Cloudflare would now like to verify a human is actually at the keyboard.

The premise is stated in one sentence in the README: "when most code is written by machines, the humans reviewing it need the language itself to answer 'what can this touch, and how sure are we' without reading every line." Jacquard's answer is threefold. Effects live in type signatures — (text) ->{net} text declares that a function may touch the network, and the runtime rejects unhandled world effects unless that authority is explicitly granted, so the blast radius is a fact of the signature, not a hope about the body. World handlers are swappable — the same program runs against the real network, a scripted fake, or a recording of last week's traffic. And identity is structural: Jacquard hashes canonical resolved structure rather than source bytes, so comments, formatting, and renames don't invalidate test results — the reviewer re-checks meaning, not cosmetics. It ships as an OCaml checker, a CPS interpreter, a C-emitting native backend, a CLI (run, check, hash, diff, test, build), and a test framework with content-addressed caching; precompiled binaries for Linux and macOS. Apache-2.0, created eight days ago, 81 stars, 90 points on Show HN.

Reach for it when the question about an agent's diff isn't style but reach — did this PR grow a network dependency, and do I have to read every line to find out. The contrast with Skillscript (yesterday's pick) is the interesting part: Skillscript gave the agent a substrate to write down what it knows; Jacquard gives the human a substrate to check what the agent wrote. Delete the read-every-line audit whose only question is "what can this touch." Tradeoff: version 0.1 is a research prototype and says so — no concurrency, no package management, no production claim — and a language designed for reviewability makes the same adoption bet Skillscript made yesterday, one seat over: the value only lands if the machines actually write in it.

An npm library, CLI, and browser bundle that converts semantic HTML fragments — the format every model emits natively — into native, editable Word documents: real OOXML paragraphs, runs, lists, tables, and images, with headers, footers, cover pages, and tables of contents. The default path is pure JavaScript on Node 20+, no browser and no Playwright (Playwright is optional, for computed styles and chart rasterization), and one command works without writing code: npx dom-docx input.html -o output.docx. Fidelity is checked by an automated visual regression suite against 30+ human-validated cases rather than asserted. MIT, 287 stars, created nine days ago, 146 points on Show HN.

Reach for it when the agent's deliverable goes to someone whose review medium is tracked changes, not a diff — the report that dies as a PDF lives as a .docx the recipient can rewrite. This is the production side of a loop we covered from the grading side on 07-08: docx-cli scored an agent's Word output against Word's actual rendering; dom-docx is how that output gets made from the agent's native HTML in the first place. Delete the copy-paste-into-Word step, and the PDF nobody downstream can touch. Tradeoff: it explicitly refuses complex layout — CSS grid and float, forms, table rowspan are out of scope — so it's for documents, not pixel-perfect design; and the fidelity number rests on the author's own 30-case suite, honest but self-graded.

The Show HN pitch is "an Excel-compatible runtime for agents and humans" (80 points), and the shipped thing is a native macOS app that opens and edits .xlsx locally — formulas, styling, tables, charts, pivots, "every Windows Excel shortcut, preserved" — with Claude, Codex, and Gemini connecting directly to the live sheet: "Your AI connects directly to your spreadsheet. Nothing in between." Files stay standard .xlsx ("open your files in Excel anytime"), everything runs locally ("your data never leaves your Mac"), and the beta is a free .dmg download with no account.

The spreadsheet is the oldest human-legible interface in business, and the standing workaround for agent spreadsheet work is a shuttle: export CSV, hand it to the agent, re-import, re-apply the formatting the round-trip destroyed. Making the spreadsheet the agent's runtime instead of its export target is today's theme in its most literal form — human and agent operating the same artifact, in the human's format. Reach for it when that CSV shuttle is your loop. Delete the shuttle. Tradeoff: closed source, macOS-only, and a beta with core editing still landing (VBA and full editing are on the roadmap) — and the agent-integration details are the thinnest part of the site, so the "runtime for agents" claim currently outruns the documentation. Fresh bet; the pitch earned the slot, the app has to keep it.

