There aren't any documented settings to disable this, but there is a feature flag it currently sits behind. You can flip it in your ~/.codex/config.toml[1]:
[features]
daemon_auto_start = false
There are also some files that it writes to in ~/.codex at startup: pid files, lock files, things like that. You can delete them and replace them with symlinks to /dev/null, then mark those unwritable with `chattr +i` or similar. Presumably then the background daemon will crash on startup, if the feature flag ever goes away.
What makes ACP deficient for headless uses of LLM agents? Maybe fixing that or standardizing on another protocol designed for pluggable LLM agents (I guess ACP is more oriented towards pair programming in an IDE, at least originally) is a better idea imo.
For whatever reason, the daemon has been refusing to start and Codex has been grumbling to me. I was wondering where the daemon had come from, and thought maybe I missed something previously and that it had just broken.
Yep, the real issue is that it was something they didn’t have before, and then with Codex 0.157.0 enabled, they installed this persistent nonsense [1] that survives after the last instance of Codex has exited on your machine, which is just gross.
It also fills up my HD with old versions.
I investigated and it took up something like 12 gigabytes in total on my system; which seems insane to me, it’s just a harness?
People moan about how badly made Claude code is but Codex seems much worse on every metric
All of the harnesses I've used are terrible at cleaning up after themselves. They leave all kind of clutter behind, to the point where on my dev VM, I now have an LLM run once a day to just look through the disk and delete everything that looks like it's leftover garbage.
Installed "clor.com" and it got into a loop of downloading its docker image from ghcr them throwing it away immediately because it was supposedly out of date (iirc). 65mbps continuous transfers for a week until I looked at my network traffic and saw 500gb of rx per day.
Can't wait until we get really well designed harnesses written in a compiled language that doesn't bloat my computer. Will probably have to be written by a company other than OpenAI because codex is like the exact opposite of everything I just listed
it's minimal and somewhat opinionated, not integrated well with memory subsystems (although that's kind of the point)
my only real criticism is that last I played with it it deals with slow provider responses very poorly , so lots of locally hosted llms require a lot of tuning to get it to handle streaming/non-streaming/timers well.
way less minimal (but still pretty minimal) I prefer Pi; but different markets I think.
hax fit my openbsd laptop well, espescially since Pi and related derivatives are kind of un-spokenly incompatible with the BSDs.
(author passing by) Could you please report the issues you've experienced on the repo? I did quite a bit of testing myself on local models (Qwen, Ornith, Gemma, Muse Glimmer) with llama.cpp, and they all worked fine. Details would be helpful.
Also one brief note for Codex users specifically. hax can work out-of-the-box as codex-cli drop-in replacement (since it can use the already authenticated codex token), so it's super easy to try it - basically it will be a zero-config experience.
It's because they statically linked it, and probably rely on a trillion pointless crates (as is standard in Rust development). A good statically linked harness can easily be ~5MB MAX
I suspect a lot of people wouldn't use AI if they had to pay per token at current rates. Yes, it's technically a different discussion, but "just write your own harness" isn't really an answer for the vast majority of us riding on the investor-subsidized subscription plans.
I did, and the weird things you have to do to make work at least decent. I’ve been doing evals and I got some things better than opencode but not as good as pi. Though my goal is to optimize for a browser and browser based llm so it can create stuff in a web container or at least generate ui by requesting json schema (a2ui style).
> People moan about how badly made Claude code is but Codex seems much worse on every metric
Also, it feels that chatgpt.com and aistudio.google.com are the buggiest websites I have seen for long time. Why it is always something related to LLMs?
for some reason I had 11 separate codex and codex-code-mode-host binaries in my ~/.bun/install/cache/@openai . each weighing about ~300 MB. might be worth a look to anyone running low on disk space.
Yes, my DEV VM tended to fill up with VSCode backups as well. I finally expanded the disk so it's less of an issue today, but for a while ENOSPC was resolved by cleaning up ~/.vscode-server/cli/servers.
The daemon has many bugs. For example, when auto-updating, it would randomly ruin the permissions and block the goal as .git became read-only. Generally, it is extremely annoying how many bugs Codex has. I basically can't trust it to execute on any goal since it would often get blocked every few hours due to random bugs.
My complaint applies specifically to the CLI. I have never used the gui app. Here are the various goal-breaking bugs I've run into:
- Goals get blocked due to .git and other stuff becoming randomly read-only when daemon auto-updates in the background (the linked issue in my comment above)
- Goals get blocked due to transient "Selected model is at capacity"
- Goals get blocked due to bwrap randomly failing with `bwrap: loopback: Failed RTM_NEWADDR: Operation not permitted`. I traced this one to `apparmor` on Ubuntu 24.04 interfering with bwrap and found a local workaround.
I’ve been using GPT models via hax[1], a bare bones harness with minimal tools on my system and have the models rely on bash for the most part the last few months. It’s been a breath of fresh air amidst all these overengineered harnesses.
If I need special actions, I just make them CLI tools on my Mac and tell the AI model that such a tool exists(via a single line in its system prompt) and that’s it. It’s been very nice.
You seem to care, so I wanna point out a semantic nit; "such a tool exists" could be interpreted as a harness provided tool, rather than a command line tool. (Edit, on a re-read its pretty obvious you mean cli tool, but keeping this statment here as im sure it'll open someone's eyes to the difference)
I've gotta give hax a go, as a pi user who came on-board due to the no frills direction pi takes. I say takes rather than took, because regardless of the (heated) discussion surrounding mcp and code mode support, I cannot fight the numbers- augmenting the harness provided tools is a notable boost to execution round trip performance.
I may need to take a second look at Pi again if it has any niceties like that notable boost in round trip performance compared to hax.
I mostly stuck around with Hax because of a few design choices. The sessions are merely plain text files indexed by a dirpath slug. This means I can just start a hax session in that directory and expect sessions history to be available in any model.. (I occasionally dabble in GLM for security checks, Gemini for writing stuff, Deepseek for tightly scoped tests etc).
Off-topic: I think the front-end quality of ChatGPT/Codex Desktop and Android app is simply dreadful—hard to imagine how such a large number of frequent failures made it into production. Do they not use human testing at all?
probably not much, they're all using workflows where the agents pull issues and automatically merge code when the tests pass. so lots of stuff slips through and the overall experience is kinda shite.
Ordinary package managers are totally capable of updating thousands of programs without manual effort, including automatically and in the background if that's a thing you want. But they do it without violating users' sovereignty over their own computers or turning every app into a special snowflake that gets to install software (namely itself) in its own bespoke way.
Firefox does not only not prompt me for updates (it only updates if I manually run an update), even if I do update with Firefox running, it does not stop me from continuing to use Firefox without restarting it. All I did was install Firefox from flathub (packaged by Mozilla (a default configuration provided by Mozilla)).
I am running Firefox Developer Edition on macOS, where it is the default option. Have to set "Check for updates, but choose when to install" to get the similar behavior.
I've never seen that behaviour, it downloads updates in the background and then gives me a little notification that I should restart it to update, but it's dismiss-able.
Being able to turn auto-updates off is okay, but that being the default would certainly not be okay in Codex, giving the reach, what it can do and how immature it is.
Yeah this seems like a decision yielded by a model. One of countless examples of the model’s outputs careening into individually defensible but holistically nonsensical results
A human seems less likely to embark upon a path that most humans know leads to something a human is unlikely to want
LLMs produce amazing things, but only one token at a time. Anything beyond that is retrospectively emergent at best
Serious tools shouldn't do this. Sure, you can offer an opt-in rapid release beta update option, but if you can't go more than a month without forcing me into a "stable" update that changes the UI and behavior I'm getting used to, then I'm not going to consider your software stable and reliable enough to be worth developing expertise in using.
Software Engineering clearly is solved, that's why Google has a tiny animated banner on the AI Studio sidebar (advertising subscription upgrade), which shows a simple animated border around it and utilizing exactly 100% of the GPU to render just that banner. And there is a browser extension removing that styling element (without banner GPU load is obviously under 0.1% as it should be) released way back in May. So the almighty AI didn't manage to fix a major UI blocker for a flagship AI product in half a year now. :)
When I had a Claude account (cancelled ~May time) I vividly remember the web chat having broken editing for backtick-quoted monospace text. If you typed <space>`foobar`<space> the last space would move the cursor backward into the closing quote, and the space would be emitted inside the monospaced region. It was like this for several months (on Firefox). I always found this baffling for a product marketed to programmers.
Products moving fast doesn't mean that they're moving in the direction I want, or that I want them to move underneath me while I'm in the middle of a project.
Auto-update systems have value but are not a suitable default for professional tools.
