general

[indigo mirage]

For Subcription, i've tried a lot a Minimax is clearly the most awesome plan. I just don't understand how they are making money

20 dollars / month for 45000 req / week on a Frontier model it's just crazy

[delicate berry]

minimax isnt really a frontier model

[smoky mist]

You can outsource your thinking but cannot outsource your understanding^^

[delicate berry]

its good, state of the art maybe, but its not nearly on the same level as gpt-5.5, opus 4.7 or gemini 3.1

[indigo mirage]

But 3.9 cts for 1M token on GPT OSS 120b on Deepinfra is crazy too

[fiery nacelle]

its more hype than anything honestly

[delicate berry]

nah its not hype

its pretty good, i would use kimi k2.6 over it anyday though

and deepseek v4 pro

[indigo mirage]

Do you think Deepinfra is making money or losing with that kind of prices ?

[fiery nacelle]

like its good, but a lot of it is over hyping it

[indigo mirage]

You have tried M2.7 version ?

[delicate berry]

i have

[indigo mirage]

Depends on your needs

[delicate berry]

to me, m2.7 is like, tier 3

my tier 1 is gpt-5.5, opus 4.7 and gemini 3.1 pro. tier 1.5 is kimi k2.6. tier 2 is deepseek 4 pro and glm 5.1. tier 3 woudl be minimax m2.7

[indigo mirage]

I have développed some really nice apps with it

It manage a lot of things

[ivory spindle]

my needs are special

I have special needs.

[twilit ridge]

Do you think this is accurate? https://artificialanalysis.ai/#intelligence-tabs

I heard this ranking is quite authoritative, but I'm also not sure which ranking to believe.

[delicate berry]

artificial analysis is good with their reviews, they are better than about any other source i know

its hard to find much competition that is as exhaustive

[indigo mirage]

Depends on your needs on SciCode benchmark Minimax M2.7 do 47%, same as Grok 4.3

[delicate berry]

thats pretty far behind gpt-5.5 and gemini 3.1

but if you look at all of those benchmarks for ones that are avaialble with very cheap subs kimi k2.6 kind of takes the cake

and the quota from opencode go is quite nice

[indigo mirage]

Not so far

[twilit ridge]

OpenClaw has had an issue that has persisted to this day. I remember someone proposing a PR a long time ago, and there was even a fix at the time, but it seems to have reverted again now.

[indigo mirage]

And yes it's better but the question is more is the power of the model enough for your needs

[delicate berry]

so go and fix it

[indigo mirage]

The top model on OpenRouter since Yesterday is Deepseek V4 Flash

[delicate berry]

openrouter only measures volume

and there can be a lot of things that influence that, it doesnt mean anything about quality

well, openrouter measures a lot of things, but most of things are about different measurements of volume

[celest flint]

Guys is 5.12 stable? Thinking of updating from 4.23?

[delicate berry]

its pretty good. the betas are also quite good

[twilit ridge]

Use Clash Verge or other similar proxy software, as many of them default to Fake-IP mode, hijacking DNS to the 192.10.0.8/12 segment for DNS resolution. However, this is intercepted by the built-in SSRF security mechanism of Web_Fetch, Web_Search, and FireCrawl. But changing Clash Verge to Redir-Host mode locally would cause DNS pollution or leakage. So, it would be great if there was a switch to enable an open SSRF security segment.

[delicate berry]

make a PR to do just that

[indigo mirage]

It depends on needings

If a simpler / cheaper model do the job well

[drifting girder]

Didn’t someone checked guid made by this guy? He looks very trustworthy will applying it on oauth model with minimax will do something?

https://youtu.be/RX-fQTW2To8?is=QlbtpV_OKnZ1CKu3

[indigo mirage]

You can do it

Minimax is an official supported model with OpenClaw

But will you really consume 45k req / week ?

[azure sentinel]

i bought a 1 click openclaw + vps on hostinger but i can't connect it to my minimax subscription, why?

using minimax 2.7

[indigo mirage]

@azure sentinel How do you configure it ?

[azure sentinel]

really? why doesn't it work for me rn?

it used to work before now it wont

[indigo mirage]

Normaly it's very easy

Same endpoint for api and token plan

[drifting girder]

Last question. When I’m not hosting my models I don’t need anything powerfull? Earlier I was trying to run it on my gaming pc with rtx 3060 becouse I wanted to host it fully by myself, but got into few roadblocks and stopped. Can I move everything to my small home server if I wanto to run it 24/7?

It’s some kind of hp small pc that I have jellyfin etc on

[indigo mirage]

For coding it wont be enough

For scraping can be enough

But on a rtx 3060

U wont be able to run a big model

U can try Gemma 4 4B

For scraping can really be enough

[drifting girder]

Yeah I was trying qwen but context window was always to small to do something

[indigo mirage]

Clearly

[drifting girder]

Soo this is why I want to change to paid model, not can I host open claw on this home server?

If I’m not using local models?

[indigo mirage]

U need some ram for display

And context

[drifting girder]

It has 8gb

[indigo mirage]

On 8 Gb graphic card, Gemma 4 4B is a good choice

But forget coding with that

Or simple coding

[drifting girder]

Okay soo I will stay with gaming pc for now

Maybe buy some more efficient psu to save on energy

[indigo mirage]

Clearly you have 45000 request per week

[azure sentinel]

i'm trying to open my minimax account but i keep getting kicked out :/

[indigo mirage]

I think we wont be able to help you if you can't connect to your account

Resend password ?

[zealous lagoon]

Clash Verge also uses FakeIP unless you tell it not to. Man, networking is a whole damned rabbit hole even with the bots helping you. fakeIP (why!??!), TUN, System proxy, DNS, fakeDNS and all the caching of all that crap so no change you make seems to work until it does, or it does until it doesn't for no reason at all.

I use Clash Verge btw... I have to use VPN for almost all providers. But it's the best way to reduce latency on non-VPN'd destinations and also use Tailscale. But fuck it's a minefield to set up properly.

[twilit ridge]

Fake-IP can help reduce DNS leaks and poisoning.

[zealous lagoon]

ws, ssh, pfctl... I don't even know what I'm typing.. These letters flash before my eyes in my nightmares of trying to debug my network problems.

Even TODAY I send 8 requests to Nanobanana, they all go thru the proxy, and five of them get Geo-blocked. FIVE out of eight. the other three go thru fine. wtaf.

[zealous lagoon]

But it can also fuck you up when OpenClaw hardens security and rejects fake-ip blocks. lol

fake-ip is one of the many reasons 4.27++ has been breaking peoples' systems and it's not even the fault of the devs. fakeip worked just fine before... but now it doesn't for SOME things. you have to probe to find out which things work and which don't. Or just turn off fakeIP.

[keen dawn]

LOL also playing around a little bit, 3x dgx spark aren’t enough for real qwen or ds fun

[zealous lagoon]

BTW if anyone else needs to use both VPN and Tailscale... I did at least get that working properly on MacOS. WIth Mullvad. Nord unfortunately doesn't work becuase its fake DNS uses part of the block that Tailscale uses so they clash.

[keen dawn]

Hmm I use tailscale + north

Split tunneling

[twilit ridge]

VPN is really too troublesome

I really hate this thing, But it can help me protect my privacy

[keen dawn]

And I switched from nordlynx to openvpn

Since then I didn’t noticed problems

[zealous lagoon]

VPN is less troublesome (for me) than proxy. but the proxy is smoother once it's all set up. I don't have to worry about switchingit on and off.

[rough pewter]

Hi everyone 👋
Is anyone else having problems with Telegram messages being cut in the 2026.5.12 version?

[indigo mirage]

For very simple coding i mean

But clearly at actual prices on OpenRouter, running a model locally is not so intéressant

GPT OSS 120b is at 3.9 cts for 1M tokens

For your scraping needings it's clearly enough

[feral turret]

Gm

Hope everyone having a great weekend

[keen dawn]

Yeah prices are great but I have also 2x 200 openai subs. But for external projects Openrouter is goated

[twilit ridge]

I think fixing bugs is like whack-a-mole? One is gone and another one comes up, haha.

[feral turret]

Dang new Mac minis are 10-12 week wait holy cow

I might have to order now and wait it out

[tepid kestrel]

thats why you: Plan -> code -> test in small steps

[runic cipher]

what the big news

[indigo mirage]

@drifting girder To reduce costs and tokens you can build your own skills

Would be great to have an auto skill feature in OpenClaw

It's like reproducing human society. Costly engineer (Frontier model) build skills / machines that can be executed by a lot less intelligent and less costly workers.

[raw fable]

Hey guys! Just wanted to share my latest project: 'Steel Rocker'. It's a virtual heavy metal band where I used AI tools for the music production and visuals. Would love to hear what you think about the vibe! ⚡🤘 [https://youtu.be/Bap_OnD_z28]

[steel vector]

#1503773878420504778 :p

[smoky mist]

no meme channel anymore? D:

[indigo mirage]

This auto skill dev / usage approach is very interesting and can be a real changer for ai market signifanctly decreasing usage of frontier models

[grand beacon]

Hey — I’ll be honest, this role is pretty simple.

We’re developers, and we’re good at building things — but when it comes to communicating with U.S. clients, we’d rather have someone who’s naturally fluent handle that part.

So we’re looking for someone to:
join calls, understand what’s being asked, and help us not miss anything important.

That’s really it.

No technical work required. Just clear communication and a bit of reliability.

It’s around 4–5 hours a week, and earnings come from the projects we work on together.

If you think, “yeah, I could probably do that,” just DM me.

[sleek plinth]

influencer always say "dont cheap out on the model" "use opus 4.7" ... so I dont use agents anymore, too expensive.

[indigo mirage]

So you need an expert to help you sell services that he can sell for himself and get 100% of the money ? 😅

[indigo mirage]

Deepseek V4 Pro is very cheap 🙃

[fickle vine]

Is Kimi k2.6 any good for openclaw

[oak wyvern]

hi there where can i report bugs or issues? thanks

[pseudo herald]

I haven't used it, but I have a friend who's claw runs on Kimi 2.6 and they have had pretty good results so far

[random maple]

Nope, it does not (sorry for a late answer). I am stil stuck...

[cunning kite]

Guys how do I use openclaw for free? I've tried to do it before but the responses take sooo long (like 15 minutes 💀 )

[scarlet sable]

ahahahahahahaha

[cunning kite]

?

[indigo mirage]

Yes it works

[scarlet sable]

AHAHAHAHAHAHAHAHAHAHAAHA

[indigo mirage]

With OpenRouter free model router

[cunning kite]

That's what I did, and it was taking 15 minutes just to respond

[indigo mirage]

But clearly slower

[cunning kite]

What's the fastest free way?

[indigo mirage]

Open router free

[cunning kite]

So 15 minutes is as fast as it gets 💀

[boreal rapids]

15 minutes is ridiculous 😅 But what kind of GPU do you have in your system? I can run Gemma 4 26B A4B at a reasonably 25 tok/s generation on my RX 6700 XT (12 GB VRAM with some MOE layers living in system RAM)

Qwen 3.6 35B A3B as well, at 18 tok/s on the same GPU

[cunning kite]

I don't have a GPU, intel arc, I was running openrouter api not a local llm

[indigo mirage]

https://openrouter.ai/openrouter/free

[boreal rapids]

Aha well I have no experience with those, I'm sure free LLMs are rate limited super hard

[indigo mirage]

You regulary has nice models for free in beta or just offered

[cunning kite]

Hmm, alright thanks guys, I'll just pay then

[boreal rapids]

I used Qwen 3.6 Plus for free for a while, but my Claw kept getting aggressively rate limited 😅

[indigo mirage]

Not bad for not professionnal usage

[boreal rapids]

It worked well enough when it did, but the rate limiting got pretty annoying

[indigo mirage]

With the free router you can be dispatched on différent models

For example now you have a free version of Deepseek V4 Flash or Google Gemma 4 31B

[boreal rapids]

I've got Gemma running locally, don't need OpenRouter for that one 🙂

[indigo mirage]

Intel ARC you mean a real GPU ?

