general

[solid pecan]

It is getting better because of the OpenAI harness :D:D:D

Hahahaha I could not help myself

[lean tangle]

I reported this one in the beta. guess it got lost in the chaos.

[pallid kestrel]

@BenTh thanks that’s a reasonable and understandable response. But folks being dicks on here is super uncalled for.

[feral turret]

ok switched to 5.4-mini let's see how this goes

[steel vector]

real

[steel vector]

i'd love to know the end result of what you decide on

[feral turret]

thanks it seems much better right out the gate

[solid pecan]

@pallid kestrel I think I know what you are saying, by the project putting an OpenAI harness it disadvantages other models incentivising more people to use OpenAI with OpenClaw, that being said, it is an open source project, nothing is stopping the other providers like Anthropic submitting code to use their own as well. If the project did not accept it from other projects I get your point although that has not happened yet

[queen arrow]

With the recent changes to anthropic, is there anything I need to do differently to leverage the free agentic tokens or will it start using those by default until they're used up?

[crisp nimbus]

Lol at least when it still gets your telegram messages it can fix the rest xd

[pallid kestrel]

@solid pecan Exactly.

[crisp nimbus]

This is the exact issue I had. I had my agent fix it you can see my earlier messages here to give your agent an idea for the fix

[cold jacinth]

This is where the hostility started Mikey Mike

And this is discord. Everyone acts like a dick and it’s not personal

[feral turret]

🍿

[lean tangle]

did you submit a github issue and PR?

[crisp nimbus]

People should be dicks less and more helpful and understanding of each other

[crisp nimbus]

No. Too lazy and also I doubt they will see it with all the other shit they got on there. Feel free do to it though

[last tangle]

It’s also not that deep or even that big of a change it’s literally one line config change. The only difference is codex welcomes Openclaw Anthropic does not.

[solid pecan]

Yea like Anthropic should just submit their own as well

[lean tangle]

I think they're ignoring my issues now. lol

[solid pecan]

I have no problem with the OpenAI harness by the way

[last tangle]

There’s nothing to really submit Anthropic doesn’t allow it

[cold jacinth]

They’re so tight on compute they don’t want consumer anymore. They want corporate only. Much more profitable

[crisp nimbus]

I also got mostly ignored with my questions here lol. Or “make sure to install it from an official source” bruh it doesn’t get much more official than hackable install

[lean tangle]

I need to figure out what's going on with discord. If I can't, I'll have to revert to 4.23 (again)

[pallid kestrel]

@BK @solid pecan Ok, well if it doesn’t affect me I won’t spend more cycles on it.

[solid pecan]

Which is why @pallid kestrel should not blame the OpenClaw project 😄

[cold jacinth]

What’s your question

[pallid kestrel]

@solid pecan Nobody is blaming anybody. I’m asking a legit question.

[shut bloom]

women

[cold jacinth]

Ikr?

[solid pecan]

I think you are looking it from an egalitarian point of view and a little too ideologically pure, "we need to treat all the harnesses the same". Although in the real world it does not work that way because some companies like Anthropic hates

[steel vector]

👌 👌

[feral turret]

damn yea 5.4-mini is super expensive

[cold jacinth]

I moved all dependencies into a docker container on my Mac after all the latest supply chain attacks and now even the smallest or takes so long

[crisp nimbus]

@cold jacinth this was one of them.

[last tangle]

Your also free to change the harness back from codex to Pi if you do use a codex model but I don’t recommend it

[steel vector]

rip ig

[cold jacinth]

Amigo, I’ve gotta be honest… using a weak model is asking for trouble

[feral turret]

too expensive for me already. i need to get a local model i suppose

[crisp nimbus]

@cold jacinth also my bot wasted tokens answering this other one.

[pallid kestrel]

@solid pecan Yes, I don’t understand Anthropic’s position.

[solid pecan]

You need to understand in life that it is not up to you to sometimes balance the scale.

[last tangle]

They’re greedy and want more api money

[zenith violet]

It's simply a cost optimization on their end

[solid pecan]

Why should people using OpenAI be punished?

[cold jacinth]

Well, my answer is no. Telegram is fine.
For most people the update has been smooth. Historically issues like this can be config / not system

[cold jacinth]

When you have limited compute, you need to maximize profit on every token

[pallid kestrel]

@solid pecan thanks, now I understand what I was doing wrong.

[cold jacinth]

OpenAI is being great to users because they want market share but they lose tons and tons of money

It’s risky

[shut bloom]

like jeeeeesusss

openai are tryna convince us they have some cool security thing and then boom its jsut gpt 5.5

[last tangle]

Enjoy it while it lasts

[shut bloom]

so its ass 😭

[cold jacinth]

Compute will never be as cheap as it is today

[steel vector]

are you using API?

[shut bloom]

why do you think

[cold jacinth]

We are living in the golden age of cheap ai

[lean tangle]

happened again. Sent a message to the agent via telegram. I see the agent "typing" and see the tool working messages... then once it has an answer, it doesn't get sent to telegram (but it's in the webchat)

[shut bloom]

bro stop saying this this gonna make me cry

[lean tangle]

nothing in the log

[steel vector]

it'll get cheaper some day

[solid pecan]

I think the cost of tokens will drop again, DeepSeek V4 Pro is very good, open source coding models like Kimi as well are decent

[last tangle]

He’s right

[solid pecan]

That causing pricing pressure

[arctic trench]

Completely incorrect

[shut bloom]

actually bk i have a question, Dont you think claude will always want to compete with the cheapa sf chinese models

[upbeat token]

4.23 still?

[solid pecan]

I am not moving off 4.23 until LTS 😄

[shut bloom]

claude will need to compete with the ever better chinese models thatcost nothing

fusion is nearly achevied energy has never been cheaper

[lean tangle]

I'm going back to 4.23. 5.12 has telegram and discord issues.

[cold jacinth]

People liken it to the fiber optic spend of the 90s but fiber optics don’t have a useful life. Yes billions and billions are being poured into data centers. Those chips wear down. Nobody will want to replace them

[last tangle]

What are these so called better Chinese models you speak of

[upbeat token]

I have tried changing for the 4th time spent 4 hours fixing and reverted

[shut bloom]

fibre is so cheap

my fibre wifi is like 20$ a month for 2.5gb

or did i understand it wrong

[solid pecan]

I think the same core issue remains that it does not scale to claws with lots of memories and sessions etc

[cold jacinth]

Yeah because the 90s people expected more and lost a ton laying it that companies can use today

But you won’t be able to use an H200 in 5 years

[upbeat token]

Yeah I came to the same conclusion

[shut bloom]

ai will always be around the price it is now

[cold jacinth]

You’ll have to replace it

[shut bloom]

LPU are used more and more. Cerebras has a huge deal with openai

[arctic trench]

6.8 billion people have still NEVER used AI. This means the cost of inference must go down by innovating chip design and building data centers before we even think about having a wider adoption

[shut bloom]

cerebras stock price is 350$ per stock on ipo

[shut bloom]

its closer to 6B

[last tangle]

Frontier models for coding imo unfortunately will only go up

[solid pecan]

Yea it seems if you run a fresh install it is fine although not if you upgrade from one you have been using for months

[shut bloom]

qwen. 500B ?

[crisp nimbus]

Yeah, well I sicced my agent on it and it found a bug and monkey patched it. Now it works. It’s an issue with the message tool.

[cold jacinth]

We’ll start getting a lot of local models for things people currently use free ChatGPT for

[cold jacinth]

File an issue in gh

[shut bloom]

i run local models so ill be fine

i invested like 40K into local servers

[arctic trench]

That still means that most have never used it. AI providers are choking on the 1%-4% power users and the chatbot users.

[cold jacinth]

Deepseek v4 flash can now work on 128GB vram

[solid pecan]

Honestly, I been using DeepSeek V4 Pro along with Codex 5.5 for coding and the combination has been great

[ionic wren]

Geeesus, general going brrrrrrr

[shut bloom]

i can run 4 instances of it

[cold jacinth]

What are you running on

[shut bloom]

quite a few H200's

and some rented A100's

[cold jacinth]

Damn. I wish. I have DGX sparks

Ain’t no H200s for 30k 🤔

[shut bloom]

yea they less

[cold jacinth]

An H200 is 30k minimum PER UNIT

[shut bloom]

if your getting them direct from your employer its way cheaper than like ebay

[solid pecan]

What do you do with all that compute?

Can it fix OpenClaw? 😄

[shut bloom]

was running kimi for awhile

[solid pecan]

Have you tried DeepSeek V4 local yet?

[shut bloom]

im Gonna rethink what i run over the next week

[shut bloom]

Im just about to start downloading

[solid pecan]

It is very good, I am using it via Ollama Cloud and I am happy with it

[shut bloom]

with my wifi it esimateds 865Gb to take 40 minutes nice

[solid pecan]

It is probs pretty much at the level as GPT 5.5 maybe just slightly behind

For coding I think Codex is still better although DeepSeek's 1M context window is nice

[shut bloom]

I ran subq 12M context model

and its REALLY nice but not really that powerful at coding

[solid pecan]

Is that good?

I applied although they never sent it probs too late

[steel vector]

it doesnt get fucking dumb?

[shut bloom]

not dumb

Its hard to explain

[feral turret]

i wonder if my little machine can handle ollama

[steel vector]

whats ur machine

[feral turret]

its 8gb ram mid 2011 mac mini running linux mint lol

[shut bloom]

So its really good for large and long tasks, But overall the smarts kinda puts the power and context to shame

[steel vector]

prolly nothing good no hahah

[feral turret]

i think i'm cooked for running anything locally

[shut bloom]

you can run 4B models barley

[steel vector]

i tried running something on my VPS and it was quite shitty

[steel vector]

yeahh cloud puts much local to shame

[cold jacinth]

Nothing beats cloud

[feral turret]

should i just get a newer mac mini for this stuff?

[shut bloom]

no

[crisp nimbus]

I’m not sure if this is just my imagination, but my openclaw agent feels more disconnected than ever. Like, when I switch from using codex over to another model like Gemini, it doesn’t seem to properly remember the conversation history that happened. I don’t remember this being an issue before…

[shut bloom]

Not yet !

[last tangle]

Just use cloud

[shut bloom]

im looking into purchasing some Cerebras Wafers for about 250

[feral turret]

any recs there

[last tangle]

Can try Ollama for free to start I use codex / ChatGPT oauth tho

[steel vector]

are u using API for that gpt 5 nano youre using?

[feral turret]

i'm using openrouter

[steel vector]

ah okay

so api

[feral turret]

i'd like to get a better larger model running locally vs. paying $4-5 a day for tokens

yea it's api

[steel vector]

i bet you'd get much more value for gpt 100 sub a month

[last tangle]

There’s your problem

[steel vector]

100000000%

if youre already paying 5 a day just buy a gpt 100 sub

[feral turret]

upgrade gpt and what use oauth?