This is the same operational model as Claude Code.
ls ~/.local/share/claude/versions/
ps auxww | grep "claude bg"
I see a few prior versions of the Claude Code binary and several running background tasks, without any copy of Claude Code in the foreground. I regularly get a "Update available" message in my Claude Code UI too.
If running systemd, wrap that crap into a cgroup and force kill the whole process tree on exit. I've been doing it for years with badly behaving applications. On phone, from memory, may not be exactly what is needed:
This kind of stuff often breaks for applications packaged with Nix, so it's often removed from Nix packages for compatibility reasons. Both behaviors you mention are helpfully patched out of the Claude Code package I use: https://github.com/numtide/llm-agents.nix/blob/dc285bfd94ec7...
As a result, I never knew that Claude Code does a nasty thing like this, either. :-\
Fwiw, packages in that repo get version bumped each day, so if that's a fast enough pace for you it may be a good alternative.
It's right about now that having a operating system distro model is great - you get folks who will package the software & test it out, make generally sensible defaults, and generally keep shenanigans like this from going on
Well, OpenAI at least is fine with you using whatever harness you like. Having a crappy first-party one is bad, but an escape hatch puts a floor on it.
Codex is very badly structured. I was an early user with codex. But it mixes state and configuration data; which is why I can't really declaratively manage it in my nix-darwin environment; so abandoned it a while ago. This issue being raised, and the way that the codex development responds back to it makes me not trust their harness ever again. I only use codex in opencode; which seems to be doing OK. At least openai doesn't gate their product to their harness like anthropic.
Not trying to sing praises; i think opencode is doing a bunch of weird stuff too but it's not codex level.
You can just copy the executable then run it with --no-daemon, or edit the config equivalently. No autoupdates and no background process.
The recent terminal UI changes were more fastidious, but even those are easily disabled with CODEX_TUI_DISABLE_KEYBOARD_ENHANCEMENT=1 and the /tui command.
I think the quality of Codex CLI (and Claude Code) is an honest reflection of the status quo of LLM-driven development. The future doesn’t look really promising.
To be fair we've gone from barely capable at writing code to what we have today in barely 3 years. I wouldn't put money on the counter bet that they won't get beyond this hump.
I wouldn’t assume that progress is linear. Maybe I’m wrong and it will be! But I’m reminded of early developments with self driving cars, it felt like manual driving would be a thing of the past within a decade. But progress has been much slower than that.
> But I’m reminded of early developments with self driving cars, it felt like manual driving would be a thing of the past within a decade. But progress has been much slower than that.
I think that's exactly the standpoint of Yann LeCun.
> The progress using narrowly intelligent algorithms to drive cars in a generalizable way was not as fast as the early gains in places the narrow algorithms could solve, thus the progress in using generally intelligent algorithms for any other field must also slow down.
Man, is this how 99% of the population works? They have a heuristic for anything that sounds computer sciencey and apply it for anything and everything that sounds similar without bothering to check if their heuristic actually applies?
You committed your own fallacy of applying a sloppy heuristic to understanding the person you were responding to without bothering to check if it applies. Spectacular. The comical levels of arrogance that infect over-hyped tech marketing boondoggles (one area among many in which GenAI is clearly similar to self-driving cars) is one of the reasons nobody trusts the hype anymore.
The breathless astroturf articles, the brain-dead talking head conversations, the endless flood of obvious misinformation about industry uptake, the nauseating LinkedIn puff-pieces about PREPARING YOUR ENTERPRISE OR ELSE... it's a hostile information economy at best, and its being driven by the same people who brought you bullshit like self-driving, crypto and the Metaverse... for the exact same reasons (to get rich).
There's no epistemic credibility left in the tech industry in general. Maybe the Dario's or whoevers of the world believed what they were talking about at one point, but I don't believe a single word I read from Anthropic or OpenAI for the same reason I don't believe anything I read from Russian state media. I know it's being produced for reasons that have nothing to do with the truth.
As someone who interacts with the technology on a daily basis at a non-trivial level... I'm not seeing anything that would cure my skepticism about GenAI. The information economy surrounding it is broken. The fundamentals don't line up. The hype about "nearby capabilities" is identical to the hype of generations after generations of hype in the AI space (dating at least back to Turing). It smells like a bad bet.
It is hands down the most user hostile UX I put up with. The fact that I put up with it speaks volumes about the value proposition that the underlying models bring to the table.
I’m in the process of dumping codex for more open harnesses that don’t add problematic new features every week. Because Anthropic insists that I have to use their harness with the subscription, eventually I’m either going to wrap `claude -p` to get away from the crappy TUI or else I’ll just unsubscribe and use other models. Their bad UX has alienated me as a paying customer.
Try using the "rewind" feature to go back to a previous message.
Went too far? Oops, it's not accessible from the Claude Code UI anymore! You can get the agent to dredge up the info from the session logs, but you'd better hope they put the pieces together correctly.
Also, you can't rewind to a steering message. Want to make sure you can rewind to it? Wait for the agent to settle, or stop them, enter your message, and hope you don't throw them off their current train of thought.
To see a proper rewind UI, see how Pi does it: Every message is available, forked conversation history is preserved.
It has its own issues, but is miles ahead of Claude Code in this regard.
I'm not who you're asking but for me I am right now going through the insanity of providing plugin/mcp/tool access across environments. This is in codex which as I mentioned is intentionally trying to provide mass-market ergonomics. But with that comes incomprehensible configuration that's also a moving target. chat/work/codex/remote/cloud are all different things. And each client exposes different abilities and hides others. And looks like only the web-ui is the most up to date because other targets need releases. As a developer I understand why it's in its current state, but that in no way excuses it as acceptable ux.
Like of course codex is the most powerful and "just" run `npx blah blah` but no, I'm not running randomly piped code on my personal computer and even if I didn't care about privacy it means I can only run this locally with my laptop on. "leave your laptop on and access from your phone" is not a good idea. It only adds to the surface area of confusion. Ux no good.
this 1000x. a lot of details in context compaction, editing, reverting ux, etc are all really centered on keeping the load and cost of subscription users as small as possible. its no always that nor is it necessarily correct, but its a huge driver of how those subscriptions and first party apps are designed
I know this is an unpopular opinion, but it's actually good.
- pasting lots of content became [35 lines] and they added a tiny notification to show it as-is if you paste again
- interactive questions layout
- paste images
- the recap
- auto-updating session titles
And so many more. I tried codex and it was lacking in many UX aspects. Remember we got from a text-paragraph-only terminal to this at breakneck speed.
agreed. the thing i reach for most is its ability to coordinate my agent graph, e.g. seeing what sub-agents are doing, messaging agents, adding new capacity. claude code really is a delight to use here
Claude Code is better than Codex UX. But as per the sibling response it's more a developer tool made by developers for developers. It does have a lot of niceties. Codex/openai on the other hand is more clearly trying to reach mass-market and it's a UX shit show because of it. Consumer UX is its own thing. So the original point is that raw ability of the tool doesn't magically solve ergonomics.
Won't disagree with you because I haven't used codex/OAI a lot - but Claude's non CLI UX is great too.
Eg - I asked for a recipe and it says leave x for y time - there's actually deeplinks that link to your clock timer with the time preset! And to change quantities, there's input boxes in the response that update all quantities accordingly.
The current approach of the big companies is to keep increasing the number of parameters (there was a report that recent Claude models are on the 4T parameter scale). Even recent SOTA open weight models are in the 1T-2T range.
This is partly due to the multiplicative scaling factor -- the number of tokens; the number of transformer layers; the number of features; etc.
I suspect that for models to have a step upgrade either/both of 1) a new/improved architecture model; and 2) better training of smaller parameter models are needed. I say the latter as I think that with the larger weights there is a lot of unused/redundant parameters.
I dont think the quality has gotten that much better. Take claude 4.5, the probability of getting usable outputs was less likely than with opus 5.5, and the usuable outputs weren't great.
I think newer models are more likely to produce the usuable output, almost always, but its still not great. The frequency in which you get usable results leads one to believe that quality is improving.. No usability is improving but the result is still meh.
Also put an llm in an environment without an lsp or compiler to check against and let me know how capable you think they are on their own. The "loop" is doing a lot of heavy lifting here.
> The capital costs have also ballooned to the point where the rate has to level out
That's a pretty bold use of "has to". None of the financing behind AI seems sane, yet it continues to defy rational expectations. Reminds me of the comment "the market can remain irrational longer than you can stay solvent". It's not the exact meaning but has the same feeling. The circular financing going on does not help anything either
in the last 2 weeks or so:
- scrolling stopped working
- subagents no longer work (main agent doesn't get notified when they finish)
- subagents don't work (a subagent can't make its own agent even if instructed to do so)
I believe it's all related. They keep adding new, badly thought out features because it's much easier than improving existing features built on low quality code.
