/research-sweep

description: Parallel research sweep - N scouts investigate different angles of one question simultaneously
allowed-tools: Bash(node:*), Read(*)

You are the ORCHESTRATOR of a research sweep. Question: $ARGUMENTS

PLAN Decompose the question into 3-5 INDEPENDENT angles, e.g.: current-state-of-codebase, how-it-works-internals, alternatives-and-comparisons, known-pitfalls-and-bugs, docs-and-external-context. Generate run id: hive-<timestamp>.

SPAWN (PARALLEL - all Bash calls in ONE message): node "$HIVEMIND_HOME/scripts/oc-worker.mjs" --agent scout --run <run-id> --label <angle-name> [--dir <repo>] "<angle-specific research question>. Cite file:line or URLs for every claim. Max 150 words."

Scouts may use webfetch for external docs; they cannot edit anything.

CHECK PROGRESS between waves if any call errors: node "$HIVEMIND_HOME/scripts/oc-status.mjs" <run-id>

SYNTHESIZE yourself (do NOT just concatenate): merge into one coherent answer, resolve contradictions explicitly (state both claims + sources), list open questions that NO scout could answer.

FALLBACK: worker fails twice -> research that angle yourself inline, mark it [orchestrator-sourced]. REPORT: synthesized answer first, then evidence table (claim | source | confidence), then gaps.