I'd been running into a lot of disk pressure on my system from agents running. I discovered two main causes:
- my approach to using git worktrees for all work. Not entirely surprising, the duplicated build artifacts in each worktree being the main culprit here. There were some limited gains to be had by just being more efficient with how builds were managed but ultimately I just had to limit the number of concurrent features I was working on in a couple of particularly large repos.
- more insidious was the fact I was actually hitting memory limits more than I realised and causing my mac to have to swap heavily. The system was surprisingly still quite performant while this was happening... until it wasn't. Unfortunately as best I can tell the only "fix" when this happens is to restart the machine. There's no way to have the system reclaim any of the swap space once it's written to disk. Which if you have it thrash a couple of times a day (especially easy if you leave something running overnight) adds up very quickly.
tho, I also met this problem and found a solution for me - develop all in one worktree on a single branch, then cherry-pick to the second worktree to the needed branch.
So, only first worktree is built:
- less space needed.
- less time spent on build, which is usually caused by switching branches.
- you are not required to merge/rebase of master branch to solve conflicts, which also cause long time rebuilds.
It's so goddamn bad, I wish it was easier to turn all of this garbage off. I used to spend time defining parallel work lanes to issue to subagents as frontloaded development work with set validation checkpoints and it worked great. Now attempting to do this with 8, 16 subagents will balloon out into hundreds of gigs of fucking mess
Copy on write worktrees do absolutely nothing to solve this, either. The build artifacts alone will do it
Agreed. I thought I was being smug by telling my agents to use btrfs for specific workloads. Sure, it took me longer to run out of space, but I of course ended up running out of space
My work provides me a Mac laptop plus remote Windows VMs and I've completely stopped using my Mac for agent builds/tests - I give them ssh access to WSL instances and they can go hog wild on the VMs.
It's all a bit adhoc - agents still need to pull (or I need to keep the main enlistment up to date on these machines) - I'm thinking of using containers or something else like github actions keep things cleaner, but right now it's "good enough".
Regardless I'm pretty happy with how it's working so far. Previously I'd be letting agents do work in the background and they'd kick off builds in the middle of a meeting or something and I'd start dropping video streams until I killed everything.
switching to copy on write worktrees is the biggest game changer it should become common knowledge pretty soon and it's already available on most operating systems
You also need a deduplication, because with simple copy on write you will have a copy for writing the same - change a byte in a file, get a copy, then change it back, get a duplicate.
Deduplication in ZFS needs 5G of RAM for 1TB of logical data - if you expect your agents to produce 2TB of data, you should provide 10G of RAM for deduplication.
thanks to the parent you're replying to, I discovered there's numerous projects that try to retrofit CoW into worktrees + also fill in the gaps for mostly-shared data like build artifacts! I've decided to take lane (https://github.com/lukeed/lane) for a spin as I type to see how it goes.
I ran into similar issues. Sometimes though there are good language specific solutions. For Rust based worktrees you should definitely check out https://mr-boxington.jdx.dev/ .
Regarding swapping, are you sure it's not just from ram backed tmpfs partition? For me that was it and most of the swap was from /tmp. So I created an agressive job that runs frequently to delete all tmp files not used in the last ~2 hours without any command holding handles to it.
The other thing I found was that there were a bunch of leftover processes from prior coding agents forking commands and just leaving them. So I had to build a system into my agent spawner to automatically launch each agent session in a cgroup whose processes all get killed on task completion.
On Linux I’ve taken to disabling swap before doing memory heavy operations. Claude will carelessly run benchmarks and tasks which gobble up all available memory. When that happens, I’m not sure why but my system becomes unresponsive and I can’t kill the process. If I disable swap, the oom killer takes care of whatever Claude was doing and Claude figures it out after the fact.
You can use systemd-oomd to force OOM killing without having to completely disable swap. It's quite handy for this especially since you can tell it to aim its OOM heavy artillery at only certain tasks (i.e. only processes that are descendants of claude-cli.)
For the first part you might be interested in Anthropic recently creating their own version control system that's server based like SVN, and it seems to ease their memory pressure too. It's called `scion`.
The amount of cope of people dealing with this is sobering. One of the biggest companies out there currently has neglected to fix a seemingly trivial bug in their core offering for 4 months and everyone is just like 'well, you can delete the files'.. I'm not hopeful for the future. Software already was a terrible buggy mess, now it's just all rng i guess.
this looks like a bug, but even normal happy-path codex creates many GB of logs, openai apparently considers this fine bc it's been going on for months
two things help:
1. fs-level compression, I use ZFS bc Btrfs has massive contention problems with concurrent Rust builds
Yeah mole is brilliant, I just discovered it days after paying for Daisydisk and haven't used Daisydisk since. Mole freed up 80GB on my laptop the first time I ran it.
