research-url

One URL = one named subagent = one report file. The parent never reads the source; it verifies the report landed and folds new learnings into .superstack/learnings.md.

Workflow

1. Pick the report path: <repo>/research/<short-slug>.md. Default repo is ~/masumi/agentsmith unless the user names another. Create research/ if missing.
2. Spawn one general-purpose agent per URL (parallel when multiple URLs arrive together). Name it research-<slug>. Use the brief template below verbatim, filling URL, report path, and today's date.
3. If a later URL is another file of the SAME source (same gist, same repo), SendMessage the existing agent to fold it into the same report. Do not spawn a second agent per file.
4. When the agent reports done, run the landing check yourself: wc -l <report> and head -8 <report>. The agent's claim alone is not evidence; a dead agent and a clean result look identical.
5. Fold learnings: read the report's "Rules to adopt" section, append only entries NOT already covered in .superstack/learnings.md (check existing keys first), format per the learn skill (kebab key, one-sentence insight, confidence, Source: learn, Files: the report path, date). Tag untestable claims REPORTED in the insight and score them 6-7/10.
6. Tell the user: report path, line count, what was new, what failed to fetch, and whether all sources share one author (single school of thought is a finding).

Brief template

Fill the three slots. Keep everything else, including the style constraints, verbatim.

Research task. Source: <URL>