RepoGym

RepoGym is a standardized reinforcement learning environment for software engineering (SWE) tasks. It frames coding as an interactive, tool-using process instead of one-shot code generation, and gives agents a real (sandboxed) repository to work in.

environment design

Agents interact with the environment through structured tool actions, not raw text diffs: reading files, writing changes, running tests, and using git. Every episode executes inside a Docker-isolated sandbox, so agent-generated code never touches the host.

import repogym

env = repogym.make("BugFix-v1")
obs = env.reset()

while not done:
    action = agent.act(obs)  # read_file / write_file / run_tests / git_diff
    obs, reward, done, info = env.step(action)
    # reward: shaped from test passage + code quality signals
    # info: test results, diagnostics, trajectory metadata

reward shaping

Rewards combine test passage with code-quality signals, rather than a bare pass/fail. That shaping is what makes learning tractable in a domain where most trajectories end in failure - the agent needs partial credit for getting closer, not just a binary outcome at the end.

telemetry and dashboard

Every trajectory - actions taken, tool outputs, reward breakdown - is logged and surfaced through a React dashboard, which turns debugging an agent’s behavior from log-scraping into something you can actually look at.

status

RepoGym is a research environment: the core loop (Docker execution, structured actions, shaped rewards, telemetry) is in place, and it’s the substrate I’m using to study RL for SWE agents.