cocoaskills 0.12.5

Local skill manager for AI agent skills with reproducible per-project installs

CocoaSkills

Translations: Русский. English is the source of truth.

csk is a local skill manager for AI agent skills. It installs reusable skill packages from git repositories into your project repositories with reproducible, content-hashed installs, skill-to-skill dependencies, and multi-agent support across six environments: Claude Code, Codex CLI, Cursor, and Gemini via adapter mirrors, plus OpenCode and Windsurf, which discover the canonical .agents/skills/ directory natively.

It is an independent Python implementation of the open Curator Protocol. The csk executable, package name, and existing state directories remain implementation-specific compatibility names; portable manifest and marker names follow the shared protocol.

Why

Managing agent skills across many projects by hand falls apart fast: drift between machines, no version pinning, README files and tests leaking into the agent context, no cleanup when a skill is removed.

CocoaSkills makes per-project skill installation declarative and reproducible:

• One Skillfile.json per project, committed to version control.
• Pinned git refs (tag / branch / revision) and content-hashed installs.
• Skill-to-skill dependencies: a skill declares the skills it builds on, and csk install resolves the transitive closure with exact refs and activation modes.
• A whitelist-based stripped layout: README, tests, build files, and other non-skill content stay out of the agent's context.
• One canonical location (.agents/skills/) with per-agent adapter symlinks or copies into .claude/skills/, .codex/skills/, .cursor/rules/, .gemini/skills/. OpenCode and Windsurf read .agents/skills/ natively, so they need no mirror.
• Skill-provided command shims exposed via a project-local .agents/bin/ directory on PATH.
• Optional global skills installed once under ~/.cocoaskills/global/ and exposed to supported agents outside any project checkout.

Install

Pick whichever fits your machine. pipx is the recommended path on every platform.

pipx (recommended)

pipx install cocoaskills

uv tool

uv tool install cocoaskills

Homebrew (macOS, Linux)

brew tap ivanopcode/csk
brew install cocoaskills

mise

mise use -g pipx:cocoaskills@latest

Convenience install script

curl -fsSL https://cocoaskills.org/install.sh | sh

The script detects Python, prefers pipx or uv tool, and falls back to pip install --user. Read it before piping if you do not trust the network.

Plain pip

python -m pip install --user cocoaskills

Quick start

1. Pick or create a directory for skill git repositories. Example: ~/agents/skills/. Existing local skill repositories are read from this directory; missing repositories can be cloned automatically when a skill declaration provides git.

2. Bootstrap the global config:

   csk bootstrap
   

   This writes ~/.cocoaskills/config.json with your skills_root, preferred locale, and default agents.

   Repository automation can make this step idempotent without overwriting a developer's existing machine config:

   csk bootstrap --if-missing --non-interactive --skills-root ~/.cocoaskills/skills
   csk upgrade .
   

3. Initialize CocoaSkills in each project:

   cd /path/to/project
   csk init
   

   This creates Skillfile.json and adds the CocoaSkills generated paths to .gitignore.

4. Declare which skills you want:

   {
     "schema_version": 1,
     "project": { "alias": "demo-ios" },
     "agents": ["claude_code", "codex_cli", "cursor"],
     "locale": "en",
     "skills": [
       {
         "name": "skill-tracker",
         "git": "[e-mail]:skills/skill-tracker.git",
         "tag": "v1.0.0"
       },
       {
         "name": "skill-metrics",
         "source": "internal/skill-metrics",
         "branch": "main"
       }
     ]
   }
   

   The optional locale field only affects skills that ship localized metadata (locales/metadata.json plus .skill_triggers/<locale>.md). Skills without localization files install unchanged.

5. Run csk install inside the checkout.

For multi-project sync, explicitly register projects with csk project add and run csk install --all or csk upgrade --all.

Skill dependencies

Since v0.9.0 a skill can require other skills (RFC 0007). A requirement lives in agent-skill.json schema v4 under dependencies.skills, is self-contained (git URL plus an exact tag or revision ref), and carries an activation mode:

{
  "schema_version": 4,
  "runtime_roots": ["scripts"],
  "capabilities": { "exec": ["trk", "git"], "network": "none" },
  "commands": {
    "report": { "type": "script", "unix_path": "scripts/report" }
  },
  "dependencies": {
    "skills": {
      "skill-tracker": {
        "git": "[e-mail]:skills/skill-tracker.git",
        "ref": { "kind": "tag", "value": "v1.4.2" },
        "mode": "runtime",
        "commands": ["trk"]
      }
    }
  }
}

Activation modes select what a provider contributes to the consumer:

• full (default) activates the provider prompt context and all exported commands.
• runtime activates commands only; the optional commands list narrows the activation to the named exports.
• context activates the provider prompt context only.

csk install resolves the transitive closure: providers are fetched, unified to one commit and one canonical source per name, ordered before their consumers, and audited together. Version conflicts, source conflicts, and dependency cycles fail with the full requirement chains.

