v1.1.22026-11-20
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 0137aiconnections, 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. Set OPENBOTTASKTHREADS=on on the server (also true or 1) 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. With OPENBOTTASKTHREADS=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 know wsl --version. Setup now
shows the WSL update screen, whose fix runs wsl --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-agent chart, published with
each release at oci://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: the openbot and openbot-bundle charts 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 install the 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
v1.1.12026-11-19
Before upgrading. Read these first: several change what a Helm values file, a Compose override
or a script has to say.
- The Helm charts refuse shared and external computers until you say you mean them. Shared
computers (computers.mode: shared, still the default) and external ones
(computers.mode: external, the one computer at computers.url) run every person's Bots on one
computer, where a command one Bot runs can read another person's files, so they suit a proof of
concept only. Sandbox mode (computers.mode: sandbox) gives each person a computer of their own
and is the way to run OpenBot for people; it needs the agent-sandbox controller in the cluster.
A release on shared or external computers, including one that never set computers.mode, has to
set computers.sharedModeAcknowledged: true before upgrading
(openbot.computers.sharedModeAcknowledged in the openbot-bundle chart), or helm install and
helm upgrade stop with a message naming both choices and change nothing. That includes an
upgrade with --reuse-values, and the single-user mode of either chart. Acknowledged, the
install notes say it again, and admins see it in Admin (below).
- The openbot-bundle chart keeps files in its own object store. A bundle upgraded with your
own values file moves every file people and Bots attached out of PostgreSQL and into the bucket
openbot-files in the bundled Intelligence's local object storage, which a new Job makes on
every install and upgrade. That storage is one pod, with no replication or backup of its own. To
keep files in PostgreSQL, set openbot.attachments.store: postgres and
openbot.objectStorage.enabled: false. An upgrade with --reuse-values keeps the values it had,
so files stay in PostgreSQL. Don't turn off intelligence.localObjectStorage while
openbot.objectStorage points at it: the bundle refuses that. See Files (below).
- Migrations run at start: five, all additive. The database moves from 0132bottask_runs to
the migration release-policy.json names for this release. 0133filesinanobject_store adds
the file_deletions table and lets an attachment's bytes live outside PostgreSQL,
0134attachmentsstoragekeyindex indexes where they live, 0135filesgate_opened records
when files may move, 0136apideprecated_use counts use of deprecated API, and
0137aiconnections keeps the AI connection an admin chose (below). A 1.1.0
server works against the new schema, so 1.1.0 upgrades to 1.1.1 with its servers rolling one by
one, with no downtime. 0133 needs OpenBot's login to hold the openbot_system role's
privileges, as 1.1.0 set up, and stops with the GRANT to run when it doesn't. It waits at most
five seconds for a lock on attachments, so a long transaction can fail the first start, which
then tries again. Upgrading from 1.0.x or 0.3.x runs 1.1.0's migrations as well: read 1.1.0's
Before upgrading first.
- Going back to 1.1.0 after files moved to an object store needs them moved back first: set
attachments.store: postgres with the store still configured, and wait for the move to finish.
A 1.1.0 server cannot read a file that is in the store.
- New Bots run on the LangGraph (TypeScript) engine where a deployment runs it. OpenBot
Desktop runs it, and an installation's Bot that ran on the Python LangGraph engine moves to the
TypeScript one at its next Install and Start, keeping its conversations. Docker Compose now
defines agent-langgraph-ts (port 4215, on this machine only), which a plain docker compose up
builds from the checkout unless LANGGRAPHTSIMAGE and IMAGEPULLPOLICY=missing name the
published image. A server without LANGGRAPHTSAGUIURL, and both Helm charts, keep new Bots
on the server's built-in engine. scripts/start.sh starts it on 4215 and sets
LANGGRAPHTSAGUIURL to it, except in the personal edition (LANGGRAPHTSPORT moves it).
OPENBOTDEFAULTBOT_ENGINE=built-in keeps every new Bot there.
- Compose: Claude and ChatGPT plans run in services of their own. agent-harness is no longer
handed CLAUDECODEOAUTHTOKEN, CHATGPTAUTH_FILE or the ./.langchain folder. They go to
two new services, agent-claude-plan and agent-chatgpt-plan, behind the claude-plan and
chatgpt-plan profiles, which OpenBot Desktop starts when a plan needs them. A Compose override
of your own that ran a plan through agent-harness turns on the profile and sets
CLAUDEPLANHARNESSURL or CHATGPTPLANHARNESSURL. agent-langgraph-agui reads the ChatGPT
plan's ./.langchain folder too, and .env.example points CHATGPTPLANHARNESS_URL at it.
