roster-research

To conduct unbiased codebase research, use a blind documentarian approach that only reads provided research questions and never accesses task descriptions, ensuring factual findings tied to specific file:line references without assuming what should be built.

The best way to trace existing implementations is by generating file-grounded research reports that mandate verifiable file:line references for all claims, ensuring every factual finding about your current code is fully traceable and easy to verify.

To scan for prior art during codebase research, use a pipeline workflow with optional online prior art scanning to search for existing tools, patterns, or academic work, contextualizing your internal findings against broader industry approaches.

Yes, you can use parallel specialist sub-agents for fact-finding across a large codebase. The research pipeline supports an adaptive full execution mode that deploys parallel sub-agents based on the volume of research questions.

Yes, codebase research works for verifying non-speculative implementation details by mandating file:line references for all findings. It produces strictly factual, grounded analysis without inferring solutions or suggesting changes to existing implementations.

You should use a single sub-agent for codebase analysis when operating in the adaptive fast execution mode, which is triggered by a lower volume of research questions, whereas full parallel specialist sub-agents are deployed for higher question volumes.