The latter requires modifying existing, badly architectured and tightly coupled code that will almost definitely break something somewhere else if touched, so what do you do? Just keep adding new code on top and hope for the best.
This is not a new phenomenon by any means, but AI amplifies it by 100x.
It has pretty rapidly reached the point where I do not think any acceptable LLM-harness is widely available. Every single one I've tried has broken in bizarre ways on even minor updates.
I used codex the first time yesterday and quickly decided I need to sandbox it and came up with a pretty easy and straightforward solution (I think).
I use nixos and use a little wrapper that wraps Codex CLI in Docker. Run codex-box /path/to/project and it starts or reuses a container with only that project bind-mounted, keeping login and history in a separate volume. It uses tmux for detaching and reattaching
Very much this... The utility of Claude has grown exponentially with the amount of work I've put into memory/context management, tooling, skills development and CLAUDE.md (and friends). I freely admit I may just be fortunate, but I've yet to hit a bug in a new release that rose above the level of minor annoyance.
The thing that really validated this point was when I recently tried Codex again. Out of the box it was OK, but not great. I swung over the memory apparatus and had it import directives from my existing Claude install and it was almost immediately quite useful in its own right and, especially, as a counterpoint to Claude when dealing with some fairly complex operational issues.
Pi went searching across my entire home directory unprompted... Idk to do that without some kind of sandboxing builtin is rough and chews through tokens on things I don't really want polluting my context.
A coworker of mine has a pretty neat script that uses bubblewrap to wrap any application. I'm using that one too for pi and it's pretty great since you don't have to trust the harness and can quite easily customize the sandbox to your needs, e.g. disallow reading to certain directories etc.
I've used the CLIs for Codex, Gemini and Claude and they all give off the feeling of being on an very rickety, very fast runaway train.
You install an update. Oops, they added a sandbox that doesn't work inside a docker container, it's stopped working.
You install an update. Oops, you can't log in inside a docker container any more, it's stopped working.
You install an update. Oops, they started catching the copy+paste shortcut and replaced it with one that doesn't work.
You revert a broken update. Oops, the update changed the format of your login details, the old version can't read them, gotta log in again.
You install an update. Oops, can't scroll with the mouse wheel any more.
Not showstopping bugs (if you're used to persuading broken oauth logins to work) but it certainly makes updating an "ugh, what have they broken this time?" experience.
You got further than I did. I checked them out and went back to manual prompting. Just don't have a need for codegen that much, and when I do I'd prefer to be in control.
It does some neat things that, ultimately, are super obnoxious. Background animations and "sparkle" effects that make my terminal indicate activity at all times. Just subtly broken behavior.
I think the (lack of) quality of these tools shows that software engineering is maturing as an engineering discipline.
We were overbuilding trivial parts of the software like the Romans overbuilt small bridges.
Modern engineering is about making the thing that barely works. If you disagree, learn a little bit about bit narrow the tolerances of modern cars and bridges are.
> learn a little bit about bit narrow the tolerances of modern cars and bridges are
Buildings typically use a safety factor of 1.5 to 2.0, and cars typically use a safety factor of 2.0 to 3.0. These things are being built twice as strong as they need to be. I wouldn't call that narrow.
And with software, we're at a point where it can't even properly do what it needs to. If we were overbuilding it, damn software wouldn't crash so much.
I think all engineering has safety factors of 2 or more. Maybe with the exception of aerospace (mostly space) where materials might be at the limits of what's possible right now so the safety factors are much reduced.
> I think all engineering has safety factors of 2 or more.
Not even close.
For example, a BMW X5 weighs about 2200 Kgs without passengers, and it's GVWR is 2800 Kgs.
If you have five people and heavy bags, you can comfortably exceed the GVWR. You can absolutely break control arms and suspension components in that scenario if you go over a speed bump or hit a pothole (which are real things that exist in the real world, like people doing something unintended and crashing software).
If this needed a 2x safety factor, there would have been no way to build this car.
Even the toughest cars like land cruisers are limited in this way.
I think there's a difference between something that's engineered and precise vs something that's rickety. Just because you built something that might fall apart at any moment doesn't mean it's been engineered with narrow tolerances.
A 100 percent of modern bridges are rickety. There has been dozens of high profile incidents because of this.
> Just because you built something that might fall apart at any moment doesn't mean it's been engineered with narrow tolerances.
Here the software they are building falls apart in ..say.. two days of use, but can be updated in 5 seconds at close to zero marginal cost. The software is useful enough that people keep updating and using it. This iteration gives them a competitive advantage over the safer choice of rigorously testing every new thing for several weeks.
I don't disagree but it's clearly a problem that already existed for decades in software and is just being 10x'ed. how many different frontend frameworks or AWS regurgitated and enshittified cloud services did we have to endure over the years.
There is actually even worse code quality written by humans, is that small sample also a reflection of human-driven development? And therefore, the future doesn't look promising that way either?
The difference is the source of the sample. This is a flagship product by a frontier lab. It should be reflective of the best that is possible. It should be biased in the most favorable possible way. It was designed by people without the normal limitations and quotas of outsiders in building it, as well as in learning experiences in driving it. This is not a random cherry-pick of bad code.
The restraint, rigour and responsibility you showed prior to LLMs relates directly to how much of that you apply now that they’re here.
Given how rare it was previously, and how little it was cultivated by shortsighted management, I’m not surprised with all the slop we see today.
The truth is there’s nothing stopping someone from using these tools to remove dead code, simplify project structure, remove dependencies, etc. but I don’t see it happening very often. I have done it a bunch myself though so it’s absolutely possible, it’s all a matter of investment and prioritization.
Seems like its just a bunch of young engineers making bad architecture decisions. Not sure I would blame the culture of vibe coding everything. It might just be bad taste.
i'm personally quite perplexed how sloppy the harnesses are from the leading AI labs, you'd think these are their primary products and would have some human attention devoted to them
it's embarrassing when every claude code release has obvious bugs in it like telling you to upgrade when you're on the latest version or codex opening visible powershell windows for every command that pop up on your screen
like 5 minutes of human testing would catch this obvious shit but obviously they don't bother because vibes are the future and stuff? idk
Solar power has become a real transformative technology and instead of trying to recreate a better world we tariffed them and the execs of this country spent all our coins on power hungry data-centers that write mid code.
It's not necessarily hostility. We considered different updating approaches with our legal IDE where security, stability, etc. are paramount, and having a buggy application out in the wild is user-hostile. Sure, in a dream world we'd say "don't ship until it's perfect", but that's just not realistic.
We considered the daemon thing, but ended up with a hard-update check on every boot. It's thus a user choice: if you don't like the automatic updating, you can just not run the application anymore.
And even with all that, it's hard to stay on the right side of AV algorithms, etc.
> It's thus a user choice: if you don't like the automatic updating, you can just not run the application anymore.
That's not a choice. It s just what most apps do these days, but it's the first time I've seen anyone stating this so directly, or trying to frame it as a choice.
(And behaviorally, this makes your product just a SaaS website with extra steps.)
I guess I mean it's a choice in the sense that it's a transparent part of the product that one chooses how and when to use (or not). That's opposed to something running in the background on its own schedule, with its own agenda whether the user is running the application or not.
[Edit: yes, that is kind of the idea. Native version has better features because it can access them.]
My 2-cents: never install those things into your host. A locked down Docker container is the bare minimum. Pin your versions and never update them without checking the changelog.
Cool thing is that I can just ask the llm to modify the harness to what I need.
I discovered this daemon thing about two weeks ago and had codex modify it's own custom launcher that I have to bypass the daemon thing. If as a recreational code farmer can do it then most "pros" should be able to handle this and other inconveniences. Cc has been pushing that no scrolling feature that codex recently introduced for a half a year now but again I had them put a stop to that. Software has never been this easy to configure.
I discovered this while benchmarking the tool I'm developing. I realized Codex was surprisingly fast after the initial run. Some other harnesses do similar things, so I decided to use bwrap around these tools for accurate benchmarking.
I feel like so many of the new AI tooling is reinventing the wheel. We've figured out how to update software via package managers. Do we need a daemon to continually check for updates and do we need shell scripts to circumvent package managers?
Have a look at https://chatgpt.com/codex/install.sh. It's managing the installation by itself not using existing package managers. There's a good point to making it easy to install, but it leads to managing an installation when you shouldn't need to. Time would be better spent elsewhere and better solutions exist.