I'm working on a side project that requires a lot of fine tuning of 3D worlds. Codex ended up creating thousands of artifacts, basically loads of screenshots, in its worktrees, and doesn't clean up after itself.
On a work project, we're moving financial processes from Excel to code, and Codex ended up creating and storing thousands of validation tests unnecessarily. That ended up eating 90GB after we let it loose on a Windows VM that we connected with a custom agent between codex and the VM.
This is definitely a Codex thing and I've never seen this happen in Claude Code or Antigravity.
It's silly I have to tell it to make part of the build process to clean old artifacts or it'll just default to keeping endless folders of test builds and old builds without cleaning them up. The same seems to usually be true with documentation and reference material.
There was a REALLY annoying period where I'd give it full control of a VM and a dedicated project folder and it'd continually refuse to delete anything I asked it to or just silently archive it, even though everything was already in a git repo. It seems to have gotten somewhat better about that now, but I still include "create cleanup scripts to delete unnecessary project content" in the default steps before doing the final merge on a change.
I would argue that the better you can “close the loop” the better. One of my favourite agent activities is giving it a firmware bug and JTAG access + a serial port. That setup means that flashing, debugging (inspecting variables and whatnot, setting breakpoints, etc), and logging can all happen completely autonomously.
Works especially well for transient/hard to repro bugs. It’ll either set breakpoints or write a monitoring script, go to sleep, get woken up by the harness when the bug trips, inspect state, put together a new build based on a hypothesis built from the debugging, flash it, and loop. I’ve had sessions run completely autonomously for over a day and come with a beautiful one-line solution to a gnarly race condition.
In that regard it’s like a human: you can read code all day if you want and you might get somewhere, but being able to inspect what happened, build a new theory, test it, and iterate is way more powerful 99% of the time.
I was trying to upload HDR photos to Instagram and couldn't figure out how to make it preserve the HDR-ness of my photos, and was going back and forth with an agent - step by step we'd slowly iterate on a photo, I'd upload, then I'd download the jpg and give it to the agent and it'd think, change something in the format, and we'd try again.
I got impatient and just said fuck it and gave it chrome-devtools and a blank instagram account and a test photo, and it figured the whole thing out in about a half hour. I built a little browser-based tool from what it discovered, and I use it to run my HDR photos through to format them correctly for instagram and it works a treat.
> I built a little browser-based tool from what it discovered, and I use it to run my HDR photos through to format them correctly for instagram and it works a treat.
I was kinda expecting: "And when my agent built a perfect clone of Instagram and social-engineered all my contact list to join"
Rust is notorious for this, one thing some people do is force Rust to have a salvage yard for builds so all those dependencies are shared by everything regardless of what you are building or where, but holy moly, can someone make Rust output smaller files? Hard drives are always having some reason to never go down to sane prices.
When I was younger I thought we would be having exabyte hard drives by now.
Before Windows XP, desktop Windows users were accustomed to frequent system crashes. Viruses in the 1990s would take out offices around the world. Nothing collapsed, and nobody switched to Linux. Everyone was just used to it. I think most people will accept the tradeoffs, because it used to be much worse than it is now.
If OpenAI's super-intelligence can solve Navier-Stokes why can't they fix this stupid bug that ate up 330GB of my Mac's disk space? come one please triage your github issues?
>
If OpenAI's super-intelligence can solve Navier-Stokes why can't they fix this stupid bug that ate up 330GB of my Mac's disk space?
Because the former can be marketed very prestigiously, the latter hardly can.
According to https://www.briefia.fr/en/article/openai-revendique-la-resol... it is estimated that solving Navier-Stokes did cost OpenAI 40 million USD. I would bet that if OpenAI spent the same amount of money using its super-intelligence to fix this bug, this would work.
- my approach to using git worktrees for all work. Not entirely surprising, the duplicated build artifacts in each worktree being the main culprit here. There were some limited gains to be had by just being more efficient with how builds were managed but ultimately I just had to limit the number of concurrent features I was working on in a couple of particularly large repos.
- more insidious was the fact I was actually hitting memory limits more than I realised and causing my mac to have to swap heavily. The system was surprisingly still quite performant while this was happening... until it wasn't. Unfortunately as best I can tell the only "fix" when this happens is to restart the machine. There's no way to have the system reclaim any of the swap space once it's written to disk. Which if you have it thrash a couple of times a day (especially easy if you leave something running overnight) adds up very quickly.
So, only first worktree is built: - less space needed. - less time spent on build, which is usually caused by switching branches. - you are not required to merge/rebase of master branch to solve conflicts, which also cause long time rebuilds.