Sx is an open-source package manager for AI skills, and 2.0's move is entirely about who gets to participate: a native desktop app for Mac, Windows, and Linux that you point at a shared folder in Dropbox, Google Drive, OneDrive, or iCloud. Drag skills in as markdown files; teammates' apps sync the folder and auto-translate each skill into the native format of Claude Code, Cursor, Copilot, Codex, Gemini, Cline, or Kiro on install. No git, no terminal, no server, no accounts — the vault is plain markdown on disk, greppable and editable in anything. Apache-2.0, brew-installable CLI for those who want it. Disclosure: sx crossed our pool on 05-16 as a v1 developer CLI; the 2.0 story is that it stopped requiring you to be a developer.

This is the distribution beat of the packaged-expertise thread this newsletter has tracked since 06-11 (security → standardization → supply → optimization → typed skills), and it suggests the answer to our 07-01 watch — "where's the skills marketplace?" — might be: nowhere. Skills are org-internal documents, and Sx bets they should travel like documents, through the file share the org already trusts, written by the operations person who was never going to open a pull request. Delete the "paste this into your CLAUDE.md" Slack message. Tradeoff: folder sync means folder-sync semantics — no versioning, no review gate, no signature on what your teammates' agents auto-ingest. A shared folder of instructions that seven harnesses execute is a supply chain with no check on it, which is the 06-11 skills-security edition's warning wearing a friendlier UI.

The slate's quietest pick (9 points, 70 stars) and its most precise design. Kassette makes agent workflows durable by decomposing durable execution into two halves and only shipping the one you're missing: it provides the journal — an append-only, immutable record of completed steps on your filesystem or any S3-compatible store — and your existing infrastructure (a queue, a webhook, a job runner) is the dispatcher that re-invokes the run after a crash. Replay skips finished steps and resumes at the first incomplete one. ctx.step() marks recordable work, ctx.parallel() runs branches, ctx.suspend() parks a workflow awaiting human input with the process exited entirely — "no compute while waiting" — and fork() branches a new run from any earlier journal point. It embeds like SQLite: no server, no dedicated runtime, no new database. MIT, npm install @usekassette/kassette, created four weeks ago.

Reach for it when an agent workflow dies in a serverless function after the expensive model call and the side effect that already happened. Two threads land here. suspend() is today's theme as a primitive — human-in-the-loop as an engineered zero-cost state, not a polling loop. And fork() is the verb this newsletter has watched for since 07-08: replay-as-execution, branching a recorded run down a different path — shipped one layer below where we expected it, against workflow journals you wrote with kassette's primitives rather than your harness's transcript, so the watch stays half-open. Delete the retry-from-scratch after step nine of ten. Tradeoff: TypeScript-only; journaling is at-most-once, so side-effect idempotency stays your job; remote journal uploads grow quadratically with run length and nothing cleans them up — all stated in the README, which is why a 70-star bet reads better than most launches.

Off the thread but worth knowing: Clawk (github.com/clawkwork/clawk, Apache-2.0, 382 stars, created eight days ago) was the day's biggest product at 200 points — disposable, network-restricted Linux VMs for coding agents, no Docker, no qemu, no sudo. The differentiated part is where enforcement lives: the VM's entire network layer (gateway, DHCP, DNS, NAT) is a userspace stack inside the host daemon, so the DNS-aware allow-list sits where even root inside the guest can't touch it — the enforcement-beats-instruction pattern applied to the network. It's off today's axis and the sandbox shelf here is well-stocked (06-18, 06-23), but it's the strongest new entry in the category since; macOS Apple silicon is stable, Linux experimental. Meanwhile the skills supply side stirred on the same day Sx shipped distribution: google/skills (14.7k stars, created March, Apache-2.0, first crossed our pool 06-08) is trending again — Google's first-party skills library re-asserting the supply beat — and cangjie-skill (2.8k stars, created April, MIT, Chinese-language docs) distills books, long videos, and podcasts into executable agent skills, a sibling of our 07-01 pick book-to-skill. Both established, both disclosed as such.