I use it at work. On Windows it also launches a new console window every it executes (?) a command. You should definitely disabled this using --no-daemon.
I don't know how and why they ruined this CLI with needless "features". It was perfectly usable two weeks ago.
A bit off-topic, but on the rare occasions I help some acquaintance with Windows I always feel a bit upset that random split second console windows still pop up when installing everyday software, even that from Microsoft.
I remember them from way back on Windows 98 and no one seems to care to hide them (?).
Is anyone using the vscode extension of Codex? Its insane how brittle it is.
Just a couple of days ago, it was completely broken for about a week before it got somewhat fixed again. When I say fixed, I mean basic functionality is working again but bugs and regressions are still left. There are multiple issues tracking the same problems. The community have even started to patch the extension themselves as a workaround until its adressed.
You know whats even crazier? There are like a new release every other day with lots of commits and changes, yet it still in a bad shape. I have turned off autoupdate for Codex extension for this reason. The only extension I’ve had to.
I mean, it doesn’t take that much work to have some tests in the pipeline to make sure the extension continue to work, it almost feels like they are purposely doing this.
I have now experienced the consequences of ”move fast and break things”. I realize how much software quality means to me. Its not fun being on the receiving end, especially when it affects my daily workflow.
If any of you, see anyone hype up LLMs again, please refer them to the state of Codex vscode extension.
not to mention the daemon mode breaks a bunch of assumptions one might have about about codex CLI e.g. that codex CLI inherits the env vars of the shell that invokes it.
—no-daemon is the way, but still feels like the noose is tightening …
Refusal to adhere to the XDG specification is a big signal that the app is going to be a general bull in a china shop. It's not that hard to do OS- and distribution-appropriate package management.
A well-behaved application should not only use XDG on Linux, but it should use platform-native conventions on macOS as well: "~/Library/Application Support", etc.
Reminder that OpenAI allows you to use the Pi Coding Agent, which is a lot more lightweight, and in my experience much better than any of the proprietary harnesses.
I've been using Pi nearly every day for 4-5 months now, and even though it is still very much in development with new updates arriving nearly every week, it has to my mind never regressed. I've never lamented that a Pi update ever broke anything I depend on or made anything worse.
I immediately adopted the new tuiMode "fullscreen" feature the moment it landed, and it's been exceptional. I love the way it now manages the viewport -- this was also right about the time I migrated to Ghostty so that also made a big difference. Showing inline images with the Kitty graphics protocol has been a huge improvement too, I love it when a model calls `read` on an image file and I get to see the results in my terminal. The OSC-8 hyperlink support is super cool too. I love the way it auto hides/shows the scrollbar now too, and recently they've even made it so concurrent tool calls show different "themes" when pending/running vs success vs error.
Yeah, and for the contrary I spotted Claude app being ugly. When I turned off Claude, then put my computer to sleep and later woke it up, it made my fans spin loudly, and after I checked, there were like 2 dozen Claude processes and a lot of Python. What was it doing - I have no idea, but it was crazy. I'm setting up a VM with nothing but Claude and necessary tools. I feel Claude is not honest with everything it is doing.
Today I used Claude Desktop after not having used it for a while. It ran bash commands on my machine, like starting brew services etc. It did not have that capability before (I used to hook up some mcp to allow selective file editing), and nobody warned me that it currently has. Shady tactics, very invasive, freaked me out.
Can you share some more detail of how you would implement this? Are you thinking something like VirtualBox (?), running a virtual machine, and you're using - for example, OpenCode terminal in the VM...?
There aren't any documented settings to disable this, but there is a feature flag it currently sits behind. You can flip it in your ~/.codex/config.toml[1]:
There are also some files that it writes to in ~/.codex at startup: pid files, lock files, things like that. You can delete them and replace them with symlinks to /dev/null, then mark those unwritable with `chattr +i` or similar. Presumably then the background daemon will crash on startup, if the feature flag ever goes away.--
1: https://github.com/openai/codex/blob/0c329ad2c9d5a2592128a1d...
Yes this should be turned off by default. The daemon is useful when building apps on top of the daemon server, e.g. pi or pi-durable custom harnesses.
What makes ACP deficient for headless uses of LLM agents? Maybe fixing that or standardizing on another protocol designed for pluggable LLM agents (I guess ACP is more oriented towards pair programming in an IDE, at least originally) is a better idea imo.
For whatever reason, the daemon has been refusing to start and Codex has been grumbling to me. I was wondering where the daemon had come from, and thought maybe I missed something previously and that it had just broken.
Yep, the real issue is that it was something they didn’t have before, and then with Codex 0.157.0 enabled, they installed this persistent nonsense [1] that survives after the last instance of Codex has exited on your machine, which is just gross.
[1] - https://github.com/openai/codex/issues/48195It also fills up my HD with old versions. I investigated and it took up something like 12 gigabytes in total on my system; which seems insane to me, it’s just a harness?
People moan about how badly made Claude code is but Codex seems much worse on every metric
All of the harnesses I've used are terrible at cleaning up after themselves. They leave all kind of clutter behind, to the point where on my dev VM, I now have an LLM run once a day to just look through the disk and delete everything that looks like it's leftover garbage.
It's like fighting fire by employing an arsonist.
Installed "clor.com" and it got into a loop of downloading its docker image from ghcr them throwing it away immediately because it was supposedly out of date (iirc). 65mbps continuous transfers for a week until I looked at my network traffic and saw 500gb of rx per day.
Is that cruft or past conversation history / subagent history and observability logs?
Just garbage. Temp files, worktrees that weren't removed after subagents finished, build files... all kinds of things.
But that's agent behavior, and not harness responsibility...?
I just erased 20GB of Cursor’s cruft.
Harness problem, or model problem?
Let's be real, what do you think wrote the harness?
> I now have an LLM run once a day to just look through the disk and delete everything that looks like it's leftover garbage
...are you sure it wouldn't consider your personal files garbage?
"In my dev VM". Everything of value in there is in a git repo.
Can't wait until we get really well designed harnesses written in a compiled language that doesn't bloat my computer. Will probably have to be written by a company other than OpenAI because codex is like the exact opposite of everything I just listed
Found this one: https://usehax.dev/
Anyone experiences with it?
it's okay.
it's minimal and somewhat opinionated, not integrated well with memory subsystems (although that's kind of the point)
my only real criticism is that last I played with it it deals with slow provider responses very poorly , so lots of locally hosted llms require a lot of tuning to get it to handle streaming/non-streaming/timers well.
way less minimal (but still pretty minimal) I prefer Pi; but different markets I think.
hax fit my openbsd laptop well, espescially since Pi and related derivatives are kind of un-spokenly incompatible with the BSDs.
(author passing by) Could you please report the issues you've experienced on the repo? I did quite a bit of testing myself on local models (Qwen, Ornith, Gemma, Muse Glimmer) with llama.cpp, and they all worked fine. Details would be helpful.
Also one brief note for Codex users specifically. hax can work out-of-the-box as codex-cli drop-in replacement (since it can use the already authenticated codex token), so it's super easy to try it - basically it will be a zero-config experience.
Codex is written in rust, fyi... Not that I disagree.
Oh, I haven't been keeping up. I assumed it was still TS since it prompts me to run npm all the time.
They just use npm as a binary delivery mechanism, for whatever reason (I guess they expect all their users to have npm?)
It's pretty darn ubiquitous, yeah. Plus enterprises tend to run npm mirrors, so delivery integrates into existing IT compliance frameworks.
Which generates unnecessarily huge binaries.
It's because they statically linked it, and probably rely on a trillion pointless crates (as is standard in Rust development). A good statically linked harness can easily be ~5MB MAX
Just write your own harness. Ironically Codex/Claude are both abysmal.
How do you even write a harness that can use a Claude subscription? I thought it was not allowed by Anthropic's ToS and can get you banned?
It's trivial to bypass this by tweaking the system prompt in other harnesses. They use string matching and a blacklist.
https://x.com/FlorianKluge/status/2041855675295318039
(Whether that's against the ToS is not clear.)
That’s a different discussion.
I suspect a lot of people wouldn't use AI if they had to pay per token at current rates. Yes, it's technically a different discussion, but "just write your own harness" isn't really an answer for the vast majority of us riding on the investor-subsidized subscription plans.
If you're using the sub, only OpenAI (via the OAuth token) is viable. Claude is API_KEY only
It's allowed, but you get bumped to per token pricing.
That defeats the purpose of using the Claude subscription.
I did, and the weird things you have to do to make work at least decent. I’ve been doing evals and I got some things better than opencode but not as good as pi. Though my goal is to optimize for a browser and browser based llm so it can create stuff in a web container or at least generate ui by requesting json schema (a2ui style).
