@toragonite/agent-mesh-gemini

The Google Antigravity adapter for agent-mesh. It drives Antigravity's official agy CLI headlessly, under your own login, behind the frozen agent-mesh Adapter contract — so a Fleet can route a task to Antigravity exactly the way it routes to Claude Code or Codex.

Unofficial. This project is not affiliated with or endorsed by Google. It shells out to the agy binary you already have installed and authenticated; it never impersonates the service, shares credentials, or circumvents rate limits.

Install

npm install @toragonite/agent-mesh @toragonite/agent-mesh-gemini

You also need the agy CLI installed and logged in (Antigravity keeps its credentials in the Electron app's data directory — there is no CLI login/status subcommand).

Run and resume

import { Fleet } from '@toragonite/agent-mesh';
import { GeminiAdapter } from '@toragonite/agent-mesh-gemini';

const fleet = new Fleet().register(new GeminiAdapter());

// Antigravity answers a one-word prompt in ~12–15s; set timeoutMs accordingly (see Latency).
const res = await fleet.run(
  { prompt: 'Reply with exactly: MESH-OK', timeoutMs: 60_000 },
  { policy: { prefer: ['gemini'] } },
);
console.log(res.status, res.conversationId, res.text);

// Continue that conversation by id.
const more = await fleet.resume('gemini', res.conversationId, 'Now say it backwards.', {
  timeoutMs: 60_000,
});
console.log(more.text);

You can also use the adapter directly:

const agy = new GeminiAdapter({ binary: 'agy' /* default */ });
const res = await agy.run({ prompt: 'hello', model: 'gemini-3.6-flash-high' });

Stream

const agy = new GeminiAdapter();
for await (const ev of agy.stream({ prompt: 'Write a haiku about routing.', timeoutMs: 60_000 })) {
  if (ev.type === 'text') process.stdout.write(ev.text);
  else if (ev.type === 'usage') console.error('usage', ev.usage);
  else if (ev.type === 'done') console.error('\ndone', ev.result.status, ev.result.conversationId);
  else if (ev.type === 'error') console.error('error', ev.message);
}

Streaming maps agy --output-format stream-json to agent-mesh RunEvents:

• The stream's init event only records the conversation id.
• Only an agent_response step with a non-empty text_delta becomes a text event. Other step types (user_input, checkpoint, …) are skipped — the contract has no event for them, and their per-step usage is partial, not authoritative.
• A delta's single trailing newline is held back and re-emitted only when more text follows, then dropped at end of stream. This makes the concatenation of all text events exactly equal the newline-normalized done.text (interior newlines are content and are never touched). Earlier versions passed deltas through raw, so a consumer concatenating them ended up with one extra trailing newline versus done.text.
• The terminal payload emits a usage event (the authoritative total) followed by a done event carrying the same RunResult a non-streaming run() would produce. The terminal payload is recognized whether it arrives wrapped in a result event or as a bare object carrying status/response (the error path and older CLI generations emit the bare form as the last line) — both are mapped by the same mapper.
• The done event's note carries every note run() would produce for the same request (allowedTools ignored, unknown model, dropped extra.args, AGY_HOME, vendor error text).
• An empty prompt throws synchronously at the stream() call, before any process spawns.
• If the stream ends with no terminal payload: a timeout yields a done result with status: 'incomplete' and a timeout note (never an error event); a non-zero exit yields an error event whose message includes the exit code and the child's stderr tail.

Models

availableModels() returns a static catalog and spawns nothing, because Fleet.route() calls it on the routing hot path. There is no auth-mode gating — every entry is selectable under one Google login.

| id | latency | note | | --- | --- | --- | | gemini-3.6-flash-high | fast | | | gemini-3.6-flash-medium | fast | | | gemini-3.6-flash-low | fast | | | gemini-3.5-flash-high | fast | | | gemini-3.5-flash-medium | fast | | | gemini-3.5-flash-low | fast | | | gemini-3.1-pro-high | slow | | | gemini-3.1-pro-low | slow | | | claude-sonnet-4-6 | fast | resold — selecting it spends Anthropic quota, not Gemini's | | claude-opus-4-6-thinking | slow | resold — selecting it spends Anthropic quota, not Gemini's | | gpt-oss-120b-medium | fast | resold — selecting it spends the provider's quota, not Gemini's |

The latency classes are a provisional pin — a best guess at interactive-vs-deep-reasoning behaviour, not measured. No entry is marked as the adapter default: when a task names no model, run() omits --model and the CLI applies its own default, so claiming a default here would be false. The model is trimmed before it is used, and a model that is only whitespace is treated as absent (no --model, no note). A non-empty model id that is not in this catalog still runs (with the trimmed value), carrying a model <id> not in known catalog note on the result.

