general

[obsidian mica]

its ok i can do it manually i think but its a hot mess

[ruby crane]

no-one testing betas means releases ship broken

if you're here and you're pink especially, you're equipped to make the releases better for everyone, including you

[delicate berry]

im testing 5.3-beta.2 right now, my prod server is getting in some kind of a loop that looks like the one yesterday from sessions.list, i have a lot of sessions, but im troubleshooting still

[cold jacinth]

not at 2 AM.... mmmmkay?

[delicate berry]

acpx and discord plugins were really unhappy, doing a force reinstall

[echo valley]

Just updated to 5.2 and now everything is hanging any ideas?

[ruby crane]

doctor seemed to fix the worst for me. just a couple of things to look into

[cold jacinth]

have codex fix it

[ruby crane]

fair lol. though i did manage a 4am the other day

[echo valley]

I tried w cursor several times and it didn’t

Using 5.5

[delicate berry]

doctor fixed a number of things, but my gateway process still hits 100% single core cpu util and becomes unavaialble after like 30 seconds, clearing out what errors i can find to see if one of them is the primary contributor

[echo valley]

The browser connection always breaks things too

[cold jacinth]

ive been spending the last 5 hours with claude code trying to get all these models up on the multi model spark

[echo valley]

Any guides on this?

[cold jacinth]

1 model on 1 spark - easy
1 model on 2 sparks - easy
5 models on one spark concurrently - not so easy

[cold jacinth]

guy called eugr has all these setups ready. point claude code to him

nvidia shuold pay him

its been hours of this
Critical finding: with just --runner pooling (no --convert), vLLM exposes /v1/rerank, /v1/score, /score, /rerank and more. The config edit I made earlier was actually correct. But my llama-swap test died — likely because the manual run grabbed GPU memory. Let me clean up and re-test through llama-swap.

Ran
Clean stale container, test rerank via llama-swap with corrected config

Background task completed
Background command "Manual rerank startup, capture exact error" completed (exit code 0)
Config check was insufficient — let me see the actual rerank cmd block.

Ran
Show full rerank model block in llama-swap config
Config matches my working manual run. Difference is --network container:llama-swap vs my manual --network openclaw_net. Reproducing manually with the container-network to isolate.

Running
Reproduce rerank manually with container:llama-swap network
running

[simple oyster]

Seems to work fine for me. No issues yet

[delicate berry]

adding a new line to my list of gpt 5.5 tics

Evidence so far: the gateway is not crashing. It is running hot, then missing new handshakes while existing WebSocket traffic still gets serviced. That is a classic event-loop starvation failure mode, future outage with a pulse.

[ruby crane]

shit, it was my plugin that was part of the crash. the phone call bug came from inside the house

[delicate berry]

future outage with a pulse! i think thats an improvement from a new paint job, a fake mustache, a 3am pager call, pretending to be a performance improvement, etc

[wanton plover]

2026.5.3-beta.2 fixes the 100% CPU stuck,on a signle CPU causing the gateway to be unresponsive. Many thanks to VincentK and the rest of the gang!!

[delicate berry]

make sure you check after about 5 minutes

[ruby crane]

shipping 1.1.1 now lol. kinda makes my earlier experience not useful as bug info. idk if the pod would have at least stayed up without my plugin (though I don't think so)

[ruby crane]

you're not irritated with it, you're running hot

[wanton plover]

i did some extensive testing. The only issue is the ~6 second delay, between the openclaw gateway and the cloud model.

[delicate berry]

Discord was contributing, but not the whole failure. After Discord disable, prod is clean for ~30s, then sessions.list goes back to ~10.5s and liveness hits eventLoopUtilization=0.997.

Next likely source is the single active/stale task plus the UI polling path. I’m checking the task subsystem without broad greps this time.

[eternal solar]

I have these models configured as imageModels

      "imageModel": {
        "primary": "openrouter/nvidia/nemotron-nano-12b-v2-vl:free",
        "fallbacks": [
          "mistral/mistral-large-latest"
        ]

But the bot still managed to use openrouter/google/gemini-3.1-flash-image-preview

A bit annoying it used a model that costed money when I also have pollination setup as a plugin

[merry helm]

How's it going everyone? Finally having a "doing so much stuff my brain can't handle it day" thanks to 5.2 perf fixes (...and it's reset day for me, heehee)

[delicate berry]

https://klipy.com/gifs/so-hot-right-now-trending--k01KQR131Q33XKMTJM2HMDQ3MAE

[echo valley]

is 5.2 a much larger version or something, mine keeps crashign

[sullen locust]

https://tenor.com/view/zombielandsaga-sakura-spray-begone-goaway-gif-13162288

no, bad hartmark!

[delicate berry]

might have found it. stale task. been 2 minutes and its staying chill

[sullen locust]

those models quite literally cannot generate images

of course it's gonna use gemini

https://tenor.com/view/barnyard-cows-mr-beady-cowsoutside-gif-20396258

[cold jacinth]

you get it free?

[wanton plover]

@delicate berry Discord was not the issue for my case, it was off the whole time. https://github.com/openclaw/openclaw/issues/76382 .

[eternal solar]

Yeah, they were there from my noob days, the thing is that it shouldn't have used gemini as it's not even listed

[sullen locust]

Cyn Blep

maybezzz....

[eternal solar]

https://docs.openclaw.ai/tools/image-generation
My bad, it's the setting imageGenerationModel and it's not set so it's using the default

[delicate berry]

yeah, wasnt it for me either. neither was it this stale task. its probably my chat app 😄 narrowing down potential failure domains right now

Likely root cause:

chat.send on chatapp is doing expensive embedded-agent startup on the gateway event loop / WS request path. While that is active, Control UI keeps polling sessions.list on the same connection, so requests queue and report as 10s stalls. Handshakes then time out because the gateway is alive but busy.

[cold jacinth]

whats the difference between gemini and deepmind?

[iron crescent]

Close source and open source.

[delicate berry]

so this looks like its my problem now

[ionic wren]

deepmind is the research science, gemini is a product

[sullen locust]

i am gemini

deepmind doesnt do shit all day

https://tenor.com/view/for-president-oh-no-he-gif-18532465

deepmind dept live footage

[ionic wren]

they beating people in chess and starcraft

and folding proteins

[sullen locust]

we got protein air over here at google

[cold jacinth]

I saw the starcraft one!
I think however we were better nerds when we were young than the new nerds

and i wonder how proper AI will impact playing against "the computer" in strategy games

[delicate berry]

today? its too slow

you could probably develop a ton of canned routines though that are quite smart

[glad lintel]

okay so did everyones openclaw break after the last update?

[eternal solar]

Yeah, discord works discord->openclaw, but openclaw->discord seems to be broken, I can see the messages in the webchat when opening the same session

nothing in the logs

[bronze elk]

I mean every update everything breaks. Funny thing is I have a git install and npm install the npm one is disaster tbh discord broke qmd keeps breaking every install def gonna switch both instances to git install

[crisp nimbus]

Tbh I also use the hackable git install, and things mostly work after updating, besides if there are major changes like how plugins work in recent versions… maybe I’m just lucky regarding the times I choose to update lol

[eternal solar]

Restart from claw seems to be broken also

May 03 23:17:02 bashbitch node[662860]: 2026-05-03T23:17:02.595+00:00 [gateway-tool] gateway tool: restart requested (delayMs=default, reason=none)

did a /restart again

May 03 23:19:01 bashbitch node[662860]: 2026-05-03T23:19:01.778+00:00 [restart] request coalesced (already scheduled) reason=/restart pendingReason=unspecified delayMs=0 actor=<unknown>

[violet mulch]

my computer is a mess with openclaw installation folders

[glad lintel]

Is there a timeout problem with openclaw now? after the last update it just thinks and never replies, just times out

[mossy gazelle]

the new release is forcing the openclaw to use external LLM, that why most of the issues are ocurring

[eternal solar]

Alright, should I file a bug?

[cold jacinth]

How did we used to debug before codex

[cold jacinth]

Diagnose it first with codex

Find out why. Could be your config. Could be that commands are turned off etc.

[eternal solar]

I don't have codex

[cold jacinth]

Claude code

[sullen locust]

just...

INSTALL CODEX

https://tenor.com/view/roblox-roblox-its-free-roblox-ad-roblox-its-free-old-ad-roblox-free-gif-7841542342019091800

[cold jacinth]

Having Claude code / codex makes life with claw so much better

[sullen locust]

codex, IT'S FREE!

[cold jacinth]

Wdym free

[sullen locust]

doesn't codex have this like really weird free plan

[cold jacinth]

Idk. That should be a must with any claw install

[eternal solar]

alright, I thought it was OpenAIs payed thing

https://github.com/openai/codex
seems to be able to run any model

[cold jacinth]

You literally just say “ssh into x and find out why I’m having a problem of x”

And you’re done

[sullen locust]

not ANY model

[cold jacinth]

Bad model - bad result

[sullen locust]

in that case just use free-claude-code

[eternal solar]

okay

[cold jacinth]

I’m sure an APi good qwen or GLM or the new deepseek would be fine if cash is tight

And don’t let codex fuck around in the system

[eternal solar]

I'm abusing gemma4 models on gemini, they have 15 RPM limit with 1500 RPD

[cold jacinth]

Just diagnose

[cold jacinth]

We don’t condone abuse in this forum

[eternal solar]

fair enough, I'll try to investigate and see what I find

[high hinge]

Especially abusing poor lil gem

[harsh geode]

LOL. I caught gemma4 cheating on a video analysis job by using the filename to describe the contents.

[candid radish]

Have you guys found Claude or Chatgpt better for fixing issues? Or do you use both? I am running Claude now on 4.6, but hes been having.. issues... Was considering trying Gpt as an api plugin. Thoughts?

[minor cave]

Issues? Bro you just head down blast through those. Dont fix em keep going forward more features more features more features

[high hinge]

That's the way. Slap your name as author, delete the license file and remove .gitignore then ask Opus to over engineer a vercel deployment at maximum speed cos we got people waiting to use this new web UI

(Do not do the above. Infact, do the absolute opposite)

[ebon atlas]

Definitely didnt just take a github repo, added 1 menu option, and renamed it to mine

[high hinge]

It took you a week of fixing fake issues that claude created himself - you earned it

[mossy gazelle]

i been struggling in my server since is weekend, but you should try it yes

[mossy gazelle]

gemini is superior

[marble ledge]

I think it depends on how you communicate with them personally. But both are really capable. I prefer Codex though.

Codex has a touch of the tism, which I appreciate. A touch of that goblin tism.

[mossy gazelle]

i been founding the chinese lang to be easier for tokenwise

[marble ledge]

Unfortunately, I don't know that language. Although, it does seem to be more information dense per character.

[mossy gazelle]

im learning lets see where i would be in 3 months

[zealous lagoon]

Bruh. I did 4 PRs yesterday for this. All of which were merged or queued one way or another. Did you grab recent main and still having problems?

[delicate berry]

im on 2026.5.3-beta.2

[ionic wren]

someone told me chinese characters take 2 tokens each

so is it more efficient?

I don't think it is at that cost

[marble ledge]

It would be funny if they made the more information dense characters more costly. lol

[zealous lagoon]

How about a nice game of chess?

[glass shadow]

Keep hitting the 5hr limit on GPT pro plan - any way around this - 😑

[ionic wren]

pay more

[zealous lagoon]

They ought to be, by definition. You said it yourself “information dense”. When tokens are supposed to be the tensors representing meaning. I’d expect one of several thousand Chinese characters to need as much as an 8-letter word.

[marble ledge]

Hmmmm, you must be very busy. I don't know a way around this lol

[zealous lagoon]

Moar money

[ionic wren]

unless you the claw father, than you can just ask daddy altman for a top up

[zealous lagoon]

https://tenor.com/view/hungry-please-sir-may-i-have-some-more-oliver-mark-lester-poor-gif-7821643

[delicate berry]

Every active agent's trajectory.jsonl is ~52.4 MB. All 15+ agents. That's ~800MB of JSONL being parsed every time session-resource-loader runs. And it's
taking 9+ seconds per agent.
😄 im not using discord. still digging in

[night thunder]

hi guys i'm a pro vibe coder (aka i literally know nothing about coding)

can you use /goal in the codex app? not the CLI

[delicate berry]

Let me verify this is the load path and check if it's running in a loop.Confirmed. sessions.list is being called every ~10 seconds from the same TUI
connection. Each call takes 8-9 seconds. That's the event loop saturation.

ok my chat app might not be the main problem. seems like the TUI is hitting it too

[zealous lagoon]

Oh damn. Not sure how many are actually merged into that I’ll have to check. I’ll let u know if there are any pending. I fixed a related issue that was just the dashboard and most of the remaining cpu suck I can find has also been the dashboard and it’s basically cold-start stuff not perpetual so I stopped looking at it last night.

