Today's issue is about tools that check their own work — an image model that reasons before it renders, a coding agent that audits its own setup, and a resolver that finally admits when it stopped verifying.
An image model that reasons before it renders.
Meta's Superintelligence Labs shipped Muse Image, its first image model — and it doesn't map a prompt straight to pixels. It runs as an agent: searching the web, calling coding tools, and self-refining draft after draft until the composition holds. That's how it lands legible QR codes, real mathematical diagrams, and clean multi-photo blends.
One command to de-bloat your setup.
Claude Code rebuilt /doctor into a full setup checkup — /checkup is the alias — that reads your install back to you and offers to fix it. It prunes unused skills, MCPs, and plugins that eat context, dedupes local CLAUDE.md against the committed copy, splits an overgrown root file into nested ones, and disables slow hooks. Every change is a plan you approve first; edits land in your working tree for a normal git diff.
Your resolver now admits when it stopped checking.
On 3 July, Albania's .AL registry rolled a new DNSSEC key but never updated the matching DS record at the root — so every .AL domain failed validation and went dark for validating resolvers. Cloudflare restored access with a Negative Trust Anchor that suspends validation, but that hid a problem: you couldn't tell a verified answer from a bypassed one. Now 1.1.1.1 returns Extended DNS Error 33 whenever an NTA is in play.
That's today.
A thin, deliberate three — the strong signals that cleared the ledger without repeating a story already shipped this fortnight.