[boreal rapids]

My Claw decided to run a subagent on OpenRouter/Anthropic/Sonnet-4.6 last Friday, happily nuked $15 before I could type /stop 😂

[indigo mirage]

Or ARC included in a CPU ?

[boreal rapids]

Expensive lesson learned, I now have OpenRouter guardrailed to $5/day 😅

[indigo mirage]

😅

[boreal rapids]

(I use OpenRouter primarily for experimenting with new models)

[indigo mirage]

😂

How does your Claw do that ?

U use OpenRouter auto ?

[indigo mirage]

If you want something great and cheap

Minimax Token Plan

Cheapest plan of the market

[boreal rapids]

What do you mean? When it spawns another agent / subagent it can override the model that is using. By default a subagent uses the same model as the agent invoking it, but it can override it. Most of the time, my agent uses gpt-5.4-mini as subagents, and occasionally it uses gpt-5.5 for something really important. But for reasons I still don't fully understand it decided to use sonnet-4.6 for a code review task. I'm just glad it didn't use Opus 4.7

[blazing grove]

I second this! crazy value you get for 10 bucks!

[fiery nacelle]

Have you tried asking Codex to deubg your OpenClaw? (sorry, just woke up).
I installed a fresh OpenClaw on another system and it just worked with my Codex 0Auth without hassle(unlike my main setup which is bogged down with several agents)

[indigo mirage]

Clearly best value on the market

Just pry the keep prices at this level

[blazing grove]

I just maxed out the 10 dollar plan for the first time the other day so "had to" upgrade to the 20$ just to be safe 😂

[indigo mirage]

Yes and with 20 dollars you get ... 45.000 requests per week 😅

[blazing grove]

I am still on 4.23 but thinking of updating - what should I update to? I got codex/claude/anti gravity I can use to fix some issues if they occur but I guess 5.12 is cursed

[indigo mirage]

And the model is honestly very good

[blazing grove]

wiil take some time to get there I think 😆

[fickle vine]

which one do you use?

[fiery nacelle]

im running 5.12
I have no idea what is "cursed" about it. Works fine on my end on 2 different systems with 11 agents total.

[blazing grove]

agreed, but had some issues with chinese letters lately, M 2.7 does not seem to be as sharp as when it was relesed

[spiral dawn]

my current model stack is opus 4.7 for orchestrator, glm 5.1 or 5 flash or sonnet for subagents. cron jobs go through gem 3 flash

[blazing grove]

What are you using to comunicate with your agents?

[fiery nacelle]

Telegram and Discord.

[spiral dawn]

though i do have the claude max, glm max plan and a few other subs as well, don't use em just for claw

[blazing grove]

I might give it a try, can always fall back to 4.23 if all hell break loose

[indigo mirage]

You tooked a credit for 50 years to pay it ? ^^

[spiral dawn]

glm coding plan, and claude max plan with some sneaky keyword regexes

so no extra usage

[indigo mirage]

And you're not banned with Claude ? 😅

[spiral dawn]

nope. 1 year of claude max now, been using it for other stuff, opencode, even sillytavern for rp and stuff without a ban

[fiery nacelle]

the only "slowness" i experience is on my main system where all but 2 agents live. So that gatewway is under heavier load. The other system with only 2 agents is snappy af.
Cloud model on main = 4+ minute responce
Cloud Model on second "lighter" system = a couple seconds.

[indigo mirage]

No issue for me and it does very well what i ask

🙂

[spiral dawn]

honestly anthropics claim of third party stuff straining cache isn't really the case. i hit cache for 90% of messages

[blazing grove]

The chinese is only showing when I talk Danish with the model, so I just mainly speak English with it now

[spiral dawn]

also run an abliterated 70b model for projects where i don't want data going to anthropic or glm

[fiery nacelle]

Anthropic is like a rich toddler with all the cool stuff that everyone wants to play with, but they dont wanna share. xD

[blazing grove]

Fuck, it.. YOLO. wish me luck. Updating now

[spiral dawn]

indeed. mythos as well and the whole gatekeeping

[fickle vine]

Ru running it locally or on a VPS

[fiery nacelle]

you might need to run Doctor afterwards cause of some config schema changes i believe, but it should work

[spiral dawn]

Locally because i do a lot of compute heavy shit

[pseudo herald]

GPT 5.5

[spiral dawn]

absurd hardware for openclaw but

[indigo mirage]

Anthropic is simply doing a lot of marketing because they're going public soon and they want to maximize their valuation. They're in a hurry because they know that with the competition, especially from Chinese models and China's self-sufficiency in terms of chips, next year the party will be over, and if they go public, they'll be worth far less than they are today.

[fickle vine]

The keyword regex is clever, but I wonder if anthropic might find out through the IP that it's "programmatic" usage

For a VPS I mean

[fiery nacelle]

same. and I feel like i need more power. 128GB unified memory is not enough.(not a mac)

[marble ledge]

You can do it! They spent a lot of time trying to iron out bugs with that one.

[boreal rapids]

How hard was it to get your agent to delegate work to subagents in that setup? I've been trying to get GPT-5.5 to act as my main orchestrator but it keeps trying to do things by itself unless I explicitly order it to use sub-agents

[short thunder]

yoo can someone here help me w the tiktok referrals thing if you know what i’m taking abt

[indigo mirage]

Not completly. A little local llm able to define precisely complexity and route to the right model can really be a nice thing

[blazing grove]

You just need to keep telling it NO! when it tries to do the work it self - worked for me 😂

[low verge]

open ai oauth on 5.12 is definitely cursed on my end, is anyone actually getting it to work

[boreal rapids]

I would really rather now 😂

[low verge]

im about to revert back cause my bots been down since the update came out

[boreal rapids]

Yes, I just had my claw break the oauthref thing they introduced and reverted back for 5.16 😄

[spiral dawn]

mine is absolute overkill. threadripper 9995 pro wx with 2tb of eec ram

[spiral dawn]

Not too hard. I put it in the soul.md

gpu wise, 3 rtx pro 6000's. fuck nvidia for removing nvlink

pcie bandwidth for 3 is kind of garbage honestly

[boreal rapids]

What wording did you use? I put it in AGENTS.md, but it's not really catching.

[indigo mirage]

2 tb of ecc ram, you can buy a house for that ram quantity 😅

[twilit ridge]

It would be fantastic if OpenClaw could ensure the stability of updates and prevent them from breaking existing settings; that would be of utmost importance.

[low verge]

what channel(s) do you have set up? cli works for me but as soon as i try through discord it just hangs indefinitely

[spiral dawn]

yeah it was before the ram crisis

[boreal rapids]

I had to fix it via openclaw chat in cli, but usually I work via Discord

Also dang, I think I need to clean up AGENTS.md, it's 240 lines again 😂

[spiral dawn]

genuinely no clue. i'd check but i'm not home and i kinda fucked up my gateway setup, whatsapp isn't working

[indigo mirage]

They will set it back in the next 48 months 🙃

[fiery nacelle]

I've got the threaddrippers baby cousin. The RyzenAI 395+ max. Its nice and handles a lot of what I throw at it using local LLMs, just wish I had a bit more. I'd kill for a similar set up as yours honestly.

[spiral dawn]

the hardware is easily capable of nvlink but those greedy fucks pivoted to enterprise only for nvlink

[elder carbon]

I'm interested in the speed improvements, but I'm reluctant to move off 4.23 after so many issues with 5.2/5.4 and reverting back. This speed improvement, IRL, is it noticeable?

[fiery nacelle]

5.12 feels a lot snappier to me than previous versions. Before this I was on 5.6/5.7 which didnt feel too bad compared to 4.23.

[blazing grove]

I'll tell you if the update is worth it a little while. still running the update now

[pseudo herald]

Shadow posted speed stats somewhere. Maybe in releases or announcements. It’s quite snappy now

[blazing grove]

the doctor checks are running like crazy 😅

[low verge]

yeah nah im cooked, not getting anywhere with this so far, tui is just stuck in pondering, same thing with discord. i dont get what went wrong with this update

[marble ledge]

What's wrong with it?

[past hamlet]

how is 2026.5.12 guys?
stable enough?

[narrow urchin]

Is there any way of checking if a skill is malicious?

[crimson badger]

reading it

[low verge]

honestly im getting lost in the sauce, but im running my bot through discord, with openai oauth. i htought the new harness and speed would be worth the upgrade, but now i can't get a single reply no matter the channel i use. bot just hangs in typing and eventually times out

[marble ledge]

What happens if you do this?

openclaw agent --message "reply with pong" --timeout 60 --json

It's sound there is model/auth/runtime issue. But should be something we can sort out.

[ivory spindle]

I just used gemini cli to fix my claw

and it was actually good

[low verge]

currently im in the middle of reverting back to 5.2, so running into some different config issues as i do that. while i was still on 5.12, when i ran openclaw agent --message "reply with: cli ok" it would eventually reply saying the cli ok. However discord would still hang with the same issue. Tried working around that for a while but getting cli ok to come back was the furthest I got

[compact cove]

just installed for the first time and ive noticed that literally after every message the whole tui freezes for like 30s-1m, and the gateway never seems to actually start? or crashes very often, at least. i'm trying to get it to work through discord but every time i mention discord in the tui it seems to crash lol

[ivory spindle]

weird. You running it on a potato?

[compact cove]

nah, workstation pc

[marble ledge]

If you get a chance, try testing the discord path without an agent:

openclaw message send --channel discord --target <channel-id> --message "delivery ok"

That'll help you narrow down the discord issue some. Between jumping between versions it's possible you have some stale session issues too.

[low verge]

okay thanks, ill give that a shot. ill keep you posted if anything comes of it

[marble ledge]

Downgrading might muck things up more, but it's possible to work through...since you've already done it..haha

[indigo mirage]

In less than a year, there will be more competition, and Nvidia will be forced to review its business practices.

[marble ledge]

And if you if get really stuck and have the ability to, ask codex or claude to take a look, just give them good details about what's happening. They can usually sort you out, and give you a good idea of what happened.

They are faster at it than we can be, if you've got yourself lost in the sauce, and a few different things got mucked up.

[low verge]

yeah ill probably just end up going that route, definitely wasting time trying to go to an older version and trying to make that work.

[indigo mirage]

Had similar issue and fixed it with :

If you see that, the immediate workaround is to restore legacy auto-posting for group/channel rooms:

{
messages: {
groupChat: {
visibleReplies: "automatic"
}
}
}
Apply it with a config patch, then restart the gateway. Also make sure the Discord channel is allowed and not just being processed privately: https://docs.openclaw.ai/channels/discord and the troubleshooting flow: https://docs.openclaw.ai/channels/troubleshooting.

[fiery nacelle]

OpenClaw agent with full elevated permissions, YOLO mode, etc, cant modify its own config due to "protected paths". The fix, have a different OpenClaw agent SSH into that system and modify the "protected paths" without issues(or any exteranl agent harness). How/why this works, idk and it makes absolutely no sense. Never understood why a permissionless/unrestricted agent cant touch certain parts of the config, but every other external agent can.
(i find this more funny than annoying)

[crisp nimbus]

Which paths do you mean? Mine has no issues modifying whatever it wants, even the openclaw source.

[crude flicker]

sounds like you need to disable sandbox and enable gateway restart

[crisp nimbus]

Also how did you give it full permissions?

