Release candidate 2.0.0-rc2 — one skill, host-specific packages, explicit evidence limits.
One vendor-neutral operating skill helps an agent choose relevant available skills/tools, load only needed instructions, and verify results. It is one skill, not a bundle of thousands. The 15 capability entries in the reference catalog are test metadata, not installed skills, connectors or credentials.
Implicit selection is enabled where the host supports it; the host still decides whether to invoke a skill. Installation, authorization and actual activation are different states. There is no always-on daemon and no guarantee of automatic activation in every chat.
- Download release packages
- Cross-platform CI
- Uzbek guide
- Installation, migration and recovery
- Platform-specific scope and official sources
- Validation and remaining release gates
- Independent pre-publication audit
- Release notes
- Original technical specification
Python 3.12 or 3.13 is the declared target. Use a fresh virtual environment. Only the development/validation toolchain needs the pinned dependencies; the installer and reference router use the Python standard library.
python -m venv .venv
# Activate .venv using your operating system's normal activation command.
python -m pip install -r requirements-dev.txt
python scripts/validate_skill.py
python -m pytest -q
python scripts/benchmark.py
python scripts/benchmark.py --dataset tests/routing/holdout.json
python scripts/package_release.py --output release/rc2-localBuilds use new output directories and refuse to overwrite existing release artifacts. Tests build their own fixtures; no pre-existing dist or release directory is required. Release output contains 14 ZIPs plus SHA256SUMS.txt, each ZIP below 25,000,000 bytes. Version and runtime mappings come from project.json and core/adapters.json.
Dry-run is the default. Replace the example workspace with the intended project:
python installers/install.py claude --workspace ./example-project --action plan
python installers/install.py claude --workspace ./example-project --action install
python installers/install.py claude --workspace ./example-project --action verifyUse the same runtime, scope, workspace/home and state directory for later operations. Destination and state must be on the same filesystem; for a workspace on a different volume from your home, pass an explicit same-volume --state-dir. Plan detects this before writes. Supported filesystem runtime identifiers: codex, claude, gemini, antigravity, cursor, copilot, opencode, generic. A runtime identifier is a packaging target, not a claim of tested host support. Generic requires a host that discovers .agents/skills.
Use the standalone skill ZIP if your account provides a compatible skill importer. If the selected surface only accepts plugins, use its separately generated plugin artifact; a plugin is not mislabeled as a standalone skill. Claude's custom skill upload has its own ZIP. The source ZIP is for developers, not the skill importer.
Uploading a ZIP as a conversation attachment is not necessarily skill installation. Web/mobile import, enablement, discovery and new-chat activation require account-level validation. Local installation does not synchronize to a cloud/mobile account.
Canonical SKILL.md → native metadata selection → selected instructions → scoped action → outcome checks. The Python keyword router is an offline reference implementation, not connected to native model routing. Its benchmark scores must not be advertised as ChatGPT/Claude activation accuracy.
Cross-platform CI runs on pushes and pull requests across Windows, Linux and macOS with Python 3.12 and 3.13. See the linked Actions page for actual results; configuration alone does not prove a passing run. No live model calls or real agent-home installations were made for this candidate. The readiness and audit documents record pre-publication evidence. See the repository's Actions and Releases pages for actual remote CI and publication status. No automatic deployment or release publishing is configured in CI.