v5.0.0 - New React App, Conversation-First Workspace & One-Click Fixes

v5.0.0 - New React App, Conversation-First Workspace & One-Click Fixes

This major release replaces Row-Bot's NiceGUI interface with one React app
and rebuilds the workspace around the conversation. Designs, code folders,
goals, delegated agents, approvals and the terminal now live with the
conversation that uses them; Home opens on an Overview of what needs you; a
fresh profile asks one question, "How should Row-Bot think?", and never runs
on a model you didn't choose. Approvals can be answered wherever you are,
Monitor keeps its checks and offers one fix per problem, phones and other
computers connect with a self-renewing QR code, and your profile can be backed
up and restored. Underneath, the server is plain FastAPI run by uvicorn, about
65,000 lines of NiceGUI-era code are gone, and the test and release pipeline
runs each genuine test once. Local-first defaults, approval gates and the
single-owner access model are unchanged. Read Upgrade Notes And Breaking
Changes below before upgrading.

One App, Rebuilt In React

• React is the only interface - the desktop window, browsers, phones,
  tablets and server mode all open the same React app at /app-v2/. /
  redirects there and keeps its query string; a remote browser without a
  session goes to /connect first.
• Plain FastAPI server - the app server is a FastAPI app run by uvicorn,
  using HTTP and server-sent events only (no WebSockets). Start-up, shutdown
  and cleanup keep their order, responses are compressed and the event stream
  is never compressed or buffered. NiceGUI and 16 locked packages left the
  dependency set.
• Background work sees what you are doing - Dream Cycle, memory
  extraction, checkpoint and thread cleanup, browser tab eviction and the
  status tool now see turns running in the app, and sending a message or
  opening a conversation marks you active, so background jobs wait while you
  work and never extract from the conversation you have open.
• One launcher per data folder - starting Row-Bot again from a shortcut
  while it runs brings the running app forward instead of starting a second
  instance against the same profile.
• Reconnects by itself - after a server restart or a lost connection, open
  windows reconnect, reopen the conversation and keep unsent drafts, with
  Reconnect offered instead of a dead end.
• Errors that say what to do - every server error reads as one sentence
  with at most one fix (Retry, Reconnect, Choose a model, open the exact
  setting, or Send now); an unexpected one shows its code under Details.
• Quiet background notices - plugin load failures, tunnel start-up
  failures and token warnings arrive as short floating notices, once per
  device per start, and start-up warnings are listed in Monitor.
• A desktop window that behaves like an app - native windows get a
  right-click menu (Cut, Copy, Paste, Select All, never on password fields),
  message text can be selected and copied, the Windows clipboard keeps accents
  intact, and exports and downloads use the native Save dialog and land where
  you chose.
• Buddy desktop overlay in React - the torn-off Buddy is now part of the
  React app: it follows the selected conversation's turns, drafts and
  approvals, docks back reliably, keeps its desktop features through long
  sessions, and places itself correctly on high-DPI and multi-monitor setups.
  The sidebar Buddy is larger and wears an activity ring.