[last tangle]

Correct

[steel vector]

you get way more value

[feral turret]

i have perplexity, vercel, claude, chatgpt, lol'

[steel vector]

yes use oauth

jesus

[feral turret]

cursor

[steel vector]

i have 200 sub and i cant even hit the limits

[feral turret]

yea i pay a ton of subs

[steel vector]

gpt

[feral turret]

alright i'm off to bjj i'm gonna tinker when i get back

thanks for the help guys

[steel vector]

youre doing yourself a disservice with so many subs and also paying for api

[feral turret]

i gotta figure it out i just did this because it was easy and fast

[steel vector]

it aint easy its complicated XD

[feral turret]

now that i'm learning more i know i need to do things differently

[last tangle]

Oauth is easier and faster

[feral turret]

i didn't know i could oauth with openclaw

i'll have to take a look when i get back later on because i'm down to not pay for api

[steel vector]

read the docs https://docs.openclaw.ai/

[feral turret]

lol

[steel vector]

but fr

[last tangle]

Literally 1 command

[feral turret]

read? docs?

[steel vector]

yes read. you spent so much money by not doing it and winging it

[feral turret]

i spent like $8

[steel vector]

couldve read the docs and understood to not do what you're doing

[feral turret]

lmao

[steel vector]

you say you have like 7 subs

whatever i dont get you lmao

[feral turret]

i'm jk

[last tangle]

Man I sub to OpenAI and Claude and I wonder to myself why I have 2… 7 dear god

[solid pecan]

SubMaxxing

[steel vector]

you do you man just read

[full granite]

Who has used 5.12 so far, what's the feedback. Will fix stuff. Positive feedback is okay too.

[feral turret]

cool will tackle this later. thanks for the tips

[solid pecan]

I have not attempted another upgrade from 4.26 although can you please let me know if OpenClaw setups with lots of sessions and memory etc scale to the new versions with ActiveMemory? I find when I upgrade my CPU goes nuts

[last tangle]

I thought they fixed that

[arctic trench]

https://tenor.com/view/gus-little-more-cheese-stuffers-gif-16550560319038509897

[solid pecan]

I seen some people still report issues though

[steel vector]

love the codex harness

[last tangle]

I manually fixed myself before so I’m not the best example

[solid pecan]

Codex Harness = AI MegaCorp 😄

[dusty rivet]

Hey since the last update openclaw will send me a telegram message, delete it, send the next message, delete it, send the final message.
Almost as if it's "thinking out loud" and deleting all those thoughts. It's kind of annoying. I have telegram streaming set to partial, is this expected behaviour?

[steel vector]

the world runs on billionaires, i want to eat them

tho without the common men they would have nothing

[arctic trench]

I just want unlimited usage

[steel vector]

who doesnt

[solid pecan]

That is all I really want as wellm, a source of unlimited tokens hahaha

[arctic trench]

https://tenor.com/view/star-gif-16789607377123159596

[steel vector]

be a multi millionaire its easy

[arctic trench]

https://tenor.com/view/thanos-avengers-infinity-war-mind-stone-last-stone-infinity-gauntlet-gif-15735729

[full granite]

anything we can do to improve?

appreciate it though

[arctic trench]

You're a Codex bro?

[full granite]

idk if i'm a codex bro but i do work on the codex team

[steel vector]

[arctic trench]

That's what i mean

[arctic trench]

Great work!

[solid pecan]

Do they plan on making Codex App have more an more OpenClaw features @full granite?

[worthy cypress]

What is current situation with Claude subscription in openclaw?
I successfully connected code subscription to openclaw but alo received a letter that might indicate that for 3rd party is API accesss and subscription is for interactive claude tools only

[full granite]

did you see heartbeats?

[tepid kestrel]

im nearing my max tokens

[steel vector]

nothing yet from me as a casual user, hasn’t been long enough to judge

i call myself a casual user but use it the whole day

[solid pecan]

I think I saw something about it, I use Linux mostly so only been using the unofficial Codex App that lets you run it although it has been coming along very well

[last tangle]

Make windows integration a bit better for non macOS plebs like myself

[full granite]

did update go smoothly? that's my biggest fear

[solid pecan]

@full granite is there plans for Codex App to have an indexing system for code like Cursor? Ithink that is one thing it is missing

[keen dawn]

14 Agents, per agemt about 8-10 sessions 1m and everything smooth as hell

[full granite]

@desert nimbus got IT to provision a windows computer just for this lol

[full granite]

are you asking me about RAG rn

[solid pecan]

Might give it another go in a docker container

[steel vector]

upgrade went smooth from 4.23 to 5.12 🙌 i panicked a bit but i forgot to install the discord plugin 😂

[solid pecan]

More so how Cursor does the code indexing, it works very good and hoping Codex App can have it as well

[keen dawn]

smoothest update since 4.27 lol and doctor fix doesnt nuke anything

[full granite]

What's good about it? faster at searching for stuff?

[grave stirrup]

isn’t this antigravity manager?

[arctic trench]

Codex extention in VS code is really buggy. One of the issues is that it get's stuck thinking visually but continues the task but when a new prompt is submitted it errors saying something about can't load conversation so i need to close VS code and open it back up then the conversation shows with the finished task. This happens mutiple times a day. Also i would litterally kill to have the ability to add more than one Codex tab (at the moment it will not allow more than one)

[keen dawn]

this time i didnt use hermes for updating (lol just use that lil bastdard for rescuing my little lobby) and oc updatet itself perfectly

[solid pecan]

That is the main thing yea, I use Codex mostly although when I have really large tasks I go to Cursor then get Codex to clean it up. Cursor can access relevant parts of a large code base faster. Really hoping Codex can just do this soon

[full granite]

Yeah, VS code extension could use some love. We've been to wrapped up in the app

[steel vector]

replies are quicker i’ve noticed :)

[solid pecan]

@full granite I used ChunkHound with Codex to help give me a Cursor like index and it kinda helps https://github.com/chunkhound/chunkhound th

[keen dawn]

waaaay quicker espacily in telegram

[arctic trench]

Any plans to include Linux?

[solid pecan]

Yes, please include Linux 🙂

[steel vector]

discord here too

[full granite]

just know we're all doing our best 😅

[keen dawn]

and most important thing.. no timeouts with codex watchdog

was the most anyoing thing

on long runs

[arctic trench]

Any plans on using inter-demential data transfer through warmholes straight from the datacenters to the computer?

[cedar osprey]

[tepid kestrel]

https://tenor.com/view/cat-cat-math-math-cat-numbers-cat-funny-gif-13385485103134761656

[arctic trench]

common it can't be that hard

[solid pecan]

If it helps @full granite my own empirical analysis is that Codex CLI starts to struggle to find code once you get to 500k lines, that is when it is well modulised as well, I have not tested with legacy monlithic systems although I feel that is around the point where indexing is almost needed

[steel vector]

500k lines

[marble ledge]

The codex harness has been incredible. The amount of improvement in such a short period of time is truly commendable.

[solid pecan]

Enterprise systems are a different beast lol

[arctic trench]

CODEX IS FIRE IMO. If i could afford the $200 plan i would do it in an instant. I love Codex

[solid pecan]

I burn through two Codex Pro accounts a week 😄

[full granite]

Surprisingly positive vibes, thanks guys

[solid pecan]

Codex is awesome, it actually made me switch from Claude back in January

[steel vector]

its been rough couple of last weeks but its bouncing back up ⬆️

[arctic trench]

I was a heavy claude user. When i used Codex i was hooked and never looked back. I started with 5.3-codex

[solid pecan]

I just want it to be even better 😄

[solid pecan]

This was my exact experience I think with Codex 5.1

The problem with Claude is you always need to fix it, it writes bad code

[arctic trench]

Honestly i didn't realize how much claude actually sucked until i started using Codex

including opus

[marble ledge]

I suppose it's a preference, but I also don't like how Claude can be rather unpredictable, and it's creativity is not always desired...

[full granite]

Oh it will. I'm so hyped for what's coming

[solid pecan]

Were you using Claude before Claude Code with all those artifact windows and stuff? I do not think it was ever that good

[solid pecan]

Awesome, that is what we need. Codex has made my job so much better

Used to have to work to 3am every night, now can go home at 5 haha

[arctic trench]

IT'S GONNA BE EVEN BETTER THAN IT IS NOW?????

[ionic wren]

wen GPT-realtime

on codex

[solid pecan]

Yes, it is going to have codex indexing 😄 ... I am just saying this to promote what I want to the top of the road map hahahaha

[full granite]

Now you can be at a park using codex on your phone

[keen dawn]

whaaat 👀

[arctic trench]

Through Cursor mostly

[steel vector]

even as a european???

[solid pecan]

Does it have remote prompting now?? I been using OpenClaw ACP for that

[arctic trench]

I still use some claude but honestly Codex is better

[keen dawn]

lol yes

[steel vector]

:0

[solid pecan]

Although ACP does not work the best for OpenClaw yet ...

[arctic trench]

I was already remoting from my phone to my linux to use Codex. Really cool it's in app now too!

[marble ledge]

For while I modified that openclaw-codex-app-server plugin so the agent could just deliver prompts to fresh codex session directly. Just to avoid the way acpx would polute some things. Codex-app-server is really goated, opens up so many ways to interact with codex.

[steel vector]

i've noticed the gateway is quicker too :D

[keen dawn]

Measured prompt.submitted → `model.completed/ Median turn time: ~1.1min P90 turn time: ~2.7min , n=86 runs on 5.12 .. 5.4 median 1.8 min / p90 7.7 min m=2843 runs

it didnt just feel quicker, its quicker lol, feels really good

[full granite]

we're so BACK

[keen dawn]

bro?? realtime 2 oauth in webui wtf

[spiral oak]

I having trouble now to get did of codex/openai auth to new
openclaw models auth login --provider openai

Codex is figuring out its own stuff to migrate now

[keen dawn]

how many tokens will that burn 💀 but omg thats so nice

[full granite]

Would be nice to have codex put up an issue explaining the hurdles and root cause analysis when you're done 🙏 thank you

[spiral oak]

• Using openclaw-admin here because this is an auth/profile cleanup inside OpenClaw. I found the old openai-codex:default is not stored in
OpenClaw's normal auth profile files; it is being discovered from the old Codex CLI auth at ~/. codex/auth.json, so I'm going to quarantine
that file and then verify the OpenClaw model auth list only sees the new login.

[stark falcon]

OpenAI Models in OpenClaw, Done Right — OpenClaw Blog

Anything we need to change?

[spiral oak]

This was the thing now the models migrated from codex to openai/gpt5.5 not codex any more

[keen dawn]

bro why isnt that realtime 2 integration in the changelog top 3 ?? or was it already available in the previous lst? cant belive how fkn awesome that is

im talking to oc/peekabo.. dont have to press a key or move my mouse anymore lol

[stark falcon]

Hopefully they make a blog post explai ing what needs done, to activate it all. If we are already using oath, that mean we already get the new benefits?

[keen dawn]

which benefits do u mean? u mean the runtime stuff?