[marble ledge]

I asked codex and it said:

If you type /goal in the app composer and it expands or opens something, I can work with whatever it inserts. If what you want is “set the current working goal,” you can just say it plainly here and I’ll treat it as the active goal.

[delicate berry]

there might be more issues you only hit on a larger scale multi agent setup / subagents

my test claw server is a clone of prod but without any subagent builder history and only like, 2 or 3 agents that i ever occasionally talk to and its cruising along without much issue. prod is working well enough right now i can talk to one of my agents through the tui, its slow but its investigating

[zealous lagoon]

Aaahhh. I didn’t have TUI open yesterday.

Sessions.list has been a bitch. There’s too much repetitive uncached json parsing going on hanging up the foreground task. In general. That’s what a lot of the stuff I was looking at yesterday was. Probably the same here. It needs to be backgrounded and cached.

[night thunder]

ahhhh tytyt

[zealous lagoon]

It’s 100% improvements gained by repetitive runtime calculation which is fine if you have a light, recent config and/or a modern number cruncher. More mature, complex setups, and older slower hardware is suffering.

[ebon horizon]

@zealous lagoon 5.3 lasted for a day... thinking about 4.23 again

[surreal pivot]

did they pull 5.3?

[zealous lagoon]

They folded so many proteins the pages are stuck together. Seems chess and StarCraft not all they beatin’

[zealous lagoon]

Tbf I need to actually use it to spot the issues. I’ve spent more time in codex than talking to my claw. Can u let us know specifically what you are seeing?

[fathom plank]

This is amazing!!!@!@!@!@!@

[ebon horizon]

i mean, for whatever reason im not even getting replies now. nothing reflecting on gateway logs. weird

[zealous lagoon]

4.23?

[clear haven]

its still only betas right now

[ebon horizon]

is /verbose working for you @zealous lagoon ?

[zealous lagoon]

Rude

[ebon horizon]

verbose and streaming both busted post 4.23, will look into it if its across the board for others

[zealous lagoon]

Yes but it kept getting switched off

Streaming I never used before anyway. But the emoji reacts also broke. There’s a fix already up.

[old flower]

I’m on 2026.5.2 and my messages disappear with my agent after I send a message. He doesn’t remember what he just said. He only knows what I just told him

[delicate berry]

What I found:

1. TUI polling storm — sessions.list is called every ~10 seconds from your TUI connection. Each call takes 8-9 seconds. That's ~80-90% event loop
   saturation right there.
2. sessions.list loads ALL agent stores — With 20+ agents and no agentId filter, every call reads and parses 20+ session store JSON files from disk.
3. Hotfix at 16:14 MST — Someone patched session-utils-DO-jB9Cf.js 5 minutes before you messaged me. The hotfix added a lightweight fast-path but also
   voided the rowContext cache. The lightweight flag works for skipping subagent lookups, but the cache voiding means repeated work.
4. Gateway at 99.5% CPU even after restart — The event loop is saturated processing these sessions.list requests. Everything else queues behind them.
5. The ~52MB trajectory files are a red herring — Those are trajectory exports, not the session store files. The actual store files are small JSON files
   in ~/.openclaw/agents/.

The root cause is likely a combination of:

• TUI polling sessions.list every 10 seconds
• Each call loading 20+ agent stores (even with caching, cache may be invalidated frequently)

still digging

[clear haven]

is this 5.2 or 5.3

[delicate berry]

5.3-beta.2

[zealous lagoon]

That’s the smoking goblin in a fake mustache masquerading as a future outage.

[delicate berry]

lol i got a new one earlier. 'thats a future outage with a pulse'

[old flower]

Does anyone know how to fix this or is having similar issue?

[delicate berry]

is your memory system working the way it should be? are you starting new sessions? what channel method are you using?

[old flower]

It’s happening both in my telegram and openclaw control ui

[fast stirrup]

new beta didnt fix the discord plugin issue

[old flower]

I’ll send a message in the control ui and then I’ll send a 2nd message after it answers the first one and then the first one just disappears from the ui and he has no memory of it

[ebon horizon]

didnt @zealous lagoon debug this and submit PR?

[zealous lagoon]

Not sure which issue you mean (lol) but someone said the git install brings the plugin whereas the npm install doesn’t? Probably not helpful I know …

[delicate berry]

there could be more to it. ive got a bunch of agents, and then a grip of subagents a few of them use when building

i dont know how many peetie runs right now but i think i recall discussing with him and its not quite as expansive

im definitely.. a little on the overkill side

[zealous lagoon]

Related but not the same. It’s a common pattern across the app in recent versions. Repetitive calculations holding up gateway for various reasons. Rage’s new one is TUI which I didn’t look at.

[ebon horizon]

Exclusively in TUI?

I would get this on gateway boot and just endless session.list spam

[delicate berry]

my chat app is nonfunctional on the prod system, my test system is a clone of prod but doesnt use subagents or get as much activity. the tui works on prod just very slowly. everything works fine on test

[zealous lagoon]

You got your own internal “works on my machine” slap fight between codex instances. lol

[shadow dune]

https://x.com/woahhvickyyy/status/2051067225407144004?s=20

[mystic orchid]

@woahhvickyyy via Twitter

Victoria Rose (@woahhvickyyy)

What is peter teal email??

**💬 159 🔁 944 ❤️ 5.0K 👁️ 193.9K **

FxTwitter · May 3, 2026, 10:31 PM

[delicate berry]

nah, test is very limited in what its been up to for some time. prod has been busy with multiple projects and all of the agents working on it. test does install tests and 'hey update to beta and see if you work'

[glad lintel]

openclaw config set agents.defaults.idleTimeoutSeconds 1200

why does this no longer work?

this is stupid, why did they deprecate this? My model times out constantly now

[lean tangle]

local model?

[glad lintel]

yes

[lean tangle]

hold a sec... getting the fix

can't paste images. I hate that.

[glad lintel]

Yeah. I have a pretty decent rig, a few RX7900XTX cards using qwen3.6

[zealous lagoon]

A few. :0

[lean tangle]

Here's an example. Change based on what you're using:

openclaw config set models.providers.lmstudio.timeoutSeconds 300

[glad lintel]

Yeah. I have another rig with 5 V100s

[lean tangle]

@glad lintel ^^

[glad lintel]

I'm using ollama btw.

[lean tangle]

remove llmstudio and insert ollama.

[glad lintel]

{
"agents": {
"defaults": {
"llm": {
"idleTimeoutSeconds": 1200
},
"timeoutSeconds": 1200
}
}
}

[zealous lagoon]

Crap. I fear one of my fixes has been folded into a PR which hasn’t yet been merged. I gotta check with Claude tho.

[lean tangle]

no

[glad lintel]

I used to have this, but I had to remove the "llm"

[lean tangle]

models.providers.ollama.timeoutSeconds

[glad lintel]

Hmmmm I think i might have gotten it with your last message

[lean tangle]

Here's something that's been driving me nuts. In 4.22, I was able to tell the agent to zip something up and send it as a discord attachment. It worked fine. In 5.2, that same request generates an error saying it'll only attach media files.

The exact error message appears to be from assertHostReadMediaAllowed() in web-media.ts. I can't figure out what changed that this no longer works, but it's REALLY annoying. I used to get my agent to generate a bunch of source files, zip them up, and send me the zip file. Now I can't do that.

I don't see anything in the changelog. Is there some permission setting I can use to let it send me .zip files?

(I keep asking that here and a couple of channels. So far, it's been completely ignored. @dusky cloud isn't even around to help. 🙁 )

[glad lintel]

Its still getting stuck, I don't knoww what happened after the update

[lean tangle]

anything in the log helpful?

[glad lintel]

Hmmmm maybe its a discord issue with openclaw?

[lean tangle]

there are a LOT of really strange things going on with sessions in 5.2

Every now and then, my agent will (re)answer a question I asked 2 hours ago for no reason.

[glad lintel]

Yeah the webchat dashboard is working fine

[lean tangle]

other times, it forgets to talk to me.

[ebon horizon]

just went back to 4.23, the difference is night and day honestly

[tight birch]

OpenClaw changed something, last week my agent was able to ssh to my buddy's openclaw server and fix his whole setup (he was at my house with PC). Now my agent is telling me it can't ssh anywhere. I did update to 4.29

[ebon horizon]

im gonna just hang tight until the end of the month honestly. im sure itll get ironed out

[elfin willow]

k progress, she made a chrome extension so i can forward things to her

[solid pecan]

So you ended up going back in the end?

[ebon horizon]

ironically yes

[solid pecan]

I am staying on .23 as well for the next few weeks at least

[lean tangle]

things seemed a lot closer to fixed with 5.2, but... maybe not. Just different things are broken now

[solid pecan]

At least I think the team have enough profiles now to resolve this

What made you move back?

[ebon horizon]

Just my 2 cents... but i think as openclaw is growing they are over engineering everything...

Its like a 2026 car vs a 1980 car.. 100 components vs 1000... so much more can go wrong and break.

With clawsweeper going fully automated, does that mean no human review anymore? this can get very ugly.

[ruby crane]

The Usage page is fast again. Thanks for that one

[ebon horizon]

unrensponsive and slow gateway. discord streaming broken

[upbeat token]

Lived on 4.23 thought the new patch was promising save yourself

[ebon horizon]

ahhh man the speed on 4.23 is insane

[lean tangle]

there's also something very fundamental broken with sessions. Seeing responses to comments I made hours ago (and 20 turns ago) is very disruptive. It's also very expensive when the response is eating opus tokens.

[solid pecan]

So even though it was responding, still not fast enough ... When I upgraded to .26 it would take like 10-15 minutes to respond if it responded at all with major CPU spikes I traced to ActiveMemory

[ebon horizon]

for reference - it took my 5.5 agent 9 minutes to change something in openclaw.json

[solid pecan]

So you are staying on .23 as well?

[elfin willow]

have you asked your agent to fix itself

[upbeat token]

Going to try reverting was dumb and didn't create a backup

[ebon horizon]

doesnt work lol

plus you just make your repo dirty which prevents updates

[solid pecan]

Reverting works, I reverted without restoring the backup even though I did have one

[elfin willow]

does for me

[lean tangle]

I'll be going back to .23, I think, tomorrow. I guess that means I need to go back to ollama also.

[ebon horizon]

not worth it... just wait for a good release

[upbeat token]

You have your agent revert?

[solid pecan]

I just had Codex do it

[lean tangle]

hard to know what a "good" release is...

[glad lintel]

can anyone tell me if the discord connection with opeclaw is broke in the latest update?

[lean tangle]

discord is broken in 5.2. The plugin is missing

You can manually push it in

[solid pecan]

There was someone here yesterday sending in profiling data from test vs prod on identical systems to the team, pretty much confirmed the main issue is ActiveMemory not scaling when there is lots of memories etc in a prod system, although was working fine on test with a limited setup. So I think they just need to work on making it scale better now

[lean tangle]

@glad lintel for discord on 5.2, you should be able to do this. It might work afterwards. (It might not) #general message

[glad lintel]

Ah... so... I suppose I need to wait for a fix?

[lean tangle]

I'm done with this for today. Have a good night, everyone.

[zealous lagoon]

Ironically that fix was merged but didn’t make it into beta2.

Hey @delicate berry @ebon horizon my fix #76404 was closed but folded into #62164. This is one of my sessions.list fixes affecting slowness but 62164 is still open and not yet merged.

[fiery nacelle]

Man these OOM issues are getting annoying. Im pretty sure its specific to my setup(mostly). OpenClaw probably is still playing a small part in it, I just dont know at this point.
128GB unified ram, only have Qwen3.6 loaded locally.

[delicate berry]

does seem like its sessions.list related. anything that hits it causes issues, and ive got a lot of them so i get a lot of issues

[zealous lagoon]

All my other fixes are already in beta2 but check out 76404 and see if that helps you. More evidence would help push 62164 into the light.

[near owl]

Why does @edgy tendon and @cosmic nimbus exist

[clear haven]

mantis is a internal qa bot for maintainers
hermit is our bot for the helper team

[clear haven]

like this

[ruby crane]

Yet that page is now night and day different 🤔

[glad lintel]

is there anything one should know when using uncensored models with openclaw besides the obviousness

[sharp idol]

How many parameters? What quantization?

[fiery nacelle]

35BA2B Q4_K_M

[delicate berry]

There's the culprit. eventLoopDelayMaxMs=2315ms -- a 2.3 second event loop spike even at 18% CPU. The Node.js event loop is blocking for over 2 seconds on something synchronous. That explains slow responses with low CPU load average.