A Redesigned Workspace

• Sidebar - New chat (Mod+Shift+O) or a new chat with an agent, Home,
  Agents (favourite profiles start a chat in one click; All agents opens the
  profile library), and conversations by date with pinning and a type filter
  (All, Chats, Designs, Code, Workflows) that covers every conversation, not
  only the most recent pages. Settings and the attention indicator sit in the
  footer; the collapsed rail keeps every destination as a labelled icon.
• Home - five tabs: Overview, Workflows, Knowledge, Monitor and Insights.
  Overview greets you with Buddy and an Ask box, then Needs you (approvals,
  problems and setup, with their fixes in place), Continue where you left off,
  today's agents and workflow runs, Since yesterday evening, and Learned this
  week.
• Conversation details card - a floating card beside the chat shows what
  the conversation is Working on (designs, code folders, the browser), its
  goal, its outputs and its agents. Below desktop width it opens as a sheet.
• A calmer transcript - a centred reading column with speaker markers
  (your Buddy's image for Row-Bot, a person icon for you), one activity row
  per turn ("Used 3 tools · 1 failed") that expands into its steps,
  highlighted code with Copy and Download, tables that copy as CSV, citations
  as source chips, link chips, inline PDFs, generated media shown once, and
  agent runs marked where they start, finish or fail.
• One composer field - paste screenshots, drop files or pick several at
  once, with limits stated up front (25 MB a file, 32 files, 100 MB a
  message); attachments show as thumbnails with progress and Retry, and Send
  waits for uploads still running. /goal, /reasoning, /profile and
  /agent run with their argument, @ mentions agent profiles, write targets
  and files, active skills show as chips, and Send turns into Stop.
• Messages that wait their turn - a message sent while Row-Bot works waits
  in one list above the composer with Send now, Edit and Discard; Stop pauses
  waiting messages, and nothing is ever sent twice.
• Stop keeps what was written - a stopped reply stays in the conversation
  with a Stopped marker, also after a reload.
• Conversations name themselves - a new conversation takes its first words
  at once and a short descriptive name after the first reply.
• Command palette (Ctrl+K / ⌘K) - one search across conversations
  (including message text), commands, settings and saved workflows. It
  understands plain intents in any word order ("connect a model", "turn on
  developer tools", "phone", "connect telegram", "dark mode"), lists settings
  switches as actions showing their current state with Undo, ranks message
  matches above scattered letter matches, lists the open design's commands,
  and Enter runs only a result that matches what you typed.
• Library and conversation menus - each conversation's menu pins,
  renames, exports as Markdown or PDF in one step, and deletes. The Library
  lists every conversation with type filters, search and multi-select (Select
  all, Clear all) to pin, export or delete many at once, with deletion
  progress shown. Deleting the open conversation keeps the app connected.
• Appearance - bundled Geist type, light and dark themes that follow the
  operating system until you choose, plus accent colour, density and reduce
  transparency in Settings › Appearance.
• Phones and tablets - phones get one header per page, a one-line
  composer, panels as full-screen sheets and 44 px touch targets throughout;
  tablets keep centred dialogs; Settings keeps its search at every width.

First Run And Models

• No preset models - a fresh profile has no chat, vision, image or video
  model until you choose one; the built-in qwen3:14b, gemma3:4b,
  gpt-image-1.5 and veo-3.1 presets are gone. Channels, workflows, Dream
  Cycle, memory extraction, delegated agents and Buddy's Hatch wait or reply
  politely until a model is chosen, and removing a provider keeps your saved
  default (shown as unavailable) instead of falling back to another model.
• "How should Row-Bot think?" - the first run asks one question: On this
  computer (Ollama detected and re-checked automatically, with install steps
  for your operating system and your installed models listed), With my
  subscription (ChatGPT, Claude or Grok with a device code that is checked
  automatically), With an API key (OpenAI, Anthropic, Google Gemini and
  OpenRouter first, the rest under More providers; each key is checked before
  it is saved), or a custom endpoint. The pick becomes your default, a
  one-message test runs and Home opens; if another assistant's data is found,
  an import is offered. It replaces the setup wizard.
• Setup Center - the remaining areas live in Setup Center, which counts
  done and skipped areas separately and reads their real state.
• An honest model pill - the composer's model pill shows a local or cloud
  glyph, "Chat only", "Choose a model", or "Unavailable" with the reason and
  Reconnect or Choose another model. It never reads "Ready" for a model that
  cannot run.
• One model list everywhere - the composer and the Brain, Vision, Image
  and Video pickers share one grouped, searchable list with billing tags
  (Subscription, Pay per use, Credits, Local · free), so subscription models
  such as ChatGPT and Claude subscription models can be chosen as Brain and
  Vision. Vision defaults to "Same as chat model", and attaching an image to a
  model that can't see images says so and offers Choose a vision model.
• Settings › Models - rebuilt as plain rows with readable model names,
  labelled icon buttons and the default model renamed "Brain model".
• Voice - Settings › Voice installs Whisper at the chosen size, showing
  its download size and source first; Realtime Talk says it is paid per minute
  through OpenAI.
• Custom endpoints - refusals show inside the dialog, the list stays put
  while it refreshes, and an endpoint that doesn't answer reads Unavailable
  instead of waiting forever.

