Antigravity Automation Skill

The skill automates the Antigravity Electron app via OpenCLI, focusing on local automation. Most commands are read-only or require explicit confirmation for writes (delete needs --yes, revert needs --yes), reflecting least privilege. However, reliance on CDP debug port could be a risk if the port is exposed; the app launches with remote debugging enabled, potentially accessible by other local processes. Data flow transparency is moderate; SKILL.md does not detail data handling. No malicious behavior detected. Deductions: lack of more explicit permission isolation and user confirmation mechanisms.

Code includes good error handling, with clear error messages and postcondition verification for uncertain operations like delete. Tests cover key paths such as delete fail-closed and model ambiguity. However, the skill depends on Antigravity UI DOM structure, which could break with UI changes; CDP control may be performance-sensitive. Static review cannot execute, limiting score. Deductions: fragile UI selectors; static assessment cannot verify runtime.

The skill clearly targets the Antigravity desktop app with clear scenarios (automating chats, extracting code, managing conversations). However, it lacks declaration of non-fit ranges or environment constraints like OS compatibility or network dependencies. The skill is primarily for Chinese users but description is in English; environment fit is average. Deductions: missing explicit boundary and trigger conditions.

SKILL.md provides clear feature overview, requirements, and basic examples, but lacks detailed parameter documentation, troubleshooting, and known limitations. The repo has versioning (package.json) and CI, but no changelog or version history. Code is well-organized with tests and comments. Deductions: lack of comprehensive docs and update path.

The skill enables core tasks like sending messages, extracting code, and switching models, likely effective for automation. But static review cannot verify output quality; code extraction may include extraneous content. There is marginal value, but cost/benefit not quantified. Deductions: no execution evidence, output quality unverified.

There is a test file and CI, but tests cover only some commands and logic, not end-to-end behavior. Static review cannot independently verify actual behavior; evidence is limited to code inspection. Deductions: lack of external validation and runtime results.