mlx-model-porting/ follows the shared Agent Skills format: one SKILL.md plus optional scripts/, references/, and assets/. Use the same folder across clients; do not fork the technical corpus per agent.
Fresh repository checkouts are pre-installed for Claude Code and .agents-root clients through checked-in relative symlinks at .claude/skills/mlx-model-porting and .agents/skills/mlx-model-porting. On Windows, Git may materialize symlinks as text files depending on local configuration. Native copy mode is intentionally unavailable because the attested installer requires POSIX no-follow directory APIs. Use WSL for a manifest-attested copy, or point the client directly at the checked-out mlx-model-porting/ directory; Developer Mode symlinks are also viable when the client supports them. Do not replace the checked-in link with an unverified recursive copy.
Use the documented client preset first:
python3 mlx-model-porting/scripts/install_skill.py --client codexUse --dest when you need explicit control over a version-specific or user-scoped root:
python3 mlx-model-porting/scripts/install_skill.py --dest PATH_TO_CLIENT_SKILLS_ROOTinstall_skill.py --client uses the repo-scoped defaults below and prints the resolved destination and mode before acting. --mode can override a preset mode when a client installation requires it.
| Client | Preset | Repository root | Recommended mode | Notes |
|---|---|---|---|---|
| Claude Code | claude-code |
.claude/skills |
symlink |
Common builds use repository/user .claude/skills roots; verify with the product’s skill listing command for user-scoped installs. |
| OpenAI Codex | codex |
.agents/skills |
symlink |
User-scoped installs use ~/.agents/skills; pass that with --dest. |
| Cursor | cursor |
.cursor/skills |
copy |
Version-dependent discovery; verify through Cursor’s Skills UI/installer. |
| Gemini CLI | gemini |
.gemini/skills |
copy |
Common builds use .gemini/skills; verify discovery before relying on it. |
| Windsurf | windsurf |
.windsurf/skills |
copy |
Version-dependent discovery; verify through Windsurf’s Skills UI. |
| GitHub Copilot | copilot |
.github/skills |
copy |
Keep scripts subject to workspace trust and review in VS Code/Copilot. |
| Google Antigravity | antigravity |
.agents/skills |
symlink |
Shares the Codex .agents/skills workspace root; a fresh checkout is already linked there. User-global scope lives at ~/.gemini/config/skills — pass it with --dest; do not assume a legacy Gemini path. |
After installation, start a fresh agent session and ask:
List the loaded skill named
mlx-model-porting, then state its version and the first four non-negotiable rules without running any model code.
Expected version: 0.7.0. The response should mention static intake, source pinning/oracle, and no optimization before parity.
Compare an installed copy against the repository VERSION file without modifying anything:
python3 mlx-model-porting/scripts/install_skill.py --client codex --checkIt prints current, stale, or missing, and exits 0 only when the installed copy matches the repository VERSION file (non-zero when stale or missing). The same --check flag works with --dest.
- The shared skill is the source of truth.
- Agent-specific files may explain discovery/permissions only; they must not duplicate runbooks.
- If a client cannot execute Python scripts, the agent may read them as deterministic specifications and reproduce the steps with available tools.
- Network and shell permissions remain client-controlled. The skill never assumes pre-approval.