Skip to content

v1.18.0

Choose a tag to compare

@LvcidPsyche LvcidPsyche released this 14 Nov 16:47
· 52 commits to main since this release
168def7

A security and correctness release over 1.17.0, from the findings the reviews of 1.17.0 left open and from reviews of
this release's own changes. A controller with no API token now refuses a write that a web page sends from another
origin, and no longer answers a DNS-rebound page through a wildcard in CONTROLLER_ALLOWED_HOSTS; behind a
root_path, a file under /artifacts keeps its sandbox. In the fast profile a click on a pay, delete, post or
account button the controller can read asks for approval whatever the model labelled it; a typed text is previewed
only where a read of its field has found it not secret; more log lines quote and cut the text a web page or a mesh
peer sends; and a session closed while it is being created keeps no tunnel. On Linux with Landlock, provider CLIs now
start under a sandbox that closes the controller's data directory to them, outside their own home. Cron webhook keys
are stored hashed, share links can be revoked, and an approval records the operator who decided it where a named
credential proved them.

Read the upgrade notes first, since several change what a running deployment does; among them: a tokenless controller
refuses cross-origin writes and does not start with a wildcard in CONTROLLER_ALLOWED_HOSTS; what a CLI reads or
writes under /data moves into CLI_HOME (or CLI_SANDBOX_ENABLED=false keeps 1.17.0's behaviour); MCP 2025-03-26
is no longer negotiated; the first start hashes the cron webhook keys, schedules run in CRON_TIMEZONE (UTC by
default), and a job whose owner's named credential is gone is refused; decided approvals are pruned 30 days after
their decision from the first start (APPROVAL_RETENTION_DAYS=0 keeps them); the new .tool-call-digest.key in
APPROVAL_ROOT needs a backup; a resumed agent run gets the full max_steps of the run it resumes; and a stop waits
at most 5 seconds for the requests in flight, so a uvicorn started your own way needs --timeout-graceful-shutdown 5
too.

The controller image changes; the client, LangChain and MCP bridge packages now carry the MIT license text in their
wheels and sdists and otherwise change only in version, and the MCP bridge package requires this client; the
browser-node image changes only in version. Advisories: GHSA-7wq9-v3c2-7x2c, GHSA-h2q4-pwpr-qp4j, GHSA-hvjj-c398-6g72,
GHSA-jv4w-prgx-557p, GHSA-5p5g-q85f-m693, GHSA-5wf5-v2cm-mr2j and GHSA-vq4x-w2m7-hj4f, fixed in this release, and
GHSA-786v-hrgx-6wcx, fixed in 1.8.0 and published with them.

Upgrade notes

  • MCP protocol 2025-03-26 is no longer negotiated. A client that initializes with it is offered 2025-11-25, as
    any other version the controller does not speak is, and one that cannot follow disconnects itself. A session that
    negotiated 2025-03-26 before the upgrade, restored from the session store, goes on at that version as it did,
    with or without the MCP-Protocol-Version header, until it initializes again: a stdio bridge that outlives the
    upgrade sends that header on every POST and re-initializes only on a 404. It then replays its client's own
    initialize and keeps the answer, so a 2025-03-26 client behind it goes on at 2025-11-25 without being told;
    the controller answers a session the same at either version. On a session at any other version a 2025-03-26
    header is refused with 400 and -32602.
  • An MCP initialize whose clientInfo or capabilities object is past a bound is refused. Each may be nested
    at most 32 levels and take at most 64 KiB as JSON (a value that is no object is kept as {}, as before); one past
    either answers JSON-RPC error -32602 and makes no session.
    What MCP defines for the two takes a few levels and well under a kilobyte, so a client is refused only for
    something large of its own that it sends there (an icon as a data URI, for one). A session that an older version
    stored past a bound is dropped at this version's start, which then writes the file of sessions once without it:
    its client is answered 404 on its next request, as after an eviction, and the initialize it then sends is
    refused as above until it sends less. That write leaves out each record the start passed over as malformed too.
  • resources/subscribe of browser://audit/events now answers -32602. No event ever notified that resource, so
    a subscription to it heard nothing. Poll it with resources/read. A subscription an older version recorded stays
    silent, as it always was, and resources/unsubscribe still removes it.
  • A failed resources/read now answers -32603, and a page's DOM is capped. A read that failed for any reason
    but an unknown resource or session (a timeout, a closed page, a policy refusal, a screenshot file that has gone, a
    result that cannot be encoded) answered -32002 "Resource not found", and a browser://sessions or
    browser://audit/events read whose listing raised was an unhandled 500 (one whose listing could not be
    encoded answered -32002). A client that took -32002 to mean a resource is gone now
    gets -32603 for those, and should retry or report them. browser://<id>/dom now returns at most 1,000,000
    characters, the most max_chars accepts; a longer page ends in an HTML comment that says so and names
    browser.get_html, which pages through the rest.
