antigravity-implant

agent
Guvenlik Denetimi
Uyari
Health Uyari
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 6 GitHub stars
Code Uyari
  • process.env — Environment variable access in antigravity-implant/cli/doctor.mjs
  • process.env — Environment variable access in antigravity-implant/hooks/lib/checkpoint.mjs
  • process.env — Environment variable access in antigravity-implant/hooks/lib/mesh.mjs
  • network request — Outbound network request in antigravity-implant/hooks/lib/mesh.mjs
Permissions Gecti
  • Permissions — No dangerous permissions requested

Bu listing icin henuz AI raporu yok.

SUMMARY

Make Google Antigravity work like Claude Code / Codex: talks while it works, shows evidence, rewindable (hooks · skills · live side panel) / Antigravity 를 Claude Code·Codex 처럼

README.md

Antigravity Implant

English | 한국어

Make Google Antigravity work like Claude Code and Codex. An Antigravity plugin that turns an agent which silently fires tool calls into one that talks while it works, answers your questions first, shows evidence mid-task, and tells you what it verified before it stops. It does not patch the app — you just drop one plugin folder in.

The agent-facing texts (hook reminders, skills, rules) are written in Korean, the author's working language. Antigravity models follow them fine; translate rules/AGENTS.md and the skills if you prefer English output.

Who is this for

  • You are tired of the Antigravity agent "thinking" in private while the screen only shows a stream of tool calls.
  • You edited GEMINI.md mid-task and it was ignored until the turn ended.
  • You are used to Claude Code / Codex features like /review, /rewind, todo lists and progress updates.

What changes

Pain point Antigravity Implant
Works silently Before each tool call: a progress line (NN% …) plus one sentence on what it is about to do
No evidence mid-task After a screenshot it is told to show it; the side panel collects captures in real time
Rule edits mid-task are ignored Your edits to GEMINI.md / AGENTS.md are injected on the very next model call
Ignores you while busy Replies to mid-task messages first; send messages or ⏸ pause from the panel
Says "done" without checking If files changed but the final answer says nothing about verification, it is asked once to wrap up properly
No undo Automatic git checkpoint on every message → /rewind or a panel button to restore / undo
Runs dangerous commands Wiping a drive, formatting, etc. are always blocked; force-push, reset --hard, registry edits only pass when you asked for them
Misses syntax errors .js/.json/.py files are syntax-checked right after an edit; errors are reported on the next call
Asks permission for every file outside the workspace Once you allow a file, its folder is remembered so you are not asked again
Auto mode still asks before many commands Opt-in: when you approve a command that had to run outside the terminal sandbox, the folders it named become writable for the sandbox, so the next commands there run without asking
Panel looks out of place The panel follows the app's theme (Settings → Appearance) live, or pick one of 9 panel themes and your own accent color

What's inside

Hooks progress reporting · mid-task messages first · live rule reload · loop / stuck detection · context warnings · wrap-up check · early-stop detection · dangerous-command guard · post-edit syntax check · folder-level permission memory · automatic checkpoints
13 skills /plan-first /code-review /rewind /remember /init-project /handoff /commit /pr /worktree /doctor /status-report /computer-use evidence-checkpoint
Live Cockpit panel progress · plan (task.md) · screenshots · git diff · rewind · live rules · hook log · tokens · mode · colors · send message · ⏸ pause
CLI status line, rewind, doctor

Full reference (settings, checkpoints, guard rules, limits) — Korean: antigravity-implant/README.md

Install (Windows)

Requirements: Antigravity, node 18+, git (for checkpoints).

git clone https://github.com/SEOBLEE101/antigravity-implant.git
cd antigravity-implant
powershell -ExecutionPolicy Bypass -File .\install.ps1 -StatusLine

Optional flags:

# Add the 10 "Readable Replies" formatting rules to ~/.gemini/GEMINI.md, auto-translated to match your settings/locale
powershell -ExecutionPolicy Bypass -File .\install.ps1 -StatusLine -ReadableReplies
# Pin a specific language (built-in: ko, en, ja, zh-CN, zh-TW, es, fr, de, pt; other BCP-47 locales fall back to English rules + reply-language instruction)
powershell -ExecutionPolicy Bypass -File .\install.ps1 -ReadableReplies -Lang en
  • Restart the app, then open the panel from the right-hand pane of a conversation: + → Extensions → Live Cockpit (once per conversation). The first time, check under Customizations → antigravity-implant → UI extensions → Live Cockpit that it says Running.
  • Health check: /doctor in a new conversation
  • Uninstall: install.ps1 -Uninstall (your data folder is kept; any GEMINI.md block added by -ReadableReplies is cleanly removed by marker)

install.ps1 validates the plugin in a temporary folder and then places it in %USERPROFILE%\.gemini\config\plugins\. It does not touch admin rights, the registry, permission allow-lists or the network. -StatusLine backs up the agy CLI settings and adds a single status-line entry. -ReadableReplies backs up ~/.gemini/GEMINI.md to a .bak file before making changes and manages only its marker block (<!-- antigravity-implant:readable-replies:start ... -->), resolving the language in order: -Lang → ~/.gemini/antigravity-implant/config.json → app config.json → existing GEMINI.md prose → ANTIGRAVITY_LANG / Windows display locale / LANG. Existing manual sections with the same heading are preserved untouched.

