Research Sweep

Commits to an exact output schema (format, top-level list, ≤5 required fields, ID disambiguation, no fabrication) before launching, and pastes it verbatim into every agent prompt so outputs merge cleanly. Then splits the topic into 5–7 disjoint slices (by source, era, geography, category, or corpus) so each agent owns an angle no other touches.

Launches the agents in a single message with run_in_background: true, each prompt carrying the verbatim schema, its slice, source guidance, verification rules, a YAML-only output requirement, and a numeric volume floor. Handles the expected 1–2 partial failures by re-running, accepting partial output, or merging slices.

Extracts each agent’s final message from its JSONL transcript with a jq + awk pipeline that strips the YAML fence and redirects to one catalog file — never Read-ing the transcripts. Sanity-checks with wc/grep and commits immediately so a compaction can’t lose the result.

Runs a fresh verifier agent that didn’t produce the data: spot-checks a 10–20% sample against primary sources, hunts for fabricated URLs/dates/IDs, checks cross-agent seams for duplicates, and tests the volume claim and gaps. Dedup, normalization, and URL verification follow as separate passes.

The sweep now keeps its team in a coverage-roles registry at .claude/roles/research.md — the same mechanism crew and panel use, owned here by research-sweep. Composition proposes from it first (stable angles whose when-to-seat matches the information space) and mints fresh probationary rows only for the gaps. See The Role System.

Coverage angles are composed from the information space itself rather than fixed in advance, then presented as an editable roster the user signs off on before any agent launches. Standard angle archetypes ship as the starting palette — entity-scout, timeline-scout, source-scout, container-scout — and corpus-specific angles are probationally minted when the space needs a cut the archetypes don’t cover.

A persistent verifier — the shared skeptic role — and a synthesizer are standing seats on every sweep: the skeptic adversarially checks the merged result, the synthesizer assembles the cited output. These are the two seats people most often skipped, now seated by default.

What worked is written back to research.md as per-repo learning, so reruns in the same space start from a better roster. When shared core role files (.claude/roles/<role>.md) exist, rows link to them by name via the roles plugin and research lessons become graduation candidates for the core (user-gated). The skill is self-sufficient without the roles plugin — it manages research.md on its own and never downgrades when the shared core is absent.