v0.3.0

• Starts eleven times faster. 158 ms down to 14 ms. A launch used to read the
  system's certificate store and wait on the update check before it showed you
  anything; it now does neither. On a slow connection that wait could run into
  seconds, and it is capped at two.
• A smaller download. 63 MB down to 39 MB, and every update is a fresh
  download, so it is 24 MB off each one.
• waibee serve has a UI. Open the printed URL: every live session, its
  transcript, plan and sub-agents, in a browser. Nothing to install — the page
  ships inside the binary.
• Approve tools from the browser. Permission prompts reach the web UI and
  answer back; before this the web surface silently trusted everything.
• Review what the agent changed, in the browser. A Review tab per session:
  changed files with counts, the diff, and select-lines-and-comment sends your
  note straight to the running agent.
• Pick up a session where you left it. Past sessions are listed by folder
  and open with their transcript, plan and panel intact.
• See what the agent is doing while it works — the tool, the elapsed time,
  and a Stop button.
• A Setup tab: every instruction file, skill, command, agent, plugin, MCP
  server and tool this session is loaded with — with switches for plugins and
  MCP servers.
• Context fill sits next to the model, as a percentage of its window.
• Nine themes, five dark and four light.
• ⌘K searches everything: your sessions, the ones on disk, the actions, the
  themes.
• Type / to see what you can run — your slash commands and skills, with
  their descriptions.
• Rename a session by clicking its name.
• Drag any pane to resize it, or fold it away. A folded pane leaves a strip
  to click, and the layout is remembered.
• The agent's side panel works in the browser. Plans, dashboards and mermaid
  diagrams open beside the conversation, the way they do in the terminal.
• Sessions are named in the browser, and show what the agent is doing right
  now instead of just the folder they run in.
• Answer a permission with one key: y once, Y this run, p always, n deny, N never. Each button shows its own key.
• Files and a terminal in the browser. Browse the session's tree and open a
  real shell in it, from the same page.
• Switch model, permission mode or compact from the browser, mid-run.
• Sessions started on the web now run your hooks and skills.
• The agent sees the right filename again. The first file in its
  "uncommitted changes" list was missing a character.
• Skills load now. Waibee listed your skills but shipped without the tool that
  opens one.
• Sub-agents know the project: your CLAUDE.md, skills and environment,
  instead of a blank slate.
• Sub-agents reach your MCP servers, including a named agent that lists one in
  tools:.
• A sub-agent's plan stays its own. It no longer takes over your progress
  panel or returns after /resume as yours.
• Sub-agents stay out of your session: no side panel, no questions, no saved
  memories, no worktree switches.
• Task(capability_mode=…): read-only or read-write. execute and all
  are gone; omit the argument for the full toolset.
• Sub-agents also gained lsp, notebook edits, and reading or stopping their own
  background commands.
• The interface no longer freezes. A command the agent ran was handed your
  keyboard, so anything that read input swallowed your keystrokes. In PowerShell the
  console also started echoing its own mouse reports on screen as ^[[M.
• Copying works on macOS, and only says it worked when it did. A selection
  announced "copied" while the clipboard kept whatever it held.
• Open it on your phone. serve --host 0.0.0.0 prints a link and a QR code, and
  the browser shows the same code. From another network, one button opens a public
  address through cloudflared, tailscale or ngrok, whichever you have. A second
  link watches without acting.
• waibee storage clean asks before deleting, not after. Answering n used to
  print "Aborted." over an empty directory.
• Roughly thirty ways past the permission gate are closed, including a newline
  after an approved git status, and git reset --hard or git push --force read
  as the commands that merely list.
• Reading a file gives you the file. A 100 KB file came back in four pieces and
  was fetched again after every summarisation.
• Edits stop failing on invisible whitespace: a tab where spaces were sent, a
  trailing space, a different indent.
• A sub-agent cannot be given more than you have. An agent definition could set
  its sub-agent's permission mode without that being weighed against your session's,
  so a file in the repository asking for bypass got it from a session that never had
  it. It may now only narrow. Bypass needs a git repository, and that was checked only
  in the mode picker - the command-line flag and the settings file walked past it.
• waibee update --force reinstalls. It promised to install even when you are
  on the latest version, and did nothing but reword the reply.
• waibee plugin install <directory> works on a plain directory. It needed a
  git init first, because every source went to git clone, and the command's own
  help offered a local path. Updating one copies the directory again, so a plugin you
  are writing picks up its own changes.
• A command no longer fails with an error about a line you never wrote. Commands
  ran inside a wrapper of ours that restored the environment first, so a one-liner sat
  ninety lines into it, and anything printed by a backgrounded job could be taken for
  an environment variable and break the wrapper's quoting: unexpected EOF while looking for matching ', blamed on "line 90". Once that happened, every later
  command in the session failed the same way. On Windows with a shell other than Git
  Bash, the working directory is read correctly too.
• What the agent remembers about you stops piling up. The same fact came back
  every session in new words - one store held three things written eleven times, and
  all eleven went into the first message of every session. Before it writes, the agent
  is now shown what it already remembers, and asked for one fact per line. Your own
  notes outrank its wording of the same thing instead of being refused as a duplicate.
  Facts already saved are left as they are: guessing which of two similar ones to
  delete is not a risk worth taking with your memory.
• The REPL cannot hang a session, and its answers no longer carry the
  interpreter's own prompt: REPL(language="node") answered > 2 for 1+1.