How this site is made

Pipeline

1. Catalog. One YAML file lists the categories and the GitHub repositories in each.
2. Pinned clones. Each repository is cloned and its commit SHA is recorded. Every link on this site points at that SHA, so links never drift.
3. Machine wikis. Two open-source DeepWiki generators run on every clone: deepwiki-open (RAG over embeddings) and OpenDeepWiki (agentic). Both use DeepSeek V4 Flash. Their output is raw research material. It is not published verbatim.
4. Question answering. Each category has a fixed set of "how is X implemented?" questions. A code-reading agent answers all of them for each project, citing file and line ranges. The agent is Claude Code driving DeepSeek V4 Flash in an isolated container, with the clone mounted read-only. An answer can also be not applicable or insufficient evidence.
5. Citation check. A script resolves every citation against the pinned commit: the file must exist and the line range must be in bounds. Invalid citations are removed. An answer left with none is re-asked, and if it still has none, it is not published.
6. Editorial pass. A stronger model writes the project overviews and the cross-project verdicts. It reads the machine wikis, the README, the verified answers and the code. Where the wikis and the code disagree, the code wins.

What to trust

The answers are machine-drafted. Citation checks prove that each cited file and line range exists at the pinned commit. They do not prove that the prose is right. Use the citation chips to read the code yourself.

Inclusion rules

Vendor satellite repositories are skipped (SDKs, example collections, n8n nodes) when the vendor's main repository is already listed. Each project has one primary category.