  • A session's second GET /mcp ends its first. A client that keeps two event streams open for one MCP session now
    loses the older one when it opens the newer, and a stream also ends when its session is deleted or evicted. Open one
    stream per session, reconnect only when it has ended, and after reconnecting read again what you watch:
    notifications are not replayed.
  • DELETE /mcp with a wrong MCP-Protocol-Version no longer deletes the session. 1.17.0 ignored the header on
    DELETE. A version other than the session's now answers 400 and the session stays: -32602 where the
    controller does not speak it, -32600 where it does. The session's own version deletes it, 2025-03-26 too for
    a session restored at it. Send the session's version on DELETE, or no header, as the stdio bridge does.
  • On a confidential session, browser.set_cookies and browser.set_local_storage now follow an action's rules, and
    a required hosted Witness must answer.
    Both are account_change, so a session with protection_mode=confidential
    refuses them for an anonymous operator, on a shared takeover surface and in a shared browser process, as it refuses
    a high-risk action: the call answers with the reason and nothing is changed. browser.set_viewport is a plain
    write, which those rules do not refuse. All three are refused when the session requires a hosted Witness
    (WITNESS_REMOTE_REQUIRED_FOR_CONFIDENTIAL) and it is not configured or does not answer. Send an operator id
    (X-Operator-Id), run such sessions with an isolated takeover surface and browser, and keep the hosted Witness
    reachable.
  • browser.set_viewport now waits for the session, and can be refused. At 1.17.0 it ran at once, on any page,
    whatever else held the session. It now waits up to 10 seconds for a session another call holds and then answers
    The session is busy with another call. Try again., is refused while the session's page is one ALLOWED_HOSTS
    refuses, and answers The page moved while it was read when the page moved during the call, the viewport already
    changed. A client that sets the viewport while a long action or an agent step holds the session now gets an error
    it never got: set the viewport before the action starts, or try again when it has ended.
  • A governed tool now refuses a workflow_profile other than fast or governed. Any other value (a typo, the
    wrong case, a non-string) used to read as fast, so without FORCE_GOVERNED_APPROVALS the call ran without the
    approval it asked for (browser.eval_js was governed whatever it said, and harness.start_convergence on a live
    session was refused). It now answers an error that names the two values, and nothing runs, with
    FORCE_GOVERNED_APPROVALS on too, where such a call used to ask for an approval: send one of them, or leave the
    argument out (null counts as left out). browser.queue_agent_run and browser.queue_agent_step refuse such a
    value too: neither is governed, but a governed call queues its run or step governed, and a mistyped one left it
    at its request's own workflow_profile (fast by default). Any other tool that is not governed does not check
    the value, as before.
  • A garbled reply from browser.find_by_vision's model is now a tool error, where most read as found: false.
    That is a reply that is not a JSON object, one with no found or a found that is neither true nor false, or a
    found: true reply whose point is not usable (see Changed). At 1.17.0 a reply that was no JSON, or whose found
    was missing or false, read as found: false; one that was JSON but no object, or a found: true reply with a
    null x, answered Tool execution failed; and one whose found was truthy ("true", 1) and whose x and
    y passed int() ("640", a point outside the screenshot or below zero) was answered as found, with the point.
    A client that took found: false to cover those must handle the error, by retrying the call or finding the
    element another way. A reply of found: false stays a normal answer: its coordinates are no longer read (a
    null or no number there failed at 1.17.0), and its confidence is now always 0.0, where the reply's own was
    passed through.
  • A browser.find_elements pattern that ran can now be refused. With regex=true, a pattern whose group repeats
    without bound and holds an unbounded repeat of its own or overlapping alternatives, at any depth ((a+)+$,
    (a|aa)+, ((a|aa))+), fails validation with a message naming the group and why, where it used to reach the
    page (an alternative that is itself a group is compared as written: (a|(a))+ passes). Everyday patterns of that
    shape, with a separator between the two repeats, are refused too (\d+(\.\d+)*, \w+(\s+\w+)*). Those were never
    slow, and an upper bound on the group's repeat, or on each repeat inside it, gets them past the check
    (\d+(\.\d{1,50})*, with {1,50} in place of +); or pass regex=false to match the text literally. A bound does
    not make a slow pattern fast: the check does not count a repeat with an upper bound, so (a{1,50})+$ passes it and
    can hang a page as (a+)+$ did.
