◆ Swapping the Model Behind a Coding Agent · Part 1
How to Compare the Models Behind a Coding Agent with SWE-bench
How do you measure the model performance of a coding agent?
20 POSTS
Page 1
◆ Building coldpath · Part 4How Bundlers Remember Imports
Tracing a coldpath misclassification that read every synchronous dependency in the Turbopack analysis file as a static import. This post compares the graph information and build output of webpack, Turbopack, Vite, and esbuild, and looks at which clues reveal the original syntax and which cases cannot be recovered.
◆ Improving number-flow · Part 2Improving number-flow Performance: Moving Animations onto the Compositor Thread
In the number-flow fork I built to support old browsers, animations in modern browsers were still recalculating styles on every frame. This post records how 0.2.0 moved additive compositing into transform and opacity keyframes, how the change was measured, and the costs that remain in the mask and in interruptions.
Saving Marp Slides from a Next.js App for Offline Use
From running the existing Marp viewer offline to fixing what deployments, networks, and the presenter window exposed along the way.
Pages You Leave Don't Die, They Hide: Next.js cacheComponents and React Activity
Next.js 16's cacheComponents hides the pages you leave with React Activity instead of unmounting them. Following a three.js canvas that went blank after a back navigation, I read the router, React, and react-three-fiber source and reproduced it in a browser.
◆ Building coldpath · Part 3Tracing Toss Securities JavaScript Without Source Maps
Chuseok side quest, part 3: with love, to Toss Securities
◆ Building coldpath · Part 2Building a Bundle Analyzer with V8 Coverage and Source Maps
Chuseok side quest, part 2: building a V8 coverage analyzer, one dead end after another.
◆ Building coldpath · Part 1Tracing When This Blog's JavaScript Actually Runs
Chuseok side quest, part 1: starting with my own blog
I Grew Never-Executed JavaScript to 10MiB
How much do you lose when 10MiB of functions you never call ships to the browser? I ran 855 measurements to separate the cost created by bytes from the cost created by the shape of the code. Uncalled declarations cost about 21ms per MiB and did not grow when I slowed the CPU by 4x, while initialization code that runs the moment the file is read added 201ms per MiB to the main thread. With real libraries, merely importing a module was enough to evaluate it.
Running One Web on Multiple WebViews: Folding Environment Branches into Adapters
We supported the iOS and Android WebViews of our own app and of a partner app, and gathered the scattered environment branching into adapters. This is a record of how we organized insets, the bridge, and CSS, along with the problems we hit with the SSR seed and hydration.
The Real Cost of Drop-in Compatibility: What I Threw Away Porting markdownlint-cli2 to Rust
I ported markdownlint-cli2 to Rust and matched 264,114 diagnostics on this blog repository byte for byte. The commits that ported 51 rules landed in a single day, but turning that into a tool you could actually swap in took two weeks. A record of giving up a parser 37 times faster, correcting two performance misreadings, and finding 11 more compatibility differences in an adversarial review.
Tearing Apart a Third-Party SDK and Rebuilding It My Way: Without Touching a Single Line of Logic
I imported one constant and 97.7% of the bundle came along with it. The vendor had no timeline for a fix, so I pried open the published source maps, recovered over 400 TypeScript files, and rewrote the build, the entry points, and the dependencies however I wanted. Everything except the logic. That got /send to -77.5% raw. The hard part came after. All 1,932 tests passed, and a few of them were watching nothing at all.
◆ Service Worker Caching Deep Dive · Part 3Measuring the Cost of Going Through a Service Worker: Building in the Lab the Control Group GA4 Could Not Give Me
I set out to confirm the 500 ms hint that part 2 left behind, but the hard reloads that would form the control group arrive at under one a day. So I built the control group myself, with Playwright and a shaping proxy, and found that going through the worker costs 2 ms on a navigation, and that the cost is not latency but the bytes the worker fetches in the background on every click. The gap I had left between the lab and the field turned out, only after the post was written, to be a measurement definition difference created by 103 Early Hints. Third part of the service worker caching deep dive series.