Resold-model warning. The three resold entries are reachable through agy, but each one bills the other vendor's quota. Routing another vendor's quota through the Gemini adapter defeats the point of a multi-vendor fleet — prefer registering that vendor's own adapter.

Static vs live

• Static (availableModels()) — fast, hermetic, no subprocess. Use it for routing.
• Live (fetchLiveModels()) — the authoritative but slow path. It runs agy models, skips the Fetching available models... chatter line, parses each TAB-separated id\tlabel row, and applies the same latency heuristic. Any failure (spawn error, non-zero exit, timeout, unparseable output) resolves to []; it never throws.

import { fetchLiveModels } from '@toragonite/agent-mesh-gemini';
const live = await fetchLiveModels(); // ModelInfo[] — or [] on any failure

Auth

authStatus() runs agy models (the only auth probe Antigravity exposes — there is no readable local credential file and no status/login subcommand). Exit 0 with at least one parseable model line reports { loggedIn: true, mode: 'google' }; anything else — non-zero exit, missing binary, timeout, unparseable output — reports { loggedIn: false }. It never throws and never returns a detail (the CLI exposes no account identifier).

This costs a network round-trip and is therefore deliberately not on the routing path.

Quota

quota() returns null unconditionally, and capabilities.quota is false. Antigravity exposes no usage endpoint that has been verified. This is a deliberate "unknown headroom" signal: the Fleet treats null as "headroom unconfirmed" and still considers the adapter, rather than being told an exhausted account has full headroom. It is not a stub to be filled in casually.

Behaviour notes

• Trailing-newline normalization. agy terminates its printed answer with a newline the other agent-mesh vendors do not emit. For cross-vendor comparability, run()/resume() and the stream's done result strip exactly one trailing newline (\r\n or \n) from the final text. Interior newlines are preserved, and a second trailing newline is kept. Streaming text deltas hold a single trailing newline back (see Stream) so their concatenation equals the normalized done.text.
• resume() conversation id. The id is trimmed before validation and before it reaches the CLI (a blank or flag-shaped id still throws before any spawn). If the vendor echoes an empty conversation_id, resume() falls back to the trimmed id you passed, since that conversation is still resumable. run() does not do this — there an empty id genuinely means the vendor returned none.
• resume() working directory. ResumeOptions has no cwd field (core is frozen), so resume() reads a string opts.extra.cwd and passes it as the child process's working directory. A missing or non-string value inherits the parent working directory as before.
• usage.totalTokens. The vendor-supplied total_tokens is used verbatim when present. When it is absent but both inputTokens and outputTokens are numbers, totalTokens is their sum; otherwise it is omitted.
• extra.args filtering. Only string entries are forwarded to the CLI. When one or more non-string entries are dropped, the result carries an extra.args contained N non-string entr(y|ies); dropped note rather than dropping them silently.
• allowedTools is ignored. agy has no equivalent flag, so task.allowedTools is dropped and the result carries an allowedTools not supported by agy; ignored note.
• account.configDir → AGY_HOME. There is no documented config-dir env var for agy. When you pass account.configDir, the adapter sets AGY_HOME to it and adds a note that Antigravity account isolation via this variable is unverified. An explicitly passed account with no configDir means the ambient login and does not inherit a default account's directory. A null or otherwise non-object account is treated as absent — it falls back to the default account (never a raw error).
• Latency. A one-word answer is observed at 12–15 seconds (versus ~5s for Claude Code/Codex); agy's own --print-timeout defaults to 5m0s. Set timeoutMs generously — a timeout resolves status: 'incomplete' (it never throws), so too tight a cap silently truncates real answers.
• No secrets. No token or credential value appears in any result, note, detail, error, or raw. There are no tokens on this path, but the rule holds regardless.

License

MIT © Toragonite