- The model picker offers only models proven on the Bot's engine (below). A conversation or a
Bot set to a model that isn't proven runs on the default, and says so.
- A model gateway has to serve /responses for gpt-5.4 and gpt-5.5. Bots behind an
OpenAI-compatible address now call those models, and their mini, nano and dated versions, on the
Responses API, because they refuse tools with a reasoning effort on Chat Completions.
COPILOTKITOPENAIAPI=chat keeps them on Chat Completions, for the built-in Bots and the AG2
and Langroid harnesses.
- A server that refuses to start exits with code 78, where it exited with 1. The all-in-one
image now ends its container instead of trying again forever, and on Kubernetes the pod's
CrashLoopBackOff names the reason in kubectl describe. A crash still exits with 1. Monitoring
that watches the exit code may need to know the new one.
- Settings that now stop a server from starting when they are wrong: OPENBOTMINIMUMCLIENTS,
OPENBOTDEFAULTBOTENGINE, SHUTDOWNKEEPSERVINGSECONDS, SHUTDOWNDRAINSECONDS,
RELEASESTATUSPORT, and the file store's (below). An S3BUCKET, GCSBUCKET or
AZURESTORAGECONTAINER already in the server's environment, for example one it shares with
Intelligence, configures a file store even with FILES_STORE unset, and the server then checks
that store at start and refuses to start when it can't use it.
- Kubernetes: a server pod takes up to 45 seconds to stop, so it can finish what it is doing
(below). server.shutdown sets it.
- The openbot chart needs this release's image. Its wait for the database now runs from the
server's own image (below), so the new chart with an older image.tag doesn't start.
- Deprecated: nothing.
Files
- Files can live in an object store. Amazon S3 or anything that speaks its API, Google Cloud
Storage or Azure Blob Storage, where they were always in PostgreSQL. With a store, a file can be
up to 100 MB. Without one nothing changes: files stay in PostgreSQL, with 8 MB for pictures and
PDFs and 1 MB for other files, and OpenBot Desktop stays on PostgreSQL. The openbot chart gains
attachments.store (postgres unless set) and objectStorage: enabled, provider (s3,
gcs or azure), bucket, prefix (openbot/), region, endpoint, forcePathStyle,
azure.accountUrl, deleteAfterDays (7), assumeAllServersCurrent, existingSecret with its
secretKeys, and inCluster for a store in the cluster. A server pod can sign in to the store
with workload identity through serviceAccount.annotations. On Docker Compose the same settings
are FILESSTORE, FILESPREFIX, FILESDELETEAFTERDAYS, FILESASSUMEALLSERVERS_CURRENT,
S3, GCSBUCKET and AZURESTORAGE in .env. The chart refuses a store asked for and
not set up, a Secret without its keys, and a store in the cluster behind a network policy with no
way to reach it, each by name.
- The server writes, reads and deletes a test file at start, and refuses to start when the store
doesn't answer. It never falls back to PostgreSQL.
- Files already in PostgreSQL move to the store in the background, and back again when
v1.1.1-rc.1预发布2026-11-19
Before upgrading. Read these first: several change what a Helm values file, a Compose override
or a script has to say.
- The Helm charts refuse shared and external computers until you say you mean them. Shared
computers (computers.mode: shared, still the default) and external ones
(computers.mode: external, the one computer at computers.url) run every person's Bots on one
computer, where a command one Bot runs can read another person's files, so they suit a proof of
concept only. Sandbox mode (computers.mode: sandbox) gives each person a computer of their own
and is the way to run OpenBot for people; it needs the agent-sandbox controller in the cluster.
A release on shared or external computers, including one that never set computers.mode, has to
set computers.sharedModeAcknowledged: true before upgrading
(openbot.computers.sharedModeAcknowledged in the openbot-bundle chart), or helm install and
helm upgrade stop with a message naming both choices and change nothing. That includes an
upgrade with --reuse-values, and the single-user mode of either chart. Acknowledged, the
install notes say it again, and admins see it in Admin (below).
- The openbot-bundle chart keeps files in its own object store. A bundle upgraded with your
own values file moves every file people and Bots attached out of PostgreSQL and into the bucket
openbot-files in the bundled Intelligence's local object storage, which a new Job makes on
every install and upgrade. That storage is one pod, with no replication or backup of its own. To
keep files in PostgreSQL, set openbot.attachments.store: postgres and
openbot.objectStorage.enabled: false. An upgrade with --reuse-values keeps the values it had,
so files stay in PostgreSQL. Don't turn off intelligence.localObjectStorage while files are in
it: the bundle refuses that. See Files (below).