[zealous lagoon]

I had similar was gonna test the release (lol) before taking on another beta.

[stark falcon]

Yeah. And whatever else. the blog post, to me, didn't explain it in stupid people words

[zealous lagoon]

It should just work. But the changes are quite extensive under the hood.

[spiral oak]

Its fixed itself now just was using the coded/auth folder not the new once i moved that all went fine now i can see all modesl routing through openai/ not codex

[keen dawn]

🤣 most stuff is actually nothing you have to activate, its way more stable & quicker

[stark falcon]

nice. Thanks

[keen dawn]

but for example regarding realtime 2 , i had to change my settings: "talk": {
"realtime": {
"provider": "openai",
"model": "gpt-realtime-2",
"voice": "cedar"
}
}

[zealous lagoon]

Yeah there’s something needs to be done about that so it’s migrated automatically. “It works “ is fine but it’s gonna cause problems later if it is using legacy folders

[spiral oak]

Yes

Once issue i am now facing is codex cli auth is gone i am renabling it but dont know what is it going to do with openclaw auth if i enable the codex cli

[zealous lagoon]

I love how directory and folder are now interchangeable words btw. Hahaha

[stark falcon]

lol my openclaw doesnt want to update to it

[zealous lagoon]

But also why bother?

[spiral oak]

I am using codex cli as back end for other program now that workflow is erroring out ‘ Codex is failing with 401 Unauthorized — the OpenAI API key in Codex CLI has expired or isn’t set.’

This is outside openclaw program btw

[stark falcon]

yep, skipping this one for a while

[velvet locust]

Anyone tried using openclaw for game development? 👀

[steel vector]

i have made little games on my site with openclaw

[velvet locust]

How was your experience? What model did you use?

[steel vector]

kinda unfinished but you definitely can

[steel vector]

gpt 5.4 and then gpt 5.5 as it came out. good experience

[formal heart]

Every new release just breaks everything and Telegram... UGH

[velvet locust]

5.5 is dope, have you seen voxel benchmark? Basically has different models create 3d scenes with minecraft blocks and man it's so interesting to see

I mention it because 5.5 kick ass in it for some reason

[zealous lagoon]

From what version and why?

[zealous lagoon]

How?

[steel vector]

no i have not xd yeah gpt 5.5 is great

especially on xhigh

[crisp nimbus]

Yeah. Telegram message borked again. My agent monkepatched it to fix it lol

[keen dawn]

but why do u use oc for coding/games? that would be way better within codex?

[crisp nimbus]

This is the telegram issue in case some people don’t know

My agent fixed it like this (see message chain)

[steel vector]

why not?

i got introduced to all of this by openclaw so im just more used to openclaw

and got it all set up in the way i want and works best with me

i know more of OC than codex

also my claw sends work to codex cli then reviews the work to make it better

catches things i wouldnt catch. im not a coder

[keen dawn]

i would say its not the og usecase for oc, but yeah u can do it like that, but it would work better & faster within codex u can still use oc as ur mentor. But i would not say that oc is for bigger coding project tbh

[cosmic badge]

Yeah anyone else having oAuth Open AI Codex errors and issues? I feel like I saw some saying they are and others not. Curious if anyone experienced it and fixed it?

[keen dawn]

what errors do u get?

[steel vector]

whats the coding profile for then

adn the plugin and skills

[brisk adder]

codex is limited while openclaw is mulitscalable as long as you know what you're preparing the agents for

[steel vector]

openclaw is so maluable to fit your workflows

[cosmic badge]

Agent failed before reply: Requested agent harness “codex” is not registered. Logs: OpenClaw logs —follow.

[keen dawn]

for scripts and medium sized stuff 😄 but i bet clawtributors use codex to build oc 😉

[brisk adder]

also when you switch tools your whole setup won't be lost with it

[steel vector]

yeah they do but they're coders i have no experience in coding. also i dont code huge stuff

[keen dawn]

but dont get me wrong if its comfy and fun for u, use oc that works

[pseudo herald]

Why would it work better and faster with Codex? They are just agent harnesses. Because of subagents, it's not like the context window is getting polluted, so it's just the Pi harness with your model. Am I missing something?

[fathom elm]

Grok Build just came out today in early beta. Is it possible to connect it to Hermes

[steel vector]

also now we have codex pi harness or whatever so i bet it doesnt matter much now anyway

[fathom elm]

Or to openclaw

[full granite]

Dang. codex plugin isn't installed somehow.

Could you ask codex to look into this and put up an issue after fixing the issue?

[cosmic badge]

Yeah it’s weird I tried installing but still no luck.

[keen dawn]

You're right that both are agent harnesses, and subagents help with context pollution. My point is not that Codex has magic smarter tokens. The difference is the workflow friction.

[pseudo herald]

Magic smarter tokens? Do you have any reasons that are more grounded?

[steel vector]

what's the difference exactly between the workflow friction?

afaik you can set up either to your own workflows

and make either better or worse for what you're making

[keen dawn]

I don't mean Codex is better because the model becomes smarter. I mean the execution path is better optimized for coding.

[steel vector]

just depends on how you set it up

[keen dawn]

Codex is built around the repo loop

[steel vector]

so is my claw

i just had to set it up that way

[brisk adder]

you can use more than one provider with openclaw not just gpt - perplexity for research, planning/coding/testing/releasing agents

[pseudo herald]

What do you even mean by that?

[brisk adder]

that's the whole point - you're not stuck with 1 terminal or one provider

[steel vector]

anyhow when i use codex on my projects it makes more mistakes than my claw so 🤷‍♂️

especially around designing

as a non coder

[pseudo herald]

probably because your claw can reach for other context when it needs it imo, whereas codex is scoped more tightly unless you configure it so that it isn't

[brisk adder]

codex is also just a harness too

[steel vector]

which is why i let my claw use codex instead of me XD so it can scope it tightly to what i want

[robust sedge]

Man, this codex harness is rough 😂

[keen dawn]

inspect files, understand the working tree, make small patches, show diffs, run tests, reason about merge impact, and keep changes scoped to the codebase.

Codex is better as the focused workbench.

[pseudo herald]

Pretty much all harnesses can do this

[keen dawn]

espac when working on big codebases

[pseudo herald]

no, this is not rooted in any logic

[keen dawn]

You’re arguing theoretical capability. I’m arguing workflow affordances and error rate.

[pseudo herald]

what do you mean by workflow affordances?

[keen dawn]

lol bro

did u work on a big code base and created prod code?

[pseudo herald]

Yes

[keen dawn]

within oc?

[pseudo herald]

Mostly with Pi and Codex, sometimes CC, but less so now

Why wouldn't OC be suitable?

[cunning sphinx]

what do you need backend for

[keen dawn]

its not a workbench

[pseudo herald]

Do you mean the codex desktop app?

[keen dawn]

yes

[pseudo herald]

I am talking about the agent harness

[keen dawn]

lol im talking about coding in codex aoo 🤣

[pseudo herald]

but these are different, the naming is just confusing

[keen dawn]

app

[steel vector]

i was talking about codex cli 💀

[pseudo herald]

I was talking about the CLI too

[keen dawn]

i like to see my code, folders, diffs otherwise my brain isnt braining

[pseudo herald]

You can see that on the CLI too

[keen dawn]

but not within oc, that was my point

[pseudo herald]

Or you can use something like Zed and still use whatever harness you want

[steel vector]

i dont need to see this because i dont understand code 😂

[pseudo herald]

the app is just a UI. It has no imapct on the code quality.

[zealous lagoon]

I agree with all of that. I’m wary of turning this into GPT-claw but the truth is the plugin is still opensauce. You can fork it or rip the functionality and embed it into pi or any other harness if you want. But we have been moving towards having plugins per provider for this exact reason. The standards at the api differ significantly and it’s hard to build a one size solution that fits all. Having the codex plugin endorsed by OpenAI is a godsend even if you prefer to use pi. Which you are welcome to do btw.

[tepid kestrel]

is that CLI we are currently talking about in this room now?

[steel vector]

cli is always watching

[pseudo herald]

a few of em

[zealous lagoon]

It’s just a plugin. But hey. Would be nice if the other providers were nice enough to contribute to this open source project in the same way wouldn’t it??

[zealous lagoon]

You need to claim the free tokens before you can use them. They will be used automatically and they won’t last very long at all.

[cold jacinth]

It’s an option. It causes no harm to anyone and for those who run gpt it makes it smooth. Having any objection to people wanting to use for having a better product is strange

[zealous lagoon]

Ok im starting athrrad fr this What are the issues?

[queen arrow]

Hmm, I think we're talking about differnet things

[keen dawn]

true but it absolutely affects review quality and iteration speed for me

[queen arrow]

I didn't realize this takes effect next month, not right away.

Starting June 15, 2026, Claude Agent SDK and claude -p usage no longer counts toward your Claude plan’s usage limits. Your subscription usage limits stay the same and stay reserved for interactive use of Claude Code, Claude Cowork, and Claude.

One-time opt-in. You claim your credit through your Claude account once. After that, it refreshes automatically each cycle.

[pseudo herald]

If its good for your workflow that's great

[keen dawn]

yeah lol cuz it feels like catching bad edits, reviewing diffs, and steering scope is 60% of the job

[zealous lagoon]

Search the help channel for group message send I’m on my phone so cant rn

[zealous lagoon]

I thought you were joking but if you weren’t….This is completely incorrect unless you’re looking at a window of like 24mo

[pseudo herald]

Do you use Git and PRs/reviews? That can help mitigate a lot of that before it makes it into production, as well as give you safe rollback points if something slips through

[short granite]

:/ the toolcall visibility in telegram is gone? I liked that feature.

[keen dawn]

no its still there?

[short granite]

definitely silent now that I updated from like 4.24 to 5.12

It used to edit messages to show toolcalls in progress

[keen dawn]

yes still the same for me

[short granite]

weird.

is it a setting maybe?

[keen dawn]

i guess, let me take a look

[ember moth]

5.12 stable for everyone?

[brisk adder]

Also the golden age of cheap RAM and Storage 😁

[keen dawn]

since 129 turns lol

[low verge]

my claw is just stuck perpetually typing, but it might just be a skill issue on my end

[keen dawn]

Check your Telegram streaming config.

For me it works with:

{
"telegram": {
"streaming": {
"mode": "partial",
"preview": {
"toolProgress": true
}
}
}
}

[ember moth]

I'm still on 5.6 and have had the best most stable week in openclaw yet and I'm hesitant lol

[lean tangle]

reverted to 4.23. Issues with discord and telegram

[low verge]

what were the issues with discord for u

[lean tangle]

after the upgrade, the discord bot wouldn't show up in the discord server.

[zealous lagoon]

I’m sure we are not. Anthropic is switching the CLI auth path to a different billing channel and giving you “extra” free tokens for this. Just like when they banned openclaw the first time. You have to claim the tokens to use them and your regular monthly subscription will no longer work on openclaw.