A workflow ships as a skill that declares requirements and exports no commands; a consumer installs the whole composition with a single Skillfile.json entry.

Two supporting mechanisms:

• Skillfile.dev.json substitutes providers locally during development: a checkout path or a git ref, branches included. The file stays out of version control, installs print every active substitution, and strict audit refuses substituted installs.
• allowed_sources in ~/.cocoaskills/config.json lists canonical host/path prefixes and gates every clone. SSH and HTTPS URLs of one repository normalize to one identity.

Global skills

Global skills are user-wide baseline skills. They are installed under ~/.cocoaskills/global/ and linked into user-level agent directories such as ~/.claude/skills/ and ~/.codex/skills/. When OpenCode or Windsurf is among the target agents, global skills are also linked into ~/.agents/skills/, which both discover natively.

csk global init
csk global add skill-metrics \
  --git [e-mail]:skills/skill-metrics.git \
  --tag v1.0.0
csk global install

Global commands are exposed through ~/.cocoaskills/global/bin. During csk global install, CocoaSkills also publishes forwarding shims into a safe user bin that is already on PATH, such as ~/.local/bin, so global commands work from any directory without per-project activation.

Agent execution never depends on shell profile activation. Installed skills resolve project shims explicitly from <repo>/.agents/bin/<command> (<command>.cmd on Windows), then global shims from <csk-home>/global/bin, and only then a validated bare command. This contract works unchanged from zsh, bash, PowerShell, Git Bash, CI, and agent processes that were not launched from an initialized interactive shell.

Generated runtime shims prepend only the paths needed by the installed skill: the current project/global shim directory, the Python environment running csk, and directories of declared system command dependencies. The inherited PATH remains available, but skill-to-skill calls and Python launchers do not depend on a shell hook.

On Windows, PowerShell 5.1, PowerShell 7, and cmd.exe can all execute the generated .cmd shims directly. Optional directory-change activation is available for PowerShell and Git Bash; cmd.exe has no profile hook and does not need one for agent execution.

If no safe user bin is available, global install still succeeds and prints a warning. Agents continue to use the explicit global path. Humans can set CSK_GLOBAL_USER_BIN to a writable PATH directory or invoke the generated shim explicitly.

Shell activation is optional human convenience for bare project commands and project-over-global command shadowing. auto detects zsh or bash from SHELL, PowerShell on Windows, and Git Bash on Windows before the platform fallback:

csk shell-init --install
# Or choose explicitly: zsh, bash, powershell

The command atomically caches the hook and prints the correctly quoted source line for .zshrc, .bashrc, or the PowerShell profile. Never put eval "$(csk shell-init ...)" in a profile: that starts Python for every new shell. Run --install again after upgrading CocoaSkills so the optional cached hook receives fixes.

Set CSK_AUTO_ENV=0 before sourcing the optional hook to disable project directory scanning on an unhealthy or blocking filesystem. Global commands remain active; project commands remain available by explicit .agents/bin path. Global skills never replace committed project Skillfile.json declarations.

Hybrid skills

Hybrid skills are stored once per machine and activated for selected projects only, with nothing committed to the target repositories. The declaration lives in ~/.cocoaskills/hybrid/Skillfile.json and names its targets by project alias, absolute path, or path glob:

csk hybrid add skill-conventions \
  --git [e-mail]:skills/skill-conventions.git \
  --tag v1.0.0 \
  --target demo-ios \
  --target "/Users/me/work/*-service"
csk hybrid list

csk install in a targeted project picks applicable hybrid skills up automatically: the prompt context materializes once under ~/.cocoaskills/hybrid/skills/ and reaches the project through managed adapter links, command shims land in the project .agents/bin, and the dependency closure and audit gates apply exactly as for project skills. Shadowing order is project, then hybrid, then global. This scope fits skills a platform team rolls out to selected repositories when committing anything to those repositories is undesirable.

