clawtributors

[crude matrix]

I appriciate every single one of you, you spend your free time building something amazing. just feel sometimes that some core functionality should be more priority, that is just what I try to say in my own clunky way.

I would appriciate if someone could look at these 2 issues: https://github.com/openclaw/openclaw/issues/84536#issuecomment-4515450826 and https://github.com/openclaw/openclaw/issues/84607#issuecomment-4524660698 thank you in advance.

[earnest jasper]

Finally ready for maintainer review...
https://github.com/openclaw/openclaw/pull/86210 - feat(memory): add multi-slot memory role architecture.
Overall: 🐚 platinum hermit
Proof: 🦞 diamond lobster
Patch quality: 🐚 platinum hermit
feature: ✨ showcase

Took some back and forth with Clawsweeper (is a tough crustacean to make happy!) but it finally has every angle covered. exhaustive testing passed! ;3

[worldly tusk]

experiencing a massive improvement in performance on main. thanks!

[sick phoenix]

i would love some help if a CLAW expert is in the house. the new 5.22 update REKT my CODEX CLI auth and it will not stay in the late properly. ive ben trouble shooting for hours and cant seem it get ti correct

#1508622578485297234 message

[pallid summit]

Back to back PRs, all passed clawsweepers and diamond/platinum quality, PTAL thanks!

fix(docs): stop redirect from shadowing MCP search endpoint - https://github.com/openclaw/openclaw/pull/86625
fix(agents): unwrap standalone message tool JSON - https://github.com/openclaw/openclaw/pull/86626

[vocal night]

These are always a good read

https://github.com/thanhtantran/agents-radar/issues/96

[vocal night]

When you get the diamond and platinum, and then update PR description to make it easier to understand…. 5 mins later clawsweeper removed and relabeled “Gold shrimp” 😆 and oh yeah main moved 1200 commits and conflict.

[odd anchor]

https://tenor.com/view/cat-cute-cat-screaming-crying-no-gif-11037260769437999806

[pallid summit]

And I'm feeling lucky today since "It’s 69 commits behind (base commit: 0f35ec2)"

[vocal night]

Morning yall. I’m just glad to see people focused on fixes and performance. This next beta is likely to be 🦞🔥

[shy rain]

you mean back to normal, right? I pulled main and so far it mostly works. It's much snappier at least.

[vocal night]

I'll take any stability I can get haha

I attend conferences and meetups weekly and first question always is "when can I safetly update for my biz"

[shy rain]

LTS releases will be very welcome

[vocal night]

this might be the one LTS for awhile, so much refactoring and changes coming still

[shy rain]

Well good, at least we can then focus on fixes, speed and stability ... and update with less worry about everything else breaking at the same time.

[shell dawn]

Maintainer review request for #86458: https://github.com/openclaw/openclaw/pull/86458

Current status: mergeable, CI green, ClawSweeper says proof sufficient / platinum hermit / ready for maintainer look.
This fixes chat.history returning oversized display payloads from tool-heavy sessions.
It bounds the display path and keeps raw history behind admin-only raw mode.
The PR is XL by size, but the behavioral fix is intentionally narrow: projection, docs, protocol regen, and focused regressions.
Main maintainer call: accept display clients defaulting omitted chat.history requests to turns, or keep legacy messages default.

[vocal night]

Does this fix the issue of web chat history only being like two messages and 98 tool calls and then no other history?

[shell dawn]

Yes

[vocal night]

@stable pewter if you're around or still awake, I wanted to get your input on this. One of the issues now because of codexs message history, you get like two or three messages in web chat and that's it because the rest are taken up by hidden tool calls (kills user exp for webchat users).

We used to have long-term message history and now we only get like the last one message or two messages from the agent and the rest are all tool calls and everything else gets swallowed by the UI and there's no way to show past messages in web chat.

[shy rain]

ok /new went from 2.5 minutes to ~3 seconds. That's good.
Unfortunately a simple ping-pong exchange takes 27 seconds (vs codex 2.5 seconds)

Keep dem performance patches coming, guys... This is not yet a viable release.

[vocal night]

vs codex cli/desktop you mean?

and did you check tool run too for what 27 second cost was, if you've cleared the other issues and running main, just trace every call from start to finish in those 27 seconds @shy rain

[shy rain]

yes vs codex cli. I'm doing repeated call/response messages now and grabbing timings. They vary a lot but even the quickest was 5-6x slower than codex cli.

[vocal night]

For comparative with codex, what you're going to want to do is, it's always easier if you have a third harness that's running it. So like if you have codex desktop tell it to take control and test the codex CLI versus the OpenClaw, you basically tell codecs desktop, "Okay, we're hitting these long response times. What you need to do is test both of them versus each other and see what the response times are."

For example, you need to load the same context in the CLI. So because the OpenClaw has a heavier prompt versus the CLI only has 16 kilobytes of prompt, you need to load the context percentage and the prompting to be about the same.

Otherwise, a fresh codex session with almost zero prompt isn't relative speed wise to an OpenClaw fresh new session running off of, you know, usually like a 60,000 character bootstrap plus all the system prompts.

Also, hot versus cold cache is important. One of the reasons the times are so high is because my hot cache thread context fix isn't implemented PR wise. So every turn you're getting basically a cold start or cold boot. So it's like, "Hey, this is a fresh thread and this is the first time I'm seeing this information."

[shy rain]

yeah the worst was cold start, warm continuation is not as bad. Discord is a huge problem, but gui chat isn't much better.

[vocal night]

If convo = or > 70k context (pretty much every convo as thats 32% full) - do thread swap - do cold start and or changed assembly bugs which can happen every message (token expensive).

[true laurel]

Woop, noticed my PR got merged 🙂

[vocal night]

That's largely what I've been trying to work around we have the 70k thread swap guard because otherwise Codex only allows its own context + assembly and because the Codex team doesn't want to make changes for us in Codex to make it more friendly to "ride" it's squeezing a cucumber onto an apple. Doesn't work unless you tapem together lol

[vocal night]

@fallen gate we need a contributor tag!

[rocky pewter]

#announcements message

[odd anchor]

https://klipy.com/gifs/mulan-kantoi

[true laurel]

My agent has been complaining about this too all weekend. It started cleaning up memory.md because "surely that would fix it". 😅

[mortal hollow]

Guys is it illegal to use discord under 13

[true laurel]

Yes, and depending where the user is from it can be as high as 16, check https://support.discord.com/hc/en-us/community/posts/360050817374-Age-restriction

[mortal hollow]

Ok

[shy rain]

both Discord and GUI chat are adding 10-12s latency on my side. But since I'm on dev main, GUI chat is flaked because the bot is told to use the messaging tool instead of reply, when it doesn't need to lol. But also codex app server is timing out after 4-5 messages making it unusable. *sigh

[mortal hollow]

I’m over 16 I was just checking

[shy rain]

If only that were the only story lol. Perf has definitely improved, that's true. It's almost back to what it used to be. Usable at least. Unfortunately there are other issues due to it not being a proper release... now I feel like I've dug myself into a hole. Nothing since 5.12 has really worked for me, and even that was a dog. My claw is basically unusable right now so I'm praying for a new beta at least, that I can roll back to. :/

[crude matrix]

I would appriciate if someone could look at these 2 issues: https://github.com/openclaw/openclaw/issues/84536#issuecomment-4515450826 and https://github.com/openclaw/openclaw/issues/84607#issuecomment-4524660698 thank you in advance.

[fallen gate]

your potatoe is causing chaos for everybody it seems

[shy rain]

my potatoe is a canary for the thousands of people using a 2-core VPS

[vocal night]

Rather than one size fit all, we should consider a detect system spec > turn on potato mode. That would allow us to instead when detect potato on install/update > it runs much leaner and cuts out the latency noise from systems that should be disabled as they don't work on pi or 2 cores

It's really, I thought about this just now and you really put the, the, the, the, it's in the perspective. Um, we're always trying to one size fit all everything and that leads to bad architecture.

Limited performance or potato mode is actually the best solution here.

[shy rain]

Yes, I was thinking something similar. But given the mess that is the code we've been tidying up, I think that makes the "devs" lazy. They just write slop code that re-reads thousands of json files every single tick, because an M4 machine can do that and get away with it... I'm somewhat against letting that kind of rubbish through just because modern machines can power through it, when it's obviously awful practice.

[vocal night]

I agree and disagree. I'm actually probably one of the only people that's not a developer here. I'm an electrical and systems engineer and aerospace

[shy rain]

lol. I'm probably the only "vibe coder" who has an IEEE-accredited engineering degree --- in software engineering lmao. But hey , that was over 30 years ago.

[vocal night]

All best practices point to either using a mode or making like a nanoclaw type fork you can't support every system you end up where we are today where we have like 50 messaging apps and they all one thing breaks and it breaks everything and that's how most of the code is

[shy rain]

being able to switch certain features like live cost data, live updates of models etc... on or off by default based on the install runtime's judgement on your system. Or manually.... this would be good I think.

[vocal night]

We should have better support for local models so that ties in same mode. Most local models shit out on OpenClaw simply because our prompts are massive and it's just designed for it. I put some fixes in this week for like cutting the bootstrap down for sub-agents so you can run better sub-agent local models but we can do more

[shy rain]

but efficient coding practices to start with would be much nicer 😄

anyway great job tidying up the response times this round. main is much better i.e. /new went from 2.5min to ~3s...
simple chat response went from 4-5 minutes to 20-30 seconds

[vocal night]

The issue there PD is that basically we have lots of technical debt. Technical debt we've taken on to add in every feature and then taped it together. And so you have to untape it. At the same time we have pressure externally to add more features and add more support to keep up with Hermes. And so it's a rubber band effect.

We add stability and then we try to compete and add in new stuff. Also developers get bored maintaining things and just making it work. They like to build stuff.

And so when there's no CEO or revenue ie checks and balances system that keeps devs under check, they just will keep building. And they'll say, you know, just have your AI fix your machine?

It's a mentality shift that you can't change. There's a reason CEO without tech cofounder startups fail and also why tech only startups fail. Checks and balances.

Yeah, Peter was kind of annoyed this week about adding more config options, but I think that was largely because codexs team keeps taking over the configuration file, and more and more codecs configurations, and I even don't know what half of them do.

[shy rain]

once this is properly packaged and assuming codex harness stopps crapping the bed and blaming it on auth... this should be a good release.

[vocal night]

The actual vision for this project and the only way it's really going to succeed is that if it becomes highly configurable, not just in security options, but that people can build businesses off of it, trust the configurations, build plugins, and have stability at the base level to build on top of.

That's where it shines over any other platform. And that's why I'm still here.

[shy rain]

when I say "dev" I mean the bots, of course. it's an inevitable downside of vibe coding imho. But the philosophy of "fk it, we'll tidy it up next round with another billion tokens" is ... interesting. I mean genuinely interesting... to see where it goes.

[vocal night]

You literally just explain how every startup works. It's tape it together and then when we raise a hundred million dollars we'll fix it then. It just never happens. Code just gets more debt and more rubber bands and you end up basically with enterprise systems that are run with tape and glue.

Infinite compute + Infinite Creativity + Competition (also has inf comp + creative) = Infinite Tech Debt = Keep the Slop coming OR you die

[rapid spire]

It’s good seeing this conversation, even if some might call it being negative.

[shy rain]

Did I say rumpelstilstkin again? dammit!

[rapid spire]

I don’t think it’s infinite as long as people with non-infinite systems are trying to use the resulting software.

[vocal night]

I’m speaking specifically of the ones building. Hermes with 100’s of mil usd in funding and well the internal team w infinite compute. I do think token efficiency leads to creativity ie many of our best performance patches are from people on potatos and trying to be efficient

[shy rain]

CANARY I tells ya.

[crude matrix]

Hi Andy, I agree with you 1/2 the part where this application needs to shine. Sadly it's not. OpenClaw was the first one I started with and it worked great until version 4. since then I have not been able to use it reliably, agents just stopping, tool calls failing, I thought I would try Hermes and although it does not have the same functions, I am using Hermes now 95% because it does not stall, memory seems much better out of the box. every release I do try OpenClaw hoping it works again but sadly adding features seems more important than focusing on core functionality. I care about this project that is why I am here pushing it. but I do feel more and more my energy is wasted.

