repo2rlenv 0.8.0

Turn any repository into an RL environment for training and evaluation.

Repo2RLEnv

Turn any repository into an RL environment for training and evaluation.

⚠️ Experimental. This is a research project in active development. APIs, spec fields, and CLI flags change between minor versions. Pin to a specific release if you depend on it; expect breaking changes on main.

Repo2RLEnv synthesizes verifiable data from existing repositories using pluggable pipelines, exports it into a uniform spec, and pushes straight to the Hugging Face Hub. End-to-end — synthesis → standardize → train + eval — with the focus on training. The uniform spec is Harbor's, so the datasets you produce drop straight into any Harbor-compatible runtime.

  ╭──────────────╮     ╭──────────────╮     ╭──────────────╮     ╭──────────────────╮
  │     any      │ ──▶ │  synthesize  │ ──▶ │ uniform spec │ ──▶ │ train · eval ·   │
  │     repo     │     │  (pipelines) │     │   (Harbor)   │     │  push to HF Hub  │
  ╰──────────────╯     ╰──────────────╯     ╰──────────────╯     ╰──────────────────╯
                       └──────────────────────── Repo2RLEnv ────────────────────────┘

---

Quickstart

# 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_diff \
  --pipeline-opt limit=5 \
  --llm anthropic/claude-sonnet-4-6 \
  --out ./datasets/<dataset-name>

# Or push straight to HF Hub with --out hf://<your-org>/<dataset-name>

# Validate a local dataset against the spec
repo2rlenv validate ./path/to/dataset

# Score a candidate diff against a task's oracle (diff-similarity reward)
repo2rlenv reward --task ./datasets/<dataset-name>/<task-id> --prediction ./candidate.diff

# Or write a sample config first and use --config
repo2rlenv init && repo2rlenv generate --config repo2rlenv.config.yaml

Full walkthrough in docs/quickstart.md.

---

Pipelines

Different methods to manufacture verifiable tasks from a repo. Pick one, run it, push the dataset.

Pipeline	What it does	Sandbox	LLM	Inspiration	Docs
pr_diff	Mine merged PR diffs (text only, no execution)	—	—	SWE-RL	📄
pr_runtime	Mine merged PRs; sandbox-verify F2P/P2P oracle	✅	✅	SWE-bench	📄
pr_stream	Continuous PR mining (watermark-based, monthly cron)	✅	✅	SWE-bench-Live	📄
commit_runtime	Commit-level mining (bypass PR-review filters)	✅	✅	R2E-Gym SWE-GEN	📄
mutation_bugs	Inject bugs via AST mutations; tests must break	✅	✅	SWE-smith	📄
code_instruct	Repo-anchored OSS-Instruct with executable verifiers	✅	✅	Magicoder / OSS-Instruct	📄
equivalence_tests	Extract a function; LLM writes equivalence tests	✅	✅	R2E	📄
cve_patches	Map OSV CVEs to fix commits in the target repo	✅	✅	PatchSeeker / CVE-Bench	📄

Every pipeline flows through the same QA gate (determinism, oracle consistency, LLM judge, false-negative filter) before tasks are admitted to a dataset. Text-only pipelines skip the heavy QA layers since there's no execution to validate. See docs/pipelines/README.md for reward kinds + GPU requirements.

---

Bootstrap (sandbox-required pipelines)

Pipelines marked with a sandbox ✅ above need a working Docker environment for the target repo before they can run. Repo2RLEnv's bootstrap phase handles this automatically — an LLM agent iterates shell commands inside a fresh Docker container until the repo builds and the test suite collects. The working image is committed, content-addressed, and cached, so the expensive env-construction step runs once per (repo, ref) and every downstream task reuses it. Pure text pipelines (pr_diff) skip it entirely.

You don't normally invoke it directly — repo2rlenv generate --pipeline pr_runtime ... auto-triggers a cache lookup and runs bootstrap on miss. But you can pre-warm it or use it standalone for debugging:

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

Full design + cache layout + cost-tracking + spec extension fields: docs/reference/BOOTSTRAP.md.

---

What you get out

A dataset format that:

• Is verifiable — every task carries either an executable test (test_execution) or a stored oracle diff (diff_similarity); your trainer picks the reward type
• Is content-addressed — content_hash over each task; same artifacts ⇒ same 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 Dockerfile + shell verifier; _diff pipelines are pure text and work for any language with no extra config
• Publishes natively to Hugging Face Hub — --out hf://owner/name writes a Harbor-compatible registry.json so consumers can harbor download without any glue
• Supports private repos end-to-end — gh auth token resolved automatically; build secrets declared by name; verifier-time secrets forbidden by spec

---

Under the hood

Repo2RLEnv emits datasets in the Harbor task format. We don't ship our own sandbox, agent harness, or registry — Harbor already has those. We focus on synthesis: turning a real repo into verifiable, reproducible Harbor tasks. A small [metadata.repo2env] extension inside Harbor's task.toml carries provenance (pipeline name, base commit, PR URL, content hash, reward kinds, etc.).

