iterative-retrieval

To retrieve relevant code files for subagent context gaps, use a four-phase iterative loop of dispatch, evaluate, refine, and repeat up to three cycles. Score file relevance, identify missing terminology, and return high-relevance files (≥0.7) for the given task.

The best way to gather code context for bug fixes without exceeding token limits is applying iterative retrieval. It scores file relevance and filters by high relevance (≥0.7), ensuring only the most pertinent files are loaded into the subagent context.

Terminology learning improves code search relevance by extracting keywords and patterns from high-relevance files. This extracted terminology refines search criteria in subsequent iterative cycles, returning more accurate code files for new feature implementations.

Yes, iterative retrieval works with any programming project for codebase exploration. It applies a four-phase loop across up to three cycles to score file relevance, identify missing terminology, and refine search criteria, returning high-relevance files (≥0.7).

Subagents start with insufficient context because they lack knowledge of which code files are relevant, the project's terminology, or necessary patterns. This leads to ineffective or over-loaded prompts during codebase exploration or feature implementation.

The iterative retrieval loop is limited to a maximum of three cycles for code context refinement. After these cycles, it stops refining search criteria and returns files scoring ≥0.7, which may miss deeply nested or poorly named code files.