Each ADR records a fork that was taken with a real trade-off, alternatives considered, and consequences worth surfacing. Future agents land cold and need to know why, not just what.
Next-generation from collapsed (not V1-original), git substrate, pure-workspace projection, no JSONB during generation.
One TypeScript Task primitive + N markdown prompts. Adding a new evaluation context = writing a new markdown file, not adding a TS class.
Per-teacher KB repo following Karpathy's LLM Wiki gist. Three layers (sources / pages / schema), four verbs (ingest, query, lint, file-back).
The original deletion mistook the baseline AND mistook Claude Code's pattern. Evaluator capability returns as gate prompts under ADR 0005.
Within-parent sub-agents dispatch sequentially; next sibling fires when prior reaches APPROVED. Cross-parent siblings parallelize.
The five ADRs interlock — 0004 establishes the substrate, 0005 the agent layer, and 0006-0008 layer specific capabilities on top. Reading order matches the build order (Wave R1 lands all five together).
Per docs/adr/README.md, ADRs are written lazily — only when ALL three conditions hold. Routine decisions stay out.
The 2026-05-27 redesign does not touch the existing ADRs. They remain canonical for their respective scopes.