Agent runtimes

Knowledge refresh

Good agents go stale if their examples, tools, and repo patterns never change. Agentlas can load a live knowledge-refresh overlay before the meta-agent designs a draft.

What refresh means

The built-in knowledge base stays stable. A separate refresh file adds recent GitHub repository search results and notes about patterns worth inspecting. The meta-agent reads both when it creates agents.

How to run it

npm run knowledge:refresh
# optional: add GITHUB_TOKEN for higher GitHub API limits
GITHUB_TOKEN=ghp_... npm run knowledge:refresh

How to make it periodic

Run the command from cron, a scheduled GitHub Action, Railway cron, or another job runner. The output file is .data/knowledge-refresh.md. The app loads it automatically when present.

HEADS UP · Refresh is not blind copying

The generated agent should borrow patterns, not prose. For any current fact, official docs and live repo files still win over cached notes.