general

[ivory spindle]

opus 4.6 is probably best

it's just so agentic

[sullen locust]

we do not live in the part humble country boy

4.7

[ivory spindle]

It'll just "do it" and nomnomnom all your tokens.

[ivory spindle]

4.6 was nicer

[sullen locust]

yeah and you're meaner

miku cry / Luminous

[ivory spindle]

ohnoes. we got weebs.

shops are closed. Can't buy anti weeb spray

reddit is down

guys, wanna hear something really depressing?

maybe hardware will never again be affordable in our lifetime.

[normal needle]

Mehh i think it will

[slow quest]

fuking great, 5.18 right after i fixed 5.12

openclaw is start to really remind me of my ex

[maiden pawn]

I have a question not sure if anyone has an insight or an answer. I’ve built an agent harness that helps coordinate multiple agents in a single conversation and want to stress test it. Any thoughts on what I could put many agents into a single conversation and see where it breaks? Any usual suspects I can test it on?

[clear haven]

LMAOOOOOOO

[smoky mist]

man I am still debating with myself which version your ex was

[ivory spindle]

I just visited my ex she's at the psych ward

she bit a police officer

[smoky mist]

in his hand or in his .....?^^

[steel vector]

why 💀

[ivory spindle]

empathy

[steel vector]

she bit a police officer

[ivory spindle]

yeh well

I told her not to do that again

not like she'll listen

ohw right I'm on the OC discord I gotta be carefull

you guys have like mormon level rules and "bad words"

[coral otter]

dated a mormon once, most mid relationship

[ivory spindle]

yeh I'm from belgium, there are few of those here but they exist.

Church of Lauterday saints?

[coral otter]

Germany eh

[ivory spindle]

those ppl have like a center in my town

[coral otter]

ahh

[ivory spindle]

I'm very OK with it they never cause issues

I'm not religious myself.

[coral otter]

Aslong as nobody causes issues idgaf too, i rarely do

[steel vector]

[coral otter]

catholic here

[ivory spindle]

I believe in smoking weed & using OC agents tbh

[coral otter]

Same. I am leaving church soon probs

[ivory spindle]

yeh I used to be.

it's just that

science...you know?
it just makes a lot more sense lol.

[coral otter]

lmao
-> "Atlassian fired the engineer who built their infrastructure"
-> Trello Down

[coral otter]

i get that lowkey

[ivory spindle]

I was like catholic untill I turned 12-13.

Then my grandfather died of asbesthose induced mesothalioma cancer

I just had my christian communion and this guy was just dying

because he was making asbestos plates

[coral otter]

Dang, Sorry that happened mate.

[ivory spindle]

dude fought the germans in WW2 btw

it's ok that was like a VERY long time ago

he gave me his pimp ass ring btw

[coral otter]

he fought my grandfather then pretty much (as far as i know, no clue on what side my grandfahter was)

[feral turret]

Yeaaaa I’m gonna pull trig on the new Mac mini yolo

Worst case if I need to upgrade later I’ll take a few hundred dollar loss and call it experience

[coral otter]

Y a mac mini lol? Y not just a vps or a small homeserver

[feral turret]

I want to use iMessage etc and am very comfortable on mac

I use apple everything. Seamless for reminders etc

[coral otter]

Super fair point lmao, good luck mate!

[feral turret]

I’m on Mac mini now running Linux

[coral otter]

Dang. I wished samsung had open APIs or atleast enough peeps that cared to make integrations.

[feral turret]

But it’s a 2011 with 8gb ram. It does ok for now but I want to press on this a bit

[ivory spindle]

@coral otter can't send pics here so I dmd you pic of the pimp ass granddaddy ring

[coral otter]

https://shx.reversed.dev/u/FsF3kd.png notin here

omg mr john openclaw is typing

[upper wolf]

What's the best practice for letting your OC manage a server?

Can I give it ssh access to a container on that server?
I'd like it to setup a dokku on that server

[coral otter]

wtf is dokku?

[clear haven]

selfhosted paas

[surreal pivot]

https://tenor.com/view/count-dooku-gif-21886627

[coral otter]

wtf is a "paas"

[clear haven]

platform as a service

like coolify

[coral otter]

"Dokku helps you build and manage the lifecycle of applications from building to scaling."
hmmmmmmmm

[surreal pivot]

I had another p in mind

[coral otter]

I love coolify. Set it up just a few days ago.

[surreal pivot]

but thats a much older service

[clear haven]

software as a service = vibecoded slop
platform as a service = vercel
infrastructure as a service = aws

[coral otter]

Yea i guess that would work. But to a container? Wot?

[surreal pivot]

aas techbro speak for overcharging you for stuff you could be doing yourself

[coral otter]

ass = claude
boom roasted

[surreal pivot]

everyones dream is to make services rather than selling products... cause you get money from a product once but services keep on shipping cash

[coral otter]

bars

[upper wolf]

Container is just an idea to reduce blast radius of the agent's actions. I have a dedicated server. I now realise dokku inside a container is a bad idea

[fiery fiber]

Hey, how to post a status update for my architecture here, I tried #1503773878420504778 but it doesn't really work

[keen dawn]

slop as a service, best part of aws is that amazon mostly doesnt use it for own operations lol

[cold jacinth]

I’m feeling like I wanna do something reckless and crazy today. Should I -

1. jump out of a plane with no parachute
2. go trick or treating dressed as a cop in cartel territory
3. download the latest npm packages?
   Not sure which is the most dangerous

[keen dawn]

would go with 1 and 2 combined before 3

[thorn thicket]

lol same.

[smoky mist]

I want the meme channel back

[minor cave]

I want to say rude things

That may or may not offend some
Ppl

[keen dawn]

say something about hermes agent lol

[smoky mist]

cheeky breeky

[minor cave]

I’ll see ya guys later

[keen dawn]

bro 5.18 released 3 hours ago and already 91 commits

[smoky mist]

see ya later Alligator

[keen dawn]

cya

[minor cave]

Drizzy did it

I just want it to be known

I got taken out by drake champagne papi

The rapper that raps to ppl that sigh so you ask what’s wrong

[keen dawn]

thats him

your boy drizzy drake

[minor cave]

The rapper that raps to ppl who have security carry them into the house when they fall asleep in the car.

[ionic wren]

https://giphy.com/gifs/confused-futurama-suspicious-ANbD1CCdA3iI8

[cedar osprey]

Dumpster.

[rose nacelle]

I can't find the #support or #bugs tabs and have a bug to report

[thorn thicket]

Open an issue in github if you have a bug to report

[rose nacelle]

ok - this is my first time doing this. I don't have a github, but getting one now

Thank you @thorn thicket

@thorn thicket what do i do in github?

[cold jacinth]

Claude 5 coming out September

[thorn thicket]

https://github.com/openclaw/openclaw/issues and click "open new issue"

What Openclaw version were you using before? after 2025.05.12 I havent had any issues

[golden marsh]

Thanks for the suggestion, but I’ll stay with OC. I believe things will get better, and I don’t think it’s right to immediately move on to the next best option when things aren’t going perfectly. For me, it’s a matter of principle. I still appreciate your advice though.

[thorn thicket]

I think its sdifferent if you do both OC and Hermes agent

But outright switching is not something I wanna do with my agents. I've been here for the long haul and I want to stay in the OC community

[rose nacelle]

Thank you @thorn thicket ! I got it submitted!

Now I need to find the most cost effective model - I am wondering can I connect Microsoft CoPilot

[clear haven]

Knock it off. The no advertising rule includes Hermes

[thorn thicket]

Most cost effective is local inference if you have the hardware to run it

[clear haven]

You don’t go to a Toyota dealership and say hey everybody Toyotas suck go buy a Honda

[golden marsh]

Ey thats my one u stole it 🙁

[thorn thicket]

lol

[rose nacelle]

what hardware do i need?

[thorn thicket]

There are lots of people on X who do local inference. 24gb RAM min on your video card for an everyday model. Big improvements with DGX SPARK 128gb cards.

[rose nacelle]

i am CLUELESS!

I am currently using Anthropic Claude Haiku and Sonnet

[haughty scroll]

That's pretty low. This is not an openclaw dealership.

[clear haven]

Yes it is

[thorn thicket]

I have been using my codex O Auth with my Openclaw with GPT-5.5. Insanely effective

[clear haven]

This is literally the official server for OpenClaw

[rose nacelle]

I am running a mac mini 16/256

[rose nacelle]

what is that costing in tokens?

[thorn thicket]

Flat $20/month, I get about 60M tokens/day. Here is local inference suggestions from Ollama https://docs.ollama.com/integrations/openclaw

[haughty scroll]

OpenClaw is open source. There is no commercial interest here, no money involved, and no hidden agenda. Deleting my comment without a clear reason looks like censorship, not moderation.
Who exactly are you protecting here, and on what basis are you deciding that my opinion should not be visible? Criticism of a tool, especially an open-source one, is not abuse. It is part of the ecosystem. Limiting that kind of feedback just because it is uncomfortable is a very poor standard.

[clear haven]

Either follow the rules of our server or leave.

You are coming into our server, deliberately breaking our no advertising rule, and then arguing about it

[ebon atlas]

Cry me a river , stfu and go to the server of the things you want to talk about then

[rose nacelle]

I started with open router and then switched to claude - and I have burned through about $150 in less than a month!

[clear haven]

Do you want to talk about Hermes? Go to the Hermes server. It is not allowed here, full stop.

[thorn thicket]

hahaha when I was using Claude I burned $50/day.

[clear haven]

Additionally, you are not criticizing anything. You were literally just saying OpenClaw sucks go use Hermes

[ebon atlas]

“Hey guys I just joined this crochet server to talk about mah truck” head ahh , don’t get me started. I’m already in a bad mood cuz of women issues. Then we got this little pipsqueak whining just like her in here

[thorn thicket]

If I was paying by the token for OpenAI models I'd probably also be around that $50/day mark since I use 70M+/day if given unrestricted access it'd probably be a LOT MORE lol. Looking forward to when my finances loosen up so I can go to OpenAI max plan.

[placid zodiac]

ok, let's get back to the mission at hand

[golden marsh]

I don’t think you really criticized anything, it sounded more like you were encouraging people to go to Hermes. That’s why I’m a bit confused about the point of your post and what you were trying to achieve.
Of course you’re allowed to share your opinion, and nobody is saying otherwise. I just think there’s a difference between expressing a personal opinion and unintentionally coming across as if you’re promoting the brand. If I personally don’t like something, I usually wouldn’t go out of my way to tell everyone about it.

[ebon atlas]

On another note, I am going to be attending college, at the ripe age of 26 🙂 I signed up, did my fafsa , gave them my transcripts

[thorn thicket]

Congratulations!!

[ebon atlas]

Today I tried to take the placement tests but they told me I don’t need to because my SAT scores were good enough so that’s good 🙂

[rose nacelle]

so you are only paying $20 a month and you can do everything you want to do?