◆ Service Worker Caching Deep Dive · Part 2Applying Service Worker Caching: App Router Traps and GA4 Field Data
Armed with the theory from Part 1, I made this blog (Next.js App Router) open offline. On the first deploy, the post I had just read would not open offline; on the second, posts opened but every image was broken. This is a chronicle of fixing, one deploy at a time, the traps created by soft navigation, prefetching, and next/image, and a record of settling the results with GA4 real-user data. Returning-visitor FCP improved by 634ms on average, while TTFB worsened by 525ms on average. The second post of the Service Worker Caching Deep Dive series.
◆ OG Scraping Server Design Notes · Part 2Building an OG Scraping Server in Node.js (2): How SSRF Gets Through
A feature where the server opens a URL the user handed it has the textbook conditions for SSRF written into its spec. Six ways a whitelist gets bypassed first, then five defensive principles that block them, all actually run on Node. Strip IPv4-mapped by hand and it gets through in hex notation, undici lookup hook is never called when the host is an IP literal, and URL.hostname keeps the brackets on an IPv6 literal. The final post of a two-part design note on OG scraping servers.
◆ OG Scraping Server Design Notes · Part 1Building an OG Scraping Server in Node.js (1): From Runtime Choice to Error Rate and Latency
The "10% error rate" of a link preview server is a single number that five different kinds of failure got mashed into. This post works out why this workload is I/O bound at that TPS, where runtime choice actually diverges across four points, and then moves on to lowering the error rate with User-Agent and encoding. Node built-in TextDecoder turns CP949 extension characters into different characters without raising an error, and a scraped og:title is not an API response but user input. It also covers cache stampedes, negative caching, and a two-million-run simulation that verifies "P95 under one second" by working backwards from the cache hit rate. The first post of a two-part design note on OG scraping servers.
Next.js Turbopack Turned One Singleton into Two: A Scope Hoisting Bug and a Circular Import
In a Next.js 16 Turbopack production build, a module-scope singleton became two live instances at runtime. Inside the same synchronous block, one condition contradicted the other, and responses that arrived in 30ms still timed out. This is the record of tracing the cause through the bundle output: a partial scope hoisting merge, a circular import, an upstream bug that had already been fixed, and the single-variable experiment I ran too late.
Killing Frame Drops in a framer-motion Banner: Two Wrong Turns and One Easing Function
A banner built with framer-motion made the entire home screen stutter for the 0.6 seconds it took to open. The record of guessing the cause from the code, having that guess overturned twice by measurement, and finally removing the reflow with a single easing function. Plus what the work left behind: the gap between declaration and execution, the principle that properties decide performance, preserving mechanisms, suspecting your instruments first, and proving sameness with a curve.
◆ Service Worker Caching Deep Dive · Part 1How Service Worker Caching Works: The Proxy, the Lifecycle, and Five Strategies
A service worker is a programmable proxy standing between your site and the network. Where does it stand, why does the cache rot, why am I seeing the old version after deploying, what goes in under which strategy, and so, should you use it? Holding on to five questions you actually meet in practice, this post goes down to the details of state transitions and to a real measurement in which a 104KB opaque response was accounted as 6.6MB of storage. It is the general theory that did not fit into the cache chapter of Frontend Performance Optimization Deep Dive (published in Korean), and the first post of the Service Worker Caching Deep Dive series.
◆ Improving number-flow · Part 1Porting number-flow to Old Browsers: Five Decisions and Two Reversals
The minimum versions where number-flow turns animations on are Chrome 125 and Safari 17.2. A record of the decisions made while building a fork that lowers this floor to Chrome 66 and WebKit 16.4, the two judgment calls I ended up reversing, and the Safari bug investigation that made me give up on automatic downgrading.