Added
- Five more agent backends —
kimi,grok,opencode,piand Google's
Antigravity — each tracking status and launching, resuming and (where the
agent can) forking, with its own known limits listed in the README. - An iTerm2 terminal backend (macOS, iTerm2 ≥ 3.0): hand launches bind to
their window,Enterfocuses,Dcloses andmiao focusworks, and previews
come back as plain text since iTerm2 reports no colour. - The resume picker (
r) works on Reasonix, Kimi, Grok and opencode. - Token and model columns for Kimi, opencode, Pi and Antigravity, taken
straight off the hook payload where the agent knows the numbers. Space mopens a log of the footer's status messages (last 200, in
memory), so a message that scrolled past is still readable.- The detail panel names a session's worktree when it has one.
Changed
- Agents launch through one subcommand,
miao launch <agent> [dir]— the
per-agent spellings (miao claude,miao codex, …) are gone rather than
aliased. - An agent that logs in for the first time inside a captain-miao session keeps
its credentials, which move into your real agent home on the next launch
instead of staying invisible in the synthetic one. - The columns ahead of Status are four cells tighter: the follow-up bell is
now a one-cell dot, the override indicator folds into the status cell, and the
icon column is fixed-width, so nothing shifts when a host connects or a wider
mark scrolls into view. - A session running in another terminal instance no longer gets a glyph in the
icon column — the row still dims, and the preview and detail panel both name
the instance it lives in. - A hand-typed
miao launch <agent>in a bare Ghostty window is refused,
namingo/Oinstead, since Ghostty exports no per-surface id to bind such
a session to its window. - The agent cycle key offers only backends that are installed, instead of
cycling through seven and failing at launch. fhides itself on a backend whose resume cannot branch, and says why on a
session with no known id, rather than offering a key that does nothing.- The hosts row editor draws as a card over the list, wraps long values
instead of truncating them, and carries the clipboard offer as a field rather
than a separate key. - The dashboard no longer implies a signal its backend cannot send —
s
names the backend that has no approval prompt to report, and a context column
with no number readsn/arather than blank.
Fixed
- Ending or restarting a session closes its window after the process in it
exits, instead of tearing the window down around a live process — which on
Ghostty raised a confirmation dialog on the doomed tab. - A Claude session started from the dashboard no longer inherits the launching
session's environment, whoseCLAUDE_CODE_CHILD_SESSIONmarker turned off
the transcript captain-miao folds for status, title and context tokens. - Ghostty could not launch a session at all, from two faults found driving a
real Ghostty 1.3.1: the command carried ashell:prefix only the config file
accepts, andnew tabwas asked for without naming a window. miao focusworks on Ghostty, where the dashboard could not tell which
surface it was in and so printedNo dashboard runningwhile one was.- The header paw and its cat no longer draw on iTerm2, where inherited
KITTY_*variables had the dashboard believe kitty was painting its cells. - opencode sessions no longer read as "working" forever — the generated
plugin registered six handler keys opencode never looks for, and fixing that
also turned on the session id, title, tool name, and token and model columns. - Kimi hooks never fired at all, because the
matcher = "*"every other
backend spells "every tool" fails to compile as the JS regex Kimi expects. - An opencode approval prompt now reaches the dashboard, which had subscribed
an event that exists only in opencode's drifted SDK types. - An opencode subagent no longer wears the parent's row, whose status,
tokens, model and resume id it used to overwrite mid-turn. - Pressing Esc in opencode no longer parks an error on the row — the abort it
publishes beside the idle events is the turn ending, not failing. - A first-ever Kimi session's transcripts and login land in the real
~/.kimi-code, not inside the synthetic home where the dashboard never saw
them. - A plugin or hook deleted from the real config dir no longer leaves a dangling
symlink in the opencode and Grok mirrors for the agent's loader to trip on. direnv's allow-status is checked on every backend, not only Claude, where
a blocked.envrcelsewhere produced no agent and no visible reason.- The resume picker no longer answers
opencode not found in PATHin place
of "No resumable sessions" on a machine that never installed it. - Closing a remote session's window takes its row with it, instead of leaving
it up for a second reading as attached in another terminal and offering to
steal a session nobody holds. D, and a closed window underon_window_close = "detach", no longer offer
to steal the session they just put down.- Editing one row in the hosts panel reconnects only that host, instead of
re-dialling every box and losing every remote row for the multi-second storm. - A host still fetching its sessions keeps its loading line, instead of
briefly presenting a host that had told us nothing as a host with nothing on
it. - A host that settled on a musl server is offered musl again after the
dashboard's version moves, and a refused deploy retries rather than leaving the
host on its old server. - A background shell that outlives a compaction shows through
Compacted,
instead of parking the row there for as long as the shell runs.
Full Changelog: v0.4.0...v0.5.0