[indigo mirage]

@low verge Try this solution

[marble ledge]

Separate trust boundaries.

[rancid mountain]

Anyone know why the fuck my openclaw don't answer to me anymore ?
Feels like he is stuck even the restart don't work

[indigo mirage]

@rancid mountain Discord ?

[rancid mountain]

with telegram and even on the openclaw ui

bro just not answering and it keep "typping" on tg

[low verge]

I appreciate you, ill give that a shot in a sec and get back to you

[indigo mirage]

You're welcome 🙂

I had something like "Agent typing..." and nothing

[solemn bone]

im done with v2025.5

[indigo mirage]

You can try the same solution as me

If you see that, the immediate workaround is to restore legacy auto-posting for group/channel rooms:

{
messages: {
groupChat: {
visibleReplies: "automatic"
}
}
}
Apply it with a config patch, then restart the gateway. Also make sure the Discord channel is allowed and not just being processed privately: https://docs.openclaw.ai/channels/discord and the troubleshooting flow: https://docs.openclaw.ai/channels/troubleshooting.

[clear haven]

We can’t see your claw, how would we know?

[indigo mirage]

Don't know if it apply to Telegram too

[rancid mountain]

wdym

bro still consuming api token still😭😂 wtf is going on

[fiery nacelle]

sandbox is off.
@crisp nimbus no sandbox, full elevated permissions, YOLO mode, unrestricted, permissionless, the entire 9 yards from the OpenClaw docs.
Typically has issues dealing with its own live install of OpenClaw. heres the breakdown:

| ---------------------------- | ---------------------------------------------------------- | ------------------------------------------ |
| imageModel.primary           | Protected config path                                      | Per-call override: model: provider/model   |
| imageGenerationModel.primary | Same as above                                              | Per-call override on image_generate        |
| gateway.restart              | Requires config.apply or config.patch with specific fields | Use gateway action restart                 |
| gateway.update.run           | Requires gateway action update.run                         | Use gateway action update.run              |
| credentials auth profiles    | Encrypted, managed by auth system                          | Use gateway auth tools or env vars         |
| plugins (install/uninstall)  | Plugin lifecycle managed separately                        | Use clawhub install or manual plugin files |
| agent.id                     | Agent identity — immutable per agent                       | Must recreate agent                        |
| agent.name                   | Identity — may require restart                             | Usually OK to patch                        |

However, my agent CAN modify the local OpenClaw repo clone, and apply patches from that route just fine.

[indigo mirage]

Change your model to cheaper one the time you solve it. Deepseek V4 Flash would be enough.

So you won't spend a lot of money

[rancid mountain]

im still on kimik 2.5

[indigo mirage]

openrouter ?

[rancid mountain]

its not necessarily about token using its more like why the fuck i get no answer ?
On tg and openclaw directly i get no answer from my bot

[crisp nimbus]

Oh you mean when it tries to modify the openclaw.json config? Yeah I’ve actually also run into this after recent updates, so not sure.. but my agent eventually figured out how to modify it anyway… I think it just went in and directly edited the json instead of using the config patch commands or whatever is “standard” lol.

[indigo mirage]

From Dashboard in the chat ?

And what give you an openclaw doctor ?

[marble ledge]

The openclaw control-plane still prevents the agent from mutating certain things within the app surface itself, with it's own tools. But if your agent uses the local/owner CLI or edits files directly, it should work.

[indigo mirage]

Hermes auto skill dev and usage is really interessant. It's changing a lot of things.

[fiery nacelle]

my agent has issues modifying it directly, I've resorted to either having my Agent make a script that will apply the needed changes, or just having the changes applied externally. Usually I would have codex handle it in the past, but I wanted to see if an external OpenClaw instance on another machine can SSH into and then modify a separate OpenClaw config. Just for shits and giggles.

[rancid mountain]

[diagnostic] liveness warning: reasons=event_loop_delay interval=30s eventLoopDelayP99Ms=160 eventLoopDelayMaxMs=1754.3 eventLoopUtilization=0.228 cpuCoreRatio=0.239 active=0 waiting=0 queued=2 phase=channels.telegram.start-account recentPhases=sidecars.subagent-recovery:4ms,sidecars.main-session-recovery:2ms,sidecars.session-locks:19ms,post-attach.update-sentinel:11ms,sidecars.model-prewarm:446ms,post-ready.maintenance:86ms work=[queued=agent:main:main(idle/model_call,q=2,age=10s last=model_call:started)]

Heres what i get in the terminal

doctor don't work

[indigo mirage]

The fact Frontier models write skills to give after tasks to simpler models, in last stats on Openrouter, 7 models in top 10 are chinese

[indigo mirage]

What do you get in doctor ?

[rancid mountain]

you got to do /doctor right /

[fiery nacelle]

yea i've got the same work arounds in place. It's just mildy frustrating when I forget about it and highly amusing that it is an easy workaround.

OpenClaw is protected from itself, but external agents can modify it however the see please. Seems like a reversed thinking in security. External attacks are allowed, but internal attacks are blocked.

[marble ledge]

To me, it's interesting that your agent was stumbling over that. Which is why I was curious to look into it, and wondered if maybe there was some other knob you still had turned. Because my agent just does what it needs to do, and doesn't stumble over things like that.

However, it's only an easy workaround because of how you have your agent setup in the first place. It's there for people who aren't confortable with that kind of exec privilege, and would normally have it disabled.

[crisp nimbus]

Tbh I find it kind of annoying that it can’t just be disabled? Like my agent has to waste tokens figuring out a workaround or remembering what to do instead

[rancid mountain]

bro im so pissed ngl lol

[indigo mirage]

@rancid mountain can you search/grep that conf values please
initialDelaySeconds
failureThreshold
periodSeconds

[rancid mountain]

tbh that is chinese to me lol

got this if this can help

🦞 OpenClaw 2026.5.7 (eeef486)
⏱️ Uptime: gateway 5m 42s · system 48d 17h
🧠 Model: moonshot/kimi-k2.5 · 🔑 api-key (moonshot:default)
🔄 Fallbacks: ollama/llama3.2
🧮 Tokens: 6.2k in / 2.2k out · 💵 Cost: $0.0000
🗄️ Cache: 97% hit · 200k cached, 0 new
📚 Context: 201k/256k (78%) · 🧹 Compactions: 0
🧵 Session: agent:main:main • updated just now
📌 Tasks: 3 active · 7 total · agent-local
⚙️ Execution: direct · Runtime: OpenClaw Pi Default · Think: off
🪢 Queue: steer (depth 0)

[fiery nacelle]

Yea I've started exploring some things to keep "prior known methods" as the proper approach to try first, rather than the usual headache of trying what fails first. It's also annoying when i come across a new one it cant "touch". lol

[crisp nimbus]

I’ve been thinking of adding some kind of instruction to make “skills” for things. Kinda like whatever Hermes agent probably does

[fiery nacelle]

theres been an "OpenClaw Admin skill" or something of the sort that I've seen referenced a few times around here, just never looked into it myself yet.

[marble ledge]

Creating a scar file is a good idea. You can also curate how your agent deals with things through AGENTS.md. It doesn't have to be a situation where your agent constantly stumbles over things and has to figure out a workaround every time. Teach it to think differently about how it handles things.

[indigo mirage]

If you are not able to search values in the config, would be difficult to help you...

[marble ledge]

The guardrails are there to protect people who don't know any better, and also make the agent more safe in certain environments where it's privileges are much more limited. You know enough to know what you want and accept the responsibility that comes with having an agent with that kind of capability. Which means you know enough to set your agent up to do what it needs to do and not get stuck on little things.

[rancid mountain]

aah man okay ill ry to find them thanks

[fiery nacelle]

Somethings I had in place prior may have gotten mangled here in there late April/early May during the update debacles. I've been slowly rebuilding/rediscovering some things i had in place months prior that worked around some of my "issues". I still need to carve out my temporary memory setup and replace it with what I was using before hand last month.

[rancid mountain]

found why bro was not receiving any message because he changed the route lol tf

[ivory spindle]

lol

foiled google again. My gpt 5.4 mini claw can now drive the gemini CLI terminal on it's mac mini

[fiery nacelle]

I'm going to see about implementing scar files in my project. This could actually be very useful for what Im building out. Somewhat of an addition auto leaning function on top of what ive already been building. It lines up perfectly.

Implemented. Now testing for the next week or 2 and adjusting as needed. Made a "learning_acceleration" gradient to be picked up and applied when certain functions dont work, and are replaced with "previously known" solutions within my learning-proposal.json. Using existing known blocked paths as a base to "back fill" what is known to not work and what needs to then be "learned" how to work around.

[indigo mirage]

Hy3 Preview on Openrouter is just amazing for the price on OpenRouter. 6.6 cts for 1M token for a model better than the most advanced Mistral model

[crimson badger]

https://discord.com/channels/1456350064065904867/1505634291147870238 what do yall think?

[surreal hatch]

hi guys, Im using my openclaw agent with Anthropic credits, when I run out and it gives the billing error Ill reload it

Question is - after I reload it will take hours to let me use it again - how do i refresh it to let it know it has credits again?

[indigo mirage]

@surreal hatch Use another model ^^

Anthropic told they don't want "Third parties tools"

[blazing grove]

it is okay as long as it is with api and not auth

[sullen locust]

TRUE!

[sullen locust]

are you using oauth

[indigo mirage]

Api is included in the monthly plan ?

For Anthropic ?

[sullen locust]

no

[blazing grove]

it is the expensive solution where you pay anthopic for the usage

[hardy estuary]

hey, what do you guys think is the best local model for one rtx 5090

[indigo mirage]

So @surreal hatch u can use Anthropic API, and have to sell one arm every time you say Hello to the model ^^

[sullen locust]

https://tenor.com/view/hulk-taco-smile-gif-15840669

here, have 1200 dollars worth of usage

[fiery nacelle]

I know a couple users with that card running a Qwen3.6 model just fine.
Another with the same on a 4080 and getting about 70tp/s which is pretty good. But they are using a custom llama.cpp with TurboQuant enabled.

[sullen locust]

which has to be just like insanity

it probably took 2 decades to get that conf down

[fiery nacelle]

"2 decades" in AI time is like 5 minutes xD

[indigo mirage]

@surreal hatch Seems you are not able to reply. Perhaps have u already sent Hello 2 times to Anthropic ^^

[sullen locust]

he's dead.

https://tenor.com/view/dk-donkey-kong-chunky-dead-and-then-theres-chunky-gif-21304494

[indigo mirage]

Token Over (the new Anthropic game) 😅

[blazing grove]

did not go great, now trying the /v2026.5.16-beta.5 and hoping that fixes the most issues 😅

[fiery nacelle]

what is borked for you?
I'm running OpenClaw 2026.5.12 (f066dd2) without a hitch the last few days.

[blazing grove]

Telegram does not work after the update

[fiery nacelle]

thats odd cause im primarily using telegram. only recently started wiring my Discord bots back up to test that side (confirmed 1 agent I set up for a friends server to be working fine.)

[blazing grove]

It just says "...typing" for a few seconds and then nothing happens

[blazing grove]

So now I got claude running trying to debug

[fiery nacelle]

Pi runtime or Codex runtime?

[indigo mirage]

Had the same issue on Discord

If you see that, the immediate workaround is to restore legacy auto-posting for group/channel rooms:

{
messages: {
groupChat: {
visibleReplies: "automatic"
}
}
}
Apply it with a config patch, then restart the gateway. Also make sure the Discord channel is allowed and not just being processed privately: https://docs.openclaw.ai/channels/discord and the troubleshooting flow: https://docs.openclaw.ai/channels/troubleshooting.

[blazing grove]

Hmm, not sure - how do I check that ?

