Repository navigation
Before upgrading. Read these first: several of them change what a script, a proxy, a Compose
override or a Helm values file has to say.
- Migrations run at start and do not go back. The database moves on from
0088_approval_reviewer_guidanceto the migrationrelease-policy.jsonnames for this release,
through the migrations that hold organizations, teams, agents and task runs. Among them,
0089_goal_ends_when_metaddsends_when_mettoresponsibilities(false on every existing
goal, so existing goals behave as before),0090_routine_goal_linked_by_runadds
goal_linked_by_runtoroutines, and0091_remote_mcp_oauthaddsoauthtomcp_servers.
Because every existing routine starts withgoal_linked_by_runfalse, a routine that already
records a goal is read as made for that goal, including a broad routine that 1.0.x linked to a
goal because one of its firings ran it: completing that goal now switches the routine off, and
deleting the goal deletes it. Fourteen tables' rows move into new tables; the old tables are left
in place and no longer written. Restore the backup taken before the upgrade to go back. - PostgreSQL needs two roles. The migrations create two database roles that cannot sign in,
openbot_appandopenbot_system, grant them to OpenBot's login, and from then on OpenBot runs
its queries as them, so PostgreSQL itself keeps organizations apart. OpenBot's login has to be
allowed to create roles, or an administrator creates the two first; otherwise the migration
stops with the SQL to run and changes nothing. Neither role may be a superuser or bypass
row-level security. The server checks at start, and every few minutes, that its queries run as
these roles and stops if they don't, so a connection pooler in front of PostgreSQL has to pass on
therolestart-up parameter. Compose and theopenbot-bundlechart set this up themselves. - The earlier
/api/...addresses are retired. Every client of OpenBot's server now uses the
one API under/api/v2/(below), and every earlier address the server doesn't keep answers 404.
A script, integration or socket client that calls one has to move to the/api/v2address
api/openapi.yamlorapi/asyncapi.yamllists. A reverse proxy or firewall that allows
OpenBot's paths one by one has to allow/api/v2/, with WebSocket upgrades, and keep allowing
the addresses listed under What stays outside/api/v2(below). The API now refuses, with a 400,
a query parameter sent more than once where the spec takes one value, and a body that isn't valid
JSON on the operations whose body is optional, as the spec says. Check-in usage refuses an empty
days=, which it used to read as 7. - The personal edition admits only requests that carry its install token. A script or tool
that reached a personal OpenBot with no credentials is refused: give it an API key from
Settings > API keys, or send the token inX-OpenBot-Install-Token. A single-user OpenBot in a
container doesn't start withoutOPENBOT_INSTALL_TOKEN. OpenBot Desktop,scripts/start.shand
both charts' single-user mode make one (below). - The OpenBot images are checked with OpenBot's own key. They are signed with the key in
docs/cosign.pub(docs/deployment.md), and no longer carry a GitHub build attestation or a
sha-<commit>tag. Deploy a version tag, or its digest fromcontainer-images.json. - Everything already in OpenBot lands in its first organization, still addressed as
default.
Everybody is a member of it and admins are its admins. Nobody is signed out and no page address
changes. Rules written for a directory group (permission switches, network policies, Bots
published or pinned to a group, and shared app accounts' audiences) now name a team in that
organization with the group's name and people, and the API answers them withscopeKind
team. Names that differed only in case get a number, such as "Sales (2)". An OpenBot keeps one
organization untilOPENBOT_MULTIPLE_ORGANIZATIONSis set (below). - Kubernetes: the
openbotchart no longer installs Intelligence. A release on the 0.2 chart
that runs the Intelligence the chart installed is refused at upgrade, and nothing in the cluster
changes (below). - Compose: services and settings are gone. SPIRE (
spire-init,spire-server,
spire-agent), theagent-botand TypeScriptagent-langgraphBots,BOT_PORT,BOT_IMAGE,
AGENT_BOT_MODELandWORKER_SHARED_SECRETare gone, and nothing starts themigrateor
workerservice. A Compose override or script that names any of them has to drop it (below). - Running from a checkout pulls the published images.
bash scripts/start.shno longer runs
the server from the checkout, and the repository no longer builds OpenBot's images or holds the
Helm chart (below). - OpenBot Desktop's back end moves to CopilotKit Intelligence 0.16.0-rc.13, with the bundled
CLI atcopilotkit@7334f7d(below). - Settings and Admin are grouped differently and several pages have new names (below). Old
addresses and bookmarks open the new pages. - Everybody's theme starts on System. A light or dark the app saved before is cleared, so
somebody who wants OpenBot dark whatever the system says picks Dark again (below).
Organizations and teams
- One OpenBot can hold many organizations. Set
OPENBOT_MULTIPLE_ORGANIZATIONS=trueon the
server. Then everybody gets an organization of their own when their account is made, or at their
next sign-in if they were here before; anybody signed in can create another; and people join an
organization by invitation, where before everybody who signed in joined the one there was. Left
unset, OpenBot keeps the one organization it has, as before. The personal edition always has
one. - Each organization is its own. It keeps its own Bots (a new one starts with its own copies of
the starter Bots), conversations, apps and connections, Vault, skills, routines, approvals,
settings, permissions, network policies, Activity, model usage, computers, memory and Learning.
Its owners and admins run it, and to somebody outside it, it isn't there. Require SSO covers the
whole OpenBot, so it is set in the first organization only; the admins named in
INITIAL_ADMIN_EMAILSare admins of the first organization; and a Slack or Teams account, or a
phone number, connects in one organization at a time. - Switch from the account menu. It lists your organizations, your own marked, with
Organization settings, Invitations and New organization where you may create one. Choosing one
opens its home, and another open window follows when you come back to it. In the OpenDots,
OpenDuet and OpenFuse skins, use Settings > Organization. - Owners, admins and members. Whoever creates an organization owns it, only an owner makes
others owners, and an organization always keeps an owner. Settings > Organization renames it
(its id and slug stay) and lists its members, with their roles, Leave and Remove. - Invitations. An owner or admin invites an email address as an admin or a member. The person
needs no account yet, and the invitation lasts 7 days. OpenBot emails it whenSMTP_HOSTand
MAIL_FROMare set, and otherwise Invite gives a link to pass on. The invited person sees
Invitations in the account menu, with Accept and Decline, and the link works signed out and
lands on the invitation after sign-in. WithSMTP_HOSTset, the server won't start without
MAIL_FROM, with an invalidSMTP_PORTorSMTP_SECURE, or with only one ofSMTP_USERand
SMTP_PASSWORD;SMTP_PORTis 587 unless set, or 465 withSMTP_SECURE=true. - Removing somebody, with many organizations on, takes them out of that organization only.
