iterative-retrieval

Use iterative retrieval: dispatch a broad keyword search, evaluate 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.

Missing context failures happen when subagents cannot predict needed files upfront. Iterative retrieval addresses this by learning codebase terminology in early cycles and targeting explicit gaps in later refinement passes.

Files scoring 0.7 or higher are considered high relevance and should be included. Files below 0.2 should be confidently excluded, since low-relevance files will not become relevant in later cycles.

Stop when you have at least 3 high-relevance files and no critical gaps, or after a maximum of 3 cycles. Three high-relevance files provide better context than ten mediocre ones.

The loop caps at 3 cycles, so tasks needing deeper exploration may return incomplete context. Initial broad queries can also miss domain-specific terminology until the first evaluation cycle reveals naming conventions.