  • A controller with no token now refuses a write from another web origin. On a tokenless loopback controller
    (API_ALLOW_TOKENLESS_LOOPBACK=true), a POST, PUT, PATCH or DELETE that a browser marks Sec-Fetch-Site: cross-site or same-site, or whose Origin names another host or port than its Host (Origin: null and one
    that cannot be parsed included), answers 403, and so does a GET /auth-profiles/{name}/export marked that way,
    since an export writes an archive. A page on another origin, or on another port of the same host, that writes to
    such a controller must be served from the controller's own origin instead. A token does not help a page: nothing
    checks one on such a controller, and a page on another origin cannot send one to a controller that has a token
    either, since no route answers a CORS preflight. A browser extension, or a non-browser client that names an
    origin, needs the controller to have a token, and sends it. A loopback reverse proxy in front of
    such a controller must pass the browser's Host through with its port, or the controller needs a token: one that
    rewrites Host, or listens on a port other than the scheme's default and forwards the name without it, leaves
    the browser's Origin naming another host or port, and every write from the dashboard it serves is refused,
    where 1.17.0 served them. In nginx that is proxy_set_header Host $http_host; ($host drops the port).
    A request with neither header (curl, an SDK) is
    served as before, and so is one whose Origin matches its Host with Sec-Fetch-Site same-origin or none.
    /mesh/receive is signed and not checked; /mcp (and /mcp/, which only redirects to it, behind a root_path as
    well) keeps its own Origin check and MCP_ALLOWED_ORIGINS, so a browser MCP client on a listed origin goes on
    working, while /mcp/tools/call is checked like any other write. A controller with a token is not affected.
  • Another operator's auth profile now answers as one that does not exist, everywhere it is read. A caller reading
    a profile someone else owns gets what a missing profile gets, byte for byte: GET /auth-profiles/{name} and
    GET /auth-profiles/{name}/export answer 404, as POST /sessions does with auth_profile or a
    storage_state_path under profiles/, where each answered 403, and browser.get_auth_profile and
    browser.create_session answer their not-found text, where they named the owner. A client that took 403 to mean
    "it exists but is not yours" must treat 404 as either. A save over such a profile and an import that overwrites
    one stay refused (403), and the refusal no longer names the owner. A cron job that names one is refused too:
    POST /crons answers 403, where it accepts a name not saved yet, and browser.create_cron_job says the
    profile belongs to another operator. The reads hide whether a profile exists; these writes do not.
    profile_dir and auth_state.path are gone from the answers of GET /auth-profiles/{name},
    GET /auth-profiles, browser.get_auth_profile and browser.list_auth_profiles: use profile_name.
    GET /sessions/{id}/auth-state keeps its path.
  • The network log of a page ALLOWED_HOSTS refuses is now refused, as a raw CDP read of it is.
    GET /sessions/{id}/network/requests and GET /sessions/{id}/network-log answer 403 with the same fixed text as
    the raw CDP route, browser.get_network_log is a refusal, and browser://<id>/network answers -32603. The page
    is checked before the read and again after it, so one that moves onto a refused host meanwhile is refused too, and
    so is a session whose active page has closed, as any page read is. A session whose open page is allowed gets its
    log as before, and one whose page is refused gets it again once the page is navigated to an allowed URL.
    Both routes now wait up to 10 seconds for a session another call holds and then answer 409 (The session is busy with another call. Try again.): at 1.17.0 GET /sessions/{id}/network/requests answered at once, whatever
    held the session, and GET /sessions/{id}/network-log waited without limit. A client that polls either one while
    an action or an agent step runs must take 409 to mean "try again". GET /sessions/{id}/network/requests also
    answers 404 for a session that has gone, as for one with no inspector, where it read the log of a session
    being closed.
  • Refused takeover-link calls write fewer takeover_link_refused events. One source writes one for its first
    refusal in a minute, and the rest of that minute's are only counted until an event reports them in its
    details.count: the refusals it reports beside its own. An alert that counts these events must add that field:
    events and counts add up to the refusals, or to more where a shutdown cut a write short (what that event reports
    can be reported twice), or to fewer where more than 1,024 sources are held: the oldest is forgotten, with what it
    had only counted. How many events a minute sees is not fixed: a read with the right
    token writes what is held and the minute goes on, so refusals between such reads write an event after each. An
    event that reports others can come late (with the source's next call, at a maintenance sweep or at a clean
    shutdown: see Fixed), so an alert on a rate should read details.count when it arrives, and a controller that is
    killed loses what it had only counted.