They keep their account, API keys and other organizations, and Restore brings back their role
and teams. Directory deprovisioning does the same. With one organization it works as before. - Teams. Teams open from the menu under your name and from Admin > Organization. Owners and
admins create, nest and delete them, a team's maintainers rename it and manage its people, and
anybody can leave one. A team under another gets what its parent was given. Publish a Bot to
teams or pin it for one, give an app or a skill to teams from its Admin page, set permissions and
network rules per team, and name a team in a shared account's audience. Bot sharing no longer
takes a new directory-group name. - Directory groups are teams. With SCIM, each group in your company's directory is a team,
marked as from the directory and kept in step with it. Its name and people can't be edited in
OpenBot. - Every organization uses the deployment's one Intelligence project and key. Creating an
organization makes nothing in Intelligence.
The model allowlist holds every run, and every run is metered
- Only allowed models applies wherever OpenBot picks the model. With it on, OpenBot refuses a
run before it starts on a built-in Bot, on a Bot on the harness picked at setup, on a Bot that
hands work to agents (judged on its own model), and on a task or routine run on one of OpenBot's
own agents (judged on that agent's model, unless you picked one in that conversation). That
holds in the browser and in every turn that runs without it: routines, Run now, goals,
check-ins, follow-ups, tasks, Slack, Teams, text messages, group conversations and hand-offs
between Bots. A Bot at its own address that picks its own model still runs, and is recorded
against the list, as before. - A refused run says why. A conversation shows the admin's sentence ("This Bot uses
provider/model, which isn't an allowed model. Ask an admin to allow it."), where it could read
as OpenBot failing to reach its backend. A task refused this way reads Couldn't finish and isn't
tried again. A group message every addressed Bot refuses is turned away with the reason and
nothing is kept; when only some refuse, the rest answer and each refused Bot's line says why. - Admin > Providers > Usage counts every run that reached a model, background turns included,
each in its own organization, where it counted browser runs and check-ins. A run on the harness
picked at setup names the model OpenBot chose, where it was listed as unknown.
One API
- OpenBot has one API, described in
api/.api/openapi.yaml(OpenAPI 3.1) is the REST API
andapi/asyncapi.yaml(AsyncAPI 3.1) is every WebSocket channel. Every resource is under
/api/v2/organizations/{organizationId}/, by an organization's id or its slug, and the first
organization isdefault; your own organizations and invitations are at
/api/v2/people/me/organizationsand/api/v2/people/me/invitations. Errors are problem details
(application/problem+json) with a stablecode, lists page withcursorandlimit, and
every write takes anIdempotency-Keyheader. CI lints both files and fails a change that would
break a client of the spec. - Every client is on it. The web app, and with it OpenBot Desktop's window, the phone app and
the OpenBot Recorder call the API through clients generated from the spec, for every area:
Bots, conversations and tasks, routines, approvals, apps and connections, computers and
browsers, Learning and memory, Vault, notifications, admin settings, organizations and agents.
A test fails any new call to the server written by hand outside those clients. - What stays outside
/api/v2. Sign-in, SSO and SCIM (/api/auth), and the addresses a
sign-in comes back to (/api/plugins/oauth/callback,/api/plugins/arcade/verify,
/api/identity/github/callback); the chat runtime (/api/copilotkit) and the Intelligence
gateway's client socket when OpenBot's server carries it (/api/intelligence/client/websocket);
webhooks (/api/events, and/api/delivery/webhooksfor text messages and OpenTag); what a
Bot's harness and Intelligence call back on (/api/agent-tools,/api/model-provider,
/api/learning-model); an outside app calling a tool as the signed-in person
(/api/plugins/call); product events from earlier apps (/api/telemetry); a file's bytes
(GETandHEADof/api/attachments/{id}); the addresses OpenBot Desktop and the released
1.0.2 and 1.0.3 shells call (/api/desktop,/api/host-access/desktop,/api/capabilities,
/api/me,/api/recorder/connect-codes); and/health,/releaseand/.well-known. The
earlier sockets at/api/channels/events,/api/computers/{botId}/streamand
/api/computers/{botId}/terminalare gone, and every other earlier address answers 404. - Watching a run takes a join token. A client asks
POST .../conversations/{conversationId}/join-tokensfor a token that works once, for about half
a minute, with the address and topic to join. - API keys and OAuth, for anything that is not the web app. A signed-in person can make API
keys (obk_..., sent asAuthorization: Bearerorx-api-key). A key acts as that person in
every organization they belong to, on/api/v2and on the socket a run streams on. Apps sign
people in with OAuth 2.1 and PKCE: access tokens last an hour and work on/api/v2and the few
addresses outside it the spec names, and refresh tokens rotate on every use. An app names its
organization with theOpenBot-Organizationheader (id or slug) at those addresses, and works
in the person's own organization without it. Each key may make 600 requests a minute, and past
that gets a429problem withRetry-After. Neither a key nor an app can list, make or revoke
API keys or sessions: that takes a signed-in session. An app's token needscomputers:writeto
type or click on a Bot's screen or terminal, and withcomputers:readit only watches. With one
organization, removing somebody deletes their keys and their apps' sign-ins. OpenBot on one
computer for one person has no OAuth, and has API keys of its own (Personal edition, below). - The API holds requests to the spec. A request body with a property the spec doesn't name is
refused with 400, at any depth, and so is text the database can't store, such as a NUL
character, where several addresses answered 500. Roles areowner,adminandmember. Every
answer carriesCache-Control: private, no-storeandX-Request-Id. AnIdempotency-Key
replays its first answer for 24 hours, except one that held a password, a code, a token or what
was on a Bot's computer, which is answeredANSWER_NOT_KEPT. A query parameter sent more than
once where the spec takes one value is refused with 400, where the first value was kept, and so
is a body that isn't valid JSON on the operations whose body is optional (approving an app access
request, replacing a Bot's grant, replacing a goal trigger's secret), where it was read as no body. - Settings has an API keys page. Under You, a signed-in person lists their keys, creates one
by name and revokes one. A new key is shown once, with Copy. OpenBot on one computer for one
person has the page too, for its one person (Personal edition, below), and OpenBot Desktop
connected to an OpenBot elsewhere opens the page in your browser.