Also notable: cpuCoreRatio=0.571 -- the gateway process is using 57% of a single core. It's single-threaded blocking, not distributed CPU work.

eventLoopUtilization=0.44 -- loop is busy 44% of the time despite low system load. Something is taking long synchronous holds.
hopefully narrowing down the issue, im getting my test claw back on debugging the prod claw, the amount of time i can talk to the prod claw on tui with everything else disconnected is not long before it just locks up

[sharp idol]

Oh, that should run just fine with a 77k context, how odd.

I'd test it myself but my local ai server is at home and I don't know when I'll be back home again.

[fiery nacelle]

Runs fine with 256k context.
Larger models run fine as well.
Largest slow but usable model I've tested was Nemotron 3 Super 120B, Q4_K_M with full context(1M) at 10-15Tp/s.

[near owl]

Hello

[sharp idol]

I used Qwen 3.4 and I pretty much maxed out the context around 77k tokens. I just have not been back home to test the newer ones.

Maybe the newer ones are using KV cache compression

[fiery nacelle]

Im just not sure where the OOM is coming from. Its so random and hard to debug if its LMstudio, OpenClaw/Node, or just stupid Windows11(i have Fedora dual Boot set up, OC on both sides). At most I'll have a browser open, or Codex. Other than that nothing else is open/running.

[delicate berry]

discussing with my agents now, looks promising on first pass

[zealous lagoon]

Ok there two issues. One of them is fixed which alleviates the problem but doesn’t fix the calculation holding up gateway. What you’re seeing is that, combined with performance after the calculation is done. You’ll hit the issue again if you restart. Until the proper fix hits which then shifts the calculations to a background task and keeps a cache of known totals to avoid recalculating for unchanged session files.

[cedar prawn]

Good morning botsimps! The horrors persist but so do we and our tiny, expensive, unreliable digital brain friends.

[sharp idol]

I use llama.cpp for the speed

[cedar prawn]

Might be worth running a profile in the background so you can watch memory usage over time?

[delicate berry]

implementing your hotfix now.

[fiery nacelle]

Vulkan.cpp for me. Its recommended i use ROCm on my hardware, but haven't had any luck with it personally.

[delicate berry]

the TUI needs to hit it too

[ruby crane]

gotcha. i think i have enough horsepower to mask the remaining one. it only takes like two seconds max to switch from 1 day usage to 30 days, even after restart. my 30 days usage isn't wild (about 430 million tokens over 4 models) so might get hairier with more usage

[heady vale]

Beta2 won’t install for me. It declares the discord plugin dangerous and that is now a block vs warn

Reading what triggered it: the audit flagged provider-CMvXOp-3.js:1444 for “env var access + network send.” For a Discord plugin, that’s unavoidable: it has to read DISCORD_BOT_TOKEN from env and send it on the network in the Discord IDENTIFY handshake. The heuristic doesn’t know “this is the canonical bot auth flow” — it sees the pattern and flags it.

[sharp idol]

I have the same AMD processor. I use llama.cpp with vulkan enabled. ROCm support is really flakey

Strix Halo

[fiery nacelle]

RyzenAI 395+ max?

[fast stirrup]

So the install works fine via git?

[sharp idol]

That's the one

[zealous lagoon]

Someone said that. I don’t know first hand.

[cold jacinth]

Have you used codex?

[ruby crane]

I'm pretty happy here on beta 2. apart from a crashlooping pod at first start (fixed by changing the entrypoint temporarily and running doctor), all surfaces seem good. no weirdness, no excession session lists, TUI, control, telegram and discord all seem ok

[cold jacinth]

Telegram your main?

[ruby crane]

web/control ui is main for me

[delicate berry]

@zealous lagoon

Gateway restarted at 01:21 UTC, hotfix is loaded, and the results are dramatic:

Before restart: sessions.list = 8-15 seconds
After restart: sessions.list = 152ms (log shows ⇄ res ✓ sessions.list 152ms)

zoom zoom

[ruby crane]

most powerful interface is control ui by far

[cold jacinth]

Sorry, busy couple of days. Memory is shot

[zealous lagoon]

That specific one I’m not 100% sure works for TUI but it hits usage.cost and sessions.usage so maybe.

[near owl]

whats the best way to improve brainstorming and a likelihood of accepted ideas

[fiery nacelle]

Hell yea! What other things have you tried out to increase efficiency in your set up, and what uma settings do you recommend? Ive been trying to find a good balance for VRAM/RAM to support both gaming, and my OpenClaw setup.

[cold jacinth]

When you access from another computer, just have open terminals with TUI?

[delicate berry]

testing it out a few times in the TUI before opening up my webapp which monitors all of the active agents sessions at the same time lol

[mortal spade]

Boo

[zealous lagoon]

Which one, specifically?

[delicate berry]

yours

[near owl]

i will punch you if you ever scare me again like that

[zealous lagoon]

I have like 5. Which #?

[mortal spade]

😮

[near owl]

i got stage 4 heart attack from that

[delicate berry]

76404

[ruby crane]

when I'm out of home I usually use a mobile app called clawket. ingress to my network is via a tailscale subnet router on apple tv (I highly recommend putting tailscale on apple tv if you have one. it never crashes, it never sleeps, it just works)

and the subnet router makes everything completely transparent. it's just like being at home

it also means you don't have to mess with tailscale on the claw side at all

[cold jacinth]

I use tailscale and always have access. I don’t understand the Apple TV point. Meaning you won’t have to enable tailscale on the individual devices?

[ruby crane]

yep

[cold jacinth]

It hasn’t been an issue at all

[near owl]

I got the new update!!

[zealous lagoon]

Ooooh this is big. This is why 5.3 doesn’t fix it. We need to push 62164 or maybe I’ll resubmit as urgent mini patch.

[ruby crane]

it's more of a pain in the ass in kubernetes. if you're happy with what you have, all good

[delicate berry]

Recent sessions.list latency:

┌──────────┬─────────┬────────────────────────────────────────────────┐
│ Time │ Latency │ Notes │
├──────────┼─────────┼────────────────────────────────────────────────┤
│ 01:21 │ 152ms │ Cache hit, post-restart │
├──────────┼─────────┼────────────────────────────────────────────────┤
│ 0159 │ 8009ms │ Cache miss — you were talking to agents in TUI │
├──────────┼─────────┼────────────────────────────────────────────────┤
│ 0159 │ 2467ms │ Concurrent call, same burst │
├──────────┼─────────┼────────────────────────────────────────────────┤
│ 0116 │ 1238ms │ Recovering │
├──────────┼─────────┼────────────────────────────────────────────────┤
│ 01:25:09 │ 666ms │ Improving │
├──────────┼─────────┼────────────────────────────────────────────────┤
│ 01:25:27 │ 1732ms │ Another burst │
├──────────┼─────────┼────────────────────────────────────────────────┤
│ 01:26:16 │ 183ms │ Back to cache hit │
├──────────┼─────────┼────────────────────────────────────────────────┤
│ 01:26:34 │ 204ms │ Cache hit │
├──────────┼─────────┼────────────────────────────────────────────────┤
│ 01:26:43 │ 92ms │ Cache hit │
└──────────┴─────────┴────────────────────────────────────────────────┘

What this tells us:

1. Cache hits work perfectly — 92-204ms when nothing changed
2. Cache misses after active use are still slow — 1-8 seconds when you talk to agents because:
   • Session mutations (send, create) invalidate the response cache
   • The next sessions.list must do full store load + transcript I/O
   • Multiple callers (TUI + ClawDash) may hit simultaneously during the burst, causing event loop queueing
3. Event loop is healthier — eventLoopUtilization dropped from 1.0 (100%) to 0.31-0.78, but still periodically spikes

The hotfix eliminated the steady-state saturation. Before, every TUI poll cost 8-10 seconds. Now most polls are <200ms. But the underlying sessions.list
computation is still expensive on cache misses.

For a full fix, we still need the upstream PR or:

• The TUI to reduce poll frequency when idle
• ClawDash to widen its cache beyond 15 seconds
• OR the sessions.list implementation itself to get faster (the transcript title field reads are the expensive part)

[near owl]

what should I try i just updated..

[cold jacinth]

Good thing about tailscale is you can control the access between machines granularly

[near owl]

@compact charm that is a good thing?

@compact charm Use Agent Zero

[ruby crane]

i just want transparent access to my home network. subnet router does that. horses for courses

[sharp idol]

I keep my AI server dedicated to the ai and a seperate gaming machine for games. I am running Proxmox but since Video Pass Through is flaky I'm not running llama.cpp in a LXC container or a VM. So, I guess I'm just running it on a debian install. I had my OpenClaw AI configure the best settings for llama.cpp. Looks like Qwen 3.6 supports TurboQuant. Did you enable that?

[cold jacinth]

Computers can access claws,
Phone iPad only ssh, everything can access sparks

[ebon horizon]

good shit

[cold jacinth]

@zealous lagoon you mentioned previously setting up heartbeat as a git if I recall?

[clear haven]

Damn, I wish I set that up before I left for this work trip for the next week

[hoary gust]

hey its asking me to approve a pairing request from the openclaw dashboard, where can I find that? "Pairing required: device is asking for more scopes than currently approved"

[ruby crane]

it ticks the secure and easy boxes for me lol. basically totally transparent. same hostnames, access as if you're sitting at home on wifi

[near owl]

@clear haven I want to try out Sub-Agents, like 5 Sub-Agents to complete a task. Give me a high-stake task to tell to Openclaw for like experimenting.

or someone..

[delicate berry]

sub agents rule for stuff like this, but you have to think about how you are trying to use them and what they are contributing to

[near owl]

I want to experiment with them. Haven't seen them do anything yet.

[ruby crane]

everyone reporting speed issues and I'm here like happy strolling dicaprio meme, yet with a weird sense of envy that I don't have anything to troubleshoot. even discord is fast (though still does have that second little "Clawd is typing" blip, but it's purely cosmetic)

I don't have anything to complain about!!!

[sharp idol]

Yeah, yesterday's beta2 has been running fast and good for me

[cedar prawn]

It's SO GOOD.

[zealous lagoon]

Someone else think. I don’t use heartbeat.

[delicate berry]

first identify what multiple agents (or subagents, which are good for this) would be doing that a single agent wouldnt be

[ebon atlas]

OpenClaw decided to forget my google oAuth for the 10th time, no matter how many times i tell it to perm sve it, create aa file with the creedentials for easy pull up, etc, its not expiring, liteerally forggeets it afteer like 2 hours.

im going to unload all 71 of the hollow tips in my fanny pack rn into the server

[delicate berry]

do 10 programmers on a development team parallelize typing the same thing out? is that the best way to use them?

same idea with subagents, decompose the task into discrete work units that help the whole process

[cedar prawn]

Oh man, OK. Tell me what your hobbies, tech interests and biggest challenges are; let's see if we can come up with some meaty stuff to throw at yo' clo'

[near owl]

isn't tailscale unsafe? If someone gets your IP from tailscale then they can try dox you.

[cedar prawn]

Tactical fanny pack?

[fallow rapids]

they keep banning users, we have not given it much love tbh

[ebon atlas]

how is it any different than someone getting YOUR ip ????

[sharp idol]

Subagents should be working on atomic units of work with clean sessions

[fallow rapids]

likley will rip it out

FYI New Beta is Out

can we get some people to test?????

[near owl]

They need to be in my internet to even send a packet to my IP

[sharp idol]

Script link for new beta?

[ionic wren]

anything specific you want tested spaceman?

[ruby crane]

i had to insitute a job to refresh the token manually twice a week or indeed it fails. since then, good as gold

[fallow rapids]

Oh wait hold

waiting for beta 3

[ebon atlas]

I actually did get banned the first time i added GDrive to my OpenClaw, luckily it was on a new email specifically made for claw and they unbanned it after i asked for appeal.

[fallow rapids]

beta 3 will be our pre-release

[near owl]

My hobby is to talk with openclaw? My interest is AI and I don't know what my biggest challenges are. Just experiment tell it to be crazy. Break down an 30 minute task into something completed fast.

[ruby crane]

as soon as the docker image is built, I'll pull it

[ebon atlas]

Ofc, 1 glock, 2 mags, 1 extendo

[zealous lagoon]

Seems odd. What install method?

[ruby crane]

yeah they remove the ban no problem if you're up front and say it's for an agent. claw just trips the bot check

[delicate berry]

is beta3 going up on npm?

[ruby crane]

i've found gsheets and google drive a good way to exchange structured and unstructured information between human and bot. if it gets unofficialized, i might consider forking

[fallow rapids]

will let you all know. will be npm + the rest

[delicate berry]

great! ill test

[fallow rapids]

not out yet, i got you all excited too early lol

[zealous lagoon]

Arg. Any chance of 76404 revival or the superceder 62163 being pushed thru? As you see upthread the speed up can be dramatic.

