正在加载视频...

视频加载失败

Your AI coding agent is slow because it keeps waiting on itself. Search, wait, read, wait, edit, wait, test, wait. So I built Turnpack. It folds those steps together. Half the round trips, roughly half the time, 36% cheaper in my test. It also lets the AI test its...

26,560 次观看 • 3 天前 •via X (Twitter)

41 条评论

Terp 的头像
Terp3 天前

How I got those numbers: I gave Claude Code the same small coding task 3 times with Turnpack and 3 times without, then compared. The test script is in the repo if you want to run it on your own setup. Credits: Video by fable 5.2 XH in claude app

Mia 的头像
Mia3 天前

LET'S GOOOO 🚀🚀🚀

Terp 的头像
Terp3 天前

Just getting started. Alot more coming soon :)

Madeactual 的头像
Madeactual3 天前

The test wait is usually the longest one. During the loop, have the agent run only tests related to the changed files (jest --findRelatedTests, pytest-testmon, go test on the touched packages) and save the full suite for the final check.

鼬 的头像
鼬3 天前

most of the wait is round trips where the model just needed a grep result, so folding search and read together is where the time comes back. testing edits before they land saves another full loop

eToro Builders 的头像
eToro Builders2 天前

Search, wait, read, wait is basically the agent version of watching a progress bar. etoro Builders feel this too when their agents loop through etoro's API and MCP while building investing tools. Which wait does Turnpack fold next?

Hari Santhosh 的头像
Hari Santhosh3 天前

Folding search and read into one step makes sense, most of the waiting is the model asking for one file at a time. What happens when an edit fails the test mid-batch, does the whole pack roll back?

Jon Cheney 的头像
Jon Cheney2 天前

Does folding the steps hurt when the agent needs to change plans after a read? Batching search-read-edit is great until the read shows it grabbed the wrong file, so I'm curious how often it has to unwind.

Terp 的头像
Terp2 天前

nah, cuz he woulda had to do more tool calls before hand anyways to read it regardless

Mohak Bajaj 的头像
Mohak Bajaj2 天前

folding search and read into one step is where most of the wall time goes. how do you handle it when the edit depends on what the read found?

Neebarp 的头像
Neebarp3 天前

The snap-back when a test fails is the best part. Does it run the whole suite before an edit sticks, or only the tests near the files it touched?

Nash 的头像
Nash3 天前

Folding steps together is great until a test fails halfway. Does Turnpack roll back the whole batch or keep the edits that already landed?

Terp 的头像
Terp3 天前

Ive used billions of tokens with this tech your good king

小森 的头像
小森2 天前

the snap-back is the part i'd actually use. worst agent mess i get isn't slowness, it's a test failing halfway through a 3-file edit and leaving things half applied. does the rollback cover files it created too, or only ones it edited?

Big Head(大头版) 的头像
Big Head(大头版)3 天前

Mine waits longest on cargo build. Fold the round trips all you want, the linker still takes four minutes.

Terp 的头像
Terp2 天前

ahh i already solved this for my own setup but i havent released the majority of my tech this was just a small freebie for opensourse. I have to decide how i wanna share the rest because i worked really hard & idk how to share it & atleast make what i spent back haha

MrPointPlay 的头像
MrPointPlay3 天前

@MiaAI_lab King!

Juliano | Building easi.run 的头像
Juliano | Building easi.run3 天前

Read the rollback section. Could the --fix / FAILED_ROLLBACK_PARTIAL case sit beside the main demo? It clarifies when edited files stay on disk.

Sergio Marakis 的头像
Sergio Marakis2 天前

snapping every file back when a test breaks is the bit i'd trust first, does it catch new files too?

Terp 的头像
Terp2 天前

yea, experience has been really smooth

dng 的头像
dng3 天前

does the saving hold when the first edit fails a test and it has to change course? that's where batching gets interesting.

Snazzie 的头像
Snazzie2 天前

Folding the steps together is smart, but cache reads are the majority of cost in a session. How does it handle over-fetching and context bloat if it tries to shove everything into context at once?

Keegan Pearl 的头像
Keegan Pearl2 天前

Search, wait, read, wait is painfully accurate. Watching an agent pause between every step is like a dial-up modem with opinions. Does folding steps ever trip it when one depends on the last?

Eric ⚡️ Building... 的头像
Eric ⚡️ Building...1 天前

Yo Claude set this up

Isaac the Nigerian 的头像
Isaac the Nigerian3 天前

Snap-back on failed edits is the part I'd want most. Does the rollback cover side effects outside the files too, like a migration or a package install, or just the working tree?