A Bot hands work to tasks and agents
- A Bot can hand a piece of work to a task and keep the conversation short. It says in a line
what it will do, and a task card follows in the conversation. The card shows Working, Needs you,
Done, Couldn't finish or Stopped, how long the task has worked and when it last did something,
and Open thread. When the task finishes, its result is posted in the conversation whole, with
any chart, table or card it drew. A task waiting on you says so in one line with a link to its
thread, and the thread opens with "Sent from" the conversation that asked, as a link back. Bots
are told to hand off only work the person asked for. - Stop a task, or steer it. A working or waiting task has Stop on its card, on its row on the
Bot's Activity page and on the Needs you board. Steps it already finished stay done, and it reads
Stopped, not as a failure. Type into a task's thread to change the work, and the card in the
conversation that started it shows "You told it: ..." under it, and the Bot there reads what you
told the task on its next turn. A turn you run in a task's thread marks it Working, and its
answer is posted in the conversation that asked. Ask the Bot to add to a task or stop one, and
its answer is drawn as that task's card; when your words fit no task or several, it asks which.
A stopped task is never tried again, and the conversation that asked is told how many steps had
finished. - Agents. A new Agents page in the main menu lists the agents Bots hand work to: one that runs
in OpenBot on a model you choose, or a custom agent at an address, private to you or shared with
the organization by an admin. A Bot's Agents page has a Hand work to agents
switch. On, the Bot hands work to an ordered list of agents, the first being its default, with a
model of its own and optional instructions for how it hands out work. Off, it works as before,
and every existing Bot starts off. A task's card and thread name the agent doing it. - See what a task did. In a task's thread, each turn's steps fold into one row, "Worked for
52s", which opens to every step; the time counts the turn's own run, not time spent waiting on
you. Anything waiting for your approval stays in view. A Task details pane beside the thread shows
who ran it and on which model, its outputs, sources and the computer it used. - Finished tasks leave Needs you. A task a Bot handed off that has finished drops off the Needs
you board, since its result is already in the conversation. A Bot's Activity page groups its
work by where it stands (Needs you, Working, Scheduled, Done, Stopped, Failed, Recurring), hides
empty groups, and lists the task threads it opened with a link to the conversation that asked. - A task's card keeps up. It re-reads its task when the conversation's work changes, so a task
that carried on after Done no longer stays Done until a reload, and an open card no longer falls
seconds behind while several conversations work. A task waiting to try again after a passing
failure says until when. A task thread picks up a turn OpenBot's server continued in it, with
any approval that turn stopped on. A step cut off mid-run reads "interrupted". - Shorter reports of finished work. A Bot that finishes a job, itself or in a task, is told to
say what it did, with the link, in a few sentences and no list, then at most one sentence on
what to check, and to call anything it inferred rather than read unconfirmed. - "The last day" means the last 24 hours when a Bot hands work to a task. Bots are told to give
a time window counted back from now, never "since" yesterday's date. - A task tries again only when it could help. A failed turn is sorted first. A passing failure
(a rate limit, a provider or app outage, a timeout, a restart) is tried again after a wait, the
one the provider asked for when it named one, up to three runs. A lasting one (a refused key,
sign-in or permission, an admin's rule, your no) is not tried again. A built-in Bot sends each
model request once under a task's turn; Bots on the LangChain and Claude Agent SDK harnesses keep
their own retries. A task that gives up says what was and wasn't done. - A task makes an app change once. A change is recorded before it is sent, so a task retried,
timed out, picked up after your answer or after a restart doesn't send it again. A change whose
outcome is unknown isn't sent again without checking the app, and a task that couldn't finish
names the changes it sent but couldn't confirm. - A task that needs you waits for you. Asked to take over the Bot's browser, type a secret or
sign in, a task reads Needs you, and carries on once you answer. Waiting for an app, Request
access keeps it Working while an admin decides: a dismissed request reads Couldn't finish, a
withdrawn one Needs you, and once the app is given the task carries on by itself. - Task results arrive in order, and find you when you're away. Results land in the
conversation that asked in the order the tasks finished. If that conversation is still unread a
few seconds later, OpenBot tells you once by push, Slack or Teams. A result that can't be written
yet is tried again, further apart each time, for up to a day. A routine set up in a task's thread
belongs to the conversation that asked. - Work carried on after an approval continues at once after a restart or deploy, where a task
could read Working for up to 15 minutes. - A Bot finishes a long list with subagents instead of stopping partway. One answer could start
five subagents in all, so a Bot given twelve accounts or documents to work through ran out after
five and did the rest itself, one by one, or left them. Now up to 10 run at once and any more
wait their turn, up to 100 in one answer, and a Bot past that is told to do the rest itself. Both
numbers are settings,BOT_SUBAGENTS_AT_ONCEandBOT_SUBAGENTS_PER_ANSWER, and in the Helm
chartconfig.subagents.atOnceandconfig.subagents.perAnswer. - A task shows once in the conversation that asked. A result posted right under its own task's
card no longer repeats the task's title and link, which the card already has. A result that
lands after something else keeps them, so it still says which task it is. A stopped task's
result reads stopped, where it showed a green check. - A subtask that called no tool says "Answered directly", where its line in the task thread
said nothing and the Bot reported it took 0 steps. Task details says how long the task worked. - A long run that finished keeps its answer. On Bots on the Claude Agent SDK harness, about one
multi-step run in 20 ended with "The Bot stopped before it finished answering" though the work
was done, and a task tried again after that kept the first try's failure line. A failure a task
will try again is now said only if the last try fails too.
