Loops
Agents on a schedule, with a human approval gate
Skills are capabilities: instructions a coding agent follows when invoked. Loops are what happens when a skill runs on a schedule. An agent does real work against real data and brings the maintainer a proposal. Nothing merges or deploys itself; every loop ends in the same place, a human reviewing a branch. This page documents the loops this repo defines.
chat-qualityTwice a month an agent will read the site chat's own report card once the chat is live: every thumbs verdict a visitor left on an answer, joined back to the logged exchange it rates. Each disliked answer becomes a regression case in the chat eval's golden set, the eval runs against the updated set, and the fixes land on a branch: a new case, a corpus correction, a persona tweak. I read the report and approve or reject. Visitors' questions are paraphrased before anything is committed.
- Read-only on the live logs: it observes, never rewrites
- Visitor questions are paraphrased before they reach the repo
- One eval run per loop: the eval spends real API budget
- “No failures worth encoding” is a valid outcome
- A human approves every merge
link-rotOnce a month an agent collects every external link the built site renders, from docs references to footer links, and checks each one still resolves. The build already proves internal links can never break; the outside world offers no such guarantee. Every failure is verified by hand before anything is called dead, because bot-blockers fake failures, and real rot is fixed on a branch with a report I approve or reject.
- A failed probe is a lead, not a verdict: bot-blockers fake failures
- Link surgery only: swap or remove the href, never rewrite the prose around it
- Never pushes, merges, or deploys itself
- “Every link resolves” is a valid outcome
- A human approves every merge