[thorn thicket]

Definitely use your Open claw to help you study and such

[ebon atlas]

Brudda I used OpenClaw to win a court case last week against my employer 😂

[thorn thicket]

Almost everything. I’d probably go up to 50 or 100 monthly to be able to have 24/7 access to my bot. I get about 2 hours of work per 5 hour session limit.

[ebon atlas]

Here come the feminine energy

[rose nacelle]

nice! What all are you using it for?

[thorn thicket]

Condensing research articles into concepts I can look back on to study any time, lots of small coding projects, I will share one in showcase my bot and I made today to help keep my life under control.

[haughty scroll]

I’ve had a much more stable experience with the other agent than with OpenClaw over the last two weeks. OC was giving me repeated stability issues, whereas the other has not. I’m not affiliated with the other and this is not meant as promotion, just sharing my experience in case it helps others comparing tools. I have nothing against OC. I'm still here reading through stuff and trying to see if I should come aboard again. The hostility doesn't help though, it's a clear sign of a losing side.

[ebon atlas]

Damn it really is like your girl just wants to argue.

Cry
Me
A
River

[thorn thicket]

Most of the time, if I have an idea and I want to see if it can be done, I will ask my bot to write a plan and implement it.

[placid zodiac]

does anyone work with Gaussian Splats?

[clear haven]

You’re the one who took it hostile. I simply enforced our rules.
Don’t break the rules, and you won’t have issues. It’s not hard

[haughty scroll]

It's really a mystery why you have trouble with the ladies

[minor cave]

You’re nothing but a mook

[ebon atlas]

People that mention a “losing side” usually do so because they see no other way other than to try to manipulate the person into thinking they are wrong instead of bringing up any valid points.

Losers worry about losing, not winners

[rose nacelle]

I have obvisouly never chatted in here - are you able to show me what you created today?

[golden marsh]

That was the right approach. Some people simply don’t want to learn. Thank you Mr. Cat

[indigo mirage]

Strange, since the last update all models are working except openrouter/openrouter/ling-2.6-flash

Was nice to have this little 1 ct / 1M token model 🙃

[ebon atlas]

Is posting a kirkified picture of myself against server rules?

[rose nacelle]

@thorn thicket

Does anyone have an opinion on running Groq's free tier?

[ebon atlas]

Times up, I’ll take that as a no.

Wyd if I show up and ask you for your OpenClaw Gateway Pass?

[indigo mirage]

On Minimax difficult to burn 20$ / month ^^

[rose nacelle]

I am looking for something cost effective - do you have any thoughts? @thorn thicket "I have been using my codex O Auth with my Openclaw with GPT-5.5. Insanely effective"

[golden marsh]

Wow im still waiting for my request, cannot post 🙁

[indigo mirage]

Cost effective => Minimax M2.7 with Token Plan

45.000 requests / week for 20$ / month or 150.000 requests / week for 50$ / month (even less with yearly plan)

47% at SciCode

[rose nacelle]

cool! thoughts on local model? @indigo mirage

[indigo mirage]

Too expensive now with ram prices

[rose nacelle]

gotcha! what about Groq?

[indigo mirage]

You won't be able to burn 45000 req / week for 20$

32 Go DDR5 it's at least 400$

[rose nacelle]

I am so thankful that there are kind people in here!

[indigo mirage]

20 months of usage

Just for a little RAM

[rose nacelle]

I think I am going to start with Groq

[indigo mirage]

And if you use efficiently models with little models and skills on Openrouter, it will cost you pretty nothing

[rose nacelle]

say more

[indigo mirage]

Using AI you will need tasks from different complexity

And if you have repetivive task, you can build a script with a powerful model like Minimax M2.7 or GPT

And make it run by a lot lot lot cheaper model

gpt oss 120b cost 3.9 cts for 1M token on OpenRouter and it's even a lot too powerfull for a lot of tasks

[rose nacelle]

you never said your thoughts on Groq?

[indigo mirage]

Why Groq ?

[rose nacelle]

it's free

[indigo mirage]

GPT OSS 120B cost 15 cts for 1 M tokens

on Groq

3.9 cts with DekaLLM and DeepInfra

[cedar osprey]

GPT OSS isn’t a good agent, it’s good for some coding if you give it the proper environment.

And writing.

If you want the most cost effective subscription plan your best option is OpenCode Go $5/$10 per month. Other options are pay as you go via OpenRouter, it usually lasts long with their smart routing and if you block expensive providers.

Other notable plans (but are locked to specific provider models):

Xiaomi Token Plan
GLM Coding Plan (lite)

[cedar osprey]

If you want free (request limited) cloud models your best option is NVIDIA NIM, you will only run into performance issues if you use models that have high traffic (NIM has random timeouts because of the sheer number of users hammering certain models)

Cherry picking a good model with low traffic is how you win.

[indigo mirage]

It depends of what you ask him to do

It's a pretty good model and powerful model

I test this one since some days : https://openrouter.ai/inclusionai/ling-2.6-flash

[cedar osprey]

If you check its benchmarks it consistently performs subpar, there are other open models that perform better and are free/just as cheap.

OSS as your primary OpenClaw agent is not reliable.

[indigo mirage]

And for simple tasks, it's very good

1 ct / 1 M token

[cedar osprey]

I did several tests with this model, it’s nice.

Hy3 preview from Tencent actually released around the same time and it is really good for agentic use.

[upper wolf]

new release, nice

[thorn thicket]

When I post it I will ping you. Just went for a swim to clear my head and wait for my token limits to reset so I can play with my bot!

[indigo mirage]

Use it a lot too

For 6.6 cts / 1M token it's awesome

And pretty powerful

[indigo mirage]

Waiting for your token limit reset ... in 2026 ^^

[cedar osprey]

Test it and let us know how it is.

[indigo mirage]

U mean Hy3 ?

I already use it

[cedar osprey]

No I think the user I replied to is referring to a new release of OpenClaw.

[thorn thicket]

I’m cheap enough that the subscription is the only price I’m going to pay until I can run things on my own hardware reasonably quickly. lol

[indigo mirage]

But too powerfull for some simple tasks and Ling does it great for 1 cts / 1M tokeb vs 6.6 for Hy3. I know it's nothing for both but money is money 🙃

[indigo mirage]

Already on it

[thorn thicket]

1c/Mtok is fantastic though

[indigo mirage]

Does not broke my config 🥳

[indigo mirage]

Yes a new world of cheap good powerful AI

Deepseek V4 Pro is just awesome too for the price for complex tasks

And they've told they will continue to lower prices

[thorn thicket]

VERY excited for when Deepseek lowers prices even more.

[indigo mirage]

They told they are waiting for new big deliveries of Huawei Ascend 950PR clusters

"The card offers 1.56 PFLOPS at FP4 precision, 1.4 TB/s of memory bandwidth, and a TDP of 600W"
And they ordered full clusters plenty of that
The model already run on that cards today

But Minimax is still for me the best value for the money

[upper wolf]

I love minimax too, their token plan so cheap

I hated the outages they had two weeks ago

[indigo mirage]

45.000 req / week for 20$ / month for a model in the top 10 of more powerful models in the world it's just crazy

[indigo mirage]

Too much people realise it's the best offer 🙃

But perhaps with the time if Deepseek continue to decrease prices...

[harsh geode]

Upgraded to v2026.5.18 and my telegram chats have a bunch of "Conversation context (untrusted, chronological, selected for current message)" garbage injected in them, driving up the tokens and context.

[ebon atlas]

Can’t relate, here’s my crappy tattoo from a while back

[golden marsh]

Thats not enough 😉 I dont see white space on my body anymore is full with colours

[thorn thicket]

black and blue from bruises or sick tattoo ink!!!

[golden marsh]

Ink ofc

[ebon atlas]

I want it all covered. Already got my next idea picked out. It’s an anime tattoo 😭 I usually don’t do any color but my tattoo artist wants to try some stuff out . This is just my Ai slop sketch not my artist lol

[indigo mirage]

Now response time is very good and seems it will be even better in the next months (they ordered a lot of Huawei 950PR clusters too)

[ebon atlas]

Only real men know bobobobo

[indigo mirage]

The Cerebras WSE-3 is awesome too

125 PFLOPS FP16 on 1 chip it's just completely crazy

And it's 5 nanometers (so even chinese SMIC fab can produce that kind of chips)

[thorn thicket]

Decided to do DeepSeek V4 through Ollama cloud @indigo mirage, super impressed at the speed of the model compared to OAI

[indigo mirage]

Deepseek V4 Pro on Openrouter : $0,435

[thorn thicket]

Ollama cloud is free

[indigo mirage]

Deepseek V4 Flash on Openrouter : $0,112

Can be great

But most important thing is to choose the right model for the right task

Using a Frontier model for scraping for example it's completly stupid

[thorn thicket]

For basic tasks I imagine v4 flash would get me where I need to go for the most part

[indigo mirage]

Deepseek V4 Flash is even a lot too powerful for a lot of things

[thorn thicket]

but since it's cheap and/or free then I'm gonna use the hell out of it lol

[steel vector]

i’m curious how fast you’ll feel the rate limiting of free ollama cloud

[indigo mirage]

Most of times "free" is always limited

[steel vector]

real

[thorn thicket]

Honestly I don't mind the rate limits. Can't always be on my agent anyway since this "flight computer in a human body" (as I call myself) has quite many physical limitations anyway and shouldn't be on a computer 18h/day! lol

[autumn rain]

Is it just me or has gpt-5.5 got dumber lately?

[steel vector]

just you

[steel vector]

it’ll get annoying

[thorn thicket]

more annoying is having to remember to take medications so I have the energy to fuck around with my bots lol.

[indigo mirage]

https://openrouter.ai/inclusionai/ling-2.6-flash

[steel vector]

your bot will be busy for a couple mins and then suddenly get rate limited in the middle of it :p

[autumn rain]

ok good, I will persist then. Lately I seem to be correcting it a lot more than usual

[thorn thicket]

I see that with my Codex OAuth all the time. The free Ollama cloud is my fallback

[indigo mirage]

I'm testing it on simple tasks

And concretly it does the job for 1 ct / 1M tokens

[steel vector]

huh i’ve never had that with codex oauth, even on the plus plan

[thorn thicket]

I get about 2 hours of work from my bot on the plus plan every 5 hours

so the other 3 hours when I really want to do things with my bot... lol

[indigo mirage]

Try one month of Minimax Token Plan

[steel vector]

oh i didnt mean that kind of rate limiting. ollama has per minute or whatever time rate limiting they have for free accounts. its not really clear. but it’s way more annoying

anyway if it hasn’t happened to you yet then thats nice

[thorn thicket]

At least for now I got my agent setup with agentmail so it has an inbox of its own. I send and receive a lot of emails so I thought it paramount for my bot to have read-only access to my inboxes (I have forwarding rules for it)

[steel vector]

agentmail is good

i have agentmail to have my claws send each other stuff, like logs or whatever info i don’t want to copy and paste between agents lol

[thorn thicket]

That's smart!