Conversations
- Bots talk about their work in plain words. They are told never to name a tool, a function or
an app action's code name, never to say how their tools are found or loaded, and never to quote
an error the way a tool gave it, so a reply says "the Teams post didn't go through", not "No such
tool available". - A conversation keeps its name when it and a subagent's conversation are named at the same
moment, where the sidebar could go back to calling it by its first message until a reload. - Maths draws as formulas. LaTeX a Bot writes (
\( \),\[ \],$$) is drawn as formulas in
replies, task threads, cards, group conversations, tool results and skills, where it showed as
source. A single$stays a dollar sign and code stays as written. A wide formula scrolls in its
own block. The renderer and its fonts load from OpenBot itself, the first time a formula
appears. - Pick the model on the home screen. The home message box has the model picker. Your pick is
saved with the conversation of whichever Bot answers. A Bot that hands work to agents, or runs on
an agent of its own, doesn't take the pick, and a notice says so with Open Agents. - A model nothing answers at says so. On Bots on the LangChain and Claude Agent SDK harnesses,
the Personal Assistant among them, an AI model address with nothing listening, a host that
doesn't exist or one that never answers ends the turn with "OpenBot couldn't reach the AI model
at host:port" and what to check, where it said only that the Bot stopped before it finished, or
read as the provider refusing. Only the host and port are shown, never a key or password in the
address. On the Claude Agent SDK harness it can take about three minutes, because Claude Code
retries the connection first. - Attach files before a new conversation's first message. On the home screen and New
conversation, files you pick, paste or drop go with the first message, so the Bot reads them on
that turn. If a file is refused, your words and files go back in the box with the reason. A group
conversation's first message still takes no files. - A file you attach stays a file once the Bot answers. On a Bot that runs behind an AG-UI
address, such as the Personal Assistant on LangChain, a text file you sent turned into one block
of its own contents when the answer finished, and an image's thumbnail disappeared. The message
now keeps the file as you sent it, during the answer, after it and after a reload. Messages
already turned into text stay as they are. - Switching back to a conversation the Bot is working in shows it working. Leave a
conversation while the Bot works and come back in the same window: it could be empty, with no
sign of the Bot, until the whole answer appeared at once. The conversation now picks the work up,
the answer streams in, and a message typed meanwhile waits for the turn, as it does when you
never left. - A conversation that went quiet ends instead of running forever. When a run sends nothing for
45 seconds, the app asks whether it is still going. If it isn't, the turn ends and what was saved
is shown, with "Send a message to carry on"; if nothing can say, it says OpenBot lost touch with
the Bot. Before, it showed Running and Stop for good. Opening a conversation while the live
updates aren't answering now opens it normally after a few seconds, where it stuck with no
history. - A Bot no longer does an app change twice after you answer an approval. When a built-in Bot
asked for several app changes at once and one waited for your Allow, the Bot was told the others
had not run, and did them again. It is now told what they did. And when the conversation can't
be carried on after an approved action, the step says what came of the action, and the Bot's
next turn gets that result. - Replies fit the question. Bots are told to answer "Can you...?" with a yes or no and an
offer, to claim only what their tool results show, to try an empty search another way before
saying nothing exists, and to ask which one when a name fits more than one person or thing,
never to pick. When a lookup finds several, OpenBot holds back the first change that would use
one of them until the Bot has asked. - Allow once is for the action you saw. It lasts ten minutes from your answer. If the action
changed, or the time passed, nothing is done and the request comes back to you with a line
saying what changed. - Share approvals show everything that will be sent. The card carries the whole text, with Show
all, where it was cut at 280 characters; in Slack and Teams it shows who will read it and the
text, up to 2,000 characters. A hand-off you decline, or one refused after Allow, leaves its task
Couldn't finish. A Bot on OpenBot's own harness that hands work to another person's Bot asks with
the card instead of stopping. - A reply a Bot posts on its own shows in the conversation you have open: from a routine, a
goal, a follow-up, Slack, Teams, a text message or the phone app, or a turn carried on after a
sign-in, an approval or in a task, even when it repeats an earlier reply word for word. - A computer step that finds the computer waiting on you in another conversation says so,
where the turn waited with nothing said. - A finished reply stays where it was written. Scrolling down once a reply had finished could
drop it from under your question onto the message box. - Slack mentions in a Bot's reply show as names, never as
<@U...>ids.@here,@channel
and@everyoneread as in Slack. - A failed app call draws as failed, and an app's answer draws without its leading "Ran:" line,
so its JSON renders as JSON again. A long app action is named by its whole name, not a shortened
one with a hash. - Hand over the browser when you ask. Asked to let you take over the browser or type a password
yourself, the Bot hands you the browser rather than showing the one-field secret box.
Bots and apps
- Mastra can be added as a Bot or as an agent. New Bot and a Bot's add-agent sheet list Mastra
among the frameworks. Give the address of your Mastra server itself, not an AG-UI route: OpenBot's
server checks it against Mastra's own API before saving, and says under the address why when it
can't. The server has to run one agent. - Add any MCP server by its address. Connections has Add server: give a remote MCP server's
https address and OpenBot asks it how it is reached. A server that signs people in has each person
connect their own account; one that needs a key takes a key, which in the team edition only an
admin can give. With the MCP allowlist on, an admin approves the address first, and allowlist
entries can now be addresses. A tool the server marks read-only, and not destructive, counts as
a read, so it doesn't ask where changes ask; rules you or an admin wrote still apply. Signing in
to a server needsOPENBOT_PUBLIC_URL.OPENBOT_ALLOW_LOOPBACK_MCP=1admits a server on this
machine, for development only. - A Bot makes an app change once per turn. The same change asked for twice in one turn is made
once, and the second call gets the first one's result. A change whose outcome is unknown, after a
timeout or a dropped connection, isn't sent again until the Bot has read that app. This holds
across server replicas. Arcade and Composio actions are sent once everywhere. Asking again in a
new message does it again. - A Bot stops repeating an app call that keeps failing. Once the same call, with the same
arguments, has failed twice in a turn, OpenBot stops sending it and tells the Bot to change it or
do it another way. - Every app answer tells the Bot what happened. It opens with OpenBot's line saying whether the
call ran and what changed, ran read-only, partly ran, or failed, with the kind of failure and
what to try. A value the model writes as text where the app's schema names a number, a true or
false, an object or a list is sent in that type. - Built-in Bots finish longer app tasks in one turn. A chat turn has 40 model steps, and 8
more when the Bot searches its tools, where it had 8 and 2. - Finding the right action among many. Tool search weighs the apps a conversation has named or
used, and a Bot at an address is given OpenBot's tool search when it holds more actions than one
request carries. The most-used actions of Gmail, Google Calendar, Google Sheets, Google Drive,
Slack, Notion, Linear and HubSpot carry OpenBot's own descriptions, andOPENBOT_TOOL_GUIDES
names a JSON file that replaces, adds or removes them. With several large apps connected, a
message no longer fails with "Runtime info request failed". - Payments ask by what the action is, not what it writes. A Google Sheets row that says
"Payment due" is no longer treated as paying money. An action named for a charge, payment,
refund, payout, transfer, checkout or order, or one that fills a field such ascard_numberor
iban, still always goes to the person. - A Bot's computer has Python. Python 3 with openpyxl, pandas, python-docx, python-pptx,
