OpenEnv Compliant

🔍 CodeReviewEnv

The first RL benchmark for structured knowledge work

Train and evaluate LLM agents on real code review tasks — severity triage, queue prioritization, and actionable feedback generation — with deterministic grading and trajectory export for world model research.

3
Tasks
50
PR Templates
7
Languages
0.69
GPT-4o-mini

Three Difficulty Levels

⭐ Easy

Severity Labeling

Classify each PR's bug severity: critical, high, medium, low, or none.

5 steps GPT-4o-mini: 1.00
⭐⭐ Medium

Queue Prioritization

Sort the review queue by urgency — security first, junior devs next.

3 steps GPT-4o-mini: 0.68
⭐⭐⭐ Hard

Feedback Generation

Write actionable review comments targeting specific buggy lines.

18 steps max GPT-4o-mini: 0.38

Interactive Demo

🎮

Select a task above to start an interactive demo

Research: Knowledge-Work World Models

🧠

Semantic MDP

States are structured text (code diffs, bug categories). Transitions depend on professional judgment, not physics.

📊

Trajectory Export

Every episode exports (s, a, r, s') transitions in JSONL for training Knowledge-Work World Models.

🔬

Deterministic Grading

No LLM-as-judge. Ordinal matching, Kendall Tau correlation, and 5-component weighted scorers.

🛡️

Anti-Exploit

Spam penalties, consistency checks, and decaying rewards prevent trivial gaming strategies.

Benchmark Comparison

Benchmark State Space Transition World Model? Domain
MuJoCoℝⁿ (joints)Physics sim✅ DreamerRobotics
AtariPixelsGame engine✅ MuZeroGames
TextWorldSynthetic textGame rules⚠️ Li et al.Text games
SWE-benchCodeN/A❌ Eval onlySE
CodeReviewEnvStructured textProfessional judgment✅ KW-WMKnowledge work

API Endpoints

POST /reset

Start new episode

POST /step

Take an action

GET /state

Current state

GET /health

Health check

GET /export_trajectory

JSONL trajectory

GET /docs

OpenAPI docs