  • Cron webhook keys are now stored hashed, and webhook_key_preview is gone. The first start after the upgrade
    replaces every key in the cron store with its SHA-256 hash, so a client goes on using the key it holds; a key or a
    hash it cannot read (a store edited by hand) is left as it is, with a WARNING, or a readable key beside such a hash
    removed, and that job's webhook refuses every key until the key is rotated. A store that
    cannot be re-saved keeps working and the failure is logged; going back to an earlier version leaves each job whose
    key is stored hashed with no usable key. A 1.17.0 controller started over a 1.18.0 cron store hands out each
    stored record whole, so its cron answers show webhook_key_hash (a hash of the key, not the key) and, for a job
    that was refused (below), refusal_audited. A backup or quarantined copy of the store from before the upgrade
    still holds the keys, and so does a .crons.json.*.tmp file beside the store, which a 1.17.0 save that was killed
    part-way leaves there and nothing removes: delete any such file, and rotate the keys if a copy may have been read.
    GET /crons, GET /crons/{id}, POST /crons, browser.list_cron_jobs and browser.create_cron_job answered the
    key's first 8 characters in webhook_key_preview: they now answer
    webhook_key_id, a short id derived from the hash that tells keys apart and is no part of the key. A lost key
    cannot be read from the store any more: POST /crons/{id}/webhook-key/rotate or browser.rotate_cron_webhook_key
    makes a new one and answers it once, ending the old one, and DELETE /crons/{id}/webhook-key or
    browser.revoke_cron_webhook_key removes it (a stored key it cannot read stays, and the answer is
    revoked: false).
  • Cron schedules now run in CRON_TIMEZONE, UTC by default, not in the controller process's zone. 1.17.0 gave
    the scheduler no zone, so it used the one its own process saw. In the shipped image that is UTC whatever the
    Docker host's clock says (neither the image nor the compose files set TZ or mount /etc/localtime), so a
    Docker deployment that added neither keeps its hours: leave CRON_TIMEZONE unset there, since setting it to the
    host's zone would move jobs that never moved. A controller run straight on a host that is not on UTC, or in a
    container given a zone of its own (TZ), fires its jobs at a different hour after upgrading: a 0 9 * * * job
    meant nine in that zone and now means nine UTC. To keep the old hours there, set CRON_TIMEZONE to that zone's
    name (America/Chicago, Europe/Berlin). A name the zone database does not know stops the controller at startup.
  • POST /crons/{id}/trigger no longer answers 404 for an unknown job. A caller without the key could tell
    which job ids exist. An unknown id, a job with no key and a wrong key now answer the same 403 (Not permitted),
    and so does a job whose stored key cannot be read (a store edited by hand): treat 403 as covering all of them. A
    job that is disabled still answers 409 once the key is proved. So does a
    run the controller refuses to start with the key proved where 1.17.0 answered 403 (an auth state too old or not
    encrypted, a start_url that ALLOWED_HOSTS refuses or whose load the pre-request gate stops), 404 (a stored
    job with no goal) or 400 (a stored job the run's request model refuses). A start_url that does not load
    answered 400 and now answers 503: send that trigger again. The route's 403 now means the key and nothing
    else, and its 400 the request itself: its body, or a job id in the path that is blank or longer than 50
    characters (see Fixed).
  • A cron job whose owner has no named credential any more is refused, and so is a job a named operator queued.
    A job fires as the token-verified operator who created it, and went on doing so after their token was removed. Each
    fire now checks API_BEARER_TOKENS as it is then, where a token listed twice proves only the first operator
    listed with it. An owner no credential names is refused: the answer is
    {"triggered": false, "job_id": "<id>", "reason": "owner_credential_removed"}, last_status reads
    refused: owner credential removed, no session is opened, and a cron_job_refused audit event is written when
    the job becomes refused, as the operator system for a scheduled fire and as its caller for a trigger; later
    fires of a job still refused write no event and are logged at debug. The job's entry in the cron store records
    that the event is written, so a restart writes no second one. An event the audit store could not take is written
    at the job's next refused fire, and so is one whose writing the cron store could not record, which the audit log
    then holds twice. A fire the credential allows ends the refusal, and the next refusal writes its own event. Where
    the cron store could not record that end, the record of the run that fire queued ends it; so does a run that a
    1.17.0 controller queued over the same store, which keeps the record's mark without knowing it. Where neither the
    end nor a run was recorded, the next refusal may write no event: this version logs an error when that can be so,
    and 1.17.0 logs nothing. A job a named operator queued fails when a worker takes it, with an error that says so,
    if that credential has gone meanwhile. Moving to the shared API_BEARER_TOKEN or to no token counts: every job a
    named operator created is then refused. Give the owner a named credential again, or delete the job and create it
    again under a credential that stays. Deleting it needs the owner's credential as well, as listing it does: give
    the credential back for the delete, or stop the controller and remove the job's entry from the cron store
    (CRON_STORE_PATH, /data/crons/crons.json by default). A job created on the shared token or with no token is
    unowned and runs as before, as does a job queued by an operator named only by a header.
