Repository navigation
Before upgrading. Nothing in a Helm values file, a Compose override or a script has to change,
unless you rely on task threads.
- No migrations. The database stays at
0137_ai_connections, where 1.1.1 left it, so 1.1.1
upgrades to 1.1.2 with its servers rolling one by one, with no downtime, and goes back to 1.1.1
the same way. Upgrading from 1.1.0 or older runs 1.1.1's migrations as well: read 1.1.1's Before
upgrading first. - Task threads are off unless you turn them on. No Bot hands a person's request to a task
thread any more: every Bot does the work in the conversation, including one set to hand work to
agents. SetOPENBOT_TASK_THREADS=onon the server (alsotrueor1) to bring them back as
they were. Nothing is removed: a Bot keeps what it is set to, and task cards and threads already
in conversations still show and can still be followed up or stopped. - Deprecated: nothing.
Bots and models
- Bots do the work in the conversation. With task threads off (above), no Bot is offered a task
thread to hand work to, and a Bot set to hand work to agents runs on its default agent with its
apps, as a Bot that does the work itself does. WithOPENBOT_TASK_THREADS=on, only a Bot set to
hand work to agents does so: since 1.1.0 every Bot, the General and Personal Assistants included,
handed requests in a person's own conversation to task threads even with Hand work to agents off. - A Bot's Agents page has no Hand work to agents switch. Every Bot shows the Runs on card, a Bot
set to hand work out included, with its default agent chosen. Saving the page keeps what the Bot
is set to, its model, effort, instructions and agents too, so turning task threads back on brings
it back as it was. A Bot's profile says what a Bot that does the work itself says, such as "Does
the work itself on OpenBot", and a conversation with such a Bot takes a model pick like any other. - A Bot asks for the app a request needs instead of searching around it. A Bot is told that what
is private to the person or their team in an app it has no tools for, such as the team's own
GitHub repositories, starts with that app's card, which connects it or asks an admin for it, and
not with the public web, which shows only what is public. A page of theirs that comes back not
found on the web and could be private, such as a private repository's 404, gets the same card
instead of "it may not exist", and so does the person saying something is private. Anything
public, such as an open-source project's release or a dead public link, is still answered from
the web, and a routine, which nobody is watching, says which app needs connecting instead of
showing a card. - Whoever manages a Bot that runs at an address of its own can change its Agent URL on Setup.
The row opens a dialog that checks the address's shape before saving it, and shows the server's
refusal under the field; saving the same address sends nothing. For anyone who can't manage the
Bot it stays read-only, and a built-in Bot is unchanged. Other Setup edits no longer send the
stored address, so a Bot saved at an address the deployment's rules have since closed can still
be renamed. - A Bot asked to watch Slack says what works. Without a Slack pairing, which the personal edition
never has, no Slack event reaches OpenBot. A Bot asked to act on Slack messages was sent to a
Reaching you page that edition does not show; it is now refused before anything else, and the Bot
runs the goal from a routine that checks the person's connected Slack app instead, with Slack's
connect card in the conversation when it is not connected. Where a pairing exists, the owner's
Slack account linked under Connections counts as well as a Reaching you binding: only the owner,
only channels that account is in, checked again on every event.
OpenBot Desktop
- Approved folders survive a server restart. A restart of OpenBot's server, Restart in the Status
window among them, dropped every folder approved in Settings > Computer: Settings read "No folders
approved" and a Bot was told its folder access had ended. OpenBot Desktop now holds the folders
for up to two minutes while the server is away, ending only the work that was running, and hands
them back when it answers. The server takes back only the folders OpenBot Desktop held for it, and
never one revoked or stopped in the meantime. Stop and Quit still end them, as does a server that
stays away for more than two minutes. - Install sets OpenBot up again over an earlier install whose engine is gone. On Windows, with
Podman removed after an earlier install, Install took the old install as done and the window
stayed on "Setting up this computer" after sign-in. Install now always checks: it brings the
engine back first, and skips the download when the install is still whole. - On Windows, setup no longer waits for ever on "Checking what OpenBot needs to download". A
command OpenBot runs in its engine's machine now stops after 60 seconds. If Windows refused the
machine's key because other accounts can open it, setup says so and names the file and the
account. - On Windows, the engine's disk goes in the install folder. With the install folder on F:,
the engine's 40 GB disk goes on F: too, and the space check measures F:. A full C: no longer
stops the install. A drive that cannot hold the disk (a network share, a removable drive, exFAT)
keeps it on C:, and Install says so. An engine already on C: stays there. When C: is short and
F: has room, OpenBot offers Move OpenBot's storage to F:, which stops every WSL distro for
a moment, Docker Desktop included. - On Windows with only the WSL that ships with Windows, setup offers the WSL update. It stopped
on "could not check the WSL version", because that WSL does not knowwsl --version. Setup now
shows the WSL update screen, whose fix runswsl --update. - Stop and Quit stop OpenBot's own Podman machine. They left it running, so on Windows its WSL
distro kept holding memory after Stop or Quit. Both now stop it last, once everything of OpenBot's
in it has stopped, and never touch your own machines; Start brings it back. An update's stop
leaves it running for the restart that follows, and Quit shows the step on the Shutting down
screen. - On Windows, uninstalling with "Delete the application data" deletes OpenBot's downloads too.