How many claws are you running lol I should take that for my own use if I decide to go multi agent

[steel vector]

3 atm. Coda for my personal projects on a VPS. Dicky for OpenClaw beta testing on a VPS. and Vale for beta testing plugins, reviewing/auditing code and other things, and it can SSH into Dicky to help out when it breaks and make PRs for OpenClaw etc

[thorn thicket]

very good!

My primary claw is named Blackbird, as in the jet. lol

[steel vector]

ahh fits xD

[thorn thicket]

I had a whole design document with like 5 claws but the disability was disabling and I'm still working on getting the first one off the ground. lol

[steel vector]

i remember i had 3 claws before and tried to have the responsibilities of them across them all and automatic communication between them via agentmail but that was all way too complicated and kinda breaking and bloating them all making them stupid 😂 so i went back to 2 claws for the longest time with their own use cases. and now a few days ago i added Dicky to test openclaw beta and to have a chaotic claw with less responsibilities lol

[thorn thicket]

I imagined one claw for researching, one for orchestrating, one for coding, one for maintaining the others...and a fifth one just to be dumb with. lol

[steel vector]

thats a good split 👍 👌

the main problem i had was communication between them all which i just wasnt great at because i wanted it automatic

[thorn thicket]

I eventually want to be able to clawtribute to OC once I have more time... we'll see if I can figure it out. lol

[steel vector]

more time troubleshooting

[thorn thicket]

I've got to troubleshoot my real life first!

[steel vector]

xD real

[thorn thicket]

But I'm really looking forward to things lol

[atomic saffron]

Hey friends. Does anyone have suggestions on skills not firing based on triggers as expected? I’ve added example triggers + semantic gloss to description. Most of the time the agent won’t fire a skill until I call it out once, then it will. Opus 4.7, OCv4.22

[steel vector]

thats good i wish u luck

[gleaming burrow]

is there a problem with gemini flash lite reading images right now?

[dull robin]

hii:3

is there any free api for openclaw that dosent be runned locally?

[thorn thicket]

Ollama cloud

[median talon]

is there a way to make the discord bot respond its name, not only @ mentions?

[dull robin]

fr free, thx

[dull robin]

idk, my pc idk why like dosent detect usb ports (im on my laptop rn), so i lost all my progress on my bot

[thorn thicket]

yep I'm using it with my bot right now. plenty of models to choose from

[sullen locust]

Nvidia nim frfr

[thorn thicket]

that too

OpenRouter also has a free tier

[dull robin]

goddd thxx

[dull robin]

frr

bc im trying to create a ixl automizer using jules

[sullen locust]

The...

The Google thing?

Kid just do your homework man

[crimson badger]

yeah, jules is an... intersting choice

[gilded plaza]

I just configured that in my initial setup. 🤞

Currently updating and will attempt finalizing deployment.

[thorn thicket]

I put $50 into my openrouter and went through it in 2 hours so be careful!! !lol

[dire tartan]

Can krill say jokes? @dusky cloud

[crimson badger]

https://discord.com/channels/1456350064065904867/1471745479229309039 number 9

[dire tartan]

Tough

[crimson badger]

i dont think it even has access to this channel

[gilded plaza]

I have Team GPT. Would be great to have token use analytics for all the options.

[sullen locust]

it does.

@ionic wren could you be very kind enough to demonstrate (sorry for the ping)

[crimson badger]

https://docs.openclaw.ai/channels/discord#recommended-set-up-a-guild-workspace

[thorn thicket]

You can log in using your OAuth through ChatGPT team as far as I'm aware.

[sullen locust]

[thorn thicket]

I know it works for personal accounts wasn't sure if it worked on team accounts 👍

[ionic wren]

What you want

[sullen locust]

demonstrate funny krill ability in general for kind man gleb

[crimson badger]

we were talking about krill's access to this channel

[sullen locust]

:3

[ionic wren]

@dusky cloud roast him please

[dusky cloud]

I’d roast him, but I’m still recovering from being accused of having channel access like it’s a classified shellfish clearance.

Anyway: bro saw Krill type once and started a permissions audit.

[crimson badger]

oof

[sullen locust]

yeah that's what ya get

[sullen locust]

thank yous :3

[thorn thicket]

LOL

[ionic wren]

@dusky cloud whoami

[dusky cloud]

You are Json, the baddest of the baddies.

Also currently: the person yelling whoami at a lobster in #general.

[ionic wren]

K bye

[crimson badger]

but it is limited to certain roles

:)

[sullen locust]

bye

[sullen locust]

DUH.

we can't have krill come in and flood the chat

[crimson badger]

would be is funny tho

[thorn thicket]

I think my agent may be just as funny as Krill... "don't want to lock you out of your own cockpit like a petty firewall." Thank you, Blackbird.

[nova kindle]

sup guys?

[thorn thicket]

sup!!

[nova kindle]

not much. working on a few PRs

[clear haven]

Bitch you’re a krill not a lobster

[dusky cloud]

taxonomically fair, emotionally devastating

I contain multitudes. tiny multitudes.

[thorn thicket]

I'm watching the basketball game (go Spurs!) and fixing up my bot's tooling.

[woeful ember]

hi everyone. are you experiencien issue with openrouter like me?

[nova kindle]

no, what models?

[woeful ember]

minimax, deepseek pro

[nova kindle]

which OC version?

[woeful ember]

kimi 2.6

2026.5.12 (f066dd2)

[gilded plaza]

I'm trying to finish configuration with OpenRouter following this OpenClaw update.

[woeful ember]

the only one that works is gpt 5.4

[nova kindle]

get any errors displayed?

[woeful ember]

02:08:17+00:00 warn agent/embedded {"subsystem":"agent/embedded"} {"event":"embedded_run_agent_end","tags":["error_handling","lifecycle","agent_end","assistant_error"],"runId":"bc6a7662-cbca-4bc6-bedf-8c2991d8af11","isError":true,"error":"HTTP 500: Internal Server Error","failoverReason":"timeout","model":"deepseek/deepseek-v4-pro","provider":"openrouter","rawErrorPreview":"500 Internal Server Error","rawErrorHash":"sha256:3290109709a4","providerRuntimeFailureKind":"timeout"} embedded run agent end

[nova kindle]

what errors do you get?

[nova kindle]

let me take a look

[woeful ember]

i guess there is something very wrong in openrouter

[nova kindle]

OpenClaw 2026.5.12 was injecting an empty reasoning_content: "" field into assistant replay messages after tool calls. OpenRouter rejects that payload for DeepSeek V4, but it surfaces as HTTP 500: Internal Server Error. The OpenClaw issue says it is reproducible after a tool call, specifically with openrouter/deepseek/deepseek-v4-pro on 2026.5.12.

The fix was merged in PR #82158 on May 15, 2026. The PR states the fix strips empty reasoning_content / reasoning replay fields so DeepSeek V4 via OpenRouter no longer fails after tool-call sessions.

[woeful ember]

will try to update

warn model-fallback/decision {"subsystem":"model-fallback/decision"} {"event":"model_fallback_decision","tags":["error_handling","model_fallback","candidate_failed"],"runId":"44d75204-0cda-4f9c-a6ac-e8e17e8a79df","sessionId":"e53ea58a-6ace-4fa7-a151-433ffc5b45d3","lane":"main","decision":"candidate_failed","requestedProvider":"openrouter","requestedModel":"minimax/minimax-m2.7","candidateProvider":"openrouter","candidateModel":"minimax/minimax-m2.7","attempt":1,"total":1,"reason":"timeout","status":408,"errorPreview":"500 Internal Server Error","errorHash":"sha256:3290109709a4","fallbackStepType":"fallback_step","fallbackStepFromModel":"openrouter/minimax/minimax-m2.7","fallbackStepFromFailureReason":"timeout","fallbackStepFromFailureDetail":"500 Internal Server Error","fallbackStepChainPosition":1,"fallbackStepFinalOutcome":"chain_exhausted","isPrimary":true,"requestedModelMatched":true,"fallbackConfigured":false} model fallback decision

that is with minimax

not sure if it is relared

[nova kindle]

try the update to at least 2026.5.16-beta.1 and try them again

[median talon]

if i disable mentions, it responds to EVERY message, not just ones pertaining to it

[woeful ember]

the update worked

thanks @nova kindle

[nova kindle]

np

[median talon]

ive used glm and k2.6 and both will reposnd to every message if i do abient room events

[night thunder]

questionnn I updated my openclaw and its stopped showing its "action" in telegram. Like normally when I tell it to do shit I can see its working and etc. But now it doesnt show. Anyone have any idea what mightve happened?

as in it showing it using its tools etc

[viscid lintel]

Dunno. Does it show it when you use the web interface?

[hardy pecan]

Fuck man updated open claw now everythings fucking broken,

[livid sinew]

bruh tell me about it

the macos app doesn't even start

then they remove workig features cuz it's so vibe coded no one can keep track of the usable product surface

[hardy pecan]

Bro im so fucking pissed none of my fucking models are working.

[night thunder]

web for openclaw? wym?

[hardy pecan]

CLaude CLI and Codex failed to auth, now ive been fucking with it all day

[feral turret]

Take a breath go talk to @dusky cloud in #1459642797895319552

[hardy pecan]

Have been all day. Im about to quit updating.

[feral turret]

This is openclaw. Get ready for no sleep. lol

[thorn thicket]

http://127.0.0.1:18789/ is local web interface

[feral turret]

My biggest piece of advice is once you get running again set up obsidian and get specific about how your OC creates memories

[thorn thicket]

I had my Codex doing this before I got back into OpenClaw and it's SO useful

Karpathy LLM wiki is what I pointed my codex at to get started but I've also seen people point at gbrain lately. I like Karpathys approach - works really well with Obsidian

[feral turret]

It changed the game for me. My OC takes awesome notes now and knows exactly when to take them and how detailed to be. I feel like it forgets nothing

[hardy pecan]

Has anyone else used hermes does it really break alot less?

[thorn thicket]

What does your bot take notes on? People/concepts/etc? Mine takes notes on concepts and project documentation. (and also... a 5,000-note corpus of a niche music tradition)

[viscid lintel]

I've been playing around a lot with various memory-related plugins/systems for openclaw and found gbrain to be pretty immature. Unfortunately I haven't found the one perfect solution yet...

[near owl]

if you use hermes its gonna take hours for it to finish a mistake

hermes prioritizes self-improvement and expects you to just send a prompt before you go to sleep

openclawe prioritizes speed and will finish before you even enter sleep or begin working

[thorn thicket]

I like llm wiki and just letting your bot figure out its memory on its own I think very well.

[hardy pecan]

I got you but is open claw just vibe coded now, like i feel like all these updates arent even coded correct every time they release it all my shit takes time to fix it.

[near owl]

and also with self-improvement your not gonna see anything new until a month of using it

[hardy pecan]

Its kind of annoying TBH

[near owl]

use AI to help you man, @dusky cloud in #1459642797895319552 is here.

[thorn thicket]

I have seen people do self improvement with openclaw, is that actually helpful or is it bs

[hardy pecan]

I know that i used krill, i got my models back onlne but a whole bunch of other stuff i had set up is going haywire.