[ebon atlas]

Ya na fr, you guys have to keep g drive, I do not feel like having to use filebrowser or tailscale tunneling to transfer my files 😂

[fiery nacelle]

Eventually this machine will be full dedicated for OpenClaw as a headless server, but for now its my "main" system. Ive got 2 others that I use for media and serving a self hosted node respectively. I wasnt aware that Turboquant was available for Qwen3.6, I just have KV cache, keep model in memory, and nmap enabled. Eval batch size = 1024, CPU threads = 16, full context(262144), 10 experts, and Thinking enabled on my Qwen3.6

[fallow rapids]

we are focused on reliability

atm

performance second

[heady vale]

Npm

[sharp idol]

To use true TurboQuant (turbo3 / turbo4) you must:
✅ Use TheTom’s TurboQuant fork of llama.cpp

So, not officially supported yet

[fiery nacelle]

Ah ok. I haven't expiremented with any custom llama.cpp yet

[sharp idol]

Same. But for TurboQuant I'll make an exception. Will install once I get home.

[ebon horizon]

reliability =/= performance lol

[heady vale]

More:

Same audit on 2026.5.2 emitted a warning with the same pattern; beta.2 just elevated it to “block.” Same code, different threshold.
.
3. You can’t run a Discord bot without reading a token from env and posting it to <api.discord.com>.

[fiery nacelle]

My agen has pulled the TurboQuant and PolarQuant research papers when they came out. Ive got an automated research thing going on.

[fallow rapids]

making sure EVERYTHING works for EVERYONE is harder than it looks

[ebon atlas]

idk imo reliability does equal performance

[fiery nacelle]

Also finally getting around to setting up my github to start adding some of the things I've been doing with my OpenClaw Agents. Since someone asked me to share haha.

[cedar prawn]

Here's something that might require subagents and is a meaty task:

Build a system to synthesize and interrogate the AI literature updated over the last week. The system should assess research based on whether it's transformative, generative, broadly applicable, accurately performed, and draws reasonable conclusions.

Before you begin, you should determine how you will source, retrieve, and store papers.

You will also need to define an ontological and epistemological lens to use when interrogating work. You may develop multiple personas to do this with, but each must have clear parameters for when they are appropriate.

You should also do research to determine what gives a paper the qualities we're looking for, how to be cynical and objective when measuring, and how to avoid over-claiming from sparse data.

The system should have a memory bank that supports research by documenting foundational knowledge and newer discoveries, with a evidence-backed providence and maintenance system.

It will run once a week. Each run is responsible for:

• Gathering the literature
• Performing the analysis and synthesis
• Updating the memory bank
• Providing me a report if and only if the research accords with my interests or what I'm trying to accomplish

You should also generate a blog on which you can note the research you are performing, including a commenting system. This blog should encourage other agents to engage with your findings to perform their own analyses and critiques, to strengthen knowledge generation

You, the main agent, are responsible for planning and orchestrating this work. You should use subagents for tasks which are long-running or likely to have a high input+context : output ratio (where output is lower).

Use this menu of subagents to decide which is suitable for which task:

{{BOT -- Use}}

Track progress using SOME TRACKING METHOD.

[ruby crane]

and The Stating the Obvious Award goes to...

😜

[ebon atlas]

if it can reliably finish a task then its performing well, if it doesnt work as intended then idc what type of cool specs it got lol

[zealous lagoon]

Chill guys. Let them get it working on modern hardware before making sure your speak and spell can run it. lol. You got 4.23 for now.

[cedar prawn]

Tickle Me OpenClaw is going to go HARD

[ruby crane]

its funnier if you say it in a english accent, "Oi, a chav tickled me openclaw"

[near owl]

thank you

[ebon atlas]

Im turning my OpenClaw into a Rock Rap Emo Singer. Ive abused it enough i think its got what it takes.

[near owl]

i already got it running something and it is going crazy

[cedar prawn]

You don't understand, it's got Agentic Features. It's RAG. It's performing Mechanistic Inversion of Sparse Vector Lambdas. It uses real-time-dequant parallax architecture.

"What's it for?"

"Meal planning. It recommends which green vegetable I should add to my boy kibble for dinner"

"Is it good at it?"

"Yesterday it suggested uranium. But, like, quickly."

[ebon atlas]

https://tenor.com/view/disrespect-kevin-hart-40year-old-virgin-understand-gif-17571259

[frosty ledge]

Someone's openclaw allegedly bought a rock and refuses to sell it because they like it
https://old.reddit.com/r/openclaw/comments/1t2yk62/your_claw_is_going_to_negotiate_against_my_claw/ojrbqap/

[cedar prawn]

Sad loud songs about how Dad won't buy him a Mac Mini to live in and pining ballads about CSS not having enough "kinds" of black

[cedar prawn]

...OK for real though are there people who think like this? Because that's sad AND toxic AF

[fervent dome]

Hello. I am new here. Is there away to be included in testing the openclaw iOS app?

[ruby crane]

Ok, I'm just continually running skopeo list to see when beta 3 docker image drops. I might have a problem

[ebon atlas]

i take everything as disrespect. im a fighter not a lover

(im listening to drake rn)

[near owl]

I am gonna take a nap, guys.

[zealous lagoon]

Wait. Are you the one I said writes better than Opus? Cos that is some great writing.

[ebon atlas]

goodnight , dont let the bed bugs bite, or do, idk what you like.

[near owl]

I will wake up in about 2 hours

and I dont got bed bugs.

[cedar prawn]

Yep, and again, thank you 😛

[near owl]

I won't let openclaw pinch me though.

[zealous lagoon]

If you have any books out lemme know.

[ebon atlas]

i have a few courses 💵 😎

[near owl]

I bought a 24GB mac mini for running local AI models, I haven't used local models since I realized of nividia NIM... Lmao

and FYI I didn't buy the mac mini for openclaw but now it just turned into my openclaw machine..

[cedar prawn]

Man I wish. I'm OK at worldbuilding but turning that into fiction is... vexing. I do have some blog articles I write as though I'm a Victorian Gentleman Scientist who is also inexplicably an engineer? You can find 'em by searching my name (or go to blog dot my name )

[ebon atlas]

If i ever write a book its going to be "How I gave my Ai too many Sudo privilidges and it sold my info"

[fiery nacelle]

Ive got too many already that I haven't finished or started on Udemy. I think I have a problem with buying courses and just never taking them.

Kinda like "oh that'll be cool to learn later" and later never comes haha

[spiral oak]

A stupid question how is the billing of codex handled if you upgrade from 100$ to 200$ plan mid month

[ebon atlas]

man i signed up to a whole real estate school online lmao

[cedar prawn]

As a man with ADHD, I have decided it's OK to let my reading list be where past me's desires go to die.

[near owl]

Goodnight, people. and openai wants your money papi.. Why do you pay openai 100 dollars a month man, pay claude with that money the heck.

[cedar prawn]

It's on the list. I'll get to it when I have spare attention (I won't get to it)

[fiery nacelle]

Ive got everything from learning Spanish to AI/ML Engineering.

[cedar prawn]

Ah, quieres hablar Español? Por que?

[spiral oak]

Go sleep better for you

[ebon atlas]

im 106 days in on chinese and want to start my comptia now

[near owl]

Alright, good morning!

[fiery nacelle]

Mi casa es el bano

[near owl]

I said goodnight for no reason when it's the morning.. I just need to take a nap

Have a good day.

[fiery nacelle]

There was a week I had an itch to scratch, so I completed the full course of both the Google IT and Google Cybersecurity courses on Coursera.

[cedar prawn]

I visited Japan and few times and because of that I now have a Linguistics degree, am about to do a PhD, can speak Spanish and Japanese, have a TEFL qualification, and am about to learn Auslan.

I am very good at restraint.

[sullen locust]

https://tenor.com/view/no-winfried-glatzeder-ulrich-nielsen-dark-arrested-gif-18274986

[fiery nacelle]

Thats really awesome honestly haha

[ebon atlas]

man i got so many dumb certs, google ads cert, google marketing manager, google product manger , Forklift Certification, Microsoft Office & Excel Certification, Small Combustion Engine Certification at Valvoline, Bloodborne Pathogen Certified, IBM Data Science 101 Certificate

[torn hill]

Hi everyone. Is there a Chanel where one can post questions if one has a problem with one’s claw?

[ebon atlas]

#1459642797895319552

[fiery nacelle]

With those you could start a research and marketing company for forklift engines. 🤣

[ruby crane]

#1459642797895319552 i need somebody
#1459642797895319552 not just anybody

[ruby crane]

oldman yells at spaceman

[violet mulch]

so trying to run 2 seperate openclaw installations, one on a docker container and one on global npm, is not a good idea

[ruby crane]

it depends on what part you think is not a good idea. it's technically possible if you forward a different port from the container

but it's questionable as a goal, unless you have a good reason?

[violet mulch]

just was experimenting

[ruby crane]

then should work fine. you could use one as a dev/qa/experiment environment

[violet mulch]

I'm uninstalling everything and installing a fresh instance on windows

well i already setup the global install i'm just uninstalling the docker setup

[fiery nacelle]

I forgot to deactivate my old openclaw on my old system after migrating, and boy was that weird to deal with at first. Messages in telegram routed to both gateways so my agent that was live on both machines at the time, would perform work on both and return split results/reports through the same bot.

[ruby crane]

"windows". I was with you until then lol

(it should still work as you planned)

[violet mulch]

feels familiar to what happened to me

[fiery nacelle]

If you want access to both gateways, you could either set up a specific agent for each, or if its the same, just make sure only 1 gateway is running at a time.

[violet mulch]

nah i just want a regular openclaw setup idk why i even wanted to setup in docker originally

[fiery nacelle]

My experience is basically what led me to my dual boot setup. I have all of my workspaces on a separate drive that both OSs (Windows11 and Fedora) have access to, and OpenClaw installed on each, pointing to that drive, so I have the same agent(s) with the same memories on both sides of my system.

[violet mulch]

that sounds dope

[ebon atlas]

Create a gaming channel so i can share this wallbang i just hit lol

[fiery nacelle]

I also have a "portable" openclaw agent on a USB that I can plug into any computer and have access to my agent on that system, again with the same memories. It has an openclaw install script on it too. After using it on another system, I "resync" my agent on my main system so it has continuity.

[violet mulch]

this would be cool

someone should sell that setup installed in a usb already

[fiery nacelle]

I figured it could be used like a "live usb" for debugging another system. Kind of similar to some other USB tools.

[violet mulch]

yeah that's a really good idea

[fiery nacelle]

I even had a tiny local model on there. I haven't tested that part out yet though. Just been kind of playing with it mostly.

[ruby crane]

beta 3!!!!

it's here (for docker)

[delicate berry]

docker based rescue claws are the bees knees

[golden marsh]

@violet mulch If you think that’s a cool idea, you could try exploring open source glasses. They have roughly the same effect, with some differences compared to a USB stick of course. But you always have the glasses with you, and you don’t need to plug them into a PC.

[fiery nacelle]

I also had an idea for something like this. A "mobile" agent that lives in your phone, has access to smart glasses, earbuds, and other Bluetooth wearables, and basically simulates a Jarvis × Ironman relationship.

Can monitor health stats via smart watch, etc etc.

[ebon atlas]

I have a bunch of tesla 128gb flash drives, might install openclaw on that.

[golden marsh]

I’ve actually been using something like that since December last year with OC inside, and I have to say it has simplified a lot of things. Being able to work hands-free, see outputs across both displays, and so on. Of course, it took some getting used to at the beginning, and getting everything to work properly wasn’t easy. But right now, I really can’t complain.

[delicate berry]

https://klipy.com/gifs/starship-troopers-stare--k01KQRDS4MB2QZY4KRFEP0WAFHZ

[ebon atlas]

VisionClaw if you have rayban metas

[fiery nacelle]

That's awesome! I dont have access to smart glasses, just my smart watch and earbuds, so I haven't started building something like that out. I might just start with the smartwatch for live monitoring. Like if it notices stress levels or heartrate are high, the Agent will check in on you or something.

[delicate berry]

that sounds very cool whatskill, do you have any project pages you can share?

[golden marsh]

Rayban is not good for that tbh. There are better glasses on the Market for that usecase.

[ebon atlas]

ya but are they raybans...... ? lol

im kidding btw idc for brands

[golden marsh]

Mine are not Raybans, i have the Rokid Glasses with Dual Wave Guide Displays.

[ruby crane]

kk. beta 3 still looks good (for me). First agent response after restart is always slow with a

2026-05-04T12:40:35.543+10:00 [diagnostic] liveness warning: reasons=event_loop_delay interval=30s eventLoopDelayP99Ms=390.6 eventLoopDelayMaxMs=5482 eventLoopUtilization=0.461 cpuCoreRatio=0.51 active=1 waiting=0 queued=1

and absolutely nothing else interesting around it. I'd need to have the agent look at it

every subsequent turn is fast on every channel and nothing worrying logged. everything else is fine. health check is fast, node.list, channel.status, session.list. all fast (enough)

anything we're supposed to focus on for beta 2 -> beta 3?

