feat(cli): add agy (Antigravity CLI) as an agent backend

[charliezong18]

What & why

Adds agy — Google's Antigravity CLI — as a first-class agent backend, so it can be driven from Happy (terminal + phone/web app) as a fallback when Claude Code hits its usage limit.

Additive: the existing Gemini integration is untouched.

Design

agy has no ACP / JSON streaming mode (unlike Gemini = ACP, Codex = app-server JSON-RPC). Its only stable non-interactive surface is agy --print "<prompt>", which streams plain text and persists a per-cwd conversation for resume. So it can't reuse AcpBackend — it follows the bespoke (non-ACP) openclaw pattern instead: a backend that translates its own output into Happy's AgentMessage stream and rides the same session pipeline, app-selectable without the ACP / MessageAdapter / agentRegistry machinery.

New packages/happy-cli/src/agy/:

• constants, types, cliArgs (buildAgyArgs), conversationStore
• AgyBackend — spawns agy --print per turn, streams stdout → AgentMessage
• runAgy — orchestrator mirroring runOpenClaw
• ui/ink/AgyDisplay.tsx — terminal status view

Key details:

• stdin EOF: agy --print blocks until stdin reaches EOF, so AgyBackend spawns with stdio: ['ignore', 'pipe', 'pipe']. An open stdin pipe hangs the turn forever (caught during testing).
• Permissions: Happy yolo / safe-yolo / bypassPermissions / acceptEdits → --dangerously-skip-permissions; otherwise --sandbox. (Print mode is one-shot and can't surface an interactive approval prompt, so non-skip modes defer to agy's own settings.json.)
• Resume: agy doesn't print the conversation id; it's read back from agy's per-cwd cache and passed via --conversation.
• Model: Happy's model string → --model "<display name>"; defaults to a Gemini model (the fallback exists because Claude is rate-limited).

Registration

'agy' added everywhere 'openclaw' already appears:

• happy-cli: detectCLI, index.ts subcommand + help, daemon/run.ts dispatch, AgentId, BackendFlavor, controlServer zod enum, registerCommonHandlers.
• happy-agent: SupportedAgent, SUPPORTED_AGENTS.
• happy-app: new-session picker (agentDefaults, persistence, picker + icon, settings labels, storageTypes cliAvailability, ops spawn unions, AgentInput).

Tests

agy/cliArgs.test.ts + agy/AgyBackend.test.ts (10 tests, incl. the stdin-EOF guard). Full happy-cli unit suite passes (691/691); happy-cli, happy-agent, and happy-app all typecheck.

Notes / follow-ups (not done here)

• The happy-app agy icon is a placeholder (copied from the Gemini icon) — wants a real asset.
• Retiring the Gemini integration is intentionally out of scope (this PR is additive).
• A richer integration via agy's embedded agentapi HTTP server (structured tool / permission events) is possible later, but it needs an undocumented running language-server — out of scope.

Visual proof

agy driven end-to-end from the Happy app, not just unit tests: a Happy session running the agy backend — online, answering a prompt and correctly identifying itself as Antigravity (Gemini 3.1 Pro). This shows the backend spawns, streams its output back into Happy's AgentMessage pipeline, and is usable from the app/browser.

---

Closes #1429
Refs #1313

[charliezong18]

hey @bra1nDump — thanks again for the quick turnaround on #1440! 🙏 whenever you've got bandwidth, #1430 (agy / Antigravity CLI backend) is also ready for a look. it's bigger (~1.2k lines) but fully additive — the gemini provider is untouched, it just registers agy everywhere openclaw already appears. tests pass, all three packages typecheck, and there's end-to-end visual proof in the description (driven from the app).

i wrote up the design rationale in #1429 (tl;dr: agy has no ACP/JSON mode, so it follows the bespoke openclaw pattern rather than the gemini ACP one). totally open to splitting it up or reworking the approach if the size/shape is a concern — just let me know what'd make it easiest to review.