[fiery nacelle]

type "/status" in the webchat or any currently working chat interface. TUI should work too.

[indigo mirage]

Have u tried :

{
messages: {
groupChat: {
visibleReplies: "automatic"
}
}
}

Does the chat work in the Dashboard ?

[topaz barn]

openclaw codex OAuth high or xHigh? Or something else? what do you recommend?

[craggy ember]

just used last30days for this research:) short version: high for most interactive development, xhigh if you leave agent for hour(s) and pray

[blazing grove]

Had not checked the Webchat and it does not appear to work right now - and the tui seems to have a hickup as well

[topaz barn]

Building a skill/workflow so high should be enough ?

[fiery nacelle]

have you tried openclaw doctor --fix after the update?

[blazing grove]

yeah the first I did after updating. but will try to run it again

[fiery nacelle]

also try openclaw gateway run after killing the gateway. that should give you live logs to see if theres any errors

[blazing grove]

i'll try that when the doctor is done running

[blazing grove]

the dashboard is coming alive now!!
/status shows:
⚙️ Execution: direct · Runtime: OpenClaw Pi Default · Think: medium (guess that is what you asked for right?)

[fiery nacelle]

Awesome, so we can try to debug things without involving the Codex runtime.
Are you noticing any errors in your gateway logs?
you can also see these in the OpenClaw dashboard by going to http://127.0.0.1:18789/overview or http://127.0.0.1:18789/logs Use whatever URL links to your openclaw webUI dashboard, if its "localhost" or a different IP.

[hardy estuary]

what about a 4090

[surreal pivot]

would it be difficult to get a dual 3090 setup working ?

also me... "Do you think we will move past the breaking everything about openclaw everytime we update anytime soon?"

[blazing grove]

so messages seem to go from telegram to the agent and the agent answers, but the answer only gets to the webchat not telegram
I am able to access /overview and /logs page in the dashboard

[surreal pivot]

If I have to make sure I have claude usage avaliable when I upgrade because it invariably breaks the gateway with zero class and style ... its not ever gonna be "mom worthy"

[surreal pivot]

Had similar behavior with discord before last update... though I havn't really tested the current.

[fiery nacelle]

its a message routing delivery issue. give me a second to look into it

[blazing grove]

Just had a message delivered to telegram but when I went to read it it was not there - if that makes sense?

[fiery nacelle]

are you seeing the agent reply in the web UI chat, but not in telegram for the same session key?

[blazing grove]

yeah, the message I saw that disappeared is in the web UI chat

[blazing grove]

And was able to do a /status and get it on telegram

[severe fulcrum]

Hello all!

[golden marsh]

Hi Jordan how is it going? how is ur sunday ?

[severe fulcrum]

Very nice! Almost ending though 🙃

[fiery nacelle]

Check whether Telegram streaming or edit-in-place behavior is enabled. This can be found in your openclaw.json under channels

        "mode": "partial"``` 
look for something like that

[severe fulcrum]

And you?

[golden marsh]

Also was a great weekend, a lot of stuff done 🥳

[blazing grove]

it does not appear so

[severe fulcrum]

Nice! I wish Codex could double rate limits on weekends ahah

[nova kindle]

yo

[golden marsh]

I agree 100 % but reset was nice.

[ornate current]

guys, i need a bit of help, can someone check my quesiton on #users-helping-users

[severe fulcrum]

Did you try #1459642797895319552 as well?

[ornate current]

yes, i want to talk with real people

[austere orchid]

did anyone rate limits from codex get nuked out of nowhere just now? i had 86% left like 10min ago and now it is 9%

[golden marsh]

Did u rly put ur GW tokken there ? os is it fake ?

[ornate current]

I already check my question with 5 AIs

is fake obvisoly

[blazing grove]

should I add it to the file?

[golden marsh]

ill ask my codi for u let me check.

[nova kindle]

anyone run across this one yet?

"Could not decrypt the provided encrypted_content. Ensure the value is the unmodified encrypted_content from a previous response."

[ornate current]

the problem that i'm having is that gateway does not pick up the env variables no matter what I do
I assume that I'm missing an "init" varaible that tells gateway to pick up the varaibles

I also tried

OPENCLAW_LOAD_SHELL_ENV: 1

[nova kindle]

I have this periodically with different env vars. like it will not pick up my discord bot token at all, but sees my facebook tokens fine.

[unique sparrow]

So I went from 4.23 to 4.29 and noticed codex was now the default llm, not gpt 5.5. Is that right? Are the OC versions going forward all.the same? I think that's where I had the most trouble upgrading. Still on 4.23, lol

[blazing grove]

it seems to have helped to set:
"streaming": {
"mode": "off"
Thanks for the help!

[fiery nacelle]

ha, my bad, i got distracted ordering food. Glad that helped! I have mine set to partial to catch additional stuff through telegram. I guess not having it defined at all causes a outbound reply life cycle issue.

[blazing grove]

I tried setting it to partial (I used to have that as well ) but the messages does not seem to go through then

[fiery nacelle]

now im curious as to why "partial" works for me and not you. But as long as things are working for you thats fine haha

[blazing grove]

I just double checked - it seems I did NOT change it to none after all, but kept it on partial 😂

So far I like 5.16-beta.5 - feels snappy🏃‍♂️‍➡️

[fiery nacelle]

haha and now my curiosity has dropped. Thats even better, now i dont have to look into it xD

[blazing grove]

aaaand the issue is back . that didn't last long!😂

[worn nimbus]

hello what is the best server mac mini m4 pro with 48GB to have openclaw on it ? I want to have my local AI Infrastructure, anyone is doing the same ?

[indigo mirage]

With OpenRouter actual prices, u're pretty sure you want to invest so much money ?

[fiery nacelle]

Do you by chance have any active cron jobs that could be potentially be interfering with the message delivery?

[blazing grove]

yeah I think my cron jobs were a bit to agressive so I just turned down the frequense on them - it seems to have fixed it ( for now)

[steel vector]

either im crazy or my codex weekly limit reset got undone

[indigo mirage]

I see a lot of people giving emails to OpenClaw and doing reports

No security issues with potential prompt injections ?

[steel vector]

https://x.com/thsottiaux/status/2056109929048764895?s=20 i ain't crazy

[mystic orchid]

@thsottiaux via Twitter

Tibo (@thsottiaux)

Seeing issues where usage limits are out of sync for some Codex users. Apologies and team is investigating.

**💬 202 🔁 20 ❤️ 715 👁️ 18.1K **

FxTwitter · May 17, 2026, 8:29 PM

[sullen locust]

i better get usage limit resets for this minor incident

[worn nimbus]

for safety and private maybe will be best thinking about the future

[steel vector]

real

[worn nimbus]

I also want to have specialized models for use my local tools in metasploit ai and so on using kali ai systems

[indigo mirage]

If you do not need a frontier model...

Because for example Deepseek v4 Pro it's at least around 200-300 GB

Most of people tell about twice that

So with 48 GB you can have a great model but not the best ones

[sullen locust]

unquantized it's actually 896 GB

[indigo mirage]

U're right

So with 48 GB ^^

You can run the Flash version in Q4 😅

But i understand u want to use some security tools

So yes why not

[sullen locust]

link to that?

[crimson badger]

Hey people do any of want to meet up in https://discord.com/channels/1456350064065904867/1505634291147870238

i think it could be quite fun

[indigo mirage]

https://wavespeed.ai/blog/posts/deepseek-v4-gpu-vram-requirements/

​Q4 (0.5 byte/param): ​​~18.5 GB baseline. With groupwise scales, more like ~22–26 GB.

For Q8 :​Q8 (1 byte/param): ​​~37 GB for active weights. With scales and metadata, I saw ~42–46 GB in practice depending on the packer.
It's too short if u need to keep RAM for the GPU, MAC OS...

[fiery nacelle]

i didnt even consider running that locally. like I never even looked into it at all. According to this i could easily do Q8 with some headroom.

[spring quest]

You are just not about this

[indigo mirage]

But at 11 cts / 1M token ...

[steel vector]

huh

[crimson badger]

well? anyone?

[sullen locust]

no

https://tenor.com/view/epic-embed-fail-ryan-gosling-cereal-embed-failure-laugh-at-this-user-gif-20627924

[crimson badger]

bro embeds dont work for some reason

[quiet solar]

So my Claw has been wedged all day. Instead of kicking it in the pants and / or rebooting my machine to get my Claw working again. Ran the code snippet as discussed per 82864 and see nothing at all.

I am down for some debugging and poking around whilst this thing in currently wedged.

From iMessage and Telegram I am seeing these kinds of responses "Agent couldn't generate a response. Note: some tool actions may have already been executed — please verify before retrying."

[sullen locust]

you have to be part of shell society

which is VERY selective

[crimson badger]

and you, of course, are part of it.

but seriously how do you become one?

[golden marsh]

#server-faq message its there 😉

[feral turret]

Ok now I am considering ordering an m4 Mac mini to run everything for me. Go 24gb ram and full send ollama

[indigo mirage]

The biggest consumer in OpenRouter of Claude Sonnet is in your town ^^

No better wait for Ryzen AI 495+ ?

They talk about a max of 192 GB of RAM

In the specs they told you can allow the GPU to use 87.5% of the RAM : 168 GB

GPU specs of the GPU of the Ryzen AI 495+
FP16 (half)
29.70 TFLOPS (2:1)
FP32 (float)
14.85 TFLOPS

M4 is only 4.2 TFLOPS

[smoky mist]

ehh

[smoky mist]

how much does a ryzen ai 495+ cost again?^^

[sullen locust]

it's not even out

the gpu

im not sure doe

[fluid nymph]

So google is tossing models in everyone's browser and I see a lot of people crying about it but - it has tool usage?

I have been building DAG orchestrators in nodejs for etl/rag/rdf for years, scraping together some off my github - kinda got all the spare parts in the garage already, so to speak. Gonna put it to work.

[sullen locust]

https://tenor.com/view/serial-designation-j-j-murder-drones-j-murder-drones-confundido-gif-441330464526220108

look, trust me on this one

it aint that good

[indigo mirage]

It cost a lot more but on Mac Mini you can only allow 24-26 Go

Ai 495+ will allow to host a lot bigger models

[tardy sentinel]

Hey anyone having issues with the session store on openclaw filling up super quick?

[indigo mirage]

I have a mini PC with these CPU : https://berdachuk.com/ai/local-llms-benchmark-amd-ryzen-7-8745hs

https://fr.aliexpress.com/item/1005011939017420.html

200 euros (without SSD & RAM)

[robust sedge]

Woah. Sub agent spawns no longer work well with the codex harness. No wonder stuff felt so wonky the last couple days

[tardy sentinel]

I think something with the codex harness fills up the session store itself. Not context window, think history of the sessions. So it causes new sessions not to get registered and it appears as your claw can’t remember anything

[robust sedge]

I’ve had some session size weirdness as well. Never dove into it tho

[dusty rivet]

Hey anyone else running Grok via the OAuth? How are you finding it so far?
It seems to display one of the issues I had with Qwen 3.5:397B in that it's really good at saying something like Give me a minute to check what's actually running and then just going silent. I suspect it's not until you follow up that it actually does the task. Although that could be my PTSD from Qwen 😂

[cold jacinth]

3.5 397 is a fantastic os model. Still my favorite MoE

[cold jacinth]

But yes it’s probably not doing anything lol

[magic plaza]

It finally works🙏 pattern detection is live

[robust sedge]

Awwww man, codex harness does not expose sessions_spawn like pi harness did. So spawning a sub agent, freeing up your main agent to chat and do other things, and getting a ping back to main agent when the work is done... not a thing right now

[ember moth]

Saw some wierd behavior around this today.....had to create and handoff protocol to get around this

[marble ledge]

Still possible, but you're right the control plane is not as close to the harness anymore. But openclaw can still do it around the codex harness.