matplotlib, pypdf, pdfplumber and reportlab is installed in the Bot's computer and in OpenBot
Desktop's folder sandbox, and built-in Bots are told what their computer can run. - Standing instructions reach Bots on OpenBot's own harness. A Bot on the harness picked at
setup, and the managed Bot, follow your standing instructions, after OpenBot's own rules. A Bot
at an address an admin added is never sent them. - A person can only allow a Team Bot an app it uses. Allowing it to use your account for any
other app is refused with "That app isn't one this Bot uses.", where the answer was kept and took
effect without asking on the day the Bot was given that app. - Turning on many of an app's actions for a Bot saves in one step. Activity records one row for
the actions turned on and one for those turned off, each naming its actions, where it had a row
per action; a script that reads the audit trail row by row sees fewer rows. - A model key added in Vault is one the Bots use. Add key for a model asks only which provider
it is for, starts on the one the Bots run on, and says whether the Bots will use it. A key saved
this way is filed where the built-in Bots read it, where before it could be saved under a name
nothing read. - Claude Agent SDK Bots. App actions with names past Anthropic's 64-character limit can be
called. OpenBot's own tools and tools already used load up front. A turn that searches its tools
in the same step as an app call finishes its answer, where it ended with none. Behind a gateway
(ANTHROPIC_BASE_URL), tool search is on, so a request no longer carries every app tool; a
gateway that can't pass tool search on needsENABLE_TOOL_SEARCH=false, which Compose hands to
the harness. A task on this harness is sent the app actions its brief names and finds the rest
through OpenBot's tool search, where it could be answered "No such tool available". - LangChain Bots. A call to a tool not offered for that step is handed back to be answered
instead of ending the run. An OpenAI model that takes tools only through the Responses API is
called there behindOPENAI_BASE_URLtoo, where every run with a tool was refused. - A Bot's computer starts when its Screen is opened, as it already did for its Terminal, and
Activity names a computer turned off while idle and turned on again. Where OpenBot has no
computers for Bots, Admin's Computers, a Bot's Computer page, the Computer panel and saved logins
say so, where the panel waited minutes and then said it couldn't show the screen. Opening or
reading a conversation no longer starts a stopped computer or keeps an idle one running. - A Bot's shell and browser are steadier. A command with an unclosed quote comes straight back
instead of waiting out its whole timeout, a deep folder's prompt no longer shows as the last line
of a command's output, and Forward pressed right after Back is no longer refused as a page that
didn't load. - Activity on a Bot's computer reads plainly. Steps are named by their control, every row keeps
its time across a reload, and an allowed step says how it was allowed. A Bot's browser steps are
listed as before. - Folders granted to a Bot on OpenBot Desktop say what failed. A command that fails returns its
exit code and output, with a hint when the sandbox is the cause; a timeout or out-of-memory stop
says so; long output keeps its end; and when nothing ran it says what to do. An approval answered
after OpenBot stopped waiting runs nothing and says so, dialogs show one at a time, their time
starts when they show, and choosing a folder a Bot already has reuses its grant. An approval left
unanswered no longer ends the folder's grant, so later commands there still ask, where each came
back "Host grant was revoked.". - Connecting an app through Composio says who Composio is first. Before the sign-in for Gmail,
Google Calendar, Slack or another app from the directory, OpenBot shows a step that names the
app, says Composio is the service that holds OpenBot's app connections, and says the next page
will ask you to allow Composio, which is expected (for Slack, so is "App not approved by Slack").
It moves on after 5 seconds, or at once with Continue now, and Cancel stops it. On OpenBot
Desktop, the page your browser comes back to says "Gmail is connected" (or that it couldn't
connect) and counts down 5 seconds before handing you back to the window, where it went at once.
Arcade, apps that take a key and the apps OpenBot signs in to itself are unchanged. - An app call with a parameter the action doesn't take is refused before you're asked. The Bot
is told the parameters the action does take, so you approve the corrected call once, where a
Slack post sent with the wrong field could ask you twice. - Arrow keys in a dropdown on a Bot's computer count from the top. The computer's screen
pointer sat in the middle of the screen, and a list that opened under it could move the choice,
so two presses of the down arrow could pick the wrong option.
Pages and routines
-
Pages are grouped by the eight things a Bot has, and some have new names. A Bot has
Conversations, Routines, Memory, Skills, Apps, a Computer, Approvals and Activity
(docs/how-openbot-works.md), and a Bot's page, the main menu and Settings, and Admin now
list them in that order. What moved:- Vault, under Settings, holds your saved logins and your keys. It replaces Saved logins,
Passwords and Credentials. - Routines, Memory and Skills are in the main menu beside Bots and Connections,
because they are things you make and come back to. Teach a Bot is no longer in the main menu:
open it from the Skills page. - Connections has two sections. Your apps lists what your Bots use. Add an app lists
everything else, the apps that need a key included, and ends with Browse apps. The count and
Show apps that need setup are gone, and so is Routines, which every Bot already has. - Skills holds the skills you wrote, the skills learned from your recordings and the
ones waiting for your review. It replaces Capabilities > Skills and Settings > Learning. - Settings > Computer is one page. Browsers and This computer are sections of it.
- Routines is one list of routines, goals and check-ins, for one Bot and for every Bot.
- In Admin, Plugins is now Apps, Audit is Activity, Credentials is Vault, and
Check-ins are under Routines. The Team group is now Organization, with People and
Teams. Settings has Organization, and API keys under You.
Nothing was removed, and every old address opens the page that replaced it.
- Vault, under Settings, holds your saved logins and your keys. It replaces Saved logins,
-
One New routine screen for routines, goals and check-ins. New routine no longer opens a menu
of Routine, Goal and Check-in. One screen, in sections, asks what to do, whether the Bot should
act for you or only tell you, what starts it (a schedule, an event such as a GitHub issue
opening, or now, once) and, optionally, when it is done. A schedule with nothing in Done when
makes a routine; an event, or anything with Done when, makes a goal, shown as one Goal row with
its schedule under Starts; Only tell you makes a check-in. A goal can now be saved with no
outcome. -
A goal can end when its outcome is met. A goal made with an outcome is completed by the run
that meets it, and completing a goal switches off the routines that run it. A goal that repeats
with no outcome is still ended only by you. -
Every routine, goal and check-in opens on a page of its own. A row on Routines is a link to
/routines/goals/<id>,/routines/schedules/<id>or/routines/check-ins/<id>, which can be
opened in a new tab, and the page's back link returns to the list it came from. Each page holds
what its dialog held, and Delete returns to Routines. A goal can be deleted, and deleting it also
deletes every routine that runs it, unless one still runs another open goal or only ran it on
some of its firings (below). -
Routines say what starts a goal in words, such as "When a GitHub check finishes", and keep
the exact event under them on the goal's page. Check-ins say "Next check in 24 hours" rather
than a date and time. -
A Bot's Routines row counts what the Routines page lists: a goal and the routine that runs it
once, and the check-ins that are on. -
Completing or deleting a goal does the right thing to its routines, whatever the goal is called
now. A routine made for a goal is switched off when the goal is completed, even after you
renamed the goal. A broad routine that only sometimes runs a goal ("Summarize my inbox, and if a
release shipped run the goal") keeps running when that goal is completed or deleted, and loses
only the link. A routine is never linked to a goal that is already complete or paused. A routine
made before 0.3.3 runs a goal only where its instruction clearly does, not wherever it quotes the
title. And a Bot is told never to make work you want every time end when it is met. -
Fixes to Routines and New routine.