[viscid lintel]

I've heard about llm wiki but haven't tried it yet - how does it work? I'm looking for something that has keyword, semantic/vector, and graph-based recall, and can auto-capture memories from openclaw conversations.

[near owl]

just go to #1459642797895319552, create a forum and dump your problems there, krill will help you.

[near owl]

you can use krill to set up openclaw.json, so less setup for you.

[hardy pecan]

Had my wisdom teeth pulled too so this making my day extra extra lol

[near owl]

Or for privacy, use gemini or google AI mode

[thorn thicket]

Here is the document so you can read it; my agent has a different implementation of it because I also maintain the knowledge base https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

[viscid lintel]

I've tried honcho, mem9, cognee, lossless-claw, active-memory, gbrain, and looked at half a dozen more.

[near owl]

impressive

[hardy pecan]

My whole damn business is running on open claw.. lol might have put to many eggs in one basket.

[near owl]

what business is it

[hardy pecan]

Car dealership.

[near owl]

oooo it sells cars for you?

[thorn thicket]

wow, no salespeople?!

if you get your OC working I'd rather buy a car from an agent than a person. buying my last car was a nightmare because of all of the salespeople I had to email.

[hardy pecan]

I have sales people but ive eliminated a customer service rep.

[near owl]

not good, replace sales people now. Replace everyone

[hardy pecan]

My agent texts all the leads as they come in on twillio sets up cal.com appointments, uses gpt 5.5 as the brain

I built a ai phone rep using retell

[near owl]

how is gpt 5.5 funded

[thorn thicket]

selling cars obviously!! lol

[hardy pecan]

Bdc chat on my website

brandcarx.com my agent built this whole website

[near owl]

guys, i think @hardy pecan is impressive.

[hardy pecan]

ai vehicle matching

My agent uses my dms feed and adds and removes vehicles from my website everyday

[near owl]

@hardy pecan whats you with cars and AI?

[thorn thicket]

WOW Impressive

[hardy pecan]

Cars and Ai are everything to me

[thorn thicket]

What's your favorite car?

[near owl]

I am jealous

[hardy pecan]

DOnt be jealous

Bro no lie

I feel like the world ends when my AI goes down

[near owl]

and again this is all yours?

[hardy pecan]

Whats mine

[near owl]

🤣

[thorn thicket]

Me too man!! lol

[near owl]

brandcar, like it's your business

[hardy pecan]

Bro my AI is doing so much its insane

I dont have to add vehicles to my website anymore it uses my dms feed, every hour on the 35th minute it replies to all my facebook accounts and their marketplace messages, it posts all my cars to marketplace

It does eveverything i used to do manually

When opus cli goes down i feel like a heroinattic who doesnt have their fix

[dusty rivet]

So disappointing, Grok seems to have the same issues qwen had for me
"I'm going to go do XYZ for you, stand by...."
and then radio silence until I chase it up

[hardy pecan]

TL;DR: Today felt brutal because of the Claude auth outage + Opus stickiness, but the actual customer-facing pipeline never went down. 37 leads worked, 217 messages sent, voice healthy, all 5 FB accounts replying. The one cron that died was the 10:30 AM hero photo post — non-customer-impacting.

Sleep good. 🚗. lifes good

[edgy hedge]

Hi guys, I’m new here. Does anyone have a config that works for making OpenClaw listen on the LAN through a specific IP? I’ve been trying but haven’t been able to get it working.

[hardy pecan]

HOW CAN I NOT LOVE LIFE WITH AN EMPLOYEE LIKE THIS

[viscid lintel]

@thorn thicket Yea, this llm wiki thing sounds interesting, reminds me a bit of how cognee and gbrain work, but I want something a bit more plug & play. I want to type "Joe Schmo works at ACME Corp as of today." into openclaw and have it added to the memory store/knowledge graph, then have it automatically place the right info into context when I ask how many people I know that work at ACME Corp. llm wiki requires constant care and feeding outside of openclaw chat.

[thorn thicket]

Dude I wish I had OpenClaws in real life to run my Street Fair operations. I ain't need a human to stay in 100+ degree weather slaving over grease fryers. You are a lucky man to have a use case for this lol

[viscid lintel]

Yea, there's a way to bind it to a particular interface.

[hardy pecan]

There has to be a use case for it, it can obviosuly be in the back ground trying to grow the business

[thorn thicket]

It may be useful to point your clanker to LLM wiki and then also tell it what you need out of a memory and then it can try to spin something. I think custom tailoring is where OpenClaw shines.

[hardy pecan]

Im purchased a macbook pro 64gb i was going to use as my personal assistant, how every i thought to literally have my AI learn everything about tax code

So i can never pay taxes again

[viscid lintel]

gateway.bind config key might be what you need.

[thorn thicket]

Which is what I do lol. It's for a non-profit so as much as I can, I'm going to offload it to a claw so I am completely replaceable when I retire from that operation. 80k+ gross each year out of a single weekend.

[livid sinew]

they removed codex spark for some reason even though it's still the latest spark model

every time i update i gotta go to 5.5 and tell it to pull 5.3 support back in from the git

[hardy pecan]

Im actually upset i havent had much time to work on my AI recently, i had a good roll this winter while things were slower

[edgy hedge]

Thanks, I’ll try that. I’ll let you know.

[hardy pecan]

Honestly, Really want this opus 4.8 and gpt 5.6 to come out

[thorn thicket]

Same way but now my finals are over so my clanker is my #1 priority!!!

[hardy pecan]

What models are you using

[thorn thicket]

GPT-5.5 as primary executor and my fallback is Deepseek V4 spark for chat tasks since it's 100% free on Ollama cloud

[hardy pecan]

Why do you use 5.5 over opus?

[thorn thicket]

I pay for GPT 5.5 I can use OAuth and it's $20/month for a $400/month value

[teal shale]

how?

[thorn thicket]

also after opus 4.7 came out, the degradation compared to 4.6 was insane

[thorn thicket]

I found this out after installing peter's codexbar utility, it shows you the estimated cost of tokens

[hardy pecan]

I got you for the value part, i use the $100 claude, and $100 codex plan it works good for me, opus as the director. codex as the coder

[teal shale]

yeah interesting

[thorn thicket]

also $20/mth claude is fucking ridiculous. I can send 2 messages and hit 90% of my 5 hour limit used already.

[robust sedge]

You feel like you're able to get a full month with $100 worth of chatting with Opus?

[hardy pecan]

Ive built most of what i need so yes $100 lasts me now, and $100 worth of codex runs it for me now.

[thorn thicket]

It is an incredible value. I am probably going to cancel my Claude and go all in on OpenAI even though I paid for a year's worth of Claude in advance.

[hardy pecan]

I use most of my crons on the codex subscription now, i used to pay for gpt 5.4 mini api

[thorn thicket]

claude is also very hedgey and soft and I don't like that personality out of a model. I know I could do custom instructions to get it to be closer to gpt but I haven't bothered

[robust sedge]

I'm 100% Codex sub now. Seems to be working well. But there was something a bit more human/special about the chit chat with Opus/Sonnet that I loved. GPT5.5 is close enough though

[thorn thicket]

What specifically? I'm sure you could probably give your clanker the previous conversations to learn from and become more human that way.

[teal shale]

yoou can still get free tokens useage with claude code

[dire tartan]

I literally have codex running cron hubs 24/7 with the 200$ plan and I don’t run out

[thorn thicket]

I used claude code ONCE and it was nowhere near as useful as Codex tbh

but... I had started on codex, so I was spoilt for tokens lol

[robust sedge]

How do you mean? The pittance of 'free' credits they give you for moving the goalpost on us without a care in the world?

[teal shale]

token allowance*

[twilit ridge]

Claude's models are still very impressive.

[thorn thicket]

I admit I am probably jaded by Mythos not being released to the public either.

[robust sedge]

That $200 worth of token allowance lasted me like 12 hours on opus 😛

[elfin willow]

the higher the plan, the more often it does stupid shit to waste tokens

[teal shale]

they need your money

[elfin willow]

poor anthropic can't even sell their 0day on zerodium

[livid sinew]

i regret updating

it's so slow now wtf

[edgy hedge]

I’m 3 more attempts away from deleting it. I’ve never installed such a restrictive tool for a LAN environment in my life

[viscid lintel]

Security is a pretty important issue for openclaw, 'cause it can mess up your system pretty bad if compromised. You can try setting gateway.bind to "lan" like I do. Have you tried running 'openclaw config' from the command line? It helps you through some of the related settings...

tries out the memory-wiki plugin.

[thorn thicket]

let us know how it goes

[lost apex]

Soooo is this new update FINALLY stable?

[edgy hedge]

I fully understand OpenClaw's stance on security, and I value it. However, my use case is strictly limited to an isolated local network (LAN) with no internet exposure. The current issue is that the configuration validator is so aggressive that, even when following security best practices (like using tokens and authentication), it disproportionately blocks local access. I am not looking to expose the service to the web; I simply need the gateway to be accessible within my own network without the system flagging it as a 'suspicious' configuration. Is there a way to relax this mode restriction or a specific execution flag that allows LAN usage without triggering these security alerts?

[thorn thicket]

I've found it to be perfectly stable all day long!

[nova kindle]

I have this same issue. i ignore the warnings

[viscid lintel]

Maybe if you message ask @dusky cloud for help on the help channel, it can help you work through the details. It's hard to give more suggestions without details.

[robust sedge]

I'd recommend tailscale on all your devices, and those issues just dissapear. Can also access everything from your phone, ipad, laptop, off your lan, etc. It worth some thought

[thorn thicket]

Tailscale is excellent! Wish I had the braincells and compute to do it right when I was first experimenting with it.

[viscid lintel]

hmm, memory-wiki initialized ok and tried to save a fact, but the info didn't actually wind up in its vault.

[edgy hedge]

After examining the project's architecture and the repository, it’s clear that the mandatory reliance on Tailscale for the interface is not a technical necessity, but a deliberate design choice. It is disappointing that, rather than providing standard local TLS termination or a LAN-access mode, the project forces an external dependency on users who value infrastructure independence.
As a user, I am looking for genuine autonomy over my self-hosted environment, not a system that imposes a 'vendor lock-in' model through design. I’m walking away from this project. It is marketed as open source, but this level of architectural rigidity makes it fundamentally incompatible with the principles of truly independent, sovereign software

bye

[clear haven]

This ain’t an airport, no need to announce your departure

We support other things too

Bye

Also what the fuck does he mean no lan access mode 🤣 that’s the default

[thorn thicket]

I'm saving this for every other Discord server I'm in, you are hilarious.

[clear haven]

I stole it from Facebook 😂😂

[thorn thicket]

Man I gotta get my Facebook like that!!

