AgentLore
Working NotPlanned taste test for issue #87. Public URL: https://notplanned.teahub.xyz/v/agentlore/
Buildless HTML/CSS/ES modules. Serve site/ with any static HTTP server; run node test.mjs for core checks. Reuses hb-5 mac-sites and the existing /v/<slug>/ venture pipeline. No backend, tokens, paid APIs, outbound notifications, or repository write path. GitHub requests explicitly omit credentials.
Scan up to three public GitHub repository URLs. Resolve the default branch commit, discover named instruction documents in its Git tree, read bounded blobs, extract exact instruction-like lines outside fenced code. Review candidate scope and duplicate/possible-conflict flags, select rules, preview and download AGENTS.md with pinned source links and original scope. Issue-based early access uses the existing teahubxyz/time intake; the user reviews and submits the issue themselves.
Limits: unauthenticated GitHub rate limit; recursive trees may truncate; 12 files/repo, 64 KiB/file, 240 candidates total. No semantic rewriting or language model. Heuristics can miss prose and subtle conflicts, and scope classification needs human judgment. Referenced docs are not traversed. License/context review remains with the user. Coverage and failed reads are shown. The app uses textContent for untrusted source text and never evaluates it.
Ready to send: “AgentLore finds useful agent instructions in public GitHub repositories and lets you carry a reviewed set into your next project. Every candidate links to its original commit and line. Try it, select what fits, and download AGENTS.md. It never writes to your repositories. Share a real use case through the GitHub early-access issue link.”