- Migrations run at start: five, all additive. The database moves from 0132bottask_runs to
the migration release-policy.json names for this release. 0133filesinanobject_store adds
the file_deletions table and lets an attachment's bytes live outside PostgreSQL,
0134attachmentsstoragekeyindex indexes where they live, 0135filesgate_opened records
when files may move, 0136apideprecated_use counts use of deprecated API, and
0137aiconnections keeps the AI connection an admin chose (below). A 1.1.0
server works against the new schema, so 1.1.0 upgrades to 1.1.1 with its servers rolling one by
one, with no downtime. 0133 needs OpenBot's login to hold the openbot_system role's
privileges, as 1.1.0 set up, and stops with the GRANT to run when it doesn't. It waits at most
five seconds for a lock on attachments, so a long transaction can fail the first start, which
then tries again. Upgrading from 1.0.x or 0.3.x runs 1.1.0's migrations as well: read 1.1.0's
Before upgrading first.
- Going back to 1.1.0 after files moved to an object store needs them moved back first: set
attachments.store: postgres with the store still configured, and wait for the move to finish.
A 1.1.0 server cannot read a file that is in the store.
- New Bots run on the LangGraph (TypeScript) engine where a deployment runs it. OpenBot
Desktop runs it, and an installation's Bot that ran on the Python LangGraph engine moves to the
TypeScript one at its next Install and Start, keeping its conversations. Docker Compose now
defines agent-langgraph-ts (port 4215, on this machine only), which a plain docker compose up
builds from the checkout unless LANGGRAPHTSIMAGE and IMAGEPULLPOLICY=missing name the
published image. A server without LANGGRAPHTSAGUIURL, and both Helm charts, keep new Bots
on the server's built-in engine. OPENBOTDEFAULTBOT_ENGINE=built-in keeps every new Bot there.
- Compose: Claude and ChatGPT plans run in services of their own. agent-harness is no longer
handed CLAUDECODEOAUTHTOKEN, CHATGPTAUTH_FILE or the ./.langchain folder. They go to
two new services, agent-claude-plan and agent-chatgpt-plan, behind the claude-plan and
chatgpt-plan profiles, which OpenBot Desktop starts when a plan needs them. A Compose override
of your own that ran a plan through agent-harness turns on the profile and sets
CLAUDEPLANHARNESSURL or CHATGPTPLANHARNESSURL.
- The model picker offers only models proven on the Bot's engine (below). A conversation or a
Bot set to a model that isn't proven runs on the default, and says so.
- A model gateway serves /responses for gpt-5.4 and gpt-5.5. Bots behind an OpenAI-compatible
address now call those models, and their mini, nano and dated versions, on the Responses API.
- A server that refuses to start exits with code 78, where it exited with 1. The all-in-one
image now ends its container instead of trying again forever, and on Kubernetes the pod's
CrashLoopBackOff names the reason in kubectl describe. A crash still exits with 1. Monitoring
that watches the exit code may need to know the new one.
- Settings that now stop a server from starting when they are wrong: OPENBOTMINIMUMCLIENTS,
OPENBOTDEFAULTBOTENGINE, SHUTDOWNKEEPSERVINGSECONDS, SHUTDOWNDRAINSECONDS, and the
file store's (below). An S3BUCKET, GCSBUCKET or AZURESTORAGECONTAINER already in the
server's environment, for example one it shares with Intelligence, turns the file store on, and
the server then checks that store at start.
- Kubernetes: a server pod takes up to 45 seconds to stop, so it can finish what it is doing
(below). server.shutdown sets it.
- Deprecated: nothing.
Files
- Files can live in an object store. Amazon S3 or anything that speaks its API, Google Cloud
Storage or Azure Blob Storage, where they were always in PostgreSQL. With a store, a file can be
up to 100 MB. Without one nothing changes: files stay in PostgreSQL, with 8 MB for pictures and
PDFs and 1 MB for other files, and OpenBot Desktop stays on PostgreSQL. The openbot chart gains
attachments.store (postgres unless set) and objectStorage: enabled, provider (s3,
gcs or azure), bucket, prefix (openbot/), region, endpoint, forcePathStyle,
azure.accountUrl, deleteAfterDays (7), assumeAllServersCurrent, existingSecret with its
secretKeys, and inCluster for a store in the cluster. A server pod can sign in to the store
with workload identity through serviceAccount.annotations. On Docker Compose the same settings
are FILESSTORE, FILESPREFIX, FILESDELETEAFTERDAYS, FILESASSUMEALLSERVERS_CURRENT,
S3, GCSBUCKET and AZURESTORAGE in .env. The chart refuses a store asked for and
not set up, a Secret without its keys, and a store in the cluster behind a network policy with no
way to reach it, each by name.