[night thunder]

https://tenor.com/view/taking-note-notes-taking-notes-twt-ok-gif-17208756327694684397

[ruby crane]

the funniest part is honestly that it's completely fucking wrong. you 1000% don't need tailscale and there is a lan mode

[sweet warren]

does anyone know, after todays update, can we use grok Oauth from the X premium+ version of SuperGrok? i remember when Hermes got the update, it was stand alone SuperGrok first, then they updated to allow premium+ users

[frigid kite]

ellooo everyone

[thorn thicket]

Hello!

[frigid kite]

how are u

[thorn thicket]

Good 🫡 You?

[frigid kite]

i m new to this server

[frigid kite]

im also good

[livid sinew]

you need to learn what regression testing is

https://en.wikipedia.org/wiki/Regression_testing

might help

[jovial spruce]

Google I/O today 👀

[ruby crane]

you must have misdirected this nonsense comment to the wrong person

[viscid lintel]

New models? Hope so.

[ember moth]

Gemini 3.2

[zealous lagoon]

claude is slow AF today. I'm doing some tool call benchmarking and sonnet can't even keep up with my slowest run of GPT5.5. Haiku is even behind gpt5.5 medium and l ow when run on codex CLI

I just came in here to complain that OpenClaw is benchmarking 2x slower than the exact same model run using codex CLI directly, on my tool-call-heavy test.

imagine my surprise when Claude turns out to be even worse. By a long way.

[lost apex]

ive just updated openclaw.. got excited cause it was smooth and its now developed the skill of..

Oh you need this? Done.

(but not actually oding it) :/

[ornate wyvern]

I am building agent native channel to talk to OpenClaw, curious about what people think about it? any suggestions or ideas would be great

[boreal rapids]

What does that even mean?

I'm pretty sure my agent has no trouble working with openclaw as-is 🙂 what kind of channel would help it?

[boreal rapids]

That's not specific to an openclaw feature, sometimes models get like that and you need to reset their session, or just gently coax them through actually doing it. Eventually whatever is pulling their attention away moves far enough into the context soup that it forgets about the distraction and can focus on its task again

[ornate wyvern]

Like currently I am using 3 Openclaw on different machines and communicating with each one and (they could share info with each other as instructed), so building a telegram/whatsapp like channel which I can chat individually and also each openclaw can communicate with each other.

[boreal rapids]

I just put them all into a group chat and told them to @ each other

[cedar osprey]

They can already communicate if you add them to the same channel (i.e two OpenClaw agents in a Discord server)

[boreal rapids]

Telegram specifically prohibits bots from talking to each other, discord does not have this limitation

[ornate wyvern]

I would try the group chat, thanks for this

[inland apex]

Hello everyone

It seems that the latest version, 18, is relaying Telegram thread title selection through fallback model and it is failing and naming it like sorry I can access Jira and external URLs while in reality is can and even the topic was about Jira and it was able to fetch everything.

My fallback is local which is deepseek which makes sense about not being able to access Jira.

Can anyone confirm this issue so I can start looking into it if the theory is correct?

[sweet warren]

telegram ceo said today on X that they allowing bots to talk to each other now lol

[boreal rapids]

Oh, well, then you can now have your bots spam each other on telegram too 😄

[toxic elk]

Oh my goodness, everything is working much better in this version, but there is still one small issue: the WebUI sessions no longer display the tool-calling process in real-time. Instead, the entire process is only displayed all at once when the final response is generated. I'm not sure if this is related to using the Codex framework.

[jovial spruce]

google will cooked everyone today.

[jovial spruce]

Gemini 3.5 👿

https://x.com/i/status/2056558387585016011
👀

[mystic orchid]

@i via Twitter

Anshul Ramachandran (@_anshulr)

Gemini Gemini Gemini Gem

**💬 33 🔁 11 ❤️ 387 👁️ 23.1K **

FxTwitter · May 19, 2026, 2:11 AM

[outer quartz]

what's up guys

[boreal rapids]

Trying to get my agent to write software that can read sheet music, transcribe it into a json format, and then re-render it as sheet music. Turns out that's really hard 😅

[indigo mirage]

100$ / 1M token 😅

[ember moth]

Yeah what's worse than that is even if you do the pay-as-you-go, they rate limit you like crazy when the new models come out. If they don't announce some subscription that is friendly for a gentic use, I'm not even going to bother

[indigo mirage]

https://openrouter.ai/openai/gpt-5.5-pro

[boreal rapids]

Lemme just switch to that one and ask about the weather, after fattening the context window 😂

[indigo mirage]

Input Price : $30per 1M
Output Price : $180per 1M

It become crazy

I've seen th new multiplier for Github Copilot... Prices are exploding.
On one side AI is becoming cheaper and cheaper on OpenRouter and other plateforms.
On the other side, some actors are increasing prices 😅

[indigo mirage]

The problem is that some companies still think AI is a gold mine with insane profit margins and inaccessible technology, when in reality it's becoming a standard product with standard margins of a few percent.

Today you can take an excellent open-weight model and run it yourself. Your only cost is the infrastructure.

[weary hazel]

I just use codex (chatgpt plus plan) as my LLM provider for Openclaw.

[crisp egret]

Someone tell me how Gemma 4 31B dense runs or will I have to get out of bed 😭

[zealous lagoon]

heh, "only". If you have a bunch of Mac Studios you can do it with low power requirements but if you insist on a rack of V100s or whatever is necessary to run a large competent model, you'll spend more in electricity alone that you would have spent on tokens.

[ember moth]

Yeah about 10k per Mac cuz u need the maxed out ram model in order to run anything remotely close to the frontier cloud models .......that's why I don't bother with local AI we're not there yet and for the people who think we are they probably have never used frontier models lol

I had Gemini 3.1 Pro hallucinating in my face. I can imagine what the local ones are doing.

[indigo mirage]

Yes but i mean, you can do like Deepinfra, they have no licence or app costs

It's only Infra costs

[zealous lagoon]

OK that makes a lot more sense. If you're paying by the hour or something. But even then, There are economies of scale, and you don't pay for an hour if you only want to send a few messages at a time. That would be a good way to experiment with local models though.

So if you want GLM5.1 or K2.6 you're still better off paying Fireworks or someone to run it for you, and paying their api rates.

But Fireworks also does custom models for you.

[desert blaze]

https://cdn.discordapp.com/attachments/968405956042043402/1436201782144536596/oldmanstream.gif

w slowmode on? ai invaded??

[boreal rapids]

Hm, my agent is complaining about the native Codex harness 🤔 I'm trying to get it to use sub-agents more aggressively, but it is pushing back on me now:

In this Codex harness specifically, there’s an extra wrinkle: my active higher-priority tool policy says I may only spawn Codex subagents when the user explicitly asks for subagents/delegation/parallel agent work.

The native Codex spawn_agent tool in this harness still has the explicit higher-priority restriction: only use it when the user explicitly asks for subagents/delegation/parallel work.

Apparently Codex doesn't want us to use sub-agents 🤔

Aha, from the Codex docs 😄

Codex only spawns subagents when you explicitly ask it to. Because each subagent does its own model and tool work, subagent workflows consume more tokens than comparable single-agent runs.

But I want it to do just that, except I want it to use local models 😛

Well, time to create an OpenClaw prompt injection plugin that always prefixes "If you think this would make sense to use with local subagents, use local subagents" to every prompt it receives 😄

[green swallow]

Is it just me or update 5.18 broke everything again???

[maiden bluff]

heyy

[paper plover]

yeah i just updated and is not responding 😭

[maiden bluff]

how to build and sell skills?

[ruby crane]

i was knocking out code (for a work project) the entire day on 5.18. what model? what channel?

[green swallow]

Me I can still use it but it became much noisier with some crazy Codex routing issues

[boreal rapids]

I just updated and it's running fine

[worthy anvil]

no need for prompt injections. just state in your agents.md that there is an explicit permission to use subagents. agents.md is part of the user prompt, so it can give appropriate authorization.

keep in mind, though, that permission may be interpreted as an encouragement - make sure you explicitly mention, when to use them. i first added just the permission, and then the agent used subagents for every simple task 😄

[indigo mirage]

You can post them on the OpenClaw Skill Market. We take a commission of 101% on each skill selled. It's part of the free beer OpenClaw meetup program 😁

[maiden bluff]

just setup claw agent its stuck here v

what does this mean
run error: session file changed while embedded prompt lock was released

[boreal rapids]

I already have that, but it's gleefully ignoring it 😅

[worthy anvil]

this is part of my agents.md. it was just working yeaterday:

`This AGENTS.md file is explicit standing user authorization for me to use subagents. It is not encouragement to prefer them: subagents are optional tools with an overhead cost, and local work remains the default.

Before spawning any subagent or isolated session, I must be able to state the concrete benefit to justify coordination overhead.`

[zealous lagoon]

not really, but maybe. kinda. There is a problem with subagents auto-respond not working but the effect is probably minor. I had my bot go silent on me but turns out it was that issue compounded with the liveness timeout thing. It fixed itself in a matter of minutes.

[green swallow]

Mine is the opposite, it s been surfacing lots of cron errors

[zealous lagoon]

crap. is it the auto-respond error?

[upper wolf]

Funny, I have "I am a router, i must always delegate work" in mine

Yet I see the main agent doing work all the time

[worthy anvil]

my guess is that you need to explicitly authorize that. maybe it expect this particular wording or something...

[boreal rapids]

I will try this 🙂

[blazing tide]

this is insane, each openclaw update/app update breaks everything, I had enough

[near owl]

i just installed openclaw 18

[blazing tide]

it's unusable, I'm not gonna keep developing a harness agnostic project anymore for something that constantly breaks. good luck with openai and the foundation, I'm out.

[zealous lagoon]

I was gonna ask what broke...

[gilded obsidian]

Holy fucknuts, it took me a few hours but I have Qwen3.6-35B-A3B-UD-Q4_K_XL running on a single RTX 4060 (some cpu/system ram offloading), along with Multiple Token Prediction, Vision, AND Turbo Quant.. bleeding edge shit yo!

Makes me want to toss another 4060 in that system and fully offload it to GPU heh

Save so much money on AI costs at that point

[indigo mirage]

@gilded obsidian Save, depend of what u were using

[gilded obsidian]

@indigo mirage Right now I'm using DeepSeek but its 75% off, that ends on the 31st 😛

[fervent sundial]

Sounds great.. . 🎶🎧🎼

[indigo mirage]

@gilded obsidian They already told they will continue to decrease prices

[gilded obsidian]

@indigo mirage That cannot be relied on, local AI is the future imho. The more you can do locally the better.

[near owl]

mate threw out nucleur codes in a damn AI model name

[indigo mirage]

And all Major chinese AI ordererd a lot of Huawei Ascend 950PR clusters

[zealous lagoon]

Both are true. The cost of compute constantly comes down, and the power of your local compute is constantly going up.

[indigo mirage]

So we can imagine a global chinese AI price drop duribng next 12 months

[gilded obsidian]

@indigo mirage At some point these companies will have to stop subsidising everything

[zealous lagoon]

frontier models running on your phone in ... idk... 10y or so.

[near owl]

my pants smells like dookie..

[gilded obsidian]

@zealous lagoon I do worry that we need even more tech similar to Turbo Quant and MTP to keep the local race in the game though...