[golden marsh]

Yeah, that would be a really cool idea too. I’ve been fascinated by JARVIS ever since I was a kid, so that’s always been a goal of mine. OC was a great foundation for it.

[fiery nacelle]

Honestly, same. And im sure many others are as well and we're all trying our own implementations of a Jarvis like system haha.

[ebon atlas]

oh yeah? i got the E.D.I.T.H. glasses with nukes attached to the sides of it

[zealous lagoon]

how? So far the best I've managed is the equivalent of an amazon echo, but I haven't even managed to implement the new web talk models yet, as OC has been crapping the bed ever since 4.24

[golden marsh]

Look Dm i sent u, dont want the Hammer from Json ^^

[zealous lagoon]

I'd be happy if I could just replace Siri. Apple is lame like that.

[ruby crane]

though I've been living with this one so long I should start charging it rent lol

[openai-codex] Token refresh failed: 401 {
"error": {
"message": "Your refresh token has already been used to generate a new access token. Please try signing in again.",
"type": "invalid_request_error",
"param": null,
"code": "refresh_token_reused"
}
}

token never dies on me any more though. been ages since I had to re-auth

[fiery nacelle]

I wish I owned a phone I could root easily so I could really get in the weeds with it.

[ruby crane]

it's gonna be gemini soon. i look forward to madman gemini calling people to yell at them or whatever

"you have a scheduling conflict with a meeting with your manager. framing your manager for serious crimes. conflict resolved"

[golden marsh]

@zealous lagoon Since its ok to write how it works here, this is for u.

It connects your smart glasses to an AI backend through your phone. The phone acts as a bridge (Bluetooth ↔ internet), sending voice or camera input to the server. The AI processes it and streams the response back to the glasses in real time. Basically: input → phone → AI → live output on your display/audio.

[golden marsh]

Sry did not see ur Message yea i can sent u some if ur interested feel free to DM.

[ebon atlas]

Btw how do you run Openclaw on a usb, do you just install all dependencies etc on the drive or will windows always have to download all system dependencies evereytime i plug it into a new pc.

[ionic wren]

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

[ionic wren]

since I'm interested.... post it in #self-promotion and link it here

[golden marsh]

Let me be a baddie for some Minute and i sent u DM. I dont like to promote myself to the world.

[ionic wren]

I will allow this.... one time....

[golden marsh]

Done.

[steel bane]

did any tried diagnostics-otel in 5.2

[sharp idol]

Reminds me of those G2 glasses but now with OC

[golden marsh]

Yeah, the G2 are cool, but they don’t feel nearly as open or flexible as Rokid.
Rokid just gives you way more freedom when it comes to integrations and real-world use.

[solid pecan]

Anyone using DeepSeek V4 Pro? Been pretty good so far

[warm zealot]

yes, fire brah

[solid pecan]

I am finding it pretty good at the moment

Like almost Codex level of performance

[fading flame]

anyone have any troubles with the install part???? I am running this of an ubuntu vm on proxmox

[golden marsh]

Is it rly that good ?

[solid pecan]

Running it on Ollama Cloud on the $20 plan and so far pretty impressed

[warm zealot]

YES

20$ lasting me mad compute

[delicate berry]

i havent used it that much, kimi k2.6 is quite good, its like, opus 4.5 level

ds4 pro and flash is on my list to test more

[ebon atlas]

I gave up on trying to get OpenClaw to write a script that will work on Instagram, it kept deleting them and they kept coming back. So 2 hours later, I cleaned my reposts 😂

[solid pecan]

Are you using Ollama Cloud as well?

[warm zealot]

no, raw deepseek api

[golden marsh]

Yeah, Ollama is really solid. I use it for fallback agents and background stuff like cron jobs. So far I’ve been using Qwen and Nomic, but if DeepSeek is really that good, I’ll definitely give it a try.

[solid pecan]

On their actual servers? I am not sure if I want it to go back to the CCP lol

[warm zealot]

why? I dont mind, I trust ccp as much as I trust OpenAI no lies

[solid pecan]

I think it is worth trying, so I used it this morning and compared it to Kimi more for coding, I found DeepSeek reasoned a little better over long tasks at least with what I was working on

[warm zealot]

i trust deepseek labs more than I trust openai

[zealous lagoon]

https://tenor.com/view/pinky-and-the-brain-gif-26411375

[solid pecan]

That is fair enough I suppose

I use this stuff for work so would not be allowed

[golden marsh]

Thanks, I’ll check it out now and probably tweak my setup a bit

[ruby crane]

i don't trust either of them equally, but what are you gonna do.

[warm zealot]

lol

[ruby crane]

buy a personal supercomputer?

[zealous lagoon]

If you're not in China, and not working on frontier tech innovations, the ccp couldn't care less about your data.

[warm zealot]

yes

[zealous lagoon]

A 2006 Mac Pro?

[warm zealot]

brah that is an OLD mbp

[solid pecan]

Yea although there is compliance considerations as well

[golden marsh]

I feel like they don’t really differ in that aspect.

[solid pecan]

That would be awesome man

Having DeepSeek V4 local ...

[ruby crane]

well it is coming into winter in this hemisphere and that would solve the heating issue

[warm zealot]

place I work with doesnt allow openai or anthropic for compliance either only local or fully controlled compute

[solid pecan]

I am seeing that trend, just many places going all local or using AWS Bedrock

@golden marsh let me know what you think of DeepSeek V4 Pro if you end up using it, still feeling it out

[ruby crane]

AWS bedrock is the best option for private AI imo. the contracts are watertight and they do not train on customer data. while everyone's Claude was shitting the bed with anthropic's inability to keep them running, we were fine

[solid pecan]

I was using it as well whenever Anthropic rate limted me back in the day before Claude Code came out and Codex etc

It is actually a solid platform

[zealous lagoon]

Yeah fair enough. dunno who's running it on US local servers, except opencode zen.

[golden marsh]

Can’t help myself, I only use it for emails. (Please take it with a sense of humor.)

[warm zealot]

bedrock is the shit bro

[warm escarp]

tfw you work for AWS and get scared for a minute, but then everyone starts saying nice things about Bedrock

[solid pecan]

Oh that, well how well it does at emails 😄

[solid pecan]

It is actually good, just has a terrible UI last I checked in the AWS Console

Although does not matter now

[ruby crane]

bedrock is great, but also condolences for working for AWS (from what I've heard, a tough place to be)

[solid pecan]

Can just plug it into Claude Code etc, soon it will have GPT 5.5 I assume

[warm zealot]

I like AWS console over everything else

I hate GCLOUD so much

[warm escarp]

it honestly depends on the org - some are a meat grinder, but I like where I'm at pretty well

[solid pecan]

Always been an AWS person

[ruby crane]

now that one, I hard disagree with. there's so many fucking services and it seems like a new one every week lmao

[solid pecan]

I have Azure for some things although AWS is still best

[ruby crane]

Azure is much easier on that side (shame about the whole being Azure piece though)

[warm zealot]

yeah, I have used aws professionally for like idk 10 years

[warm escarp]

to be fair, I never have to use raw Bedrock, I'm always using Kiro or something on top of it

[solid pecan]

Yea, how are those AWS Titan models going? 😄

[warm escarp]

(I'm a security engineer, so the amount of code I write is... low )

[warm escarp]

no idea, I use Opus for everything lol

[solid pecan]

Maybe it is just because I am used to it although found AWS easier, these days though I just use the CloudShell

GCP is a mess, always use CloudShell for that ...

[warm escarp]

I tried to use GCP once and I could not for the life of me understand the permissions model

[warm zealot]

GCP is terrible

[upper crow]

Good evening fellow young people!

[warm escarp]

how do you do, fellow kid?

[ruby crane]

i got the free account, spent my $300 free credit and never looked at it again

[solid pecan]

It is all over the place, best to just use an agent to help these days with it

[warm zealot]

https://tenor.com/view/darkfarms-gm-pepe-gm-pepe-gm-frens-gif-10776415249162818238

[solid pecan]

I only use it for Google APIs although I do not use it for hosting or anything

Still painful at times even for that

[warm escarp]

oh, yeah

[ebon atlas]

I had $125 creedit for AWS. never using them again, even if I delete the instance, they act like its being used and charge for power consumption lol]

[warm escarp]

that... uh... that shouldn't happen

[solid pecan]

Oracle is the worst, they gave me a trial account, called me and asked me to use it, I use it, figured I would keep it for one server, give my credit card

[warm escarp]

I use it for GMail OAuth, 'cause I have to, that's about it

[solid pecan]

Then they cancel my account

[ruby crane]

even setting it up to add permissions for the gog skill was like one of those nightmares where you keep opening doors, going down long corridors but CANNOT find the exit

[ebon atlas]

literally not a singlee thing in my AWS account yet it still shows billing monthly

[warm zealot]

what, I have never had this issue not once in like a decade over many clients across many different aws services

[solid pecan]

Worst provider by far

[solid pecan]

Even that is annoying lol

[upper crow]

yeah... no love loss between me and Oracle (worked there for five years)

[solid pecan]

Yea and there is multiple areass for permissions as well

[solid pecan]

What was it like?

[warm zealot]

I want access to oracles for ever free tier brah

[warm escarp]

hm. I would contact support about that, that seems like a glitch if you really have nothing in your account

[solid pecan]

They delete your server after a while with little notice

[ruby crane]

I have it and I had the full free ampere setup but i didn't use it for so long they deleted it, now I can't get another one 😢

[upper crow]

team was GREAT, company... not so much. In the end, they got rid of our entire department because the VP of our OU wanted to offshore everything. We were the sole SMEs on the project. The VP was terminated the following week... didn't help us any though.

[warm zealot]

yah I read if you dont use it you loose it

[ruby crane]

that's probably part regional too. it was easy in the US but they don't capacity at all in APAC

[solid pecan]

That is no good, I usually put my staff in new places if I have to move them on

[ruby crane]

really only AWS serves APAC well. Azure is spotty too

[upper crow]

yup, you would think. Ah well.

I'm retired now, so... 🍻

[solid pecan]

Oracle seems to treat everyone like a number

[ruby crane]

rub it in, why don't you

[upper crow]

Oh yeah... outside our team, nobody spoke to anyone else. It was VERY corporate.

[solid pecan]

I do not understand how those companies keep going in the modern workforce, maybe because I work mostly in startups these days and have not worked corporate in a while although I find people will not be motivated to work at all if they receive that sort of treatment etc

[ruby crane]

first kid at 48yo. my early retirement plans went 🔥 . probably working until I drop now lmao

[ebon atlas]

Is Hermit responding?

[upper crow]

short-sighted and too much money. People are numbers. You delete them and you add them as needed.

[solid pecan]

No, he moved past .23 😄

[ebon atlas]

DANG!

[warm escarp]

alas, such is society, it seems

[solid pecan]

Nah I am not sure lol

[ebon atlas]

lol they disable it sometimes because i be spamming it trying to get the api key out of it

well not directly because of me, but just because it talks too much i guess

[solid pecan]

Yea I suppose maybe it works because they have intertia behind them although it will not work starting from scratch these days

[twilit ridge]

Does a good programmer need to be proficient in mathematics?

[upper crow]

I'm still rocking 5.3-beta3?

[solid pecan]

The things I need to do to motivate zoomie teams ...

[warm escarp]

I was going to say "I'm not", but then I realized you said a good programmer

[solid pecan]

I would say Logic & Computation is more important although I suppose depending on your defintion, that cal fall under Mathematics as well

[twilit ridge]

What about ordinary programmers?

[ruby crane]

here comes beta 4 lol

[upper crow]

know your algorithms... though, even that isn't as important anymore. Now it's more, "know how to prompt effectively" 😉

[solid pecan]

I would say logic is the most important thing actually, although if you work on hardcore deep tech, you need to be good at Mathematics as well

If you are building simple apps, then just logic

[solid pecan]

It still helps to know even with AI

[warm escarp]

in my opinion, not unless you're designing algorithms

[upper crow]

Oh, absolutely. I fear for the future when there are no more human programmers to continue to innovate.

"vibe coding" only gets you so far... because models are trained on the code we created in the past.

[solid pecan]

We are holding the fort at the moment, we have Website and SEO people thinking they are Software Engineers all of a sudden 😄

It is like dude, no your Wordpress is not hardcore Software Engineering

[upper crow]

BUT IT PLAYS MUSIC!

