The headline this week belongs to supply-chain hygiene: Andrey Sitnik — the person behind postcss and nanoid — explains staged publishing, trusted publishing, and provenance in the clearest current npm release playbook out there. And as if on cue, GitHub announced npm now scans every package for malware at publish time — new versions can be held or blocked outright, which is either reassuring or slightly terrifying depending on your deploy pipeline.
The context engineering conversation keeps evolving. Anthropic cut over 80% of Claude Code's system prompt for the newer models — smaller prompts, progressive disclosure through skills, CLAUDE.md and memory instead of prompt bloat. Meanwhile, Alessandro Fazio walked through every journal file, cache, and index behind a Codex conversation — a file-level tour that changes how you think about what "delete thread" actually means.
Tooling was unusually interesting this week. Vercel Labs' scriptc compiles ordinary TypeScript into native executables with ~2ms startup and 170–200KB static builds — no Node, no V8. Octane compiles React's programming model away entirely ahead of time, no VDOM, no hooks bookkeeping. TurboFieldfare runs Gemma 4 26B on Apple Silicon in roughly 2GB of RAM by streaming weights rather than loading all 14GB at once — a cheat code for local-model experiments. And Claude Code Merge Queue solves a very 2026 problem: parallel coding agents racing each other to push.
On the security and observability side: Figma's agent-based vulnerability workflow hit 70% precision before allowing agents to comment on PRs, and a shared policy became a de facto threat model. Anthropic's Claude weakened HAWK and sped up an AES attack by 800x, mostly unsupervised for a week — the strongest public example yet of AI doing real applied cryptanalysis. And MCP's largest update since launch goes stateless, finally letting remote servers run on serverless and edge infra.
The standout read is Canva's session revocation teardown: encrypted cookies, in-memory lookups, and why moving revocation records to S3 cut database load at hundreds-of-millions-of-users scale.
Also worth a look: Harry Roberts' Container Timing API piece brings per-component render timings straight from the browser (currently an origin trial), and Sentry's log structure guide makes the case for stable event names and lint-enforced attribute keys — the boring discipline that makes logs usable months later.
Enjoy!
Signup here for the newsletter to get the weekly digest right into your inbox.
Find the 15 highlighted links of weeklyfoo #148:
Session revocations at scale
by Canva
How Canva revokes sessions for hundreds of millions of users with encrypted cookies and in-memory lookups, and why moving revocation records to S3 cut database load and shrank the cache
🚀 Read it!, engineering, security, scale
The new rules of context engineering for Claude 5 generation models
by Anthropic
Anthropic cut over 80% of Claude Code's system prompt for the newer models — smaller prompts, progressive disclosure through skills, and context assembled from skills, CLAUDE.md and memory instead
📰 Good to know, ai, agents, claude
The life of a Codex conversation on disk
by Alessandro Fazio
Walkthrough of the journal file, caches and indexes behind a Codex chat — and why clearing a thread is not the same as deleting it
📰 Good to know, ai, agents, tools
How to structure a log
by Sentry
Stable event names, scoped attribute keys, primitive values — and enforcing the conventions with ESLint so logs stay queryable months later
📰 Good to know, observability, logging
Measuring component performance with the Container Timing API
by Harry Roberts
Per-component render timings straight from the browser, now in a Chrome origin trial
📰 Good to know, performance, web
Discovering cryptographic weaknesses with Claude
by Anthropic
A model halved the effective key strength of the HAWK signature scheme and sped up an attack on reduced-round AES by up to 800x, working mostly on its own for a week
📰 Good to know, ai, security, cryptography
MCP gets its largest update since launch
by MCP
The protocol goes stateless, so remote servers run on serverless and edge infra or scale horizontally behind a load balancer
📰 Good to know, ai, mcp, protocols
How Figma stays ahead of vulnerabilities with agents
by Figma
Precision over recall was the whole game — 70% before agents were allowed to comment on PRs, and the shared policy became a de facto threat model
📰 Good to know, security, ai, agents
The secure way to release an npm package in 2026
by Andrey Sitnik
Staged publishing, trusted publishing and provenance, explained by the person behind postcss and nanoid
📰 Good to know, npm, security, javascript
scriptc
by Vercel Labs
Compiles ordinary TypeScript into small native executables — no Node, no V8 in the binary, ~2ms startup and 170–200KB static builds
🧰 Tools, typescript, compiler, tools
Octane
by Dominic Gannaway
React's programming model compiled ahead of time — no VDOM, no rules-of-hooks bookkeeping, JSX and TSX supported
🧰 Tools, javascript, react, framework
Codex Security
by OpenAI
CLI and TypeScript SDK that scans repos for vulnerabilities, validates them, and tracks findings over time
🧰 Tools, security, cli, tools
TurboFieldfare
by drumih
Runs Gemma 4 26B on Apple Silicon in roughly 2GB of RAM by streaming weights instead of loading all 14.3GB at once
🧰 Tools, ai, llm, tools
Claude Code Merge Queue
by funador
Local merge queue for parallel coding agents, so simultaneous landings stop racing each other on push
🧰 Tools, ai, agents, tools
PGSimCity
by Nikolay Samokhvalov
Postgres internals as a 3D city you can walk around in and break on purpose
🤪 Fun, postgres, fun
Want to read more? Check out the full article here.
To sign up for the weekly newsletter, visit weeklyfoo.com.
Top comments (0)