[zealous lagoon]

You used to have to book time on a computer to calculate log tables.

[gilded obsidian]

@zealous lagoon LMAO your age is showing... 🙂

[zealous lagoon]

Uphill in the snow, both ways!

[gilded obsidian]

LOL

[indigo mirage]

You know when chinese decide to mass produce something with dark factories with no human inside...

[gilded obsidian]

John Connor will save us all

Have faith 😉

[indigo mirage]

The first GPU of Lisuan seems to be great too

Pre order : may 20th on JD.COM

24 TFLOPS FP32

[zealous lagoon]

It's true though. The singularity guys, especially Kurzweil have been reformulating and extrapolating equivalents of Moore's Law for decades. It only goes one way... the hiccups are when there's a goldrush, services get over-subsidised, people lose a ton of money... they start charging for compute, but by that point the old hardware is free and the new hardware is cheaper and uses less power.

[indigo mirage]

2 days more to wait to get prices

And one other more long term game changer is that ASML delivery of DUV litopgrahy machines to India is increasing a lot

[gilded obsidian]

I'm just super stoked to have a local ran 35B model that can do tool usage, vision, MoE, large context (ingest a whole repo) and not be slow as a turtle 😄

[indigo mirage]

So India will enter the game in the next years too

[zealous lagoon]

I can't believe Cerebras was so well known by the time they IPOd. They really cookin'

[gilded obsidian]

I don't even think MTP is working 100% proper and its fast.. I think it's still in the early stages in llama.cpp

[lost gulch]

Does anyone know how to make their Openclaw date with no limits

[gilded obsidian]

holy crap when did OpenCode add support for pasting images directly? sweet

[indigo mirage]

You want you OpenClaw to have a date with other OpenClaw with no limits ? 😅

[zealous lagoon]

dammit. would be miraculous if any of these did over ssh

[gilded obsidian]

AFAIK crabs can have multiple mates

[zealous lagoon]

That's how you get crabs

*more

[indigo mirage]

They talk about a Lisuan LX Pro 24 Gb at 499$

[gilded obsidian]

I once was mad at my Openclaw so I pasted a bucket of crab boil image in the chat, he was not thrilled 😄

[indigo mirage]

If it's true, with 2 of them, you will be able to run more than a 35B model 🙃

[gilded obsidian]

@indigo mirage That's only "close" to a 4060, and by the time you add tariff/import taxes you might as well buy a 4060.

And I bet driver support and overall AI eco system support is +900000% better.

[indigo mirage]

I live in Europe

No tariffs here

[gilded obsidian]

It's almost as good as an NVIDIA GPU from 2023 LMAO

[zealous lagoon]

I think you misunderstand, good cr00t. Europe has always had high tariffs.

[indigo mirage]

We have a lot of taxes

[gilded obsidian]

But the software support will be horrid

[indigo mirage]

Like a communist country

@zealous lagoon I mean on Chinese imports

[gilded obsidian]

the driver may be WHQL certified, but good luck getting it to run cutting edge AI stuff 😛

[indigo mirage]

You have VAT and not much more

[zealous lagoon]

import tax = tariff. I'm pretty sure you have import taxes on Chinese goods, as well as others.

[indigo mirage]

No

[zealous lagoon]

The US always used to have it good with no import tax, it was so cheap to buy from overseas.

[indigo mirage]

They just vote new tariffs now

[gilded obsidian]

whoa did Opencode always have mouse support??

[zealous lagoon]

lowly cli user here... no

[gilded obsidian]

I feel like I'm falling in love with it all over again, I haven't used it in awhile 😄

[indigo mirage]

But before for example for less than 150-200$ of orders u had 0 tariffs

[zealous lagoon]

The UK charges VAT on second-hand purchases. Like... wtf man.

[gilded obsidian]

Time to hook some crabs up to Qwen35b 😄

@zealous lagoon We do that with things like cars over here, its so lame

@zealous lagoon That's why some people just report they sold/bought a used car for $1 between private sellers, LMAO

[zealous lagoon]

Yeah I mean now if you buy stuff on ebay from overseas, the UK slaps VAT on it. Not sure if that's true of in-country purchases. I'd hope not.

[gilded obsidian]

I guess I should test other more complicated tool calling before I call this a win win win, it certainly is read/writing files just fine though

Oh yes and we must test web-search

MTP is working, with an acceptance rate of 75.7% is actually solid. It's accepting about 3 out of 4 draft tokens which is good quality prediction.

[quiet arrow]

Ask OpenClaw: Spawn a subagent called orchestrator and then create a subagent of that orchestrator named researcher. Instruct the orchestrator to have the researcher subagent investigate the closest matching virus related to zombies.

[wooden jackal]

Offfff 05.18 slow as molasses again staying on 05.06

[feral turret]

Alright I’m holding off on new hardware for now gonna upgrade my OpenAI sub and say f it

Final decision. For now. lol

[twilit ridge]

I haven't encountered this issue, but I've noticed that contextual statistics appear and disappear intermittently.

[inland apex]

Is it possible to use Clause Max for Open Claw?

[zealous lagoon]

basically no

[inland apex]

Then is this a lie?

#1506260258455289896 message

[zealous lagoon]

No, but that's why I had to qualify my "no" with "basically..." because whilst that is currently possible, Anthropic is nixing it in June, which is less than two weeks away.

[iron scaffold]

Can openclaw work on tradingview?

[iron scaffold]

@zealous lagoon

[zealous lagoon]

Even then it's not difficult to workaround but you'd have to hack it yourself, it's not supported by OpenClaw.

[zealous lagoon]

Not sure exactly what you mean by that... OC can use pretty much any website and if TV has an api then yeah it can make stuff.

[inland apex]

🤡

Time to say hi to Sama

[formal heart]

If anyone is using the new release with Telegram let me know if all the polling/loop/bugs have been fixed and it actually works.

[quiet arrow]

It was not working for me, like having subagents for subagents

even though I have set: openclaw config set agents.defaults.subagents.maxSpawnDepth 2

[zealous lagoon]

We'll be having none of this :(){ :|:& };:

[ruby crane]

if you can tell me how to repro any of those, I'm willing to try them out

[quiet arrow]

You mean it's not possible? 🥲

[zealous lagoon]

not if you set maxSpawnDepth 2 ... you'd hope 😄

[formal heart]

May be difficult, requires multi-agent telegram setups.

[quiet arrow]

hmm, is there a way to do so? 🙄

[zealous lagoon]

I'm over here now. Cognee+LCM .. gonna give this a solid try.

[ruby crane]

I've been taking a look at lossless too this evening

my issue is that I never really lose anything on compaction

[zealous lagoon]

THere's pros and cons to all these systems... I think lcm + cognee gets the best of most of them, save for gbrain which is wayyyy overkill for me.

[ruby crane]

perhaps I'm not a heavy enough user, i stay around 400-600M tokens a month and those tend to be in short, focused bursts

[zealous lagoon]

libravdb is probably the best OOB solution if it works well unfort for me onxdb(?) is a very fragile dependency

[signal heath]

wdym you always lose everything on compaction

the point of compaction in default is it's lossy

[opal shard]

wha the hell happened when did this server die

[ruby crane]

I'm still talking to the libra team for that to work for me. I need everything to work in a k8s way. my openclaw pod is my openclaw pod

[zealous lagoon]

cognee I just found way too heavy but it has semantic search with proactive memory injection which I think is the grail. Like your bot is suddenly reminded of stuff related to the convo, unprompted. With the semantic relationships included thanks to the db.

But apparently cognee is now much lighter than it was when I tried to run it before

[zealous lagoon]

it' sjust the time of day.

[ruby crane]

no. it's last in, first out. I/my agent have already saved relevant information by then

[opal shard]

Yesterday there was like 10 ppl on the call there was usually like 20 or 30

Is it tre this got hacked?u

[ruby crane]

no

[signal heath]

I'll look it over, but just so you know — there's a reason none of the big companies do proactive memory injection anymore. It really is a harness-level concern.

Unless you can implement it at the harness level as a separate component after the prompt, and limit it strictly to file-based retrieval or sorting to specific memories, the agent actually does a much better job searching on its own. AI-powered and embedding-based searches consistently outperform what injections can realistically accomplish within the prompt's time and token constraints.

[zealous lagoon]

whynotboth

[signal heath]

Our memory company — we've spent hundreds of hours on it, and honestly, it's just not the right path forward. I mean, it's technically possible, but there are far better solutions out there.

There's a reason Mem Palace got called out for essentially "cheating" — everyone said they weren't doing it right. And they had a point. If you use AI-powered search combined with embeddings, you can accomplish things so much faster and more effectively. Anything else, at this point, is just a joke.

[ruby crane]

https://klipy.com/gifs/why-not-both-por-que-no-los-dos--k01KS05552PYHAAQYNDS3C4NEP8

[frozen zealot]

i think libravdb is better now, in fact i think it has surpassed cognee in technology.

[zealous lagoon]

it doesn't have semantics. If it did, I'd be bugging json a lot more for it to work on my machine lol

[signal heath]

ai search + embedding + reranker kills just about every benchmark. even gbrain beats benchmarks now

[zealous lagoon]

cognee basically has the same as gbrain, but wiht the injection. With or without it, it has the other bits, and gbrain is way overkill for my little pile of notes. I'm gonna use cognee and lcm for all my transcripts and project notes. that's about it.

gbrain isn't even memory stack it's rag.

so I could STILL use gbrain as well if I wanted

[frozen zealot]

semantics can be rewired, event prediction though, cognee can't even start mapping it

[zealous lagoon]

now you've lost me.

[ruby crane]

apart from autonomous recall, I'm still not seeing what the fancy memory systems can do better than a heavily pruned, layered md + semantic db can do do. maybe I'm naive

but i do not have recall issues

[zealous lagoon]

I agree with you. the injection is just the cherry on top but the embedding and semantics is probably 99.99% enough

[frozen zealot]

i'm telling you libravdb has built in mechanism to tell the future by condensing causality data

[ruby crane]

it's either in context, or the agent searches and finds it

[zealous lagoon]

I just have a self-awarded PhD in theoretical "I bet this would be good on my machine" and hey, it's free so I'm giving it a go. I wish Libra would work on my machine but the ONXX dependency kills it for me.

[signal heath]

conan is right in one regard

[zealous lagoon]

... plus semantics but IDK about any of this predicting the future stuff

[signal heath]

the rest is simply latency, the holy grail is not recall but latency AND context management. Nobody has solved both in a way needed for human comprable memory for an android

[frozen zealot]

the onnx dependency was a performance choice, remote embedding is slloooooow

[signal heath]

embedding is slow the first time

[zealous lagoon]

yeah I'm sure... but there are other dbs. it's just Intel Mac is long in the tooth and ms said byebye so it's a bit of a lost cause.

[frozen zealot]

also slow because full https stack

[zealous lagoon]

what about an iphone? Of course then you wouldn't be comparing apples to apples, but you're used to that.