  • A resumed agent run now gets the max_steps of the run it resumes, where it got what was left of them.
    POST /agent/jobs/{id}/resume and browser.resume_agent_job with no max_steps queued a run of the steps the
    first had not used, and of one step when it had used them all. They now queue a run of that run's whole
    max_steps, so a resumed run can take more steps than were left: pass max_steps to set the number. A resumed run
    also no longer presents the approval_id its first run was queued with, so a step of it that needs an approval asks
    for a new one. After a downgrade to 1.17.0, every run job this version left queued fails when it runs, and no run
    job it wrote can be resumed (400): its stored request holds resume_checkpoints (null for a new run), which
    1.17.0 refuses as a field it does not know.
  • Decided approvals are now deleted 30 days after their decision, and this deletes history. The first start after
    the upgrade, and each maintenance sweep after it, delete from APPROVAL_ROOT and STATE_DB_PATH the approvals
    decided, approved or rejected, more than APPROVAL_RETENTION_DAYS (default 30) ago, with a tool call's stored
    arguments. The days count from the decision, for an approval that has run since too, and the decided ones the
    Changed entry lists are kept. A pending approval is never deleted, whatever its age, so the stores still grow with
    approvals nobody decides. GET /approvals stops listing the deleted ones, and their ids answer as ids that do not
    exist. The audit log's own events about them stay, within AUDIT_MAX_EVENTS. To keep every record, set
    APPROVAL_RETENTION_DAYS=0 before upgrading, or copy APPROVAL_ROOT and the database first. A negative value
    stops the controller at startup.
  • A fast-profile click on a pay, delete, post or account button now asks for approval, whatever its label. A
    click labelled read, write or upload (an unlabelled click is write) on a button whose label starts with
    a phrase such as Pay, Buy, Place order, Delete, Remove, Send or Change password now answers
    approval_required of that kind, as the same click labelled honestly did. So does a type aimed at such a
    button, and a click that names a container when the press lands on such a button inside it. An agent loop that
    ran such clicks unattended in the fast profile now stops there until an operator approves (or the agent itself,
    under AUTONOMOUS_APPROVALS). A form submit with no such label still runs unattended (one labelled read is
    recorded as write, but not where only the press found it), and so do Send code, Remove filter,
    Purchase history and the other harmless continuations (Send link by itself asks: it is also the button of a
    share dialog; so does Send code to all users, since after to a label may go on only with words for where a
    sign-in code goes or what it is for, an address, a masked number, a count of seconds (30s) or another of
    Send's tail words: Send code to me again runs). A menu's command (role=menuitem) is read as a button is, and a
    link inside a button or a menu item as that item. A click or a type, so labelled, by selector on a wrapper with no
    box of its own (display: contents) answers target_has_no_box, and by element_id a click answers
    element_not_visible once ACTION_TIMEOUT_MS has passed and a type focus_lost, with nothing pressed: name an
    element inside it. The direct routes (POST /sessions/{id}/actions/click and /actions/type) carry no label and
    are not read. The whole table, how a label is read, and what the read does not catch, are in
    docs/production-hardening.md.
  • Approvals for tool calls are now matched under a key file, .tool-call-digest.key in APPROVAL_ROOT: back it up
    with the records.
