Agent Skill Sync

agent-skill-sync

skillsync — scan, classify and sync AI-agent skills (SKILL.md files) across toolchains.

If you use several AI coding agents (OpenAI Codex, Claude Code, WorkBuddy, …), you end up with skill libraries scattered across different directories, in different layouts, some copied, some stale, some requiring dependencies you never installed. skillsync turns that mess into one honest inventory — automatically.

Zero dependencies. Pure Python ≥ 3.11 stdlib.

What it does

┌──────────────┐   ┌──────────────┐   ┌────────────────┐
│ ~/.codex/    │   │ marketplace  │   │ ~/.claude/     │   … any number of
│   skills/    │   │ plugin cache │   │   skills/      │       source roots
└──────┬───────┘   └──────┬───────┘   └───────┬────────┘
       └──────────────┬───┴───────────────────┘
                      ▼
              ┌───────────────┐     classify every skill:
              │  scan + parse │     A portable · B missing deps
              │  frontmatter  │     C rewrite needed · D platform-private
              └───────┬───────┘
                      ▼
              ┌───────────────┐     diff against target:
              │ INDEX.md      │     identical · drifted · missing
              │ inventory.json│
              └───────┬───────┘
                      ▼
              ┌───────────────┐     optional, safe copy:
              │ skillsync sync│     dry-run first, backups, never deletes,
              └───────────────┘     skips .venv/.env/node_modules

The A/B/C/D migration taxonomy

Category	Meaning	Action
A	Portable — no platform-private references, every referenced command exists	copy & use
B	Portable but missing external deps (a CLI, an MCP server)	install deps, then copy
C	References another agent's conventions (AGENTS.md, image_gen, …)	rewrite for your target
D	Bound to the source platform's private runtime	don't migrate

Classification is heuristic, conservative, and every decision ships a reason — the index shows why a skill landed in each bucket, and which dependencies are missing (probed live with shutil.which / your [deps] table).

Install

pipx install agent-skill-sync   # recommended: isolated CLI install
pip install agent-skill-sync    # or into your environment

# from source, without installing:
git clone https://github.com/kina-cmd/agent-skill-sync && cd agent-skill-sync
python -m skillsync.cli --help

Usage

# What do I have, and in what state?
skillsync status

# One line per skill: category, sync state, source, missing deps
skillsync scan
skillsync scan --category B          # only the ones needing deps
skillsync scan --json                # machine-readable inventory

# Generate a full index (INDEX.md + inventory.json) you can commit
# or paste into a "tool reuse" skill for your agent to read
skillsync index --out ./output --lang zh

# What changed upstream since I last looked?
skillsync check          # diffs against the last snapshot, then records a new one
skillsync check --quiet  # summary line only (cron-friendly)

# Sync portable skills into the target root — plan first, always
skillsync sync --dry-run
skillsync sync --categories A
skillsync sync --categories A,B --force    # update drifted copies (backs them up)

Sync safety rules

• Never overwrites without --force; a forced update first moves the old copy to <target>/.skillsync-backup/ — nothing is ever deleted.
• .venv/, .env, node_modules/, __pycache__/ are always excluded from copies, so secrets and 500 MB virtualenvs never travel silently.
• --dry-run prints the exact plan and touches nothing.

Configuration

By default, skillsync auto-discovers the well-known roots on your machine — every major adopter of the Agent Skills open standard:

Root	Platform
~/.codex/skills + ~/.codex/plugins/cache	OpenAI Codex (+ marketplace plugin caches)
~/.claude/skills	Claude Code (reference implementation)
~/.copilot/skills	GitHub Copilot
~/.gemini/skills	Gemini CLI
~/.cursor/skills	Cursor
~/.config/opencode/skills	OpenCode

Target defaults to ~/.workbuddy/skills. Roots that don't exist are skipped silently.

Override or extend with skillsync.toml (looked up in ./ then $XDG_CONFIG_HOME/skillsync/):

[target]
label = "workbuddy"
path  = "~/.workbuddy/skills"

[[sources]]
label = "codex"
path  = "~/.codex/skills"

[[sources]]
label = "codex-marketplace"
path  = "~/.codex/plugins/cache"
glob  = "**/skills/*/SKILL.md"

[[sources]]
label  = "codex-system"
path   = "~/.codex/skills/.system"
system = true              # platform-private → always category D

[deps]
# Teach the classifier that a dep is satisfied even if not on PATH:
voicebox = { kind = "path", value = "~/App/Voicebox/voicebox.exe" }
ffmpeg   = "ffmpeg"                                   # shorthand: check PATH
notion   = { kind = "env",  value = "NOTION_TOKEN" }  # check env var

See skillsync.example.toml for a complete example.

Typical workflow: keeping an index skill fresh

Many people maintain a hand-written "local tool reuse" index for their agent. It rots the moment a skill is added upstream. Instead:

skillsync index --out ~/.workbuddy/skills/local-tool-reuse/generated --lang zh

…on a schedule or a git hook, and let the agent read a generated file that is always true. inventory.json is stable, machine-readable output for further tooling.

Development

python -m unittest discover tests -v

Design notes

• No PyYAML. The frontmatter parser implements exactly the YAML subset skill files use (scalars, folded/literal blocks, inline and block lists), and fails soft — unparsable lines become warnings, never crashes.
• No network. Everything is local filesystem inspection.
• Qualified names. Marketplace plugin caches nest skills as <plugin>/<version>/skills/<name>/; these are reported as plugin:name so collisions are visible.

License

MIT © 2026 kina-cmd