Skip to content

CHANGELOG.md

Latest commit

 

History

History
52 lines (42 loc) · 1.77 KB

CHANGELOG.md

File metadata and controls

52 lines (42 loc) · 1.77 KB
Raw

Changelog

All notable changes to CodeFix-Env are documented here. Format: Keep a Changelog Versioning: Semantic Versioning

[0.2.0] — 2026-04-22

Added

  • Full src/ layout with proper Python packaging
  • Pydantic v2 models: CodeFixAction, CodeFixObservation, CodeFixState
  • 21 tasks across easy (8), medium (8), hard (5) difficulties
  • 11 bug categories: syntax, logic, off-by-one, type, algorithm, scope, etc.
  • CodeFixClient with async + sync (.sync()) support
  • RewardPipeline blending rule-based + optional neural reward model
  • RewardMLP — PyTorch neural reward model (8-feature input)
  • SessionManager for concurrent multi-client server support
  • Sandbox executor with subprocess isolation and security policies
  • GitHub Actions CI: lint, test (Py 3.10/3.11/3.12), Docker build, type check
  • Docker multi-stage production image with non-root user
  • Benchmark runner with RandomAgent, GreedyAgent, OracleAgent
  • GRPO training example with TRL
  • Full API docs and quickstart guide
  • OpenEnv YAML manifest

Changed

  • Removed committed .venv/ directory
  • Restructured server into server/ module
  • Separated environment logic from FastAPI app

Fixed

  • Session handling: X-Session-ID header pattern
  • Reward normalisation clipped to [0, 1]
  • Sandbox import whitelist updated to include all stdlib math/collections

[0.1.0] — 2026-03-15

Added

  • Initial MVP: 3 tasks (easy, medium, hard)
  • Basic FastAPI server
  • CodeFixAction with run_tests, edit_line, submit_fix
  • Docker deployment
  • OpenEnv YAML config

Features

  • Multi-factor reward signals
  • Sandboxed code execution with timeout protection
  • Support for SFT, DPO, GRPO, PPO training frameworks