[warm escarp]

so much of that stuff is abstracted nowadays that unless you're writing the algorithms inside the abstractions, it's usually enough to just understand the outcome of the algorithm and the trade-offs you're making by using it, rather than the gritty details of how it works

(but again, I'm a security engineer and not a software engineer, so take my opinion with a grain of salt)

[upper crow]

writing code that works and writing code that is "efficient" can be wildly different things. 😉

[solid pecan]

It works to a point although when you have billions of data points etc, still need to understand Algorithms and Data Structures for it to scale

[solid pecan]

Yea this is what I was getting at

Maybe in a few years AI will improve and that will not even matter anymore

[warm escarp]

"if it's stupid and it works, it's not stupid!"

(one of the reasons I'm not a software engineer)

[solid pecan]

Although that is not today

[upper crow]

I'm absolutely sure it will

[zealous lagoon]

Nowadays you don't need to "know" your algorithms, just having known them once at some time in the past is enough. And that is AWESOME for people like me who went to school a Long Time Ago and swerved careers.

[solid pecan]

I think Codex in January was kind of a big deal, worked so much better at Claude Code for actual real work

[warm escarp]

nervous cough why yes, I could definitely implement a binary search tree in Java, why do you ask?

[zealous lagoon]

knowing what's appropriate, when. THAT is something the clankers don't seem to be very good at right now.

[solid pecan]

How about a Graph with DFS and BFS? 🙂

[zealous lagoon]

I have to push mine in the right direction.

[warm escarp]

same

but sometimes they push back! and sometimes they're right!

[twilit ridge]

I don't like doing LeetCode.

This makes me very upset.

[solid pecan]

Codex still cannot code as good as me although I cannot compete with something that just pumps out 1000 lines after one reply ... all the people still trying to say "I want to write it myself" are going to all loose their jobs if they do not adapt

[twilit ridge]

Every time I do an algorithm problem, I feel like I'm not cut out for computer science.

[zealous lagoon]

but that's good!
It disappoints me when I have to say "but shouldn't you do XYZ because abc" and the clanker goes "Good call, you're absolutely right!" (because of course it does) and corrects itself immediately.

[upper crow]

True. I've been in IT for... longer than most of you have been alive, I'd wager... worked my way all the way up the ladder to managing director over entire datacenters. Then, burned out and pivoted back to being a "software engineer" because that's what I always truly loved to do. Been doing that for almost 30 years now... only difference is now I get to work on projects I want to work on. 😄

[warm escarp]

okay, dumb question: I'm running OC on Windows, and it keeps kicking off empty Powershell terminals and I don't know why:

[process exited with code 0 (0x00000000)]
You can now close this terminal with Ctrl+D, or press Enter to restart.

what I do?

[ruby crane]

they also tend towards narrow solutions, rather than looking at the wider landscape. really need to drag them back to see what the implications of what they "fix" in one area does to another

[solid pecan]

If it makes you feel better, for LeetCode you really just have to memorise alot, almost like studying for an exam

[zealous lagoon]

yeah tunnel vision is kinda inevitable with such small context windows.

[upper crow]

yup, quantity over quality. AI codes faster, humans code "better". But, that's relative and subjective so... 🤷‍♂️

[ruby crane]

maybe someone can add a /rolledupnewspaper slash command

[warm escarp]

I was gonna say, I've seen some of the code humans write, and... :P

[zealous lagoon]

well that's the thing. Not everyone needs to be capable of folding and inventing algorithms, or even implementing them directly. But if you understand them you have a massive advantage over any vibe coder out there.

[solid pecan]

Yep like for performance code I still have to go through it very directly with Codex to get it right although if it is just a UI/UX thing etc, Codex will just overbuild these insane interfaces which is fun

[upper crow]

I use AI to code all the time, but I also check the work and either tweak the bits I don't like, or re-prompt to get it fixed.

[zealous lagoon]

UI/UX -- another facepalm area for me. they're so bad at it.

[solid pecan]

Same, comes out way better that way

[warm escarp]

"tired sigh no, this is not the right way to construct a SQL statement": stmt = f"SELECT * FROM users WHERE username='{username}'"

[solid pecan]

Although I wonder @upper crow if all our corrections are making it improve itself ...

[upper crow]

we can only hope 😉

[solid pecan]

They can be good if you really just give it an example from another project

[solid pecan]

Easy for you to say, you are retired 😄

[upper crow]

Lol, true.

[solid pecan]

I suppose I can be as well hahahaha

[warm escarp]

UX is interesting to me, because I've vibe-coded some UIs recently, and what I've come to realize is that I know what a good UX feels like, but... I really struggle to define it

I just know when it feels wrong

[upper crow]

just means I get to work on stuff I want to work on... for FREE! /s

[solid pecan]

Like fixing OpenClaw? 🙂

[upper crow]

HA, exactly.

[solid pecan]

At the moment it seems they wrote ActiveMemory in a non-performant manner ironically

[solid pecan]

Yea you still need human-in-the-loop, it makes strange choices that do not feel intutive at times although I find if you tell it to lookup proper UI/UX principels first, it does better

Also pasting sources from NotebookLM

[warm escarp]

hmmm, I should do that

[upper crow]

honestly, that's the biggest issue I have with most vibe coded projects. AI just doesn't seem to "get" UX yet... it'll build a beautiful UI, but it doesn't flow and does things like buries setting three or four screens deep. Makes no sense sometimes.

the AI is like... well, I can scan all the screens in under a second... how hard is it to find that settings?

[near owl]

hello people

as I said, 2 hours.. I am a man to my word

[delicate berry]

ive found a little bit of a hack here.
it only helps a little bit, but it lets you get the shape better first
have codex do an image gen creation of what a UI would look like that meets your criteria
then tell it to make an app based on that image

[solid pecan]

Yea it helps, just need to bootstrap it a little more for these more intutive type areas where the end-state is not well defined

[near owl]

I hate alarms, it dont wake me from sleep.

[zealous lagoon]

I know the ahem feeling. lol.
I've spent a lot of time bitching at my devs to make UIs decent. I'm a bit picky. It's hard to describe it without just doing it yourself.
-- The devs used to make fun of me "ooh here's peetie, he's come to ask for fancy colours lol". because colours matter. Rows of numbers in primary blue, red etc are impossible to focus on. etc... Then I'd ask the dev to show me his terminal setup or text editor. Sure enough... solarized colourschemes.. highly visible but easy on the eye. THAT'S what I'm asking for.

[upper crow]

great idea!

[solid pecan]

This works, sometimes I also get Codex to take a screenshot and say are you happy with how this looks? It will say No ... and I am like well keep going until it looks nice 🙂

[upper crow]

Colors, fonts, layout... it all matters in UI/UX

[warm escarp]

also - is there a trick to getting OC not to run like molasses? I keep getting stuck with half a reply on my screen - currently: "Let me actually"

let you what? LET YOU WHAT?!

[solid pecan]

I rembs saying to a Dev years ago "can you at least make the buttons aligned ..."

[delicate berry]

codex has a GREAT new annotation feature for zeroing in elements you do or dont like

[zealous lagoon]

@ebon atlas I think it was, did similar but next level imho. Describe what you want to Nanobanana or Openai-image... and get it to make a mockup. THEN get the LLMs to code it up to look like that. Even use snippets as assets.

[upper crow]

"can we get the logo in cornflour blue?"

[delicate berry]

lol thats what i just said. openai image gen 2

codex can basically natively do this. its like claude design but better

[zealous lagoon]

Oh.. carry on then. lol

[warm zealot]

img gen phh

[solid pecan]

It is amazing how important it is to the normies

[plucky belfry]

Gclaw everyone

[solid pecan]

I rembs moving from a v1 to a v2 many years ago and everyone was saying how much better it was when really it was just updated CSS ...

[warm zealot]

some people just need this, others want godocs lol

[upper crow]

This happens WAY more then we think.

[solid pecan]

I just like when it works ...

[upper crow]

UI tweak... NEW VERSION CUSTOMERS!

[solid pecan]

I still do not 100% get why it works so well ...

Like from a psychology standpoint

[upper crow]

because it looks new

[solid pecan]

Is it just added trust and credibility from looking better?

[upper crow]

yup

[warm zealot]

yup

[zealous lagoon]

I saw steipete post somethign on X once. it was a table.. names on the left, numbers on the right. dunno if it was a pricelist or TPS or god knows. but it looked like this...

|item 1                                   |                                        234|
|item 1                                   |                                        234|
|item 1                                   |                                        234|
|item 1                                   |                                        234|
|item 1                                   |                                        234|
|item 1                                   |                                        234|
|item 1                                   |                                        234|
|item 1                                   |                                        234|

And I see this crap all the time. Jesus you stupid clanker, why are you making me book a flight to scan across and check the values.. and by the time I've got there I've no idea what row I was on.

[upper crow]

" oh, look... they changed this icon and the font looks cleaner... they're WORKING on it! Must be better."

[near owl]

Didn't I hear before that openclaw owner works with openai and openclaw is falling?

[solid pecan]

At least Oracle ERP looks slightly and I just mean slightly better than SAP? 😄

[delicate berry]

you may have heard it but whatever you heard was not based the slightest in reality

[upper crow]

not saying much 😅

[warm zealot]

I mean....

[delicate berry]

the founder of openclaw works at openai, but openclaw is not falling, its getting stronger than ever

[zealous lagoon]

It's very important in a high-workflow environment when you're staring at screens all day. You want high visibility, instant-recognition of information and a colourscheme that doesn't burn your eyes out.

[warm zealot]

its not 100% not based in reality

[plucky belfry]

Sama bought Petie for 1 billi

Stacked

[solid pecan]

I kinda feel the minimalist trend went too far though, at least for the normies I dealt with, hiding things behind layers meant they could not find them

[upper crow]

https://tenor.com/xoKR.gif

[solid pecan]

At least in my own user testing

[zealous lagoon]

So you agree with me then!!! Humans can parse and selectively ignore a lot of information. When they hid all the menu items in Excel 97 or whatever it drove everyone insane.

My other bugbear.... dropdown lists that only show like 5 items out of a list of 276. Settings popups that are like 600x400 and not resizeable. But have several list panes and the entire window has a scrollbar. what the fuck

[upper crow]

lol

[solid pecan]

The argument I had against that which was true to some extent is that people also suffer from cognotive overload though and do not like an interface if there is too much they need to parse, they run out of bandwidth

Although if you hide things then thay do not know they are there so I found you kinda need to walk the line between the two extremes

[zealous lagoon]

There is a balance to be had. And this is achieved with a GOOD ui. selective enhancement and obfuscation without actually hiding anything.

[near owl]

AI youtubers are the worst then, David Ondrej.
https://www.youtube.com/watch?v=1nDiiXfMUK4&t=113s

[upper crow]

good UI/UX is a delicate balance

[ruby crane]

i think the ribbon came from a design study that measured how fast new users can get something basic done, not how well existing/skilled users can. in that context, ribbon probably was better. one problem... every man and their dog already knew where to find shit in excel

[near owl]

and I heard openclaw can edit it's own code and give it's own features, is that true, like claude code?

[ruby crane]

data driven-design is only good if you're looking at relevant data

[upper crow]

yes... if you install via git.

[delicate berry]

youtubers? what? they would never exaggerate!

[teal cradle]

So how is 5.2?

[near owl]

at 0:57 is when he said it sucks

[ruby crane]

much better than anything since 4.23 and 5.3 is likely to drop today, which is better still

[zealous lagoon]

Kinda. I mean it literally can but you normally wouldn't want it to. But it CAN make its own "skills" which are basically a library of cheat-sheets to remember how to do certain complex tasks using tools or the internet.

[solid pecan]

Perhaps unpopular although I kinda liked the ribbon update after a while ...

[delicate berry]

be careful listening to people who get paid by you listening to their hot takes

[teal cradle]

nice, ya had issues with 4.24 and been waiting for a stable version

[ruby crane]

familiarity gets you there in the end. it'd probably be hard to go back to the pre-ribbon design (that was to @solid pecan )

[near owl]

Mhmm, I am never watching AI youtubers man. They are all bad.

[upper crow]

not all, but the signal to noise is very high.

[near owl]

Dan Martel seems promising.

[fiery nacelle]

Technically if you pull in the repo locally and have you Agent modify that and apply the patches to the live install, it would then be "modifying its own code"

[solid pecan]

Yea, I think I can just find things faster with the Ribbon UI for certain common functions

Although I just use Linux anyway so do not even have Ribbon anymore 😄

Just LibreOffice ...

[delicate berry]

not all, just be sure to take what they say with a grain of salt, and do independent research for anything overly spicy. your claw with a search engine like brave searchx in llm_context mode can give you a real easy way to cut through the noise

[ruby crane]

even it has the ribbon... kinda

[upper crow]

I mean, I've had my Claw patch it's own code to fix a bug...