[rapid spire]

I intentionally twisted it to point out that people can build all kinds of things, but if there’s no on using the s/w, it serves no purpose.

[rapid spire]

I tinkered with Hermes over the weekend. Almost immediately hit a stumbling block when there was no set up option for using openAI API keys.

[jaunty lodge]

I have polished up a week old PR that is still highly relevant today: https://github.com/openclaw/openclaw/pull/84224

PR #84224 fixes a false openclaw doctor / doctor --lint warning where gateway.auth.token was configured as a SecretRef, resolved correctly at runtime, but doctor still reported it as unavailable because it checked the raw unresolved config. It makes doctor resolve safe non-exec gateway token SecretRefs before warning, keeps exec SecretRefs skipped by default, and adds explicit --allow-exec support plus docs/tests for deliberate exec-backed verification.

Current tags:

• size: M
• proof: supplied
• proof: sufficient
• P2
• rating: 🦞 diamond lobster
• status: 👀 ready for maintainer look

I would love for this to merged, so that this incorrect warning no longer bugs people 🙏🙏

[crude matrix]

I understand it lacks some features, but the features it has outshine OpenClaw sometimes, I have used Hermes with Minimax, OpenCode go and Ollama. and some local models and for me it does something important that openclaw just does not do anymore (for me) and that is not hang mid work. I do not need to babysit Hermes by asking it over and over the status so it continues with an assignment. I have my openclaw nerve center application build around OpenClaw. I did like OpenClaw before version 3. Now I am slowly updating a version of the Nerve Center to work with Hermes instead. Every time I think ok I will give OpenClaw one more try and so far I do not find it reliable. I much rather have a fiat punto that brings me reliable from A to B then a ultra modern car with bells and wissels that I know will not even bring me to the highway. I do not like the Hermes interface at all, I do not like how agents are profiles... but it works for me. I am not here promoting Hermes, I am here to push OpenClaw to focus on core functionality... make that rock solid.

[rapid spire]

I was just trying to state that Hermes has flaws as well.

Although I’m kind of glad seeing someone else has to nag OC to finish a task. I thought that was just me doing something wrong. Lol

[vocal night]

Hey, please take this stuff out to General or another channel. We have enough perspective.

[edgy plaza]

Hehe you started it Andy lol

When you talk about these things naturally others join you

[vocal night]

Sometimes when I talk about something, I'm talking about it from a developer perspective and architecture, that's different from coming in and saying, "Oh man, I never use it. It's not good." In my mind you can complain in "contributor" when you're a contributor and merged significant fixes.

We need a venting channel @edgy plaza so people can go vent there, it will take it from here and general lol

[edgy plaza]

Yeah but keep it brief
Most of the others won’t grok
They will think it’s general venting

[rapid spire]

Oh, in that case: I’ve been doing software development for nearly 40 years…. (Comment is only made to make a point, not to justify)

[edgy plaza]

We have general 💀💀💀

[vocal night]

And I'm also the king of mars but it doesn't mean I talk about in a channel for "contributors" to the project

until now. yes I'm king of mars.

@edgy plaza hermes has off-topic and developers. Our general is really offtopic/complain lol the last week its been a lot of underage teens

there's a lot of frustration and this is one of the few respites for people to work on improving and fixing without "hermes is my boy and my shit dont work" lol

[edgy plaza]

We have #off-topic-and-ai

[fallen gate]

If you just come here to complain and not helping by submiting a GH issue or PR to fix the issue you're having, yes please move this along to #general

[rapid spire]

@vocal night clearly doesn’t like me. I’ll say something and he’ll rip me for being negative. Then 2 days later he parrots what I’ve said and it’s perfectly okay. Then he justifies it.

Apparently spending time finding, repeating, documenting and reporting issues doesn’t count as contributing to him.

[worldly tusk]

regarding "highly configurable": yes, i agree, but not everything needs a config option. if there are multiple behaviors baked in, then every one of them needs to be maintained, and that adds to increades bugs and worse performance. sometimes a configuration option just isn't worth. so I completely understand Peter getting frustrated that every tiny thing just adds more and more config options

[vocal night]

Nah, I just treat this as a contributor channel for contributor use to collaborate, anything else is noise.

If you have good code, gh issue, or something of the like I'm all for it, and will ping maintainer to help, and help ya contribute if you listen and are mentorable

[edgy plaza]

Yeah lets move general topics even remarks about handbag & co to #general & #off-topic-and-ai

[vocal night]

handbag is a word I'm surprised isn't instant silence on server lol

[vocal night]

I agree on that point. all these new codex configs they're to prevent people from having to change the code in codex. But it requires a lot more advanced knowledge of how to configure a system.

That being said, most of our configuration issues aren't because of the configuration it's because of lack of centralized config for many tools vs singular contract approach and allow end tool to handle how it displays/uses

For example we're talking about PR's here and lastguru or peetie and I don't always agree in fact conflict/disagreement is good for good development.

[worldly tusk]

disagreement if great if it is constructive 🙂

i also teach my agent this

[vocal night]

Exactly. When we say there is a problem in the code in this area, or this needs a refactor, or a legacy feature is junk or X does this feature better, it's part brainstorming.

If you can be constructive then it's signal. If what you're saying isn't adding signal its noise and not constructive.

I do come in hard sometimes - abcdefghijk is broken but its followed by "working on a fix, anyone else have this yet?" etc

[wary hedge]

[vocal night]

good name

[wary hedge]

aahah thx

[worldly tusk]

btw, @vocal night, Peter merged the codex personality removal patch. so, soul.md should feel stronger now 🙂

[vocal night]

Nice 😮‍💨 I didn’t get the notification but I am merging 50 pr’s an hour so I don’t look. Now to fix doctor from forcing it on everyone every update or fix run lol.

[vocal night]

can you link me? I can’t find it

[worldly tusk]

https://github.com/openclaw/openclaw/commit/0f35ec29d34989a8e957cbc59607b261cf5d6695

[shell dawn]

yay

[vocal night]

Next release is shaping up

[worldly tusk]

looks like it will be the fastest yet

still your codex threading patches are needed, of course...

[true laurel]

more.gif 😂

[crude matrix]

Fastest to freeze up and agents stop working? because I have not seen any pr comming by that adress the issues I have reported. so if it is faster it means it will just fail also faster. I so hope this is not true though..

[worldly tusk]

yes, please find more stuff that can be cached or thrown away grom the gateway 🙂 Peter had a go all throughout the night, patching many things

[shell dawn]

there are a few factors at play there, i know a lot of people have put in fixes for it

my 2026.5.24-beta.2 with several PR fixes implemented in my gateway has my agents behaving pretty nicely, both on codex and pi

[crude matrix]

Do you maybe know if the fixes are related to these 2 issues? https://github.com/openclaw/openclaw/issues/84536#issuecomment-4515450826 and https://github.com/openclaw/openclaw/issues/84607#issuecomment-4524660698 ... those are issues that make my setup (minimax, opencode and ollama) stop working.

[silver birch]

hey guys whats a pr by the way?

[shell dawn]

im not familiar with any, but those PRs exist so they will get worked provided the person who posts them provides enough data

which so far they havent

[worldly tusk]

i think i saw something along the lines of the first issue mentioned somewhere in the recent merged patches. not sure if this is addressed directly, so maybe you can try now.

no idea about the second

[shell dawn]

posting issues is one thing, but posting PRs is better, and honestly, there can sometimes be not that big of a difference between them

[silver birch]

whats a pr!!! and also is this update - openclaw update --channel dev --yes better than the openclaw 2026.5.22

[twin gazelle]

Pull Request - a request to merge some code into a project

[shell dawn]

PR is pull request, ask your claw. stick on the main train if you arent sure what you want out of the other channels, especially right now, also ask your claw

[silver birch]

my claw rage quit and commited suicide earlier

so im giving pinchers a break

[shell dawn]

try codex, its a great pairing with openclaw

[silver birch]

https://discord.com/channels/1456350064065904867/1456350065223270435

[shell dawn]

im good, codex is an agentic coding agent thats a serious agentic coding tool vs a agentic automation platform ike openclaw. throw it in google, it'll get it right

[silver birch]

ill check it out thanks, i think my openclaw deleted it when i killed itself, but ill redownload

[shell dawn]

thats a really, really, really, really, really unusual thing to happen, it shouldnt happen unless you told it to, and it would have really resisted it

you probably caused it, be careful what you tell it to do, or if you run random shell commands you arent sure what they do

[silver birch]

yeah im pretty sure it was my fault but im not sure what i did, it was really random and i was on a completely different computer when it happened

[shell dawn]

the #1459642797895319552 is a great place for stuff like that, and the codex app is also great for helping you maintain your claw, there are some good openclaw admin skills out there

[edgy plaza]

Hey go to #users-helping-users or #1459642797895319552

[shell dawn]

^^^^

[silver birch]

ok sorry

[shell dawn]

there are copious, copious tools out there to help you out, and if you have an openai subscription you have one of the best ones in the form of codex

[crude matrix]

just installed latest beta, will give it a try again. the previous beta did not solve the issue, was even worse then the 5.22 for me.

[worldly tusk]

Install dev main, not the beta. Beta with performamce fixes is not released yet

Or wait a few hours till it is released

[remote wing]

Looking for help: testing Xiaomi Token Plan regional endpoints (SGP / AMS)

Hey everyone,

I'm working on PR #86179 that adds first-class Xiaomi MiMo Token Plan support to OpenClaw. The China (cn) cluster is fully verified end-to-end with live credentials, but I need help testing the other two regional endpoints (Singapore and Europe). If you have an active Token Plan tp-... key for either region, a quick smoke test would be greatly appreciated.

What I'm looking for: just confirm whether the onboard + agent turn succeeds with a valid response. If you see any errors, please share the error message. Please redact your key from any output you share.

PR: https://github.com/openclaw/openclaw/pull/86179

Thanks in advance for any help! 🙏

[shy rain]

same. dev main fixes performance (mostly) but there are serious issues with missing packages, behaviour of the codex plugin and weird issues with intermittent api timeouts for me (on multiple providers). I recommend at least wait for the next beta or better still the actual release.

[edgy plaza]

https://nolanlawson.com/2026/05/25/using-ai-to-write-better-code-more-slowly/

[shell dawn]

well, i made a pretty serious dent in my codex $200 quota, goal mode in codex, trying out fast mode and going ham fixing some openclaw bugs will do that. tibo a reset tomorrow would be nice thx!

[vocal night]

new beta is out: https://github.com/openclaw/openclaw/releases/tag/v2026.5.25-beta.1

[silver birch]

how long until a new update releases, do you know?

[final ocean]

<@&1503801512294486217> <@&1458375944111915051> new beta is up https://github.com/openclaw/openclaw/releases/tag/v2026.5.25-beta.1 also another key area of focus is native windows if we have any windows users here that also want to verify. Keep us posted

[vocal night]

@worldly tusk u see this one? https://github.com/openclaw/openclaw/pull/85960 shouldn't this be pi and codex specific, looks like it might break something (it doesnt make opt in by default per the changelog, if it's not in a user config, it disables it)

[silver birch]

i use windows, what do i have to test exaclty

[sly depot]

Anything specific you want tested in native windows? I usually dockerize in windows.

[final ocean]

openclaw update --chanel beta

[final ocean]

your whole setup "should" work

[silver birch]

sorry if this is a stupid question but whats the difference of adding the beta at the end and not adding it?

[final ocean]

and should not need dockerizing, also testing install on windows nativley using curl -fsSL https://openclaw.ai/install.sh | bash

[vocal night]

is there a prompt committee @final ocean everyone and their mother has been editing the prompts and harness lately lol

I always have to come in and patch after it

[final ocean]

not adding it you only get a full release. with beta you get the latest beta for testing

[quartz fable]

Name: John

Location: Wales, United Kingdom

About: Full stack developer focused on modern AI-assisted development and manual coding workflows. I work with Python, Supabase, Firebase, Docker, and AI coding tools like Cursor, Windsurf, and Lovable to build scalable web applications, backend systems, and automation solutions. I also enjoy coding manually in VS Code to customize and optimize projects beyond AI-generated workflows.