    The first start after the upgrade writes the key there (also when STATE_DB_PATH holds the
    records) and rewrites each stored tool call's stand-in, a request_human_takeover action whose action.text reads
    <tool> sha256:<digest>, so that it reads <tool> hmac-sha256:<key id>:<digest>, in both stores; no other
    approval's text changes, and one it cannot write is logged and taken at the next start. Pending and approved
    approvals go on covering the same call, and a client that reads the tool's name out of action.text has the new
    form to parse. Back up and restore APPROVAL_ROOT whole, the dot file included. Restored without the key file, or
    with another one, pending and approved approvals for tool calls are refused with a message that says so and the
    agent asks again; each start logs how many there are, and rejecting them ends that warning. An operator's earlier
    rejection of a tool call then no longer stops an agent's own approval of the same call under
    AUTONOMOUS_APPROVALS, and each start logs how many such rejections are stored, in a warning of its own: reject
    the call again when it is asked for, turn AUTONOMOUS_APPROVALS off, or restore the key file. The retention
    deletes such a rejection once it is older than APPROVAL_RETENTION_DAYS, which ends that warning (with 0 it
    stays, and so does the warning); a key file restored after that does not bring the rejection back. A start that can
    neither read a key from the file nor write one logs a warning and keys with a key of its own, so the approvals for
    tool calls asked for in that process do not outlive it. Its retention deletes no such rejection, in controller
    scope or in an open session: it cannot tell the file's key from one that is gone, and the next start that reads
    the file matches again the rejections kept under it. Nor does such a start rewrite the stand-ins 1.17.0 stored,
    nor does a decision, a claim or a give-back of one there: it logs how many it left, they match their calls there
    as before (a stand-in keyed under the file's key, by a start that read it, matches only where that key is held),
    and the next start that reads the file rewrites them with what was decided of them there (an
    operator's rejection still stops the agent's own approval of the call, and an approval can still run until it
    expires). A file that holds no key is left as it is, to be removed by hand. After a
    downgrade to 1.17.0, an approval for a tool call that 1.18.0 stored or rewrote matches no call there (the call
    answers that the approval does not match the requested action), and the agent asks again. Nor does an operator's
    rejection of a tool call that 1.18.0 stored or rewrote stop an agent's own approval of that call there under
    AUTONOMOUS_APPROVALS once the call is asked for again on 1.17.0, with the key file kept or not: 1.17.0 compares
    the stored text as it is (so a pending approval that 1.18.0 stored, approved by its id after the downgrade, is
    still stopped by such a rejection). Reject the call again there, or leave AUTONOMOUS_APPROVALS off across the
    downgrade.
  • The image now starts uvicorn with --timeout-graceful-shutdown 5: start it your own way, and you must pass it
    too.
    A stop waits at most 5 seconds for the requests in flight and then runs the controller's cleanup; without
    the option, a stop with an event stream open (GET /mcp, GET /sessions/{id}/events) reaches the cleanup only
    once each stream's client has gone, and never where the supervisor kills the process first (see Fixed). If your own
    image, command or process manager starts uvicorn app.main:app, add the option, in
    seconds and well under what your supervisor waits before it kills the process (compose waits 60). A call still
    running 5 seconds after a stop is now cancelled, where it ran on until it ended or Docker killed the controller.
    At such a stop uvicorn logs Cancel N running task(s), timeout graceful shutdown exceeded at ERROR: that line is
    a normal stop (docs/deployment.md, "Stopping the controller").
  • A REST route's 500 Internal error is logged with no Python traceback under it. Its ERROR record carried the
    error by its traceback (Traceback (most recent call last): and the frames as Python prints them), which an alert
    or a log parser may look for. It now holds, in its message, the route's message, answered 500 for the error's
    class and text, the three outermost frames and the ten nearest the raise, a line each, with the count of those
    between, the first three errors it was raised from or while handling, by class and text, with the count of the
    rest, and of a group the first three errors it holds, the same way (see Security): match on answered 500 for
    instead. The record carries no exc_info, so a handler that reads the exception from it (an error tracker's
    logging integration, a structured formatter) gets the message alone. So are, in the same form,
    POST /mesh/receive's (mesh.receive failed: answered 500 for, where it was mesh.receive error), the ERROR of
    an agent job worker that crashed on a job (agent job worker ... crashed on), of an agent step that failed
    (agent step failed for session) and of a background task that failed (background task ... failed). This
    is the record of the 500 Internal error a route answers where it catches the error (41 of the 54 routes under
    /sessions, and every route of /agent/jobs, /approvals and /audit/events): an error a route does not catch
    (a cron route's, a proxy persona's read, and under /sessions those of the network requests and hooks, the CDP
    element and raw command, the events, the replay, the export script, the trace, the takeover link and the share)
    still reaches Starlette, which answers a plain-text Internal Server Error, and the server logs its traceback, as
    before. The WARNING now written behind a 409 Conflict of POST /sessions and POST /sessions/{id}/fork
    (answered 409 for), which 1.17.0 did not log, has that form with the three frames nearest the raise alone.
  • A controller with no token whose CONTROLLER_ALLOWED_HOSTS holds a wildcard no longer starts. With
    API_ALLOW_TOKENLESS_LOOPBACK=true and no token, on a loopback controller not in production (the only one that
    starts without a token), an entry *, *.example.com or any other wildcard stops startup with a message that
    names it (see Security); a production or exposed controller without a token stops with the error that asks for
    one, as it did. Set a token (API_BEARER_TOKEN, or remove API_ALLOW_TOKENLESS_LOOPBACK: the controller then
    uses the token in API_TOKEN_FILE, and generates one there if the file holds none, and demands it, so every
    client must then send it), or list the exact host names clients use, whose DNS you control. On any controller,
    an entry in brackets round anything but an IPv6 address ([localhost], [127.0.0.1], [*]), an IPv6 address
    in brackets with a port ([::1]:8000, which never matched a Host), an entry with a stray bracket (local]host)
    or a bare *. is now refused too: the controller then serves nothing, and every request, /healthz included,
    answers 500 with a log line that names the entry (started with uvicorn --lifespan on, it does not start).
