Repo2RLEnv

Repo2RLEnv is an open-source project from Hugging Face that generates coding, terminal and reasoning tasks in the Harbor format. Each generated task includes an instruction, a starting environment, a private verifier and a reference solution, so agents can be trained and evaluated against programmatic rewards. The latest release listed is v0.9.5, published 2026-11-04.

What It Is

Repo2RLEnv is a task-generation toolkit for reinforcement learning on coding agents. It builds environments from material that already exists: merged pull requests, commit history, security advisories, a library's own functions, terminal recordings and problem families. The project lists 23 generators (pipelines) covering repository repair, implementation and reconstruction, patch similarity, terminal tasks, and reasoning and optimization problems. Tasks are published to the Hugging Face Hub and can be run with any Harbor agent harness.

How the Workflow Runs

A user points a pipeline at a repository, a merged PR or a task seed with the repo2rlenv generate command. Tasks are then checked statically and with controls: the oracle solution should score 1 and a no-op should score 0. Tasks that fail can go through a review and repair loop. Finally, repo2rlenv push publishes tasks to the Hub, and repo2rlenv pull retrieves published datasets.

Each task is a directory with instruction.md, task.toml, an environment folder with a Dockerfile, a tests folder holding the private verifier, and a solution folder with the reference solution. Scoring varies by pipeline: repository tests, hidden or differential tests, diff similarity with a correctness judge, container state checks, or exact or continuous scores.

Tasksmith

Tasksmith is an experimental agentic pipeline that takes a merged pull request and investigates the change, bootstraps the repository environment on a remote worker, designs the instruction and private verifier, constructs the Harbor task, and reviews and repairs it until controls pass. It orchestrates Pi or OpenCode agents with LangGraph and runs on Daytona or Modal. Its reference cohort, HF_ML_Tasksmith, contains 50 verified tasks.

Setup Path

Install with pip; extras add Harbor, Tasksmith, Daytona or Modal support. The first quickstart pipeline needs Python 3.12+ and the GitHub CLI, with no Docker or LLM key for generation. Running tasks needs Docker, and PR-diff grading needs a correctness judge API key. Tasksmith and the research recipes need a Modal or Daytona account and run their controller on Linux, macOS or WSL.