v0.3.1

Before upgrading. There are no migrations: the database stays at 0077_channel_opening. Ten
things change for an existing deployment:

• On OpenBot Desktop, the first Start after the app updates to this release moves the installed
  OpenBot to it too. Until now an installation kept the release it was installed at until you moved
  it under Runtime updates. While the old release keeps running, Start checks the installation
  against its release's published manifest and downloads the new release and its images; then it
  stops OpenBot, backs up the database, installs the new release and starts it, which can take a few
  minutes. A move that fails is undone, database included, and the old release starts. Pinned builds
  (internal and candidates), and an installation already on a newer release than the app, are left
  as they are (below).
• OpenBot Desktop installations signed in to CopilotKit with no Composio key get one at their next
  Start: the bundled CopilotKit CLI asks CopilotKit for this installation's own Composio project,
  and the key goes to the app's protected store, never to .env. Apps, unavailable on Desktop since
  0.3.0, become available. A COMPOSIO_API_KEY in ~/OpenBot/.env or the environment, or an Arcade
  choice, still wins, and then nothing is asked (below).
• When nobody has picked a model, built-in Bots now answer on BOT_MODEL while the provider lists
  it, and otherwise on the middle of the provider's current line: the newest model of each line,
  ordered by price, and the one in the middle (on 2026-11-05, gpt-6.1-terra, claude-sonnet-5-5,
  gemini-3.8-flash and grok-4.7). With no BOT_MODEL, that replaces the tenant package's
  default_model (gpt-5.6-terra in the bundled packages) or, on Anthropic, claude-sonnet-5-5.
  On OpenAI's own API, BOT_MODEL now applies to built-in Bots, where it used to be ignored. A
  BOT_MODEL the provider doesn't list gives way to the middle: a ChatGPT plan given gpt-5.5 runs
  the plan's middle model. Behind a gateway in OPENAI_BASE_URL, the configured model stays the
  default while the gateway lists it, and until the provider's list has been read the configured
  model is used. The Bot framework picked at setup, when it is OpenBot's own LangGraph or Claude
  Agent SDK harness, is sent the same model on every run (below).
• Nothing writes a BOT_MODEL on your behalf any more, so an unset one means the middle of the
  line. docker-compose.yml gives the picked harness an empty BOT_MODEL when it is unset, rather
  than gpt-5.5. OpenBot Desktop no longer writes BOT_MODEL=claude-sonnet-5-5 for an Anthropic
  key, and takes out the one it wrote at the next Start; a model chosen on a Desktop setup screen
  (an OpenAI-compatible endpoint's, or a Google or xAI sign-in's) is still written and stays the
  default. The harness answers on the model the server sends each run, and only when the server
  sends none, because it has not read the provider's list, on the provider spec file's model for it
  (gpt-4o-mini on OpenAI).
• The server now reads the provider's model list (a plan's own, through its harness) and
  https://models.dev/api.json when it starts and every six hours after, whether or not anybody
  picks, and again at once, at most once a minute, when a provider refuses a model as unknown or
  unsupported. A refresh that fails keeps the last good list. OPENBOT_MODEL_CAPABILITIES_URL
  points at a mirror of models.dev, or off reads none, and then the default is the configured
  model while the provider lists it.
• With OPENBOT_SINGLE_USER=true (OpenBot Desktop and the personal edition among them), the one
  person is offered every model the provider lists, with no switch and no allowlist. Admin →
  Providers shows neither, and a model allowlist saved earlier no longer applies.
• In a team, people pick among the models an admin allows under Admin → Providers → Allowed models,
  plus the default, and "Let people choose a model" is gone. Until an admin allows one, people get
  the default and see no picker: a team that let people pick before upgrading has no picker until
  models are allowed there, and one whose list already holds models now shows them, whatever the old
  switch said. A pick on a model no longer allowed, or no longer listed by the provider, runs on the
  default and is kept for when it is offered again. "Only allowed models" still refuses a run on any
  model not allowed, the default included, so with it on, allow the new default (above) or runs
  nobody picked a model for are refused.
• A Bot's commands now run in one interactive bash on its computer, one for each Bot a person uses,
  instead of a fresh bash -c each (below). cd and exported variables carry over to the next
  command. A command's stdout and stderr come back together as its output, with stderr empty, as a
  terminal shows them. Something a command starts in the background with & keeps running after the
  command returns, even while it still prints; such a process used to hold the command open until
  its time limit stopped both. Two commands from the same Bot and person run one after the other,
  and the wait counts against each one's time limit. Commands see TERM=xterm-256color, and
  PAGER=cat, GIT_PAGER=cat and GIT_TERMINAL_PROMPT=0 unless COMPUTER_SHELL_ENV passes
  others.
• Upgrade the server and its computers together. A computer from 0.3.0 has no terminal for the
  Computer panel's new Terminal tab to connect to, and refuses every pointer move made with no
  button held, which the live screen now sends as you hover.
• The Python LangGraph and Pydantic AI Bots set no upper bound on google-genai, so an image of
  either built from source since 3.0.0 came out on 2026-11-06 installed it, and 3.0.0 sends Gemini
  requests to /v1 instead of /v1beta. Both now hold google-genai at 2.34 or later and below 3,
  and the Google ADK Bot names the same range, so Gemini is called on /v1beta as before.