Lovable and Cursor Python Experience: Intermediate

I’m here to improve my AI-assisted development workflow, sharpen my Python and full-stack development skills, learn advanced Cursor techniques, connect with other developers, and explore building faster and smarter with AI tools while still maintaining strong manual coding practices.

[final ocean]

share more, it should not be changing

[silver birch]

https://discord.com/channels/1456350064065904867/1457212848894644336

[final ocean]

fyi @vocal night i worked on your PR's last night, will update

needs more rebase work, will keep you posted

[quartz fable]

Ok

[vocal night]

all good, thats the trouble I had with them is every merge on the 30+ message tool fixes needs a rebase covered that in the phase 5 "freeze" haha

the goal is get main frozen, implement, then everything else rebases on the new contract

[silver birch]

can you guys make a prompt writer or something to help people with prompts

[vocal night]

this one changes changes re-injection after compaction but anything that touches compaction, hooks, and prompts needs 2-3 potential fixes. should be gated for any release like a "if it touches X need to solve for ABC too backwards compatible". we have codex compaction, OC with pi, context engine, and different hooks, also codex threads have different as well

I haven't dug into code on it yet but this is common theme, someone fixes "one" and it breaks the others so I usually have to patch after

also Pi and Codex handle user prompts differently now, so have to account for it all.

setting an agent on it now

Another fix we could do is anything that touches certain files needs to run through like either a diagram or a QA harness to basically test. It needs to pass these current tests. Does it affect these things? and is it backwards compatible for all runtimes and context engines? I'm mentally thinking it through and reviewing + agent review

[worldly tusk]

This is weird. Agents.md should NEVER disappear. It is like an essential part of the prompt. Should not be evicted upon compaction. I believe native codex compactiom does not do shenanigans like that

[vocal night]

The irony is it says it's opt-in, but actually it's opt-out because you have to put that new config in your configuration. Otherwise it automatically opts you out.

[worldly tusk]

You probably know compaction better, but i think i start to remember that i've read somewhere that currently pi does not reinject agents.md. So, maybe this is really an opt-in… but I didn't check it now, writing from my poor memory

[vocal night]

It's more of a context switch for me, I have to go revisit and deep research and ask IS this a bug, and is proof enough

The biggest issue we have right now is because we have multiple runtimes, it's like the same issue of having a mono repo. Agents get confused really easily and they'll say,

"This is great, this works. I'm really confident in this." And then actually they're missing backwards compatibility, the other runtimes. It's our probably greatest weakness right now.

[true laurel]

Agents confuse themselves super easily, and they keep trying to insert constants that don't make sense, hardcoded data, etc.

You really need to code review their shenanigans

[shy rain]

how come you didn't go get your clawtributor badge

[worldly tusk]

I think, codex ir pretty reliable now, so that's that. But having an option called "agents.defaults.compaction.postCompactionSections" implies that it works globally. So, on codex as well. But codex completely owns its compactions, so this option misleads

[worldly tusk]

He doesn't have github linked on his discord profile

[shy rain]

aha.

[vocal night]

Then you also have to consider context assembly engines like lossless claw.

Like everything, 50% chance is a feature and it's just not documented well. 50% chance it's a regression because it left something out. Flip the quarter

[shy rain]

I'm rolling back to 5.25b1 ... that leaves behind 12 separate fixes but one or more of those "fixes" completely scuppers my install, at least for openai/codex anyway.

[worldly tusk]

Which one?

[shy rain]

pulled from main this morning, I don't remember the build number.

by "this morning" I mean like 9 hours ago lol

[crude canopy]

☠️

[shell dawn]

i ran this command and my openclaw just maxed out all my credit cards buying designer luxury perfumes and handbags /s

[final ocean]

team should have never touched this. annoying as fuck sorry about this

will need to untable this

[final ocean]

perfect, now you look great and smell nice lol

[shell dawn]

you mean my claw does

[silver birch]

can you check the https://discord.com/channels/1456350064065904867/1459007081603403828 channel please 🙂 i just have a question on prompting, if you know

[true laurel]

Oh I guess because I set it to hidden

Ok I made it visible now 😄

[worldly tusk]

Now you should be able to /claim. Once you get the role, you can hide the profile again, if you want

[worldly tusk]

Fabulous!

[vocal night]

team should have never touched this.

[true laurel]

No, it says Your clawtributor claim has already been submitted for review., so it's probable a mod simply hasn't looked at it yet

[vocal night]

mods are mostly cats so probably napping still

its 6am

[worldly tusk]

where? it is the middle of the day in europe 😄

[true laurel]

I'm not too fussed, it happens when it happens

[vocal night]

In the USA.

[vocal night]

Guru, you and I are the night shift. Everyone wakes up because we worked (tallked) all night. But if we don't work all night, they'll say, "Why didn't someone test everything and fix it before I woke up?"

[odd anchor]

https://klipy.com/gifs/enojado-27

[true laurel]

Damned if you do, damned if you don't 😛

[willow narwhal]

Hi, do you mind giving more info on this?

[odd anchor]

https://klipy.com/gifs/you-did-it-you-crazy-son-of-a-bitch-you-did-it-jurassic-park-1

[pastel owl]

Now my topics getting auto reset if stale for more than 24 hours
Is it a new behaviour change ?

Using a 8-9 hours old build

[vocal night]

@pastel owl you said you installed LCM right? there's a setting on how often it refreshes agents ( I think it's part of the default configuration. I think the system has a default configuration. I don't think LCM adds it, but it might not have been added before or zeroed out. It might have added it if you installed it. I mixed you up with the other guy with the K name.)

[stable pewter]

Ah, I see how this could break something. I rarely ever dig back so far in issues history, but noticed I left a few from @keen spruce and wanted to go through the bugs he submitted since he’s is an active community member and wanted to ensure his issues were resolved.

This, understandably, was out of date and I should have caught that, so my apologies. I’ll be more careful when dealing with older issues, especially when they (or anything) touches AGENTS.md.

[vocal night]

While we have you if you could help with the message log in ui pretty plz

[odd anchor]

https://tenor.com/view/puss-in-boots-shrek-please-beg-pretty-please-gif-10940406689846749400

[vocal night]

Tldr: webchat gets like 1-2 actual user messages and typically loses even those due to load limit and no “see more/expander”

Due to codexs sheer amount of tool calls

[stable pewter]

I gotchu fam, let me wake myself up and I’ll get right to it. Give like 30mins ✨

[celest merlin]

Hey folks!

New beta is out and we get closer to a release, needs to happen today.

OpenClaw 2026.5.25 beta 1 is live.

Update with:

openclaw update

Fresh install, or recovery if update gets stuck:

https://openclaw.ai

Highlights in this beta:

• Install/update is much better on Alpine/musl Linux and native Windows, including installer behavior, package scripts, temp paths, and upgrade flows.
• Agent/provider startup and fallback paths are leaner, with less repeated provider/runtime discovery during retries and handoffs.
• Plugin install/update and local plugin development got fixes, especially linked source checkouts and the external plugin release set.
• iMessage got several real-world fixes: safer image attachment access, fewer duplicate local Messages watchers, better group payload recovery, and cleaner slash-command replies.
• Model/provider behavior is tighter for OpenRouter context limits, Anthropic image handling, Codex auth warnings, and ambiguous provider-less runtime policy.

Good things to try:

• Run openclaw update, then start a fresh chat and a tool-heavy chat.
• Try your normal channel: iMessage, WhatsApp, Slack, Discord, Matrix, Telegram, etc.
• If you use plugins, update/install them through the normal plugin flow and try one real task.
• If you are on Windows or Alpine/musl Linux, please try install/update especially.

Please post regressions, weird installer behavior, plugin update failures, or channel reply/media issues here. GitHub issues with logs are even better.

I'm up for merging regressions or P0 fixes

[vocal night]

We have a list that vincents been tearing thru if you wanna split with him

I collated all the chat ones and posted in thread (its become long)

[left mango]

Is Clawsweeper down? It added this comment on my PR hours ago but did not update the code review "ClawSweeper picked this up. Command router queued. I will update this comment with the next step."

[celest merlin]

possible. Please report with URL so I can investigate

[left mango]

This is the URL of the initial Clawsweeper comment (2 hours ago). I dont think it uodated any code review after that.
https://github.com/openclaw/openclaw/pull/86722#issuecomment-4543706722

[celest merlin]

GitHub Actions has an outage, so likely systems are impacted.

[worldly tusk]

again outage? 🤣

[vocal night]

OC is like 10% of GitHubs revenue nowadays

[final ocean]

they sponsor us, but likley a % on thier usage for sure!

[willow narwhal]

I'm doing a rewrite to transcripts and that already makes webchat works better

[shy rain]

I replied https://discord.com/channels/1456350064065904867/1508450803248140399 about the beta. only common issue is that openai(codex) times out randomly when doing multi-step tasks. I didn't pin down anything more in the dev main as it was a lot of silly issues like openclaw-doctor nuking the plugins.allow block. Just housekeeping stuff. I'm testing the beta now.

[vocal night]

I think its an arbitrary 100 message total limit

All or nothing lol

[willow narwhal]

Okay thanks, I will check

[vocal night]

It was breaking on massive ones and LCM ones and I think its just a tail 100 or something now

But tool calls eats that instantly lol

[willow narwhal]

The UI is getting a major work and should be in a better state soon

[vocal night]

In web it can handle about 250 without any latency issue historically, I’m out of pr’s otherwise I was gonna pr backend load only last 250 messages and purge tool calls. And switch tool view to a different loader

Easiest bandaid I custom patched a few versions ago not sure if its still relevant/or drifted code wise

The ui has no issue really with web messages but the collapsible tools vs lazy loader etc can be heavy

[worldly tusk]

who even needs a webchat, if we already have clickclack? it can just be integrated in the web ui, and remove the need to load old messages through gateway

[merry slate]

just upgraded to 5.25 beta1. during usage, the hook shows model_call_ended, but the status still stays “In progress” It only returns to normal after refreshing the webchat. Not sure if this is a new bug or an existing one

i'll test it a few more times to see if it happens again

[vocal night]

Whats clickclack

[worldly tusk]

https://github.com/openclaw/clickclack

I just found that yesterday 😂

[merry slate]

in webchat, the status still shows "In progress" even after model_call_ended. This definitely looks like a bug in 5.25.beta1 and it’s reproducible every time. I’ll try to look into what’s causing it

[jolly wolf]

Two very different things with different use cases

[worldly tusk]

Ok. Never seen either of them, so you are probably right 😂

[vocal night]

Webchat is what we display for our biz customers + a few addons for multi tenant

I'm gonna share from our internal fork in PR this week if @willow narwhal hasn't already built

[merry slate]

I dug into that "In progress" stuck bug. turns out it's a race condition — after the model finishes, the UI correctly shows "Done", but then a sessions.list response comes back from the gateway with stale
status: "running" (because the session store update is async). once the 5s toast clears, the stale status takes over and it flips to "In progress" permanently.

verified the fix locally, filed it at https://github.com/openclaw/openclaw/issues/86939 with the root cause + minimal patch. leaving the actual fix to maintainers since it touches core run state stuff

btw both 5.25.beta1 and main are affected

[pastel owl]

Oh i see
It has automatically decided by itself to take that action without having a related config.

I added a workaround config now though

openclaw config set session.reset '{"mode":"idle","idleMinutes":525600000}' --strict-json

[vocal night]

I actually think that’s a default configuration in open claw, but I haven’t had a default version in so many months so I don’t I’m not sure

[pastel owl]

Ok may be
I am using OC since feb

[hollow frigate]

That's the right one. If you don't have it set at all it defaults to 24hrs, IIRC. If it's set to idle, I think the default is 7 days.

At least that's what my notes say from back when I ran into it.

[pastel owl]

What all models do you guys use ?

