Skip to content

akayamishurui42/skill-orchestrator

v0.2.1

ChatGPT + Codex skill discovery and multi-skill execution with approval gates, seven-day GitHub caching, task audits and MCP tools.

Skill Orchestrator v0.2.1 — portable ChatGPT / Codex plugin

A packaged skill-driven orchestrator plus optional local Codex hooks and an MCP server adapter. This package is a plugin folder, not an automatically connected ChatGPT account.

Codex Desktop on Windows

Extract the ZIP. In PowerShell from the extracted ZIP parent:

py -3 .\skill-orchestrator\install.py --dry-run
py -3 .\skill-orchestrator\install.py

The installer copies it to %USERPROFILE%\.codex\plugins\skill-orchestrator and registers a personal marketplace entry in %USERPROFILE%\.agents\plugins\marketplace.json. Existing installations require replace, keep both, or deny selection; no silent upgrade. It also adds an identifiable policy section to .codex\AGENTS.md (omit via --no-global-policy). Restart Codex and explicitly review/trust its hooks in the local Plugins Directory.

For a collision, pass --on-conflict replace, --on-conflict keep-both or --on-conflict deny only after choosing that outcome.

ChatGPT

See chatgpt/CONNECT.md. The included server/mcp_server.py must be separately deployed over an authenticated, secure MCP connection; a ZIP alone does not activate backend tools in ChatGPT. A skills-only local plugin may be loaded on eligible desktop/Work surfaces. Hook automation is Codex-specific and host-dependent.

Validation

Unzip the source/test package, or run python -m unittest discover -s tests -v in the development distribution. The packaged files have been checked for valid manifests and Python syntax.

No newly discovered GitHub skills are installed without explicit user approval. Skill activation is required every task when this plugin is actually invoked, even while the discovery cache remains fresh. Execution logs are self-reports, not proof that work was done.