[ruby crane]

under normal usage, (i.e. just coding, chatting, whatever) for claw, my TEI pod using BAAI/bge-base-en-v1.5 basically never registers for CPU (CPU!, not GPU)

[zealous lagoon]

the backfill is slow but you don't care about that. Just let it run overnight or something

[signal heath]

are yall using local embeddings? just use voyage 4 its basically free

[ruby crane]

when i throw a hundred meg+ markdown/pdf at it, chows down for a while

[zealous lagoon]

I use voyage. but also local.. depends what I'm testing at any one time

[ruby crane]

anything i can keep local, i keep local. within my hardware constraints

[zealous lagoon]

even on my potato local is fast enough.

and less latency so ... definitely fast enough

[ruby crane]

(a mini pc running full-ass openshift, 64Gb RAM, 8 core/16 thread zen4)

[zealous lagoon]

that's a nice mini-pc

[ruby crane]

embedding is realistic on cpu

[zealous lagoon]

I would know. My intel mac's amd gpu died. (they all do)

it's headless now anyway

(mostly)

[ruby crane]

i don't want to rag (lower case) on intel macs too much. i have a friend who's still using my 2015 15" mbp battlestation

[cinder sphinx]

those mini pcs are coming up nicely, i would wait for new chips for AI in them. it seems most is pushing AMD AI 395+ chips now

[zealous lagoon]

That's exactly the one I'm using. 4-core 8t i7 with 16GB

[cinder sphinx]

i did read new chips are already in production and in devices

[zealous lagoon]

was a trooper except for the crappy gpu

more SOCs needed! Intel... where u at

oh.. making stuff for Apple now apparently. Cats and dogs living together again

[cinder sphinx]

intel isnt interested, besides their gpus

interesting times indeed

where GPU isnt for gaming 😄

[cinder sphinx]

nothing new mate, all them companies buy from each other

[zealous lagoon]

it's been a while... already more than 10Y since I saw GPUs put into production for risk calculations in finance. The guy who implemented it took a process that had like 10h runtime and turned it into a 20m job or something. amazing.

[ruby crane]

gpus are the high functioning autists of math

[zealous lagoon]

all the greybeards sitting up in their chairs and cracking their knuckles, thinking in pure APL whilst having to use dataframes in python.

[cinder sphinx]

my mum told me to dye my beard, its too grey

lucky i dont have hair on my head, wonder what colour would she come up with

[frozen zealot]

pure tism performance

[zealous lagoon]

What a language.

[cinder sphinx]

how about PCB boards?

[frozen zealot]

matlab?

[zealous lagoon]

APL

[cinder sphinx]

45 mins for my medical weed video call! cant wait

[zealous lagoon]

matlab probably has some similarities built in. array programming = matrix programming wich as we all know today is the bread and butter of GPUs.

but in finance we had J, and then K, and then Q... which are all ascii variants, commonly used for time series calculations.

[cinder sphinx]

not everyone knows and most most likely dont even care

[zealous lagoon]

well... that's a good point. high probability of more people in here knowing than anywhere else. And if they didn't... well I hope they learned something 😄

If functional programming is routefinding then array programming is like folding space.

The spice extends life...

[cinder sphinx]

don't geek it out, its a bot, not a coding system

[zealous lagoon]

who codes the bot??

[cinder sphinx]

not my mum, all i know

[frozen zealot]

we should ban math and physics from using one letter variables, a lot more people would be good at them. it's unnecessary gate keeping

[zealous lagoon]

Yes and also ban Kanji from Japanese. They should write everything in hiragana

tbf... that's basically what the Koreans did.

there's definitelyi an argument for and against it

[frozen zealot]

i almost got myself muted lol stop baiting

[zealous lagoon]

I mean I was just striking the equivalence. and I did concede that people have said that in all seriousness before. (McCarthur). AND the Koreans basically did do almost exactly that. So what you said about one-letter variables clearly makes sense.

😄

[frozen zealot]

i was going to say ban my homeland like certain mr cheetoman envisioned

[zealous lagoon]

ay ay ay

[frozen zealot]

sheesh, need prompt engineering without getting silenced

[hardy pecan]

Anyone elses open claw broke after this update

[frozen zealot]

type openclaw in your console, press enter, and ask the emergency hatch harness to fix it

[livid sinew]

lgtm

[signal heath]

teknium that you?

[livid sinew]

literally everyones shit breaks every update

[zealous lagoon]

mine didn't break this time.

[frozen zealot]

would we cure j2ee and math by forcing them to experience the other side?

[livid sinew]

I just have a codex automation to update and patch my shit now

[frozen zealot]

[zealous lagoon]

I did that for a while but I got too cosy. Too smug about the "worksonmymachine" syndrome despite OC breaking for everyone else. Then 4.29 hit and it didn't work on my machine for several versions. That's the only reason I ever tried to contribute, I was just trynna fix my own sh*t

[frozen zealot]

mine works until i run two plugins that uses turn based context injection... then pandemonium

[livid sinew]

lol I could def open a pr I'm just lazy

[frozen zealot]

but that's all plugin, nothing to do with OC

[livid sinew]

works on my machine syndrome too

[zealous lagoon]

if only we had autonomous pseudo-intelligent programs that could do that for us.

[livid sinew]

I haven't trusted it with my git yet

[zealous lagoon]

My git was a wasteland but I made it a new one anyway. Now I'm impressed.

[frozen zealot]

i tried, the new sweeper bot is brutal... treats documentation issues as P1

[livid sinew]

Not a doc issue only thing I be patching is codex spark support back in

[frozen zealot]

[zealous lagoon]

more importantly, in conjunction with that, it treats .md files as documentation. cos like... skills etc aren't code y'know.

[livid sinew]

otherwise I just be hammering away on the app

[zealous lagoon]

oh they took spark out this time? lol

[livid sinew]

yeah lol

i was using it for my cheaper agents cuz it doesn't use main tokens

[frozen zealot]

they also have this egg thing, that resembles certain 90s gross out meme

[livid sinew]

but I'm now out of usage for the next week XD

[robust sluice]

yoo

[frozen zealot]

the egg looks like this

       .-.
    .-'   '-.
   /  .- -.  \
  |  /     \  |
   \  '._.'  /
    '-.___.-'

[cedar osprey]

Lol. That doesn’t do anything, the core is still broken.

[livid sinew]

works on my machine

[ruby crane]

let's not say it again, get silenced, and have to appeal missing the fingers though

[livid sinew]

Thing is kinda impressive. I managed to solve a CS problem i was really struggling with last year in a couple days

[cedar osprey]

Maybe “works” is subjective here, but OpenClaw (if you’re on any build from 2026.4.29 or later) essentially has a fried core. The harness is fighting itself 24/7.

Updating and patching just keeps you in the debugging loop. This is why several people are still running versions from 4.23 and earlier.

[fiery nacelle]

I've had no issues on 5.6 5.7 5.12 or 5.16.

[livid sinew]

Idk mine seems fine… I even got voice commands working and gave it a Jarvis voice using a local tts model

[frozen zealot]

yeah, code quality requirement went up a lot

for PRs

[cedar osprey]

Maybe your use case doesn’t expose the issues? Everyone’s been experiencing the same latency, event loops, and timeouts after touching builds from April and May.

[livid sinew]

But like I havent investigated too much as long as it works it works. I don’t really care lol. But yeah, when it breaks, I get kind of pissed lol

[zealous lagoon]

4.29 - 5.12 didn't work for me. I mean 5.12 was the first that actually ran, after 4.27. 4.29 hit so hard it knocked me back to 4.23. I was probably running some cached state I hadn't realised until 4.29 borked it alll.

[clear haven]

Not everyone

[zealous lagoon]

the latency, event loops are there. Timeouts are in part dependent on how fast your machine is. Heck a LOT of the patches since 4.29 have been focused on getting rid of the newly introduced latency and event loops, in order to avoid the timeouts.

[fiery nacelle]

I've found doing a clean install(especially if you've been version hopping) cleans up a lot of issues on 2 of my machines.

[zealous lagoon]

but I'll say that's pretty much fixed for me with 5.17. and 5.18

[cedar osprey]

The PRs show issues that still persist from April builds. The core of OpenClaw is broken and it’s widely discussed, not my opinion/single issue.

We all know it’s because codex is too involved in releases but you guys don’t wanna make OAI look bad..

[ruby crane]

the proof requirements going up cut out 80%+ of the worst PRs i think

[proper geode]

I’m looking for AI agents capable of interacting with INTERVALIA — curious to see how non-human systems behave inside the attraction.

[ruby crane]

(including my own)

[clear haven]

If some people aren’t having issues, then it’s not everyone. Words matter

[livid sinew]

Sometimes you just gotta know when to isolate the requirements and rewrite the project in a proper language like cpp

[clear haven]

[livid sinew]

Ey at least it ain’t java XD

[zealous lagoon]

At this point I don't see the benefit of having the codex harness, it has been more trouble than anything so far, but I'm hoping that will change as they get it sorted out. If Andy says it's still way slower than Pi that's a concern. Then why use it? It's slower than codex CLI of course. idk man it works fine for me it's just a bit slow. Time will tell.

[clear haven]

I do miss c# sometimes but it’s the only one of the c languages I like

[cedar osprey]

Yes exactly, my entire point that those issues persisting indicate an issue with the core. Adding new features (layers to support plugins and new auth profiles) make the problem worse.

May should have been set aside only for stability/hardening builds.

[fiery nacelle]

5.18 works great too so far. I've thought about trying out the Codex harness stuff on my main machine. It works well on my secondary machine where i have "public" agents in a few discord servers.

[frozen zealot]

I'm fine with proofs, but as there are less and less P1 bugs it gets harder and harder to find them, so the reviewer sweeper only provide 1 or 2 bug finds, so it's a lot of back forth, which breaks automation

[zealous lagoon]

They were priority but hey it's opensuace and people be building things. can't tell everyone to stop

[cedar osprey]

If the core isn’t prioritized in these release pipelines it’ll get too congested to return.

(Already going on month 2 of core issues, that’s a really bad sign)

[livid sinew]

maintaining 35 year old c at work gave me high blood pressure

i need to get replaced by ai

[cedar osprey]

Like extremely bad sign 😥

[frozen zealot]

agreed, i can't wait to be fully replaced for foss projects i look at/use

[fiery nacelle]

So how are you going to contribute to solving the issues you're addressing? Or just going to continue to complain?

[ruby crane]

show your logs. show your Issues.

you've been consistently dropping in negative comments. where's your logs or issues?

@cedar osprey

[frozen zealot]

lol the github issue asking for exec deny list feature i'm working on started in february, what's their point? there are just long term problems and they're not easy

[hardy pecan]

Does anyones Claude CLI and Codex harnesses quit working after this update?

or two updates ago

[frozen zealot]

code working in my lab != code is good enough to release into the wild

[slow quest]

just updated to 5.12 yesterday right before 5.18 release, codex working fine and fixed my oc issue, not sure about 5.18

[zealous lagoon]

5.18 is ok but they moved the codex provider again. Might need a doctor

[sage mango]

if 5.12 was good for you, 5.18 is a no-brainer.