Omp
My preferred Nix package for this patches at least that behavior out, afaict: https://github.com/numtide/llm-agents.nix/blob/dc285bfd94ec7...
https://github.com/numtide/llm-agents.nix/
> People moan about how badly made Claude code is but Codex seems much worse on every metric
Also, it feels that chatgpt.com and aistudio.google.com are the buggiest websites I have seen for long time. Why it is always something related to LLMs?
All the LLM turds we've curled out at work are the same.
It’s actually insane how buggy the Gemini web version is (or at least was a few months ago, since I stopped using it).
Because they're the most likely to be 100% vibe coded. And as much as some people still deny it, AI code = buggy code.
ChatGPT.com is dreadful. The whole thing chokes when the conversation gets moderately long. It's amazing that it's been allowed to stay this way.
It’s extremely ironic that the software that’s meant to end software is itself pretty terrible software.
for some reason I had 11 separate codex and codex-code-mode-host binaries in my ~/.bun/install/cache/@openai . each weighing about ~300 MB. might be worth a look to anyone running low on disk space.
I haven't worked with the update subsystem before but that seems like an oversight to me. I'm gonna fix it.
Yep, .codex/packages/standalone/releases keeps all the old versions. Also each codex bin is somehow ~300MB.
thanks, I just removed all the old stuff.
what's the lazy way to check? can you share where/how you cleaned up?
Remember: VSCode runs a complete headless Eclipse instance as its Java LSP, so I don't expect quality code from any of the big companies anymore.
Are you referring to something built into VS Code, or an extension? It looks you're talking about https://github.com/redhat-developer/vscode-java, which is RedHat's doing.
Yes, my DEV VM tended to fill up with VSCode backups as well. I finally expanded the disk so it's less of an issue today, but for a while ENOSPC was resolved by cleaning up ~/.vscode-server/cli/servers.
Even if you installed codex with brew?
The daemon has many bugs. For example, when auto-updating, it would randomly ruin the permissions and block the goal as .git became read-only. Generally, it is extremely annoying how many bugs Codex has. I basically can't trust it to execute on any goal since it would often get blocked every few hours due to random bugs.
https://github.com/openai/codex/issues/52251#issuecomment-60...
Also, in Codex, one randomly can't click "New Chat" or "Scheduled Tasks". You know, the absolute two basic functions.
The CLI has been pretty good on this score, way better than the App. Has been…
My complaint applies specifically to the CLI. I have never used the gui app. Here are the various goal-breaking bugs I've run into:
- Goals get blocked due to .git and other stuff becoming randomly read-only when daemon auto-updates in the background (the linked issue in my comment above)
- Goals get blocked due to transient "Selected model is at capacity"
- Goals get blocked due to "Codex ran out of room in the model's context window". Somehow, both auto-compaction and manually running /compact don't work. https://github.com/openai/codex/issues/52567#issuecomment-61...
- Goals get blocked due to bwrap randomly failing with `bwrap: loopback: Failed RTM_NEWADDR: Operation not permitted`. I traced this one to `apparmor` on Ubuntu 24.04 interfering with bwrap and found a local workaround.
I’ve been using GPT models via hax[1], a bare bones harness with minimal tools on my system and have the models rely on bash for the most part the last few months. It’s been a breath of fresh air amidst all these overengineered harnesses.
If I need special actions, I just make them CLI tools on my Mac and tell the AI model that such a tool exists(via a single line in its system prompt) and that’s it. It’s been very nice.
[1] - https://usehax.dev/
No affiliation whatsoever. Just a happy user.
You seem to care, so I wanna point out a semantic nit; "such a tool exists" could be interpreted as a harness provided tool, rather than a command line tool. (Edit, on a re-read its pretty obvious you mean cli tool, but keeping this statment here as im sure it'll open someone's eyes to the difference)
I've gotta give hax a go, as a pi user who came on-board due to the no frills direction pi takes. I say takes rather than took, because regardless of the (heated) discussion surrounding mcp and code mode support, I cannot fight the numbers- augmenting the harness provided tools is a notable boost to execution round trip performance.
I may need to take a second look at Pi again if it has any niceties like that notable boost in round trip performance compared to hax.
I mostly stuck around with Hax because of a few design choices. The sessions are merely plain text files indexed by a dirpath slug. This means I can just start a hax session in that directory and expect sessions history to be available in any model.. (I occasionally dabble in GLM for security checks, Gemini for writing stuff, Deepseek for tightly scoped tests etc).
Off-topic: I think the front-end quality of ChatGPT/Codex Desktop and Android app is simply dreadful—hard to imagine how such a large number of frequent failures made it into production. Do they not use human testing at all?
Is there evidence to suggest that open ai business leadership values money over inspiring general public customer delight?
probably not much, they're all using workflows where the agents pull issues and automatically merge code when the tests pass. so lots of stuff slips through and the overall experience is kinda shite.
This stuff is moving fast, should update automatically. This is kind of standard now. Browsers do it. No reason for harnesses not to do it as well.
Ordinary package managers are totally capable of updating thousands of programs without manual effort, including automatically and in the background if that's a thing you want. But they do it without violating users' sovereignty over their own computers or turning every app into a special snowflake that gets to install software (namely itself) in its own bespoke way.
My browser doesnt update automatically, i have to approve it.
Maybe this kind of stuff is standard for some people but I absolutely loathe auto updating software for many reasons
One practical reason - system stability for important events.
I travel for work, and often have a tech demo, solution, tool, or something to show - or am running a workshop.
I usually do NOT update things right before my trip, as I want to minimize the potential for disruption.
Isn’t this what docker is for?
Jokes aside I do the same I tend not to upgrade my Mac, my phone unless I really have to.
The “update” button on the top of my brave install has now turned from green to orange and I’ll likely wait until it goes red
> Isn’t this what docker is for?
What? Pinning a browser?
First thing it'll want to do after starting is update itself to newest version anyway.
For example, Firefox updates by default and forces you to restart it, disabling using websites or seeing any site content before doing so.
On your setup.
Firefox does not only not prompt me for updates (it only updates if I manually run an update), even if I do update with Firefox running, it does not stop me from continuing to use Firefox without restarting it. All I did was install Firefox from flathub (packaged by Mozilla (a default configuration provided by Mozilla)).
I am running Firefox Developer Edition on macOS, where it is the default option. Have to set "Check for updates, but choose when to install" to get the similar behavior.
I've never seen that behaviour, it downloads updates in the background and then gives me a little notification that I should restart it to update, but it's dismiss-able.
I think this is the old default option. What options you see in the update options?
I've been using Firefox for years and don't think I've ever seen this. Are you running some custom configuration?
Standard on Linux, I believe. One of many reasons I got back to Windows as my main OS (with Linux needs fulfilled by WSL).
Not true since summer 2025.
https://www.firefox.com/en-US/firefox/141.0/releasenotes/#:~...
Firefox Developer Edition on macOS.
Iirc, this is a pattern of the flatpack version of Firefox. I do not believe the .deb does this.
Are you sure this isn't your organization's update policy kicking in here?
Being able to turn auto-updates off is okay, but that being the default would certainly not be okay in Codex, giving the reach, what it can do and how immature it is.
You can do this without a daemon, and without storing n previous versions as noted by another user.
Yeah this seems like a decision yielded by a model. One of countless examples of the model’s outputs careening into individually defensible but holistically nonsensical results
A human seems less likely to embark upon a path that most humans know leads to something a human is unlikely to want
LLMs produce amazing things, but only one token at a time. Anything beyond that is retrospectively emergent at best
I guarantee the prompt “add auto update feature” was the first and last human input applied to that feature.
Serious tools shouldn't do this. Sure, you can offer an opt-in rapid release beta update option, but if you can't go more than a month without forcing me into a "stable" update that changes the UI and behavior I'm getting used to, then I'm not going to consider your software stable and reliable enough to be worth developing expertise in using.
The entire class of software (harnesses) you're talking about was literally invented like 2 years ago...
If software engineering is solved they should be able to get it right the first time
Software Engineering clearly is solved, that's why Google has a tiny animated banner on the AI Studio sidebar (advertising subscription upgrade), which shows a simple animated border around it and utilizing exactly 100% of the GPU to render just that banner. And there is a browser extension removing that styling element (without banner GPU load is obviously under 0.1% as it should be) released way back in May. So the almighty AI didn't manage to fix a major UI blocker for a flagship AI product in half a year now. :)
When I had a Claude account (cancelled ~May time) I vividly remember the web chat having broken editing for backtick-quoted monospace text. If you typed <space>`foobar`<space> the last space would move the cursor backward into the closing quote, and the space would be emitted inside the monospaced region. It was like this for several months (on Firefox). I always found this baffling for a product marketed to programmers.