Doing Things By Conversation

• Designs and code folders on request - ask for a deck or an app and the
  assistant creates the design or a code folder named after the request, then
  keeps working in it. A card in the transcript offers Open, Rename and Undo.
  Wording alone never creates or binds anything, and the old keyword-based
  setup before a turn is gone.
• Existing folders and clones - "use my folder" binds a registered
  code folder by name, or offers Choose folder when none matches; "clone " shows a clone card, you choose where it goes, and the conversation continues in the new folder. A "Using code folder" card offers Open and Undo, which unbinds the folder and keeps its files. The model never picks a path.
• Setup cards - when work needs a tool that is off, such as web search or
  Developer tools, the transcript asks "Turn on ?" with Turn on and Not
  now through the approval-gated settings change; a missing account or channel
  shows a Connect card that opens its setup. Not now is final for that
  request.
• Goals without a turn limit - set a goal from the conversation details
  card or with /goal. It starts at once and continues after each turn until
  it is done, with no turn limit by default and optional turn and time limits.
  The goal card shows the turn, time running, tokens, the latest progress
  reason, Pause or Resume, and Stop.
• Goals that know when to stop - a goal pauses with its reason after two
  turns without progress or when the same step fails three times; it waits
  out a provider's rate or usage limit and continues when the limit resets
  (or pauses and says so when the provider gives no time); approvals pause it;
  a conversation's goal continues after Row-Bot restarts, while a channel goal
  pauses.
• Delegated agents in reach - every agent appears in the transcript as it
  starts, finishes or fails, and in the details card's Agents with Stop and
  Message. An agent's own conversation opens with "Task from the parent
  conversation" instead of its internal handoff prompt and offers a compact way
  back to the parent. Agent work interrupted by a restart can be resumed or
  dismissed.

Approvals Answered Where You Are

• One place for every approval - approvals from workflows, other
  conversations and delegated agents appear in the attention indicator's list,
  in Overview's Needs you, as a floating notice and in Buddy, each with
  Approve and Deny in place. A delegated agent's approval can be answered in
  its own conversation or in the parent's.
• Approvals wait until answered - approvals no longer expire after 30
  minutes by default, and each shows how long it has waited. A timeout set
  explicitly on a workflow step still applies, and while a run waits, the same
  workflow's next scheduled run is skipped and says so.
• Truthful outcomes - a denied action ends the turn with your answer
  instead of letting the model try another way, its step reads as skipped
  rather than spinning, and tool results tell the model whether approval was
  needed, given or denied. Stop withdraws every approval the turn was waiting
  on.
• Survives restarts - a pending approval comes back with its card after a
  restart, and conversations waiting for one are marked in the sidebar.
  Ctrl+Enter approves only outside text fields.

Design, Code, Terminal And Computer Use Panels

• Design panel - a design opens in its conversation's Design panel: a
  canvas with a page strip, Edit with an inspector that uses font and logo
  pickers instead of raw fields, drafting from a brief in one step with
  "Drafting · …" progress, add and delete pages with Undo, a Size menu (16:9,
  4:3, 1:1, A4, 9:16 phone) that re-fits every page, full-screen Present,
  Review with per-issue Fix and "Fix all safe issues", brand colours and fonts
  "From a website", and Duplicate.
• Export and publish - Export is four format buttons; on this computer one
  click saves into the workspace's Exports folder with Open and Show in
  folder. Publish asks once, then Share shows the link with Copy, Open, a QR
  code drawn locally, and Unpublish.
• Developer panel - a code folder opens in its conversation's Developer
  panel with a status strip (folder and branch) and Changes, Files, Run and
  Git tabs; it opens on Changes when there are some and reads the folder
  afresh each time. Shell operators such as && and | are explained before
  a command is sent, Stop all processes ends everything the folder started,
  and a missing or signed-out GitHub CLI shows the Connect GitHub card.
• Custom tools - Settings › Tools lists custom tools with add from a
  folder, Test, on and off, and remove; test commands that need approval use
  the standard approval card.
• Interactive terminal - in the desktop app, a real terminal docks under
  the conversation: keystrokes, colours, history, tab completion, Ctrl+C and
  Stop, a resizable height remembered per device, Ctrl+` to show or hide it,
  and a full-screen sheet on phones. "Open in your terminal" opens your own
  terminal app at the code folder without Row-Bot's keys in its environment.
• Computer use card - a Computer Use turn shows "Using your computer" with
  the app, the latest picture (kept in memory only), Pause to take over,
  Resume and Stop.