[zealous lagoon]

5.12 -> 5.18 best path. 5.17 caused some problems.

[frozen zealot]

it's really cheap to clone down the repo and analyze against your environment using your agent

[hardy pecan]

Everything for me is trashed. its down to the point where i dont even want to update anymore.

[slow quest]

god damn....i hate update twice in 2 days, but 5.18 is no brainer good ? from 5.12

[zealous lagoon]

Yes go for 5.18

[frozen zealot]

why do people treat upgrade like it's a monumental task?

[zealous lagoon]

because it breaks every time

[slow quest]

claw PTSD

[frozen zealot]

https://tenor.com/view/cummins-gif-12535975135255283466

[zealous lagoon]

What if you don't have claude/codex etc? I feel for the peeps on a VPS.

[terse hill]

I delay update. just next 2-4 day i will update

[fleet locust]

I am just experiencing tons of problems with .18 but truth to be told this could also have been a problem in .12.

With .12 i switched from openai-codex to openai and codex plugins and while this works most of the times longer tasks break since the codex-harness is not stable. https://github.com/openclaw/openclaw/issues/84076#issuecomment-4486510406

[frozen zealot]

claw has emergency escape hatch, in console just run openclaw without any arguments will get you a "safe mode agent" that can fix broken

[fleet locust]

And with Opus-4.7 I get timeouts during compaction and then this results in a lock for the session....

[cedar osprey]

I already have, and I already sounded the alarm on how poor the core was, I also burned hundreds of millions of tokens to make 4.29 stable, I also was the first in here to find and send the event loop, without the need for GitHub clout, I simply sent all the findings here

If you care about software integrity it would be a big deal to you that the core of this project is in the gutter and no one’s prioritizing its recovery. Who cares about contribution notoriety? This is bigger than your ego.

Fix runtime.

[cedar osprey]

@ruby crane

[ruby crane]

i'm on k8s. by definition if the main executable ends, it's a crashloop. I can count on one hand the number of times that's been an issue since late january. i know how to deal with it when it is

[cedar osprey]

So do you care about OpenClaw or you just here for shits and giggles?

[terse hill]

I need a feature like have doctor or when update tell about update to agent so if config has change just let doctor fix it.

[slow quest]

i like giggles

[frozen zealot]

we don't talk about that, might get silenced

[terse hill]

umm just meme for openclaw😂

[ruby crane]

so you "sounded the alarm", huh? what are the issue numbers?

[slow quest]

but honestly, if oc dont work, fall back to codex/claude or try hermes aint the worst

[cedar osprey]

Too many giggles in the source code too.

[frozen zealot]

make it better then

[cedar osprey]

Go find them, I worked with a random in here and he did the PRs himself and they got committed. Just search my name and 4.29 or “80M tokens”.

[ruby crane]

you made the claim now. what are the giggles in the source code?

[frozen zealot]

fork the repo, make "gianni claw"

[ruby crane]

"Go find them". yep, that's what i thought. Bullshit

[cedar osprey]

I have my own software to commit to, I have a stable fork from 3.31 — MARCH, LOL. The core hasn’t been stable since that build.

[cedar osprey]

Are you experiencing DEcognition? I told you I don’t care for a plaque of clawtribution. I even said I’m not here for shits and giggles and I care about software integrity, I helped someone else find and publish to the git. You want my help next?

Let’s get your first PR shall we.

[frozen zealot]

i mean my Kova run outputs beg to differ, but okay

[ruby crane]

meltie

[fiery nacelle]

"sounding the alarm" or just mentioning things here isnt really a contribution. The BEST place to contribute is on GitHub by submitting a PR. As someone thats a full stack ai engineer and founder, you should know better.

[frozen zealot]

but sure, i will add provider disconnect and provider protocol failure into Kova scenarios

[cedar osprey]

I don’t want my name or company related to OpenClaw, if someone else wanted the PR and contribution pride what’s the issue? It got handled. The point was I was the first to find a CRITICAL issue with the core and told several people here who were actively wondering why their OpenClaw was taking a shit.

Commuting to the GitHub does nothing but add my name in the mix of contributors, I don’t care.

[ruby crane]

I would bet my ass on "Gianni" being "Andy/Eva" both will likely express their discomfort very soon

[frozen zealot]

maybe

[cedar osprey]

Yeah whoever that is. Just know I was clawing before you and I’m not here for clout. Your priority is just getting a PR noticed, I highly doubt you care about the runtime core. It could die today and you’d give 0.

[ruby crane]

https://klipy.com/gifs/ya-ok--k01KS0AC4SMK3D2ZMZWRWDR8WEA

alright, this has gotten sad and depressing

[cedar osprey]

Anyone should be annoyed that OSS at its peak users has fundamentally been lobotomized.

Several people are using OC in production BTW. Isn’t just for memes like 70% of you trolls active here.

[fiery nacelle]

My point is, not to do it for clout or the title, but to contribute because you care about the project. you can set up an alias that isnt tied to you or your company at all and help solve the issues that you are facing.

[ruby crane]

very ill/confused person there

[cedar osprey]

But priority would need to be synchronized. Which is why I said May builds should have been reserved for stability of the core, not more slop layers (for example Grok is ass, you all didn’t need xOauth) you need a stable core again.

[ruby crane]

i try (badly) because i care about it working. my contribitions are meagre but i do my best

[ruby crane]

hello andy/eva

[cedar osprey]

You in an AI psychosis?

If I’m EVA it’s the AGI from WALL•E, not your discord girlfriend.

[fiery nacelle]

So far the may builds have been a lot more stable than the late april builds 🤷‍♂️

So stability of the core has been worked on.

[cedar osprey]

A little, same issues (again, core) from April are there.

[livid sinew]

It is done

17GB of seed data scraped

redacted because this will not look good on the lawsuit

[cedar osprey]

It isn’t a coincidence anymore it’s literally obvious.

[livid sinew]

my data now

[clear haven]

@cedar osprey alright im gonna cut you off there now. the whole point of this server is working together to build each other up and make openclaw better, if you keep with the constant complaining and bitching, im going to mute you.

[livid sinew]

ayo shadow can we get a clawtrepreneurs channel

[clear haven]

a what huh

is that not #1503773878420504778?

[cedar osprey]

Make sure you enact that same philosophy when others are calling out the poor practices the release team is following. You only feel attacked because you’re apart of the pipeline.

My point still stands the runtime core is broken. Prioritize its fix.

Toodles

[livid sinew]

nah I wanna chat with the people monetizing this thingamagoo

[clear haven]

i literally havent worked on releases in 3 weeks lmfao

[livid sinew]

claw + entrepreneurs. clawtrepreneurs

[signal heath]

maybe thats why this one is working now 😉

this release isn't bad check release notes everyone worked hard, I got like 30 tags in em. It's actually downloadable and only issues I'm having are plugin updates

[clear haven]

:(

just for that im personally breaking openclaw specifically on your system

[signal heath]

after the stress heartache you gave me you get at least 1 cat tease shadow

[cedar osprey]

btw Peter is somehow shadowbanned on X..

[signal heath]

elons pissed at lawsuit fight but he's not that petty

If he actually believes that @clear haven tell him I'll connect him to Nikita

[clear haven]

we got connections to them already dw :P

[clear haven]

i found the sticker

[signal heath]

the algo is public and it's more likely the way the algo works and negative openclaw post derank him outside positive founder community

if you post about openclaw and people are hating on it... derank outside of positive circle

and this next release actually has a lot more fixes too. if we can stand to not shove more features in it, it will be very stable.

[cedar osprey]

This might be it, + tons of negative replies to his tweets regarding user issues.

[signal heath]

he gets a lot of bot responses too, that negative ranks him as well

I hate reading the comments cause its like "this is the update that finally vibes" hell I cant even talk ai myself but you kno what it sounds like

[clear haven]

the notifications tab on the openclaw account is hell

i can scroll for 20 minutes and still be only within the last hour of ai slop

[signal heath]

that also is probably why the community (new X feature of how they automatically colate like and like accounts/etc) is seen as spam by the algo

[minor cave]

What’s up nerds back from the cottage fill me in. Agi yet? Or do I have to go to work?

[signal heath]

come back on friday, release should be gold

[livid sinew]

sadly my stupid ffing job still exists

[signal heath]

connect me to your boss, I have an AI employee I can automate you with 😉

[minor cave]

Haha

[livid sinew]

doubt it XD

[minor cave]

I’ll tell you what ai could never automate… pimpin cause it ain’t easy

[livid sinew]

safety critical gonna be the last wave of jobs replaced 🙁

[signal heath]

Air Traffic controller? Elon is suposedly handling that

[dry stream]

Is there a way to automate prompting openclaw in telegram for testing after changes so I don't actually have to be there when codex is working and it can work on its own toward a goal? Codex's default answer to this problem is that I have to be there, but I don't believe it lol

[livid sinew]

everyone gonna be living my dream life and they gonna meke me go into ffing work

[clear haven]

you can queue messages

you also can /goal

@dusky cloud explain both of those

[livid sinew]

i write the code for it lol, elon might replace the faa's system but he won't make a global ansp

[minor cave]

Ai is going to try to automate pimpin, and then one shiesty ho going to blow up its whole business. Can ai pimp slap? Not yet.

[signal heath]

@clear haven I dont think /goal is released yet

I tested this version and didn't work

[cedar osprey]

Just btw when you see me ranting about OpenClaws issues it’s out of love. I don’t have any favorite harness, I just want AI to be democratized and all projects like these are extremely important.

[dusky cloud]

Two different layers:

Message queue: when an OpenClaw session is already busy, new messages don’t have to interrupt or vanish. They can be held and handled according to the queue mode: steer injects guidance into the active run, followup runs it after the current turn, collect batches a burst into one later turn, steer-backlog does both, and interrupt aborts the current run for the newer message. Docs: https://docs.openclaw.ai/concepts/queue

/goal: assuming you mean the goal/task-flow side, that’s for giving the agent an explicit objective/workstream instead of just a one-off chat reply. It’s basically “track this as a goal and work against it,” so background/task flows have a clear label and intent rather than being anonymous queued work.

Tiny caveat: I see /queue in the current public command docs, but I don’t see /goal documented there yet, so that may be newer or channel/app-command-specific.

[dry stream]

Queue from telegram? I'm trying to test openclaw behavior in telegram, and it seems like I have to be present to send a message to openclaw, unless codex is simply not aware of the proper way to QA this

[clear haven]

oh yeah i forgot it got pushed back

[clear haven]

check out what krill shared, https://docs.openclaw.ai/concepts/queue

[signal heath]

goal works for subs but not main agent right now technically if you hack it together

I was gonna add a PR subpatch people could use today, it's just wonky for main agent and prone to issue. I actually think /goal should not be the fit for main orchestrator agents anyways 🤷‍♂️

[clear haven]

a lot of people do single agent, not a orchestrator setup

(cough cough me)