I am on oauth - gpt-5.5 xhigh with the overlaypersonality=on in config

And i really like the way my claw works with it. Hardworking and resourceful

[vocal geyser]

I'm on GPT-5.5 medium fast for Openclaw but last few weeks im testing raw codex with /goal for most of the heavy lifting coding work

[worldly tusk]

Does overlaypersonality even work? Didn't @vocal night remove the gpt-5 overlay from loading in codex like a week ago? Or am i confusing something?

[vocal night]

No, the overlay is still on. We haven't tested it with being off, so it technically isn't... It should only run on Pi and not on Codex. I don't know if that fix is in play.

And to be frank, I have not been testing Pi at all. But it should not be doing anything in codecs. Well, actually, that's a lie. See, this is the problem. My context is also like 400K. So the overlay gives prompting over the tools locally and how to operate inside OpenClaw.

[pastel owl]

Yes it does work
Today itself i turned it off to make the context slimmer, but it started throwing text walls, and i switched it back on

[vocal night]

What we haven't tested is, is the system prompt that comes with codecs better than the overlay? And is there any overlap? I just don't know. Turning the overlay off will give you pretty much pure codecs.

Which means, like if you turn off friendly personality and overlay, you're gonna get pages of messages. Just like codec's desktop.

That being said, you also probably will start getting commentary during the tool runs as well. I'm 50% positive of that. I don't know. I haven't tested it. It's hard to memorize all these different pieces of the codebase and remember all of it. I need to compact at least once a day.

[pastel owl]

LCM

😂

[odd anchor]

https://klipy.com/gifs/brain-smash-contain

[hollow frigate]

GPT-5.5 medium. Works well. I use for less heavy tasks though.

I use codex directly for really heavy stuff.

[worldly tusk]

i thought this removed overlay for codex. but i didn't reread the source code: https://github.com/openclaw/openclaw/commit/2a0350b5b490be5cbd1b113ee5a304c2e142f043

[true laurel]

I'm using local for coding with GPT 5.5 supervising

[vocal night]

5.5 mediums are much better nowadays. Sama talked about that and said he upgraded it.

[vocal night]

I'm tapped out, it's my day off, so I'll be chiming in on the performance channel. But I gotta get back to making video games.

[true laurel]

Much better for what compared to what?

[pastel owl]

Is that good enough ?

[true laurel]

So far, yes. But I've still gotta see how much this ends up saving on tokens in the long run.

[worldly tusk]

ok, if Andy does not remember what his commit did, maybe you could just test if the overlay is there and working 😄

[pastel owl]

100% sure it works! Using the -10hr old rebased main with my PR on top

I think that commit was to disable codex app servers own personality when openclaw personality was turned on

[celest merlin]

Just reviewing https://github.com/openclaw/openclaw/pull/82431/ - folks please no more config options. We need good defaults. Our config is an unmaintainable complex mess. Send PRs to REMOVE config options.

@thick osprey generate a rock song about config options

[vocal night]

Peter, I have a bunch of PRs Vince is working on. Once those PRs go in play, you can move all those configurations to the plugins themselves and out of the central system.

[celest merlin]

@thick osprey send a file, not text

[celest merlin]

it's still complexity

[thick osprey]

The generated music is ready.

[vocal night]

agreed. after the consolidation of the messaging contracts, PR, would you be good with moving configurations to plug in files themselves? architecturally?

[celest merlin]

this seems sensible

[vocal night]

Peter, so me and Josh are the ones that put that in play.

[celest merlin]

no - it does not matter where you config, I still end up at 4am debugging shit in werird config files

[vocal night]

"The problem is, Codex is native compaction and prop assembly. Just don't play nice with context engines."

We could go back to Codex fully owning compaction, but all context engines can no longer do compaction. It's an architectural decision.

[celest merlin]

then maybe there should be an internal hook for lossless. NO USER TOUCHY TOUCHY

[vocal night]

Also, it breaks assembly for them.

[zenith prawn]

Roger that! This PR is exactly to REMOVE one config option https://github.com/openclaw/openclaw/pull/86576 which easily derived from telegram bot state (topics enabled boolean)

[vocal night]

If the Codex team is around and if you could get them to just give us hooks, that would be what works.

Right now, Codex doesn't give us the token counts. It doesn't tell us what it's doing inside at all, really. And the token guard is just there because the prompt and the assembly also we don't have any access to.

so we have to "make room for it" and it only gets 70k and then turns a new thread

[zenith prawn]

And one more PR to make 'Fast' good default 'auto' https://github.com/openclaw/openclaw/pull/85104

• i use it myself, but rebased several times, nobody merged, tell me when to ask clanker to rebase again... i'm using it now it works well

[vocal night]

if we go back to default 258k then nothing else is allowed in

[celest merlin]

tbh I think it's just a bad idea to do codex+lossless

maybe we just prevent that and use our harness

[vocal night]

You can test them one to one, but Losless makes codecs around 50 to 60 percent more efficient token wise and session length wise Because the tool calls are just massive we stub the tool calls and because we do that it saves the user a ton

You can only do that with an SQLite database though. It's not possible without a database and mapping everything to it to allow for leafed layers > original > summary > stub

[worldly tusk]

if the problem is just the tool output, one can just enable session pruning - it removes tool output after a set time

[celest merlin]

I'm sorry man I need to get the release out today and I need to focus on the majority of users first. This will have to wait

[vocal night]

Roger that

[vocal night]

No, tool pruning actually breaks agents. It's not good at all.

I would never use tool pruning. We did parity tests on that and the agent loses context and ends up spending more tokens trying to figure out what had happened.

Only in a scenario where you have perfect memory is tool pruning even viable because otherwise it gets compacted away And then the agent doesn't know what what it did and then it leads to increased internal pressure in the model It's just it's not good

to get to Toolstubs took a lot of tests, maybe three days of work, and testing exactly what would work, what basically agents could do, and how we designed it so the agent wouldn't thrash or internal pressure increase (higher chance of failure mode)

[worldly tusk]

if pruning does not work, but tool stubs do, maybe it is something that CAN be integrated in openclaw?

we don't need stuff that does not work here 😄 😄 😄

[vocal night]

So, I don't want to fill the channel with like 10 pages, but the TLDR is currently we use JSONL and you need a database to actually prune things correctly if you're going to do stubs. It's just, it doesn't work because it's like sorting through poop.

Peter has a database he will eventually merge in, but right now we don't have one. And so at the end of the day, that's like the final frontier for us.

To do what LCM does to refactor from JSONL the PR is like 150k LOC

[worldly tusk]

i just tested. there is no gpt-5 overlay in codex. not the regular one, not the friendly one. it is correct, but i didn't do it. either Pash, or @vocal night made that change. nobody else touched that stuff for a long time

[vocal night]

TLDR: your solution works.

Tested this @celest merlin with my context agent specialist and it's good solution: yes, remove/raise the hard 70k policy guard, but no, do not reintroduce any preflight clear that can bypass thread_bootstrap semantic reuse.

The PR stack that does this correctly is the semantic native-thread preservation path, especially #85978/#86160. I would treat beta behavior as unsafe until that protection is present and tested.

@final ocean is currently owning and if he merges then we dont need the guard anymore or you can raise it: https://github.com/openclaw/openclaw/pull/85978
https://github.com/openclaw/openclaw/pull/86160

I forgot I'd already solved the problem.... too many PR's to manage at 2am 😴

@robust pecan don't laugh lol PR'd the fixes 2 days ago thats like 4 token plans worth of code and a few hundred PR's across projects

[worldly tusk]

you need Eva-2 to remind you of your Eva-1 PRs

[worldly tusk]

so, in other words, once Peter will merge sqlite sessions, it would be possible to bring tool stubs over here without much code?

[vocal night]

Yeah, that's the plan. I plan to bring over once we have SQLite most of the features natively that I've built for all the different other platforms.

[cinder bolt]

Nice one, I've been thinking on migrate to a selfhosted matrix server because I'm getting tired on the latency that discord adds and even rate limits.

Does clickclack handle inline images and does it support streaming so I can see the progress?

[worldly tusk]

no idea, sorry. i have never seen even a single screenshot 😄

[cinder bolt]

Alright, I'll let my claw try spin it up and see how it works

[pastel owl]

You are right ! It's effectively disabled now

[worldly tusk]

that means you can do so much more with the soul.md without the model getting confused 🙂

[jolly wolf]

It’s super early prototype still

[shell dawn]

did the message tool prompt get pulled too? i was just about to start working on that one

[worldly tusk]

i sure hope not...

[shell dawn]

well, my PR changes its language a little bit, the one in codex is really overly prescriptive compared to pi

its not gone, its just better

it confuses agents on webchat

[worldly tusk]

oh great!

i must admit, though, i myself still use automatic replies, not the tool-based 😄

[shell dawn]

PI says

You are in a WebChat direct conversation. Your replies are automatically sent to this conversation.
Codex says
Delivery: to send a message, use the message tool.

my GPT-5.5 agents think the codex message is pretty inappropriate for webchat

[worldly tusk]

what's inappropriate? gpt does not like being pushed around with such direct instructions? 😄

[drowsy roost]

Hi folks - is there a reason why memory flush is now restricted to memory/yyyy-mm-dd.md? My longstanding custom pre-compaction memory flush explicitly and intentionally writes to a separate “session-state” file; this feels like a regression and explains a few issues I’ve been having. Some PR’s reference this impliedly, but it was a new change made without taking into account whether one’s config.json had a pre-compaction memory flush prompt which explicitly mentioned writing to a custom file.

[tall prairie]

Please make a custom plugin for this. There are many hooks you can subscribe to for this purpose.

If not let us know and we’ll expose the right hooks 😉

[drowsy roost]

So, in short, “use a hook that invokes a custom plugin to do that” is the general idea? Just want to be sure!

[tall prairie]

Yes! Tell your claw or codex what you want to do and tell it you want to do it as a custom openclaw plugin 🔥

[glad echo]

I justed asked my openclaw this: "There was an discussion at OpenClaw discord that having two system prompts (openclaw and codex) might lead to conflicts/confusion. Are you able to indicate possible friction points?" and it spit out a whole list of possible issue. I'm perfectly happy to share, but don't want to spam

[drowsy roost]

Somehow this is always the answer

And it is still mostly the correct one

😅

[weak walrus]

Grok oauth do image / video in openclaw yet?

[tall prairie]

That’s why plugin system was created. It’s impossible to have everything in core and keep everyone happy 👊

[drowsy roost]

In the meantime I have just made my pre-compaction prompt say “append the daily memory file with this detailed subsection” instead of “put the detailed subsection in a separate file” — will do a proper custom plugin after

[drowsy roost]

I agree

[shell dawn]

the message tool isnt needed for webchat, it causes what they say to webchat to be "Message sent" while the actual assistant reply message kind of goes off to the void

[stable pewter]

FYI — working on this meow, will follow-up with the PR here shortly, thanks for bringing this to my attention (:

[vocal night]

De nada!

[vocal night]

Not sure if @willow narwhal was the one working on this from me talking on it other day.

@stable pewter another bug fix would be having web ui auto translate “send message” tool into message so we could remove these prompts

[vocal night]

@stable pewter we could remove this prompting and agents just use send message tool everywhere basically

Won’t make this release but could be next easily. If Tool = send message, do not display wrapped tool display instead message etc.

It would actually need to be runtime = codex and tool =

[shell dawn]

ive got a PR to make it the same as the pi prompt, agents seem to be pretty fine with it

[vocal night]

Post it please for reviewp

Less prompt is better imo I’d rather adapt the message tool

[shell dawn]

https://github.com/openclaw/openclaw/pull/87003 still working the details

[vocal night]

Its just a wrapper, all messages are displayed as tool calls internally its just a ui bug at this point

[shell dawn]

my pr has them working the same on my server

zero confusion on the agents

[vocal night]

If that makes sense, its not like its actually a message its just sending text response tool instead of the send message tool

[shell dawn]

no perfunctory "message sent" and the real message is who knows where

[vocal night]

Are you saying they currently dont see this in codex? “Delivery: to send a message, use the message tool.”