Workflows

• Run in one click - Run starts a workflow at once and the run drawer
  follows it live until "Run finished"; New workflow opens on the page, like
  Edit; Overview offers "Run again" for a failed run.
• Clearer editor - "Send results to" is In this app plus a checklist of
  your configured channels, prefilled from your defaults; agent profile and
  model pickers; Duplicate (the copy starts switched off, without a
  schedule); typing {{ suggests date variables and earlier steps' results;
  leaving with unsaved changes asks first; a one-off time that has already
  passed is refused.
• Webhook triggers - the trigger shows its address with Copy and the
  secret sent as the X-Row-Bot-Webhook-Secret header, plus an explicit
  "Make reachable from the internet" that asks first.
• Honest run history - runs that a previous process left unfinished are
  settled as stopped at start-up instead of reading "running" for months.

Knowledge And Memory

• Knowledge graph - Home › Knowledge draws the graph with WebGL and lets
  it settle visibly, shows up to 2,000 memories by default and up to 5,000
  with Show all, and searches the whole library.
• Everything about memories in one place - status and type filters, the
  Review queue, bulk select and delete, full details, Add memory and the
  activity logs moved into Knowledge; Settings › Memory keeps the settings.
• Reliable recall with parallel agents - recalling a memory no longer
  counts as a change to it, so agents recalling at the same time keep using
  semantic recall instead of falling back to word search, and recalls no
  longer rewrite wiki vault files.
• Readable wiki vault names - articles keep readable <Subject>.md names
  under an ownership manifest, so the app only writes or removes articles it
  created. A renamed memory's article is renamed with it, a clash gets a short
  suffix instead of overwriting anything, and the vault status reads quickly
  on large vaults.
• Document processing model - documents are processed with the "Model for
  documents" chosen beside the queue rather than the last conversation's
  model.
• Clearer relation editing - vague relation types are refused with an
  explanation before anything is saved.

Monitor, Insights And One-Click Fixes

• Checks that run and are kept - Monitor's local checks run shortly after
  start and every 15 minutes; connection checks (Ollama, the Google, X and
  GitHub sign-ins, and internet reachability) run hourly while "Check
  connections every hour" is on, and whenever you ask. Results are kept with
  their time, so tiles no longer fall back to "Not checked".
• A fix for every problem - each warning or failure comes with one fix on
  Monitor's tiles, its detail drawer, Needs attention and Overview: restart a
  stopped channel, renew a Google or X sign-in, choose the default model, open
  the exact Settings row, or check again.
• Attention indicator - a sidebar indicator appears only when something
  needs you (a problem, an approval or an update) and stays quiet otherwise;
  an update offer can be put off with Remind me later.
• Honest Insights - review-only proposals offer no Apply, an applied
  proposal says what it did, and each insight records the model that found it
  and says when it may be out of date.

Devices And Remote Access

• Devices & remote access - the Settings page that replaces Remote Access.
  Connect a phone or computer detects the ways in (Tailscale, Same Wi-Fi, or
  the internet through your public link), shows one QR code backed by a
  one-time, 10-minute invitation that renews itself, waits for the device and
  then shows Connected with Rename and how to install Row-Bot on that phone.
• Your devices - each device with "This device", when and where it was
  last seen in words, Rename, and Sign out, which ends every session of that
  device.
• Advanced - where Row-Bot listens, one list of allowed addresses applied
  live, the Tailscale share with Stop sharing, the tunnel provider and token,
  and one Public line with the address, Copy and Stop.
• Sessions renew themselves - invited devices renew their session in the
  app and never revive a revoked one. Tailscale's own consent page is linked
  when it asks, a Serve conflict comes with next steps, and Same Wi-Fi is
  offered only while Row-Bot listens on the network.
• Tunnels never outlive Row-Bot - tunnels close on every exit, agents a
  crash left behind are stopped at the next start, agents Row-Bot didn't start
  are never touched, and Devices & remote access and Monitor report the real
  tunnel state.
• SMS and plugin webhooks through the tunnel - Twilio messages and plugin
  webhooks reach Row-Bot through the public link and are checked by their own
  Twilio signature or webhook secret; every other route still needs a
  session.