A Bot's commands run in one live shell you can watch and take control of

The Computer panel has a Terminal tab beside the Screen: a real terminal where you watch each of the
Bot's commands typed and its output stream, in colour. Whoever may watch a Bot's screen may watch
its terminal. Take control gives you its keyboard, as it does the browser's, and the Bot waits until
you hand back, then carries on in the same shell. The panel opens on the view the Bot last worked
in, the terminal after a command and the screen after a page; a tab you pick, or a command you click
in the conversation, holds until the panel closes.

• The Bot gets what it got before. Each command's output and exit code come back within the same
  time limit (120 seconds unless the Bot asks for up to 600) and the same 64 KB of output, and
  approvals and policy for commands are unchanged. What changes is under Before upgrading.
• Commands in the conversation and in Activity are terminal blocks: the prompt and the command,
  the last lines of output with "Show all N lines" for the rest, and how it ended, as "Exit 0",
  "Exit 1", "Running", "Blocked", "Declined" or "Done by you". Clicking one opens the Computer panel
  on its terminal.
• A shell left busy is not typed over. If you hand back with something still running in the
  foreground, the Bot's next command is not run and the Bot is told why. A command the Bot sends
  unfinished (an unclosed quote, bracket or here-document) is abandoned at once with exit code 2,
  instead of waiting out its time limit.
• The shell ends when the computer is stopped or reset, or after an hour nobody used or watched
  it, and a new one starts when it is next needed.

Taking over a Bot's screen feels like using your own browser

• Clicks land at once. Your browser sends at most one pointer position per screen refresh and
  adds scrolls together, and the Bot's computer replaces a waiting move with the newer one instead
  of playing every move before your click. Clicks, keys and pastes are never merged or dropped, and
  land in the order you made them.
• The picture is the page as it is now. The computer sends the next frame only once the last has
  gone out, with at most two on their way to your browser, and a newer frame replaces one still
  waiting, so a slow connection no longer shows frames seconds old. The once-a-second screenshot
  stops while the full-size live view is open, and a hidden tab draws nothing until you come back.
• Keys and the mouse behave as on your own page. Hovering is a hover, not a drag with the left
  button held. Letting go of the mouse outside the screen lets go on the page, and releasing Command
  releases the keys held with it. Emoji type as text, Firefox scrolls normal distances, and the
  cursor is an arrow rather than a crosshair.
• A missed input is a note, not an error screen. One key or click that doesn't land shows a
  short note in the corner for a few seconds, and the screen stays live. One missed check of who has
  control no longer pauses your input; two in a row do, and the computer still refuses input from
  anyone not in control. A click made the moment the screen appears is held until the connection to
  the computer opens, not lost.

OpenBot Desktop moves your OpenBot to the app's release when it starts

An updated app used to keep running the OpenBot it was first installed with, so server fixes never
reached an existing installation. Now Start moves a stable app's installation to the app's own
release, through the same path as Runtime updates, and says "Updating OpenBot from v0.3.0 to v0.3.1.
Your Bots, sign-ins and data stay as they are." while it runs. A modified installation is refused
before anything stops.

• A move that fails before OpenBot stopped changes nothing, and the next Start tries again. One that
  fails after is undone, the database restored from its backup, and the old release starts with a
  note saying so. Start does not try that release again by itself; Runtime updates can.
• Stop or Quit during a move stops it, and the next Start recovers it as any interrupted update.
• Runtime updates, and this move, download only the images this installation runs: eight for 0.3.0,
  where the release names nineteen. Staging an update no longer fails on an Apple silicon Mac with
  "no matching manifest for linux/arm64/v8" over the all-in-one openbot image, which Desktop never
  runs and which 0.3.0 published for x86-64 only.

