the session and at the idle limit) and never closes Chrome; the
    tab governor leaves your tabs alone, and browser_run_code_unsafe is refused.

    turno browser doctor gains an Engines section (Lightpanda found or how to install it, a
    Lightpanda left running, which --fix stops) and checks each identity's attach target without
    connecting. identity.json is now format 6 (engine, attach), and the turno-browser API is 1.5.

    Read mode works with "sharing": "tabs". Two pairs are refused with the reason, in identity.json
    and from server flags: attach mode with "sharing": "tabs" (turno keeps one agent's tabs apart
    from yours in your Chrome, not several sessions apart), and read mode with humanize (Lightpanda
    lays out no page, so there is no pointer path to play). Attach mode works with humanize.

  • 4814fb8: Human-like input for browser identities, off by default. turno identity set browser:<id> --humanize light|full|adaptive (or turno-browser-mcp --humanize <level>, which wins) makes the stock click, hover, type, fill form, select, drag, press key and coordinate tools, and browser_batch, move the pointer along a curve from where it last was to a point inside the target, hold the button for tens of milliseconds and type short text key by key with per-key delays. full adds a move time that follows distance and target size, overshoot and correction, a reading pause on a new page, longer gaps between words and wheel scrolling; adaptive is light, and full on a site where the identity met a challenge page. Tool names, arguments and results do not change. The setting is the humanize key of identity.json (format 4, with an optional seed), and turno-cli needs turno-browser API 1.3 for it.

  • bf1cafb: turno browser do <id> <action> drives a browser identity from the shell, for coding agents that
    prefer commands to MCP tools: open, back, snapshot, read, find, click, type,
    press, select, hover, wait, screenshot, tabs, login (a vault sign-in), call (any
    tool of the server), close, session and end. Each action is one tool call through mcpd to the
    identity's turno-browser-mcp, so the tab locks or the one-session lease, attach mode, read mode's
    refusals, humanize, recordings and the vault apply as they do to an agent. The commands one agent
    or one terminal shell runs are one session and keep their tab: turno keeps the mcpd session open
    between them and ties it to the agent (or the shell), and end closes it; --session <name> or
    TURNO_BROWSER_SESSION keeps several apart. Output is url: and title: lines and what the
    action returned, errors on stderr with exit 1, or one object with --json. turno browser <id> <action> is the short form when the id is not a subcommand. The Claude Code plugin adds a short
    browser-identity skill that teaches the loop and reads the installed help first.

  • 6698608: Live view and take-over for browser identities, off by default. turno browser set <id> --live true (or turno-browser-mcp --live) serves a view of the agent's tab on 127.0.0.1 with a per-server token; turno browser live <id> opens it through a link that works once and is never printed. turno browser pause <id> (or Pause in the view) holds the agent's next call, up to live.wait (2 minutes by default), while you act in the window or through the view; turno browser resume <id> lets it go, and each held call is told what changed (URL, title, tabs, your input) and a call that would act on stale refs is not run: it returns an error with a fresh snapshot. The natural language and captcha tools are held too, and a browser_act run stops at its next step. A page that holds a vault value or a password is never shown, including a sign-in typed key by key with humanize. Read mode and attach mode refuse the view and keep pause and resume. turno-browser API 1.9, identity.json format 10.

  • c716051: Opt-in network and user agent settings, and a Camoufox engine, all off by default and not
    recommended for your own accounts. identity.json format 11 adds userAgent and proxy, and the
    engine key takes "camoufox" (also format 11); API_VERSION is 1.10 and turno-cli needs it.
    "userAgent": "coherent" takes no typed value: a headless Chrome gets the user agent a headed one
    of the same version sends, with its real client hints metadata, and a launch check refuses any
    contradiction between the user agent, client hints, platform, headers, languages and time zone.
    proxy sends the identity's traffic (sign-ins included) through an http, https or SOCKS proxy,
    with its password only from the secret store, hidden from every tool result, and WebRTC limited
    to the proxy. --engine camoufox runs an identity on an installed Camoufox through stock
    playwright-core, with its own Firefox profile; turno never downloads it, and every stock tool but
    browser_run_code_unsafe runs, as do natural language actions, browser_login and
    turno browser do. Captcha solving is refused with it, and the live view shows no picture there
    (pause and resume work). The act model's prompt and the live view's tab list hide the proxy
    credentials. New flags on turno identity set (--user-agent,
    --proxy-server, --proxy-username, --proxy-password, --no-proxy), doctor checks for
    Camoufox and the proxy credentials, and fingerprint bench checks that compare headers and a
    worker with the page.

  • 4b9556f: Several agent sessions can now share one browser identity, each in its own tabs. Turn it on per identity with "sharing": "tabs" in identity.json (turno identity set browser:<id> --sharing tabs, or the server's --sharing tabs) and serve the identity with exclusive = false in mcp.toml; the default stays one session at a time. Each session gets its own current tab and holds the tabs it works on: another session's call on one of them waits up to 10 seconds and then fails with a message that names the holder. A session lets go of its tabs when it ends (mcpd now tells servers that declare the io.turno/sessions capability), when a tab closes, or after 5 minutes without a call. browser_close closes only the calling session's tabs, tab lists mark the tabs other sessions hold, and the tab governor's cap counts each session's tabs apart. turno identity add browser:<id> --sharing tabs prints the matching mcp.toml block. turno-browser's API version is now 1.4.

  • 3bd7bc2: Follow-ups for browser identities.

    A tool call no longer starts while the tab governor is still closing a popup or a parked tab, so
    the call cannot fail with "Target page, context or browser has been closed". A Chrome that does
    not stop after SIGKILL gets 5 seconds instead of 2 before turno calls it a survivor.

    A server in read mode stops a Lightpanda that a killed server left running on the same cookie jar
    when it starts, and removes its socket folder.

    turno browser doctor warns when an identity that serves one session at a time
    ("sharing": "session") is served by an mcp.toml server with exclusive = false, and says how to
    switch it to "sharing": "tabs".

    The tab lock wait (10 seconds) and the idle release (5 minutes) of "sharing": "tabs" can be set
    per identity: "tabLocks": { "wait": "30s", "release": "10m" } in identity.json, or
    turno browser set <id> --tab-lock-wait 30s --tab-lock-release 10m. identity.json is now format
    9 and the turno-browser API is 1.8.

    The light humanize level moves the pointer on a visible arc (some 4 to 8% longer than the
    straight line) instead of an almost straight path. With humanize on, browser_login clicks into
    each field and types the value key by key; the keys go to Chrome only, never to a result, a
    recording or a log.

    When another element covers the target of a click, hover, type, select or drag before the call,
    and still covers it 300 ms later, the error names that element at once instead of after the 5
    second action timeout.

    Chrome keeps its third-party storage partitioning on, as a person's Chrome does: Playwright turns
    it off, and turno now passes Playwright's list of disabled features without it.

Patch Changes

  • 31a9a36: turno doctor, turno status, turno mcp daemon status and turno doctor --security no longer
    wait for lsof to tell who holds mcpd's port. A free port is told by a refused connection, and a
    running mcpd's port is checked against the pid it reports, so lsof's scan of every process (which
    takes seconds on a busy Mac) runs only when another process holds the port. When lsof cannot say
    which process that is, doctor still reports the port as in use.

A CycloneDX SBOM is attached.