Connections, Plugins, MCP And Skills

• One connect sheet - Telegram, Slack, Discord, SMS, WhatsApp, Google, X,
  GitHub and plugin setup each use numbered steps with links. Channels add
  "Send a test message to me" (confirmed, one message to your own account),
  WhatsApp shows its live QR code and Reset session, a channel that needs a
  public address opens the tunnel and shows "Reachable at ", and X
  shows its callback address with Copy.
• Accounts that agree with Monitor - one GitHub status shared with
  Monitor, stored Check results, and Google's short-lived access tokens and
  X's expiry read correctly.
• Reviewed plugin changes - install, update, prepare and uninstall show a
  review (version, source, checksum, permissions and disclosures) and run only
  after you confirm. Marketplace downloads require the index's checksum and
  verify it, remote index entries download their source archive, plugins that
  run in their own worker get a reviewed Prepare step, and a plugin that fails
  to load says so.
• MCP servers - Add a server takes arguments one per line and environment
  values and headers as masked name and value rows; "Add and connect" saves,
  tests, accepts tools, turns the server on and connects, stopping at the
  first step that needs you. Enable in chat is on the MCP page, the Node.js
  runtime install recovers from an interrupted install, and saved servers
  connect reliably.
• Skills - public skill search answers quickly (GitHub keyword search uses
  the browse index, sources are read in parallel and partial results show as
  they arrive); a skill opens in its own dialog with Install, and new skills
  are available in chats by default. Skills created or edited in Settings
  reach the agent without a restart.

Settings And Your Data

• Settings regrouped - six groups (General, Models, Knowledge,
  Capabilities, Connections, System) with 19 pages and a search over pages
  and individual rows.
• Saved as you go, with Undo - every field saves at once and shows "Saved
  · Undo"; the Save and Revert buttons are gone except for typed credentials
  and editors such as a custom endpoint or MCP server. Unbinding a resource and
  dismissing an insight can be undone too, and Settings shows times, states
  and names in words rather than raw dates and ids.
• Back up and restore - Settings › Data › Back up now writes one zip to
  the workspace's Backups folder, without keys, sign-ins, sessions, caches,
  logs or runtimes, with webhook secrets and MCP headers and environment
  values blanked, and lists what to set up again. Restore from backup, in the
  desktop app, checks the archive (a foreign, newer or unsafe one is refused),
  applies it at the next start, keeps the current profile aside in
  before-restore-<time> and rolls back if anything fails.
• Export conversations - as Markdown or PDF. The PDF is printed offline in
  Chromium with scripts off and every request refused, or as a text PDF when
  Chromium isn't installed.
• Import from another assistant - Hermes Agent and OpenClaw data is found
  in its usual folder or picked in the desktop app, with Select all and Clear
  all.
• Tidier data folder - start-up removes temporary and splash files
  Row-Bot itself left behind, by exact name pattern and age only.