Skill command manifests

Skills declare commands, capabilities, and dependencies through agent-skill.json. Schema v2 supports multi-file runtimes: runtime_roots are copied into ~/.cocoaskills/runtime/<skill>/<commit>/ and excluded from agent prompt context. Schema v3 adds the capabilities envelope used by csk audit and strict install gates. Schema v4 adds skill requirements (see Skill dependencies).

Existing packages named csk-skill.json remain readable. New and updated packages should write only agent-skill.json. During a staged rename, both files may coexist only when their decoded JSON values are equal; conflicting files fail installation instead of selecting one silently.

{
  "schema_version": 4,
  "runtime_roots": ["scripts"],
  "capabilities": {
    "network": ["gitlab.example.com"],
    "filesystem": "repo",
    "exec": ["review-cli"],
    "secrets": "none",
    "env_read": ["HOME"],
    "prompt_scope": "Review merge request metadata and produce local advice."
  },
  "commands": {
    "mr": {
      "type": "script",
      "unix_path": "scripts/mr"
    },
    "review-cli": {
      "type": "system",
      "command": "review-cli",
      "hint": "Install the review CLI through project bootstrap tooling"
    }
  }
}

system commands are only checked with shutil.which; CocoaSkills never installs system tools, and manifests carry no install hooks or version probes.

Skill audit

csk audit runs security checks against the same committed skill snapshot that csk install would use. Static detectors always run. Optional command and codex backends extract additional structured findings; the install decision stays deterministic inside CocoaSkills.

csk audit
csk audit . --json
csk audit --global

Install gates are opt-in per command or through config:

csk install --audit
csk install --audit strict
csk global install --audit

Advisory audit prints warnings and continues. Strict audit blocks findings at or above the configured threshold. Schema v1/v2 skills declare no capabilities; strict audit requires migrating them to schema v3 or newer, or pinning the content hash through the trust workflow when that workflow is enabled.

Backend safety rules:

• Local command backends receive raw skill files and are treated as trusted local tools.
• Local codex backends require oss=true and an explicit local_provider.
• Cloud backends require audit.allow_cloud=true and a public source policy match. File contents are redacted before they are sent to a cloud-capable backend.
• Unverifiable backend findings are shown in reports and never block strict installs.

Audit registry

An audit registry serves signed statements that a skill, at a specific commit and content hash, was audited or revoked (RFC 0008). A machine pins the registries it trusts in ~/.cocoaskills/config.json:

{
  "audit_registries": [
    {
      "name": "internal",
      "url": "https://registry.example.com",
      "public_keys": ["ed25519:base64key..."]
    }
  ],
  "disable_builtin_registries": false
}

csk install resolves each skill against the trusted registries and verifies every record against the pinned keys before trusting it. A verified revocation in any trusted registry denies the install; a verified audit is recorded as an attestation in the install marker. Registry lookups are advisory unless a skill is revoked, and organizations pin only their internal registry with disable_builtin_registries. Signature verification uses a standard-library Ed25519 implementation, so the runtime keeps no third-party dependency.

Snapshot rollback and equivocation state is keyed by canonical registry URL under the configuration home (~/.cocoaskills/state/registry by default), outside the disposable response cache. It survives signing-key rotation and is written atomically before a snapshot is accepted. Back up this directory with the machine configuration; existing corruption, deletion after prior use, or an unwritable state directory disables the affected registry. A protected catalog distinguishes deletion from genuine first use. Record reads reject cursor cycles, oversized cursors, more than 10,000 records, and responses larger than 16 MiB. Network retries have three total attempts and finite deadlines. GET retries only network failures, 429, and 503; record publication retries only the identical idempotent request. Redirects are rejected.

For managed fleets, a system configuration at /etc/cocoaskills/config.json (or %ProgramData%\cocoaskills\config.json on Windows) is read before the user config. Keys it lists under locked cannot be overridden from the user config, so registry trust, the source allowlist, and the audit policy can be distributed through device management. Set audit.registry_policy to strict to fail any install that is not audited by a trusted registry, and run csk status --attest to re-check installed skills against the registries. An auditor submits a signed record with csk audit --publish <record> --registry <url> --token <token>. The production service, including stable pagination, durable append, backup verification, and air-gapped bundle import for closed networks, is Curator Skill Registry.

