Ryan Dahl says the whole Deno team is joining Cloudflare. Deno keeps monthly bug and security releases for a year, then Deno development ends. Deploy runs for six months.

Today’s other thread is cheaper judgment: a developer who stopped caring which model he was talking to, a Mac arrow that points at the button only you can press, and a Microsoft model that scores options instead of writing prose.

Cheap, fast and contained beats clever, when the work is routine.

🦕 Deno’s clock just started

Ryan Dahl’s post says the entire Deno team is joining Cloudflare. They will keep shipping monthly bug and security releases for another year. After that they will end development of the Deno runtime. Deno stays open source, and he says others are welcome to continue it. Deno Deploy shuts down after six months.

The stated plan: paying Deploy customers get migration support to Cloudflare Workers. JSR keeps running, with its infrastructure moving to Cloudflare. The team will keep supporting rusty_v8 and work toward integrating it into workerd. Dahl says the new direction is celld, built on the Workers programming model, with Durable Objects as the base for agent harnesses. That is his plan, not a shipped product.

Announcement  ·  HN discussion

💸 He stopped checking the model name

A developer who writes at dgt.is says he used DeepSeek 4.1 Flash heavily for a month across a dozen projects. Mid-session, he says, he often could not tell whether he was using it or Opus. He pays $10 a month for OpenCode Go and says he has rarely gone over about $1 of expected cost in a session. For occasional critical work he asks Opus 5.5 for a final review, then has DeepSeek apply the fixes.

The change he describes is behavioral: cheap models make it easy to start scratch tasks, UI monkey testing and tidy-up chores. He calls this subjective experience and links a benchmark chart we did not verify. The post also makes claims about cache size, energy use and training data. We did not check those. HN gave the post a lot of attention; attention is not evidence.

Author’s post  ·  HN discussion

➡️ A big arrow for the button only you can press

Franz Enzenhofer’s README starts with the familiar failure: the agent can refactor a repository, then asks you to click Allow in a terminal you are not watching. His macOS tool draws an arrow and a sign over every window. Clicks pass through, keyboard focus stays put, and the arrow removes itself.

bigarrow is a Swift command-line tool plus a skill for Claude Code and Codex. It points at a named button, such as the Allow button in System Settings. It never clicks it. The README says drawing needs no macOS permission, while finding a button or window can require Accessibility or Screen Recording for the app that starts it. Several demos are staged scenes, labeled that way in the README. Project tests and CI numbers are the author’s. We did not run it.

Repo + README  ·  HN discussion

🧭 A model that only answers with options

Microsoft says Xbox Research used Microsoft-Decision-1 on more than 10,000 open-ended pieces of feedback from surveys, Steam and X. The job was sorting them into themes the researchers had already chosen. Microsoft reports quality competitive with GPT-6 Sol, 14 times faster and 200 times less expensive.

Decision-1 is Microsoft’s post-trained Qwen3.5-9B. Given a fixed list of answers, it returns a calibrated probability for each option. It handles yes or no, multiple choice, ratings and rubric grading through a structured API call. Microsoft says it tested 36 additional public and private benchmarks across nearly 150,000 questions, and that decisions flipped on 1.3% of perturbations. All of those figures are Microsoft’s own. We did not call the model.

Microsoft launch post  ·  Foundry model page

🚶 Turn the reading pile into a walk

Nazar Boyko built Tabs to Trails for the reading backlog he never got through. He picks a walk length and the app chooses saved articles that fit. His reported test: a ten-minute request produced 10:15 of audio, and a twenty-minute request produced 19:39, on an M5 Max.

A bookmarklet saves pages. Local Gemma 4 condenses text only when it does not fit. A Kokoro voice reads it, and ffmpeg builds one MP3 with a halfway cue. The README says its number guard checks numbers in condensed text against the source. It does not claim to catch a changed meaning, which is exactly where a rewrite can go wrong. Setup needs Node.js 22 or newer, Ollama, a 6.6 GB Gemma pull and a first voice download of about 330 MB. It serves a phone page on your local network. We did not run it.

Author’s write-up  ·  Repo + install  ·  Issues

📦 Microsoft’s sandbox has an audit switch

The MXC README presents a containment SDK for code execution rather than another agent harness. Think of the moment an agent runs a command you have not read. MXC is meant to place that command inside a policy-controlled boundary.

The Rust, .NET and Node SDK uses a Windows process container, Linux bubblewrap or macOS seatbelt by default. The README marks some backends experimental, and network support depends on the backend. Its --audit mode disables sandbox security and is not for untrusted code. Optional Windows telemetry has several opt-in conditions. We read the README and issue list, but did not execute the SDK.

Repo + README  ·  Open issues  ·  HN discussion