OpenBot Desktop gets your apps from your CopilotKit sign-in again

0.3.0 started every Desktop installation without a Composio key, so apps were unavailable unless you
put your own COMPOSIO_API_KEY in ~/OpenBot/.env. The Account step now connects your apps once
you have signed in ("Connecting your apps…"): the bundled CopilotKit CLI, with its own sign-in,
makes or reuses this installation's Composio project, and the app keeps the key in its protected
store. If that fails, the step says so, with Connect apps to try again and Continue to go on without
apps, and every Start tries again by itself without ever stopping Start. An installation signed in
under 0.3.0 gets its key at its next Start, without signing in again.

Signing in with ChatGPT in Desktop setup no longer ends on "localhost refused to connect"

• A browser tab that comes back after the sign-in finished (from "Open the sign-in page", or from
  approving again at OpenAI) gets an OpenBot page saying OpenBot is signed in and the tab isn't
  needed. OpenBot answers that address for 15 minutes after a sign-in ends, then gives the port up,
  and gives it up sooner to a new sign-in or to the Codex CLI.
• Starting a sign-in again no longer fails with "port is already allocated". The earlier sign-in is
  stopped first, and so is one you leave by picking another provider, Change, Escape or leaving the
  screen, whose tab then says the sign-in ended rather than "You're signed in". Quit stops it too.
• A tab left over from an earlier attempt can no longer end the attempt in progress: it is told its
  link is out of date.

Pick the model and thinking effort for every Bot, from the provider's live list

The composer's model picker is now a text button naming the model and its thinking effort ("Claude
Sonnet 5.5 · High"), and its list adds a Thinking group with the levels the model in use accepts,
its default marked. A model is only ever shown, saved and sent the levels it takes, and switching to
a model that doesn't take the chosen effort clears it. The pick is still kept per person and
conversation, a change applies from the next message, and a new conversation saves both before its
first message.

• Who picks from what. One person on their own machine picks from every model the provider
  lists. In a team, people pick among the models an admin allows under Admin → Providers → Allowed
  models, which lists each with its maker's mark and thinking levels, and see no picker until one is
  allowed (above).
• More Bots and every sign-in. It is offered for the Bot framework picked at setup when it is
  OpenBot's own LangGraph or Claude Agent SDK harness, and on OpenAI, Anthropic, Google and xAI keys
  and sign-ins, any OpenAI-compatible endpoint, and ChatGPT and Claude plans, each listed from its
  own source (see The model list in docs/configuration.md). A Bot on any other framework, or a
  harness somebody brought, still picks its own model.
• Fixed: every message failed on a ChatGPT plan with "The 'gpt-5.5' model is not supported when
  using Codex with a ChatGPT account". A plan now runs only a model its own list offers.
• Out of credit says so. A provider refusing because the account is out of credit or at its
  spending limit now reads as that, naming the provider and quoting it, on built-in Bots (which
  showed "Forbidden" or a retry count) and on the LangGraph harness (which asked you to sign in
  again, and flagged the sign-in as failed). Any other refusal from the provider ends the harness's
  run with the provider's own words, where the run used to stop with no answer.
• Admin → Providers names who serves the model: OpenAI, Anthropic, Google, xAI, ChatGPT plan, Claude
  plan or OpenAI-compatible endpoint, with Google, xAI and Claude marks beside OpenAI's and
  Anthropic's. Each run logs one run-model line with its Bot, run id, model and effort.
• CHATGPT_CODEX_CLIENT_VERSION (default 0.167.0) is the Codex client version the LangGraph
  harness names when it reads a ChatGPT plan's list.

Every step a Bot takes with one of OpenBot's own tools reads as words

Each tool OpenBot itself gives a Bot is named in words chosen for it, in the present while it runs
and the past once done ("Using a skill", then "Used a skill"; "Saved a memory"; "Checked existing
Bots"), where most showed words split from their identifier ("Copilotkit load skill", "Remember
personal fact"). That covers skills, memory, follow-ups, check-ins, responsibilities, asking a
person or another Bot, subtasks, the browser, and files and commands on the Bot's computer or yours.

The Mastra Bot's image installs from a lockfile

The Mastra Bot was the one Bot image built from version ranges with no lockfile, so two builds of
one release could ship different code, and a build could fail on a package npm had only just
published. It now installs from agent-mastra/bun.lock with --frozen-lockfile.