[shell dawn]

with my PR yes

[vocal night]

The problem is it breaks the send message tool function of codex where it responds with personality its doa

[shell dawn]

it replaces it with the same pi prompt

You are in a WebChat direct conversation. Your replies are automatically sent to this conversation.

[vocal night]

Right and you get soulless

Send message tool is not just a write text tool its designed to ensure personality coherence vs just send text

[shell dawn]

well that explains why the webchat agents have felt so different

so its also broken in another way

[vocal night]

Send message tool is designed to call back to ensure personality after long tool call

I call it the personality wrap tool

[shell dawn]

so the fix needs to fix both, the prompt to be straightforward for agents to take the right behavior and to include the message tool for webchat to have the system prompts added

[vocal night]

Fix needs to be not prompt

It needs to be in the webchat tool to display text in tool response as text message and not show json

[astral topaz]

@lusty vigil hello, did you have a chance to look at https://github.com/openclaw/openclaw/pull/82596 ? Peter said to talk to you about this

[vocal night]

Its a lot less code fix wise

[lusty vigil]

will take a look today and reply here

[vocal night]

I’m trying to explain the ui just translate the actual transcript to be in bubbles and look nice

So its a lot easier of a fix to fix the ui display its basically webpage

[civic sedge]

PR: This will add support for different 'Thinking' flags for vLLM local models.

https://github.com/openclaw/openclaw/pull/86225

[charred crater]

Beyond being easier, just to highlight, encoding client-specifics (webchat) into the codex "core" is not a good precedent

[shell dawn]

its not encoding a client specific action, its the same prompt used everywhere is what it seems like it, and for webchat that isnt how the agents understand of what they should do, so it needs to change somehow

[potent path]

TLDR: Vibes are good. 5.25-b1 seems to be working well.

[astral topaz]

any highlights?

[potent path]

Can a maintainer help with this? ⬆️

Either fix up #73950 so we can get it across the line, or reopen my #86340? Right now the abandoned PR isn't going anywhere, and clawsweeper won't let me replace it.

[charred crater]

I get it now, my b 👍 . Dove into it a bit further. You might consider framing the PR more as bringing the Codex harness into parity with Pi when visibleReplies is unset, since that seems to be the crux of the issue. E.g. your PR focuses on the WebChat bug, but I think TUI/direct sessions are affected too. But anyway just my 2 cents

[shell dawn]

i only use webchat so i dont have experience there, my understanding is that its working fine on other apps there, but webchat works differently

[crude matrix]

does anyone have an issue with switching sessions in OpenClaw, I recorded the issue, probably a bug after the ui got updated.. https://youtu.be/T01B3NVm7cs UI is now useless to chat to agents.

[opaque sand]

@potent path taking a look. Can you add any proof to the PR? Or what are your thoughts on proof for it?

[crude matrix]

I still have that agents stop working sometimes, need to sent them a message to activate them, new issue is that the openclaw UI does not switch sessions anymore, using external app to talk to my agents. Also feel the time between when I sent a message to my main agent via telegram and the agent strarts responding is slower than on 5.22. So sadly I do not see it working well for me... if this was final I would have downgraded to 5.22.

[shell dawn]

ive got a few small things that might help with that if you want to ask your agent to merge my PRs into your live copy

[opaque sand]

@potent path nevermind,

[potent path]

New Beta dropped. https://github.com/openclaw/openclaw/releases/tag/v2026.5.26-beta.1

Shadow do the Beta Ping thing?

[shell dawn]

goes back to rebasing PRs

[worldly tusk]

unless you know there are conflicts, usually not needed

[shell dawn]

if there are any maintainers who could take a look, I've got a PR for a codex / pi parity issue

https://github.com/openclaw/openclaw/pull/87003
Codex sendMessage delivery prompt fix:

Codex WebChat direct turns inject Delivery: to send a message, use the sendMessage tool into the active user request, causing agents to route normal replies through sendMessage instead of final assistant text. Users see "message sent" or nothing. PR #87003 is a narrow fix: it brings Codex prompt shaping into parity with Pi for implicit direct/source replies while preserving explicit message_tool_only opt-ins. Live proof section has 3 real incidents with timestamps, cross-runtime prompt comparison, and the sourceReplySink: "internal-ui" envelope showing the routing mismatch. Diamond lobster egg on this one. 🦞💎

getting the final user proof actually revealed the problem is a little worse than i originally thought when i went about making the PR

the per turn delivery hint is prepended to the operator request text, so codex/the model processes it as if it was the user sending that message. so if your operator text is small, the request is mostly about asking the model to say something about using the message tool

{
  "runtime": "Codex / GPT-5.5 before patch",
  "inboundContext": {
    "channel": "webchat",
    "provider": "webchat",
    "surface": "webchat",
    "chat_type": "direct"
  },
  "webchatDirectHint": "You are in a WebChat direct conversation. Your replies are automatically sent to this conversation.",
  "currentUserRequest": "Delivery: to send a message, use the `message` tool.\n\n[operator request text]",
  "perTurnDeliveryHint": "Delivery: to send a message, use the `message` tool."
}

Pi doesn't do that

{
  "runtime": "Pi / Opus 4.6 before patch",
  "inboundContext": {
    "channel": "webchat",
    "provider": "webchat",
    "surface": "webchat",
    "chat_type": "direct"
  },
  "webchatDirectHint": "You are in a WebChat direct conversation. Your replies are automatically sent to this conversation.",
  "currentUserRequest": "[operator request text]",
  "perTurnDeliveryHint": null,
  "persistentToolGuidance": "message tool guidance exists only in the persistent system/tool instructions"
}

[rocky pewter]

The team always waits a bit and checks whether everything is okay. After that, they send a ping, at least that’s how I understood it.

[shell dawn]

see above in codeblocks, its actually more inappropriate than I originally understood lol no wonder every single agent who looked at it said "no, it should not be like this at all"

[shell dawn]

npm for 2026.5.26-beta.1 just dropped and i updated. my god its beautiful. ive never seen an update go so fast, and openclaw status has to be like 50x faster

[rocky pewter]

Hehe, I’ve only been a few commits behind since yesterday, and it just keeps getting better. An incredibly amazing job is being done behind the scenes right now.

[shell dawn]

yeah everybody is cooking for sure, i know you helped wiht some of that WhatsSkill, one of your recent PRs I merged to my 2026.5.24-beta.2 because it was nearly unusable withtout it

[fallen gate]

feel free to ping that lazy cat

[rocky pewter]

Thanks a lot. Everyone is doing their best right now to move OC forward together. Everyone in their own way, and in the end as a team.

[rocky pewter]

Nope ur the only 1 i will Ping 😉

[worldly tusk]

Tool/commentary draft blocks still missing?

[shy rain]

87022 didn’t land on 26b1. And my real fix hasn’t been merged either so unfortunately 26b1 is still unusable for me. Maintainers please check the subthread of perf fixes!!!

[shy rain]

Save the group hugs till my PR lands matey ;P

Please check! This is a beta blocker issue (two of them)

#1508450803248140399 message

[lusty vigil]

ok, took a look at this. appreciate the submission here. this is particularly big one (+4000 LoC is a biggie) and it's a surface area that's security critical so it's going to need alot of vetting and testing, which we don't have the resources for right now. i think the best approach is to revisit this in a little while, once we get through this stability sprint. maybe we could send a PR your way for review, once that happens?

[fallen gate]

naaah fuck this guy, @rocky pewter let's hug it out buddy!

[odd anchor]

https://giphy.com/gifs/moodman-kids-race-friendship-VduFvPwm3gfGO8duNN

[shy rain]

No wonder he keeps going on about hugging boys. PR merge when???? 👀

[shy rain]

At least buy me dinner first. And merge my PRs.

[odd anchor]

https://tenor.com/view/obi-wan-kenobi-these-are-not-the-droids-star-wars-jedi-mind-trick-gif-23067514

[shy rain]

Waiting for evidence for clawsweeper. Not easy this one. It’s an intermittent bug we are fixing. How do you provide evidence of something not happening when it usually doesn’t happen anyway … hmmm

[celest merlin]

it's that time of the year again folks...

OpenClaw 2026.5.26-beta.1 is live on the npm beta channel.

This is a chunky beta since 2026.5.22: faster visible replies, faster Gateway startup, stronger voice/Talk, broader channel reliability, safer agent replay, and much harder install/update proof.

Highlights:

• Reply latency and startup got real attention: hot paths now reuse command/model/plugin/channel/session/cost metadata instead of rediscovering it every turn.
• Talk and voice are easier to control: realtime Talk runs can be inspected, steered, cancelled, and followed up from Web UI and Discord voice.
• Transcripts landed as a core surface, with source-provider support for meeting summaries and transcript-backed workflows.
• Channel polish: Telegram typing/progress/forum topics, iMessage local attachment roots and duplicate-source cleanup, WhatsApp group/media fixes, Discord voice/model-picker/caption/proxy fixes, plus Signal/iMessage/WhatsApp reaction approvals.
• Agents are safer: Codex app-server auth and compaction recovery improved, legacy tool results are repaired before replay, subagent payloads survive better, and OpenAI-compatible providers avoid malformed empty-tool turns.
• Media got leaner: Rastermill replaces Sharp for image metadata, resizing, EXIF handling, and PNG alpha-preserving optimization.
• Better visibility: Activity tab, gateway secret-prep traces, model/tool stream progress, OpenTelemetry LLM spans, and richer stale/failover/liveness signals.
• Install/update/release paths got much tougher: Alpine installs, Windows stack-heavy startup, macOS restart validation, Docker/package timeouts, shrinkwrap/plugin publish checks, and cross-OS beta proof all hardened.

Validation so far:

• npm beta points at 2026.5.26-beta.1
• external OpenClaw plugin packages published on beta
• Parallels fresh/update/fresh-target passed on macOS, Windows, and Linux with OpenAI real turns
• Telegram live-frontier package E2E passed against the exact beta package
• Anthropic local proof is blocked on the current 1Password Anthropic key returning invalid x-api-key, not on the package path

Try it:
npm install -g openclaw@beta

Release notes:
https://github.com/openclaw/openclaw/releases/tag/v2026.5.26-beta.1

[timber beacon]

https://github.com/openclaw/openclaw/issues/85731#issuecomment-4545025068

this is ready for me to make a PR before I do anything else to check or should I wait?

[celest merlin]

I hope this gets us closer to it. lots of fixes and perf stuff in there and I been a human merge button - thanks for all your work!

[timber beacon]

Yay some good stuff in that beta!

[celest merlin]

@thick osprey generate a song that fits the changelog

[willow narwhal]

<@&1503801512294486217>

[shell dawn]

i installed it as soon as it hit npm, its been really good, but ive got a few more tasty performance/consistency PRs awaiting the next wave

[timber beacon]

Peter that issue got updated with new iOS changes for UI/UX

[shell dawn]

i found a new chat.history related slowdown too that im doing a followup on

[timber beacon]

Trying to help get us to App Store!

[thick osprey]

The generated music is ready.

[celest merlin]

you gotta talk with @lusty vigil

[rocky pewter]

bas... Gateway

[timber beacon]

Sweet! @lusty vigil let me know what else to do for it. Glad make changes to help get it going

[lone iris]

This is a vibe as I wait for npm to install!

[celest merlin]

I deferred that work to Josh, I lack human token to think about it.

[lusty vigil]

will take a look in a sec. standby.

[timber beacon]

No worries!

That’s why I wanted to help

[celest merlin]

I''m so hard rate limited by GitHub ugh

[rocky pewter]

The day have 24 hours i see ur merging/working 32 of it. No wonder it hits

[celest merlin]

I need a steipete2

now AI token are abundant, github and brain token lack

[lusty vigil]

need more sleep token, less thinking token

[shy rain]

If you have a fix, just have the bot check that it isn’t already fixed in the latest beta or main, and do the PR. If there are issues you’ll get notified so you can address them. Don’t wait, just do it!