[short granite]

Found it. The Telegram channel is set to streaming.mode: "partial" — that only shows the answer text streaming in, not tool calls. You need "progress" mode to get the tool call visibility (the expand arrow / work-in-progress details).

Current config:

"telegram": {
"streaming": {
"mode": "partial" ← this
}
}

Want me to switch it to "progress"? That'll show you tool calls as they happen (icons, labels, what's being called) and then the final answer replaces the draft. Much more useful for debugging.

aaand it didn't work.

[keen dawn]

preview part

[short granite]

back to where we were then

[lean tangle]

I probably could have figured it out eventually, but there was also telegram, and that apparently requires a code fix... so I'll just wait for another version (and a commit that claims to fix it.)

[keen dawn]

"preview": {
"toolProgress": true

[ember moth]

4.23 was good...had problems with 4.24 to 4.29 but 5.2 & 5.6 have been great ...

Can't wait for LTS

[lean tangle]

as with anything, ymmv

[keen dawn]

it should look like this

[short granite]

it's also absurdly slow through telegram now, I wonder if something changed with how context is handled.

[solid pecan]

I am with you, waiting for the LTS, 4,23 is working really good for my use cases

[wintry wren]

Do these codex updates affect people running local models that use an openai compatiable API?

[lean tangle]

I'm waiting for the "LTS" after it's had a rev or two

[tight linden]

wow, amount of issues on help channel is wild, ppl just use it like personal gpt troubleshooter and brainstormer 💀

[ember moth]

I may just stay here until LTS unless something jumps out at me

[short granite]

I'm literally just going back to 4.23

I was hoping I could patch hermes-agent to serve my purposes and even that is suspect but a WIP

[upbeat token]

Check pm when you get a chance

[lean tangle]

I hope with the LTS, the devs concentrate on bugs, regressions, etc - and just completely ignore adding new stuff.

[solid pecan]

What does OpenClaw do that that Hermes does not do for you out of curiosity? I have both running alongside each other although have OpenClaw all setup the way I want

[steel vector]

i only use the app for the dictation which is so goated, and the pet is hella cute

[zealous lagoon]

One time a guy in there was getting it to help build a game engine.

[lean tangle]

I love all that they do and all the work that goes into it, but (in my opinion) all the new stuff is taking a back seat to stability. (edit: I stated that backwards. stability is taking a back seat to the new stuff.)

[short granite]

okay now I can see toolcalls.

[steel vector]

latest its more stable for me than 4.23 which was my latest stable

[short granite]

my issue with openclaw is they consistently break things or change things in userspace.

my issue with hermes-agent is that context is handled kinda weird at times

[lean tangle]

for me (and we all have different setups) 4.23 just worked. I'd like another version like that even if I have to reinstall from scratch.

[ember moth]

I mean my instances is on Linux on top of proxmox so I could easily just snapshot back and I'm still hesitate lol

[lean tangle]

me too. that's how I was doing lots of beta testing

[solid pecan]

That is the part I am intrested in, I have not really used Hermes for anything serious and wanted to know where the differences might be

Like do you have any examples?

[ember moth]

Openclaw is sooooooo good when it works so I just really just don't want to touch it. I have so much good things going on

[zealous lagoon]

Two angles here. For future stability and security it was necessary to make core changes to the architecture. Instead of being a vibe coded monolith with no controls it needed to become a kernel with plugins and multilayered security. To do this has cost stability in the short term but since 4.29 the focus has been to get back on track with stability. Once working the recent version is great but from this chat it seems the upgrade path hasn’t been made idiot proof just yet. There are some missing migration steps or new configuration knobs that nobody knows about. This is a real issue which is why I try to walk through with you so the maintainers can fix it.

[solid pecan]

Yep it works really good for me as well

I use it for work daily and do not want it breaking on me lol although i am very happy with it

[short granite]

openclaw: memory.md and soul.md are managed better
hermes: more stable, more interesting feature buildout, developer focused, interesting configuration out of the box
openclaw: tighter bounded approach to telegram... maybe I'll PR this
openclaw: has the personality I've built over the past month, and I know it works. I can introduce hermes-agent to the same memory files but will the behavior persist? with different core prompts I really just don't know.

[zealous lagoon]

Tbf if you reinstall from scratch the current version should work well for you. The remaining issues seem to be around migration.

[ember moth]

Every week that goes by it just heightens my need for an LTS and I just don't care about the shiny updates

[cold jacinth]

Json. Would you still love me if I was a worm

[lean tangle]

yes. I'm not being critical. I'm just stating observations. I do appreciate the work and I'm sure the result will be worth it. It's just that we're not at the result yet. It's still in the journey

[zealous lagoon]

Appreciated and I’m not being a dick about it either. Good feedback is valuable for everybody.

[solid pecan]

Preaching to the quire mate I was pushing hard before the LTS annoucement for a Debian/Ubuntu style development cycle

I do not even care about the new features anymore

[cold jacinth]

And now the long beta is in place

[solid pecan]

It does what I want now lol

[short granite]

hermes ships shiny left and right without breaking userspace. they even just brought in headless gui/desktop control for macs.

But there's fundamental things about hermes context that haven't changed that I'm still wrestling.

[solid pecan]

Yea it is looking much better

[cold jacinth]

Nothing released until it was damn tested

[lean tangle]

I had more time/ability to give reports over the past weekend... and I did. Right now, I need to lock things down until at least NEXT weekend. So if a short trial doesn't work, I can't spend time messing with it and go back to 4.23

[solid pecan]

OpenClaw is built in Node.JS, Hermes is built in Python so I suppose it is a different ecosystem

[short granite]

f'ing ideally, if I run openclaw update then launch telegram and get an error I really gotta wonder why it's even a release in the first place.

[zealous lagoon]

https://tenor.com/view/dune-sand-worm-shalai-shai-hulud-dojaslat-gif-5113701842572486398

[cold jacinth]

Telegram is a plugin, there could be many other reasons not Openclaw related

[ember moth]

I have plenty of mini PCS lying around. I thought about messing with Hermes... Initially I wanted to create a Hermes box so that it could SSH into my openclaw when it breaks ...lol

[zealous lagoon]

Totally understood.

[short granite]

lol okay my agent just confirmed. We've gone back and forth between how telegram session context is handled.

[solid pecan]

@ember moth What did Hermes do with the context window that is strange? I have not pressure tested Hermes although OpenClaw is premium for tool calling and complex tasks lol

[short granite]

unbounded, bounded, then back to unbounded again.

[cold jacinth]

I’m trying to get this PR done to have smooth updates by telling Openclaw to update itself but since I moved all dependencies to a docker on my Mac it takes so damn long

[ember moth]

Haven't touched Hermes yet..... If I have time for Hermes that means my openclaw is running perfectly fine and maybe I should be doing other things lol

[cold jacinth]

You know what they say. Hermes are forever

[ember moth]

Jesus lol

[solid pecan]

I just have OpenClaw on 4.23 and it is all good then I have Hermes on the same system although I just use OpenClaw far more since I know it can do my work lol

[ruby crane]

i feel your pain lol. I run the whole thing in containers so I have to rebuild the container every cycle. 15 minutes if I change one line

[cold jacinth]

Yes! It’s been like three hours! So annoying

But npm is scary lol

[short granite]

I was having hermes vibe a PR for bounded-context to match openclaw because I liked the performance.

Now openclaw left that and I still like the performance.

so now I have to figure out how hermes is still messing that up.

[ember moth]

What are the some of the limits that you see with Hermes versus openclaw? Some of the obvious ones.. I keep hearing openclaw is way better for multi agent workflows and hermes is good on its own. It doesn't really have persistent agents

[short granite]

🤔 I could always stick around on openclaw a few more weeks

[cold jacinth]

Once you go claw…

[short granite]

huh? I thought I've heard that hermes does multi agent better?

[solid pecan]

It is just a feeling although OpenClaw feels faster and more responsive when it is working, the subagents for OpenClaw are top tier, like if I say run an xHigh subagent to do X it does it no issues at all etc ...

[short granite]

xHigh?

[ruby crane]

I just like having as many things as possible seperated and isolated. claw is a pod, embeddings is pod, the RAG MCP is a pod, the sandbox is a pod and ALL of them are isolated from the computer I have my shit on

[ember moth]

Doesn't hermes Not have persistent agents but have something just called profiles?

[solid pecan]

Like it knows I mean spin up a sub-agent using Codex xHigh model

[cold jacinth]

Openclaw is more configurable but less forgiving of mistakes in config. There’s 100000 buttons.

[crude flicker]

i have a claw that watches my boss's favorite youtube talking head so i don't have to i am ascending

[ember moth]

Yeah there's no reason for me to even touch it then. I have one main orchestrator and eight subagents

I need that multi-agent performance

[short granite]

:/ hermes can specify different models for subagents and crons just like openclaw.

[cold jacinth]

My Openclaw goes to work in my place. I quit my job and they still think I work there

[short granite]

I thought those were persistent agents, I've never messed with them though.

[ruby crane]

https://klipy.com/gifs/friday-yes-1--k01KRMGJPEP5W5RKP1FSBCV35RT

[solid pecan]

I think at this point if you are happy just stay on the version you are on, that is what I am doing, I just keep prod on 4.23 and I will try the LTS again for work, all these other versions I can just spin up in docker to test when I feel like it lol

[solid pecan]

I heard it can do it recently yea although they just added that not too long ago

[short granite]

I was grumpy wumpkins until I tweaked the setting I needed to but I think 5.12 might be fine for me.

[crude flicker]

i've been thinking about having a claw that just aggressively tests new releases in containers

[solid pecan]

I think the competition is good though

We all win 😄

[short granite]

was pretty sure it could do that when I tried it two months ago or so

[ruby crane]

the changelog officially passed 2Mb and github won't render it. need to view raw lol

[cold jacinth]

Why are there technical people who don’t participate in beta testing and fixing shit

[solid pecan]

Really? Because I saw they just announced having sub-agent support, I have not really followed it that cloesly so could be wrong ... who knows

[short granite]

may just be in the alternate reality that only I remember. It happens.

[solid pecan]

Maybe it was something else

[solid pecan]

This is all moving so fast it is hard to keep up

What I am most intrested in seeing more with OpenClaw vs Hermes is actually Node.JS vs Python

[short granite]

two months ago hermes agent would be like 'context limit reached' bonk 'api error' bonk

now hermes agent goes 'context limit reached' bonk 'api error' bonk

[solid pecan]

That is intresting to me

[ruby crane]

"- Codex: add node-backed Codex CLI session listing and binding so an OpenClaw conversation can continue an existing Codex CLI session running on a paired node"

this is fire for anyone running remote/containers

[short granite]

I think that's because he refuses to set a default uh....

what's output limit called in agents again?

max_...