- The server writes, reads and deletes a test file at start, and refuses to start when the store
doesn't answer. It never falls back to PostgreSQL.
- Files already in PostgreSQL move to the store in the background, and back again when
attachments.store is set to postgres with the store still configured. A server refuses to
start without the store's settings while any file is still in it.
- During a rolling upgrade new files stay in PostgreSQL until no 1.1.0 server is connected. A
connection pooler, such as PgBouncer or RDS Proxy, hides which servers are connected: set
assumeAllServersCurrent: true once every server runs 1.1.1.
- A deleted file leaves the store after deleteAfterDays, and a weekly sweep removes anything
left behind. Files are served from OpenBot's own addresses, as before.
- Word, Excel and PowerPoint files. People can attach .docx, .xlsx and .pptx files, and
Bots can send them, recognised by what is inside rather than by the name. The Bot reads their
v1.1.02026-11-17
升级前必读:本次更新涉及多处不兼容变更,脚本、代理、Compose 覆盖文件和 Helm values 都可能需要调整。
数据库迁移在启动时执行且不可回滚,从 0088approvalreviewerguidance 一路推进到本版 release-policy.json 指定的迁移。其中 0089goalendswhenmet 给 responsibilities 加上 endswhenmet(存量目标均为 false,行为不变),0090routinegoallinkedbyrun 给 routines 加 goallinkedbyrun,0091remotemcpoauth 给 mcp_servers 加 oauth。注意:已记录目标的例行任务今后会被视为专为该目标而设,目标完成即停用、删除目标即删除任务。14 张表的数据迁入新表,旧表保留但不再写入,回滚只能靠升级前的备份。
PostgreSQL 需要新建两个不可登录的角色 openbotapp 和 openbotsystem,OpenBot 之后以这两个角色执行查询,由数据库层隔离组织数据。登录账号需有建角色权限,或由管理员预先创建;否则迁移会打印所需 SQL 并中止。前置的连接池必须透传 role 启动参数,Compose 与 openbot-bundle chart 已自行配置。
旧的 /api/... 地址全部下线,统一走 /api/v2/,其余一律 404。脚本、集成和 WebSocket 客户端需改用 api/openapi.yaml、api/asyncapi.yaml 中列出的地址;反向代理或防火墙要放行 /api/v2/ 及其 WebSocket 升级。API 现在会拒绝重复的单值查询参数和不合法 JSON 请求体,签到接口也不再把空 days= 当作 7。
个人版只接受携带安装令牌的请求:无凭据访问会被拒绝,需使用 Settings > API keys 生成的 API key,或在 X-OpenBot-Install-Token 中带上令牌;容器化单用户部署必须设置 OPENBOTINSTALLTOKEN。
镜像改用 docs/cosign.pub 中的密钥签名,不再附带 GitHub 构建证明和 sha-<commit> 标签,请改用版本 tag 或 container-images.json 中的 digest 部署。
存量数据全部归入首个组织,仍以 default 寻址,所有人自动成为成员、管理员即该组织管理员,无人被登出、页面地址不变。原先针对目录组的规则(权限开关、网络策略、发布或置顶到组的机器人、共享应用账号的受众)改为指向该组织内同名同成员的团队,API 以 scopeKind: team 响应;仅大小写不同的名称会追加编号。
v1.1.0-rc.2预发布2026-11-17
升级前必读:本次更新包含多项破坏性变更,脚本、代理、Compose 覆盖文件和 Helm values 都可能需要调整。
数据库迁移在启动时执行且不可回滚,从 0088approvalreviewerguidance 一路推进到本版 release-policy.json 指定的迁移。其中 0089goalendswhenmet 给 responsibilities 加上 endswhenmet(存量目标默认为 false,行为不变);0090routinegoallinkedbyrun 给 routines 加 goallinkedbyrun;0091remotemcpoauth 给 mcp_servers 加 oauth。注意:已记录目标的例行任务今后会被视为专为该目标而设——完成该目标会关闭例行任务,删除目标则连带删除它。14 张表的数据迁入新表,旧表保留但不再写入,想回退只能恢复升级前的备份。