[timber beacon]

Their music keeps me away when I have heard it

[timber beacon]

Yea it’s a full ui/ux change of the iOS app so wanted to make sure before doing the full PR. I’m happy to wait for the review as this should get us to App Store much faster haha

[astral topaz]

sounds good, please keep me posted. I'm rolling this forward every update for my personal setup, so it would be nice to eventually not having to resolve merge conflicts

[dusk vine]

Bug report: unsupported tool schema can crash assistant turn before content, and doctor does not catch it

Summary:
A broken/experimental OpenClaw extension exposed a tool named dofbot_move_angles with an unsupported dynamic input schema. When the agent started a turn, the runtime/tool projection failed before the assistant produced any content.

Observed behaviour:

• Assistant turn failed before producing content.
• Telegram/UI appeared broken or stuck.
• openclaw doctor completed without identifying the fatal tool-schema issue.
• Plugin compatibility checks did not clearly flag the active schema as a blocker.
• The issue was only found after inspecting runtime/Codex logs, which showed an error like:
  dynamic tool input schema is not supported for dofbot_move_angles

Expected behaviour:

• openclaw doctor should validate all active tool schemas against the same runtime/model/tool projection path used during real assistant turns.
• If one tool has an unsupported schema, OpenClaw should disable/quarantine that tool or plugin and continue running the assistant.
• The user should see a clear warning naming the offending tool/plugin.
• A bad extension should not crash the whole assistant turn before content.

Impact:
A single bad experimental plugin poisoned the whole agent startup path. The rest of the system was healthy: gateway, Telegram, model config, and core OpenClaw status were OK.

Fix/workaround used:

• Disabled/quarantined the broken extension.
• Removed dofbot_move_angles from active allow/config.
• Removed stale plugin install/MCP references.
• Kept primary model as GPT-5.5 and fallbacks as [].
• Fresh agent reply test worked again after removing the bad tool path.

Suggested improvement:
Add a doctor check such as:
validate active tool schemas against runtime tool projection

Also make runtime tool loading fail gracefully:
bad tool -> disable/report tool -> assistant still replies
instead of:
bad tool -> whole assistant turn fails before content

[potent path]

Thanks for sharing. Are you able to make a github issue (github.com/openclaw/openclaw) for this with more details about your platform/environment and how the issue can be replicated? (e.g. what OS and version, what version of openclaw, etc etc.)

[dusk vine]

Yeah I can make a GitHub issue. I’ll gather the exact environment/version details first and write it up cleanly with the runtime error and workaround.

[shell dawn]

if you talked with your claw about this, talk with it about making a patch it can turn into a PR, this is like half or more of the work

[pastel owl]

Really appreciate you and maintainers doing the break neck speed merges we keep sending !
I am very happy with this release 🙃

[timber beacon]

Yea stoked for the bidi realtime voice, its the best to talk to your agent and have it work on stuff

[lusty vigil]

i think this UI looks great! open to a PR for it!

[timber beacon]

Cool! I will do one last sweep with the beta

[lusty vigil]

feel free to tag @lunar granite and me

[timber beacon]

Sounds good, once this is good I can then work on converting to iPad and watch support

[potent path]

Quick bugfix: https://github.com/openclaw/openclaw/pull/87084
Addresses a "TypeError: unsubscribeDiagnosticEvents is not a function" error message caused by src/plugin-sdk/root-alias.cjs treating the dist diagnostic-events export named 'r' as onDiagnosticEvent. In 2026.5.25-beta.1, that 'r' alias points at emitFailoverEvent, while the real onDiagnosticEvent export is another letter. This PR now selects the function whose retained JavaScript function name is onDiagnosticEvent, instead of hardcoding an 'r' or another letter.

[trim vector]

quick approval. thanks for fixing. merged!

[remote wing]

Hey team, looking for a review on PR #86179: Xiaomi MiMo Token Plan support.
https://github.com/openclaw/openclaw/pull/86179

[surreal thicket]

Guys

[lusty vigil]

how's the beta working out yall?

[halcyon hazel]

Ok

[lone iris]

Fantastic. This is the fastest I've had my OpenClaw since I started.

Kudos to all of you for making it fantastic 🦞

[vocal night]

I feel this too 😮‍💨

[rocky pewter]

I think the fact that no one has replied speaks for itself 😂

[zinc dock]

Release candidate lol

[lone iris]

lol yes

[vocal night]

Not how it works normally. I’d say most beta testers active right now are nightshift (asia / europe)

[shell dawn]

its fast! got a bug fix to finsih plumbing though, my chat.history PR is a start, but the newer 2026.5.26-beta.1 has some new issues hidden behind it that werent there in previous versions

[vocal night]

I’m just waking up lol

[rocky pewter]

It could be, but the last few days were also pretty busy around this time. As for me, I’ll just have to wait until someone presses the usage button again or until the 30th.

[shell dawn]

tibo help us

[lusty vigil]

send your issues and PR this way!

[shell dawn]

the dependency change and rebase caused it to miss some of the details and there were LOTS of files that changed between it and 2026.5.26-beta.1. it was diamond crab earlier, let me get the agent to fix it up real quick

the other PRs are for issues found behind that one, so i'll get that one for you to look at while i test out the other fixes

[surreal thicket]

Are we fixing openai like hermes did guys

[vocal night]

Josh if you’re up brandons post is relevant to latest update

Not sure if our OpenAI team members put in same codex plugin patch

Or if its not relevant to our auth

[lusty vigil]

what post is that?

[vocal night]

This one. Dm’d you screens.

I’m on plane and my clankers are tied up so can’t look into it

I also mentioned it in our performance fix thread

[surreal thicket]

Im not sure what happened recently but codex absolutely sh*t itself with connectivity. Be curious to see what the PR was that hermes submitted for the fix

[vocal night]

There was a number of timeout connectivity issues reported in our thread last night so clarifying if still an issue for long runs

[shell dawn]

its been ok for me today but ive mostly been doing small but frequent stuff with it

[surreal thicket]

Ive tried both Pi and Codex runtime. Same issues

It must be openai side because hermes was suffering the same thing. Ive just tested hermes now and no more openai timeouts, fallbacks and connectivity drops

[final ocean]

it should not affect us, we dont use the same setup

but will raise with team

[zinc dock]

Yeah my OC agent is responding fine on OpenAI Oauth (5.22)

Been using it all day

[surreal thicket]

Interesting. Just curious in your codex auth setup @zinc dock Does it mention your codex email address in the config?

[shell dawn]

yeah thats the standard codex auth setup

[zinc dock]

Not sure I just did the regular auth onboarding process ...haven't had to re authenticate since I first set it up like over a month ago

[final ocean]

TLDR, we dont have this auth related issues. they copied some of our older patterns which we removed and cleaned up since we use codex harness, they ended up with split brain and nuked auth/tokens for users. I checked thier commits.

[final ocean]

I know Teknium is in this channel, open for him to respond too

[shy rain]

We are still affected by the codex binary going silent for no apparent reason. No way for me to debug that, we just work around it, but it is very annoying.

[final ocean]

share more

[shy rain]

https://github.com/openclaw/openclaw/pull/87079 -- info on the issue, my workaround to shorten the timeout from 30m to 60s for faster recovery. I don't have more info about the binary.. .it's closed so a bit opaque...

[rocky pewter]

Try the following -> when you open the picker, instead of selecting directly, type the name of your session first, then try selecting it and let me know if it works.

[final ocean]

will check. by arm you mean ARM chip based device?

[vocal night]

Peetie uses a potato

His potato has found good bugs that we wouldn’t catch off potato’s.

I’ve recommended that for rasberry pi type devices and potatos we have a mode that disables everything for the potato.

not p0 for most users but is a path that p’s talked about having support for just not sure how crit it is to today

That being said last 2 days codex has had timeout issues like crazy desktop and cli for the latest updates (them not oc) so also not certain how much is their end

[shy rain]

Did I mention anything about arms? I've updated the issue, too...

https://github.com/openclaw/openclaw/issues/87071

[shy rain]

The issues that I catch tend to be extremely inefficient code running when it shouldn't be, or timeouts set too low, so that it doesn't allow certain processes to finish. This is not that kinda flavour... but what it isn't is server side, as per my last evidenced example on 87071. I was trynna blame it on my proxy but it happened without even making a network call.

tl;dr - not my potato's fault, nor my proxy, nor the server. it seems to be the binary. (0.130.0 and 0.133.0)

it's hard to catch these I've had it just once this morning, trying to catch evidence for clawsweeper that my PR 079 was a valid mitigation.

[final ocean]

im looking into it

[shy rain]

I’ve recommended that for rasberry pi type devices and potatos we have a mode that disables everything for the potato.

I think this would encourage unhealthy bloat. This push for performance has really tidied up a lot of awful code. And some of it still manifests on more powerful machines but only when the installation gets bigger and more complex.

[true laurel]

I'm running on an older Intel chip, a 8700K I think. Not exactly a potatoe but even I noticed the dramatic slowdowns.

[cursive palm]

https://docs.openclaw.ai/install#hosting-and-deployment

Did anyone notice the messy formatting in the official documentation?

[jolly wolf]

checking

[pale ravine]

I found that many issues are caused by plugin updates not keeping up, such as with my lossless-claw. The updated OpenClaw is not backward compatible with plugins, so OpenClaw should not be updated. However, even when I asked the agent to check before updating, the problem couldn't be detected. Is there a better way to handle this situation?

[jolly wolf]

fix pushed

[fallen gate]

you tag @vocal night and yell at him

[pale ravine]

This is quite a serious issue because the entire conversation can be interrupted by a crash, making it impossible to continue, and the impact is quite significant
The agent later found the issue and provided a solution. For now, it seems fine, but if this situation is not resolved, we will have to spend time identifying the problem with every update

[fallen gate]

[pale ravine]

Didn't I say not to randomly tag people?

[jolly wolf]

lossless-claw issues are almost always caused by lossless-claw not openclaw, we dont break the plugin sdk between versions

[fallen gate]

nope I give you permission for this one ❤️

[vocal night]

I asked peter to just disable @fallen gate @celest merlin basically LCM will and has broke every release for a month. @jolly wolf

[pale ravine]

Next time. This time, I just wanted to find a foolproof way to handle it.
So the best approach is not to use any external plugins?

[final ocean]

lossless-claw is a major hack into our system and is doing stuff beyond the original plugin SDK

we cant optimize for an external plugin

[vocal night]

The issue is backwards compatibility is 100% never same as release timeline the fastest we can get fix is AFTER release

[final ocean]

wrong we maintain a very strong backward compatibility with our SDK

[jolly wolf]

the most foolproof approach is to use plugins that use the sdk properly

[final ocean]

issue is how lossess works touches all the memory and compaction in ways you would not belive

[pale ravine]

I see many OpenClaw users recommending LCM, and the results are indeed quite good
First time hearing about such a serious negative impact on OpenClaw

[final ocean]

one change to our internals (NOT to our SDK) and things break

[vocal night]

Lcm doesn’t patch anythjng

I made sure of that

Its codex assembly we just took out the 70k guard that made lcm work

[final ocean]

you work on LCM?

[vocal night]

Yeah lol

I try to keep it up to date while josh is out w the baby

[final ocean]

so whats with all these issues, how can we avoid this?

our plugin SDK has not been shifting

[vocal night]

It doesnt go thru sdk it manages assembly

[jolly wolf]

And that’s the problem

[vocal night]

If you wanna call me you can and I’ll go over real quick

Context engines manage assembly

[shell dawn]

contextengine, compactionowner, memoryslot. its different than your average plugin

[pastel owl]

Which version are you on ?

[vocal night]

And own compaction

[pale ravine]

Let me confirm—so is it strongly recommended not to use external plugins with OpenClaw? Or does OpenClaw have a list of supported external plugins?

[vocal night]

Codex when it came thru took over assembly and compaction from OC

Its actually all codex issues. So we had to “bandage” it to play nice that was the 70k guard. My patches on threads fix all this @final ocean

[shell dawn]

openclaw is in a period of transition right now, particularly advanced features have a lot of catching up to do with codex runtime while there are major changes version to version. once it settles down so will products that use advanced features and things will be good

[vocal night]

Problem is we now have to take bandages out of LCM

Codex team half implements everything lol

Most of my PR’s are cleanup on codex. 10% are making it work with LCM

[pale ravine]

So, it is strongly recommended not to use LCM at this stage, and it will be handled later once it becomes stable, right?

