dream-rsi - Run recursive policy improvement over coding-agent discovery trees
Runs online discovery cycles and deterministic offline replays to improve an exploration policy, rank candidates, report status, and optionally apply a recorded candidate.
Tags
Updated: 2026-09-28Capabilities
Typical Inputs
Typical Outputs
What this skill does
- Initialize the Dream-RSI task
- Run online discovery cycles
- Replay recorded worlds offline
- Improve the exploration policy
- Sweep beta parameters
- Rank recorded candidates
- Report loop status
- Apply a recorded candidate
Inputs
- Task configuration
- Problem statement file
- Discovery workspace
- Candidate program path
- Scoring command
- Loop budgets
- Beta parameters
- Apply confirmation
Outputs
- Recorded discovery worlds
- Policy versions
- Candidate evaluation results
- Beta sweep results
- Loop status reports
- Applied candidate program
Requirements
- Initialized Dream-RSI task
- Dream-RSI mode enabled
- Configured workspace and scorer
- Available agent command
- Compatible active model
- Apply confirmation for source changes