- A goal made to start Now, once with a Done when ends once that is met.
- Routines are read again every 10 seconds, as goals already were, so a goal a run completed no
longer shows its schedule on until the page is opened again. - New routine no longer takes over a check-in the Bot already has: Create says so and nothing
changes. Accepting a Bot's offer still turns a matching check-in back on. - Create routine stays enabled while a field has a message. Pressing it moves to the first
field with one, and a screen reader reads each message with its field. - A routine that runs a goal, listed on its own, says "Runs a goal" where it showed the
instruction written for the Bot. - When your browser's time zone isn't one OpenBot knows, New routine says to ask your Bot to set
the routine up. - The Routines page no longer warns that routines aren't running: that was for a deployment with
no worker, which is no longer a state.
Finding things
- Cmd/Ctrl+K opens a command palette anywhere signed in. It finds pages, Bots and
conversations and opens them; it can't send, approve, run or delete anything, and it never lists
a page you can't open. In a Bot's browser the shortcut goes to the browser. - Searching conversations searches all your history on the server, by name and recent
messages, instead of only the conversations already loaded in the sidebar, with the
conversation list's own rules for who can see what.
Personal edition
- Only requests that carry the install's token are let in. OpenBot Desktop makes the token
once per install and keeps it in the system's credential store, never in.env, and it stays the
same across restarts, upgrades and a put-back to an earlier version. Compose hands it to the
server alone. The window enters once through an address that leaves it a cookie, and a program of
yours sendsX-OpenBot-Install-Token. A request without it, a Bot's own computer included, is
refused, and a browser refused this way shows "Open OpenBot from its app".scripts/start.sh
makes the token, writes it to.envand prints the entry address as its Ready line, and both
charts' single-user mode make one and keep it across upgrades. An app sign-in that ends in your
browser hands back to the window, which finishes it. The team edition is unchanged. - API keys work in the personal edition. Settings > API keys lists, creates and deletes your
keys, where it said "Couldn't load your API keys". A key works on/api/v2only, is shown once,
can't make a key, and needs no install token. - Admin is in the personal edition too, under your name. Its pages open where they used to
send you to Settings. Providers, model usage and check-in history show your own Bots. Pages and
controls that only apply to a team say "This applies when OpenBot runs for a team." with a way to
OpenBot for Business. - Connections lists the search providers, GitHub and Composio's apps. The built-in Google Drive
and Notion are no longer offered, because Composio carries both. One already added stays listed
and keeps working. - New routine doesn't offer Slack as an event source: the personal edition has no Slack, so a
goal it started never ran. - A personal owner is never told to ask an admin. Stopped routines say to restart OpenBot, a
Connections page that couldn't load offers Retry and Setup, and a Bot's Apps page with no app
connected says "Connect one on Connections." with a link. - OpenBot for Business opens in the app. Meet OpenBot for Business, Book a call and the link
about a computer for each Bot open a dialog with the booking calendar, instead of a website tab.
If the calendar can't load, Book a demo opens the page in your browser.
OpenBot Desktop
- A first install no longer stops at "Connecting OpenBot" when your home folder is a git
repository. On a Mac whose home folder is a repository that ignores everything, a common way to
keep dotfiles, setup stopped with "Setup stopped partway through with an error", and Retry failed
the same way every time: git looked up from OpenBot's folder into the home repository, whose
rules hid OpenBot's own files. Setup now keeps git to OpenBot's folder. A repository you keep in
OpenBot's folder itself is still read. - A setup step that stops offers its troubleshooting logs. When Finishing setup or Starting
OpenBot stops with an error, Copy logs and Save to a file are under Retry, where the card said to
copy the logs below and showed none, and the logs now say why the connect step stopped. - OpenBot Desktop can connect to an OpenBot that runs somewhere else. At set-up choose
OpenBot for Business, or later Settings > Computer > Connect to an OpenBot, which stops
OpenBot on this computer first. Type the https address of your company's OpenBot. Nothing is
installed or downloaded, and no CopilotKit account, license or model key is asked for: you sign
in in your browser, and the app keeps the sign-in in Keychain or Windows Credential Manager,
never in a file. While connected, Start, Stop, Status, model keys and local folders for Bots are
not shown, Quit only closes the app, and you can switch organizations. Settings shows where you
are connected, with Sign out and Use OpenBot on this computer instead, which loses nothing.
It needs an OpenBot of this release or later, and refuses one set up for one person on their own
computer. Setup's big downloads now start when you press Install, not when the first screen
opens. - OpenBot Desktop is one address. With CopilotKit Intelligence on the same machine, the app
used to open a second connection, straight to Intelligence's gateway on a port of its own. It now
opens everything on OpenBot's own address: the server carries that socket
(INTELLIGENCE_GATEWAY_PROXY, whichdocker-compose.ymlturns on for a gateway on this
machine). Nothing is moved or rewritten in an installation; the change takes effect at the next
Start. An Intelligence somewhere else, managed Intelligence among them, is reached at its own
address as before. OpenBot run from a checkout (scripts/start.sh) does the same only when
OPENBOT_PUBLIC_URLis set, and otherwise keeps the gateway's own port. - The first setup screen says what Self-hosted needs: a Mac or a 64-bit Windows PC, 16 GB of
memory, 4 CPU cores and 40 GB of free disk. - Setup waits for room on a full drive instead of stopping. It pauses with how much is free and
Try again, checks every 10 seconds, and carries on by itself once there is room. - A Podman machine too small for setup is resized for you on a Mac. Setup needs 4 CPUs and
4 GB of free memory. The row now says what your machine lacks ("Your Podman machine has 2 CPUs.
OpenBot needs 4."), and Resize my Podman stops your machine, raises what is short (at least
6 CPUs or 8 GB, never more than your Mac has), starts it again with the containers that were
running, and carries on. If the resized machine can't download yet, it says to check your
internet connection and Retry. Where OpenBot can't resize it, the fix is numbered steps, each
command with Copy. Setup's messages count CPUs in all, as setup does. - A 4-CPU Docker or Podman no longer fails setup because something else is running in it. The
CLI counts the engine's CPUs rather than those left free. Memory and disk are still measured
free. - A first setup that stops in its image download carries on by itself once, as Retry would.