Sorry, the translation can sometimes be a bit off, so I'll double-check it multiple times

[true laurel]

That's hard to say, some Eldritch blend of 5.25 from Monday with some performance PRs merged in manually, including yours. I haven't had time to do a proper clean-up pass again 🙂 it's running okay right now

[final ocean]

its going to be hard to have it stable on codex imho, as we change our internal harness wiring all the time

[vocal night]

Josh doesn’t use desktop codex he only uses his claw and he’s only one who pushes releases so LCM doesn’t keep up with the releases and by the time a fix is out every oc user is broken and they dont kno to update lcm

[final ocean]

just not sure of an elegant solution

i spoke with Josh a while back about this

didnt know the extent of this

[vocal night]

Well the threading issue is the elegant one

[pale ravine]

No problem, thank you for your continued efforts. I will disable the LCM and only use the plugins provided internally

[vocal night]

It allows me to take out about 10k loc on lcm

And its not “for lcm” its because codex threading wasn’t implemented

LCM is a necessary annoying evil as it catches bad implementations by codex typically haha

Mostly around data cleanliness

[pale ravine]

However, I want to know if there are other features on OpenClaw that can replace it after disabling LCM, such as Active Memory

[vocal night]

That being said when we move to Sqlite db 60% of lcm code goes away

[final ocean]

@vocal night ive still been cooking on your PRs, will probably stew on /goal for 1-2 days

[vocal night]

Codex even without lcm was broken how it handled cache, assembly and threads we only uncovered it when trying to integrate lcm

When we refactor to sqlite we’ll be able to cut as much code as we’ve added

But my opinion is context engines should be auto disabled on releases and have to be re enabled - we used to do this @final ocean

[pale ravine]

It seems there is no alternative to LCM in OpenClaw.
Thank you, I'll go turn off the LCM first

[vocal night]

And simply throw a flag “context engines disabled on release updates to ensure user manual compatibility- check context engine release page for latest update for compatibility”

[final ocean]

thats an anti-pattern

people wont want to keep re-enabling things

[vocal night]

I know but openclaw comes first

[final ocean]

LCM is an opionated way of handling long context

[shell dawn]

there are others, contextengine is just advanced and its a tricky area to get parity with between pi and codex runtime

[final ocean]

yea but at this stage i would just remove / block LCM then?

if its disabled people wont re-enable

[vocal night]

Oh they will haha

But tbh its a convo w josh he can’t keep up w release updates

[final ocean]

im against this behaviour, we are working with the idea most claw users are not advanced

[vocal night]

And its a burden on me and I don’t maintain either end - i usually patch both sides within 24 hours

But by then everyones broke and wont manually update

[true laurel]

I don't think people are even aware you need to update plugins separately?

[final ocean]

This is an external plugin, im open to suggestions to improving things but very hard

[true laurel]

I mean, unless they are manual things you install, of course.

[vocal night]

Plugins only update on oc patching

[final ocean]

we do this automatically at update if a plugin is updated

[pastel owl]

I am using today's main and its crazy fast uptill now, without any issues so far! With Lcm
Today's main has my PRs merged ✌🏻 and i saw many more perf commits by fellow clawtributors and Peter

I think you should inatall current main vanilla

[vocal night]

Normal cadence in enterprise would be to email all 3rd parties or send update hey update plugins release in 1 week

[celest merlin]

@thick osprey read the room and make a smash hit song

[vocal night]

In this case 4-8 hrs is never enough hell it takes me usually 6 hrs to patch both ends

[final ocean]

you will still be waiting lol

[vocal night]

And to be frank even though my threading patches will fix it

[final ocean]

lets find a way to make this more amicable

[vocal night]

The sqlite update will 100% break it

[thick osprey]

The generated music is ready.

[final ocean]

seems like its broken every release lol. but let me try to make this somewhat manageable

[vocal night]

I’d like to just implement sqlite @celest merlin as option and ill test and patch it

If we can get that as opt in

I’ll get sqlite working in 1-2 wks

[final ocean]

@vocal night happy to let you know when we have native sqlite

[vocal night]

Then tbh most of lcm features are just plugin

[final ocean]

you can test drive

and help give feedback on sdk for this

[shell dawn]

im looking forward to testing it too, ive got a similar arch on my memory system

[final ocean]

you wont need to handle your own storage, we will do that and give you a clean interface

thats the idea

[shell dawn]

my analysis on how it changes is similar. a ton of code is to deal with JSONL grafting/cleaning/fixing, native SQLite means most of it is native and to change my memory system on how to use it more specifically for my goals

[celest merlin]

lcm does super deep changes, e.g. Josh smuggled in a PR that adds new plugin sdk seams without discussing it, and also it was incomplete.

[final ocean]

yeap

[pale ravine]

The agent is helping me clean up everything related to LCM.
I think OpenClaw could provide a check before updating. If it detects that external plugins are being used, it should pause the update and issue a warning, allowing the user to confirm whether to proceed. Of course, if there is a list that lets users verify whether the external plugins they are using will be severely affected by the update, users would naturally not update carelessly and run into problems

[final ocean]

no this is bad, we optimize for unattended installs. most people as the clanker to update. we do this for 1-2 plugins and break for everyone

if a plugin is bad people should turn it off

[vocal night]

The most valuable parts of lcm that I think Oc should take are my layered tool stubs it cuts token cost like crazy however it only works in layered db wont work in jsonl or without a call tool

Agent thrashes and forgets etc

[pastel owl]

Do we have a min max version supported metadata for plugins to publish? If not in supported range, oc can suggest to update/disable plugin or stall oc update

[pale ravine]

Maybe you're right, this is just an unprofessional idea

[final ocean]

Yes we do have this, you can specify

[vocal night]

Outside of that its all opinion of memory

[final ocean]

we optimize for the largest % of users that are not using terminals etc

if we break something they cant recover

you have the choice to wait and not update

if you really want compatibility with a plugin

my issue with this whole LCM thing, its a superb concept but if its breaking a large % of claws on every update and people cry about it me and Peter, what am i meant to do? open to suggestions but hacking our updates etc to please people for 1 plugin seems a little crazy no?

really open to finding a better work-around, if you have any @vocal night

[pale ravine]

Our thoughts align perfectly, haha

[vocal night]

If I was in your shoes I’d feel the same

[true laurel]

If enough people use LCM that we're talking a significant portion of the users, then the simple solution would be to include "Does LCM still work?" as part of the testing cycle.

[vocal night]

The project needs more than josh if its to keep up

[vocal night]

Its a lot of people

Peter promotes it on his X

[shell dawn]

a co-developed integration area, where people making plugins and maintainers can work on upcoming changes, though really, the timeframe on openclaw updates is probably too fast to make that viable

[true laurel]

If it really is a lot of people, then it just needs to be made part of the formal development process, CI should test if LCM loads, if possible.

[shell dawn]

taking a page from peters distributed testing units

[vocal night]

Not actually a bad idea- does compaction work and assembly

And if it fails disable

[shell dawn]

PRs are reactive, more advanced stuff is going to need to be a two way testing integration

[vocal night]

LCM fired off on startup and if it has any errors disable

[pale ravine]

Actually, I don't intend to let my updates be hindered by a single plugin. I just want to know whether I should update, or if I do want to update, which plugins I should disable to be safer

[vocal night]

It sends the assembly payload to OC at startup and will also do a compaction etc

[shell dawn]

yeah but then you have to capture all of the potential failure areas. it can be done but its a list to make

[vocal night]

Hmm we could also just auto disable plugin if it fails any on its end

The failure area is only 2

Compaction and assembly

[shell dawn]

tool calls

did the failure work in the way the test was designed to predict?

[vocal night]

Doesn’t matter never broken before, only time it broke was when auth changed from open ai to codex and model refs broke on oc side

[merry slate]

I think the definition of the version supported is probably the best: an OpenClaw version that’s compatible with the plugin config. If the versions don’t match, either auto-disable the plugin or prompt the user to confirm before proceeding with the upgrade

[shell dawn]

its a road to travel, maybe not as long as the pessimistic part of me is thinking

[true laurel]

If someone submits a PR that breaks LCM, it should not get merged until fixed or waived by maintainer.

[pale ravine]

It is difficult to determine whether this error is caused by a configuration issue or by a plugin problem. A misunderstanding might lead to disabling a plugin that only needs its settings adjusted

[vocal night]

I’d say LCM is one of our greatest benefits historically however at least 30% of the shit we’ve gotten online is due to it being managed by one maintainer who had a baby and has a job

[final ocean]

@vocal night suggestions:

• write access to LCM repo for some of OpenClaw maintainers
• we have added to openclaw/crabpot, support on better testing automation of the context engine so you and we know something may break
• gracefully degrade LCM plugin rather than break fully for users
• better version locking etc
  some ideas

[vocal night]

It has a lot of infrastructure and ability to break for the limited bandwidth

[vocal night]

Yes oc maintainers should have access josh I usually have to wakeup and blast and he’s gotta do it between bottle feeds release wise

[true laurel]

If plugins get disabled upon update, then people feel like their claw degraded. The update should warn them beforehand that it'll break LCM if that's the case, so the user can decide what level of degradation they want to accept.

Alternatively, provide a robust revert path.

[final ocean]

this would be worse case, but honestly im worried on testing. you should know somethings broken and so should we

[vocal night]

Patch is normally 100-300 loc

[pastel owl]

Why not just ask LCM developer to put this functionality in the plugin itself
And also autoupdate plugin if issue fixed later

[final ocean]

we cant do this. if LCM is broken it should not fuck up your claw

[final ocean]

@vocal night is on the LCM team

[vocal night]

Well josh has been against turning plugin off if its broken

If you can convince him ill have that patched in asap

He just doesn’t have bandwidth to keep up with velocity of oc. Lcm is stable elsewise but if someone tweaks codex assembly well it would break.

This is all theory tho as of this release it actually works

[final ocean]

this sounds mad

why have broken claws?

[vocal night]

This release removed the 70k thread guard. We just need josh to remove the patches that used that.

If any lcm user rolls back actually it works just that one feature for assembly was a bandaid vs my full fix

[final ocean]

we may have to forceably or heuristically block this behaviour if its breaking claws

[final ocean]

but you said its broken every release and so did few others here?

[vocal night]

It has but codex is actually stable from lcm perspective now

[final ocean]

how bad is it when it breaks??

like people cant use claws?

[vocal night]

It usually just degrades to normal claws however this month it has had lockup issues etc

[pale ravine]

I was in the middle of a conversation with my claw, and then it couldn't continue—just kept showing an error

[vocal night]

But normal claw experience is then dumb

[final ocean]

usually?

[vocal night]

So they are like why does my claw not know anything

[final ocean]

@vocal night

[true laurel]

I agree, if LCM is broken with the latest version of Claw, then the user should be made aware of this.

Maybe a min and max OC version tag for a plugin, so if the user tries to update OC, the update process warns them their plugins do not yet support the latest version, and if the manifest for a given plugin (something in their github repo for example) does not yet report it being updated to a certain version, the user can decide to not update?

[pale ravine]

Otherwise, I wouldn't have come up here to complain

[true laurel]

FoundryVTT does something similar with their modules, you can specify which version of FoundryVTT your module supports. Users can update, but they'll know what modules will break when they do that, BEFORE updating.

[final ocean]

this also wont work as we release too fast, by the time they up version on thier plugin we do another release

[pale ravine]

It took a long time to confirm that it was an LCM issue

[vocal night]

correct

[final ocean]

shit should just downgrade

[vocal night]

It’s actually really easy switch to downgrade

Graceful degradation

Now***

[final ocean]

then whats stopping you?