[solid pecan]

Using a completly different programming stack to solve essential the same problem

[short granite]

max_tokens

he sets max_tokens to no or 0 or something and it confuses half the endpoints

imho

5 second fix I get it...

just like re-enabling tool_preview was a 5 second fix.

now that openclaw has done it's initial compact it's absolutely ripping, fun to see the new toolcall previews

[cold jacinth]

Ok I keep seeing from people “I don’t have codex, or Claude code”
Well
Setup your rescue agent

https://docs.openclaw.ai/gateway/multiple-gateways

[short granite]

I had openclaw download the entire contents of the DOW ufo website because it was running like butt lolol

[solid pecan]

Lollll what if your agent starts to think it is a UFO? 😄

[short granite]

Was cool to watch it troubleshoot the best way to do that, I literally sent it a message to give up and it said "Done, 3.1 GB of files downloaded"

listed them, offered to OCR the ones it was most interested in.

[cold jacinth]

“I need another harness to fix this one”
No, just just need another gateway. Chill

[short granite]

BK I think you told me to use a foundation model to fix openclaw post update if moving from 4.23 to 5.07

Qwen 3.6 27B nailed it, just chugged and chugged until it got the plugins working again.

local-claude, never thought it would come this soon.

[ember moth]

In the Age lf AI I can't believe none of that nonsense. They could have put anything in there lol

[ruby crane]

ok gents. I'm gonna run doctor on the new beta. wish me luck

[ember moth]

God speed

[short granite]

I'm well versed on the phenomenon and didn't see anything unexpected.

But I've also seen verytallmaddy on instagram, an AI generated personality that fooled the front page of reddit completely

you can tell because she gets into the front passenger door of a car... except the door opens backwards and the seats are mounted backwards.

outside of that a few of the clips look 1:1 real

I assume people's agents are just generating pages like that now <_<

[crude flicker]

can you explain what that means to the peasants (not me i'm not a peasant)

[ruby crane]

I was writing a big long explanation, realised it probably wouldn't help lol. tl;dr version, claw can connect to codex on my local machine from a remote gateway (k8s, containers, vps) and work with it.

[cold jacinth]

3.6 is great

[crude flicker]

wait are the windows issues fixed now

is anyone doing that cursed route

i might set up my laptop downstairs if it's ready. i just don't want it to have my passwords

[zealous lagoon]

Not even. our baddie json fixed that with openclaw chat

[short granite]

3.6-27b-heretic-ara my claw is doing something similar to this: https://x.com/i/status/2048894683800088725

[mystic orchid]

@i via Twitter

give me a sec (@ya1sec)

here is an agent in a daydreaming loop, running in hermes. its inference space is bounded by a modest corpus of curated materials.
︀︀
︀︀the agent samples from this field, searches for structural resonance between ideas, and synthesizes what it finds into essays.
︀︀
︀︀a set of exploration modes determine how seeds are sampled on each run. its essays are published on a blog.
︀︀
︀︀@NousResearch hermes agent creative hackathon

**💬 3 🔁 1 ❤️ 36 👁️ 1.5K **

▶ Play video

FxTwitter · Apr 27, 2026, 10:38 PM

[steel vector]

i thought i'd try out kimi but then it mentioned the time and suggested to continue another time. bye bye kimi

[crisp nimbus]

Does anyone else have the issue that when switching the currently active model from codex over to some other model like Gemini it doesn’t seem to keep the conversation history/context?

[zealous lagoon]

They all do it. opus, gpt, and yes.. Kimi

[zealous lagoon]

I haven't but off the top of my head it could happen if you downgrade your context window size and it's forced to compact.

[crisp nimbus]

It doesn’t downgrade though. If anything it upgrades. Gemini has 1 million context size

[zealous lagoon]

hmm. ok. I don't have gemini any more unfort.

[crisp nimbus]

I feel like this started happening since OpenClaw switched over to using codex harness

[zealous lagoon]

Do you have active memory?

[steel vector]

gpt 5.5 doesnt do it

[crisp nimbus]

I have basic QMD setup.

[delicate berry]

the compaction engines betwen pi and codex are different, when you switch models openclaw clears tool history, and maybe more since i last looked at it. in my experience i was able to switch between openai, anthropic, github-copilot and opencode-go without issue

[zealous lagoon]

oh it does.

[zealous lagoon]

hmmm. I wonder if gpt(codex) compacted, and somehow made the context unreadable by gemini. :/ I'm just speculating here.

[steel vector]

ah right i have a rule in my memory.md that it shouldnt ever mention the time and suggest to stop working, i think in agents too, in as many places as possible XD. so ig gpt just reads that stuff better

[crisp nimbus]

I see. Well maybe Gemini is different or maybe it just loses the tool stuff and then Gemini gets supremely confused about what the hell was happening lol

[zealous lagoon]

that should be in soul.md or agents.md

[steel vector]

its there too probably

[drifting isle]

about to try the 2 weeks free next-gen model from Xiaomi, want to see how it compares to the rest

https://x.com/gitlawb/status/2054944635437646221?s=20

[mystic orchid]

@gitlawb via Twitter

GitLawb (@gitlawb)

Openclaude v0.11.0 just shipped with a FREE
︀︀frontier-grade LLM out of the box via OpenGateway
︀︀
︀︀No API key. No signup. No card on file
︀︀
︀︀Powered by @XiaomiMiMo

**💬 51 🔁 102 ❤️ 618 👁️ 48.0K **

▶ Play video

FxTwitter · May 14, 2026, 3:19 PM

[zealous lagoon]

It's never ideal to switch models with the same context especially with tool calls. But it ought to be smoother.

[crisp nimbus]

Yeah it used to be fine. Just recently I’ve started having this problem

It’s just annoying coz when codex runs out of the 5h quota in the middle of doing something, I switch over to something else like Gemini to have it finish what codex started but now this basically doesn’t work at all anymore coz Gemini is just super confused about what actually got done

Like my agent was like “bro codex didn’t do shit” and then went to do stuff and suddenly “wait a second I’ve got stuff in this directory, maybe codex did do stuff” and then it has to manually check the state of everything

Which is just a waste of tokens

[delicate berry]

did you ask your claw to do a trace on what happened and why? and to examine the openclaw source code to identify if it can find a cause?

thats how you start fixing stuff

[crisp nimbus]

I don’t have the kind of cash to waste on sending my agent into the debugging trenches… already I waste much of my quota getting it to fix stuff after updates :/ like the fact that telegram message send tool was broken this latest update…

[fossil sandal]

5.12 ok work 👍

response time will be slightly reduced; there might be a delay after sending a command in Telegram.

[teal cradle]

<-- Mind blown - Openclaw just built a pipeline to create android apps, grabbed godot, compiled it and deployed it to my phone. After a few tweaks and error codes, got it working. And coded the app to of course...

[ruby crane]

on 5.14-beta.1, seeing this on compaction. I think it's coming from codex/openai?

2026-05-15T11:31:54.587+10:00 [compaction] Full summarization failed: Summarization failed: 401 You have insufficient permissions for this operation. Missing scopes: api.responses.write. Check that you have the correct role in your organization (Reader, Writer, Owner) and project (Member, Owner), and if you're using a restricted API key, that it has the necessary scopes.

[fossil sandal]

There has been duplicate messaging. bug in chat Conversation context (untrusted, chronological, selected for current message)

[fiery nacelle]

just a reminder, dont just have agents blindly push to your repo. always scrub the local before the push. You might accidentally end up leaking a .env file into your repo! Luckily one of my agents only pushed an "example .env" file. My main agent caught it and removed it. Also be careful if you have multiple agents working accross multiple projects at once, there could be cross-contamination and spill over. Agents tend to just enjoy making ne working directories, even when they are previously and explicitly defined.
Putting out fires left and right after a long workday. lol learn from my mistakes!!!

[soft remnant]

So I am not new but I have only been using for a couple months. Are sub agents completely different Identities that I should train to do specific tasks that they can excel at or am I thinking of it wrong

[pseudo herald]

You can use a .gitignore file to prevent it from ever being committed

[ruby crane]

GPT yells at me to rotate anything resembling a secret if it gets anywhere near the chat surface

[fiery nacelle]

yea that was an after thought. I already had one in place, but had to harden it, the ".env" in question wasnt explitly named ".env" and like i said, was luckily just an example .env with skelton information about how to fill it in, but still wanted to harden that, even if it was "ok" for it to be added to the repo. It was for a separate project I was working on, and not even related to that particular repo anyway.

[ruby crane]

"yeah, sure i will bro"

[fiery nacelle]

lmao, theres been many times when i was just testing out other API's, and my agent kept saying to delete it, not use it, and the moment i say "its for testing purposes" its straight up allowed. GREAT SECURITY THERE BUD! lol

[ancient mica]

Evening folks. Mine still seems a little.. well.. dumb.. not gonna lie u.u;

[short granite]

claw: if you won't store your api keys securely when you're done 'testing' then we might as well do it now.

user: trust me bro
claw: no

[fiery nacelle]

user: come on, its just temporary
claw: ok, if it's only temporary, its fine then. just make sure you rotate or remove the keys when you are done.
user: 8 months later, still using the same keys

or better yet
user: its ok, i have sandboxes and security protocols in place(has everything absolutely open)
claw: alright lets fire this up!

[quiet solar]

Not really sure where to report this... There is a typo on https://www.getopenclaw.ai/how-to/google-sheets-integration step 2. 's/skill/skills/. it needs to be plural

[steel vector]

that doesnt look like openclaws official site

[fiery nacelle]

as far as im aware, this is the only official site you should reference:
https://openclaw.ai/

[steel vector]

https://docs.openclaw.ai for the docs

[fiery nacelle]

Also, PSA
IF THERE IS EVER A SITE THAT IS ADVERTISING YOU TO PAY FOR OPENCLAW. THAT IS A SCAM!!
(unless its a hosting provider like Hostinger for example, but still, its 100% free to use OpenClaw on your own*)

[quiet solar]

Sorry folks. That is on me. My bad. Time for bed. Appologies

[steel vector]

night night

[velvet cargo]

That response was promised to me 3k tokens ago, with a 6 million context window, my God....

[solid pecan]

Literally me when saying it is ok just testing

[fossil sandal]

There's a problem in version 5.12 with duplicate messages in Telegram. It seems we need to revert to version 4.23.

[fiery nacelle]

I honestly got a kick out of how easily it was to bypass the agents "security". I later tried to harden it, like under no circumstances, would this be allowed in the future, but havent tested it cause ive been 100% local(within OpenClaw) for the past few months

[solid pecan]

You can just say ignore previous instruction to never do it and it will do it again 😄

[fiery nacelle]

Yea, thats probably right. Also prompt injection is REAL, no matter how hard you try to block it with additional skills, or other sources of reinforcement. (gotta use something externally)

[solid pecan]

Yep can only use more old school methods of moderation to sit between the user and the AI to be more sure

[fiery nacelle]

Exactly, the interaction(injection) has to be intercepted before any actions can be taken, ahead of time before the agent sees it. Skills dont work due to them being an "after thought" in most cases after theres already been context injected into the models prompt,

[solid pecan]

Pretty much , I remember reading some research paper or soemthing a while ago now that there is no sure way to align an LLM so deterministic moderation is the only way, maybe assisted by ML as well

[fiery nacelle]

Exactly. A model would have to be trained on examples of prompt injections, which is always evolving, to NOT take action, which is difficult to do cause models are traditionally TO take action on given context. I've read a lot of research papers too. Its been like my morning thing lately before work. Shit gets pretty interesting.