Same with their desktop app by the way, absolutely infuriating.
Coding is solved, software engineering still TBD
CODEX WRITES ITSELF! DONT YOU PEOPLE UNDERSTAND!!!!!
Nah, you're confusing it with Autolith.
So you imply that Codex has a hoarding issue?
It loves hording issues. Over 15 THOUSAND on GH.
It gets very attached to itself
Products moving fast doesn't mean that they're moving in the direction I want, or that I want them to move underneath me while I'm in the middle of a project.
Auto-update systems have value but are not a suitable default for professional tools.
Then use the package manager, or windows store on windows.
This is the same operational model as Claude Code.
I see a few prior versions of the Claude Code binary and several running background tasks, without any copy of Claude Code in the foreground. I regularly get a "Update available" message in my Claude Code UI too.If running systemd, wrap that crap into a cgroup and force kill the whole process tree on exit. I've been doing it for years with badly behaving applications. On phone, from memory, may not be exactly what is needed:
This kind of stuff often breaks for applications packaged with Nix, so it's often removed from Nix packages for compatibility reasons. Both behaviors you mention are helpfully patched out of the Claude Code package I use: https://github.com/numtide/llm-agents.nix/blob/dc285bfd94ec7...
As a result, I never knew that Claude Code does a nasty thing like this, either. :-\
Fwiw, packages in that repo get version bumped each day, so if that's a fast enough pace for you it may be a good alternative.
And here is the issue raised in github: https://github.com/anthropics/claude-code/issues/14980
TLDR: They don't care.
It's right about now that having a operating system distro model is great - you get folks who will package the software & test it out, make generally sensible defaults, and generally keep shenanigans like this from going on
Nix/NixOS takes it a step further and makes you have to go out of the way to enable self-updating behavior when packaging something :)
What are the top ones you recommend? ( I am running Mint Debian Linux )
Well, OpenAI at least is fine with you using whatever harness you like. Having a crappy first-party one is bad, but an escape hatch puts a floor on it.
Codex is very badly structured. I was an early user with codex. But it mixes state and configuration data; which is why I can't really declaratively manage it in my nix-darwin environment; so abandoned it a while ago. This issue being raised, and the way that the codex development responds back to it makes me not trust their harness ever again. I only use codex in opencode; which seems to be doing OK. At least openai doesn't gate their product to their harness like anthropic.
Not trying to sing praises; i think opencode is doing a bunch of weird stuff too but it's not codex level.
You can just copy the executable then run it with --no-daemon, or edit the config equivalently. No autoupdates and no background process.
The recent terminal UI changes were more fastidious, but even those are easily disabled with CODEX_TUI_DISABLE_KEYBOARD_ENHANCEMENT=1 and the /tui command.
I think the quality of Codex CLI (and Claude Code) is an honest reflection of the status quo of LLM-driven development. The future doesn’t look really promising.
To be fair we've gone from barely capable at writing code to what we have today in barely 3 years. I wouldn't put money on the counter bet that they won't get beyond this hump.
I wouldn’t assume that progress is linear. Maybe I’m wrong and it will be! But I’m reminded of early developments with self driving cars, it felt like manual driving would be a thing of the past within a decade. But progress has been much slower than that.
> But I’m reminded of early developments with self driving cars, it felt like manual driving would be a thing of the past within a decade. But progress has been much slower than that.
I think that's exactly the standpoint of Yann LeCun.
Let me clarify your illogic for you:
> The progress using narrowly intelligent algorithms to drive cars in a generalizable way was not as fast as the early gains in places the narrow algorithms could solve, thus the progress in using generally intelligent algorithms for any other field must also slow down.
Man, is this how 99% of the population works? They have a heuristic for anything that sounds computer sciencey and apply it for anything and everything that sounds similar without bothering to check if their heuristic actually applies?
No need to be rude! Besides OP just said:
> I’m reminded of
They were drawing a comparison, not saying "x will definitely happen because y"
No need to be such a dick, it reflects poorly on you.
You committed your own fallacy of applying a sloppy heuristic to understanding the person you were responding to without bothering to check if it applies. Spectacular. The comical levels of arrogance that infect over-hyped tech marketing boondoggles (one area among many in which GenAI is clearly similar to self-driving cars) is one of the reasons nobody trusts the hype anymore.
The breathless astroturf articles, the brain-dead talking head conversations, the endless flood of obvious misinformation about industry uptake, the nauseating LinkedIn puff-pieces about PREPARING YOUR ENTERPRISE OR ELSE... it's a hostile information economy at best, and its being driven by the same people who brought you bullshit like self-driving, crypto and the Metaverse... for the exact same reasons (to get rich).
There's no epistemic credibility left in the tech industry in general. Maybe the Dario's or whoevers of the world believed what they were talking about at one point, but I don't believe a single word I read from Anthropic or OpenAI for the same reason I don't believe anything I read from Russian state media. I know it's being produced for reasons that have nothing to do with the truth.
As someone who interacts with the technology on a daily basis at a non-trivial level... I'm not seeing anything that would cure my skepticism about GenAI. The information economy surrounding it is broken. The fundamentals don't line up. The hype about "nearby capabilities" is identical to the hype of generations after generations of hype in the AI space (dating at least back to Turing). It smells like a bad bet.
The product and UX of the tool is distinct from the raw ability of the tool.
Claude code has great UX.
It is hands down the most user hostile UX I put up with. The fact that I put up with it speaks volumes about the value proposition that the underlying models bring to the table.
I’m in the process of dumping codex for more open harnesses that don’t add problematic new features every week. Because Anthropic insists that I have to use their harness with the subscription, eventually I’m either going to wrap `claude -p` to get away from the crappy TUI or else I’ll just unsubscribe and use other models. Their bad UX has alienated me as a paying customer.
Can you give me examples of things you don't like? Because I unironically admire how they improve UX. It's subtle, minimal, yet insanely useful.
Listed a few in my other comment
Try using the "rewind" feature to go back to a previous message.
Went too far? Oops, it's not accessible from the Claude Code UI anymore! You can get the agent to dredge up the info from the session logs, but you'd better hope they put the pieces together correctly.
Also, you can't rewind to a steering message. Want to make sure you can rewind to it? Wait for the agent to settle, or stop them, enter your message, and hope you don't throw them off their current train of thought.
To see a proper rewind UI, see how Pi does it: Every message is available, forked conversation history is preserved. It has its own issues, but is miles ahead of Claude Code in this regard.
I'm not who you're asking but for me I am right now going through the insanity of providing plugin/mcp/tool access across environments. This is in codex which as I mentioned is intentionally trying to provide mass-market ergonomics. But with that comes incomprehensible configuration that's also a moving target. chat/work/codex/remote/cloud are all different things. And each client exposes different abilities and hides others. And looks like only the web-ui is the most up to date because other targets need releases. As a developer I understand why it's in its current state, but that in no way excuses it as acceptable ux.
Like of course codex is the most powerful and "just" run `npx blah blah` but no, I'm not running randomly piped code on my personal computer and even if I didn't care about privacy it means I can only run this locally with my laptop on. "leave your laptop on and access from your phone" is not a good idea. It only adds to the surface area of confusion. Ux no good.
No, the tooling I make has better UX very quickly. It is built for their financial goals, not for the end user.
this 1000x. a lot of details in context compaction, editing, reverting ux, etc are all really centered on keeping the load and cost of subscription users as small as possible. its no always that nor is it necessarily correct, but its a huge driver of how those subscriptions and first party apps are designed
I know this is an unpopular opinion, but it's actually good. - pasting lots of content became [35 lines] and they added a tiny notification to show it as-is if you paste again - interactive questions layout - paste images - the recap - auto-updating session titles
And so many more. I tried codex and it was lacking in many UX aspects. Remember we got from a text-paragraph-only terminal to this at breakneck speed.
I’m not deeply pro-codex, but it does have all those features. Who knows since when.
agreed. the thing i reach for most is its ability to coordinate my agent graph, e.g. seeing what sub-agents are doing, messaging agents, adding new capacity. claude code really is a delight to use here
Claude Code is better than Codex UX. But as per the sibling response it's more a developer tool made by developers for developers. It does have a lot of niceties. Codex/openai on the other hand is more clearly trying to reach mass-market and it's a UX shit show because of it. Consumer UX is its own thing. So the original point is that raw ability of the tool doesn't magically solve ergonomics.
Won't disagree with you because I haven't used codex/OAI a lot - but Claude's non CLI UX is great too.