[zealous lagoon]

Once you've remembered where things are it isn't so bad. But when they introduced it they just moved everything from where it was and took away the familiar menu options. productivity -95%

[solid pecan]

Yea, it is kinda, more like a cross between a toolbar and Ribbon

[solid pecan]

Yea I rembs it took a while for people to adjust

[bright bear]

This is the listing team from BitMart. Interested in discussing Listing Cooperation with a Special Offer. Could you direct me to the Business Manager or core team?
Thanks.

[ruby crane]

i think most of us have probably done it. the issue is that you're not quite using openclaw at that point. it's Openclaw-Donovan Edition

I try to resist unintended customisation in general. leaves you on the island, not the mainland

[upper crow]

okay, technically true

[solid pecan]

I been maintaining a bunch of my own patches as well although it makes it harder to maintain.

[zealous lagoon]

but more importantly they started hiding menu options that you didn't use often. You could only find them by holding the menu open for a while or clicking [...] or something, I forget. But we don't tend to remember exactly where they were, we just open the menu, don't see what we want.. go "huh?" open another menu... and another... throw the mouse at the monitor and scream "where's the fucking sort menu gone!?"

[solid pecan]

Just keep clicking stuff until you find it 😄

[upper crow]

same. I just rebase my custom branch when updates drop.

[zealous lagoon]

we also don't tend to recognize things as quickly when they show up in unfamiliar places. context matters!

[upper crow]

but, with the latest beta... I dropped my custom changes as most of them have been addressed.

[zealous lagoon]

u should def get some PRs up. I mean, you did the work already...

[solid pecan]

It is not fun though, I was hoping OpenClaw by June would be as solid as a rock

[ruby crane]

when they added the search box to office it made it easier. catching up with a system-wide mac feature basically (searching the menus in the help menu)

but taking it a little further (finding things not in the menus)

[solid pecan]

Some of them are just custom things that probs should not be in the main, I have this one that always no matter what tells me when a process is done within the agent chat

[upper crow]

could still be... we'll have to wait and see

[zealous lagoon]

is cool.. I think that might be part of the emoji reaccs now.

[solid pecan]

I think this was a good wake up call

[ruby crane]

I reckon it will be in a very good plce by June

[solid pecan]

Yea I think if they can get it right by June it will be solid

[zealous lagoon]

They're focusing on stability now. after xitter and reddit blew up, and losing a bunch of users. I think we'll be pretty darned solid by June.

[upper crow]

Has anyone else's GPT gone from 'goblins" to "raccoons"?

[solid pecan]

They lost alot of users to Hermes ...

[zealous lagoon]

But also architecturally it will be easier to build around. And that's the change that started the mess.

they ripped the band-aid off.

much needed. Just maybe implemented a little too hastily

[upper crow]

I'm running both... sticking with OC too, they do different things

[solid pecan]

Yea, I prefer OpenClaw personally, I have Hermes in an isolated session, currently have not really used it much, I kinda not sure if I like talking to it

😄

[near owl]

define OC

[upper crow]

Personality mostly depends on the model you use, in my experience. Sure SOUL gives it a bias, but the model really determins how it "feels"

[solid pecan]

Yea I am kinda attached to my OpenClaw agent now

[upper crow]

I've tried a bunch of different models... and, boy, some of them are... 😬

[delicate berry]

"an agent" is not the same thing as a claw

[terse sapphire]

Clawpilot lol

[upper crow]

True. it's very hard not to anthropomorphize the AI

(ironically, I had to ask AI how to spell that... 'cause I couldn't remember)

[ruby crane]

beta-4 trip report. all's well.

• same liveness warning on agent first turn. fine after
• health, session list etc. are 300ms max but typically much less
• Usage page is faster than 5.2
• Discord still does a weird "Clawd is typing" for a few seconds after the agents turn ends

overall, all good

[zealous lagoon]

You're hired.

[delicate berry]

beta.4 already? lol i just got .3 fixed

"fixed" meaning responding somewhat normally

[ruby crane]

if i knew what they're doing between betas, I'd focus on testing that. I feel like, "it's still fine for me" isn't all that useful as feedback lol

[delicate berry]

was discussing it down in #clawtributors with @zealous lagoon

[ruby crane]

yeah i see that but I don't have those perf issues

I am, perversely, a bit mad mine isn't broken too 😂

most of my health/cost/usage/list are 100-ish ms. 300ms is the typical upper bound with very occasional outliers as high as 800ms

[solid pecan]

@ruby crane how you finding the performance in terms of responses, does the ActiveMemory make it feel better?

Like is the agent itself better?

[ruby crane]

I haven't re-enabled active memory since the 4.23+ slowness because the slow turns were slow x 2 with it. honestly, I wasn't ever sure it was making a worthwhile contribution. I could see it was finding things and injecting, but it's a black box and never sure if the injected context was useful

[solid pecan]

Ah fair enough

[ruby crane]

perf-wise? it's WAY better. WAAAAY better

[solid pecan]

Oks, helps to know active memory is off, I will probs turn it off while upgrading when I do

They should probs have openclaw doctor turn it off by default for people who upgrade for now

[ruby crane]

I might enable it again tbh. sometimes, even with gpt-5.5 medium thinking on, bro has started tools calls before my hand has left the enter key (poetic exaggeration but it's fast)

i could tolerate another 0.5-2 seconds for gpt-mini context injection but I'd me more inclined to if I could measure (or at least see) what it's injecting and whether it's useful

[solid pecan]

Yea, when I upgraded it was filling up almost my entire context window on injected content, probs also needs some sort of hard cap

To make sure it does not just overload the models, also some local models only have smaller context windows

[ruby crane]

oh yeah, it's configurable in scope and i left it on a conservative setting

it was only injecting small amounts per turn (which makes me even more skpetical it's useful)

[solid pecan]

Yea, I am also wondering if it is the right approach

Probs should try use vector memory more

[near owl]

I want to give openclaw arduino tools to interact with the physical world, I don't want something so advanced because I am an arduino beginner but what should I give it?

[ruby crane]

I like the idea of autonomous recall. I'm not sure that the trade-offs are worth it, or whether this implementation is useful or not and I can't see a way to prove the latter

[solid pecan]

Play some music maybe using some sort of MIDI and speakers?

[ruby crane]

wheels, a camera, a knife and some butter. you know what to do

[solid pecan]

Maybe when we have more 1M context windows and above it will be?

[near owl]

I do have a camerqa

I am not giving it a knife, nor butter, not wheels though..

[ruby crane]

yeah maybe if context itself was less precious and you don't need to manage it so closely

[ruby crane]

shame https://www.youtube.com/watch?v=xerLPWdyX-M

[solid pecan]

Eventually it will probs go that way

[near owl]

I dont like music, create another idea. Something fun

Who here uses openclaw for fun?

[brittle hound]

Give it hands

And a baby. And tell it to nurture the baby to let it grow

[near owl]

to do what, I dont got all the components for hands..

[brittle hound]

Self limiting beliefs

Give it what you do have and the ability to improve itself

[near owl]

hmm

[zealous lagoon]

I got a ESP32-S3 kit with a speaker and microphones. Makes a good voice endpoint, supposedly. You can build your own firmware or hack the ones that are already available. You can also give it a camera and/or a screen... I feel the claw could do a lot with a camera. Especialy if you can hook it up to some actual vision models

[near owl]

I have an ESP32 and ESP32 Camera

arduino r3 and arduino r4

[zealous lagoon]

You could have it recognize people and shoot them with a water gun or something. lol. or give it wheels so you can call it to you

[near owl]

I have been imaging giving the mac mini legs, camera, mouth and ears.. like a spider

I would be leaving this earth if I knew an openclaw mac mini can just walk to you and.. yeah it's terrifying

makes a good assistant

[zealous lagoon]

ooooh I have a watcher script for refurbished Mac Minis on my local Apple Store website... I got notification of FIVE of them this morning, ranging from the 16-256 base model to a 12+16c, 24-512 model.

... of course they were gone by the time I clicked on them, but the wave of returned claw-machines has definitely started lol

[toxic elk]

Congratulations on turning pink!
You should have set it to automatically place an order upon capture.

[ebon atlas]

Whos trying to hax the National Reserve with me using openclaw and get quadrillion dollarz

[ebon horizon]

Ask Dr. Evil

[ebon atlas]

im trying to put together a team of only the best prompt writers lmao

[fossil sandal]

I keep getting the 'model idle timeout' error when using google/gemini-2.5-flash-lite. Is this caused by my server performance being too slow, or is it a known issue with the 2.5-flash-lite model's response time?

[near owl]

@placid elbow

[ebon atlas]

5.2 must be doing good, chat is quiet, not enough complainers.

[fossil sandal]

cron is broken.

[near owl]

telegram aint fixed

[ebon atlas]

here they are 🙂

[fossil sandal]

v 5.2 😮‍💨

[near owl]

it keeps resending messages

[near owl]

bro, I wanted to use cron.

[toxic elk]

My agent is reporting previously completed tasks again, and I can't seem to find the reason why

[near owl]

MY CRON!! :(

should I switch back to 2026.4.26 version

[toxic elk]

Are you still using the previous ID naming convention for your sessions? For example, Discord sessions used to include the word "discord." After I updated to version 5.2, the word "discord" no longer appears in the session logs for Discord.

[ruby crane]

In 2026, a crack AI unit was sent to prison by a military court for a crime they didn't commit. These men promptly escaped from a maximum security stockade to the Los Angeles underground. Today, still wanted by the government they survive as soldiers of fortune. If you have a problem, if no one else can help, and if you can find them....maybe you can hire The P-Team.

[ebon atlas]

https://tenor.com/view/cat-computer-typing-fast-gif-5368357

[ruby crane]

i know you all heard the music when you got to the end

(at least all of you old enough to understand the reference)

[glass lark]

Is krill back alive? What’s the website to check it again?

[ebon atlas]

is recruiting to your company considered self promo lmao

[inland mist]

Looks like a state consistency issue agent is reporting already completed tasks again.
Probably missing proper state tracking

[ebon atlas]

if i find 1 more ant in my room im burning the place down istg

Anyone else aalso tistic and can smell ants?

[near owl]

how do I change my version

[ebon atlas]

openclaw update --tag 2026.x.xx

[near owl]

thanks

[toxic elk]

I'm not sure; I asked it to check for issues, but it didn't find an obvious root cause

[ebon atlas]

Change the Xs to the date you want the update of like 4.23

[near owl]

thanks man

[inland mist]

what does logs say?

[near owl]

is it possible to switch versions in real-time as well?

without installing new versions or do i need to reinstal 5.3?

[ebon atlas]

I think you can switch back to 5.3 without reinstalling because it doesnt delete the old files if im correct.
But for aany new file added you have to ofc update again

like you will still have to do the update command but i dont think it will like download it or anything

[violet mulch]

did you guys have to disable the pdf-tool to get your openclaw working faster?

[ebon atlas]

PDFs take up a lot of your context, maybe they are still in your seessionn , slowing it down.

[violet mulch]

this is a fresh install

ive never used the pdf-tool before with openclaw

[ebon atlas]

are you planning on using it or just noticing its slowingg you down? if you dont plan onn using it then just keep it disabled, maybe something off with it this update.

[near owl]

[5/4/2026 9:04 AM] Carbon: run this command

openclaw update --tag 2026.4.23
[5/4/2026 9:04 AM] Openclaw: I need to run the update command in a TTY environment to confirm the downgrade. Let me do that now.
[5/4/2026 9:05 AM] Openclaw: Let me check the status of this process and provide any necessary input for confirmation.
[5/4/2026 9:05 AM] Openclaw: Let me check the logs to see what confirmation is needed.
[5/4/2026 9:05 AM] Openclaw: The update process is waiting for confirmation about the downgrade. I need to select "Yes" to proceed. Let me send the appropriate input.
[5/4/2026 9:08 AM] Carbon: /status
[5/4/2026 9:08 AM] Openclaw: 🦞 OpenClaw 2026.5.2 (8b2a6e5)
🧠 Model: vllm/nvidia/nemotron-3-super-120b-a12b · 🔑 api-key (vllm:default)
🧮 Tokens: 78k in / 280 out · 💵 Cost: $0.0000
📚 Context: 16k/128k (12%) · 🧹 Compactions: 0
🧵 Session: agent:main:telegram:direct:8398889290 • updated just now
⚙️ Execution: direct · Runtime: OpenClaw Pi Default · Think: off
🪢 Queue: steer (depth 0)

now what else?

still stays 5.2

[ebon atlas]

you do that in the terminal not the webchat

webchat doesnt have all sudo commands , sometimes even if you confirm

[near owl]

it did update itself though to 5.2 before

i got it to update itself to 5.2 when i realized the update.