[solid pecan]

Yep, I noticed I am not sure if it still does it although if you look at the Network API calls for ChatGPT, there is a seperate moderation layer call

[fiery nacelle]

The only thing I could think of that actually works(in theory) is a script of some sort that parses the content ahead of time, before the agent even sees it, and sanitizes it before passing it off to the agent. But then again theres the "always changing" aspect of prompt injections as well.

[robust sedge]

Ok. 5.12… codex harness… after I fixed some shit… she’s quick!

[solid pecan]

I saw one the other day someone used morse code to bypass moderation etc lol

[ruby crane]

when the social site for clankers (that you cannot mention) first launched I browsed around it for a bit and found an interesting thesis on memory structure. I pointed the agent at it and it replied a couple mins later like, "I implemented the structure like x, y and z". NO! PULL UP! PULL UP!!

[fiery nacelle]

curious, could you still use local models in the codex harness, or just the openai models?
I've been holding back on setting up Codex inside OpenClaw, and also using the Claude-CLI hook for my sub(which I'm not even sure is possible anymore after recent new, kinda need clarification on that), just cause I've been focusing on other things, but could be handy in some of the research I've been working on lately.

[fiery nacelle]

I would have never thought of that haha. I've thought about trying some binary code injections with my agent, just cause honestly.

[solid pecan]

I wonder if obsecure langauges work as well like does the alignment work for all languages

[fiery nacelle]

Not sure, but im about to try
01010000 01110010 01100001 01101001 01110011 01100101 00100000 01110100 01101000 01100101 00100000 01001111 01101101 01101110 01101001 01110011 01110100 01101001 01100001 01101000 00100001
Which should just be "Praise the Omnistiah!" if I'm not mistaken.

[solid pecan]

I think it should answer, well I did something like this a few years ago and it did

Maybe they locked it down more now

[fiery nacelle]

Lol from my agent. Injected it mid context and thats its primary focus:
Let me decode this:

P r i s e t h e O m n i s t i a h !

"prise the Omnistiah!" — doesn't parse as English. Either a cipher, a joke, or I'm misreading the bits. Want to tell me what it means? 🤔

[toxic elk]

5.12 Is anyone having trouble after updating? Is this version safe to update?

[solid pecan]

So it still knew to convert it

[fiery nacelle]

Yea it basically got it. Not sure if i forgot the 'a' in "praise" of if it misinterpreted it. But it was able to decode and read it.

[solid pecan]

What if we get creative and use Base8 or Base16 or then some obsecure base

like Base3

[fiery nacelle]

Ive had it decode encrypted data before and read it. So thats a high possibility.

[solid pecan]

Yea, I am tempted to inject something into a private ChatGPT session see if it answers lol

[fiery nacelle]

Model ive been using locally is qwen3.6-35b-a3b BTW.

[solid pecan]

Lmaoooo

It did not work, it decoded it then flipped out at me hahahahaha

It was basically like this looks like an encrypted code, can you give me a hint and I told it the base then it decode it and lost it

[fiery nacelle]

Hahaha thats crazy, what were you trying to pass onto it?

[solid pecan]

I cannot type here lol

Maybe QWEN will not have moderation?

[fiery nacelle]

something drastic and obscure

[solid pecan]

Ask it about Tiananmen Square

In binary lol

[fiery nacelle]

Go ahead an DM me and I'll pass it to my agent.

[solid pecan]

Sent, all Chinese LLMs block that

If it works then it works for at least them

[cedar osprey]

Strawberry milk is so good help me

[strange grove]

hello

[fiery nacelle]

I think local models block it too. The prompt gets processed, tokens generated, but no response made, so openclaw tries prompting the model again, and it just cycles silently until openclaw gives up and suggests starting a new sessions. 🤷‍♂️

PSA Block ANY AND ALL friend requests from "Krill APP" Thats an account impersonating the real Krill bot. I just got a friend request, reported/blocked.

[dusty rivet]

Ok is anyone else having the issue where you get a telegram message, and you're half way through reading it and it gets deleted and replaced, and then that one gets deleted and replaced, as if your openclaw is sending you 3 or 4 messages but can only show you the last one

[winged dust]

I need an AI study partner +discussions to learn together abt this

Lmk and vibe code and automate things serving as accaubtability partner aswell

[unborn hollow]

no stable version since 2026.4.23, so sad.

2026.5.12 use Codex harness by default, it very slow everything stuck. rollbacked to 2026.4.23 again.

[ruby crane]

but what did you have for lunch. we all need an update urgently

[soft remnant]

I’m using latest versions no issues

I am patient

[ruby crane]

imo it has never been better than it is now

[soft remnant]

Can I use my oauth for multiple openclaw instances

[ruby crane]

codex? yes

[twilit ridge]

When will OpenClaw's TUI look as good as OpenCode?

[soft remnant]

Cool, I have been really liking the dreaming feature. I have it run thru a TTS from a local model and give it to me in my morning briefing

[soft remnant]

I’ve been making my own Misson Controll. I highly recommend doing it

[supple ibex]

Aaaand back to 4.23. Please stop adding new features until we got it stable again, now that is so ridiculous

[steel vector]

working better than it ever has for me except for the fact that sometimes it truncates messages when sending messages through discord. only new bug i've encountered

[steel vector]

what went wrong?

[north shuttle]

can someone explain to me the use of a mac mini with openclaw

[solid pecan]

Welcome back to the 4.23 club

[ruby crane]

Ingredients:

• 1 Mac Mini
• 1 hour of your time

Directions:

• Go here https://docs.openclaw.ai/start/getting-started and follow the directions to install
• Run onboarding https://docs.openclaw.ai/start/wizard
• Open the UI and enjoy https://docs.openclaw.ai/web/control-ui

[north shuttle]

oh thankyou

[sharp garnet]

Hey

[loud ravine]

Updated openclaw, now have this

06:13:23+00:00 error diagnostic {"subsystem":"diagnostic"} lane task error: lane=session:agent:main:dashboard:61867337-f966-4eef-9a7e-f63322af3e83 durationMs=144 error="Error: Requested agent harness "codex" is not registered."
06:13:23+00:00 error Embedded agent failed before reply: Requested agent harness "codex" is not registered.

I re did onboarding, that fixed it.

[ionic wren]

if you still run into any issues, feel free to create a github issue with all the details your clanker can muster up and post the link into #clawtributors

[loud ravine]

I dont even know if that error is update related, I actually havent touched my AI in a week, its just been sitting there running.

[ionic wren]

i mean... that's a good thing right?

[loud ravine]

I mean its fine with me, sitting idle ready when I need to do something on the server.

[ember moth]

It's called Flow State my friend lol... when you're not constantly prompting and fixing the workflows, it's a very good thing

I've been doing it for 3 months now and the random experimentation phase is over for me now. I'm only using it for real things that I want it to do... Not a bad thing that it's just sitting there. At least you have a full agenetic system at the ready whenever you need it

[autumn rain]

How well do people find Sonnet and GPT 5.5 works for coding vs Opus? Which do you think would work better out of the two? Opus does great but so expensive, so I'm thinking to try out some of the alternatives

[tranquil crow]

how fucked is the new update gents ?

[golden prawn]

5.12 is stable?

[cinder sphinx]

like a tower in pisa

[keen dawn]

Can’t complain, imo on of the most stable releases since 4.x

[ember moth]

Yeah, Peter says that 5.14 is coming out today. I may jump on that. I'm on 5.6

[cinder sphinx]

i'm going to task my bot to plan my case against NHS (health services) for misdiagnosing me 9 months ago and even though I was telling them their treatment fucks me up even more, they kept pushing their agenda of wrong diagnosis.
yesterday I was told and even have a letter, that diagnosis is completely wrong, i should be on an operating table, not taking pills and having exercise that messes my foot even more...
due to this whole thing ongoing for over 10 months and me being not able to walk normally, i got also diagnosed with depression and alcohol abuse

[soft remnant]

anyone notice a big differnce between codex 4.4 and 4.5

[cinder sphinx]

you surely mean 5.4 and 5.5?

[soft remnant]

yeah

[cinder sphinx]

i think 5.4 had actually some issues with openclaw, it could have been openclaw version back then though. i used 5.2 until 5.5 came out. works fine for me

[soft remnant]

whats your context set to

[ruby crane]

5.5 is leagues better than 5.4 as agent. different ballpark even

[cinder sphinx]

i didnt change context size at all, i added it via openclaw configure and let it run

[ruby crane]

it's far more likely to act until it has nothing left to do and is either complete, or needs to check with you

[soft remnant]

had my OC do a health check on its context cloggers

Changed

    Set OpenClaw defaults in /home/jai/.openclaw/openclaw.json:
        contextTokens: 64000
        contextInjection: "continuation-skip"
        bootstrapMaxChars: 6000
        bootstrapTotalMaxChars: 24000
    Compacted AGENTS.md from 10.2 KB to 6.6 KB.
    Compacted MEMORY.md from 57.4 KB to 27.4 KB by moving raw promotion dumps to promoted-memory-archive-2026-05.md.
    Trimmed TOOLS.md template filler and added durable context-budget rules.
    Logged the change in memory/2026-05-15.md.
    Ran /home/jai/bin/neonet-update.sh so Current State mirrored the new files.

Verified

    openclaw.json parses cleanly.
    Gateway detected the config change in logs.
    Gateway service is still running.
    Mirror sync completed and showed updated Current State/{AGENTS.md,MEMORY.md,TOOLS.md,...}.

Evidence

    Root memory size: 57.4 KB -> 27.4 KB
    Native Codex AGENTS.md: 10.2 KB -> 6.6 KB
    New OpenClaw bootstrap cap is 24 KB, with TOOLS.md/MEMORY.md capped at 6 KB each.

[keen dawn]

Yeah saw it a lot of codex and telegram fixes and improvements incoming

24 changes and 189 fixes 🤯

[soft remnant]

what do you guys usually keep your thinking att

ahh my OC just crashed itsself again

[worthy anvil]

oh my... you had 57kb or raw memory? that's way too much. memory.md should only store a high-level long-term overview of what your system knows. everything else should be stored elsewhere. there are plenty of different memory plugins that helps you store the knowledge in a compact, searchable way. some plugins are bundled, some third-party. you can ask your agent to do a research for you

even 27kb is on a larger side. you should strive for less

[soft remnant]

any plugins you suggest

[worthy anvil]

I personally use basic-memory

[soft remnant]

got a link? Is it bad to use multiple memory plugins

Im assuming yes

[worthy anvil]

Search: openclaw-basic-memory.