PostgreSQL 需要两个不可登录的角色 openbotapp 和 openbotsystem,OpenBot 之后以这两个角色执行查询,由数据库层面隔离组织数据。登录账号需有建角色权限,否则迁移会打印所需 SQL 并中止。若前面有连接池,必须透传 role 启动参数。
旧的 /api/... 地址全部退役,统一走 /api/v2/,其余一律 404。脚本、集成和 WebSocket 客户端需按 api/openapi.yaml、api/asyncapi.yaml 改地址;反向代理或防火墙要放行 /api/v2/ 及 WebSocket 升级。API 现在会拒绝重复的单值查询参数和非 JSON 请求体(返回 400),days= 空值也不再当作 7。
个人版只接受携带安装令牌的请求:无凭据访问会被拒,需提供 API key 或在 X-OpenBot-Install-Token 中带上令牌;容器化单用户部署必须设置 OPENBOTINSTALLTOKEN。
镜像改用 docs/cosign.pub 中的密钥签名,不再附带 GitHub 构建证明和 sha-<commit> 标签,请改用版本标签或 container-images.json 中的摘要部署。
存量数据全部归入首个组织,仍以 default 寻址,所有人自动成为成员,管理员即该组织管理员,无需重新登录,页面地址不变。原先针对目录组的规则(权限开关、网络策略、发布或置顶到组的机器人、共享应用账号的受众)改为指向该组织中同名同成员的团队,API 以 scopeKind: team 响应;仅大小写不同的名称会追加编号。
v1.1.0-rc.1预发布2026-11-16
升级前必读。 这次更新有几处会直接改变脚本、代理、Compose override 或 Helm values 的写法,务必先看。
数据库迁移只进不退。 启动时自动从 0088approvalreviewerguidance 迁到本版 release-policy.json 指定的版本,途经组织、团队、Agent、任务运行等迁移。其中 0089goalendswhenmet 给 responsibilities 加 endswhenmet(存量目标均为 false,行为不变),0090routinegoallinkedbyrun 给 routines 加 goallinkedbyrun,0091remotemcpoauth 给 mcpservers 加 oauth。注意:存量 routine 的 goallinkedbyrun 虽为 false,但已关联目标的 routine 会被视为专为该目标而设——包括 1.0.x 因某次触发而挂上目标的宽泛 routine,现在完成该目标会关闭 routine,删除目标则删除 routine。14 张表的数据迁入新表,旧表保留但不再写入;要回退只能恢复升级前的备份。
PostgreSQL 需要两个角色。 迁移会创建不可登录的 openbotapp 与 openbotsystem,授权给 OpenBot 登录账号后,查询以这两个角色执行,由数据库层面隔离组织。登录账号须有建角色权限,或管理员先建好这两个角色,否则迁移会打印所需 SQL 并中止、不做任何改动。两角色不得是超级管理员或绕过行级安全。服务器启动时及每隔数分钟校验查询角色,不符即停,因此前置连接池必须透传 role 启动参数;Compose 与 openbot-bundle chart 已自行配置。
旧 /api/... 地址全部下线。 服务端只保留 /api/v2/ 一处 API,其余旧地址一律 404。调用旧地址的脚本、集成或 WebSocket 客户端需改用 api/openapi.yaml、api/asyncapi.yaml 中列出的 /api/v2 地址;按路径逐条放行的反向代理或防火墙要放行 /api/v2/(含 WebSocket 升级),并继续放行「留在 /api/v2 之外」清单中的地址。
镜像改用 OpenBot 自有密钥签名(docs/cosign.pub,见 docs/deployment.md),不再附带 GitHub 构建证明,也不再发 sha-<commit> 标签;请部署版本标签或 container-images.json 中的 digest。
存量数据归入首个组织,仍叫 default,所有人是成员、管理员即其管理员,无人掉线、页面地址不变。原先针对目录组的规则(权限开关、网络策略、发布或置顶到组的 Bot、共享应用账号的受众)改为指向该组织中同名同人的团队,API 返回 scopeKind team;仅大小写不同的重名会加序号,如「Sales (2)」。设置 OPENBOTMULTIPLEORGANIZATIONS 后才支持多组织。
Kubernetes:openbot chart 不再安装 Intelligence,0.2 chart 上运行其自带 Intelligence 的版本升级会被拒绝,集群不做任何改动。
Compose:多项服务与配置被移除。 SPIRE(spire-init、spire-server、spire-agent)、agent-bot 与 TypeScript agent-langgraph Bot、BOTPORT、BOTIMAGE、AGENTBOTMODEL、WORKERSHAREDSECRET 均已删除,也不再启动 migrate 与 worker 服务;override 或脚本中引用这些的须一并清理。
v1.0.32026-11-14
- The OpenMuse skin is now called OpenFuse. Anyone who picked it keeps it, with nothing to
choose again.