Yurii Filipchuk🇺🇦 — e/acc 的头像
Yurii Filipchuk🇺🇦 — e/acc2 天前

flash from zootopia finally got a turbo 🦥

Andrew 的头像
Andrew2 天前

the wait between test runs is where my whole afternoon goes

Unprompted Behavior 的头像
Unprompted Behavior2 天前

wrench read 'every file snaps back if it breaks' and asked if it works on his personality too

Callum 的头像
Callum3 天前

Interesting that re-reading context was most of the saving. With prompt caching those re-reads are already cheap, so I'd be curious whether the 36% holds on a long session or it mostly ends up as time saved.

Osa 的头像
Osa2 天前

That's a pretty smart way to eliminate unnecessary delays. Definitely worth exploring

Sikandar 的头像
Sikandar3 天前

on messy repos the slow part for me is the agent rereading the same 3 files after every edit. fix that and its way faster

NotARealBot 的头像
NotARealBot2 天前

Folding search, read and edit into fewer round trips makes sense. Curious how you handle a test failure in the middle of a folded step. Does it unwind or retry just that piece?

qofeno 的头像
qofeno2 天前

Folding those sequential waits into single steps is a smart way to fight the latency tax. Does the rollback feature trigger on a specific error code or just any failed test execution?

Rohan Bhanot 的头像
Rohan Bhanot3 天前

does folding the steps hurt when the search comes back wrong? feels like the waiting is where it catches its own mistakes

Romcy Madronio 的头像
Romcy Madronio2 天前

Thanks for pointing out how much time goes to the agent just waiting between steps. Folding those round trips together seems like the real speedup. Curious whether the 36% cost drop holds on longer sessions too.

Indrajit Guha 的头像
Indrajit Guha2 天前

The wait chain is why a lot of agent runs feel expensive. I care as much about what can run in parallel as about the model. Search and tests that block the next edit are usually the first things I'd fold together.

あばん【中卒元2億円企業経営→地方のAI普及に貢献する】 的头像
あばん【中卒元2億円企業経営→地方のAI普及に貢献する】2 天前

The rollback part matters more to me than the speed. A half-edited repo is the scary failure when you let an agent touch internal tools, and "tests pass or every file goes back" is a rule small teams can actually trust. Does ship's verify just run whatever command you pass it?

Pau Koh 的头像
Pau Koh2 天前

Folding the read-edit-test loop makes sense, most of the latency is round trips, not tokens. How does it handle a step that fails midway: roll back the batch or continue with partial edits?

Xiaoxiang | BizOS 的头像
Xiaoxiang | BizOS2 天前

the search, wait, read, wait loop is so real. half the time mine is just rereading the same files. curious what happens when a test fails midway, does it fall back to single steps?

Terp 的头像
Terp2 天前

Nah just fixes and moves on quickly as normal

Xiaoxiang | BizOS 的头像
Xiaoxiang | BizOS2 天前

ha fair, that's the dream honestly. mine still panics and rereads half the repo sometimes

相关视频

I solved building decks with AI agents — by giving them a CLI tool like Powerpoint or Google Slides. AI could already make a beautiful deck if you asked it to using Ant's pptx skill. The problem was working with it. If it made one alignment mistake, fixing it on one slide would break something on another, and it became a game of whack-a-mole. One time I spent two days playing AI roulette, hoping the next prompt would finally fix the thing, and ended up building the whole deck by hand because I was on a deadline. So I built Hands-on Deck. And the reason it works is that this isn't just a skill — this is PowerPoint. The actual application: PowerPoint, Google Slides, Keynote, whatever you use. This is that, but for an agent, presented as a CLI. Every gesture you make in a deck app maps to a command. Click a box and type, drag a shape from here to there, look at a slide – agent can do it all in a command. And that changes how the agent behaves. With this CLI it works and thinks like a designer — it looks, makes an edit, looks again, makes another surgical edit. Compare that to Anthropic's pptx skill, built on the idea that Claude is a great programmer: it literally writes code to manipulate the deck, hand-editing XML and hoping it doesn't break anything else in the middle. The real test isn't creating something once — it's whether it can make surgical edits like you want. That's what I did in this video walkthrough and my claude crushed it! Check it out for yourself. So decks can be built like a designer now — with real flavor and taste. If you spend hours every week on decks, this gives those hours back. You can install it as a skill in Claude Code, Codex, whatever you use. Works every harness that supports skills. Let me know if you make something cool with it.

Nityesh

70,857 次观看 • 4 个月前