wiki-research-loop

Run research-loop.js with the run command, a wiki slug, and optional flags like --max-pages, --max-depth, --budget-usd, and --fetchers. The loop pops seeds from the queue, fetches sources, compiles pages, and enqueues follow-up topics until a halt condition trips.

Three built-in fetchers ship with the skill: web via DuckDuckGo Lite, arXiv via its public export API, and GitHub via repository search. You can add custom fetchers by dropping a JavaScript module into ~/.pro-workflow/fetchers/ following the documented interface.

Create a file at ~/.pro-workflow/STOP and the loop halts gracefully on its next iteration check. You can also cancel all pending seeds for a wiki with the cancel command, or rely on automatic halts from budget, depth, or convergence limits.

Private wikis marked with private: true in wiki.config.md refuse all non-local fetchers and emit a warning. Only manual seed ingestion through local raw content is allowed, preventing any data from leaving the machine.

After each compiled page, the loop computes Jaccard token overlap against the prior three pages. If fewer than 5 percent of tokens are novel for three consecutive pages, it reports convergence and halts the run.

The loop requires the built store at dist/db/store.js in the pro-workflow repository. Run npm install and npm run build from the repository root before invoking the loop, since the script exits if the compiled store is absent.