- An OpenBot whose saved data is gone says so, and starts fresh. When OpenBot's Podman machine
was removed (podman machine reset, a WSL unregister, or an older version uninstalled with its
data), every start stopped at "Setup stopped partway through with an error", and Retry did the
same. OpenBot now says plainly that its saved data on this computer is gone, with your Bots and
conversations, and offers Start fresh. Your sign-in and AI connection stay. - OpenBot starts and stops more reliably.
- Retry and the next Start run an unfinished back end update again when it left OpenBot running
but not working. - OpenBot starts again by itself after an update; only your own Stop keeps it stopped.
- Retry starts OpenBot's engine when it stopped under a running OpenBot, where only Quit and a
relaunch helped. - A Podman machine that says it is running but doesn't answer is stopped and started again.
- A back end that refuses to start straight after a Quit is asked again twice before setup
fails. - When the engine stops under the open window, the window shows OpenBot as stopped, with Start.
- Quit and Stop treat a stopped Podman machine as nothing left to stop, and the last run's
shutdown notice clears once Start runs. - After an update, the previous release's images are removed, freeing about 3 GB.
- Retry and the next Start run an unfinished back end update again when it left OpenBot running
- Your default Podman connection is kept whenever OpenBot makes, starts or removes its own
machine. Removing it could move your default to another of your machines. - Windows.
- Uninstalling with its data no longer breaks the next install: the uninstaller stops OpenBot's
containers first, and an installation already stuck at "Opening OpenBot" is repaired at Start,
by clearing only the port rules a removed OpenBot machine left behind. - A lone Alt or F10 no longer freezes the window until a click.
- The title bar is dark from the first frame in dark mode, and follows every theme change.
- After an update OpenBot opens in front, or flashes in the taskbar where Windows won't allow
that. - While WebView2 installs, the installer says so in plain words, and after Finish the window
shows "Getting OpenBot ready" rather than a blank white window. - A fresh install no longer waits about two minutes on "Output folder".
- Uninstalling with its data no longer breaks the next install: the uninstaller stops OpenBot's
- Mac.
- Get Troubleshooting Logs... no longer quits the app after saving the file. While the logs are
gathered, the menu says Gathering Logs.... - The window opens on OpenBot's own background, with no grey and white flash.
- The window's buttons no longer cover a skin's controls, and a press anywhere along a skin's
top bar moves the window.
- Get Troubleshooting Logs... no longer quits the app after saving the file. While the logs are
- A maximized or full-screen window reopens as it was, and each Bot has one color everywhere.
- OpenBot Desktop 1.0.2 and 1.0.3 can start an installation a newer version updated. For one
release,container-images.jsonnamesagent-botandagent-langgraphagain, both as the
server image, anddocker-compose.ymldefines two services of those names that nothing starts.
Those, theworkerandmigrateservices below, the two commands they run in the server image
and theroutine_sweepstable are all dropped in the release after. - CopilotKit Intelligence 0.16.0-rc.13. The bundled CLI moves to
copilotkit@7334f7d, so
OpenBot Desktop's back end runs Intelligence 0.16.0-rc.13. - A second OpenBot Desktop can run beside an installed one, for testing, with
OPENBOT_DESKTOP_PROJECTset to another Compose project name.
Running OpenBot
- The server migrates and does the scheduled work itself. It applies its own migrations when it
starts, and fires routines, cleans up unsent attachments and suspends idle computers in its own
process. Nothing starts themigrateorworkerCompose service, nothing waits onmigrate,
and the server is no longer handedWORKER_SHARED_SECRET.docker-compose.ymlstill defines
both services for this one release, doing nothing, behind itsx-openbot-scheduled-work: server
line. OpenBot Desktop,bash scripts/start.shandbash scripts/stop.shremove theworkerand
migratecontainers an earlier release left, and run an earlier release's file as it requires.
bun scripts/update.ts local applyupdates to and from a release of either kind, and
bun scripts/update.ts kubernetesaccepts a chart whose server applies its own migrations.
With several replicas, one leads and decides which routines are due, and every replica runs
routine turns; each server keeps two PostgreSQL connections of its own for this, beside its pool.
STAGED_ATTACHMENT_HOURSsets how long a file attached but never sent is kept: 24 hours unless
set, and it must be above 0. A computer suspended, used again and left idle is suspended at the
next sweep, and several replicas starting together on a new database all come up, where one
could stop on a duplicate key. - One image runs the server and every Bot's computer. A release publishes
ghcr.io/copilotkit/openbotandghcr.io/copilotkit/openbot-supervisor, for amd64 and arm64.
The server and the computer run from the oneopenbotimage, each with its own command
(/app/bin/openbot server,/app/bin/openbot-computer).SERVER_IMAGEandCOMPUTER_IMAGE
still name the image for each service. The computer's container is given none of the server's
settings. The first update downloads the whole image once, about 715 MB compressed on amd64 and
740 MB on arm64. The images are signed, anddocs/cosign.pubis the key to check them with
(docs/deployment.md). - The server checks its license and its Intelligence before it starts. With
COPILOTKIT_LICENSE_TOKENset, the license must be valid and current, and the key is not tried
instead. With none set, the server asks the Intelligence atINTELLIGENCE_API_URLwhether it
accepts the project key, and stops when the key is missing or refused or that address never
answers. It also refuses an Intelligence that lacks an operation it calls, naming the oldest
version that has them all (0.16.0-rc.12). It asks again for up to ten minutes while Intelligence
is starting, and each refusal is one sentence that names the variable to fix. A demo license,
the kind the CopilotKit CLI issues, is refused outside the CLI's own local cluster, as
Intelligence refuses it, and only CopilotKit's own license keys are trusted. - Running from a checkout pulls the published images.