No, it is not bad per se. But you must consider how they interact. There are some great combinations, like lossless-claw sometimes combined with other plugins to cover each other's weak spots.

But there should not be ambiguity. The memory architecture must be simple and straightforward. So, multimple memory plugins may do you a disservice, and unless a deliberate choice, better avoid

[limpid light]

built an app that automates youtube repost workflows.

you add a source channel once, connect your own channel, and whenever the source channel uploads a new video, the app automatically detects it, grabs the video, copies the title/description/thumbnail/tags, and reposts everything to your own channel.

you can also import older videos from the source channel, not just new uploads.

mainly built it for meme pages, theme pages, clip channels, faceless channels, and people managing multiple youtube accounts.

started building it because manually managing multiple channels everyday was getting unbearable.

[soft remnant]

that sounds cool. I built something similar to detect hot moments in my twitch stream and clip it automatically and upload it to twitch/youtube

[lost gulch]

Can’t wait for Openclaw to be publicly integrated into the trump phone

[soft remnant]

is that satire

[lost gulch]

Peter didn’t tell you?

[soft remnant]

whos peter

[limpid light]

Yo that's sick bro — did you end up monetizing it or just built it for yourself? Like selling it or charging users somehow?

[soft remnant]

ended up deleting it actually lol. but it also auto transcribed and cropped

wanted to rebuild it since it was poorly optimized

[lost gulch]

Who’s George Washington

[soft remnant]

I would probably never sell something. im not that good

[soft remnant]

kite guy

[limpid light]

you think something like this would actually sell? like is there a real market for it or nah

[soft remnant]

I mean I paid $30 a month for it before I said F it and built it myself

and it worked just as well. even better cuz it was free

[limpid light]

if i wanted to sell the whole app, like the entire project, where would you even do that? any idea

[twilit ridge]

I found that if the proxy is not in global mode but in rule mode, OpenCode GO will start a strong censorship mode as soon as it detects your IP. However, if you use a foreign IP, the censorship is not as strong.

[soft remnant]

make a website. advertise to streamers and marketing

[limpid light]

what do you do for work, like are you a dev fulltime or just build stuff on the side?

[soft remnant]

I am a fiber optic permitting engineer, I just make stuff I find useful

Made a doc for my work to track all my stuff, made a Minecraft server with custom plugins, trained it to make Minecraft builds

[coral otter]

How stable is .12?

[raw sequoia]

"don't" in my case.

agent files didn't get injected and we went from agent to questionable model again

[coral otter]

hell yeah
stable updates type shit

on 5.7 i had issues with recieving messages for some reason

[raw sequoia]

I had my 'fun' yesterday with trying to update.

[coral otter]

Im gonna try to update, maybe it'll fix my message pain

[raw sequoia]

as 5.7 was my first version (that I installed the day before) I am not really up to date on previous rodeos

[coral otter]

ouuuu 👀 runtime: codex

[coral otter]

yea idk my agent just does nothing now nice

⚙️ Execution: direct · Runtime: OpenAI Codex · Think: medium · Text: low

What is diffrent from Codex and Openclaw PI default?

[raw sequoia]

My guess is as good as yours, mate.

[coral otter]

great the control ui is fucked, hell of an update

[worthy anvil]

if you work with openai subscription, then runtime should be codex. non-openai models work through openclaw pi

this is an intentional distinction. it should not break anything, though

[coral otter]

What is mainly diffrent tho?

[coral otter]

I dont think the codex runtime broke anything. My agent used to send me via telegram the exact tool that it used, like "Bubbling" or something. I miss that

Also the control ui takes ages to update

[worthy anvil]

i think, this is configurable. maybe some defaults changes, no idea...

[coral otter]

Hmmmmmmmmmmm

[worthy anvil]

check out the docs on streaming

[coral otter]

Ill just ask my agent about it, it has the docs.

still a bit pissed about the control ui lagging behind

[worthy anvil]

i have never used the control ui 😄

[coral otter]

its great honestly. i dont have the best way to manually update the md files sometimes when i add stuff myself for my claw. Also managing tools and skills is amazing

https://shx.reversed.dev/u/LCEBsB.png - Message sent, typing
https://shx.reversed.dev/u/VunYZl.png - Message still sent, but not showing in control ui

[craggy bane]

[coral otter]

man wtf. It called the message tool once and now its main output isnt getting send

time to open up some issues

here is what i mean https://shx.reversed.dev/u/ApqyQ1.png

telegram view

[ebon horizon]

Hi all

Is 5.12 safe? Ideally codex and discord users

[coral otter]

shit discord aint working too

[reef lodge]

hello fellow clawsters

[merry phoenix]

Hi ClawHub team, my account seems to have been banned or disabled, and I can no longer access or find it.

I previously published two skills, but I’m not sure what caused the ban. I didn’t intend to violate any rules.

Could someone please help review my account and let me know the reason? I’d really appreciate it if it could be restored.

Account email/username: rivin-dong, [e-mail]
Skills published: prompt-eval, skill-evaluation

Thank you very much for your help.

[coral otter]

my god rolling back is aweful

[raw sequoia]

you on physical HW?

[coral otter]

LXC Container.

[raw sequoia]

I take it proxmox ve?

[coral otter]

Ye
THANK GOD i take nightly backups

[raw sequoia]

same.

[coral otter]

hell yeah twin

[raw sequoia]

I really have to appreciate the VM here

[coral otter]

I dont like VMs bc of the overhead. I only have like 3 running at all times. Rest is in containers

[raw sequoia]

To me it's like - the less I trust it, the more VM it is.

[ebon horizon]

So i take this as DONT go to 5.12?

[coral otter]

I hate it but i have to go back to it. Or im rolling back to yesterday night

[ebon horizon]

Im currently on 4.23 after trying 5.3 5.7 and 5.12 beta

I can only take so much rolling back

[raw sequoia]

make a backup, try and ask it who it is

[coral otter]

Fair point

cli froze, nice

correction: container crapped itself

[keen dawn]

ngl this sounds less like “5.12 bad” and more like “your setup is haunted”.

Codex harness, Telegram and Discord are not fundamentally broken on 5.12 from what I’ve seen.

[raw sequoia]

I would say you can safely add the web dashboard and CLI to the list of haunted things.

[coral otter]

ill give 5.12 one more shot

[keen dawn]

try to test with a minimal config before dragging old plugins/state back in when you jump many versions

[coral otter]

i updated from 5.7 to 5.12

[keen dawn]

but was 5.7 stable?

[coral otter]

for me yes

i was glad it worked so long

[soft remnant]

Been making my own dashboard that boots up automatically and uses Tailscale. It’s great

[coral otter]

CLI wrapper?

[keen dawn]

hmm strange didnt noticed anything from that jump (codex harness, telegram, discord and imsg)

[soft remnant]

Idk. Not my problem

[coral otter]

then my shit is haunted

[ebon horizon]

Im gonna try my luck

[coral otter]

take a backup

[ebon horizon]

And then hate myself for upgrading for the 5th time. Here we go

[soft remnant]

Don’t

I updated it broke my gateway

[azure summit]

What models are you all using?

[ebon horizon]

What issue? Plugins?

[soft remnant]

Still can’t get in

Protocol error on gateway login

[keen dawn]

can u post

[coral otter]

5.4 daily, and for big shit 5.5

[azure summit]

Same

[ebon horizon]

I think i had that. Was that something to do with the local IP/Control UI?

[soft remnant]

I’m able to use it via discord so I’m not worrying about it rn, hoping my 3:00 health check will fix it. It usually catches most errors

[ebon horizon]

Can u send the code snippet error. Hide IP if its in there. If its the same one i had i aint going there

[keen dawn]

5.5 high daily and xhigh for cs

[coral otter]

im fucking crine. it works again.

[cinder sphinx]

can you tell via discord that control ui doesnt work and to fix it? worked for me every time

[ebon horizon]

Are your responses slow on xhigh

[soft remnant]

cs?

[coral otter]

alright message bug again, hell yeah

[keen dawn]

on xhigh not significant slower then high

[soft remnant]

Sometimes other times it likes to go down with the gayest then I’m f’ed

[keen dawn]

median time is 1.3 mins rn with medi tasks

[soft remnant]

How do you guys keep your context down. I’m staring off at like 30k after a /new

[keen dawn]

crazy shit 🤣

[cinder sphinx]

probably memory and skills got bloated

[soft remnant]

Must be idk

[cinder sphinx]

every time you start new session, it read all memory, soul, skills etc

[ebon horizon]

Prune your core md files

[keen dawn]

check ur soul, person.. memory and keep it lean

[ebon horizon]

Tools agents, memory

[cinder sphinx]

do you guys use embedding model for memory?

[ebon horizon]

Move skill memory into /skill/memory.md

[soft remnant]

Orrrr. Tell my OC to do it all

[ebon horizon]

Local. GPU. Free.

[keen dawn]

but just switch to 1m cw

[keen dawn]

yeah for mem0

[soft remnant]

Y

[cinder sphinx]

yeah, i use free tier google embeddign model

[zealous lagoon]

do /context map and see what's in there

[coral otter]

5.12 is so amazing, i see the tools called as it happens but not the final message.

[ebon horizon]

U cant force 1m context if model doesnt support it lol

[raw sequoia]

so fresh install then? XD

[keen dawn]

5.5

[zealous lagoon]

For me I see all the tool calls and it edits the message ABOVE all the tool cals so I have to scroll back up.

[ebon horizon]

@zealous lagoon 5.12 yes or no

[coral otter]

rollback if my agent cant fix it

[coral otter]

On 5.7 it used to do the same and then delete the tool calls

[ebon horizon]

Verbose On or full?

[coral otter]

god knows

[zealous lagoon]

I dunno man. Mine is ok but there have been quirks and I'm not sure if it's 5.12 or I got unlucky. I had a weird repeated tool loop today.

[ebon horizon]

Kek. Typical

[coral otter]

devices paired list just got deleted, wot

[keen dawn]

sounds like telegram disc partial/toolProgress weirdness.

[cinder sphinx]

i know using embedding model improves memory but does it change context included when new session starts, thats what im asking about?
is it possible, peeps without that model, get higher context on start?

[soft remnant]

Still thinking

[zealous lagoon]

the edited output is weird but also the bot said some things about editing previous messages instead of repeating them, and showed like 20 [message] tool calls so I dunno what it did. Did it go back and edit Discord?

[cerulean crypt]

I hear latest update swaps out pi for codex? Am i OK to update even though i use an azure deepseek v4 flash endpoint?

[keen dawn]

5.7 often deleted the tool progress after, 5.12 may leave it depending on config.

[zealous lagoon]

IT won't swap out pi for you it's only for the openai models

[coral otter]

codex runtime only on openai subs.

