Antigravity Engine

TaroCub drives the official agy CLI as a native local engine. The verified compatibility baseline is Antigravity CLI 1.1.24. Install and authenticate agy locally before selecting /engine antigravity.

Official references:

• Antigravity headless CLI
• Antigravity slash commands
• Antigravity CLI changelog

The 1.1.24 baseline was live-verified with consecutive turns on one persistent worker and a fresh-process --conversation resume. The public per-platform updater manifest advertised 1.1.24 while the bundled CLI changelog still ended at 1.1.23, so TaroCub treats the protocol probe, not an inferred changelog, as the compatibility evidence.

Runtime Contract

Ordinary turns share one persistent headless worker per live conversation, with both input and output in NDJSON stream-json format. TaroCub writes one user event per turn and keeps stdin open. It accepts one process-level init, the matching user echo emitted by agy, and structured step_update and result records for each turn. An unstructured stdout line, a mismatched user echo, an error result, or a process exit before the active turn's final result fails closed instead of being posted as assistant text.

The worker remains warm for later turns and is reaped after two idle hours. Changing a startup setting such as workspace, approval mode, model, effort, or native timeout recycles an idle worker and starts a replacement with --conversation <uuid>. A crash removes the worker; the next turn resumes the same authoritative conversation ID in a fresh process.

TaroCub maps the stream as follows:

Antigravity event	TaroCub event
init.conversation_id	session binding
agent-response text_delta	answer stream
active tool step	tool_use
terminal tool step	correlated tool_result
successful result	authoritative final answer
per-step usage	current-turn token accounting

The result-level usage of a resumed conversation is cumulative. TaroCub does not record that value directly; it sums the latest usage for each step in the current turn so resumed turns are not billed twice.

/goal is the one deliberate protocol exception. Antigravity does not accept CLI slash commands through stream input, so TaroCub first recycles the idle persistent worker, then invokes native direct -p prompt mode with /goal as the first prompt token while still parsing structured output. Private bridge guidance and attachments follow the goal text and are not allowed to hide the slash command inside an XML wrapper. The next ordinary turn starts a new stream worker attached to the same conversation.

Model, Effort, and Sessions

• /model <id> persists a single-token model ID and passes it as --model when the next worker starts. Changing it recycles an idle worker while preserving the conversation. Use agy models to discover the current IDs. /model off restores the CLI default.
• /effort supports low, medium, high, and off, matching the native startup flag. Incompatible persisted values are removed before startup.
• A new turn binds the conversation_id from the structured stream. A resumed turn passes --conversation <uuid>. /resume can still scan recent CLI logs because agy does not expose a structured conversation-list command.
• The normal runtime policy is a six-hour hard cap plus a 30-minute inactivity watchdog. /timeout off removes both bridge watchdogs but explicitly passes a seven-day native ceiling; omitting the native flag would silently restore agy's five-minute default.

Capability Matrix

Capability	Status	Boundary
Structured answer streaming	Supported	Native stream-json
Structured tool lifecycle	Supported	Tool input/output is correlated by step index
Conversation resume	Supported	Explicit UUID binding; log scan for discovery
Model and effort selection	Supported	Applied at process startup
Token usage	Supported	Per-turn step usage; no USD cost
Native /goal	Supported	Direct -p prompt mode, structured output
Files, media ASR, delivery tags	Supported	Shared TaroCub channel/delivery layers
Native MCP/plugins	Supported	Managed by Antigravity's own configuration
Per-tool remote approval	Not exposed	Headless stream input has no interactive control-response protocol
Structured questions	Not exposed	Turn-level approval is the safe bridge fallback
Mid-turn steering	Not exposed	Follow-ups queue as another turn
Post-result background lifecycle	Not exposed	A result ends the turn; the worker stays warm but emits no background-task lifecycle
Manual /compact or context telemetry	Not exposed	No official headless API currently available

For approval mode normal, TaroCub asks once before the turn and then grants the whole headless process inside Antigravity's sandbox. full-auto combines the skip-permissions flag with --sandbox; only explicit bypass omits the sandbox. This is not equivalent to the per-tool approval flows available from Claude Code, Kimi ACP, or DeepSeek Harness.

中文摘要

TaroCub 已按 Antigravity 1.1.24 的原生结构化协议接入：每个活跃 conversation 维持一个 stream-json worker，后续轮次复用同一进程；回答、工具、终态和 token 分开处理，恢复会话只统计本轮 step，避免累计 token 重复记账。空闲 两小时、进程崩溃或启动参数变化时会安全回收，并用权威 conversation ID 重建。/goal 因上游限制仍使用一次性直接 -p，执行前会回收持久 worker， 下一轮再恢复同一会话。full-auto 会同时启用 --sandbox；只有显式 bypass 才跳过沙箱。

对会输出 user echo 的 CLI 版本，TaroCub 以“与当前 prompt 完全匹配的 echo” 作为持久 worker 的轮次边界：新一轮 echo 之前迟到的 step_update 会被隔离， 不会污染下一轮文本、进度或 token；缺少预期 echo 的结果会 fail closed。

尚未对齐 Codex/Claude 的部分来自当前上游边界，而不是 bridge 伪装支持： Antigravity headless 暂无单工具远程审批、运行中 steer、结果后的后台任务生命 周期，以及手动 compact/context API。TaroCub 对这些能力明确显示为不支持。