iterative-retrieval

Codebase search context for subagents is improved through iterative retrieval, which progressively refines relevant files across cycles to reduce noise while staying within context limits for multi-agent workflows.

Iterative retrieval is a cyclic process that solves the subagent context problem by progressively evaluating and refining candidate files through a four-phase loop to return high-relevance results with minimal noise.

When initial context is insufficient for bug fixes, you retrieve relevant files by starting with a broad query, running a single evaluation iteration to assess relevance, and progressively refining queries for up to three cycles.

Iterative retrieval enforces a maximum of three cycles, progressively dispatching, evaluating, and refining context-aware file candidates to identify code patterns without exceeding context limits.

Limitations of iterative retrieval include a strict maximum of three cycles for context refinement, meaning highly complex codebase searches with deep terminology gaps may not fully resolve within the enforced loop limit.

Iterative retrieval supports feature work in tasks planning by applying context-aware scoring to evaluate candidate files across dispatch, evaluate, refine, and loop phases within multi-agent workflows.