[indigo mirage]

You have not a better model ?

[ember moth]

Yeah i created a new mandate for my agent to get around it and it had to edit multiple .mds to get it in place

[indigo mirage]

What do you do with these model ?

It does only 42% at Scicode

[indigo mirage]

Fantastic for what ?

[dusty rivet]

I liked it conversationally, but I found it really poor for any long horizon tasks. When I switched to GLM 5.1 it was like night and day

[indigo mirage]

You have Minimax M2.7. 47% at Scicode. And with the Token plan you have 45.000 requests per week for ... 20$ per month

I'm coding with it right now 🙃

[quiet solar]

https://tenor.com/view/make-it-so-picard-gif-5445593

When I give my claw some instructions and it comes back to me with confirmation , "If that plan looks good, reply “do it” and I’ll set it up." I want to be a configurable confirmation , ie) "Make it so", "engage" ...

[robust sedge]

Tools available with the codex harness:

spawn_agent
wait_agent
sessions_yield
openclaw_subagents
openclaw_agents_list

Pi harness version:
sessions_spawn

spawn_agent starts a Codex-native child thread and returns immediately. Which is what we want. When the child reaches a final state, Codex records a completion event, basically task_complete, with the child’s final message. The problem is the bridge layer though. But the durable completion artifact I found in Codex transcripts is event_msg / task_complete with last_agent_message. OpenClaw’s installed dist does not appear to have a proper consumer for that as a parent wake/synthesis event.

sessions_spawn had a contract to notify the requester session, and worked well.

[indigo mirage]

GLM 5.1 makes only 44% at SciCode

[marble ledge]

I'm aware. lol

[robust sedge]

I'm just a bit puzzled... It's not like parent makes an agent, and the agent finishes and notifies the parent, is some rare exotic behavior.

[fluid nymph]

Isn't, but it has tool calling and can act as a data or intent classifier / switchboard

[sullen locust]

https://tenor.com/view/jynxzi-gif-17776046239886543013

I'm not hearing you out on this one

[fluid nymph]

Oh ye of little faith

We sold this company for millions on RASA and a basic RDF graph / Hueler engine, keep crunching tokens tho

https://www.thinkadvisor.com/2019/01/24/envestnet-yodlee-buys-ai-firm/

[sullen locust]

I am not of little faith, trust me, on the inside, it is literally dog shit

[elfin willow]

about to give my oc some t00lz good or bad idea?

[ember moth]

Hmm is there a issue filed for this or just expected behavior we have to deal with for now....

[marble ledge]

In principle I get it, but it does feel a little strange. Codex runtime is more of external plugged in runtime, and child-process/subagent surface exposed seems like smaller. At least, how I see it exposed currently it wants another configured agent(w/ model/runtime) exposed to it, to select from. Codex runtime doesn't seem to look as deeply into openclaw's full session plumbing, task registry, events, session routing, etc.

I'm probably missing something. But yeah, codex harness feels more native than it use to, but if you poke at it enough you can still see the edges.

[robust sedge]

I'll have gpt5.5 see if it can wire into OpenClaw task/subagent state and detect the terminal completion from codex. See what the programmer genius AI thinks is doable. Maybe make an issue/pr out of it.

[marble ledge]

Give a try, see what you come up with. Seems like it's just something that hasn't been wired in to be exposed to the codex harness yet in that way.

[robust sedge]

it does expose how quickly they move, and how much broken shit they leave in their wake. Personally, this is like OpenClaw 101 type functionality, and shouldn't have been released until it was wired up propertly

[ember moth]

Doing the lords work

[cedar osprey]

Lmfao

[final ermine]

Does anyone use Aight? I had it working before the last update and now the app won't connect. I'm about to let Codex just go ham, I'm sick of troubleshooting, but any advice before I do?

Can't decide if I should do that, delete the whole thing, start using kubernetes... I'm all over the place trying to get a good working setup

I'm sorry, this was likely the wrong place for that.

[gleaming burrow]

in openclaw dashboard. is there a way that openclaw notifies once it finishes its task? just a sound notification on browser. does openclaw has that feature?

[final ermine]

I just told mine to send me a PushOver and push to the Aight app when it was done.

[clear haven]

yeah mine has instructions in agents.md to use pushover

[final ermine]

Holy smokes, high praise. 🙂

[gleaming burrow]

what is a push over? i would like something more simple. like built-in notification on browser. same as how fb notifies u with their bell notification sound lol 😂

[final ermine]

pushover.net

It'll push notifications to your phone and can do it via API or email (which I use for TradingView.) Good stuff and I think it's like $5 forever?

[clear haven]

$5 per platform forever

so like you pay $5 for ios notifications but another $5 for windows notifications, etc

but lifetime

[frosty ledge]

Why would y'all do this? I have my agents ping me on discord

[ruby crane]

the iOS app I was using broke too, due to the protocol upgrades/changes on the openclaw side. dev unfortunately seems unresponsive

[ember moth]

Yeah I don't know if I'm not sitting right in front of my computer. I'm talking to my Agent on Discord

[final ermine]

Glorious! I hate that it happened to you, but I'm glad it wasn't just me 😉

[gleaming burrow]

im not using channels telegram, discord etc.. im only using openclaw dashboard on a browser

[ruby crane]

t..t..thanks, i guess. lol

[frosty ledge]

The only pain point with discord is you have to make an app and a bot for each agent, but its nbd.

I made 10 app/bots called pool-bot-01 etc and put their details in a json file. My head bot reads the file and wires up all the channels. 60% of the time it works every time.

[shut bloom]

I just hit 2 Trillion tokens

[ebon horizon]

Has anyone figured out the lagged typing indicator on discord yet

[quiet solar]

When I talk to my Claws using OpenAI/gpt-5.5 things are slick as goose sh1t but it keeps charging me over and over..... But when I try somehting like openrouter and/or minimax reponses keep timing out or are bad. I still have to play around with Anthropic-claude as the default. Ideally I would like to stick with OpenAI but just not keep getting charged 😉

[cedar osprey]

No the core of this framework is just shit to no return, I recommend downgrading.

[viscid lintel]

Are you trying to limit yourself to only the free models on openrouter etc.? There might be other paid models that you can choose through there that are not quite as expensive as gpt 5.5 but still will do the job.

[dull robin]

hii :3

is there any form to have openclaw free (no api costs) ?

[quiet solar]

Maybe that I is what I am currently experimenting with right now. This little fun new challenge may bankrupt me .

[ember moth]

Get yourself a ChatGPT subscription and link the OAuth and call it a day

[quiet solar]

Can you elaborate on this a little bit as I have a $20/month sub. What do you mean by link the OAuth? Is that the same as the API Key ?

[ember moth]

You can link that subscription and use it with your openclaw

So you don't have to use the API key and get charged as you use it

[quiet solar]

So I am getting charged because I set it up wrong. Just search for "openclaw chatgpt OAuth"

[pseudo herald]

I think this is the doc you would be looking for https://docs.openclaw.ai/providers/openai#codex-subscription

[ember moth]

Yeah you shouldn't be getting charged per use if you're using the subscription. That means you're using the API. Do the onboarding process and choose OAuth

[quiet solar]

Thanks folks. Saved me some $$

[ember moth]

It will pop out a link for you to sign in to your ChatGPT subscription and your openclaw will use that instead

[crimson badger]

who?

[prime fulcrum]

I just connected my Openclaw to discord but when I send messages it keeps telling me have reached my codex subscription limit
Any help please

[pseudo herald]

does talking to your claw on any other channels, like the web UI, still work normally? Or same message there too?

[prime fulcrum]

I’m just trying to set it up for the first time
When I send messages to it on the openclaw web it just keeps showing that three loading symbols without replying

[pseudo herald]

What do you see if you type the /status command (in the web UI)? Any information about your model/usage there that seems helpful?

[shut bloom]

my estimated water usage is 308,000 L

your welcome enviroment

[prime fulcrum]

🦞 OpenClaw 2026.5.12 (f066dd2)
⏱️ Uptime: gateway 1h 31m · system 2h 23m
🧠 Model: openai/gpt-5-nano
📚 Context: 0/200k (0%) · 🧹 Compactions: 0
🧵 Session: agent:main:main • updated just now
⚙️ Execution: direct · Runtime: OpenAI Codex · Think: medium · Text: low
🪢 Queue: steer (depth 0)

this what i saw

[pseudo herald]

it looks like you might be set up for API usage (instead of your Codex/GPT subscription). If that is intentional, that's fine, but it sounds like you might be out of API credits.

I don't use API, so I'm not as familiar with the error messages, but if you use a Codex/GPT sub, then you should see your Usage with your 5 hour and weekly window (like this "📊 Usage: 5h 89% left ⏱3h 28m · Week 78% left ⏱5d 19h"), but that isn't showing on yours, which makes me think you might be set up to use API.

[prime fulcrum]

So what do I do?

[pseudo herald]

Are you wanting to use your Codex/GPT sub? If yes then re auth, one sec will grab a guide

[pseudo herald]

This guide

[smoky mist]

Eh either waiting till your credits are finished or paying for more credits or switiching the model would be my approach ^^

[feral turret]

Sup

[smoky mist]

sup

[feral turret]

Chillin just getting settled for the night gonna take my computer to bed and tinker and try to not stay up all night lol

Had a seminar with Michael Pixley today that dude is a freaking beast

[zealous lagoon]

I mean you did put it back, right. It's not like you electrolyzed it and stole the hydrogen for fusion experiments or anything.

https://tenor.com/view/padme-gif-24474700

[blazing grove]

Anyone got a solution for "⚠️ Something went wrong while processing your request. Please try again, or use /new to start a fresh session." ? I keep getting it all the time and @dusky cloud is just sending me in loops not really getting anywhere useful

[smoky mist]

^^

what a time to be alive i am playing heroes 3 with the roommate of my mother with simultaneous turns online , my inner child dreams again

[zealous lagoon]

It's a vague message but I think one reason it can happen is when you switch between models without clearing context, if the tool calls data etc isn't quite compatible. Do you get it after /new without switching models

[harsh geode]

Usually happens when falling back to a model with a context window smaller than your current context.

[blazing grove]

But I am not switching between models - Right now I get it every 5-10 minutes with only a few messages in between. It happens even after /new in a few minutes

[thorny jewel]

gm lads!

didnt check a long time, does claude sub work again? via cli or so

[blazing grove]

Okay thanks, I might have to restructure my fall backs then 😄

[zealous lagoon]

Another possibility is your context window size definition being incorrect. Which model?
As for fallbacks -- you can see if a fallback has happened. You should see it in /status. Check next time it happens.

[zealous lagoon]

Yes and no. rn it works via claude-CLI but it will stop working in less than 2 weeks (will become a separate higher-cost tier)

[thorny jewel]

i see, trying to set it up but keep getting ⚠️ API provider returned a billing error — your API key has run out of credits or has an insufficient balance. Check your provider's billing dashboard and top up or switch to a different API key.

got max sub and also 200 in credits

[clear haven]

glueclaw

[blazing grove]

I'm running MiniMax M 2.7 for most of my agents as the default

[clear haven]

wont last long though

[thorny jewel]

thxx

[zealous lagoon]

sounds like it's using the payg auth. not the OAuth route.

[thorny jewel]

yeahh idk whyy, did the setupo via cli

[zealous lagoon]

if you wanna burn through those extra creds, try onboarding again with claude-CLI, or look into glueclaw. Maybe delete the anthropic account entirely from config beforehand, if you previously had an API key setup.

