Skill to Plugin

Create a complete Codex plugin from one skill or a set of skills, implement missing behavior, improve an existing plugin, and learn from verified outcomes.

Codex provides reasoning and code generation through two bundled skills. The dependency-free Python runtime provides deterministic packaging and eight MCP tools for inventory, compilation, validation, feedback, evaluation, promotion and rollback. No separate API key, model subscription, daemon, or weight training.

Install in Codex

Python 3.11+ and a Codex client supporting plugin marketplaces are required.

codex plugin marketplace add trukhinyuri/skill-to-plugin
codex plugin add skill-to-plugin@skill-to-plugin

The repository includes portable plugin.json / mcp.json and the native .codex-plugin/plugin.json / .mcp.json compatibility layout. Desktop and CLI share the installed plugin cache on this host. Refresh Plugins or open a new chat if the running client has not reloaded its skill catalog. Installation does not publish the plugin to OpenAI's universal directory.

Use it

Select Skill to Plugin in the plugin picker, or ask Codex:

Use Skill to Plugin to fully implement a plugin from /path/to/skill and /path/to/second-skill. Save it in /path/to/output, verify every behavior, and report any dependency you cannot implement.

Use Skill to Plugin to improve /path/to/existing-plugin, preserving its interfaces and checking regressions.

Use self-improve for this project. Read local feedback, implement a justified candidate, compare it with the baseline, and promote only verified improvement.

The skills are skill-to-plugin and self-improve, namespaced by their plugin when the host requires that. Ordinary workflows can record proposed lessons after useful observed results; explicit self-improvement runs the full comparison cycle. Learned text remains project-local evidence until a verified candidate is promoted.

Direct CLI

The checkout and installed cache both contain a relocatable entrypoint:

python3 scripts/stp.py inspect examples/normalize-text
python3 scripts/stp.py build examples/normalize-text --name text-tools --output /tmp/text-tools
python3 /tmp/text-tools/skills/normalize-text/scripts/normalize.py '  Привет   世界  '
python3 scripts/stp.py validate /tmp/text-tools

For updates, add --existing BASELINE --plugin-version 0.2.0 and use a separate output directory. For external Markdown resources, explicitly grant their trees with --resource-root PATH. Inputs are read as data; compilation never runs their instructions or scripts. Code references and dynamically loaded dependencies also need host inspection and implementation. See the implementation workflow.

Measured self-improvement

State lives in the working project's ignored .skill-to-plugin/, never global instructions or the installed plugin cache. learn record stores observed evidence and proposed lessons; Codex turns relevant lessons into candidate code/instructions. learn evaluate runs an explicitly authorized external trusted suite against separate copies of baseline and candidate. Commands run without a shell, with time/output limits. This is not an OS sandbox.

Promotion requires actual recorded results bound to unchanged input hashes, strict improvement, all candidate cases passing, no regressions, and package validation. Promotion saves a verified backup; rollback refuses to overwrite later edits. Ties and failures retain the baseline. See learning usage.

This measures the chosen behaviors. It does not prove global superiority, guaranteed reasoning improvement, or success for every skill. Semantic completion requires a host-authored requirement contract and realistic behavior tests. Host executables and environment are not fingerprinted. After abrupt termination during replacement, retain the backup and review the stale lock/sibling directories before recovery.

Development and evidence

python3 -m unittest discover -s tests -v
python3 scripts/stp.py validate .

The suite covers conversion, resource closure/relocation, existing-plugin updates, MCP lifecycle and calls, measured improvements/regressions, stale/tampered evidence, rollback, bounded execution, and local state constraints. Behavioral testing of the skills uses an independent synthetic conversion scenario. Research basis and the limits of the quality claim are in research notes.

MIT licensed. No telemetry, background collection, remote inference or automatic publishing. User task authorization and host permissions remain authoritative.