Optional: desktop app control (computer use)

Installs the open-source sandraschi/windows-computer-use-mcp (MIT) at a pinned commit. Not part of the default install.

powershell -ExecutionPolicy Bypass -File .\install-computer-use.ps1

It asks for confirmation, backs up mcp_config.json, and registers one server with keylogging, face detection, trajectory logging and clipboard reading turned off. Requires git and uv.

Fewer permission prompts in Auto mode

Why Auto mode still asks: commands run inside a terminal sandbox that can only write to the workspace and to folders you allowed. When a command needs more, the agent asks to run it outside the sandbox (BypassSandbox), and the app saves your approval as an exact full-command rule that never matches again.

What the plugin can do (opt-in): with "folderGrantsFromCommands": true in ~/.gemini/antigravity-implant/config.json, after you approve such a command and it succeeds, it adds write_file(<folder>\) for the existing folders named in the command line (at most 3). The sandbox then allows writes there, so the agent no longer needs to leave the sandbox for that folder. It is off by default because the command text alone cannot prove where the command really wrote. Folders reached through a junction, symlink or 8.3 short name are never widened, and drive roots, your home folder and the folders directly under it, AppData, dot-folders, system folders, folders of executables you ran, and odd spellings (trailing dots/spaces, ~1 short names, : streams, \\?\) are never widened. Granted folders are listed in ~/.gemini/antigravity-implant/folder-grants.json.

What it cannot fix: tools that do not exist inside the sandbox (for example a browser CLI) always need the bypass. If you want no prompts at all, that is the app's choice to make in Settings → General → Permissions → This Computer → Turbo ("Run all commands without restrictions") — a real security trade-off, so the plugin never changes it for you.

Colors

  • Antigravity itself: Settings → Appearance has 11 presets (Catppuccin, Dracula, Tokyo Night, Gruvbox Material, One Dark Pro, Solarized Dark, …) plus your own Background / Foreground / Accent for the dark and light themes, Contrast, Font and Font Size. The app stores them in ~/.gemini/config/config.json (userSettings.customThemeSeedsDark/Light); edit them in the app, not in the file — the running app does not pick up outside edits.
  • Live Cockpit panel: the 색 (color) menu at the top. Follow app (default) re-reads the app's colors every few seconds; or choose Catppuccin, Dracula, Tokyo Night, Gruvbox, Nord, Solarized Dark, One Dark, GitHub Light or Paper, and override just the accent with the color swatch. When a panel theme is selected, the panel shows the three hex values to paste into Settings → Appearance (click to copy) so the app can match it.

Safety

  • The dangerous-command guard only hooks run_command and never returns allow (your permission settings stay in charge). It also unpacks commands hidden inside cmd /c "…", powershell -c "…" and -EncodedCommand.
  • Checkpoints never touch your working tree, index or branches, skip secret-looking and large files, and never run filter / signing / hook programs from the repository's own config. Memory and time limits apply.
  • The panel binds to 127.0.0.1 only, requires a token for write requests, and rejects requests with any other host name.
  • Live rules changed while the agent was running commands are ignored (a command could have written them).
  • Reviewed twice end-to-end by another model (Gemini Carbon orchestrator swarm) and three times by an automated security review; all findings were fixed.

Tests

node tests/run-all.mjs

273 checks. Each run uses a temporary fake home folder, so your real ~/.gemini is never touched. cockpit-rewind-test needs the sidecar_sdk that the Antigravity app installs and is skipped otherwise.

Why it is built this way (verified by experiment)

  • All six Antigravity hook events (SessionStart PreInvocation PreToolUse PostToolUse PostInvocation Stop) really fire; one hook run takes about 0.1 s.
  • In PreInvocation, ephemeralMessage lasts for that call only while systemMessage stays in the conversation → per-call reminders are ephemeral; only things that must persist (rule changes) use systemMessage.
  • PreToolUse fails closed (empty output, errors, broken JSON all block the tool), and allow bypasses the user's permission settings → the guard hooks run_command only and defaults to ask.
  • PostToolUse injectSteps is ignored → syntax-check results are handed over through a file and reported by the next PreInvocation.
  • Stop with decision: "continue" works and the second stop carries continuationActive → the wrap-up check runs at most once per turn.
  • Without any rules, Gemini Flash almost always calls tools with no visible text; with the reminder it explains first every time.

Limitations

  • Tested on Windows only (Antigravity app 2.31, agy CLI 1.5.8).
  • Mid-task messages and pause never interrupt a running tool; they apply on the next model call.
  • Hooks can tell the model to show a screenshot it took, but cannot force it to take one. Captures that do exist always appear in the panel's Evidence tab.
  • The panel has to be opened once per conversation from the right-hand pane's + menu.

License

MIT. The panel uses the sidecar_sdk that Google ships with the Antigravity app; it is not included in this repository (the installed copy is used). Antigravity is a trademark of Google; this project is not affiliated with Google.

Yorumlar (0)

Sonuc bulunamadi