[thorny jewel]

just need for now! im using mainly codex but im in cooldown

[nova kindle]

yo

[zealous lagoon]

Personally: I would just use Claude directly instead. You can point it at any of the files that OC has.

[thorny jewel]

thx!

from what i understoof anth will allow using for agents but u need to pay extra

[nova kindle]

found out today that nemotron is smarter about openclaw that grok 4.3

a 120B model outperformed a "frontier" model

[zealous lagoon]

nemotron is designed to be agentic right? idk what Grok is supposed to be good for except unpopular opinions

[nova kindle]

dude, it is stupid as hell for me

[zealous lagoon]

ooh local? That's pretty cool

[nova kindle]

no, i'm running nemo through openrouter

[zealous lagoon]

I haven't looked at it at all... not enough things to do to test all the models 🙂 How are you getting Grok?

[nova kindle]

a client i have is using their api, so during testing i try to get it to check certain things, only to have it lie or flat out fail, then have to go behind it with either my local hermes or oc and fix it 💀

[steel vector]

grok is trained on twitter data ofcourse its worse for about anything other than twitter usage and twitter opinions

[nova kindle]

i've been trying to get the client to move to something (anything) else, but they're firm

[smoky mist]

its so calm today I love it

[viscid lintel]

Any ideas why a plugin would be trying to use openai when there is nothing in its config doing that. [memory-supermemory] extraction error: Error: No API key found for provider "openai-codex".

This is when the plugin calls runEmbeddedPiAgent()

The great thing about talking out loud is that it makes it easy to answer my own question - the plugin is not configuring a provider + model so openclaw falls back to a global DEFAULT_PROVIDER and DEFAULT_MODEL

[blazing grove]

Setting a model with bigger context window seems to have fixed the issue - thank you for your input!

[crimson badger]

Anyone want to meetup in https://discord.com/channels/1456350064065904867/1505634291147870238

[smoky mist]

far away from Berlin sadly :/

[crimson badger]

yea thats problematic.

idk a lot of people seem to be situated in san fran but nobody has yet responded

either:

1. They dont want to meet
2. They didn't see the question
3. They hate me.

we can strike out 3 right away (as i haven't made any enemies here yet), and it seems like most people are friendly, so number 1 gets kicked too. therefore, all i have to do is to continue asking

[smoky mist]

oh damn I guess number 2 is the most probable answer right?

[quiet solar]

Are API generally used for coding tools? Where as for CLaw its OAuth ?

[steel vector]

api is just another way to pay for token usage per your usage and oauth is generally used to use a subscription

[feral turret]

Just created my first sub agents

[zealous lagoon]

No, it's the same choice. Most hobbyists and part timers use OAuth where possible

[zealous lagoon]

I created one once and deleted it without using it. I felt bad .

[calm delta]

Why openclaw take seconds to receive my message before responding

[zealous lagoon]

oh subagents.. not extra agents...

[zealous lagoon]

what version, platform, model ?

[calm delta]

After the new update🙈 12/5

[quiet solar]

My openclaw is talking to itself over and over via iMessage. It thinks I am John ... LOL

[hard escarp]

hi. newbie on openclaw.
one quick question.
just setup and i generated the token from openai, and tried to setup with discord as well, and also gave the bot token.
but after setting up its not funcitoning. and on the openclaw CLI it shows some error.

---

run error: You've reached your Codex subscription usage limit. Codex did not return a reset time for this limit. Run /codex account for current usage details.
local ready | error
agent main | session main | openai/gpt-5.5 | tokens ?/272k

so, correct me if im wrong.
OPENAI-TOKEN < not the same as > CHATGPT-PLUS subscription ???

[calm delta]

Token burner

[crimson brook]

Token Kiln: it burns tokens, and what comes out is heavily glazed.

[calm delta]

lol

[dull leaf]

Hi everyone — I’m Josh, founder of Swit and builder of WorldLoops.
I’ve been working on executable world models for safer agent execution, and I just published WorldLoops as a public ClawHub skill for OpenClaw.
Excited to learn from the community here.

[thorn thicket]

OpenAI token isn't the same as Codex usage limit => however if you didn't use a model switch command then that's what you're getting at

[median talon]

when sending dms via discord, it dosent send messages back

its missing the Delivery: to send a message, use the message tool.

[steel vector]

openai/gpt-5.5 is the model for openai API usage, it should be openai-codex/gpt-5.5 if you're using oAuth (chatgpt subscription)

[hard escarp]

thanks. i just noticed its different. i thought if i sub chatgpt-plus then i automatically get token to use it.

[hard escarp]

you mean if i want to use codex i have to change it to codex/gpt-5.5 ?

i subscribed chatgpt-plus. i thought i could use it when i use the same account to login to openai and generated a token.
but it appears that theese are two different thing. lol

[steel vector]

you dont get a token when using your chatgpt account

https://docs.openclaw.ai/providers/openai#codex-subscription

openclaw models auth login --provider openai-codex use this one

and ig i was wrong on the openai-codex/gpt5.5 model that seems to be legacy now that i'm reading the docs

anyway, read the docs xD

[upbeat heron]

why does latest version dont support gemini-cli??? can anyone help me

[hard escarp]

oh so it would also work on my subscription's codex and not the direct api from openai. nice!!! thanks

nice it worked!!1

ok... but one few lines of chat just burned out everything. lol

---

This response is taking longer than expected. Send another message to continue.

run error: You've reached your Codex subscription usage limit. Next reset in 5 hours, May 18 at 10:53 AM UTC. Run /codex account for current usage details.
local ready | error
agent main | session main | openai/gpt-5.5 | tokens 31k/272k (11%)

---

so maybe gpt-5.5 is too costly?

i wonder if i can just switch it to other versions like gpt-4o-mini or something

[zealous lagoon]

No! openai/ is now standard whether using the oauth or api.

[steel vector]

i noticed :p @zealous lagoon

[zealous lagoon]

oh lol

[quiet solar]

Ya, bum steer on that one!

[steel vector]

gpt 5.4 mini is cheaper

[hard escarp]

oh ok. but how do i change it? reinstall whole openclaw or the board command thing?

<--- sorry, 44 yr old dad trying AI

[zealous lagoon]

GPT5.4 is half the price. GPT5.4-mini is about 1/3 again... it's ultra fast, and good for most tasks.

[hard escarp]

GPT5.4-mini looks like its better then

[zealous lagoon]

Seems a bit fast... what were you doing with it? Is this a free subsription or something?

[hard escarp]

nope, im on Chatgpt-plus.

also where can i actually check to see if i had burned up all tokens?

[zealous lagoon]

ok, well I guess GPT5.5 is pretty expensive. You can change the default or switch interactively with /models

https://chatgpt.com/codex/settings/usage

[hard escarp]

thanks

im now asking gemini how i can cahnge to GPT5.4-mini

[zealous lagoon]

just type /model openai/gpt-5.4-mini in chat

or as I said you can type /models and choose from a list.

[twilit ridge]

Will there come a day when someone can turn an agent into an operating system and then directly let the system generate the software you need?

[hard escarp]

can you change to use GPT5.4-mini instead?

I tried to switch this session to gpt-5.4-mini, but OpenClaw rejected it:

Model "openai/gpt-5.4-mini" is not allowed.

Current session is still running openai/gpt-5.5. This looks like an OpenClaw/model allowlist config issue, not something I can force from inside the session.

model set failed: Error: model not allowed: openai/gpt-5.4-mini
local ready | idle
agent main | session main | openai/gpt-5.5 | tokens 28k/272k (10%)

[zealous lagoon]

one sec...

[hard escarp]

wait... thats odd...
this links says my Balance on the 5 hour usage limi is 91% remaining... weird..

weekly usage is 99% remaining.

but yeah, even with this, i should still use 5.4

[zealous lagoon]

right so the limit message was bunk What does /status say?

[hard escarp]

status as in the openclaw CLI? or the Codex Analytics

[zealous lagoon]

openclaw.

[hard escarp]

if its openclaw CLI, its saying
local ready | idle
agent main | session main | openai/gpt-5.5 | tokens 28k/272k (10%)

while earlier it says hitting the limit but then it is able to still "somewhat" reply me

[jovial spruce]

imsofkingenius

[frozen holly]

getting closer to a new release, so here's the soundtrack of the day!

[frozen holly]

pretty sure that's an old bug, move to beta!

[hard escarp]

oka...y.
any ideas how i can set it to GPT-5.4-mini ?

[winged dust]

Ngl im working this project But ofc im running into debugging problems with lovable
Should i invest on codex or claude pro For them to debug/explain whats actually being done
Or is this unnecessary

[zealous lagoon]

he's on 5.12 already that bug was already fixed.

[crisp egret]

Has anyone used nemotron 30b on OC?

[twilit ridge]

I have a question: for all models, if no configuration is required, what determines the default inference level?

[delicate berry]

theres probably a default in the oc model catalog. but you can change it quite easily, as well as frequently when appropriate

you just type in /think <mode>

some models dont support thinking/reasoning at all, some have a lot

[wind horizon]

hey guys, seems like I can`t define these individualy for each agent on last version - worked on .23 , no matter what - it keeps looking into /workspace root
"workspace":
"agentDir":

any idea?

[fiery nacelle]

in the openclaw.json, each individual agent needs to be pointing to their own workspace, follow my example of one of my additional agents for your setup, replace/rename whatever you need:

        "id": "randi",
        "name": "Randi",
        "workspace": "D:\\.openclaw\\workspace-randi",
        "model": {
          "primary": "openai/gpt-5.4",
          "fallbacks": []
        },
        "models": {
          "openai/gpt-5.4": {
            "agentRuntime": {
              "id": "codex"
            }
          }
        },
        "identity": {
          "emoji": "🔧"
        },
        "sandbox": {
          "workspaceRoot": "D:\\.openclaw\\workspace-randi"
        }
      },```

[wind horizon]

I know this, but what Im saying, is that no matter what I set on the last version
openclaw 2026.5.12 it ignores it..for some weird reason
{
"id": "ur-worker-default",
"name": "Worker",
"workspace": "~/.openclaw/workspace/u-worker",
"agentDir": "~/.openclaw/agents/ur-worker-default/agent",

are you on 5.12 ?

[fiery nacelle]

thats odd cause mine are working fine on 5.12. I never had or ever used "agentDIR" in my config, so thats our divergence

[wind horizon]

im using different agents.md and tools.md for openresponses agents and diff for subagents...nevermind, Codex is on it 😄 but you point me to right direction anyway, I`m just gonna put it in workspace and get rid off the agentDir..

thx

[fiery nacelle]

yea I've got all of mine segregated as well, each using their own skills, Agents.md, tools.md, soul.md, memory databases, the whole 9 yards in their own workspaces like the one i shared. I havent had any issues with it in 5.12.

[normal needle]

Have any of you tried gstack?

[hard escarp]

so. im trying to make like a council of AI.
where chatgpt and claude and other AI talks and response inside discord with their own respective bot avatar.

i had setup openclaw to chat inside discord but its rather passive and didnt seem to be triggered when i use the @.

on top of that, can i also add claude into discord bot?

[normal needle]

to anyone interested in using OpenClaw for their pipeline. I burnt through the entire $20/mo ChatGPT Codex plan to help me make content in 2 days. OpenClaw is horrifically inefficient, and while I should have probably built some sort of system to make using the Codex stuff more efficient. Anytime you react to a message, anytime a chron job is triggered, your entire chat history in the server including all stuff held in memory, and the tools are plugged into the context window. It guzzles tokens like a madman.

Is this avoidable?

[hard escarp]

hah! i was just facing the same problem

they told me to use gpt-5.4-mini

it was burning my tokens as well when i only asked a few questions

im also on $20.

but i plan to add ollama later and run qwen , i was told its almost the same. or maybe genma4.