- A Microsoft Agent Framework Bot on an Anthropic key sends the key again. Its image holds
agent-framework-anthropic at 1.0.0b261106, as main does, because 1.0.0b261112 (released
2026-11-12) sends the request without it.
v1.0.3-rc.1预发布2026-11-14
- The OpenMuse skin is now called OpenFuse. Anyone who picked it keeps it, with nothing to
choose again.
- A Microsoft Agent Framework Bot on an Anthropic key sends the key again. Its image holds
agent-framework-anthropic at 1.0.0b261106, as main does, because 1.0.0b261112 (released
2026-11-12) sends the request without it.
v1.0.22026-11-11
Before upgrading. No migrations run in this release: the database stays at
0088approvalreviewer_guidance, and a 1.0.1 database is used as it is. Upgrading from 1.0.0 or
0.3.x runs the migrations listed under 1.0.1, and a database that has run them does not go back:
restore the backup taken before the upgrade instead. Three things change for an existing OpenBot
Desktop installation:
- On Windows OpenBot now always runs in a Podman machine of its own, openbot, beside any of
yours, and on a Mac it does when yours is rootless, after asking to pause it (below). An
installation that ran in a machine of yours starts fresh in openbot, and its earlier data stays
in your machine.
- At setup and at every Start, a Docker Desktop or Podman too old to run OpenBot is updated
(below). Docker Desktop quits while it updates, which stops its containers; the ones that were
running are started again afterwards.
- OpenBot Desktop's back end and the Helm chart move to CopilotKit Intelligence 0.16.0-rc.11, and
the bundled CLI to copilotkit@7bd073b (below).
OpenBot Desktop
- **Installing a new OpenBot over the old one on Windows never offers to delete your data, and
uninstalling stops what OpenBot runs.** Running a new installer over an installed OpenBot opened
the old version's uninstaller in the middle of the update, with a "Delete the application data"
box, and one wrong tick deleted your Bots and conversations. The old version is now removed
silently, as an update, on the Installing page, which says "Removing the previous version of
OpenBot. Your Bots and data stay." with its progress bar moving. If OpenBot is open, the installer
asks before it closes it. Uninstalling from Settings > Apps used to close OpenBot and leave its
containers, its services and its Podman machine (openbot) running. It now stops them first,
with the same Compose OpenBot runs, and never touches your own Podman machines or containers.
Tick "Delete the application data" to remove OpenBot's machine, the backend OpenBot created and
OpenBot's files in its installation folder as well, so installing again starts from nothing; a
file of yours in that folder stays. Otherwise they stay, stopped, for a reinstall to find. And on
Windows, an OpenBot whose own Podman machine is gone makes it again when it opens. It used to open
on "OpenBot can't run on this computer as it's set up", having looked at your own machine.
- Setup says what self-contained means up front, and asks about usage data last. The first
screen, "Set up OpenBot (self-contained)", says this version runs entirely on your computer,
backend included, that setup takes between 1 and 10 minutes depending on your internet speed,
and that in production OpenBot for Business runs on your own infrastructure with Kubernetes, or on
OpenBot Cloud, with a link to book a demo. The usage-data notice moved off
it to the foot of the last screen, Finishing setup, as a small checkbox, checked by default as
before: it is the same setting as Settings > General > Privacy > Share usage data. The Account
step's heading is now "Get your free license key".
- OpenBot uses its own Podman machine when yours can't run it, and leaves yours as it is. A
Podman machine made with podman machine init is rootless, and OpenBot used it and stopped at
"Change this computer's setup" because local Intelligence needs rootful Podman. On Windows even a
rootful machine of yours left chat unable to reach OpenBot's back end. OpenBot now makes and uses
its own rootful machine, openbot, as on a computer with none. On Windows it runs beside yours,
which keeps running untouched, and never takes Docker's default pipe (docker_engine), so your
Docker tools and Docker Desktop keep reaching what they did. On a Mac, where Podman runs one
machine at a time, OpenBot asks before it pauses your rootless machine while it runs (**Pause my
Podman**), and Quit starts yours again with its containers. Docker, and a rootful machine of
yours on a Mac, are used as before. Your default Podman connection is never changed. An
installation that ran in your own machine starts fresh in openbot; its old data stays in your
machine.