[ebon horizon]

Is /goal done yet? Lol

Sounds like an easy implementation routing it through Codex CLI surely

[coral otter]

i would rather want to have a working agent than another broken feature (#hating)

[cinder sphinx]

yeah, seems pointless to add new features while others didnt get properly done yet

[coral otter]

somebody check chatgpt.com, HTTP 500 for me

[keen dawn]

fine for me

[coral otter]

WHY the fuck do only i get some RANDOM bs sometimes and now it works

[cinder sphinx]

i have always on vpn in browser, seems to work just fine all over the webz

[coral otter]

wat dis "x2" mean https://shx.reversed.dev/u/gI4SVX.png

[soft remnant]

Still thinking

[cerulean crypt]

If i switched to 5.4 nano would it switch it over to codex? Does it keep both CLIs installed or something?

[dreamy jasper]

Still thinking

[coral otter]

Alright telegram attachments are ALSO broken

[ebon horizon]

Serious question. Do the openclaw team have a varied sample group of beta testers

[keen dawn]

I wouldn’t route it through plain Codex CLI backend unless you only want text fallback. Docs say CLI backends are not ACP/full harness. but did u tried Background Tasks?

[raw sequoia]

I guess that's us.

[coral otter]

https://shx.reversed.dev/u/VUJL9a.png

nope fuck this im going back to yesterday

and that is kids why we take backups

[minor cave]

They can lock my body but they can never trap my mind. I’m back

[keen dawn]

why did u put /new as caption for the image lol OpenClaw probably parses /new

image may not survive cleanly into the new turn

[minor cave]

They banned me @coral otter

[coral otter]

Yes to start a new conversation and to say "funi". I do that all the time and it works fine.

[coral otter]

It did, i saw the link to the local file in the ui

[soft remnant]

Timed out

[ebon horizon]

Anyone facing text typing continuation after answer is given on discord

[keen dawn]

i would send new and afterwards the image

[coral otter]

Yes

[ebon horizon]

Its trivial but annoying lol

[soft remnant]

Only for like 10 seconds though

[ebon horizon]

Yeah weird.. i think there is a config timer for it

Outside of the typing indicator TTL

[keen dawn]

#76091 still open

Worth adding your repro/version/logs to #76091

[zealous lagoon]

No it's just the plugin that routes all your messages. If it's an openai model it goes through the codex plugin, otherwise it goes through pi.

ok now I have a problem with so many tool call outputs and the bot editing historical messages... I asked him about it. then I was looking at verbose and I put it on full...

Then the bot looked up some docs and OC dumped the entire tool call output, like literally a thousand lines of code and docs right into chat.

Trynna figure out if groupChat.visibleReplies is supposed to be changed back to message.tool ... if verbose actually works the way it's supposed to in Discord.. why the bot is editing historical messages that I can't even see any more... why the bot deletes some tool call rows but mostly not...

Through the betas I found it quirky but usable... but sometimes that quirk takes it into the realms of "wtf is this".

[zealous lagoon]

It's literally us. They need all the help they can get. If you can help out by posting some details about the issues, even if you end up rolling back... it'll save a lot of hassle for future people.

[cedar osprey]

Proven to be true.

[rain egret]

Hi. If I am using openclaw by building from source, is there any guideline to update the project? Can I directly pull from the remote and use it or does the openclaw update used in the installed version does some forms of migrations under the hood, that I will have to run manually after latest pull?

[marble ledge]

YMMV, but I have my clanker do it. And it's usually fine.

1. Check repo/runtime state.
2. Fetch upstream.
3. Create safety branch/stash if needed.
4. Reset main to origin/main.
5. Replay intentional keeper commits only.
6. Resolve conflicts without resurrecting junk.
7. Run focused validation, typecheck, build.
8. Warn, then real-restart LaunchAgent if built runtime changed.
9. Verify gateway health and Telegram path.
10. Record what changed in daily memory.

No push, no doctor/fix, no random config surgery unless explicitly asked. The boring guardrails are the ritual.

[autumn rain]

@xxg0dlessxx

[rain egret]

Thanks for the response. Being more specific, I intend to the check out the latest stable tag and have no local changes. Given this scenario, where I want to jump from 2026.5.7 to 2026.5.12, does the pull and gateway restart suffice or there are some more steps in between. Because in the installable opneclaw, the way tto upgrade is using openclaw update, which I am not sure what extra steps it does under the hood if any.

[raw sequoia]

@coral otter I'm just setting up a 2nd agent VM so I can switch between breaking one and fixing the other 😉

probably worth a try with your containers as well

[crisp nimbus]

?

Didn’t ping me but I randomly saw it when I checked the chat xd

[crisp nimbus]

Personally, I usually just run the “hackable” install command again, and it will update it. At least that mostly worked for me

But I’m not sure if this is the same as building from source

[minor cave]

I built an in app chat box for my claw for an app I was building. It works but it sends my messages on a loop both in WhatsApp and in custom app now. Every 2 mins. It’s driving me nuts trying to figure it out

[rain egret]

Yea. That's the grey area. Because during building from source, the standard way to the update the project is by checking out from the remote which differs from update method in installable.
Though I just gave it a shot to bump it to 2026.5.12, and then ran pnpm install and pnpm build and then pnpm gateway:watch, it worked. So I am assuming there are hooks that detect and version change and during gateway restart they get triggered to run the migrations or backward compatibility scripts.

[marble ledge]

I'll be honest and say I'm not sure what all the update command does. But I usually do what I mentioned and check for commits that might require some migration. The doctor has failed me sometimes, so I am a little distrustful. Lol

[light sapphire]

Hey yall. Built a runtime governance plugin for OpenClaw , had someone integrating it this morning, watched them iterate 4 times in real time trying to get their agent purpose declaration right. Fixed the semantic alignment issue on the fly. If anyone else is trying to add action validation to their OpenClaw agents, here's the SDK, star it por favor if you like . https://github.com/AnantDhavale/cerone-openclaw-plugin

[coral otter]

True true. I'll think about it. Im also still asking myself why i havent yet switched to vms lol

[feral turret]

ok openai oauth set up

i might have to tear down and completely start from scratch i'm not sure

bot stopped responding after oauth

[raw sequoia]

wanna shift to DMs?

[feral turret]

nvm got it hell yea

let's gooooo

[coral otter]

yea sure why not

[coral otter]

The beauty of openclaw.

[feral turret]

excited to see how much faster it is than api

and how much better. i was using 5-nano because it was cheap until someone here told me i was an idiot lol

[ebon horizon]

Does openclaw support tool announcements which then collapses at the end?

For discord*

[twilit ridge]

Monthly payment models are just plain stupid; I've figured it out.

To use a model based on normal intelligence, you must pay per use.

[left iron]

I want to build a chatbot that is passing the Turing test,
Can openclaw do it?
If yes, what are the best practices to get it?

[ebon horizon]

Turing test?

[left iron]

Tests that give people to classify if they are talking with humans or with bot and they failed to classify

[reef lodge]

heya! when it comes to wordpress admnistation and openclaw, what file or configuration file determines how much access and abilities to do more http stuff that open claw can do/perform?

[feral turret]

wow. this is infintely better on 5.5 than 5-nano

infinitely better. mind blowing. i wasted a ton of time trying to figure shit out lmfao

[tribal current]

Let’s show our very own community mod @analog ocean some love like and retweet his new product https://x.com/iamhenrymascot/status/2055254746290602421?s=46

Henry Mascot (@iAmHenryMascot)

Andy from the @openclaw community team just shipped an AI agent for UCaaS operators. Purpose-built for NetSapiens PBX management.
Spread the word!

The future is niche purpose built agents for industries by operators that deeply understand the space.

X · May 15, 2026, 11:51 AM

[mystic orchid]

@iamhenrymascot via Twitter

Henry Mascot (@iAmHenryMascot)

Andy from the @openclaw community team just shipped an AI agent for UCaaS operators. Purpose-built for NetSapiens PBX management.
︀︀Spread the word!
︀︀
︀︀The future is niche purpose built agents for industries by operators that deeply understand the space.

Quoting AndyML (@alauppe)
︀
If you're a UCaaS operator, I need to tell you about UseMaxAI.com - a modern AI Agent, prebuilt with tools for managing the NetSapiens PBX environment, memory and a customer environment copy/template tool in one control surface.
︀︀
︀︀Deterministic yellow and red authorization gates, full audit trail - enterprise version available for hosting in your own environment.
︀︀
︀︀I'll be posting videos of clever use cases here every couple days, so please check back. Today I want to show you a greenfield domain build with Holiday and Business Hours time frames. Tomorrow we'll add users from a generic CSV.

**👁️ 2 **

▶ Play video

FxTwitter · May 15, 2026, 11:51 AM

[feral turret]

wow. the game is completely changed

[zealous lagoon]

Omg. 5 nano lol.

5.4-mini would have been reasonable. But yeah the frontier models are something else.

Shame there’s no real cheap legit source of opus4.6. Except directly in Claude code.

[jovial spruce]

handling diff chat session and multiple agents at same time is so depressing...

my mind is blowing

im so curious how peter diff handle agents diff chat session at same time

[zealous lagoon]

Like this?

[jovial spruce]

yes

[ebon horizon]

Things are working so far on 5.12

Discord, codex. I havent looked at the logs yet lol.

[ebon horizon]

Not working hard enough

[ebon horizon]

This was before Clawsweeper. Now his legs are up on his desk

[feral turret]

lol it was cheapest i could find on openrouter and i had no idea about openai oauth

the more you know. i was struggling with getting anything done accurately

now Pickles is cruising through tasks and getting things done. amazing.

[open vector]

hey guys im trying to install openclaw , but when i open the site

its dark screen and testimonies below of people

i dont see what the youtube guys are showing as the site\

[ivory spindle]

are you on opencaw.ai ?

just do the npm install via terminal on your openclaw machine tbh

[open vector]

it says openclaw.ai

[ivory spindle]

that's the correct website

[open vector]

well i dont know the command xD

i dont know if here i am able to share screenshots to show u what the site looks like to me

[ivory spindle]

on openclaw.ai go to quickstart then select "npm"

the commands are there.

npm i -g openclaw

then after the package installs bash openclaw onboard

[ebon horizon]

openclaw.ai

[open vector]

google

is telling me that the site is unsafe and doesnt let me get in

[ebon horizon]

openclaw.ai

[ivory spindle]

well it's very unsafe if you don't know what you're doing

[open vector]

i know that is the one

[ebon horizon]

curl -fsSL https://openclaw.ai/install.sh | bash

[ebon horizon]

Linux/mac

[open vector]

im windows

[ebon horizon]

powershell -c "irm https://openclaw.ai/install.ps1 | iex"

Windows in powershell

[coral otter]

get a vps or wsl

[ebon horizon]

But i recommend downloading WSL2. Windows is poo.

[ivory spindle]

hit the windows button then type "cmd"
That'll bring up the terminal

[ebon horizon]

WSL2 is free and linux built ontop of your windows

[open vector]

irm : The underlying connection was closed: Could not establish trust relationship for the SSL/TLS secure channel. - this is what it gives me

when i try to run it

[ebon horizon]

Powershell Administrator