CLI

Command	Behavior
csk bootstrap	Create machine-level global config; interactive or scripted via --skills-root, --default-agents, --non-interactive, --force. --if-missing is an idempotent no-op when config already exists and is mutually exclusive with --force.
csk init [path]	Create project Skillfile.json and the managed .gitignore block. Supports --alias, --agents, and --no-interactive for scripted setup.
csk install [target]	Apply Skillfile.json using current git refs. Missing git URL sources are cloned into skills_root; existing local repositories are not fetched. No target means current project; target may be an alias, ., or a project path.
csk install --audit [strict]	Run the audit gate for this install only. Without strict, audit is advisory and does not change config.
csk install --all	Install every project explicitly registered in global config.
csk update	Fetch all git repositories under skills_root. Does not modify projects.
csk upgrade [target]	Fetch only the selected project's direct and transitive skill repositories, then install. --dry-run does not update cached repositories or persist files.
csk upgrade --all	Fetch the union of dependency closures once, then install every registered project.
csk status [target]	Show manifest vs installed state, including active dev substitutions. --check exits non-zero unless everything is up-to-date; --json prints machine-readable output.
csk status --all	Show status for every registered project.
csk add <name> --tag/--branch/--revision ...	Add or replace a skill declaration in the project Skillfile; apply with csk install.
csk remove <name>	Remove a skill declaration from the project Skillfile; the next install cleans generated files.
csk gc	Remove unreferenced runtime entries, snapshot cache entries, and dead consumer registry entries.
csk audit [target]	Run skill security audit for the current project, an alias, ., or a project path. Supports --all, --global, and --json.
csk skill check <dir>	Validate one skill directory without requiring global config or project setup.
csk list [--paths]	List configured projects and declared skills.
csk project add <alias> <path>	Register a project for --all and create a manifest if missing.
csk project resolve [target]	Show resolved project alias, checkout alias, Skillfile, and install paths.
csk global init	Create the user-wide global Skillfile.json, global skill context, bin, and env files.
csk global add <name> --tag/--branch/--revision ...	Add or replace a global skill declaration.
csk global remove <name>	Remove a global declaration; the next global install cleans generated files.
csk global install	Install all globally declared skills without fetching.
csk global update	Fetch source repositories for globally declared skills.
csk global upgrade	Run global update, then global install. --dry-run skips the update and performs a non-persistent install plan.
csk global status	Show global manifest vs installed state.
csk global list	List global skill declarations.
csk config show	Print resolved config path and contents.
csk shell-init [auto|zsh|bash|powershell]	Optionally print shell hook code for human-facing global and project-local auto-PATH activation. The default auto detects the current environment; --install atomically caches it and prints the profile source command; --no-global limits activation to project checkouts. Agent execution does not require this hook.
csk --version	Print version and exit.

Flags shared by install and upgrade:

• --dry-run: plan work without modifying files.
• --verbose: print resolved commits and installed command shims.
• --fix-gitignore: deprecated escape hatch; prefer csk init.
• --strict-tags: fail if a tag was locally moved to another commit.

Exit codes: 0 success, 1 one or more projects or skills failed, 2 configuration error, 3 lock contention.

Development

Requires Python 3.11+.

git clone https://github.com/ivanopcode/cocoaskills.git
cd cocoaskills
python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[dev]"
pytest

Build artifacts locally:

python -m build
twine check dist/*

The runtime package is stdlib-only. Versioning is driven by setuptools-scm from git tags; the generated src/csk/_version.py is not committed.

See CONTRIBUTING.md for the contribution workflow, coding conventions, and the RFC process for design changes.

Documentation

• Architecture overview: module map, install pipeline, the context/runtime split, storage layout, and security boundaries.
• Skill dependencies, RFC 0007: schema v4 requirements, closure resolution, activation modes, dev substitutions, source allowlist. Russian translation: docs/v0.9-design.ru.md.
• Skill authoring guide: practical contract for authoring CocoaSkills-compatible skill repositories, covering schema v2 runtime roots, schema v3 capabilities, schema v4 requirements, system dependencies, audit behavior, and the release checklist.
• Skill security audit, RFC 0005: schema v3 capabilities, deterministic audit gates, verdict cache, and trust workflow.
• Audit LLM backends, RFC 0006: the command and codex audit backends, file-content redaction, timeout plumbing, and fail-open/fail-closed behavior.
• MVP design specification: the v0.1 contract; later RFCs supersede parts of it.
• CHANGELOG: release history in Keep a Changelog format.

Security

See SECURITY.md for supported versions and the vulnerability reporting process. The audit subsystem and its guarantees are described in docs/audit-design.md.

Archive extraction rejects links, unsafe or colliding paths, more than 100,000 entries, or more than 512 MiB of declared file data. Registry reads cap each response at 16 MiB and each artifact query at 10,000 records.

License

Apache-2.0. See LICENSE.