* feat: add antigravity_cli integration harness
Register the `antigravity_cli` target, resolve its configuration
directory to `~/.gemini/antigravity-cli/`, wire target actions,
and add lifecycle hook scripts for Unix and Windows.
* feat: support session resume for antigravity-cli
* fix: correct antigravity-cli hook config path, shape, and agent label
Antigravity CLI reads global customizations from ~/.gemini/config, keys
hooks.json by hook name, and only accepts the matcher/hooks wrapper for
the tool events. The previous install wrote event arrays at the top level
of ~/.gemini/antigravity-cli/hooks.json, which agy never reads and would
reject anyway, so no Herdr hook ever ran.
Session reports were dropped for a third reason: the hooks reported the
agent as antigravity-cli, but the server normalizes that to the canonical
label agy before matching, so is_official_agent_source never matched and
agy --conversation resume was unreachable.
- Resolve the config dir to ~/.gemini/config
- Nest Herdr entries under a Herdr-owned "herdr" block that install
rewrites and uninstall removes, leaving other named hooks untouched
- Emit grouped entries for PreToolUse/PostToolUse and flat handler lists
for PreInvocation/PostInvocation/Stop
- Report the canonical agy label from both hook assets and match it in
agent_resume
- Compile-gate the hook asset constants and stop shadowing $args in the
PowerShell hook
- Document the real directory, that it must already exist, and the
named-block behavior
refs #1011
refs #1571
* fix: make antigravity-cli integration session-only
antigravity cli cannot express lifecycle safely: there is no blocked
event, postinvocation is skipped on interruption, and stop fires at the
end of a turn rather than on process exit, which left stale state and
released authority at the wrong time.
report only the conversation on preinvocation and let screen detection
own agent state. resume via `agy --conversation <id>` is unchanged.
* fix: update antigravity session after conversation switch
---------
Co-authored-by: Can Celik <ogulcancelik@gmail.com>
point_is_final_atom compared an atom's terminating column against a
word-motion target's starting column. The two are equal for narrow
cells but never for wide ones, so the read window stopped expanding
and copy-mode `e` halted at the soft-wrap boundary on CJK text.
Co-authored-by: Can Celik <ogulcancelik@gmail.com>
* fix: report a clear cli error when no herdr server is running
socket cli commands surfaced a raw io::Error debug string
(`Error: Os { code: 2, ... }`) when nothing was listening on the
api socket, which read like a bad --cwd path. map dead-socket
connect failures to a `server_not_running` json error carrying the
resolved socket path, printed once at the edge that surfaces the
error; recovering callers (plugin offline registry fallback, agent
start polling) recognize the marker and keep their existing
behavior.
refs #1941
* fix: map dead-socket errors on the unchecked cli request path
refs #1941
* fix: make server-not-running guidance session-aware
refs #1941
Enable termination-signal handling so SIGHUP and SIGTERM follow the graceful
quit path and restore terminal state. Log handler registration failures and
add PTY coverage for direct SIGHUP and server EOF restoration
Co-authored-by: Matt Coles <macols@amazon.com>
handle_pane_died only clamped the stale active/selected indices after
removing a workspace that closed itself, so it never shifted them down
when the removed workspace sat before the focused one. With workspaces
[x, y, z] and focus on y, removing x left active pointing at z, silently
moving the user into a workspace they did not select.
Resolve both indices by workspace id across the removal, as
close_selected_workspace already does, keeping the clamp as the fallback
for when the removed workspace was the focused one. Also scroll the
focused workspace back into view, which the explicit-close path already
did and this one did not.
refs #1621
Co-authored-by: Can Celik <ogulcancelik@gmail.com>
herdr tab close on a workspace's only tab returned tab_close_failed,
while the TUI close-tab action on the same tab closes the tab and
cascades to closing the workspace. Cascade in the API too, emitting
tab.closed then workspace.closed. When the cascade would close a whole
worktree group, return confirmation_required instead, matching
pane.close and the TUI prompt.
refs #1760
* feat: switch IME to ASCII in prefix mode on Windows
Extend switch_ascii_input_source_in_prefix to Windows: detect IME state
via WM_IME_CONTROL and toggle to ASCII with SendInput for the Korean IME.
Other IME languages are detected and left unchanged.
refs #1802
* docs: clarify Windows (Korean IME) support for prefix ASCII switching
Follow-up to the Windows prefix ASCII switch, which extended
switch_ascii_input_source_in_prefix to Windows but left the surrounding
text describing macOS only.
- Settings row label: "(macOS)" -> "(macOS/Windows)", plus its test.
- Generalize the Windows behavior wording to "switches the IME to
English (ASCII) input" in the config-reference.json entry and the
model.rs doc comment, and state that Windows support is currently
limited to the Korean IME.
- Add a :::caution aside noting the Korean-IME-only limitation to the
en/ja/zh preview docs, keeping the three translations in sync.
refs #1802
* fix: warn when prefix IME toggle injection is incomplete
SendInput can queue fewer events than requested (for example when blocked
by UIPI), which would otherwise leave the IME silently stuck in ASCII or
native mode. Log a warning with the sent and expected counts so the failure
is diagnosable.
refs #1802
* fix: guard prefix IME restore and correct open-status naming
Address maintainer review on the Windows prefix ASCII switch:
- Restore only when safe: store the foreground window at switch time and,
before toggling back, require that the same window is still focused and
the IME is still in English. This stops restore from toggling another
application or overriding a manual return to Hangul during prefix mode.
- Report injection success from send_vk_tap and only arm restoration when
the initial toggle landed; on a partial injection retry the dropped
key-up so a key is not left logically held down. The injection policy is
split into send_vk_tap_with(inject) so the success, partial, and failure
branches are unit-tested without the real SendInput.
- Correct the WM_IME_CONTROL sub-command naming: 0x0005 is IMC_GETOPENSTATUS
(IME open state), not IMC_GETCONVERSIONMODE. Behavior is unchanged; the
constant, comments, and tests now describe the open-status read.
refs #1802
* fix: bound IME open-status read with a timeout
The prefix IME switch reads IMC_GETOPENSTATUS via WM_IME_CONTROL, which
crosses into the terminal-emulator process. A plain SendMessageW would block
herdr's client thread until that process responds, indefinitely if it hangs.
Read through SendMessageTimeoutW with SMTO_ABORTIFHUNG and a short timeout
instead; on timeout the switch is skipped and restore does nothing, so a hung
host process can no longer freeze prefix-mode entry or exit.
refs #1802
* fix: preserve windows ime restoration after partial injection
refs #1802
---------
Co-authored-by: Ogulcan Celik <ogulcancelik@gmail.com>
akbash-bot