Source discovery

How loop discovery works

The directory starts with 34 human-reviewed seed loops and can ingest public code-source candidates into an admin review queue. Public pages should only publish useful, attributed, reviewed loops with original explanation and copy-ready Markdown.

1. Search public code sources

A controlled server-side job searches public files that mention goals, exit conditions, guardrails, agent rules, hooks, or verification commands.

2. Fetch and score candidates

Candidate files are fetched server-side, hashed for deduplication, checked for safety signals, and scored before they enter the review queue.

3. Classify with AI when configured

An optional AI classifier extracts the likely goal, check command, exit condition, supported agents, guardrails, and safety flags. It never publishes automatically.

4. Human review before publishing

Admins review each candidate, rewrite the page with useful context, and publish only loops that add value beyond the upstream file.

Publishing rules

Do not publish scraped text as if it were original. A public loop page must add structure, context, safety rules, source attribution, and a Markdown protocol that helps the user run the loop correctly.

Do not publish candidates with unclear licensing, destructive commands, validation bypasses, or low confidence without manual editing and explicit warnings.

Do not claim a source is reviewed or published until the candidate exists in the admin queue and has been processed by the discovery pipeline.