NVlabs/Skill2Env

Skill2Env turns any Agent Skill into RL-ready terminal tasks in the Harbor format.

• Linux host with Docker running (docker info works).
• Python 3.12+ and uv.
• A Codex login at ~/.codex/auth.json (see below). The Codex agents inside the containers use this file for authentication.
• Internet access while generating: planner and creator containers research and download public assets.

git clone git@github.com:NVlabs/Skill2Env.git
cd Skill2Env
uv sync

Install the Codex CLI and sign in once. This writes ~/.codex/auth.json, which skill2env bind-mounts into every agent container:

npm install -g @openai/codex
codex login

On a headless server, sign in on a machine with a browser and copy ~/.codex/auth.json to the server (or pass --auth-json /path/to/auth.json).

Also log in to Docker Hub so base-image resolution uses your authenticated pull quota.

docker login

Generate up to two tasks from one small sample Skill. Expect 20 to 30 minutes on a 4-CPU host (the first run also builds the generator image, which adds a few minutes):

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

The command prints the private run directory (.skill2env/runs/<run-id>/) at startup and a JSON summary at the end. A successful run has retained_tasks >= 1, and each retained task lands under output/quickstart/:

output/quickstart/
├── _corpus_manifest.json
└── task_query_<8-char-id>/
    ├── instruction.md
    ├── task.toml
    ├── environment/
    │   ├── Dockerfile
    │   └── ... fixtures and setup files
    ├── tests/
    │   ├── test.sh
    │   ├── rubric.md
    │   └── ... optional verifier helpers
    └── solution/
        ├── solve.sh
        └── ... optional solution helpers

SkillHub/skills/ contains license-friendly Skill folders (see SkillHub/README.md). --input-root is scanned recursively, so it can point at one Skill, one family, or the whole hub:

uv run skill2env generate \
  --input-root SkillHub/skills \
  --out output/skillhub \
  --max-tasks-per-skill 2 \
  --max-parallel-workers 8 \
  --resume

Retained tasks keep the source hierarchy under --out. --resume skips every Skill that a previous run into the same --out already finished, so interrupted batches continue where they stopped.

Start with a small --max-parallel-workers (about one per two CPU cores) and raise it only after watching CPU, memory, Docker, and Codex rate-limit behavior.

Codex runs a nested sandbox inside its container. On Ubuntu hosts with AppArmor, allow unprivileged user namespaces once and make it persistent:

sudo sysctl -w kernel.apparmor_restrict_unprivileged_userns=0
echo "kernel.apparmor_restrict_unprivileged_userns = 0" | sudo tee /etc/sysctl.d/99-skill2env-userns.conf

Without this, every creator shell command fails with bwrap: loopback: Failed RTM_NEWADDR: Operation not permitted and tasks end in creator_contract_failed.

Project-owned source is licensed under Apache-2.0. See the third-party notices for skill2env's dependencies and container/build tools.

Third-party material in SkillHub/ retains its respective copyright and license terms. See the source and license inventory, preserved license files, and SkillHub README.

@misc{xu2026reinforcing,
  title = {Reinforcing Agents with Collective Skills},
  author = {Xu, Binfeng and Hu, Jian and Zhang, Hao and Zhang, Shaokun and Zhang, Yifan and Demoret, Michael and Kautz, Jan and Dong, Yi},
  year = {2026},
  publisher = {alphaXiv},
  url = {https://www.alphaxiv.org/abs/2609.reinforcing-agents-collective-skills},
  keywords = {Artificial Intelligence (cs.AI), Machine Learning (cs.LG), Computation and Language (cs.CL), Reinforcement Learning, Agentic RL, Skill2Env, FOS: Computer and information sciences}
}