Radar / AI coding tools / CodeMidas
CodeMidas: scaling agentic coding RL environments from code itself
Agentic pipeline that turns implemented functionality in open-source codebases into executable RL environments using source code as the only input.
Why it matters
Scales RL task creation without issues/commits: 5,545 verified training tasks extracted from 3,185 repos, with execution-checked tests.
What you could build with it
Mine any open-source codebase into RL environments to fine-tune a small, cheap coding agent on real tasks.
Does it hold up?
Too early to judge. The paper is a week old with no public code, dataset, or third-party replication yet — all traction so far is explainers and writeups, not builders running the pipeline.
Built with CodeMidas
- Emergent Mind deep-dive: CodeMidas — Scaling RL Environments from Source Codeemergentmind · Independent technical breakdown of the 5,545-task pipeline, positioning it against SWE-bench, R2E-Gym and SWE-smith, with full result tables (DeepSWE 10.0 to 21.7%, ProgramBench Almost Solved 4.5 to 21.5%).
- CodeMidas: Building RL Environments from Source Code Alone (video)youtube · Emergent Mind's paper-explainer video walking through how CodeMidas mines open-source codebases into executable RL environments without needing issues, commits, or existing tests.
- Hugging Face paper page — #2 Paper of the Dayhuggingface · HF Papers trending page for 2609.22068 (submitted by author Lei Li), showing early community traction around the source-code-only task-mining idea.
First spotted on arxiv: source.
More AI coding tools
mattpocock/skills: Skills for Real EngineersMatt Pocock's open collection of production agent skills, published straight from his .agents directory - 272k stars…coding · JEV 0.75Pi.dev adds MCP to core: 'You Said No MCP'Pi.dev reverses its public no-MCP stance, integrating Model Context Protocol into the core of its agentic coding…coding · JEV 0.71ZCode: Z.ai's Apache-2.0 open-source coding-agent harness, open-sourced 2026-09-21Z.ai's official coding-agent harness, an Electron desktop app plus browser interface, terminal agent and agent runtime…coding · JEV 0.7CodeGraphPre-indexed, auto-syncing Rust-kernel code knowledge graph for Claude Code, Codex, Cursor, OpenCode, Hermes Agent and…coding · JEV 0.69opencodexUniversal provider proxy for OpenAI Codex and Claude Code: use any LLM with Codex CLI, App, SDK, and Claude Code.coding · JEV 0.69Whiteboard: open-source canvas IDE for human-agent software designAn open-source, MIT-licensed desktop app where humans and agents architect software together on a shared canvas. The…coding · JEV 0.62
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.