- OpenBot updates Docker Desktop or Podman when it is too old to run OpenBot. A computer with
Docker Desktop 4.27 got through setup and then stopped at Start with "Compose config exited with
exit status: 15", an error nobody could act on. OpenBot now checks the engine at setup and at
every Start, before Compose reads anything, and updates one that is too old for you: Docker
Compose before 2.27.2 (Docker Desktop before 4.32), or Podman before 6.1. Docker Desktop's update
comes from Docker's own update feed, is checked against Docker's checksums and signature, and
installs over the one you have, keeping its images, containers and volumes; the containers that
were running before are started again once it is back. Podman's is OpenBot's own pinned
installer over the old one, and every Podman machine that was running before, yours included, is
running afterwards with its containers. A Mac asks for your password, and an all-users Docker
Desktop on Windows for Windows' approval. The row for what's installing in the background says
why it is updating and shows the update's own lines in full, and the Account step waits for the
update and says so. Where an update can't be done for you, such as a Mac or PC too old for the
newest Docker Desktop, or an Intel Mac with an old Podman (Podman publishes its Mac package for
Apple silicon only), one plain step says what to do. Separately, the Compose file now starts on
an older Compose too: with no harness Bot picked, Compose 2.25 and older refused the whole
project.
- A Podman machine that stops answering is restarted. A Podman machine OpenBot uses that had
lost its Docker API failed setup and Start until you restarted it yourself, and Quit could not
stop the back end, so the next launch said "OpenBot had trouble shutting down last time." At every
check, including Start and Retry after a relaunch, and before Stop and Quit, such a machine is now
stopped and started once, your default machine included. A failure while setup runs behind the
Account step now has Try again.
- Start says in the log what it is waiting on, and why it gave up. A Start that waited five
minutes for OpenBot's services and then said "OpenBot didn't get ready in time" left nothing in
the log in between. Each wait now writes what it asked (the address or the back end's status),
each new answer and the same one again once a minute, and why it ended, with how long it took and
the last answer.
- The big downloads start on the first screen, before you click Install. Nothing used to
download until Install, and then one thing at a time: on a new Mac, Podman, then the 937 MB
engine image with no progress shown, then 1.1 GB of OpenBot's software. Now the engine image,
Podman's installer, the deployment and, when Docker or Podman already runs, OpenBot's software
start downloading as soon as the first screen shows, side by side, while you sign in and enter
your Mac password. The first screen says how far they are ("Downloading OpenBot. About 3 minutes
left.", with the bytes under a bar), and the row for what's installing in the background carries
on with it. A download stopped by quitting carries on from where it stopped next time,
and every file is checked against its pinned digest before it is used. Podman builds its machine
from the downloaded image instead of fetching its own. With less free disk than OpenBot needs,
the first screen says how much is free and how much is needed, and the downloads start once
there is room. Nothing is installed before you are asked.
- OpenBot's own software downloads on the first screen too, before there is anything to run it.
On a new Mac or PC, or with Docker Deskto
v1.0.2-rc.3预发布2026-11-11
Before upgrading. No migrations run in this release: the database stays at
0088approvalreviewer_guidance, and a 1.0.1 database is used as it is. Upgrading from 1.0.0 or
0.3.x runs the migrations listed under 1.0.1, and a database that has run them does not go back:
restore the backup taken before the upgrade instead. Three things change for an existing OpenBot
Desktop installation:
- On Windows OpenBot now always runs in a Podman machine of its own, openbot, beside any of
yours, and on a Mac it does when yours is rootless, after asking to pause it (below). An
installation that ran in a machine of yours starts fresh in openbot, and its earlier data stays
in your machine.
- At setup and at every Start, a Docker Desktop or Podman too old to run OpenBot is updated
(below). Docker Desktop quits while it updates, which stops its containers; the ones that were
running are started again afterwards.
- OpenBot Desktop's back end and the Helm chart move to CopilotKit Intelligence 0.16.0-rc.11, and
the bundled CLI to copilotkit@7bd073b (below).
OpenBot Desktop
- **Installing a new OpenBot over the old one on Windows never offers to delete your data, and
uninstalling stops what OpenBot runs.** Running a new installer over an installed OpenBot opened
the old version's uninstaller in the middle of the update, with a "Delete the application data"
box, and one wrong tick deleted your Bots and conversations. The old version is now removed
silently, as an update, on the Installing page, which says "Removing the previous version of
OpenBot. Your Bots and data stay." with its progress bar moving. If OpenBot is open, the installer
asks before it closes it. Uninstalling from Settings > Apps used to close OpenBot and leave its
containers, its services and its Podman machine (openbot) running. It now stops them first,
with the same Compose OpenBot runs, and never touches your own Podman machines or containers.
Tick "Delete the application data" to remove OpenBot's machine, the backend OpenBot created and
OpenBot's files in its installation folder as well, so installing again starts from nothing; a
file of yours in that folder stays. Otherwise they stay, stopped, for a reinstall to find. And on
Windows, an OpenBot whose own Podman machine is gone makes it again when it opens. It used to open
on "OpenBot can't run on this computer as it's set up", having looked at your own machine.
