Port AI skills to Google Antigravity tool syntax

Skill Porter previews conservative tool-name translations and copies complete local skill bundles for manual adaptation to Google Antigravity, for a locally obtained Claude Code, Cursor, Codex or generic agent skill bundle. It preserves support files and previews limited tool-name substitutions, but does not itself prove client compatibility.

Its tool-name translation table, applied only to exact backtick-quoted identifiers in the body (frontmatter untouched): View becomes view_file, Edit and StrReplace both become replace_file_content, Write becomes write_to_file, Bash becomes run_command, Grep becomes grep_search, Find becomes find_by_name, LS becomes list_dir, WebSearch becomes search_web, and Fetch becomes read_url_content.

Its safety design, enforced in the script itself: symbolic links and non-regular files (pipes, devices) are rejected at every path component; input is bounded to 1,000 files and 20 MiB total; source and destination trees must not overlap; every bundle in a multi-skill source is validated (frontmatter present and closed) before any file is written; and writing to an existing skill destination directory is rejected rather than merged or overwritten. These boundaries are exercised directly: dry-run writes nothing, a copy preserves binary support files and untouched metadata byte-for-byte outside the tool-name substitution, an existing destination is left untouched with its own files intact, a symlinked source or destination is rejected, a FIFO is rejected as non-regular, a remote URL is rejected, overlapping source/destination trees are rejected, and a byte-budget violation is caught before any write occurs.

Its prerequisites: Python 3.9+ with only the standard library, a reviewed local skill directory containing SKILL.md (or a repository with skills/<lowercase-hyphenated-id>/SKILL.md directories), and verification of the upstream identity, pinned revision, license and complete bundle beforehand - remote material must be obtained separately through a reviewed download or clone workflow, since remote URLs are rejected outright. Use stable local directories you control: the skill is marked risk: critical and is explicit that the script is not a sandbox against concurrent hostile filesystem changes, so it must not be run against a tree being modified by another process or user (inspect scripts without executing them).

The workflow: preview the local bundle with python3 scripts/port_skill.py --source "/absolute/path/my-skill" --dry-run, reviewing a diff where only exact backtick-quoted tool identifiers change and frontmatter and support bytes stay intact; copy to a fresh staging destination with --dest, which rejects any existing destination and writes no global paths; for workspace installation, run with --workspace (dry-run first) to write only .agents/skills/<id> under the current working directory; and manually review context-file references, tool argument shapes, client configuration, dependencies, licensing and multi-agent ordering, testing actual client invocation before claiming compatibility.

Use it when adapting a locally obtained Claude Code, Cursor, Codex or generic agent skill bundle for Google Antigravity. This is conservative text adaptation only - no AST conversion, semantic optimization, automatic parallelization, artifact generation or compatibility certification - so context references and support scripts may still need manual adaptation, and it performs no downloads, credential handling, network calls, global installation, plugin manifest generation or external MCP setup. A disk or filesystem failure may leave a partial destination with no overwrite or rollback mode, so inspect it and choose a fresh output directory to retry. Never point it at a directory that another process or user may be changing while it runs.

It suits developers adapting a locally obtained Claude Code, Cursor, Codex or generic agent skill bundle for use in Google Antigravity.