iterative-retrieval

Use iterative retrieval: start with a broad keyword search, score each file's relevance on a 0-1 scale, identify missing context, then refine the query and repeat up to 3 cycles. Return files scoring 0.7 or higher.

Iterative retrieval reduces token usage by sending only high-relevance files instead of entire codebases. Low-relevance files are excluded after evaluation, and the loop stops once roughly three high-relevance files are found.

Iterative retrieval is a 4-phase loop (dispatch, evaluate, refine, loop) that progressively narrows file searches. Each cycle learns codebase terminology from previous results, so later queries match the project's actual naming conventions.

Subagents fail because they are spawned without knowing which files, patterns, or terminology the codebase uses. Sending everything exceeds context limits while sending nothing lacks critical information; iterative refinement resolves this trade-off.

Stop when you have at least three files with relevance scores of 0.7 or higher and no critical context gaps remain. The pattern caps execution at 3 cycles to prevent unbounded searching, then proceeds with the best context gathered.