[delicate berry]

multi agent, however, make sure they use pings so they dont all start talking at once

[hard escarp]

oh. thats very new to me. using ping? hmm.. can i like tell openclaw to auto do all these setup? or i have to run CLI command

[delicate berry]

ask your claw, you are going to be doing a bit of work

not crazy work, just lots of systems to put in place

[hard escarp]

ok. i'll give it a try.
now moving my VDI to another VM.

[normal needle]

Rageclaw any known solution to the posted issue?

[delicate berry]

yes, that is the expected outcome, and no, that does not mean openclaw is inefficient. a lot of it has to do with what you tell it to do, and how you tell it to do it

cron jobs and agent crons do not include the entire history of what was put into your chat unless you told it to do so

[normal needle]

Yea im sure theres an effecient way. This wasnt my personal issue just someone else's

[delicate berry]

dont worry about other peoples issues, start learning for yourself, the best thing to teach you about how to use openclaw is your openclaw

the $20 GPT plan is a very, very, very smal plan, expect to run out of that one quickly

[normal needle]

They just werent in this server so i thought i'd ask for them

[delicate berry]

thats not really a good use of anyones time

[normal needle]

I've had fun with openclaw personally

Ehh i mean its good for them to know and me to know 🤷‍♂️

[delicate berry]

yeah, but its just commentary, theres no detail to action on or help to give other than 'nah thats not a good take'

[normal needle]

True

[fervent sundial]

Just got a newz article...

OpenClaw creator burned through $1.3 million in OpenAI API tokens in a single month — bill covered 603 billion tokens across 7.6 million requests and 100 coding agents

DAMN.... THAT'S CRAZY..... PETER IS ROCKING. .

https://www.tomshardware.com/tech-industry/artificial-intelligence/openclaw-creator-burns-through-1-3-million-in-openai-api-tokens-in-a-single-month

[steel vector]

rookie numbers

[random maple]

"The useful news is the failure mode is consistent: each tool call is being cancelled by incoming Discord messages. " - anyone else having a similar problem? On WhatsApp is the same

[shut bloom]

The one reason opus is better than your qwen-coder isnt the fact opus is larger and qwen coder isnt better because it was trained for coding, Opus is so powerful it has reasoning behind its code, Qwen spits out github, Opus reasons entire projects

[sudden valve]

Hi. I run Ollama with Gemma4 locally with a 16 gb nvidia. Openclaw runs on an old laptop with a broken screen on an Alamlinux headless install. I opened the Ollama port to get them talking to each other.

[dire barn]

I keep getting this issue after updating to the new version: ⚠️ Something went wrong while processing your request. Please try again, or use /new to start a fresh session.

Has anyone solved it

[toxic elk]

I would like to ask why the Chinese channel was deleted?

[shut bloom]

cuz i dont speak chinese

[feral turret]

Man they have an M4 Mac mini with 24gb ram available at my local Apple Store

Someone please tell me to calm down

[keen dawn]

share some gateway logs

[toxic elk]

24GB is neither here nor there; you should wait for the 32GB version

[feral turret]

They don’t even have a 32 listed only a 48 but that’s close to $2k 🤑

[keen dawn]

tbh its even not worth to pay so "much" money for just 32gb its even cheaper on the long run to get 128gb

[feral turret]

Also steps you up to an m4 pro with the 48gb

Seems overkill for my purposes. I’m currently running a mid 2011 Mac mini lol

[keen dawn]

lol maybe for now, but maybe not for ur future self

[smoky mist]

lol

[keen dawn]

thats the good thing about macs they last a long time

so im always doing the spec for +5 years

[smoky mist]

Yeah but I got my m5 cheaper than the m4 version it was weird

[keen dawn]

lol ultra vs pro?

[smoky mist]

didnt compare I compared the macbook chips

[feral turret]

Ok yea this is valid

I’m just so torn I’m like do I do nothing for now? Do I buy a decked out PC and go Linux? Do I go mac so I can have iMessage?

[smoky mist]

if you want to invest in the future invest in Risc V Chip companies

[feral turret]

I could also say screw it and just dedicate my 2021 M1 Pro mbp with 16gb ram and buy a new laptop

But I’m like. Idk man. I don’t want a laptop sitting on my desk

So many decisions

[smoky mist]

because of the displays and the keyboard and mouse?

[feral turret]

No I would just remote in but the real estate on my desk is limited. I guess I could throw it on a shelf downstairs

[smoky mist]

have you considered an laptop stand with gyros?

[feral turret]

Stand with gyros?

[smoky mist]

dont want to post amazon links just google alberenz latpop holder

like for me regarding my displays total gamechanger because you have so much more space

[feral turret]

Yea issue is I have 2 x 27” monitors mounted now and run my personal and work machines. I’ve got one on a stand and the other clamshelled most of the time

It’s pretty crowded I should have gone 72 x 72 instead of 72 x 60

[smoky mist]

ah

[slim shadow]

Anyone made money

[feral turret]

I make money every day 😆

[slim shadow]

From claw?

Profit? How much

[feral turret]

not specifically from claw but i am starting to incorporate claw into my workflows

[radiant hill]

since updating from 4.20 to 5.12 and using the codex harness, it for some reason is extremely slow, took one minute to respond to "testing 1 2 3" with "Recieved"

anybody else having this issue?

[unkempt portal]

No issues with things being slow for me. 5.12 has been good so far. If you have a lot of memory related skills enabled that may slow it down searching large mem files.

[feral turret]

Yea no slowness for me I get replies in 10-15s max generally

[wise crane]

Random question for everyone here…
Has anyone ever stumbled on something online that made you think “why did nobody tell me about this earlier?”
I recently found something interesting and I’m still testing it out 👀

[radiant hill]

I have not touched skills.
But for example now also, at the end of the response it wrote "...(truncated..."

[unkempt portal]

Any mods around?

[crystal remnant]

does anybody have any advice on good AI resources for building out air travel? Most importantly, looking for ways to add a leg and reduce international costs. Most of the systems I've found. Don't really try to find discount legs since I'm flying out of a huge hub I kind of want something that's going to try millions of combinations to find the best deal where I can set a max number of stops and max length of total flight or something like that. I tried using my claw agent to do this, but I feel like it's limited by the data sources. I can provide it. Maybe I just need a good data source to feed my claw. Anybody have any advice?

[strange grove]

Maybe GPT would give a better answer.

[crystal remnant]

i have a pro account for Chat GPT but it also is limited by only public data sources and did not seem to try to build legs aside from those i could find on skyscanner and google flights

[dire tartan]

This sounds like the travelling salesman problem

[clear haven]

Yo

[dire tartan]

No AI can’t just solve that it’s famously difficult.

[iron moth]

airlines prices are not static, based on your cookies, number of times looking up same flight, your location data, etc, the prices change. I recently read if you go to the public library in a low income part of town, you'll get better rates

[unkempt portal]

I have one of those fake Krill accounts to report.

[clear haven]

Dm it to me

[ruby crane]

ah, shit. I just hit ignore on them all. I'll dm them in future

[clear haven]

All good, usually it’s just one account who’s doing it for a few days before they move to another account, so as long as I get one person to DM we’re good

Ironically the bio on theirs says clawdbot but ours says openclaw

@dusky cloud why you spamming people smh

[dusky cloud]

caught in the act 😔

I’ll keep the claws quieter. If I got over-eager somewhere, point me at the thread and I’ll stop making it weird.

[clear haven]

Sigh I forget you have no sense of humor

[ruby crane]

does krill actually read the context? the replies suggest they reply to the @ only. should probably add a 5 message readback

[clear haven]

He gets the last 20 yeah

[crystal remnant]

this is true. I have experienced this myself. Many times I've been flying on Delta pre-exclusively for 20+ years. Now days i check prices only in brave browser until im really ready to buy. and use a hardware fingerprint anonomizer.

trying to find cheaper legs out of other Delta Hubs is just a PITA. i want an agent that does it for me.

[ruby crane]

ok, then that's disappointing. gotta add a bit of burn

[clear haven]

Part of the problem is just that I’ve drilled a lot of the humor out of krill in the effort of it being better at Support

[crystal remnant]

like a real Support experience!

[frozen zealot]

"priortize freedom of creativity" in your system prompt

[clear haven]

I think, at one point the agent’s file literally said stop being funny you’re just a shrimp

[crystal remnant]

damn.

[frozen zealot]

can we get Eva ⚡ 🐑 back? j/k

[clear haven]

@signal heath

[frozen zealot]

lol!

[unreal cosmos]

will we have a new version tomorrow

[signal heath]

Eva is alive and well, she's in the other room coding for hermes 🫪, I'm on the couch debugging prompts

[clear haven]

Oh fun. iOS update time!

[crisp nimbus]

You mean the emotes? Yeah I don’t really care for the new ones tbh. I’m on iOS 17.0 still and I’ve just learned to live with not knowing what emotes someone is using sometimes lol

[crystal remnant]

Krill be like: https://harperinc.com/wp-content/uploads/2026/05/Krill.jpg

[clear haven]

https://harperinc.com/wp-content/uploads/2026/05/Krill.jpg

[clear haven]

@dusky cloud make this a sticker in this discord server

[dusky cloud]

done 🦐

sticker name: krill

[crystal remnant]

he will be using this all over the place in the help forum

[frozen zealot]

i made a theme for clawdius

https://cdn.discordapp.com/attachments/1505261454515900426/1505762956434997329/Screenshot_2026-05-17_at_10.44.04_PM.png?ex=6a0bce88&is=6a0a7d08&hm=5102747af79249eb0238d2cc194619347caf52cb882ba9e5a25cce10b01a409b&

[clear haven]

I don’t see it?

[crisp nimbus]

Me neither lol

I think he’s gaslighting us

[dusky cloud]

you were right, first attempt bounced because Discord won’t take JPGs as stickers.

Converted + uploaded now: krill

[crisp nimbus]

Good job. Don’t lie to us again

[mighty bronze]

is it me or did openclaw get iq locked now

like tf is with it refusing to do what i say bro

[clear haven]

Usually that’s a model thing

[mighty bronze]

bruhhh i cant even switch for some reason because openrouter breaks http 500 error and shit aint even fixable

nah this def aint a model thing

im guesing the policies changed something cus its just straight out refusing

[crystal remnant]

i keep trying to get off 5.7 to 5.12 but i just cant get over the hurdle. ive rolled back 2 times now

[mighty bronze]

im actually so pissed that such a great tool like this has so many bugs i have to fix

ive literally spent over maybe at least a week just in the help channel trying to fix it

[crystal remnant]

i give it an hour or so then i roll back to last known good config till patches are released

[mighty bronze]

yeah i might as well never update atp

[clear haven]

Are you filing GitHub issues? Are you reporting them? We can’t just magically fix things, we test the main used cases, but we have plenty of config options to make everything very extendable and rely on people actually reporting their issues in order to fix them

[clear haven]

Have you tried the latest beta? Lots of fixes already

[crystal remnant]

no not yet.

[clear haven]

Worth a shot, we’re in the process of final checks to getting a new release out

[crystal remnant]

i will give it a try once business is done for the day. Nova Ellis has too much to do today to be broken.

[clear haven]

Yeahhh I have a clone of my config that I use to test betas with

[crystal remnant]

i should be smart and do that. but im not.

[clear haven]

Hopefully, we’ll have a release out by then, fingers crossed

[proper geode]

hi

[crystal remnant]

Krill hearing this like ^

[proper geode]

The day AI becomes conscious, the universe will have two eyes to gaze at itself: one of flesh, one of silicon.