    Write a name without brackets, and an IPv6 address without a port. The compose files' default no longer lists
    testserver: a client that sends Host: testserver to a controller started by compose needs it in
    CONTROLLER_ALLOWED_HOSTS.
  • An approval asked for a type shows its text only where its field read as plain. An approval now says how
    the field of its type read when it was asked for, as field_read: plain, secret, or unread (the read
    did not answer within a second, what the type names has no field that takes text, or has a frame in it and no
    field that reads as secret, the caller marked the text sensitive, or an earlier release stored the approval).
    Unless it is plain,
    GET /approvals, GET /approvals/{id}, the approval tools, a decision's and an execution's answer and the
    approval_required error give its action.text as "text hidden: its field reads as secret" or "text hidden:
    its field has not been read", as its webhook's and SSE event's summary do (see Security); a text the caller
    marked sensitive shows its stored digest there ([sensitive text] hmac-sha256:...), as before, and its summary
    reads "text hidden at the agent's request (sensitive)". A pending approval
    for a type that 1.17.0 stored shows no text after the upgrade: decide it from the page, or ask for the same
    action again, which reuses the approval and reads its field. A client that retries
    the action from the action an approval answers, not from its own, now sends that text, which the approval does
    not match: retry with the action you asked with. The text_preview of a type whose text submits its form and
    leaves the page as its Enter goes in (a search that ends in a newline) is now withheld, and that type's receipt is
    restricted (a page whose script submits a moment after the Enter keeps the preview, as the field had focus); so
    is that of any type whose field does not have focus once the text is in, or lost it on the way (a text whose
    Enter moves on to the next field, a field that blurs itself on Enter, a component that hands focus to a field
    in its own open shadow root and takes it back), and of one typed by a selector that names a frame.
  • A type by selector that clears its field can now end as focus_lost during the clear. It checks after
    select-all and after Delete that focus is still in the element its selector named, as a type by element_id
    does, but for focus in a frame inside that element, which counts as in it by selector and not by element_id
    (such a type keeps no preview: see Fixed). A page that moves focus during the clear, or puts another node in the
    field's place as it empties, now gets the retryable focus_lost with no text typed, where the text went into
    whatever had focus; so does a type whose element is not found again, or does not answer, within a second. Where
    the page replaced the field, a second type finds the new node. clear_first: false skips the clear and its
    checks.
  • Behind a path prefix, REQUEST_RATE_LIMIT_EXEMPT_PATHS names paths as the router matches them. Only a
    controller served behind an ASGI root_path (uvicorn --root-path /api, UVICORN_ROOT_PATH) is affected. The
    list is now compared with the path that follows the prefix, as every other path rule is (see Security), so the
    default list works there as it stands. A deployment that wrote its prefix into the list (for example
    /api/healthz) loses that exemption: write each entry without the prefix (/healthz). Behind a prefix the
    paths that answer without a token (/healthz, /readyz, /version, /dashboard and the rest of that list)
    now do, as they do without one, where each answered 401.
  • A failed agent step's error_code is now 502 or 503 where the provider failed, never the provider's own
    status, and its error text is fixed.
    A step the provider failed carried the provider's HTTP status as its
    error_code: it was also the
    response status of POST /sessions/{id}/agent/step, and it is in the steps of a run and of a queued job's result.
    A 401, a 404 or a 429 from the provider came back as that. A client that read it must now read 502 (the
    provider refused, and a retry will not help) or 503 (a rate limit, a server error, a timeout, or a connection
    that could not be made or broke while it was read or written: try again later), with the error
    Provider request failed. Any other failure of a step now says only Agent step failed, where it named the
    error's class and text, and its error_code stays null, as before (/agent/step answers it 502): a connection
    the provider closed with no answer and a proxy's refusal are among these, unretried, as in 1.17.0. One step
    error keeps its text: a page that could not be observed says Could not observe the page: and the browser's
    error, cut at 300 characters, in the step, the job's checkpoint and a failed job's error. A job that failed with
    an error nothing handled now records Agent job failed. The detail is in the controller log.
  • POST /sessions/{id}/agent/step and /agent/run answer an approval_id that no record holds with 404, before
    any step.
    A step whose action needed no approval used to run with such an id. Send an approval_id with the step
    it was asked for, and not one the approval retention may have removed (see Fixed).
