Repo indexing was doing more harm than good, so we dropped it from our agent (thegit.ai)

In a significant reevaluation of its approach, TheGitAI team has announced the decision to drop repository indexing from their coding agent, shifting focus towards a more dynamic understanding of coding tasks. Initially, the idea was to create a comprehensive index of code to aid the model in retrieving relevant information before beginning work. However, they found that software engineering is not simply a retrieval problem, but an investigative process that requires real-time context and adaptability. Debugging often involves following clues rather than relying on a static map of the codebase, meaning the ability to search for information as tasks evolve is more valuable than pre-determined context. This new methodology emphasizes just-in-time investigation, allowing the agent to learn from the current state of the code, test results, and even errors as they occur. Instead of overwhelming the model with information that may not pertain directly to ongoing inquiries, TheGitAI now enables a feedback loop where the sequence of questioning evolves based on the evidence discovered during the coding process. By prioritizing actionable evidence over speculative context, TheGitAI aims to enhance both the accuracy and traceability of coding solutions, capturing the iterative nature of how seasoned developers solve problems.