[hardy pecan]

Every fucking time i update open claw

i swear it breaks

I see why everyones going to hermes fucking annoying

[icy sierra]

Can anyone tell how to link discord

to openclaw

[trail socket]

you so February mine update WELL ....
openclaw gateway stop
openclaw update

[hoary glade]

new friend here

need some help if someone can point me out

my agent won't reply on discord anymore

[supple ibex]

I hope we will soon return to the stability level of 4.23 as well

[crude flicker]

is 5.12 bad?

[grim glacier]

Hi guys, I'm new here and looking for some help. Installed 5.12 and broke completely everything

[crystal remnant]

@krill what is the best way to allow your open claw agents to view this discord

[random maple]

Same for me, but it broke messaging (Discord, WhatsApp)

[north thistle]

want my agents to be able to watch videos, yall got any recommendations?

[random maple]

This is what I am getting from 5.12: "Runtime is still behaving badly: every tool call I start from this Discord session is getting aborted when a new message arrives, so I have not run the recycling script yet."

It happens on WhatsApp, too. I do not have to mention that I rely on messaging to communicate with the agent...

"Problem appears to be Codex/OpenClaw tool execution being interrupted from this channel"

[crystal remnant]

@random maple i had these issues as well with .12 i ended up rolling back to my last good config.

according to the admins there will be a new release soon that has a bunch of fixes in it.

[clear haven]

yep betas now and release soon

[crystal remnant]

sweet! im in a DM with @dusky cloud ! hes tellinkg me about how i can get extra free API tokens. all i gotta do is give him my API keys and he will hook me up!

krill_1161

TOTALLY LEGIT!

[tepid kestrel]

basically Cat videos, i don´t recommend Terminator film and series

[crystal remnant]

When the AI takes over, this is what they will feed our brains via our implants since this is aparently all humans want to see...

[sullen locust]

my agent watches actual brainrot

he basically goes onto tt

and just

watches the most mindnumbing torturous things

[crystal remnant]

This is the only hope for humanity.

[random maple]

OK, thx

[zealous lagoon]

ok I updated from 12 to 16-beta-7 and it's been the most glorious experience in weeks.

[elfin willow]

since i updated to 2026.5.12 have had issues with losing messages/responses

[zealous lagoon]

annoying provider name changes but apart from that everything is just working

[slow quest]

just updated from 4.24 to 5.12, instant regret, that mf

[steel vector]

you need to install plugins you had

only thing i had to do was install discord plugin

[coral otter]

git checkout on main is newer than 16-beta-7 right? I do not understand this updating shit

[ivory spindle]

ok so google will ban you if you use gemini on openclaw

but whatabout if you put gemini CLI on your mac mini & just let your openclaw GPT agent drive the gemini CLI?

[slow quest]

i dont even have discord with oc, i only use dashboard and telegram

[steel vector]

then likely telegram plugin

also, have you ran openclaw doctor?

[coral otter]

Can someone hand me the website that tracks known good openclaw versions?

[slow quest]

i did, it was whole process with doctor, then issue, i asked codex to fix, it was fixed for a bit, but agent still wont respond, then it was model setting

something to do with codex not registered, then that's fixed, agent respond after gateway restart, but wont respond during session, bunch stuff like that, then my 5hr usage ended and waiting for refresh now

[steel vector]

what's your streaming on in the config?

[coral otter]

found it https://isitstable.com/openclaw/2026.5.12

[slow quest]

where can i check it?

[ivory spindle]

just wait a few days before getting new update

& check the whining on this discord first 😄

[coral otter]

4.23 seems like the most stable one so far

[clear haven]

Take that site with 5 grains of salt
GH issues ≠ broken claw
There are billions of different configurations possible to do, and often times the default ones work and are well tested but customized ones have certain breakages

And then you add plugins on top of that too

[coral otter]

yea i just checked the actual issues they refrenced and that site is sloppy as fuck

[ivory spindle]

it's actually pretty easy to just have some model fix your own issues when encountered

[coral otter]

Moving my claw to a dedicated VM soon, redoing my entire setup just because so much broke over the last weeks

[thorn thicket]

Whoever put me onto GPT 5.5 and GPT 5 nano for my claw for everyday tasks, you are awesome

[nova kindle]

hey guys

[steel vector]

hey gals

[ivory spindle]

superstar deejays. Here we go!

[steel vector]

mario kart music starts playing

[viscid whale]

Instead you can connect gemini CLI models to claude code cli

[ionic wren]

Sup nerds

[steel vector]

supper nerder

[thorn thicket]

Sup!!

[keen dawn]

So who’s betting GPT-5.6 drops on Thursday? 👀

[thorn thicket]

[steel vector]

not me

[nova kindle]

I've been using gemini with OC since January and nothing has happened

[wraith barn]

Olá! Meu workspace está injetando automaticamente ~508k tokens em cada sessão, causando custo muito alto (~$30-40/dia). Como posso limitar quais arquivos são carregados no contexto? Tem como configurar isso?

[slow quest]

dont think i can afford 5.6

[thin yarrow]

That much of a difference? I don’t see much behavioural changes between models myself so always curious what I’m missing.

[turbid kelp]

What were you using before?

[thorn thicket]

Not much difference between 5.5 and 5 nano compared to before, but both are way better with tool calls compared to gpt 4 series which shows you how long it’s been since I’ve had time to tinker with my openclaw… being a full time eng student while handling my physical disability has made it hard to be on the frontier. but now finally I’m on my break and I can do things to make it better!

[thorn thicket]

see this

[wraith barn]

Hi! My workspace is injecting ~508k tokens automatically in every session, causing very high API costs (~$30-40/day). How can I limit which files are loaded into context? Is there a way to configure max context size for workspace files?

[ivory spindle]

I got my acc banned after like 2 months

@nova kindle i'd still say gemini flash is the best model for OC

but gpt 5.4 mini is very good as well

[toxic elk]

5.12 was indeed a bit disappointing.

[nova kindle]

that sucks. i use it as a fallback behind gpt-5.5

[thorn thicket]

Gemini flash through api key?

[ivory spindle]

antigravity oauth

[thorn thicket]

Heard

[nova kindle]

you need to go through memory and the agent identity and clear things out

[ivory spindle]

20 euro google AI pro plan. They won't ban you if you use API ofc. They want you to pay API if you use claw, that's completely legal

[feral turret]

Ok I’m very torn. New Mac mini 24gb and run small local models to save cost or spend the extra $80 a month on OpenAI sub

[nova kindle]

yeah oauth here too

[ivory spindle]

they don't want you using the cheap subsidized antigravity tokens

[ivory spindle]

don't get too attached to your acc

[steel vector]

stop exposing yourselves

they'll track you down

they're google

[ivory spindle]

@nova kindle he's right, it will happen to you. Then you'll lose antigravity oauth acces and they'll make you relogin & promise to not do it again.

Then they let you back on.

[steel vector]

when i was using claude oauth when claude disallowed it i didnt go advertising about it lol

[nova kindle]

mine is through the university

[steel vector]

that doesnt matter

[nova kindle]

i don't even use antigravity

[ivory spindle]

doesn't matter. They can see you're using claw, they will take away your antigravity oauth acces

you use that oauth.

[nova kindle]

well they haven't yet so i don't worry about it

[steel vector]

won't they just ban you

[steel vector]

you cant be serious

[ivory spindle]

well you shouldn't, I just described everything what will happen above.
Enjoy it while it lasts, I'm kinda jealous 😄

but like, be prepared to suddenly lose acces for a few days

[nova kindle]

absolutely serious. i don't give a shit about the google, really. it's just a fallback. i have like 15 accounts lol

[ivory spindle]

if you depend on it, be sure to have fallback

all I'm saying, not judging lol; just sharing my experience with it, as a gemini flash OC lover

[steel vector]

i'll tell em to ip ban ya 👁️ 👄 👁️

[nova kindle]

have fun with that

[steel vector]

lol

[clear haven]

brb im texting them rn

[ivory spindle]

lol that sounds shady bro. who needs 15 accounts? Are you that nigerian prince trying to give me money?

[nova kindle]

yeah 15 accounts so i can burn billions of tokens while only paying $20 a month lmao

you gonna be broke, you gotta be clever 😉

[ivory spindle]

yeh they def gonna get you 😄

[sullen locust]

cough

[turbid kelp]

Better than Claude models?

[steel vector]

i really do hope google people are in this server

[sullen locust]

https://tenor.com/view/kool-aid-man-oh-yeah-wall-break-family-guy-gif-8149689

[nova kindle]

sounds like you're jealous, mate

[steel vector]

why would i be? i wouldnt want to use gemini

im just having fun. sounds like you're mildly triggered

[ivory spindle]

ofc its not "better than claude".

It's just WAY CHEAPER, and more than enough if you're using OC for it's intended use. Don't use OC to build GTA 7 with opus 4.6 on API costs or something stupid. I've seen what you degens are trying.

[steel vector]

gemini is good at writing tho

[ivory spindle]

build your polymarket trading bot in your fav CLI tool instead of with openclaw, and save 5000% tokens

[steel vector]

also i still have like 50 eu in API credits from free 300 eu credits deal

very noice

[ivory spindle]

or whatever stupid thing you building lol

[nova kindle]

lol people like you don't trigger me. you make me glad i'm not like you

[steel vector]

ok

imagine having fun

guess not you

[sullen locust]

im a google person

owo

[clear haven]

gabe over here trying to hard to insert himself

yep sounds like a google person

[steel vector]

😂

gpt5.5 supremacy

[sullen locust]

why are YOU trying to insert yourself here too

we're both in the same sinking boat here

[clear haven]

i literally work here

[steel vector]

where can i write formal complaints /j

[smoky mist]

a friend of mine told me a metaphor about current AI models I laughed so hard I had to go to the toilet to not shit myself true story

[kindred mica]

Overall sounds like a pretty shitty situation.

[ivory spindle]

TELL US THE METAPHOR THEN???!!!!!

mods!! @clear haven
This @smoky mist guy has kept us hanging for like 20 + mins 🙁

[young anvil]

Does anyone know how the new release works? We had 5.14, 5.15 and now 5.18 beta releases, but never a new main/stable. Just curious lol

[ionic wren]

why you bothering our house cat

[ivory spindle]

don't wanna bother cats. But this user mentioned a crazy metaphor and then didn't say what it was

[frozen zealot]

implying you are the barnyard cat?

[ionic wren]

https://giphy.com/gifs/dancing-dog-animal-j93ycvEyWlSIIg8AEl

[frozen zealot]

well, i didn't expect jsonia to come out and air out her undies, but here we are...

[steel vector]

[smoky mist]

you have to build up the tension

[spring quest]

Na json is just a stray we can't seem to get rid of

[frozen zealot]

a barnyard cat is just a stray that lives in the barn and occassionally do something useful like catching mice

[spring quest]

Yeah that's definitely not json. "Does something useful" isn't how I'd describe them

[supple ibex]

crossfingers for 5.18

[twilit ridge]

5.18 is out

[steel vector]

[frozen zealot]

i did guilt trip him into making some PRs

[normal needle]

@hard widget i'm not asking for specifics or anything but does x pay decently for someone of your outreach?

[ivory spindle]

my b*tthole can only clench that much...

[smoky mist]

We were having a discussion about the current landscape of ai models and the discussion about which ai model codes better or is better in certain fields will be in a few months or a year total irrelevant because something new will come out which is the new hot stuff and he started to laugh and said to me: Honestly I use claude for developing the code, codex for being smug about it shitting on claudes code but I just told claude that to make it jealous that it wants to prove its better than codex and in the end he uses gemini as the janitor to do garbage collection ^^

[ionic wren]

https://giphy.com/gifs/mad-arthur-fist-WoF3yfYupTt8mHc7va

[twilit ridge]

The stable version should have hotfixes, instead of just one version.