Copy on write worktrees do absolutely nothing to solve this, either. The build artifacts alone will do it
It's all a bit adhoc - agents still need to pull (or I need to keep the main enlistment up to date on these machines) - I'm thinking of using containers or something else like github actions keep things cleaner, but right now it's "good enough".
Regardless I'm pretty happy with how it's working so far. Previously I'd be letting agents do work in the background and they'd kick off builds in the middle of a meeting or something and I'd start dropping video streams until I killed everything.
Deduplication in ZFS needs 5G of RAM for 1TB of logical data - if you expect your agents to produce 2TB of data, you should provide 10G of RAM for deduplication.
Regarding swapping, are you sure it's not just from ram backed tmpfs partition? For me that was it and most of the swap was from /tmp. So I created an agressive job that runs frequently to delete all tmp files not used in the last ~2 hours without any command holding handles to it.
The other thing I found was that there were a bunch of leftover processes from prior coding agents forking commands and just leaving them. So I had to build a system into my agent spawner to automatically launch each agent session in a cgroup whose processes all get killed on task completion.
two things help:
1. fs-level compression, I use ZFS bc Btrfs has massive contention problems with concurrent Rust builds
2. Archive your old sessions using something like sessionwiki (https://github.com/youdie006/sessionwiki), I've integrated this into my control plane (https://github.com/BrokkAi/mjolnir/)
I haven't read about btrfs having issues with concurrent workloads like multiple concurrent rust builds. Are there notes on this somewhere I can read?
TUI version is free, though there is a paid app if that’s your thing
https://github.com/tw93/Mole
They still update it, it works perfectly, no version 2.0 forced obsolescence. Just a nice, reliable tool.
https://github.com/jesusha123/qdirstat-macos
https://github.com/tw93/Mole
And yes, codex is eating my disk, too. Ironically my solution is mostly codex - I have a task which goes around stomping on detritus.
On a work project, we're moving financial processes from Excel to code, and Codex ended up creating and storing thousands of validation tests unnecessarily. That ended up eating 90GB after we let it loose on a Windows VM that we connected with a custom agent between codex and the VM.
This is definitely a Codex thing and I've never seen this happen in Claude Code or Antigravity.
There was a REALLY annoying period where I'd give it full control of a VM and a dedicated project folder and it'd continually refuse to delete anything I asked it to or just silently archive it, even though everything was already in a git repo. It seems to have gotten somewhat better about that now, but I still include "create cleanup scripts to delete unnecessary project content" in the default steps before doing the final merge on a change.
Had to tell it to update it's agents.md to not do that...
Works especially well for transient/hard to repro bugs. It’ll either set breakpoints or write a monitoring script, go to sleep, get woken up by the harness when the bug trips, inspect state, put together a new build based on a hypothesis built from the debugging, flash it, and loop. I’ve had sessions run completely autonomously for over a day and come with a beautiful one-line solution to a gnarly race condition.
In that regard it’s like a human: you can read code all day if you want and you might get somewhere, but being able to inspect what happened, build a new theory, test it, and iterate is way more powerful 99% of the time.
I was trying to upload HDR photos to Instagram and couldn't figure out how to make it preserve the HDR-ness of my photos, and was going back and forth with an agent - step by step we'd slowly iterate on a photo, I'd upload, then I'd download the jpg and give it to the agent and it'd think, change something in the format, and we'd try again.
I got impatient and just said fuck it and gave it chrome-devtools and a blank instagram account and a test photo, and it figured the whole thing out in about a half hour. I built a little browser-based tool from what it discovered, and I use it to run my HDR photos through to format them correctly for instagram and it works a treat.
I was kinda expecting: "And when my agent built a perfect clone of Instagram and social-engineered all my contact list to join"
When I was younger I thought we would be having exabyte hard drives by now.
What people can get away with on the upper crust with end-user apps would spell disaster for the inner parts.
Home users maybe. I've started with Windows NT 4.0 and don't remember it being crash-prone.
for some reason I had 11 separate codex and codex-code-mode-host binaries in mine. each weighing about ~300 MB.
Apparently OpenAI is unable to fix this properly, even using its state of the art AI.
See this: https://github.com/openai/codex/issues/25921#issuecomment-62...
Because the former can be marketed very prestigiously, the latter hardly can.
According to https://www.briefia.fr/en/article/openai-revendique-la-resol... it is estimated that solving Navier-Stokes did cost OpenAI 40 million USD. I would bet that if OpenAI spent the same amount of money using its super-intelligence to fix this bug, this would work.
But seems not to be very economical. :-)
You don't have to give any more legitimacy to Trump's inane ramblings than they already get.