The installers, tools and machine image OpenBot downloads (about 128 MB, in~\OpenBot) stayed
in the profile whichever folder OpenBot was installed in. They now go with the data, and the
folder with them once nothing is left in it. Without the data deleted, they stay, as the machine
does.
Kubernetes
- OpenBot can be kept up to date for you, if you ask. The
openbot-agentchart, published with
each release atoci://ghcr.io/copilotkit/charts/openbot-agent, installs a management agent into
openbot-system, beside an OpenBot install, with Flux's source and Helm controllers limited to
that namespace. Nothing is managed without it: theopenbotandopenbot-bundlecharts never
reference it, and an install without it is unchanged. To opt in, CopilotKit gives you an install
id and a client certificate. Put the certificate in a Secret and your product values in another,
choose a weekly window of at least four hours (UTC) and how your database is backed up, and
helm installthe chart.- The agent only calls out: every minute, over HTTPS with that certificate, to
https://fleet.bot.copilotkit.ai. A heartbeat carries only the fields in
shared/management-contract.ts: the install id, versions and image digests, the last
migration, Flux's readiness, ready and wanted pods per workload, and the upgrade's state. A
metrics collector sends a fixed list of health metrics with nothing that names a person, an
organization or a tenant;otel.enabled: falseturns it off. - It upgrades only to a release the management service names, only inside both your window and
the service's, only along the release's upgrade paths, by pinning Flux to that release's exact
chart digest. It creates and keeps the product's Flux objects from the chart's settings, so
helm install --waitfinishes under Helm 4. Before an upgrade that changes the database it
stops writes and takes a backup, and goes on only once the backup succeeds; it refuses the
upgrade when the backup it would take is of a different database from the one the install
uses. A failed backup puts the install back as it was, and the same release is not tried again
for 30 minutes. It never rolls back by itself: an upgrade that stops after the database may
have changed waits for a person. - It runs five commands, each signed and short-lived: restart, rollback (only where the release
says the previous one runs on its database), pause, resume, and a support bundle that is
redacted and written to a volume in your cluster, never uploaded. - It installs the product as
openbot-deployer, an admin of the product namespace only, which
also holds the HTTPRoutes, ExternalSecrets and Sandboxes the product charts create there. It
turns on the server's internal status port for the release it manages, open to
openbot-system, and keeps your ownserver.internalStatus.allowFrompeers. - Pause it with
helm upgrade openbot-agent ... --set paused=true, which works under Helm 4 on an
install the agent has already upgraded.helm uninstall openbot-agentturns it off and leaves
OpenBot installed for a plainhelm upgrade. - Its image is the release's own,
ghcr.io/copilotkit/openbot-management-agent:v<release>. The
chart published with 1.1.1 named a tag that was never published.
- The agent only calls out: every minute, over HTTPS with that certificate, to
bun scripts/update.ts kubernetes applyaccepts a published release. It compared the commit
the running server reports with the one the release names, which never match for a published
release, so an upgrade that had worked ended in recovery-required with "Running server release
identity does not match the candidate", and a release that stops writes for its upgrade was
scaled back to zero. It now checks the version, distribution, repository and channel, with the
image digest the server Deployment is pinned to, as before.
Smaller changes
- The API refuses a notification route that names the same way of reaching a person twice, such as
["push", "push"], with 400VALIDATION_ERROR, and changes nothing, as the spec's unique lists
say. The web app never sends one.