AI Core Operating Protocol v2.0.0-rc2

AI Core Operating Protocol v2.0.0-rc2

Audited prerelease: one vendor-neutral operating skill with scoped execution,
metadata-first selection and evidence-based completion. Not a bundle of thousands of skills.

What is included

• Eight filesystem adapter targets: Codex, Claude Code, Gemini CLI, Antigravity IDE,
  Cursor, Copilot, OpenCode and a limited generic Agent Skills layout.
• Safe plan/install/verify/rollback lifecycle, ownership checks, hashes and retained snapshots.
• Strict offline reference routing, benchmark validation and reproducible packaging.
• English README, Uzbek guide, technical specification and independent audit report.
• Fourteen ZIP assets, each below 25 MB, each including the original MIT notice.
  Verify downloads against the accompanying SHA256SUMS.txt.

Verified release evidence

Published commit: 0d69fb6f012d3e89a95749938f28b356f8d1939e.

• All six CI matrix jobs passed:
  Windows, Linux and macOS with Python 3.12 and 3.13. Tests, validators, both router
  benchmarks, snapshot checks, package builds and source-drift checks all passed.
• Local independent-audit validation: 118 tests passed; all 11 reported findings closed.
• Offline reference-router datasets: 120 golden cases and 22 development holdout cases.
• The attached source ZIP matches all 86 files in the published commit byte-for-byte.
• Source ZIP SHA-256: bcb336e0a431446f0ecb48d69f10517cfc49bd638ca0bfc2496564d0c61258c7.

The audit report
and readiness documents are historical pre-publication records. The linked CI run above
provides the subsequent remote platform evidence.

Choose one appropriate asset

Use *-skill-* for a compatible standalone skill importer, *-claude-upload-* for
Claude custom skills, *-openai-* for an OpenAI plugin importer, and *-claude-plugin-*
for Claude Code plugins. Runtime-named ZIPs contain that filesystem layout.
Developers should use *-source-*; it is not a skill-import ZIP.
Do not install every variant together: duplicate discovery can result.

Known limits

• Native ChatGPT web/mobile and other host automatic activation are not certified.
  CI validates code and packaging, not native model invocation or cloud installation.
• Local installation does not synchronize into a ChatGPT/Claude cloud account.
• The reference catalog contains test metadata, not installed connectors or credentials.
• Destination and lifecycle state must be on the same filesystem. Abrupt-kill recovery
  requires manual review. Runtime permission and consent controls remain authoritative.
• This is a prerelease, not a stable or all-hosts-supported release.

Read the installation and recovery guide
before migration. Keep the same runtime/scope/workspace/home/state context for lifecycle operations.
Never use the older RC1 manifest-driven uninstaller.

Report vulnerabilities via private reporting,
without including live credentials.