By targeting Harbor we inherit its full stack: Local Docker / Modal / Daytona / E2B / Runloop sandboxes, every major coding-agent harness, parallel execution, the publishing CLI, and downstream hooks for OpenReward (which adds Miles, Slime to the trainer list).

---

Documentation

Docs are organized into three tiers — see docs/README.md for the index.

• 🚀 docs/quickstart.md — install → first dataset → push to Hub, in 10 minutes
• 📖 docs/pipelines/ — one page per synthesis pipeline (when to use, oracle shape, inspiration)
• 📚 Reference contracts and module-level API:
  • reference/SPEC.md — input/output contract
  • reference/API.md — Python API for src/repo2rlenv/
  • reference/AUTH.md — GitHub / HF / LLM auth resolution
  • reference/BOOTSTRAP.md — LLM-iterated per-repo Docker image
  • reference/AGENTS.md — Harbor agent harnesses + RL trace plumbing
• 🛠 CONTRIBUTING.md — dev setup, PR conventions, commit style, release flow
• 🧪 contributing/ADDING_A_PIPELINE.md — step-by-step cookbook for shipping a new pipeline

---

Adjacent projects

Beyond the per-pipeline inspirations linked in the table above, Repo2RLEnv builds on or adjacent to:

• Harbor — the task format + runtime ecosystem we adopt as our output spec
• RepoLaunch (Microsoft) — LLM-agent-driven environment setup; our bootstrap is an independent reimplementation
• OpenReward — ORS protocol + extra trainer integrations layered above Harbor
• SWE-Gym — RL-environment framing for SWE-bench-style tasks
• SWE-Bench++ — four-stage QA pipeline we'll re-implement
• 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 licensed Apache-2.0.

---

Status

Pre-alpha.

• v0.1.0 shipped on PyPI: pr_diff + HF Hub publish + diff-similarity reward, end-to-end on any GitHub repo (public or private).
• v0.2: bootstrap phase (LLM-driven Docker env), unified Rich UI, content-addressed cache, registry-qualified pullable digests. (rolled into v0.3 release)
• v0.3.0 shipped on PyPI: pr_runtime pipeline (sandbox-verified PR mining with FAIL_TO_PASS / PASS_TO_PASS oracle), auto-triggered bootstrap, structural quality filters, targeted test invocation.
• v0.4.0 shipped on PyPI: polyglot log parsers (Go / Cargo / Jest), Harbor end-to-end verification (Mean reward 1.0 on Go via urfave/cli).
• v0.5: pr_stream (continuous PR mining, watermark-based) + commit_runtime (commit-level mining, SWE-GEN style); defensive git install in emitted Dockerfile so any bootstrap base image works. Harbor-verified on both. (rolled into v0.6 release)
• v0.6.0 shipped on PyPI: first LLM-synthesized pipelines — mutation_bugs (AST-based bug injection inspired by SWE-smith) + code_instruct (repo-anchored OSS-Instruct inspired by Magicoder, with executable verifiers). Harbor-verified on pallets/click (Mean reward 1.000 on both). 271/271 tests passing.
• v0.7.0 shipped on PyPI: equivalence_tests (R2E-style function-level synthesis — extract a real function, LLM writes equivalence tests, gold patch fills in the candidate with the original) + cve_patches (OSV-driven security-fix mining — CVE → fix commit → Harbor task). Harbor-verified on pallets/click and pallets/werkzeug (Mean reward 1.000 on both).
• v0.8.0 shipped on PyPI: refactor_synthesis (Python-native rename-refactor mining — drop the JVM RefactoringMiner dep; commit-message regex + diff verification + multi-criteria verifier). Harbor-verified on pallets/click (Mean reward 1.000). All 8 originally-planned pipelines now shipped.
• v0.9 planned: LLM-judged QA gate (SWE-Bench++ four-layer recipe) + iterative refinement for equivalence_tests + LLM-synthesized PoC tests for cve_patches + HF Hub append-mode for pr_stream + polyglot mutation (Java/JS/Go via tree-sitter) + Extract Method / Inline-function refactor kinds for refactor_synthesis.

License

Apache 2.0 — see LICENSE.