“Very often, we found that the role of the engineers was to tell Claude ‘no.’”
Anthropic’s source-control team, on moving a monorepo that takes more than 30,000 merges a day off git in five months. Story 01, and the thread through the week: tools built for the volume agents now produce.
Git couldn’t keep up with Claude.
Merges into Anthropic’s monorepo rose from about 2,500 a day in May to more than 30,000 in September, and CI ran 25 million checkouts a day against nearly 100 GB of history. Seven engineers built Scion, version control for clients that are always online: the server owns history, merges and blame, and each job fetches only the files it reads. Anthropic offers its server-side RPC, cache tiers and sparse checkouts as ideas to carry back to git and the forges.
30,000merges a day in September 2,500 in May
- 200×
- smaller CPU and memory footprint than the git mirrors it replaced
- ~100×
- faster CI checkouts: seconds where they took minutes
- 7
- engineers built it and a code-review tool, and migrated in five months
Open weights for the people who patch.
Mistral Large 4 is a mixture-of-experts model with about a trillion parameters, 52 billion active, in public preview on Mistral Studio at $1.36 in and $4.18 out per million tokens. It scores 82% on a test that asks a model to reproduce a real vulnerability and then patch it, the best result among publicly available models; Mistral says several closed models score near zero there because they refuse. Import AI’s caveat: it is the strongest Western open-weight model and still trails the leading Chinese ones.
“Weights drop end of this month.”
Mistral, announcing Large 4 on 6 October. Hosting includes a European deployment Mistral runs end to end; the license is not yet named.
06·10·26
- Reproduce, then patch82%
- Cybench, 40 challenges93%
- AA Intelligence Index38
- Per M tokens, in / out$1.36 / $4.18
Ask how it works. Get the machine.
Anatomy is an MIT-licensed agent skill that answers an explanation request with an interactive isometric drawing: shaded SVG parts with a live readout, optional WebGL fire, water and steam, and a 3D mode in beta. It installs with one command into Claude Code, Codex, OpenCode and others, needs no API key, and passed 14,000 GitHub stars nine days after the repo went up. On @denissexy it was paired with a sound-layer skill to rebuild historical photos as simulated scenes.
$ npx skills add wheresryan22/anatomy -g › /anatomy explain how a token bucket rate limiter works, interactive
- SVGa standalone page, a lone .svg, or React components
- SVG + WebGLfire, water, steam and light as shaders
- 3D, in betaparts that spin or break off in true 3D
One query in 100,000 came back empty.
Hyperdrive pools database connections for Cloudflare Workers and caches query results, and its MySQL beta failed on about 0.001% of requests, only where caching was on. The cause was a capability flag: the mysql2 driver can’t read CLIENT_DEPRECATE_EOF, so the node beside the Worker rewrote result bytes while the node beside the database did not, and both wrote their own encoding under the same cache key. If you cache wire bytes behind a protocol translator, this is the bug to look for.
FRAME 01
Hyperdrive talks to the origin with CLIENT_DEPRECATE_EOF on. The Worker’s mysql2 driver can’t, so result sets must end the old way.
FRAME 02
The client node translates for the Worker. The endpoint node keeps the origin’s bytes. Both cache under one key.
FRAME 03
Each node reads the other’s bytes: an untranslated answer, an early EOF, or a hang.
FADE OUT
Store one canonical encoding, whoever writes it. Client connection success: 99.99%. MySQL is now GA.
“the difference between beta and GA is just 20 bug fixes in a trenchcoat.”
Your old sessions are the eval set.
Ibragim went through about 1,270 of his own Claude Code, Codex and Pi sessions, kept ten tasks that a script can check, and ran agent setups on them three times each. On GPT-6.1 Sol both the lightweight Pi agent and native Codex solved every task, and Pi did it for half the money. His skill, evals-from-sessions, walks your agent through the same pipeline into Harbor tasks, and stops for your approval on the candidate list, the spec and the finished task.
- 1Mine
Copy logs from
~/.claude/projects,~/.codex/sessionsor Pi, scan the copy with gitleaks, list candidates. - 2Spec
Write down what the agent is told and what it has to discover on its own.
- 3Seal
Package a Harbor task with no future commits, reference diff or visible tests to peek at.
- 4Prove the checker
A no-op fails, the oracle passes, a near miss fails, a different correct fix passes.
Cost per task, all ten solved · GPT-6.1 Sol, medium
Star history, without the stargazers.
