1st Place
thinking
We built a self-healing coding agent that searches over compact DSL implementation plans before generating code.
Core trick
An LLM proposes compact DSL plans as branches, and a Rainbow RL controller chooses which to refine, compile, or backtrack from using test feedback.
Description
We built a self-healing coding agent that searches over compact DSL implementation plans before generating code. The LLM proposes semantic reasoning branches, while a Rainbow RL controller learns which branch to select, refine, compile, backtrack from, or terminate using verifier feedback from compile and tests. This decision can be passed back to the LLM to converge to a final solution by pruning the search space. On our held-out synthetic benchmark, Rainbow with Gemma E2B (2 billion parameters) matches Codex 5.4 Low’s 80% solve rate while using roughly 140x fewer tokens and running about 4x faster. Developed and evaluated entirely on a single M3 MacBook Pro, this is an existence proof that a modular learned reasoning layer can help small local models approach frontier-agent performance at a fraction of the cost. The key idea is not to replace the LLM, but to spend expensive LLM budget only on branches that are likely to work, turning code repair from one-shot generation into a learned, verifier-guided search process.
README excerpt
# Hyperreasoning Hyperreasoning is a JetBrains/Codex hackathon project for DSL-guided code search with a Rainbow-style branch controller. This won 1st place at the Jetbrains Codex Hackathon. The core idea is to search over structured implementation plans instead of raw code edits: - generate compact candidate plans in a DSL - rank and traverse candidate branches with heuristic or Rainbow policies - compile and verify promising plans against task tests - surface the search process in a native JetBrains tool window - compare Rainbow against heuristic, random, and one-shot LLM baselines ## Current Status The repo contains an end-to-end demo stack: - Python search-control environment and verifier - plan DSL proposal and compiler interfaces - local FastAPI backend for IDE-triggered task runs - native JetBrains plugin with live search-graph visualization - offline dataset collection and Rainbow training scripts - held-out evaluation reports and benchmark artifacts The current implementation is optimized for clarity and hackathon demo reliability. ## Results Primary held-out eval split: `data/splits/eval_10.txt` Current headline result: | Method | Solve rate | Mean tests passed | Mean time | Mean tokens | | --- | ---: | ---: | ---: | ---: | | Rainbow | 80.0% | 80.0% | 15,353 ms | 1,694 | | Heuristic | 40.0% | 40.0% | 23,144 ms | 2,340 | | Random | 30.0% | 37.5% | 23,009 ms | 2,342 | | One-shot | 30.0% | 37.5% | 31,016 ms | 3,649 | See: - [Eval results](docs/markdown/res