Reliability, Security And Privacy Fixes

• Webhook secrets - compared in constant time; a workflow whose saved
  secret is empty no longer runs for any caller; the secret can travel in a
  header instead of the address, while existing ?secret= addresses keep
  working and a request carrying two different secrets runs nothing.
• No secrets in logs - MCP server environment values and headers are no
  longer written to row_bot.log in plain text.
• Windows credential storage - when Credential Manager can't store more
  secrets, keys and sign-ins are saved as Windows-protected (DPAPI) records
  for your user instead of failing; Row-Bot never falls back to plain text.
• Windows upgrades - the installer replaces the application's source
  folder, so an upgrade never leaves a removed module behind.
• Stopping commands on Windows - Stop ends workspace commands started from
  a source checkout instead of leaving them running.
• Workflow fixes - double-encoded characters in new workflows and run
  titles are fixed, and a past one-off schedule no longer fires on save.
• Approval monitor - an agent approval that timed out no longer locks the
  database every minute.
• Windows file locks - provider settings, the wiki manifest and the
  document index retry while Windows briefly holds a file that was just
  written.
• OpenCode Go - requests no longer fail with "missing x-opencode-session":
  Row-Bot now names itself and sends the session id OpenCode Go requires on
  every model route. The id is a private value per conversation, never the
  conversation's own id. OpenCode Zen requests carry the same headers.
• OpenRouter attribution - OpenRouter requests, from chat and document
  processing alike, now identify Row-Bot as the app (its website address and
  name, as OpenRouter asks apps to) instead of LangChain, which lets OpenRouter
  show Row-Bot in its public app rankings. No user, prompt or account data is
  added.
• Dependency security updates - anyio 4.14.2 (fixing a critical
  advisory), soupsieve 2.10, and image-size 2.0.4 in the client toolchain.

Testing, CI And Release Pipeline

• One test pass per pull request - CI runs the deterministic suite once,
  sharded on Linux, beside static checks, the client checks, a Windows (Python
  3.13) and macOS platform lane and a Chromium browser smoke, with one
  CI / ci-ok result; slow tests run nightly.
• Nightly and weekly runs - the full suite with slow tests on Linux,
  Windows and macOS, the browser nightly set at desktop and phone sizes, a
  Linux package smoke and the docs reference check; installer verification
  and Firefox and WebKit smokes weekly.
• Release gate - a release refuses a commit without a green CI / ci-ok,
  runs the nightly suite only if that commit has no green nightly run, and
  each Windows and macOS build install-smokes its own package.
• Genuine tests only - a guard fails any deterministic test that reaches
  the network, a missing snapshot fails instead of being recorded, the strict
  app smoke requires / to reach the React app, and low-value and
  bookkeeping tests were removed.

Documentation And Website

• Public docs for the new app - the user guide describes the React app,
  with screenshots captured from it, and the generated reference pages read
  the client's settings pages and Home tabs directly.
• row-bot.ai - a redesigned landing page with product demos.
• Developer docs - AGENTS.md, CONTRIBUTING.md, the architecture,
  source layout and client platform guides describe the React client, the
  FastAPI server and the new test lanes.

