Philip Z

Shelley's new compaction algorithm: a tool that the model calls to "squish" and "trim" prior messages. Joint work with @maisem.com. Read all about it at blog.exe.dev/compaction.

CI is the eternal battle between isolation and caching? This post brought to you by leaked processes, overflowing /tmp tmpfs, ephemeral port exhaustion, mismatched build flags, GitHub latency, and on and on and on. Also, blog.philz.dev/blog/what-is...

Certificate revocation is cool. Have we built software/library revocation? So you can prevent, say, libbrotli from running after the fact?

TIL: hovering over the preview in the menu bar green video icon is how you change your Mac camera's zoom

Really digging Shelley's new automatic compaction.

CI, ideally, is a deterministic function from your source tree to a boolean: are the tests passing or not? To make it fast, parallelize, isolate, visualize, analyze. It need only be as slow as your slowest test. My notes here— blog.philz.dev/blog/what-is...

It's "write (vibe code) your own e-mail client" o'clock over here. Trying to implement "unsubscribe and archive all messages from same sender." Google throwing up speed bumps every step of the way.

My dev process for iOS apps: shake it like a Polaroid picture. blog.philz.dev/blog/shake-t...

If I were to (as the story goes) get a job at Google, fix one bug, and quit, this is the bug: a given calendar event should only be shown once! I went through the motions; Calendear for iOS is on the App Store: apps.apple.com/app/id681658...

I’m giggling at Luna (in Shelley) playing Tetris over my shoulder. The pattern is handy for ops work wbere you want an agent to double-check the logs. Not just any agent… the one you usually use. blog.exe.dev/agent-over-y...

"Is your product legible to coding agents as users?" I can just smell Christensen's Innovator's Dilemma as I switch from better, more polished, established products to ones that let me point an agent at them. Of the nine points at blog.exe.dev/revisiting-j..., this is the most important one.

sqlite > jsonl

Experimenting with Pliable apps. apps.philz.dev has a download if you want to try it.

New prompting technique in Shelley: talk at your computer while it records your screen. I grossly underestimated how effective this would be. Here's a recording from a boring bug I fixed. It’s that easy. blog.exe.dev/show-and-tel...

I recently reverted to old habit and wrote a "Jenkins job" (it was Buildkite, but pot-ay-to pot-ah-to) to nag me about Fireworks model updates. Eventually, I came to my senses, and used a VM with a timer and an agent instead. So much better.

Group chat for agents! blog.exe.dev/mayfly-chat

Handy prompt of the day: Produce a mermaid diagram or similar with a full sequence diagram of how <X> happens including all the players and rpcs and such

TIL from @commaok.xyz: git plumbing to synthesize a commit equivalent to resetting to main and cherry-picking a single commit c. git commit-tree -p main \ $(git merge-tree --write-tree --merge-base c^ main c) c^ as "merge-base" "subtracts" the delta b/w main and c^. philz.dev/blog/git-vir...

> Cache reads now cost $0.25 per million tokens, 75% less than Fable 5 This is a *huge* price drop, given that cache reads inevitably end up most of your costs, because of the quadratic nature of LLM conversations. blog.exe.dev/expensively-...

The Joel Test was 26 (!) years ago. I revisited it at blog.exe.dev/revisiting-j.... What questions did I leave out?

A wise person once told me that the best wireless microphone is almost as good as a wired microphone. I got a new wireless split keyboard, and really feeling that right now. The right half drs es e craz My kingdom for a split traditionally staggered keyboard with a trackpoint for the index finger.

fzakaria.com/2026/08/24/a... code is data. Love me some binfmt_misc and SQLite hijinks.

Gave Shelley (gpt-5.6-sol, in this case) a pointer to www.texttocad.dev and English instructions of a keyboard tilt thingamajig, and it one-shot it. (Keyboard is NocFree&; not my favorite.)

