Repolode
Submit
NVlabs

NVlabs/Skill2Env

Turns any Agent Skill into RL-ready terminal tasks in the Harbor format, using Codex agents inside Docker containers.

View on GitHub
Stars
185
Forks
18
Open issues
0
Last commit
18 days ago

What is it?

Skill2Env takes Agent Skill folders and generates terminal-based tasks in the Harbor format, the structure used for reinforcement learning environments. Each generated task includes an instruction file, a Docker environment definition, fixtures, a verifier with a rubric, and an optional reference solution.

The pipeline runs Codex agents inside containers: a planner decides how many tasks a Skill should yield, and creator agents build each task. The tool requires a Linux host with Docker, Python 3.12+, and a Codex login. Batch runs can be scanned recursively across a directory tree, with a resume option that skips Skills already finished.

The repository also includes SkillHub, a collection of license-friendly Skill folders that can serve as input.

Key features

  • 01Converts Agent Skill folders into Harbor-format terminal tasks.
  • 02Runs planner and creator Codex agents inside Docker containers.
  • 03Generates instruction, environment, verifier and optional solution files for each task.
  • 04Scans input directories recursively and supports resuming interrupted batches.
  • 05Includes SkillHub, a set of license-friendly Skill folders as sample input.

Use cases

  • Building RL training environments for terminal agents from existing Skills.
  • Batch-generating verifiable tasks across a collection of Agent Skills.
  • Experimenting with turning reusable agent skills into training data.

Quick start

  1. 1

    Clone the repository

    git clone git@github.com:NVlabs/Skill2Env.git
  2. 2

    Enter the project directory

    cd Skill2Env
  3. 3

    Install dependencies with uv

    uv sync
  4. 4

    Generate up to two tasks from a sample Skill (requires Docker, Python 3.12+, and a Codex login)

    uv run skill2env generate --input-root SkillHub/test_samples/game-developer --out output/quickstart --max-tasks-per-skill 2 --max-parallel-workers 2

Things to know

  • Requires a Linux host with Docker running.
  • Needs a Codex login at ~/.codex/auth.json and internet access during generation.
  • A first run can take 20 to 30 minutes on a 4-CPU host.

Topics

#agents#reinforcement-learning#docker#codex#terminal#training-data

FAQ

What is Skill2Env?
Turns any Agent Skill into RL-ready terminal tasks in the Harbor format, using Codex agents inside Docker containers. Skill2Env takes Agent Skill folders and generates terminal-based tasks in the Harbor format, the structure used for reinforcement learning environments. Each generated task includes an instruction file, a Docker environment definition, fixtures, a verifier with a rubric, and an optional reference solution.
Is Skill2Env free to use? What license does it have?
Skill2Env is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is Skill2Env still maintained?
Actively maintained. The last commit to Skill2Env was on Oct 1, 2026.
What can I use Skill2Env for?
Building RL training environments for terminal agents from existing Skills. Batch-generating verifiable tasks across a collection of Agent Skills. Experimenting with turning reusable agent skills into training data.
How do I install Skill2Env?
Clone the repository: git clone git@github.com:NVlabs/Skill2Env.git Enter the project directory: cd Skill2Env Install dependencies with uv: uv sync Generate up to two tasks from a sample Skill (requires Docker, Python 3.12+, and a Codex login): uv run skill2env generate --input-root SkillHub/test_samples/game-developer --out output/quickstart --max-tasks-per-skill 2 --max-parallel-workers 2
Where can I get Skill2Env?
The source code and installation instructions are on GitHub: https://github.com/NVlabs/Skill2Env

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects