Turn any GitHub repository into a verifiable RL environment for training and evaluation.

Quickstart · Pipelines · Output · Docs

Repo2RLEnv synthesizes verifiable training and evaluation data from existing repositories, exports it into a uniform spec, and pushes it straight to the Hugging Face Hub. The output spec is Harbor's, so every dataset you produce drops directly into any Harbor-compatible runtime — no glue code.

# Install (pick one)
uv add repo2rlenv                                 # add to a uv-managed project
uvx repo2rlenv --help                             # one-shot, no install
pip install repo2rlenv                            # classic

# Auth: nothing to set up if you've done `gh auth login` and `huggingface-cli login`.
# Otherwise:  export GITHUB_TOKEN=... ; export HF_TOKEN=...

# Generate a dataset locally
repo2rlenv generate \
  --repo <owner>/<repo> \
  --pipeline pr_runtime \
  --pipeline-opt limit=5 \
  --llm anthropic/claude-sonnet-4-6 \
  --out ./datasets/<dataset-name>

# Validate (fast structural check) and publish
repo2rlenv validate ./datasets/<dataset-name>
repo2rlenv push ./datasets/<dataset-name> <your-org>/<dataset-name>

# Anyone can pull + run a published dataset on a fresh machine
repo2rlenv pull <your-org>/<dataset-name> ./datasets/<dataset-name>
harbor run -p ./datasets/<dataset-name> -a oracle --env docker

→ Explore and visualize any Harbor dataset pushed to the Hub: Harbor Visualizer

Full walkthrough in docs/quickstart.md.

Repo2RLEnv runs synthesis pipelines that read real repositories — source code, merged PRs, commits, CVEs — and use them as a seed to generate RL environments: tasks with a concrete, solvable objective and a programmatic reward (no human grading).

Input: any repo. Output: a runnable RL environment you can point any LLM or coding agent at.

# every pipeline shares one contract: read a repo, emit verifiable tasks
class Pipeline(Protocol):
    name: ClassVar[PipelineName]
    def run(self, out_dir: Path) -> PipelineResult: ...   # writes tasks/<id>/

Generate from a repo, then run any agent against the result — the environment is scored automatically:

# 1. synthesize an environment from a repo
repo2rlenv generate --repo pallets/click --pipeline pr_runtime \
  --pipeline-opt limit=10 --llm anthropic/claude-sonnet-4-6 --out ./env-click

# 2. run an agent inside the sandbox (swap -a / -m for any of 25+ harnesses)
export ANTHROPIC_API_KEY=...   OPENAI_API_KEY=...
harbor run -p ./env-click -a claude-code -m anthropic/claude-sonnet-4-6 --ae ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY --env docker
harbor run -p ./env-click -a openhands   -m openai/gpt-4o               --ae OPENAI_API_KEY=$OPENAI_API_KEY     --env docker
harbor run -p ./env-click -a codex       -m openai/o3                   --ae OPENAI_API_KEY=$OPENAI_API_KEY     --env docker
harbor run -p ./env-click -a hermes      -m anthropic/claude-sonnet-4-6 --ae ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY --env docker

Each agent's per-task reward lands in /logs/verifier/reward.json, ready for training or eval.

A pipeline turns a repo into Harbor tasks. Two are stable and recommended for production; six are experimental — usable today (the CLI prints a warning before they run), with interfaces and output quality still evolving.

pr_diff mines merged pull-request diffs into lightweight, text-only tasks. The agent proposes an edit, and a verifier scores it against the real merged diff — on format, the files it touched, how much it changed, and (via an LLM judge) whether it's semantically right. No per-repo setup: every task ships a thin python:3.12-slim image. → Reference dataset: AdithyaSK/repo2rlenv-pr-diff (100 oracle-verified tasks).

pr_runtime is the SWE-bench-style flagship. It mines merged PRs and actually runs the repo's test suite inside a Docker sandbox: the tests the PR fixed must go from failing to passing under the gold patch, while the rest keep passing. That makes it the strongest, least-gameable signal of the set. → Reference dataset: AdithyaSK/repo2rlenv-pr-runtime (100 oracle-verified tasks).

→ All reference datasets: Verifiable RL Environments collection

These run normally but emit a warning first — pin a release if you depend on them. Each links to its own page; the gist:

• commit_runtime — mines commit history directly, catching fixes that never went through a PR. Reference dataset: AdithyaSK/repo2rlenv-commit-runtime (52 oracle-verified envs).
• cve_patches — security tasks from public CVEs, mapped to their fix commits.
• mutation_bugs — injects synthetic bugs into real code; the agent must restore the tests to green.
• code_instruct — generates a problem + executable verifier from a real source file.
• equivalence_tests — the agent reimplements a real function; generated tests check it matches the original.
• refactor_synthesis — mines refactor commits and verifies behavior is preserved.

What the columns mean

• Reward signal — the verifiable signal emitted per task. test_execution = the repo's own tests gate the reward (F2P/P2P or pytest pass rate); diff_similarity = the agent's output is scored against the oracle diff (format, file targeting, region overlap, LLM judge). Pipelines that emit both use test_execution as the primary training signal.
• Sandbox — whether the task runs inside Docker. ✅ = a per-repo image is built once by the bootstrap phase and cached; thin = no bootstrap, just a generic python:3.12-slim image.
• LLM use — when a language model is invoked, which sets where your API cost goes:
  • at env build — only during bootstrap (constructing the Docker image); cached, so generation itself is LLM-free.
  • at synthesis — the model authors the task (problem + verifier) for every task generated.
  • at verify — the model judges the agent's solution at scoring time (one reward component), and degrades gracefully when no key is set.
• Languages — source languages the pipeline supports.

→ Full reference — per-pipeline options, reward design, and dataset cards: docs/pipelines/.

Sandbox pipelines need a working Docker environment for the target repo. Repo2RLEnv's bootstrap phase builds it automatically — an LLM agent iterates shell commands inside a fresh container until the repo builds and its test suite collects, then commits and content-addresses the image. The expensive step runs once per (repo, ref); every downstream task reuses the cache. pr_diff skips it entirely.

repo2rlenv bootstrap --repo <owner>/<repo> --llm anthropic/claude-sonnet-4-6

Design, cache layout, cost tracking: docs/reference/BOOTSTRAP.md.

A dataset that:

• Is verifiable — every task carries an executable test (test_execution) or a stored oracle diff (diff_similarity); your trainer picks the reward type.
• Is content-addressed — a content_hash over each task; identical artifacts ⇒ identical hash.
• Trains anywhere via Harbor — TRL, SkyRL, Prime-RL, Tinker, Miles, Slime, harbor.rl.
• Evaluates with any agent harness — Claude Code, OpenHands, Codex CLI, Gemini CLI, …
• Is language-agnostic by spec — runtime pipelines emit a Dockerfile + shell verifier; pr_diff is pure text and works for any language.
• Publishes natively to the Hub — repo2rlenv push writes a Harbor-compatible registry.json so consumers harbor download (or repo2rlenv pull) with zero glue.
• Supports private repos end-to-end — gh auth token resolved automatically; build secrets declared by name; verifier-time secrets forbidden by spec.

Our focus is synthesis — we don't reimplement sandboxes, agent harnesses, or a registry. Tasks are emitted in the Harbor format (with a small [metadata.repo2env] block for provenance: pipeline, base commit, PR URL, content hash, reward kinds), so they run on Harbor's existing stack — Local Docker / Modal / Daytona / E2B / Runloop, 25+ agent harnesses, parallel execution, and the publishing CLI.

Pipelines are pluggable by design — adding a synthesis strategy is the main way to extend Repo2RLEnv:

1. Implement the Pipeline protocol (name + run() -> PipelineResult) in src/repo2rlenv/pipelines/.
2. Register it in PIPELINES and add its options model; new pipelines start experimental = True.
3. uv run pytest tests/test_pipeline_contract.py enforces the contract.

Full cookbook (oracle invariant, reward design, QA gate): docs/contributing/ADDING_A_PIPELINE.md. Issues and PRs welcome — see CONTRIBUTING.md.

• 🚀 docs/quickstart.md — install → first dataset → push, in 10 minutes
• 📖 docs/pipelines/ — one page per pipeline (when to use, oracle shape, options)
• 📚 Reference contracts:
  • REWARD_SCHEMA.md — reward.txt + reward.json fields for every pipeline
  • SPEC.md — input/output contract
  • API.md — Python API for src/repo2rlenv/
  • AUTH.md — GitHub / HF / LLM auth resolution
  • ENV.md — every environment variable the tool reads, in one place
  • BOOTSTRAP.md — LLM-iterated per-repo Docker image
  • AGENTS.md — Harbor agent harnesses + RL trace plumbing
• 🛠 CONTRIBUTING.md — dev setup, PR conventions, release flow
• 🧪 ADDING_A_PIPELINE.md — cookbook for shipping a new pipeline
• 🔭 Harbor Visualizer — explore and inspect any Harbor dataset pushed to the Hub

• Harbor — the task format + runtime we adopt as our output spec
• RepoLaunch (Microsoft) — LLM-agent env setup; our bootstrap is an independent reimplementation
• OpenReward — ORS protocol + extra trainer integrations above Harbor
• SWE-Gym — RL-environment framing for SWE-bench-style tasks
• verifiers (Prime Intellect), OpenEnv (Meta + HF) — adjacent standardization efforts

Every pipeline that draws from external work carries an Acknowledgment block in its .py file. No code is copied — implementations are independent and Apache-2.0 licensed.

Apache 2.0. The original PR/commit contents remain under their respective source-repo licenses; datasets redistribute public commits for ML research under fair use.