Deterministic, auto-regenerated codebase inventories embedded in your docs — the structured index that lets Claude Code stop re-deriving your architecture.
Your agent already generates the data to predict its own build failures. How we score plan steps with Bayesian pattern confidence and cascade-risk analysis.
Our real production setup, published: the hook roster, five subagents with model tiers, eight MCP servers, and the machinery that makes agent output shippable.
Same model, three tools, one migration — measured. What Copilot's upgrade agent, plain Claude Code, and a guarded setup actually delivered, warts included.
Edit-thrashing is the quiet failure mode of coding agents — A→B→A loops that burn tokens without progress. The hook that stops it, full implementation included.
Why every C# coding agent needs compiler-grade context, the nine tool methods worth exposing, and the design rules that keep a Roslyn MCP server fast.
A complete, copy-pasteable CLAUDE.md for .NET 10 + Aspire + EF Core 10 — annotated section by section, with the reasoning behind every rule.
