iterative-retrieval

Use an iterative retrieval pipeline to progressively refine context for subagents over codebase files. It cycles through a four-phase loop—DISPATCH, EVALUATE, REFINE, and LOOP—to incrementally surface high-value files when the required context is not known upfront.

Apply bounded iterative retrieval to manage context size limits in multi-agent workflows. It restricts the refinement loop to a maximum of three cycles, preventing overfetching and reducing context waste while surfacing necessary files.

The four-phase loop works by dispatching an initial broad search, evaluating results against explicit criteria, refining the query to surface high-value files, and looping up to three cycles. This incremental pattern discovery handles missing context in evolving codebases.

Yes, iterative retrieval applies to both bug-fix context gathering and feature implementation in evolving codebases. The four-phase loop evaluates and refines file searches whether the subagent task requires patching existing logic or building new features.

Avoid iterative retrieval when the exact codebase files needed by a subagent are already known upfront. Since the four-phase loop adds cycles to discover missing context dynamically, using it for predictable retrieval tasks introduces unnecessary overhead.