- Setup says what self-contained means up front, and asks about usage data last. The first
screen, "Set up OpenBot (self-contained)", says this version runs entirely on your computer,
backend included, that setup takes between 1 and 10 minutes depending on your internet speed,
and that in production OpenBot for Business runs on your own infrastructure with Kubernetes, or on
OpenBot Cloud, with a link to book a demo. The usage-data notice moved off
it to the foot of the last screen, Finishing setup, as a small checkbox, checked by default as
before: it is the same setting as Settings > General > Privacy > Share usage data. The Account
step's heading is now "Get your free license key".
- OpenBot uses its own Podman machine when yours can't run it, and leaves yours as it is. A
Podman machine made with podman machine init is rootless, and OpenBot used it and stopped at
"Change this computer's setup" because local Intelligence needs rootful Podman. On Windows even a
rootful machine of yours left chat unable to reach OpenBot's back end. OpenBot now makes and uses
its own rootful machine, openbot, as on a computer with none. On Windows it runs beside yours,
which keeps running untouched, and never takes Docker's default pipe (docker_engine), so your
Docker tools and Docker Desktop keep reaching what they did. On a Mac, where Podman runs one
machine at a time, OpenBot asks before it pauses your rootless machine while it runs (**Pause my
Podman**), and Quit starts yours again with its containers. Docker, and a rootful machine of
yours on a Mac, are used as before. Your default Podman connection is never changed. An
installation that ran in your own machine starts fresh in openbot; its old data stays in your
machine.
- OpenBot updates Docker Desktop or Podman when it is too old to run OpenBot. A computer with
Docker Desktop 4.27 got through setup and then stopped at Start with "Compose config exited with
exit status: 15", an error nobody could act on. OpenBot now checks the engine at setup and at
every Start, before Compose reads anything, and updates one that is too old for you: Docker
Compose before 2.27.2 (Docker Desktop before 4.32), or Podman before 6.1. Docker Desktop's update
comes from Docker's own update feed, is checked against Docker's checksums and signature, and
installs over the one you have, keeping its images, containers and volumes; the containers that
were running before are started again once it is back. Podman's is OpenBot's own pinned
installer over the old one, and every Podman machine that was running before, yours included, is
running afterwards with its containers. A Mac asks for your password, and an all-users Docker
Desktop on Windows for Windows' approval. The row for what's installing in the background says
why it is updating and shows the update's own lines in full, and the Account step waits for the
update and says so. Where an update can't be done for you, such as a Mac or PC too old for the
newest Docker Desktop, or an Intel Mac with an old Podman (Podman publishes its Mac package for
Apple silicon only), one plain step says what to do. Separately, the Compose file now starts on
an older Compose too: with no harness Bot picked, Compose 2.25 and older refused the whole
project.
- A Podman machine that stops answering is restarted. A Podman machine OpenBot uses that had
lost its Docker API failed setup and Start until you restarted it yourself, and Quit could not
stop the back end, so the next launch said "OpenBot had trouble shutting down last time." At every
check, including Start and Retry after a relaunch, and before Stop and Quit, such a machine is now
stopped and started once, your default machine included. A failure while setup runs behind the
Account step now has Try again.
- Start says in the log what it is waiting on, and why it gave up. A Start that waited five
minutes for OpenBot's services and then said "OpenBot didn't get ready in time" left nothing in
the log in between. Each wait now writes what it asked (the address or the back end's status),
each new answer and the same one again once a minute, and why it ended, with how long it took and
the last answer.
- The big downloads start on the first screen, before you click Install. Nothing used to
download until Install, and then one thing at a time: on a new Mac, Podman, then the 937 MB
engine image with no progress shown, then 1.1 GB of OpenBot's software. Now the engine image,
Podman's installer, the deployment and, when Docker or Podman already runs, OpenBot's software
start downloading as soon as the first screen shows, side by side, while you sign in and enter
your Mac password. The first screen says how far they are ("Downloading OpenBot. About 3 minutes
left.", with the bytes under a bar), and the row for what's installing in the background carries
on with it. A download stopped by quitting carries on from where it stopped next time,
and every file is checked against its pinned digest before it is used. Podman builds its machine
from the downloaded image instead of fetching its own. With less free disk than OpenBot needs,
the first screen says how much is free and how much is needed, and the downloads start once
there is room. Nothing is installed before you are asked.
- OpenBot's own software downloads on the first screen too, before there is anything to run it.
On a new Mac or PC, or with Docker Deskto