feat(agy): support Google Antigravity CLI as a first-class agent

[jiunbae]

Why

The Gemini CLI is EOL upstream; Google replaced it with the Antigravity CLI (agy). Because agy keeps the ~/.gemini directory and still runs Gemini models, muxa's existing Gemini support looks like it carries over. It doesn't — and every mismatch fails silently.

	Gemini CLI (gemini)	Antigravity CLI (agy)
Hook config	hooks key in ~/.gemini/settings.json	its own ~/.gemini/config/hooks.json
Events	SessionStart, Before/AfterAgent, Before/AfterTool, Notification, SessionEnd	SessionStart, Pre/PostInvocation, Pre/PostToolUse, Stop
Payload	snake_case, session_id	camelCase protojson, conversationId

Confirmed against a live install: hooks written to settings.json make agy log loaded 0 named hooks from 0 hooks.json file(s) and carry on, and the Gemini adapter's required session_id would reject an agy payload outright. Meanwhile pane_current_command is agy, which classify_command didn't recognise — so agy panes weren't even discovered.

What

AgentKind::Antigravity as a separate kind, not a rename — both CLIs still work and can be installed side by side.

• Adapter (adapters/antigravity.rs) covering all six events.
• muxa hook agy CLI handler.
• Transcript reader (adapters/antigravity_transcript.rs). No agy payload carries prompt or response text — they carry a transcriptPath. PreInvocation and Stop read it, mirroring what Claude Code's Stop hook already does here.
• muxa init --component agy-hooks writing ~/.gemini/config/hooks.json, plus a matching muxa doctor check.
• Bundled agy screen manifest, which supplies the one state agy's hooks cannot — see below.
• Wired through everything else that enumerates agents: discovery, muxa agent start --agent agy, the watch spawn form, muxa timeline, MCP muxa_start_agent / muxa_call_peer, the dashboard kind filter, the herdr bridge, and the omp sink.

Docs: docs/ANTIGRAVITY.md, plus updates to README (en/ko), INSTALL, SCREEN_DETECTION, SINKS, MCP, TIMELINE, HERDR, CONFIGURATION, ARCHITECTURE, DASHBOARD.

Four findings that shaped the design

1. muxa hook agy must be fail-open and byte-silent.
agy reads a hook's stdout as a verdict. Empty stdout means "no opinion"; JSON without a valid decision blocks the call, and so does a non-zero exit. I hit this by accident during probing:

Encountered error in tool execution: tool call denied by pre-tool hook

So this handler writes zero bytes to stdout and exits 0 even on an unparseable payload, an unknown event flag, or a down daemon.

That guarantee was then being defeated from outside the handler: main() loads config.toml with ? before dispatching, so one TOML typo made every hook exit non-zero and blocked every tool call in every agy session. muxa hook … now degrades to defaults and reports the parse error on stderr; every other subcommand still fails loudly. (Released 0.8.34 → exit 1; this branch → exit 0.)

2. invocationNum is per-turn, not per-session.
Captured across a two-turn session: turn one ran invocations 0,1,2,3, turn two reset to 0,1. So PreInvocation at 0 is the turn boundary and the only place a prompt is reported; later invocations emit a Heartbeat (which also refreshes model — agy stamps modelName on every payload, more than Codex or the Gemini CLI give us).

3. Hook-authoritative precedence needed its first carve-out.
agy fires no hook when it raises an approval prompt, so an agy row could not reach WaitingInput at all. And because muxad skips screen inference on any pane a live hook row owns, installing muxa's agy hooks actively removed the one signal the manifest could have given. Hooks-or-attention is the wrong trade for the single most important thing muxa reports.

New AgentKind::hooks_report_attention() keys the exception. Claude Code, Codex, the Gemini CLI and opencode all return true and are completely unaffected; agy returns false, and for those rows the detector contributes exactly one signal — applied to the real row:

Screen says	Row is	Result
Blocked	not waiting, not Error	→ WaitingInput
Idle	waiting	→ Idle (releases a stuck wait)
anything else	—	nothing

Three exclusions, each a bug if reversed: no Heartbeat (it would overwrite the hook-supplied model), Working is left to the hooks (milliseconds vs a seconds-late tick), and no synthetic row is minted — Store::apply evicts synthetic rows from a pane a real row owns, so one would flap in and out on every hook event. Refined panes are also never tracked, since that set drives the stop-sweep and a real row isn't ours to stop.

4. Seeing the real prompts fixed the manifest.
agy's command prompt renders numbered > 1. Yes / 4. No rows — the label-based patterns would have missed it had option 2 not happened to contain "always allow". Now anchored on Requesting permission for:, the header agy prints above every request whatever the rows look like, plus the cursor-marked numbered row.

Testing

End to end against agy 1.1.17, isolated muxad, real tmux panes:

• Fresh session: idle → working → idle, with prompt, response, model, cwd, conversation id and tool events all landing on the row.
• Resumed session (agy -c, no SessionStart fires): lands on the same row via conversationId, prompts recovered from the transcript for both turns.
• Attention refinement, on both prompt shapes agy renders (numbered command prompt and file-creation prompt): working → waiting_input → approve → idle, with the hook-supplied model and response intact throughout.
• All four fail-open paths: valid payload, malformed JSON, daemon down, unknown event → exit=0, 0 bytes on stdout.
• muxa init --component agy-hooks → muxa doctor → --uninstall round-trip in an isolated HOME; doctor is silent when agy isn't installed and requires all six events when it is.
• muxa agent start --agent agy launches agy --dangerously-skip-permissions.

Suite: cargo test --workspace --no-fail-fast — everything green except upgrade::tests::dry_run_renders_plan, which asserts systemctl --user restart muxad and already fails on main under macOS. ~60 new unit tests. cargo clippy --workspace --all-targets introduces no new diagnostics versus main (client_kind_from_arg dead-code is pre-existing).

Note for reviewers: tests/e2e.rs locates muxad at target/debug/muxad, so on a tree that has only ever been release-built it fails to spawn and takes 10 tests down with it. cargo build -p muxad first.

Known gaps (agy-side, documented not worked around)

• No session-end hook — Stop is a turn boundary. agy rows are reaped by pane liveness, like Codex's.
• workspacePaths is empty in print mode, so cwd is absent for agy -p. A hook's own cwd is the hooks.json directory, so there's nothing to fall back to.
• No rate-limit or cost signal. modelName rides on every payload; the usage fields stay None.
• agy caches hooks in a long-lived backend process — editing hooks.json needs an agy restart to take effect. Called out in the docs.

🤖 Generated with Claude Code