v0.15.1

Fixed

• A session with a shell running for Claude now reads working instead of idle,
  so the board stops offering it as free. Claude Code reports shell while that
  shell runs, and vigie read the word as the operator having left for a ! prompt
  (#851).

v0.15.0

Added

• The usage strip shows the weekly limit that applies to a single model, when one
  is in force. Claude enforces it alongside the 5-hour and 7-day windows, and it
  can be the highest of the three — the board read comfortable while the limit that
  would stop the work was nearly full (#840).

Fixed

• A session that stops its own background command no longer stays working. Claude
  Code answers the stop but sends no completion notice for it, so the row stayed
  busy until the session ended, with no way to clear it (#842).

v0.14.3

Fixed

• A session watching a persistent Monitor no longer reads idle while the
  monitor is still running, so the board stops offering it as free. vigie
  recognised only one kind of background work — a Bash launched with
  run_in_background — and now reads what the call itself declares (#834).

v0.14.2

Changed

• The watcher now says so when Claude Code reports a session status this build
  does not know, instead of quietly showing those sessions as idle. A status
  renamed in a Claude Code release used to put the whole fleet at rest with
  nothing anywhere naming the cause (#822).

Fixed

• A session no longer stays working after its background command finished. When
  the command ended while Claude was still working, Claude Code delivered the
  "it is done" notice in a form vigie did not read, and the session kept reading
  busy until it closed — about one session in five on the board (#818).

• A session you have already answered no longer reads waiting while its command
  runs. Approving a permission prompt left the board showing the session as
  blocked on you until the command finished, and the indicator counted it among
  those needing attention (#816).

v0.14.1

Fixed

• A session babysitting a background command stays working when you queue the
  follow-up. Typing "when the CI is done, do X" used to put it at rest while the
  command was still running (#810).

• A session no longer stays working after a network outage swallowed the hook
  that ended its turn. When Claude Code itself reports the session at rest, that
  now clears the stale status; a guess made from a quiet transcript still does
  not (#803).

• The browser and the GNOME indicator no longer replace a fresh board with an
  older answer that arrived late, which could also announce a session twice
  (#805).

• Changing show ended or hide idle after in the browser redraws the board
  straight away, instead of leaving it contradicting the setting until the next
  refresh. Opening the Sessions tab redraws it too (#804).

v0.14.0

Changed

• The hidden-columns warning is behind i instead of standing above the session
  list. On a narrow terminal it wrapped onto two lines and cost the board three
  of its ten sessions, to repeat something already visible — the missing columns
  are absent from the header, and it is the operator who narrowed the window
  (#788).

Removed

• The remote-control column, its sort key and the rc filter token. The field
  vigie read to detect /rc is now set on every session whether or not it is
  active, so the column marked sessions nobody was driving — and there is no
  documented signal to rebuild it on (#789).

Fixed

• The terminal notices a dead event stream in seconds instead of waiting on it.
  After a suspend the connection can stay open and answer nothing; the board kept
  refreshing on its own poll, but live updates stopped until the kernel gave up
  (#793).

• A session the daemon reads as over now shows when it ended, instead of a dash.
  0.13.0 announced this for a session that ended with its process; it was true
  only where a report established the end, and not on the quieter path where the
  session simply stopped being reported (#792).

• A machine's card in the browser draws every session it counts. A session whose
  status the page did not recognise was in the total and in no pill, so the two
  numbers disagreed with nothing to say why (#791).

v0.13.2

Fixed

• The usage gauges can no longer go backwards. A machine whose turn to fetch had
  lapsed could still post the figures it had read minutes earlier, landing them
  on top of fresher ones (#774).

• A machine that loses the race for the usage lease is told so, instead of
  getting an error it cannot tell from a broken database (#779).

• A retention setting the daemon fails to read no longer deletes sessions. A
  failed read was treated as "nothing chosen", so it pruned on the default and
  then wrote that default over an operator's keep everything (#777).

v0.13.1

Fixed

• A report that carries no timestamp no longer lands its token growth on today.
  The day figures are never recomputed, so one attributed to the wrong day stayed
  wrong; the growth is now counted by the next report that names a day (#768).

• An end time recorded after a session was already over now reaches an open
  dashboard, instead of waiting for a manual reload (#765).

• An open session detail in the browser follows the session. It kept the state it
  was opened with, so a session you opened because it was waiting still said
  waiting after you had answered it (#764).

• A daemon restarted without VIGIE_TOKEN says so instead of re-keying the fleet
  in silence. It comes back on the token it had stored before the variable was
  adopted, so every machine is refused — and the failure used to look like a
  network or certificate problem (#759).

• n opens the session that is calling even when hiding ended sessions would
  keep it off the board. A call outlives the session that raised it, so a caller
  whose process had died was picked by the jump and hidden from it at the same
  time — and the panel opened on another session (#758).

v0.13.0

Changed

• A desktop notification says what is being asked of you, not what the session
  is. A session stopped on an API error was announced as is waiting in the
  terminal and as error in the browser; all three clients now name the reason,
  each in its own words (#742).

Fixed

• A request body over the size cap is refused as too large on every endpoint, not
  only on reports. The other four answered invalid json body, telling an
  operator their JSON was malformed when it was fine and the payload was simply
  too big (#740).

• A session waiting on a background command reads working, not idle. Claude
  Code answers the launch within seconds while the command runs on, so the turn
  looked finished and the board offered the session as free — an operator
  scanning for one interrupted a session that was about to resume by itself
  (#748).

• The site's status legend shows the colours the product actually uses, and
  includes error. It painted working orange where vigie paints it green and
  idle grey where vigie paints it blue, so the page taught a colour rule vigie
  does not follow (#738).

• A session that ended with its process now shows when it ended. Only a clean end
  announced by Claude Code stamped a time, so the case where nobody was there to
  close the session — machine shut down, terminal closed, Claude killed — reached
  ended carrying none (#739).

• A day's working, waiting and idle time can no longer be counted twice. The
  event log is the mark each interval is measured from, and it was written after
  the seconds and allowed to fail — so a failed write left the mark behind and the
  next event remeasured a span already counted, permanently (#737).

• n opens the session that is calling, even when a filter hides it. It jumped
  to the caller and then opened the detail panel on whatever row the cursor
  happened to be on, so the operator read another session's prompt, branch and
  last message believing they belonged to the one waiting on them (#736).

v0.12.0

Changed

• The Sessions bottom bar no longer marks the usage gauges when their refresh
  fails. The state pill already says so on every tab, and the state modal says how
  old the snapshot is and that it cannot refresh (#726).

Fixed

• A machine no longer disappears from the fleet for a quarter of an hour after
  waking from suspend. The watcher kept sending on a connection that had died with
  the machine's network state, and nothing noticed until the kernel gave up (#732).
• The state modal reads just now · cannot refresh for a usage snapshot that has
  only just failed, instead of just now old (#727).
• vigied token no longer prints a token the running daemon is not using. A
  daemon given its token through the environment left an earlier one in the
  database, and the command handed it over as though it were live (#720).
• A session that comes back no longer keeps the time it ended. Only reopening it
  cleared that, so a session resumed by typing carried a timestamp saying when it
  had stopped (#722).
• A turn stopped with Ctrl-C while Claude was thinking now shows interrupted.
  It read as still reasoning instead, and stayed that way until the next prompt
  (#721).
• The dashboard shows a status it does not recognise as it is, instead of drawing
  it as idle. A session stored under a status this build has retired was shown
  resting (#719).