What problem does it solve?
It solves unreliable “single-edit” bug fixing by treating a GitHub issue as a repository search over many independently sampled patch candidates and then selecting the best one via explicit pruning and voting.
Core Features & Use Cases
- Independent candidate generation (fresh starts): generates C1–C10 candidates without reusing intermediate reasoning, command history, or partial patches across candidates.
- Explicit pruning, selection, and majority-vote discipline: prunes weak/duplicate candidates, ranks survivors, and performs grouped selection with majority voting (default budget) instead of one-shot picking.
- Clean final revalidation: reruns reproduction steps and targeted tests in a final validation pass, recording why the chosen winner survived.
Quick Start
Ask an agent to run the trae-agent harness on the target repository issue, using the default budget of 10 independently generated candidates with grouped majority-voting selection and a clean final revalidation of the chosen winner.