Inside BMAD+: how 5 prompted personas actually delegate (and when they fail)
A technical post-mortem of Nestor's 5-persona BMAD+ stack — Atlas, Forge, Sentinel, Nexus, Shadow. How the loader assembles a layered system prompt at preflight.ts:208, how delegation works without separate processes, and the three failure modes (tool-whitelist hallucinations, persona context bleed, router cost explosion) that keep the whole thing behind NESTOR_PERSONA_V2=1.
Splitting three god-objects in a TypeScript monorepo: −56% LOC, zero API breaks
The v3.5 release is mostly a refactor. store.ts went from 5828 to 1902 lines (−67%), runtime.ts from 2107 to 1372 (−35%), and shell.ts from 2973 to 1327 (−55%) — across 50+ new focused modules with the public API unchanged and 258 tests green throughout. Here are the three patterns that made it tractable.