Star-history tools drew their charts by paging through every account that starred a repository, 100 at a time, through an endpoint that also let spam and phishing scrapers harvest usernames. GitHub now limits stargazer lists to a repository’s admins and collaborators and has added a star-history endpoint that returns daily counts with no user data, up to 30 weeks per call. If your dashboard reads the old list, it gets an empty response or a 403.
react/react
About 251,000 stars. Requests for a full history.
- before
- 2,511pages of 100 people
- after
- 24pages of 30 weeks
- Filter by date first, then group: about 251,000 rows read, against some 6 million had they grouped and then paginated.
- Stargazer requests reaching the database fell about 60%; time on count queries, about 80%.
“A smaller response does not always mean a cheaper query.”
A packet from inside the sandbox reached the host.
Vercel’s $1M bounty for escaping Vercel Sandbox drew 1,285 reports in two weeks and paid out $242,000. On day three Paulos Yibelo found a flaw in the Linux kernel’s networking stack: crafted packets from inside a microVM leaked host memory in production, and on matching KVM and Firecracker machines crashed the host or ran code on it. CVEs are pending, and the same kernel layer isolates tenants at most large clouds. If you run sandboxes for agents, the section on what Vercel stopped trusting is the part to copy.
1,285 reports, triaged
- KERNELPackets from a guest reach host memory. Exploit details: withheld until the kernel fix is public.
- STORAGEDevice-mapper handed recycled blocks from deleted sandboxes to new ones. Found by a Vercel intern.
- SNAPSHOTSThe control plane trusted storage details the guest reported. Bound to the owning project within a week.
“a boundary that depends on a secret staying secret is not a boundary”
Type a world, then walk around in it.
Odyssey-3 is a world model that turns a prompt into a place you move through in first or third person, updating as you act or drop in events, with a camera you steer separately. The free research preview runs Flash, a distilled version fast enough for real-time play. Odyssey also pitches it for training robots, and its own benchmark numbers put the Pro variant first on Physics-IQ Verified. API access is by request.
In @TochkiNadAI’s run every step on screen came from a keypress, and the character could interact with objects in the scene.
Your imports are the infra config.
VoidZero, the Vite company that joined Cloudflare, released Void: an open-source SDK, CLI and self-hostable platform shipped as one npm package and Vite plugin. Import void/db and you get a local database in development and D1 on deploy; void/ai wires up AI Gateway. void deploy builds, migrates, provisions and ships, and Void Platform lets a company give people a place to ship internal apps without a Cloudflare account per project.
$ npm install void # add the Void plugin to vite.config import … from "void/db" // D1 on deploy import … from "void/ai" // AI Gateway $ void deploy build · migrate · provision · ship $ void connect <platform-url> ship to your own Void Platform
“Void turns a Vite app into a full-stack Cloudflare Workers app.”
Works next to React Router, SvelteKit, Nuxt, TanStack Start and Astro. Auth through Better Auth; the platform needs Workers for Platforms, R2 and a Postgres database.
Two unauthenticated RCEs. Patch to 16.4.1.
The 14 October out-of-band release fixes two critical remote-code-execution bugs and one high-severity denial of service, all traced to upstream dependencies. The first hits App Router apps on Cache Components, which create-next-app has switched on by default since 16.4, and has no workaround. The second sits in image optimization of SVGs on glibc Linux, through sharp.
RCE when a use cache function returns multiple typed arrays. No workaround; upgrade even if you have no such function.GHSA-673q-mjm6-r5xr · React GHSA-xqc2-xf4c-x3h3
RCE in the Image Optimization API with SVG on glibc Linux. Stopgap: sharp 0.35.5 or later.CVE-2026-96889 · GHSA-2464-fhr4-w9xw
Crafted requests to Server Actions pin the CPU. Pages Router apps, and apps without Server Actions, are not affected.GHSA-p3m5-59mp-jv5r
npm install next@16.4.1 # 16.4, Active LTS npm install next@15.5.28 # 15.5, Maintenance LTS
That’s the week.
Issue 125, Saturday 17 October 2026. Ten picks from nine publishers, published between 6 and 16 October, seven of them since last Saturday’s issue. Three came to us through Telegram channels; Mistral Large 4 came through Import AI.
Rubric, in order: AI tools you could adopt this week · creative software and generative media · dev tools and agentic coding · privacy and security where the signal is high · science and research with a practical kernel · anything immediately actionable for a senior engineer or founder.
Also read this week and left out: The Batch, The Zvi, OpenAI News, Anthropic News, jeffbullas.com, PostHog, Sentry, @rvnikita_blog, @ProductsAndStartups. The newest fly.io post had already run here.