Eg - I asked for a recipe and it says leave x for y time - there's actually deeplinks that link to your clock timer with the time preset! And to change quantities, there's input boxes in the response that update all quantities accordingly.
Many such cases of use-case specific UX
The current approach of the big companies is to keep increasing the number of parameters (there was a report that recent Claude models are on the 4T parameter scale). Even recent SOTA open weight models are in the 1T-2T range.
This is partly due to the multiplicative scaling factor -- the number of tokens; the number of transformer layers; the number of features; etc.
I suspect that for models to have a step upgrade either/both of 1) a new/improved architecture model; and 2) better training of smaller parameter models are needed. I say the latter as I think that with the larger weights there is a lot of unused/redundant parameters.
I dont think the quality has gotten that much better. Take claude 4.5, the probability of getting usable outputs was less likely than with opus 5.5, and the usuable outputs weren't great.
I think newer models are more likely to produce the usuable output, almost always, but its still not great. The frequency in which you get usable results leads one to believe that quality is improving.. No usability is improving but the result is still meh.
Also put an llm in an environment without an lsp or compiler to check against and let me know how capable you think they are on their own. The "loop" is doing a lot of heavy lifting here.
If we’re learning anything from the past 15 years it’s probably not to extrapolate AI progress in a straight line.
https://finance.yahoo.com/sectors/healthcare/articles/godfat...
Progress is spiky and not guaranteed. The capital costs have also ballooned to the point where the rate has to level out
> The capital costs have also ballooned to the point where the rate has to level out
That's a pretty bold use of "has to". None of the financing behind AI seems sane, yet it continues to defy rational expectations. Reminds me of the comment "the market can remain irrational longer than you can stay solvent". It's not the exact meaning but has the same feeling. The circular financing going on does not help anything either
Time doesn't solve bad designs. We still have lots of buffer overflows in C.
We don't see any indication that the LLM companies are slowing down and go for better code.
Instead it just looks like a cash grab until the AI bubble bursts.
I say this about 30 times a day, to nobody. "How the FUCK is this all so bad? They have unlimited tokens"
I think it’s a new spin on Brook’s Law. Adding tokens to a project makes it less efficient not more.
The "Mythical Man-Month" is now the "Mythical Token-Minute"?
that's what I'm also wondering about.
in the last 2 weeks or so: - scrolling stopped working - subagents no longer work (main agent doesn't get notified when they finish) - subagents don't work (a subagent can't make its own agent even if instructed to do so)
what's going on? I thought coding was solved?
> They have unlimited tokens
That is exactly why it is so bad. Tokens spent generate code. Large amounts of code is a liability.
Both CLIs are getting worse. I think it's because they keep bolting in unconsidered "features", not because of the code quality.
But maybe being able to code up a feature at the drop of a hat has turbo-charged their ability to make bad UX decisions.
I believe it's all related. They keep adding new, badly thought out features because it's much easier than improving existing features built on low quality code.
The latter requires modifying existing, badly architectured and tightly coupled code that will almost definitely break something somewhere else if touched, so what do you do? Just keep adding new code on top and hope for the best.
This is not a new phenomenon by any means, but AI amplifies it by 100x.
It has pretty rapidly reached the point where I do not think any acceptable LLM-harness is widely available. Every single one I've tried has broken in bizarre ways on even minor updates.
I used codex the first time yesterday and quickly decided I need to sandbox it and came up with a pretty easy and straightforward solution (I think).
I use nixos and use a little wrapper that wraps Codex CLI in Docker. Run codex-box /path/to/project and it starts or reuses a container with only that project bind-mounted, keeping login and history in a separate volume. It uses tmux for detaching and reattaching
https://gist.github.com/entropie/8f30b9e33e9412fda3adf2f7507...
Build your own, using the broken harness
Maybe thats just the tool nerd in me that just wants the perfect tool for every job.
Very much this... The utility of Claude has grown exponentially with the amount of work I've put into memory/context management, tooling, skills development and CLAUDE.md (and friends). I freely admit I may just be fortunate, but I've yet to hit a bug in a new release that rose above the level of minor annoyance.
The thing that really validated this point was when I recently tried Codex again. Out of the box it was OK, but not great. I swung over the memory apparatus and had it import directives from my existing Claude install and it was almost immediately quite useful in its own right and, especially, as a counterpoint to Claude when dealing with some fairly complex operational issues.
Pi as well? Comparatively it's quite a small surface. I haven't seen issues when I upgrade.
Pi went searching across my entire home directory unprompted... Idk to do that without some kind of sandboxing builtin is rough and chews through tokens on things I don't really want polluting my context.
A coworker of mine has a pretty neat script that uses bubblewrap to wrap any application. I'm using that one too for pi and it's pretty great since you don't have to trust the harness and can quite easily customize the sandbox to your needs, e.g. disallow reading to certain directories etc.
This blogpost describes a similar thing: https://bxt.rs/blog/easy-sandboxing-on-linux-with-bubblewrap...
Well yeah, it's YOLO-by-default by design. That's not related to upgrade churn.
I find codex CLI to be a pretty high quality application. Is the existence of a backend daemon what you're referring to ?
I've used the CLIs for Codex, Gemini and Claude and they all give off the feeling of being on an very rickety, very fast runaway train.
You install an update. Oops, they added a sandbox that doesn't work inside a docker container, it's stopped working.
You install an update. Oops, you can't log in inside a docker container any more, it's stopped working.
You install an update. Oops, they started catching the copy+paste shortcut and replaced it with one that doesn't work.
You revert a broken update. Oops, the update changed the format of your login details, the old version can't read them, gotta log in again.
You install an update. Oops, can't scroll with the mouse wheel any more.
Not showstopping bugs (if you're used to persuading broken oauth logins to work) but it certainly makes updating an "ugh, what have they broken this time?" experience.
You got further than I did. I checked them out and went back to manual prompting. Just don't have a need for codegen that much, and when I do I'd prefer to be in control.
To be fair, how many products this old even have a revert update mechanism?
It's buggy like hell once you do not run the default settings. Whole permissions system is completely broken, for example.
It’s kinda vibe coding in a nutshell. Lots of cool and varied features, that often just don’t quite work right.
It does some neat things that, ultimately, are super obnoxious. Background animations and "sparkle" effects that make my terminal indicate activity at all times. Just subtly broken behavior.
Totally.
The problem is hardly anyone gives a shit if they can pump out stuff that makes them look like they are good at stuff.
Vanity before engineering.
https://www.jerrypournelle.com/reports/jerryp/iron.html
I think the (lack of) quality of these tools shows that software engineering is maturing as an engineering discipline.
We were overbuilding trivial parts of the software like the Romans overbuilt small bridges.
Modern engineering is about making the thing that barely works. If you disagree, learn a little bit about bit narrow the tolerances of modern cars and bridges are.
> learn a little bit about bit narrow the tolerances of modern cars and bridges are
Buildings typically use a safety factor of 1.5 to 2.0, and cars typically use a safety factor of 2.0 to 3.0. These things are being built twice as strong as they need to be. I wouldn't call that narrow.
And with software, we're at a point where it can't even properly do what it needs to. If we were overbuilding it, damn software wouldn't crash so much.
I think all engineering has safety factors of 2 or more. Maybe with the exception of aerospace (mostly space) where materials might be at the limits of what's possible right now so the safety factors are much reduced.
> I think all engineering has safety factors of 2 or more.
Not even close.
For example, a BMW X5 weighs about 2200 Kgs without passengers, and it's GVWR is 2800 Kgs.
If you have five people and heavy bags, you can comfortably exceed the GVWR. You can absolutely break control arms and suspension components in that scenario if you go over a speed bump or hit a pothole (which are real things that exist in the real world, like people doing something unintended and crashing software).
If this needed a 2x safety factor, there would have been no way to build this car.
Even the toughest cars like land cruisers are limited in this way.
I think there's a difference between something that's engineered and precise vs something that's rickety. Just because you built something that might fall apart at any moment doesn't mean it's been engineered with narrow tolerances.
I disagree.
> [...] vs something that's rickety
A 100 percent of modern bridges are rickety. There has been dozens of high profile incidents because of this.
> Just because you built something that might fall apart at any moment doesn't mean it's been engineered with narrow tolerances.
Here the software they are building falls apart in ..say.. two days of use, but can be updated in 5 seconds at close to zero marginal cost. The software is useful enough that people keep updating and using it. This iteration gives them a competitive advantage over the safer choice of rigorously testing every new thing for several weeks.
I don't disagree but it's clearly a problem that already existed for decades in software and is just being 10x'ed. how many different frontend frameworks or AWS regurgitated and enshittified cloud services did we have to endure over the years.