Ian's solution at blog.exe.dev/we-taught-sq... is clever. Associate cache entries with the WHERE clause they were produced by and invalidate when updates hit the same WHERE clause. Made possible by some code-generation and a single-writer database.

TIL: memfd_create will let me ship a single file binary that needs to bundle a different single file binary without unpacking it to tmp. As the joke goes, I had one problem. I used an obscure Linux syscall. Now I have two problems.

I wrote up what I meant by this jeremymorrell.dev/blog/extensi...

blog.exe.dev/athena-deplo... 💯 Never going without this again. Build an agent to babysit your deployment. Give it only the code, the git log, the metrics, the logs, a way to send you messages, a way to pause the deploy. This is the way.

Friends don't let friends complicate their deploys with Spinnaker. Build the 5% you need; having your own deployment console and CD system is fantastic. It's been a game-changer for our deploys at @exe.dev. turbopuffer has one too: turbopuffer.com/blog/control...

If you use gRPC this is your reminder that the default max message size fits on a very short stack of floppies.

A month in the lab can save a day in the library. But what if it’s the agent in the lab (vm)? And it doesn’t take a month? blog.exe.dev/planning

An under-appreciated benefit of running sqlite is seeing your query execution in the same profiles as your application logic. If running a database separately, you have to work to connect the dots when the database is busy.

Your agents need computers. Giving the agents computers without persistent disks would be like having to pick up a freshly blank laptop everyday. (See also blog.exe.dev/persistent .)

In which I write some reflections on No Code at blog.exe.dev/the-end-of-n....

I am adamant about keeping bots out of human spaces. Also, about linking bot messages back to their code and runtime environment. blog.exe.dev/botiquette

An embeddable terminal library is such a boon. To get scrollback to work across a handful of terminal surfaces, we built it into the session manager. Read all about it! blog.exe.dev/exe-scroll

SaaSpocalypse happens like bankruptcy: gradually first, then all at once. Just this past month, I’ve built several bots/agents for work that correspond to products I’m “supposed” to have bought. Sam Schillace wrote “what if I can just do it faster than talking to you?” That applies to SaaS.

ssh -R and -L are so powerful. Needed to let agent on my VM connect to iOS sim on Mac, far behind a NAT. 1. ssh exe.dev ssh-key add --vm vm ... 2. mac: cat >> authorized_keys 3. mac: ssh -R 1111:localhost:22 vm.exe.xyz 4. vm: ssh -p 1111 user@localhost 5. Shelley happily driving the simulator

Four quick prompts to make a custom feed bsky.app/profile/did:... for myself on an @exe.dev VM.

Once you've got a stable of agents right in your pocket, you can't go back! www.youtube.com/watch?v=BhBF...

A rapidly increasing share of the software I use on a daily basis is bespoke. It makes bugs and papercuts in other software ever more exasperating. Once you've experienced mutable software there's no going back. I think blog.exe.dev/customizing-... may be most important thing ever on the exe blog.

As far as I can tell, S3 charges you based on the uncompressed size, but surely they compress. It is a failure of their APIs that uploading uncompressed is easier than compressed. LLM providers will also eventually compress your tokens (eg call out to another model to reduce tool output).

"The solution was fairly straightforward" blog.exe.dev/a-tale-of-tw... by Maisem is very cool. Learning cool stuff at work is fun.

We 🐡🐡🐡 exe.dev -- simple, effective, fast. If you've noticed improvements in our user experience over the last 6 months, they were accelerated by how simple exe.dev made things.

Plugin architectures are awful, right? Either it’s a versioning dependency mess or you’re coding in config files. I’m tickled by how well blog.exe.dev/customizing-... works. The upgrade button does an agentic rebase.

The part about using the exe.dev iOS app to create a VM from a short idea as a replacement for just jotting down a note was a major 💡 moment for me

Coding agents: definitely on team monorepo!

For an api that’s essentially f: string -> string, the LLM providers have managed to break coding agents at basically every model release.

Always such a delight when folks use your product for cool stuff. bsky.app/profile/exe....