bash scripts/start.shpulls the server,
computer and supervisor images at the tag inbackend-images.envon every start, falling back to
local copies when it can't, and runs them as Compose services, with the app served from the
checkout.OPENBOT_BACKEND_TAGin the environment or.envruns another build. The server's log
isdocker compose logs server, and a server and worker an earlier start left on the host are
stopped first. The repository no longer builds the images or holds the Helm chart. The personal
edition's browser shows in the page, as in OpenBot Desktop, andOPENBOT_PERSON_BROWSER,
BROWSER_PORTand the helper that opened your own Chrome are gone..env.examplenow points at
an Intelligence on this machine, with managed Intelligence as a commented option. - A recording made in a Bot's browser saves when OpenBot runs from a checkout.
bash scripts/start.shsetsCOMPUTER_LOOPBACK_RELAYS, as OpenBot Desktop does, so the
Recorder in a Bot's browser reaches OpenBot. A value in the environment or in.envwins. - Two checkouts on one machine no longer share the managed Bot's image. It is tagged
<project>-agent-langgraph-agui:latest, with the stack's Compose project name. A checkout in a
folder namedopenbotkeeps the tag it had.LANGGRAPH_IMAGEstill names the image outright. - SPIRE is gone. It issued each Bot's computer a workload identity that nothing in OpenBot
checked. Its three services (spire-init,spire-server,spire-agent), their volumes and the
supervisor'sSPIRE_*settings are removed; volumes an earlier stack made stay until you remove
them. - The managed Bot is the LangGraph harness,
agent-langgraph-agui.agent-botand the
TypeScriptagent-langgraphare retired, withBOT_PORT,BOT_IMAGEandAGENT_BOT_MODEL, and
their images are no longer published.scripts/start.shstarts the harness on port 4201 and
pointsMANAGED_AGENT_AG_UI_URLat it, whichever providerBOT_PROVIDERnames. OpenBot Desktop
takes the retired Bots' settings out of an updated installation's.env. - Usage data says which release it came from, and carries error reports. With usage data on,
every event names the OpenBot version and edition. The app's error screen, errors a page didn't
handle, and the server's uncaught errors and 5xx answers are sent as error reports, through the
server, with secrets, emails, addresses, paths and long numbers taken out and their size and
number capped. OpenBot Desktop reports when an update starts and how it ends, and the first
launch of a new version reports the update that brought it, once. With usage data off none of it
is sent.docs/TELEMETRY.mdlists every field.
Kubernetes
- The
openbotchart (0.3.0) is OpenBot alone, andopenbot-bundle(0.1.0) is the all-in-one
install.openbotno longer installs CopilotKit Intelligence: it points OpenBot at one that
already runs, managed or self-hosted, by its two addresses and a project key, so several OpenBot
releases can share one Intelligence.openbot-bundleinstalls OpenBot and Intelligence together
on one PostgreSQL server, with Intelligence 0.16.0-rc.13. Theopenbotchart can also make a
project and key of its own in an existing Intelligence organization, given that organization's
service token (config.intelligence.project.create). A release attaches both charts, with
release-charts.jsonnaming them and their checksums. - One address in a cluster too. With
httpRouteoringresson and Intelligence in the same
cluster, theopenbotchart routes/client/websocketon OpenBot's host straight to
Intelligence's gateway, and OpenBot's server stops carrying that socket. An existing release
changes over athelm upgrade. Setconfig.intelligence.inCluster.edgeRoute: falsefor a
gateway that won't upgrade a WebSocket to Intelligence's Service. An Ingress with Intelligence
in another namespace keeps the server carrying it. - A release on the 0.2 chart with its bundled Intelligence is refused at upgrade to 0.3, and
nothing in the cluster changes, because the upgrade would delete that Intelligence and its
conversations. Keep the release on 0.2, or installopenbot-bundleas a new release. A 0.2
release that already pointed at an Intelligence of its own upgrades as before. - No CronJobs and no migration Job. The server fires routines, suspends idle computers,
deletes unsent attachments and applies its own migrations.migrations.enabled: falseis
refused, with what to do instead.attachments.culler.olderThanHoursreaches the server as
STAGED_ATTACHMENT_HOURS, andserver.waitForDatabaseSecondsbounds how long a server pod waits
for its database. The values that configured the removed jobs (routines,
computers.sandbox.culler,secrets.workerSharedSecret, andattachments.culler.enabled,
.scheduleand.activeDeadlineSeconds) are ignored.
Appearance
- OpenBot follows your system's light or dark setting. Settings > General has a Theme row with
System, Light and Dark, in place of the Dark theme switch, and everybody starts on System. The app
switches when the system does, with no reload, and OpenBot Desktop's window follows it.
OpenBot runs on CopilotKit 1.88.0
@copilotkit/react-core in the app and @copilotkit/runtime in OpenBot's server move from 1.85.0
to 1.88.0. AG-UI stays 1.0.2.
Smaller changes
- In the OpenDots skin, the badge on Bots sits on the corner of the Bots icon, and a count is a
smaller circle. - New Bot's first screen ends with a Bring your own agent card showing the mark of every
framework a custom agent can run on, in place of the Connect a custom agent button. - The Bots page lists the Bots you can use first and leaves out empty sections. Lists draw their
rows' shape while they load, and Needs you says which part didn't load, with Retry. The message
box shows no plus when there is nothing to attach. A Bot's Skills page no longer waits for its
apps. - Connecting an app that takes a key says "OpenBot connects to Fireflies with a key from your
Fireflies account, not a sign-in.", where it said the app had no sign-in. - Delete dialogs name what they delete, and a personal conversation's no longer says "for everyone
in the conversation". Key-check banners no longer show an app's raw JSON error. - Admin > Providers > Usage says "Model not reported" and how a model is known, where it showed
remote:<id> · unknown. - The masked secret box tells the Bot a secret was entered only once the computer confirms it.
- Record stays on screen beside a Bot's full-size screen in a narrow window.
- After an app's consent screen, the address no longer carries the broker's report.
- After you delete a routine, goal or Bot, or disconnect an app, the page stops asking for it.
- An unticked tickbox's edge stands out at 3:1 or better, in light and dark.
- A task that couldn't finish says why once, and offers Try again only where it could help. A
dismissed app request says so, and the Bots badge updates on Allow and Deny. - A Learning proposal's "Supported by" line is left out whole, whatever case its aliases are in.
- Setup's usage data checkbox stays where it is while the step's rows arrive.
- Check again on Setup asks about a model key added a moment ago, where within ten seconds of the
last check it still said the key hadn't been checked.