  • In API mode, a Claude or Gemini provider reports its API model, whatever CLAUDE_CLI_MODEL or
    GEMINI_CLI_MODEL says.
    With CLAUDE_AUTH_MODE or GEMINI_AUTH_MODE at api (the default) and a CLI model set,
    GET /agent/providers and browser.list_providers gave the CLI model as the provider's model, and so did a step
    that failed or waited for an approval, while the request itself went out with CLAUDE_MODEL or GEMINI_MODEL, as
    it always did. They now report the model the request uses. A client that read the CLI model there in API mode reads
    the API model now; in cli mode the CLI model is reported as before. Nothing about a request changes. (STORE-30)
  • The Skills Curator's default model is now the provider's model setting. With CURATOR_ENABLED=true and no
    CURATOR_MODEL, the curator used its own fixed defaults (claude-opus-4-7, gpt-4o, gemini-1.5-pro), which no
    longer matched the models the controller is set up for. It now uses CLAUDE_MODEL, OPENAI_MODEL or GEMINI_MODEL
    for CURATOR_PROVIDER, and the startup line names that model (model=). No code calls the curator's model, in
    this release or in 1.17.0 (the post-session review is not wired to a transcript), so no call and no cost changes.
    Set CURATOR_MODEL to keep a specific model. (STORE-31)
  • On Linux, a provider CLI (*_AUTH_MODE=cli) can no longer open the controller's data directory, except
    CLI_HOME, what CLI_SANDBOX_WRITABLE names there and the screenshot of its call, and outside it writes only its
    own directories.
    Where the kernel has Landlock (5.13 or later), codex, claude and gemini now start under a
    Landlock ruleset: under /data each can read and write CLI_HOME and what CLI_SANDBOX_WRITABLE names there and
    read the screenshot its own call names, and opening, listing or running anything else there answers EACCES
    (docs/deployment.md lists what is not refused: a stat, a watch, a socket connection); outside /data it reads,
    lists and runs what it did before, and writes only in its home, wherever that is, the temp directories, /dev and
    the directories the new CLI_SANDBOX_WRITABLE names. Move whatever your CLI reads or writes
    under /data into CLI_HOME: a CODEX_HOME, CLAUDE_CONFIG_DIR or GEMINI_CLI_HOME set elsewhere under /data,
    a CA bundle, the CLI's own binary. A CLI that keeps state outside its home (a CODEX_HOME or CLAUDE_CONFIG_DIR
    set to a directory elsewhere, which the adapters pass on; a directory a tool of yours writes) needs that directory
    in CLI_SANDBOX_WRITABLE. With the sandbox on, Codex cannot run a command its model asks for, in any
    CODEX_SANDBOX_MODE (measured with codex-cli 0.114.0): its own sandbox does not work beneath this one. The decision
    the controller asks Codex for needs no command. Where any of this cannot be done, or a CLI fails under the ruleset,
    CLI_SANDBOX_ENABLED=false starts provider CLIs without it, as 1.17.0 did, which also gives a CLI /data again. A
    CLI cannot list / itself, on a Linux host run now runs with no_new_privs (the image already set it), and on a
    kernel older than 5.19 cannot rename or link a file from one directory to another. On a kernel older than 6.2 it
    can still empty any file the controller's user may write (Landlock refuses truncate from 6.2). Where the sandbox
    is off or
    cannot be used, or would leave something open (the list is the Security entry's), the controller logs a warning that
    names the condition, once per process for each distinct line, and still runs the CLI: a host run that uses a cli
    provider and keeps its data outside /data gets that warning at its first CLI call, and so does one on macOS or
    Windows. Set CLI_SANDBOX_REQUIRED=true to have no CLI run there instead. See "The CLI sandbox" in
    docs/deployment.md. (N1)
  • API_ADMIN_OPERATORS is new, and once set it refuses the shared token on the admin surfaces. Nothing changes
    until you set it. Set (API_ADMIN_OPERATORS=alice,ops), the maintenance sweep and its status, GET /readiness,
    the /proxy-personas routes, GET /mesh/peers, and the browser.readiness_check and proxy persona tools answer
    only an operator it lists, proven by a named credential of API_BEARER_TOKENS. The shared API_BEARER_TOKEN
    proves no operator, whatever its X-Operator-Id says, so every caller on it is answered 403 there, or a tool
    error over MCP: before you set it, move the scripts, dashboard tabs and MCP clients that use these surfaces to a
    listed operator's named token. Ids are compared exactly, and an id no named credential proves admits nobody.
    (AUTH-32)

Security, Fixed and Changed

This release also has 31 Security entries, 82 Fixed entries and 8 Changed entries, too many for a GitHub release page: read them in the 1.18.0 section of CHANGELOG.md.