There is actually even worse code quality written by humans, is that small sample also a reflection of human-driven development? And therefore, the future doesn't look promising that way either?
The difference is the source of the sample. This is a flagship product by a frontier lab. It should be reflective of the best that is possible. It should be biased in the most favorable possible way. It was designed by people without the normal limitations and quotas of outsiders in building it, as well as in learning experiences in driving it. This is not a random cherry-pick of bad code.
The restraint, rigour and responsibility you showed prior to LLMs relates directly to how much of that you apply now that they’re here.
Given how rare it was previously, and how little it was cultivated by shortsighted management, I’m not surprised with all the slop we see today.
The truth is there’s nothing stopping someone from using these tools to remove dead code, simplify project structure, remove dependencies, etc. but I don’t see it happening very often. I have done it a bunch myself though so it’s absolutely possible, it’s all a matter of investment and prioritization.
Seems like its just a bunch of young engineers making bad architecture decisions. Not sure I would blame the culture of vibe coding everything. It might just be bad taste.
If you blindly trust frontier LLM architecture 'decisions,' they often do look a lot like what a young/junior engineer would do.
I'm inclined to agree
i'm personally quite perplexed how sloppy the harnesses are from the leading AI labs, you'd think these are their primary products and would have some human attention devoted to them
it's embarrassing when every claude code release has obvious bugs in it like telling you to upgrade when you're on the latest version or codex opening visible powershell windows for every command that pop up on your screen
like 5 minutes of human testing would catch this obvious shit but obviously they don't bother because vibes are the future and stuff? idk
Solar power has become a real transformative technology and instead of trying to recreate a better world we tariffed them and the execs of this country spent all our coins on power hungry data-centers that write mid code.
I'm sure if you file a bug on the 15k backlog they'll fix it eventually /s
Legit bugs still open 6 months on.. It's a bit ironic
It's not necessarily hostility. We considered different updating approaches with our legal IDE where security, stability, etc. are paramount, and having a buggy application out in the wild is user-hostile. Sure, in a dream world we'd say "don't ship until it's perfect", but that's just not realistic.
We considered the daemon thing, but ended up with a hard-update check on every boot. It's thus a user choice: if you don't like the automatic updating, you can just not run the application anymore.
And even with all that, it's hard to stay on the right side of AV algorithms, etc.
We wrote about it here: https://tritium.legal/blog/update
> It's thus a user choice: if you don't like the automatic updating, you can just not run the application anymore.
That's not a choice. It s just what most apps do these days, but it's the first time I've seen anyone stating this so directly, or trying to frame it as a choice.
(And behaviorally, this makes your product just a SaaS website with extra steps.)
I guess I mean it's a choice in the sense that it's a transparent part of the product that one chooses how and when to use (or not). That's opposed to something running in the background on its own schedule, with its own agenda whether the user is running the application or not.
[Edit: yes, that is kind of the idea. Native version has better features because it can access them.]
My 2-cents: never install those things into your host. A locked down Docker container is the bare minimum. Pin your versions and never update them without checking the changelog.
Cool thing is that I can just ask the llm to modify the harness to what I need. I discovered this daemon thing about two weeks ago and had codex modify it's own custom launcher that I have to bypass the daemon thing. If as a recreational code farmer can do it then most "pros" should be able to handle this and other inconveniences. Cc has been pushing that no scrolling feature that codex recently introduced for a half a year now but again I had them put a stop to that. Software has never been this easy to configure.
I discovered this while benchmarking the tool I'm developing. I realized Codex was surprisingly fast after the initial run. Some other harnesses do similar things, so I decided to use bwrap around these tools for accurate benchmarking.
while this points it out, i'll show you what I had in place:
~/.codex/config.toml
hell, I also have this redundancy configured in bash:Continuing to be glad I trust my instincts around bad management making problematic products!
I feel like so many of the new AI tooling is reinventing the wheel. We've figured out how to update software via package managers. Do we need a daemon to continually check for updates and do we need shell scripts to circumvent package managers?
Have a look at https://chatgpt.com/codex/install.sh. It's managing the installation by itself not using existing package managers. There's a good point to making it easy to install, but it leads to managing an installation when you shouldn't need to. Time would be better spent elsewhere and better solutions exist.
Am I the only one thinking this?
I use it at work. On Windows it also launches a new console window every it executes (?) a command. You should definitely disabled this using --no-daemon.
I don't know how and why they ruined this CLI with needless "features". It was perfectly usable two weeks ago.
A bit off-topic, but on the rare occasions I help some acquaintance with Windows I always feel a bit upset that random split second console windows still pop up when installing everyday software, even that from Microsoft.
I remember them from way back on Windows 98 and no one seems to care to hide them (?).
In some circles this is called a backdoor.
Is anyone using the vscode extension of Codex? Its insane how brittle it is.
Just a couple of days ago, it was completely broken for about a week before it got somewhat fixed again. When I say fixed, I mean basic functionality is working again but bugs and regressions are still left. There are multiple issues tracking the same problems. The community have even started to patch the extension themselves as a workaround until its adressed.
You know whats even crazier? There are like a new release every other day with lots of commits and changes, yet it still in a bad shape. I have turned off autoupdate for Codex extension for this reason. The only extension I’ve had to.
I mean, it doesn’t take that much work to have some tests in the pipeline to make sure the extension continue to work, it almost feels like they are purposely doing this.
I have now experienced the consequences of ”move fast and break things”. I realize how much software quality means to me. Its not fun being on the receiving end, especially when it affects my daily workflow.
If any of you, see anyone hype up LLMs again, please refer them to the state of Codex vscode extension.
It's as if there was a reason the whole industry was obsessed with quality up to like 3 years ago
So, similar as Google and Microsoft have been doing with all their tools?
All I'd want is a CLI with a few tools AND a built-in, properly configurable and working sandbox. Seems that's too much too ask.
Just be the 10th person to make one this week.
not to mention the daemon mode breaks a bunch of assumptions one might have about about codex CLI e.g. that codex CLI inherits the env vars of the shell that invokes it.
—no-daemon is the way, but still feels like the noose is tightening …
Refusal to adhere to the XDG specification is a big signal that the app is going to be a general bull in a china shop. It's not that hard to do OS- and distribution-appropriate package management.
A well-behaved application should not only use XDG on Linux, but it should use platform-native conventions on macOS as well: "~/Library/Application Support", etc.
Sandbox everything if the OS is not doing what it’s supposed to do
Exactly, that's what I do, I even up end up writing my own for this purpose - https://github.com/ashishb/amazing-sandbox
Claude has been doing this already.
ah... so this is why my agents all made a signal noise when i used chatgpt on my phone?
Reminder that OpenAI allows you to use the Pi Coding Agent, which is a lot more lightweight, and in my experience much better than any of the proprietary harnesses.
https://pi.dev/
I've been using Pi nearly every day for 4-5 months now, and even though it is still very much in development with new updates arriving nearly every week, it has to my mind never regressed. I've never lamented that a Pi update ever broke anything I depend on or made anything worse.
I immediately adopted the new tuiMode "fullscreen" feature the moment it landed, and it's been exceptional. I love the way it now manages the viewport -- this was also right about the time I migrated to Ghostty so that also made a big difference. Showing inline images with the Kitty graphics protocol has been a huge improvement too, I love it when a model calls `read` on an image file and I get to see the results in my terminal. The OSC-8 hyperlink support is super cool too. I love the way it auto hides/shows the scrollbar now too, and recently they've even made it so concurrent tool calls show different "themes" when pending/running vs success vs error.
Claude also now hacks your MacOS to put an "Ask Claude" item into every right-click menu. You can remove it through the system's keyboard preferences.
Yeah, and for the contrary I spotted Claude app being ugly. When I turned off Claude, then put my computer to sleep and later woke it up, it made my fans spin loudly, and after I checked, there were like 2 dozen Claude processes and a lot of Python. What was it doing - I have no idea, but it was crazy. I'm setting up a VM with nothing but Claude and necessary tools. I feel Claude is not honest with everything it is doing.
Today I used Claude Desktop after not having used it for a while. It ran bash commands on my machine, like starting brew services etc. It did not have that capability before (I used to hook up some mcp to allow selective file editing), and nobody warned me that it currently has. Shady tactics, very invasive, freaked me out.
Can you share some more detail of how you would implement this? Are you thinking something like VirtualBox (?), running a virtual machine, and you're using - for example, OpenCode terminal in the VM...?
I've used pretty much this type of setup:
https://michael.stapelberg.ch/posts/2026-02-01-coding-agent-...
If you want something not NixOS-related, smolmachines.com (I haven't tried this yet) gets mentioned here a lot.