Upgrade Notes And Breaking Changes

• NiceGUI is gone - the old interface and its routes are removed.
  --legacy-ui and --client-v2 are deprecated no-ops that log a warning and
  open the React app. Bookmarks to / keep working; addresses that only the
  old interface served open Home.
• Places that moved - designs and code folders are conversations with a
  Design or Developer panel (Home's Designer and Developer tabs and the
  separate studios are gone); the Activity Center is replaced by the
  attention indicator, Overview's Needs you and the conversation details card;
  agent profiles live in the sidebar's Agents dialog; Settings › Utilities is
  now Settings › Tools; Remote Access is Settings › Devices & remote access;
  Google is under Accounts; migration is under Data; goals are set in the
  conversation. Old Settings links redirect to the new pages.
• Memory editing moved - the memory list, Review queue and editing are in
  Home › Knowledge; Settings › Memory keeps memory on and off, extraction,
  graph health, the wiki vault and Delete all knowledge.
• Models after upgrading - a profile that finished setup on an earlier
  version keeps the models it was effectively using: the old presets it ran
  on are written once as its own choices. Values you saved are never changed.
  A new profile chooses its first model in the first run.
• Developer tools turned on once - the Developer tool was off by default
  in earlier versions and saved settings kept that default. The first start of
  5.0.0 turns it on once and says so ("Developer tools are now on · Settings ›
  Tools"); if you want it off, turn it off again and it stays off. Its actions
  keep their approval gates.
• Wiki vault tidy - the first sync after upgrading adopts existing
  articles in place, under the same names, so Obsidian links keep working.
  Vaults written by pre-release builds with entity-<hash>.md names get one
  readable article per memory, and the hashed copies move to
  raw/.row-bot-retired/hashed-names-<date>/. Nothing is deleted, and
  hand-edited old articles are left alone and listed for review.
• Approvals and goals - new approvals wait until answered (existing
  pending approvals keep their expiry); new goals have no turn limit (existing
  goals keep theirs).
• Webhook workflows - a webhook workflow whose saved secret is empty no
  longer runs; save its trigger again to generate a secret. Prefer sending the
  secret in the X-Row-Bot-Webhook-Secret header.
• Hourly connection checks - Monitor checks connections hourly by default:
  Ollama, the Google, X and GitHub sign-ins, and internet reachability (a
  connection to 1.1.1.1 on port 53). Turning off "Check connections every
  hour" in Monitor stops them, along with the periodic account sign-in checks.
• Plugins - marketplace plugins without a checksum in their index can no
  longer be downloaded; a plugin that runs in its own worker and reads "Needs
  preparing" needs Prepare once.
• Reverse proxies - the app uses no WebSockets; keep the event stream
  (/api/v1/events, text/event-stream) unbuffered. The Caddy example drops
  stream_close_delay. The old interface's internal routes (/_nicegui…,
  /_media/, /api/voice/local, /api/client-error) are gone; scripts use
  the authenticated /api/v1 API.
• Dropped features - plain-text conversation export (Markdown and PDF
  remain), Realtime voice diagnostics, the Designer's zero-state quick
  actions, its separate command palette (⌘K lists design commands) and its
  references list (the conversation's attachments are the references).
• Running from source - the app serves the built client from
  frontend/dist; build it once with Node.js 24.15 or later and npm 11
  (npm ci then npm run build in frontend/). Installers and Docker images
  already contain the built client.
• New files in the data folder - launcher.lock (held while Row-Bot runs),
  system_health.json (Monitor's kept checks), backup state, and
  before-restore-<time> folders after a restore.

Known Issues And Deferred Work

• Subscription sign-ins will change - the ChatGPT subscription sign-in
  will be reworked, and Claude subscription is planned to move to the Claude
  Agent SDK using your own Claude Code sign-in, in later releases; 5.0.0 keeps
  the current sign-ins. Image generation through the ChatGPT subscription is
  not offered, because OpenAI's route for outside apps doesn't support it.
• Stop and delegated agents - stopping a turn doesn't stop the delegated
  agents that turn started; stop them from the details card's Agents.
• No system notification for approvals yet - a waiting approval shows in
  the app, as a floating notice and in Buddy, but not as an operating-system
  notification when the window isn't focused.
• Goal cost - the goal card shows tokens but not cost for pay-per-use
  models.
• Private GitHub clones - cloning a private repository in conversation
  with your connected GitHub account comes in a later release.
• Buddy Hatch - Hatch doesn't show its model and number of calls before
  generating, and a missing image model gives a generic error; both come with
  the Buddy revamp after this release.
• Wiki vault recovery folder - the vault's recovery folder keeps every
  retained version and candidate link; there is no pruning yet.
• Voice - Talk shows a generic message for transcription failures such as
  a missing ffmpeg (Dictate names the problem), and voice mode doesn't speak
  the combined answer of an agent orchestration.
• Smaller items - renaming a custom endpoint shows it unchecked until its
  next check, and a workflow run that asks for a second approval after an
  approved step can read "running" while it waits.

schema: 1
files:
  Row-Bot-5.0.0-Linux-x86_64.tar.gz: sha256=e98bfa7cd2a12d27bf189d2b7d62d393bd700c7fab3f458ea7765bff0602db4a
  Row-Bot-5.0.0-Windows-x64.exe: sha256=8a0b6d99346c2c17eede503f74c89f9d29cfb2426b76965d0cd10c0834fb067e
  Row-Bot-5.0.0-macOS-arm64.dmg: sha256=8275014fa15fb42bcd023fc3cb65a4e1b284d45c08b8b7929015614ee621b769