cmun2

I pre-registered a hypothesis that a fifth of public MCP servers break strict clients. Then I measured 617 servers and 14,804 tools. The answer was zero — and the findings that mattered were somewhere else entirely.

We cut review turnaround by about 60% by putting a model between Bitbucket and the team. Comment budgets, idempotency, and never blocking a merge mattered more than the prompt.

The same protocol carried device telemetry in an energy product and agent reasoning in a security product. The failure modes were completely different.

Independent deployment sounds good on a slide. Here is what Module Federation cost us, what it bought, and the three things that broke first.

Next.js

RequestAnimationFrame

SSR

Javascript

SSE

Websocket

Polling

ISO

Typescript

CI/CD

Typescript

Javascript

Typescript

Typescript

webview

Global Error

SSR

Firebase

i18n

HTTP

SSR

Pinia

Jest

Jenkins

Storybook

webapp

MSA

MSA

MSA

MSA

MSA

Jest

Jest

Git