/diagnose

Run systematic RL debugging using the 80/20 rule - check environment and reward BEFORE changing algorithms

# Diagnose Command

Apply the 80/20 rule for RL debugging: 80% of failures are environment/reward issues, not algorithm problems.

## Core Principle

**DO NOT suggest algorithm changes until environment and reward are verified.**



## Diagnostic Process

### Phase 1: Environment Sanity (Check First - Most Common Issues)

1. **Observation Space**
   - Does the agent observe everything it needs to solve the task?
   - Are observations normalized (roughly [-1, 1] or [0, 1])?
   - Any information leakage (observing future states)?

2. **Action Space**
   - Discrete vs continuous - is algorithm...