What shipped: the weekend of August 6–August 10

What the weekend actually looked like

Not much sleep lost over any single thing this weekend, but a lot of small plumbing changes that add up if you're the one holding the pager. I spent Saturday morning reading changelogs with coffee instead of doing anything useful, which is either a good sign or a bad one depending on the week.

Google puts ordering inside Maps itself

Google has added food ordering to Ask Maps, the AI chatbot living inside the Maps app — you can type something like "order a pepperoni pizza for me to pick up on my way home" and it finds an open restaurant, applies your saved preferences, and builds a cart. Right now it only works with restaurants on Square or Toast for online ordering, with Uber Eats integration planned, and Google hasn't said when this rolls out more broadly, where, or what it costs anyone on the restaurant side.

Bun.serve ships as a native Vercel Functions entrypoint

Vercel now lets you deploy Bun.serve() directly as a function, no framework wrapper, no rewriting to Node conventions — including WebSocket handling via server.upgrade(request). You set "bunVersion": "1.x" in vercel.json and it just runs. WebSocket connections bill on Fluid compute's Active CPU model, so you pay for time spent processing messages, not for idle connection time, though each connection is pinned to a single instance so cross-instance coordination still needs an external store. This is a developer thing, not a restaurant thing — but if you've got a KDS or an order-status board pushing live updates over a socket, this is the kind of deploy target that used to mean standing up your own box. I want to try wiring a small one up just to see what the cold start looks like.

Claude Code defaults to Auto mode on Pro, Max, and Team

Starting August 14, Claude Code switches its default behavior on Pro, Max, and Team plans to Auto mode — picking models and actions without requiring a manual approval on every command. Anthropic's own numbers, from a study of 1,053 developers, found humans approved 13.6% of commands the study flagged as dangerous, while Auto mode blocked 89% of them; separately, third-party testing across 72 indirect prompt-injection scenarios found none of 720 attack attempts succeeded against Claude Fable 5, Opus 5, or Sonnet 5 running Auto mode. Simon Willison has said he wants more independent confirmation before he'd call prompt-injection risk solved, and that tracks with how I feel about it too — the numbers are the kind of thing I want to see replicated by someone with no stake in the result before I stop watching the terminal.

GitHub Copilot code review gets two effort levels, generally available

Copilot's code review now ships with Lite and Balanced effort levels out of preview — Lite for routine changes, Balanced for anything complex, security-sensitive, or crossing services. They're renamed from the preview's Low/Medium, and you can set the level per-review, per-repo, or org-wide, with the applied level showing up in the PR timeline. Available on Copilot Pro, Pro+, Max, Business, and Enterprise. This is squarely a dev-workflow knob — nothing here touches anything I'd run in a kitchen — but if you're managing review load across a team, being able to cascade an org default and still let individual repos override it is a real ergonomics improvement over a single fixed setting.

Cloudflare folds Workers AI and AI Gateway into one control plane

Cloudflare is merging Workers AI and AI Gateway into a single control plane — one AI binding, one /ai/ REST endpoint — collapsing what used to be two separately-managed products for observability, billing, security, and logging. Default gateway usage now gets automatic logging, token tracking, and cost attribution with no setup, and there's a shared credit wallet you can spend across OpenAI, Anthropic, or Workers AI under unified billing, with elevated rate limits for Workers AI specifically. Model-first routing with automatic failover and a "smart routing classifier" are both listed as coming soon, with no dates attached. No migration timeline or pricing structure yet either, so this is one to watch rather than move on — I've got enough spread across providers right now that a shared wallet with real cost attribution is exactly the kind of thing I'd switch for, once I can see the bill.

Skill packs land on skills.sh

skills.sh now supports "skill packs" — bundling multiple agent skills from community sources, local folders, zips, or GitHub repos into one shareable pack with its own unlisted URL. Install with npx skills add against the pack URL, update with npx skills update, browse or build packs at skills.sh/packs. No stated limit on how many skills go in a pack, and no pricing attached. I've got a handful of skills scattered across projects that really should be one pack instead of five separate installs — this is the kind of small utility feature that doesn't announce itself well but saves real time once you use it.

GitHub Models is retired

GitHub Models — the free playground giving unified API access to multiple LLM providers, usable straight from GitHub Actions with existing credentials — went away on July 30. One developer who wrote about it didn't notice until August 9, when an Actions workflow broke. GitHub hasn't given an official reason; the developer's own guess is that agent-driven usage made the free tier too costly to keep running, though that's speculation on their part, not anything GitHub has confirmed. Their fix was switching to an OpenAI API key with a monthly spending cap, running what they call "GPT-5.6 Luna" for a folder-summary workflow. If you had anything wired to GitHub Models in CI, this is the weekend you find out about it the same way they did — a broken build.