v0.2.1

Added

• Startup check for the terminal control channel. Being in a supported
  terminal is not the same as being able to drive it. On Kitty, a remote-control
  setup that is missing listen_on or carries a mismatched password used to
  surface one failed action at a time, with kitty's raw error in the status line;
  captain-miao now proves the channel works before the dashboard takes over the
  screen, and a failure prints which half of the setup is wrong — plus the config
  block to fix it — while stderr is still visible.

Fixed

• Codex sessions on macOS track their transcript again. Codex rows showed no
  context usage, never picked up the first-prompt title, and stayed green forever
  after an Esc-interrupt: the rollout watch matched the wrong spelling of a
  symlinked path, and macOS reports nothing at all for the way Codex writes (one
  long-held file descriptor, appended to for the whole session). Context tokens
  and interrupts now land with the event that caused them, and the watch parks
  itself while a session is idle, so a session sitting at rest costs nothing.
• Codex no longer refuses to start with "local database appears to be
  damaged." The synthetic $CODEX_HOME now repairs entries where a real file
  sits where a link into your real ~/.codex belongs — what Codex leaves behind
  whenever it adds a new state file, which eventually split a SQLite database
  across the two homes.
• Review-pending detection for commands containing a quote. An r3 watch
  wrapped in something like nix develop … --command bash -c '…' was truncated
  when normalised, so the row sat at "Task" instead of "Review" — and every
  command wrapped that way collapsed onto a single entry in the learned
  long-running set.
• CLAUDE_CONFIG_DIR is honoured when locating Claude's home. An instance
  that relocates the agent's config dir no longer has the resume picker and the
  b browser listing every project from the real ~/.claude instead.
• nix build works again. The build's source filter dropped assets/, so
  the logo the dashboard embeds at compile time was missing and the build failed.

Changed

• README rewritten around installation and configuration, gathering the
  Cargo, npm, and Nix routes under one section and documenting every CLI
  subcommand and config key in tables.

Security

• The recommended Kitty setup is now scoped by an authorization script.
  Pairing the remote-control password with a small is_cmd_allowed script shuts
  the in-terminal escape-code channel outright — the vector by which a shell
  running inside a Kitty window, including one on the far end of an ssh session,
  could otherwise drive your terminal — and confines even an authenticated
  request to the eight commands captain-miao actually issues. The README carries
  the script, and now says plainly that substituting your own password is what
  makes that check real: the built-in default is a published constant and
  authenticates nothing on its own.

Full Changelog: v0.2.0...v0.2.1

v0.2.0

First public release.

(0.1.0 was never a usable release: its four per-platform binary packages reached
npm, but the launcher that resolves them did not, and no GitHub Release was ever
cut. 0.2.0 is the first version published as a complete set.)

Added

• The whole fleet at a glance. Every agent session in one table — status,
  working directory, model, context usage, git branch, and a live transcript
  preview — so you can tell at a glance which session is working, which is
  waiting on you, and which has finished.
• Claude Code and Codex. Both agents are supported side by side, and rows of
  either kind mix freely in one dashboard.
• Full session lifecycle from the dashboard. Launch, resume, fork, and kill
  sessions without leaving it, with a filterable picker over recent working
  directories and path completion for new ones.
• Status detection that distinguishes waiting from working. Sessions blocked
  on your approval or an answer are flagged and sorted to the top; s jumps to
  the next one needing attention.
• Background jobs split into Task and Server. A build or test the agent is
  waiting on keeps the session green, grouped with the active sessions, and your
  machine awake. A dev server or watcher the agent parked does not — it goes
  yellow, sorts with the idle sessions, and flags itself for follow-up. Common
  dev servers are recognised out of the box, and anything unrecognised that runs
  past an hour is learned, so it is classified correctly from the first moment
  next time.
• Review-pending detection for r3. When
  every background job a session is waiting on is an r3 watch, the agent is
  blocked on a human review rather than working, and the row says so.
• Kitty and zellij backends. captain-miao brings no terminal of its own — it
  drives the one you already run, so every session is a native window or pane.
• Switchable session layout. Stacked (all sessions in one shared
  cm:sessions tab, the dashboard being the switcher) or per-tab (one tab each,
  switchable with your terminal's own keys). Space l toggles it, and the choice
  persists.
• Keep-awake. Prevents the machine from sleeping while any session is still
  working (caffeinate on macOS, systemd-inhibit on Linux), with an indicator
  shown only while it is actively inhibiting.
• Pin, mute, and follow-up flags, plus per-directory icon and colour marks
  with a searchable emoji picker.
• Configurable keybindings. Every Normal-mode command is remappable from a
  [keybinds] table in config.toml, including the Space-leader ones.
• Prebuilt binaries on npm and GitHub Releases for macOS (Apple silicon and
  Intel) and Linux (x86-64 and arm64). npx @hyperlogue/captain-miao runs the
  matching native binary with no Rust toolchain and no runtime download.

Security

• Session state is written owner-only — 0600 files under 0700 directories —
  because state records your prompt text, working directories, and session ids.
  A tree created by an earlier build is tightened in place on the next run.

Known limitations

• Remote hosts over SSH are experimental and off by default. The full
  cross-host lifecycle is implemented but unverified against a real remote host,
  and restart and fork remain local-only. Build with --features remote to try
  it; without it the dashboard is strictly local-only.
• t (move window to tab) is unsupported on zellij, which has no CLI to
  reparent a pane across tabs. The key reports it and the help entry is hidden.
• Linux binaries are glibc builds (built against glibc 2.35, so Ubuntu
  22.04+, Debian 12+, RHEL 9+). musl/Alpine needs a source build.

Full Changelog: https://github.com/hyperlogue/captain-miao/commits/v0.2.0