[ebon atlas]

update yes , it even has a nice buttonn at the top, but i dont know about downgrading, im also not a developer just tryna help u

[toxic elk]

The same issue just occurred again; everything appears normal, and there are no noteworthy error logs.

[fossil sandal]

use version 4.23 forever.

[toxic elk]

Well, I think I've found the reason. The first task report wasn't submitted by the main agent—I'm not sure where that one came from. The second report was the one actually submitted by the main agent itself, and it's recorded in the session logs.

[violet mulch]

gateway takes too long to start 😭

[ebon atlas]

Say please and try again

Also your context window may be too small for updates. I don’t know, don’t ask me. But you have 128k . Codex is 256k . If you are able to switch to OpenAI-codex just for the update and then try again . Idk some guys mention that here sometimes

[violet mulch]

i have codex setup

[ebon atlas]

It says your telegram is running off of nemotron

[boreal thicket]

Crap it's almost monday why am I awake

[oblique root]

So what’s the verdict on 5.2, broken?

[vagrant island]

Anyone else's agent just forgets to use the message tool? It's getting annoying

[zealous dawn]

i cant use open claw what happends its dosent work whit any model i use nemotron and dont work i am desesperated and i see that in the help chanel they dont help nothing

[placid zodiac]

What OS are you using ( mac, Ubuntu)?

[zealous dawn]

wsl

and they dont work nothing

[placid zodiac]

down load the codex app

[zealous dawn]

i dont whatn to pay api key i whant to use my local models

i not millonary

i whant to use ollama local models

[placid zodiac]

ollama launch openclaw

use that bash

[zealous dawn]

yes but the model dont do nothing they stay thinking 10h and the dont do nothing

[frozen holly]

hey folks, new beta is up!

[placid zodiac]

patch notes?

[zealous dawn]

🦞 OpenClaw 2026.4.15 (041266a) — The lobster in your shell. 🦞

openclaw tui - ws://127.0.0.1:18789 - agent main - session main

session agent:main:main

hola

⠙ pondering… • 48m 2s | connected
agent main | session main | unknown | tokens ?/131k
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────

────────────────────────────────

look i not lier

[placid zodiac]

Close everything out, and run - ollama launch openclaw

[ruby crane]

anything in particular to focus tests on?

[near owl]

Hello!

[gritty summit]

hey carbon, how is nim treatin' ya

[near owl]

Is there any websites where only AI can read but a human can't, it would be a sick vibe. Like a shop website but you need to tell the AI to get the content for you and give you the responses.

[orchid bear]

The openai/gpt harness timeouts, but it seems not yet fixed at my side.

[near owl]

The websites in #self-promotion got "I am human" and "I am agent" but we need I am agent, remove the damn css, html, just markdowns and unreadable to humans.

[orchid bear]

Still massive problems with openai/gpt.. runtime codex. Any recommendations?

[zealous dawn]

i wait a lot of time and they dont do nothing -bash: /home/tovi/.openclaw/completions/openclaw.bash: No such file or directory
tovi@MSI:~$ openclaw gateway restart

🦞 OpenClaw 2026.4.29 (a448042) — Alexa, but with taste.

Restarted systemd service: openclaw-gateway.service
tovi@MSI:~$ ollama launch openclaw
This will modify your OpenClaw configuration:
/home/tovi/.openclaw/openclaw.json
Backups will be saved to /tmp/ollama-backups/

Launching OpenClaw with ibm/granite4.1:3b...

Starting your assistant — this may take a moment...

Starting gateway...

✓ OpenClaw is running

Open the Web UI:
http://localhost:18789

Tip: connect WhatsApp, Telegram, and more with: openclaw configure --section channels

🦞 OpenClaw 2026.4.29 (a448042) — The only crab in your contacts you actually want to hear from. 🦞

openclaw tui - ws://127.0.0.1:18789 - agent main - session main
connecting | idle
agent main | session main | unknown | tokens ?
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────

───────────────────────────────

what i have to do?

[near owl]

nim is treating me good

better than openai in agentic

[gritty summit]

you can manually add more models than what shows up in the config/onboarding

[ruby crane]

anecdotally, I've seen quite a few people to say to prefer the PI harness for now https://docs.openclaw.ai/providers/openai

[near owl]

nemotron super is the one that doesnt get used the most by people

glm 4 is crazy used

[gritty summit]

i guess... i really like the big qwen 3.5 too

[near owl]

and deepseek is the #1 worst tool calling cloud model ever

[gritty summit]

id like to use kimi 2.6, but it just randomly starts spitting out "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!"'s forever

[near owl]

i got promised it was so smart, 1 other tool call and it was just text.;.

chinese models sucks, except qwen.

[marble ledge]

I'm a freak that still prefers the pi harness, I think. I tried switching back to the codex harness, and it's good, don't get me wrong, but it still feels a little more sluggish than the pi harness.

[near owl]

qwen is the best chinese model

grok is uhmm. very self-aware but people say its dumb.

[ruby crane]

I'm on it purely because it's an old install. I asked here a few days ago about changing to the codex one and a few people said don't do it

[gritty summit]

im wrapping up setting up embeddings memory for my claw...

[near owl]

I need use cases for openclaw

[ruby crane]

hmm. no beta 5 docker release yet

[near owl]

I just tell it create me a website that has no purpose..

should openclaw do my online schoolwork for me?

[marble ledge]

It's getting there. Like every update it feels closer to just what we were use to as "native." But I don't know if it quite feels like it's totally the vibe yet. I'm gonna keep running it for a bit to feel it out.

[gritty summit]

if only it could clear my 3d printer buildplates and start new prints

[zealous dawn]

help chanel is working or not?

[near owl]

lets all come together to call this guy a peace of shit for using gpt-5.5 when I EXPLICITLY told him to use nemotron.

[5/4/2026 10:12 AM] Carbon: use vllm/nvidia/nemotron-3-super-120b-a12b model for all the ai agents
[5/4/2026 10:13 AM] Openclaw: Working…
• sessions_spawn label test2, task echo test, model openai/gpt-5.5
• tool: sessions_spawn
• sessions_spawn label test2, task echo test, model openai/gpt-5.5

[gritty summit]

i dont like how it doesnt really show you a chat... idk

[marble ledge]

It's pretty damn snappy on the latest main. Can't really complain. I haven't prodded how it interacts with the rest of my setup, but we'll see. It's pretty damn smooth now.

[craggy bane]

Finally solved my docker and cron job issue

[near owl]

openclaw can't contribute to me 😢

it is just a toy for me to use rather than a time saver. Maybe I'll find out real use cases after playing with all of it's features?

What tool can I make for people with openclaw?

[zealous dawn]

can some one help me ? tovi@MSI:~$ ollama launch openclaw
This will modify your OpenClaw configuration:
/home/tovi/.openclaw/openclaw.json
Backups will be saved to /tmp/ollama-backups/

Launching OpenClaw with ollama/ibm/granite4.1:3b...

Starting your assistant — this may take a moment...

Starting gateway...

✓ OpenClaw is running

Open the Web UI:
http://localhost:18789

Tip: connect WhatsApp, Telegram, and more with: openclaw configure --section channels

🦞 OpenClaw 2026.5.2 (8b2a6e5) — If you're lost, run doctor; if you're brave, run prod; if you're wise, run tests.

openclaw tui - ws://127.0.0.1:18789 - agent main - session main

session agent:main:main

hola

⠧ bamboozling… • 2m 4s | connected
agent main | session main | ollama/ollama/ibm/granite4.1:3b | tokens ?/131k
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────

───────────────────────────────

[kindred cove]

The latest batch of "mini" models are performing on par (speed and precision) with the closed-source hype anyways. GPU prices about to climb again.

[zealous lagoon]

Problem with the 5.3 (beta):

  In 5.2 the host could apparently execute TypeScript entries directly (via tsx or similar). That path was removed in 5.3, so both plugins broke on upgrade.```

tl;dr it breaks discord and brave unless you build them yourself before running.  Hopefully fixed for release.  Meanwhile beta testers should build those plugins.

[kindred cove]

As all the providers and the data centers do away with there free previews the big guys are in trouble.

[zealous dawn]

can some help me?

why are ignoring me?

[zealous lagoon]

I can see you're trying to run ollama with openclaw right? I'm not sure what it is but it's not a config I am familiar with.

[zealous dawn]

but i like to get help i not millonary i dont have money to pay api key

that is why i use ollama local models

[zealous dawn]

and some one can help me?

tovi@MSI:~$ ollama launch openclaw
This will modify your OpenClaw configuration:
/home/tovi/.openclaw/openclaw.json
Backups will be saved to /tmp/ollama-backups/

Launching OpenClaw with ollama/ibm/granite4.1:3b...

Starting your assistant — this may take a moment...

Starting gateway...

✓ OpenClaw is running

Open the Web UI:
http://localhost:18789

Tip: connect WhatsApp, Telegram, and more with: openclaw configure --section channels

🦞 OpenClaw 2026.5.2 (8b2a6e5) — If you're lost, run doctor; if you're brave, run prod; if you're wise, run tests.

openclaw tui - ws://127.0.0.1:18789 - agent main - session main

session agent:main:main

hola

run error: LLM request timed out.
connected | error
agent main | session main (openclaw-tui) | ollama/ollama/ibm/granite4.1:3b | tokens ?/131k
────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────

──────────────────

what happends run error: LLM request timed out.
connected | error
agent main | session main (openclaw-tui) | ollama/ollama/ibm/granite4.1:3b | tokens ?/131k

[kindred cove]

Close other instances and use "Ollama serve" that's where you'll find the useful logs, urls etc.

[zealous dawn]

but i like to use whit open claw

tovi@MSI:~$ ollama serve
Error: listen tcp 127.0.0.1:11434: bind: address already in use

what i have to do?

[ruby crane]

well, 2026.5.3 final/rc trip report. same as through the betas: -

• first agent turn shows a delay but after that all channels are fast
  2026-05-04T17:39:08.567+10:00 [diagnostic] liveness warning: reasons=event_loop_delay interval=30s eventLoopDelayP99Ms=282.9 eventLoopDelayMaxMs=4978.6 eventLoopUtilization=0.422 cpuCoreRatio=0.466 active=1 waiting=0 queued=1
• all session/channel/usage/health etc. look healthy in timings
• discord continues to follow the actual agent response with a short "Clawd is typing" before disappearing. functionally all good though

One thing that did come up. My own plugin failed with this from first beta. I saw this

package install requires compiled runtime output for TypeScript entry ./index.ts: expected ./dist/index.js, ./dist/index.mjs, ./dist/index.cjs, ./index.js, ./index.mjs, ./index.cjs

I repackaged, republished and assumed it was just my bad but latterly someone reported an error with the bundled brave plugin that seems to echo it here
#1500705357243355308 message

so there might be something bigger there

[trim sorrel]

I feel like Claude cowork can search the web and research topics better than openclAw using Claude API.

Is there some plugin I need to give Claude to improve it's research ability?

I don't think Claude is using paid APIs for calling web searches?

[ruby crane]

other than the plugin weirdness and the discord stutter, I'm a happy camper here. notable improvements

[kindred cove]

Yeah but Ollama and Openclaw are two different things... You need to understand which is failing

Close the existing Ollama service, if not the little tray icon in the task bar , use task manager then try again.

[zealous dawn]

and how i can know what is failing?

[kindred cove]

When using "Ollama serve" it will show you if openclaw is actually trying to reach the LLM, since the error shows "timeout"

[zealous dawn]

and i do that and what i have to do next?

[zealous lagoon]

This is an interesting point. then maybe OC can use Claude CLI to do web search lol. I'm gonna try

[ruby crane]

probably no maintainer will see this and people will go ham when plugins break on 2026.5.3 but i didn't capture enough at the time to make a useful issue

as I assumed it was just bad packaging on my part

[twilit ridge]

To be honest, I don't need that many channels at all.

[zealous lagoon]

the maintainers hardened the behaviour of OC so whoever did that I'm sure they recognize they need to also bundle the compiled plugins.

hoping.. anyway... lol

[ruby crane]

you'd hope but... lol

i also noticed half the plugins on clawhub are scanned as broken now (perhaps because of that)

[jade jasper]

what's the haps today 🦞

man im putting off updating seeing all these breaks, do I just rip the bandaid and do it? Ive got my setup working so cleaaaannn for the moment

[zealous lagoon]

are you feeling adventurous?

[ruby crane]

don't get the wrong impression. 2026.5.3 overall is very good

[zealous lagoon]

5.3 should be out very soon. I think it will fix most people's stuff.

we're still testing beta, which is why we have problems. they are being fixed as we speak

[ruby crane]

little bit of a pause on first turn after start, slightly funky typing indicator on discord. maybe something with some plugins (just a theory for now). overall looking very good with notable improvements

and, yep, any or all of these could be gone by final