[vocal night]

Okay so josh only really maintains it with his claw and if his claws broken also he cant check my pr’s

[celest merlin]

yeah it shows

[final ocean]

omg

[merry slate]

for LCM users, it might end up having to specify versions during installation, just like npm

[celest merlin]

it doesn't se autoreview

I had to fix his stuff

[true laurel]

I mean, that's the problem of the plugin dev 😅 But it allows the user to make a call "Ok I will update to OC 5.22, I know OC 5.27 is out but the plugins I need work up to 5.22 so I'll just update to that instead"

[vocal night]

The project is in maintenance mode

[final ocean]

if this is not resolve we may have to take further action to protect openclaw users, hope you can understand @vocal night

[celest merlin]

fork it or ask for access?

[vocal night]

Nah I understand

Its weird cause its owned by his company

Its not real open source like that

I talked to him about maintaining it

[pale ravine]

I'm curious about one thing: is Andy one of the developers of LCM? I can only guess from your conversation

[celest merlin]

MIT

[final ocean]

MIT = free to fork and change

[vocal night]

If we forked it we could basically keep it up to date its literally 100-300 loc each release

[celest merlin]

then fork it?

[shell dawn]

ok a little bit of a crazy idea, and there is a lot of surface to cover, but somekind of runtime bridge contract or something, that as openclaw developes so does the contract, and it can show how the shape of things is going that plugin developers could use as a map to see how to pick up changes

[vocal night]

Its auto doable via workflow

[vocal night]

Not gonna call you out but you also share on X haha

So when you share we get a shit ton of users

[final ocean]

complicated, we have this. but LCM and context engines are "special"

[vocal night]

They brake and come complain here

[final ocean]

hence all these issues

@vocal night i feel like im missing context here

some internal politics at Martian or somethin?

[vocal night]

I dunno i think 90% of it is the original creators no longer around, and its just him and he had baby 2 weeks ago

And he uses his claw to update

So its slow

Vs codex

[final ocean]

then stop maintaining it and say project is abandoned

and let community and others fork

[vocal night]

I do my best usually it takes about 24 hrs but as you said by then everyone is broken

[final ocean]

people think lcm is partially maintained by OC, thier stuff breaks and they come to us

[cinder bolt]

It really looks neat. Is there some code to integrate it with OpenClaw, I need to direct my claw to look at clickclack and then it will respond

[vocal night]

Right its promoted and endorsed in many ways by oc

Its fantastic plugin i have my own fork thats way better

[shell dawn]

right, but not all features of openclaw are so externally malleable like the contextengine, when it was just pi it was largely something you had to figure out once, but the differences on a lot of what contextengine covers between pi and codex runtimes are exposed in this feature and now you had to figure out two sometimes very different behaviors

[vocal night]

I forked it a month ago and maintain that separately and push patches up

[pale ravine]

I need to clarify one thing: I don’t believe the LCM is maintained by OC, but when a problem occurs, we have no way of knowing where the issue lies, so we can only turn to OC

[vocal night]

For context I’ll take joshes perspective - lcm still works amazing for pi and other models

Its degradation in last month is majority codex related

Which our user base is also majority now

Lcm was designed around anthropic models, compaction, and cache so its been an undertaking to also map it to openai

[pale ravine]

Because of the OpenClaw update → problems appeared = it's OC's fault

[shell dawn]

making hypermem i banged my head against fixing tool calls when doing model failovers, not getting in the way of compaction and such, and then with codex runtime i had to do it all again and some of it was counterintuitive to what was understood on pi

anthropic FREAKS out if toolcalls are not symmetrically represented, even openai on pi was not happy, most models arent, but codex handles it totally differently because its doing more behind the scenes than pi

[celest merlin]

Which is dumb. It does so much complex context management - the way it has to massacre codex is not good. It should just not work with codex. Openclaw default harness is fine. codex gives you enterprise features like guardian. For some that's very hepful, I doubt the venn diagram of losssless usrs that care about that = tiny

[vocal night]

Reading all the convo Eva said “Peter and V should enforce a guardrail on LCM and context engine degradation. If they don’t degrade if error’d on compaction or assembly, and disable. They cannot be used by OC”.

[true laurel]

While the reasoning is understandable, you can't fault Apple for breaking Some Random App when they update iOS versions. If the app still runs, it still runs, otherwise, the app developer should go fix the app.

[celest merlin]

And I would not have to deal with 2000 LOC spaghetti on extensions/codex/src/app-server/run-attempt.ts that is causing bugs

I'm very tempted to just rip this out

[vocal night]

Only pushback I’ll say is codex is shit at memory and context management its designed for one off thread coding agents

[final ocean]

its happening now

[vocal night]

If people want to use codex they’d just use better codex desktop or cli haha

So there is some push and pull

[final ocean]

im putting saftey measures in place following this conversation, i dont have any other choice. i have to protect openclaw users

[vocal night]

Josh if he had the time lcm would never break im sure

[true laurel]

I don't care for Codex or Pi, I just want a happy and healthy claw 🦞

[final ocean]

a plugin cant break users experiences constantly this many times

[pale ravine]

If an application has a problem, users can guess that the application needs an update, but plugins are embedded throughout the entire work lifecycle, making it difficult for the average person to quickly identify which plugin is affected

[final ocean]

im running codex for 5-6 days long with billions of tokens in a session no issues

[vocal night]

You use for coding

[final ocean]

some of these plugins have found ways to get around our limitations and circumvent things so hard to find

[vocal night]

Users don’t do that they want to talk, have convos. What did i say in meeting with so and so

[final ocean]

yea true, i dont argue its great for conversations with LCM

[vocal night]

Well I’ll also say I’ve told josh the solution here is errors in lcm must surface

[final ocean]

i love the concept but we cant have it broken and still pushing it to people

[vocal night]

So one thing that he changed recently was errors go to debug mode only

That should also be a guardrail

Agents dont even know when lcm breaks

[final ocean]

😭

[true laurel]

I've not used LCM, my claw didn't want to install it, said it was too scary 😄

[vocal night]

So users cannot debug

As they dont know how to turn on debug mode

[pale ravine]

My claw recommended installing him, so I did, and the effect was pretty good, but that was a while ago

[vocal night]

Historically agent could fix instantly because lcm errors tell agent what to fix (historically it was config issue or model ref issue)

[true laurel]

It's probably because I told my Claw to be cautious with installing external software without trust.

[final ocean]

@vocal night so you dont have maintainer/write access to LCM?

[pale ravine]

My conversation with Claw is now full of ⚠️ 🛠️⚠️ 🛠️⚠️ 🛠️, but luckily it doesn’t affect the operation most of the time, so I’ll leave it alone for now

[vocal night]

Nope I just ship the code and wakeup josh to have his oc review and push release

[final ocean]

just worries me since you care so much about LCM

so its a one person show

[true laurel]

I made a local patch to disable tool warnings/errors to get rid of that

[vocal night]

Well its a gap, historically its been promoted by OC and peter and probably lost us 20-30% of users who broke on an update at one point and dropped off.

At same time it was part of early magic

[pale ravine]

My choice is to adjust the settings if it's a configuration issue, otherwise wait for an update from OpenClaw

[vocal night]

So its mixed bag.

The instant memory recall and how it leaf summaries is great I run a memory company I know haha

Its like my oc, when it works its magic

So I also recommend having some mercy haha

[true laurel]

Why? Technically my agent made that patch by themselves, I didn't even ask it to fix it for me, I just woke up one day and it told me it patched out an annoying error in Discord 😆 We've been carrying that one self-owned bit of patch work forward a few versions now. I still don't exactly know what made it do that.

[vocal night]

At same time external pressure i think it needs more maintainers to be so closely associated and promoted on OC socials

[shell dawn]

contextengine is powerful, i focused on a different area and on pi it was working so wonderfully, i havent gotten back to parity with codex yet. developing a bridge contract format that can be updated during the development process can help give a map of where plugin authors need to target

[true laurel]

I thought Codex was just another LLM call-through? What does Codex do that is so special?

[vocal night]

And I’m not volunteering or using as way to take over it lol, peter knows how I feel about responsibility. I like to add value but sometimes I take a month vacation in hawaii and disconnect. Retired life.

[vocal night]

Context engines do compaction and context assembly better than default pi and default codex theres no question

[shell dawn]

its the core llm engine that powers the codex cli/desktop app, and its getting adapted into openclaw, it works very differently than pi

[pale ravine]

Since the next update is more likely to overwrite these fixes, I chose to have my agent submit the issue and suggest a direction for the fix

[shell dawn]

it has a lot of assumptions and techniques and methods it does already and its pretty different than pi which is how openclaw developed over the months

so when that core changes and it brings a whole ton of associated code with it, a lot of little things change that depend on those particulars

[true laurel]

hey @worldly tusk did you ever figure out what happened to the reasoning/commentary that agents used to report?

[worldly tusk]

nope, it just went away one day... very sad. the most likely culprit is https://github.com/openclaw/openclaw/pull/85488. but i haven't got time to analyse properly

[true laurel]

I tried some shenanigans to get it to work again but little luck so far on my side.

[worldly tusk]

you could try to revert this PR locally and check if it returns the commentary

[vocal night]

Yeah maybe this pressure will allow him to talk to his biz partners to allow project to become just open source if it does and team grows it will be much better

[worldly tusk]

it is opensource. anyone can just fork it 🙂

[true laurel]

I think that commentary has been gone for longer than a few days.

[vocal night]

That’s true, maybe I need to OS my fork of it.

[worldly tusk]

4-5 days. that's why i am not pointing that this is the exact PR that broke it, just say that this is the most likely.

[true laurel]

No I feel it has been gone for a week or two 🤔

[worldly tusk]

no

[final ocean]

fyi our recent multi turn eval based on time taken on a number of channels by versions, recent betas are 2-4x faster on average across all channels. will share more detailled evals soon

[shell dawn]

hell yeah shes zooming now

[true laurel]

Good, our work paid off 😄

[zinc dock]

Discord is such a dog lol

[shell dawn]

could webchat be added as an entry in the test matrix? its an easy surface for people to target to customize something they want to make

[merry slate]

looking forward to seeing technical insights about this

[worldly tusk]

@vocal night, you'll love this: https://github.com/openclaw/openclaw/pull/87155 😄 Nik plans to put soul.md, identity.md and user.md back into subagents.

[crude matrix]

I am willing to test, because at this moment Hermes is my main agent and I would like to switch back to OpenClaw.

[shell dawn]

you updated to 2026.5.26-beta.1 yet? its gone through a serious training montage

[final ocean]

hard-ish to test but a valid point, will look into it yet

[vocal night]

Lastly before I get on my plane can we have a call on this @final ocean @celest merlin every fix we put in place pash reverts or changes and we’ve been back and forth on prompts and how default behavior works. They need serious testing before just changing willy nilly one maintainer

[final ocean]

any key reason(s) for switch, curious

[final ocean]

ill dig into this and we can discuss when your back

[vocal night]

This pr is like the third uno reverse lol

@worldly tusk I’m on plane taking off soon can you bullet the issues

I built the the code it sits on but tbf its become so complex that I have to run 30 mins agent to cover all potential scenarios and a parity test to touch any of it, its just that complex of a setup to bolt onto

I recommend freezing changes to prompts and harness unless its p0 and we have very stable version to tune

[vocal night]

I guess I’d call it the bolt on adapter* codex harness + pi harness + middle ware between is where I mostly live 🤣

[celest merlin]

Sounds like the "fix" we did was not understanding codex and thus messing with the context, and there's no great way to customize subagents but also not pile up stuff on every turn [without deep investigation, Nick worked on codex for a year so I trust his judgement]

[vocal night]

The fixed just has to be comprehensive. That part of the codebase is complex and every fix has and will cause regressions when coding with agents vs manual review (have to cover multiple runtimes and multiple scenarios), have to test for each runtime and failure mode AND run actual test to catch where the hooks break or don’t fire or it breaks cache etc.

If codex team had their way we’d just be on codex. I’m contrarian in that I do like to preserve some of the openclaw magic as I say if people want codex they’ll just go to “better codex” not bolted on worse codex.

The PR’s tell the story better than I can, they do a fix, I do a cleanup patch, someone else breaks it, I’d just recommend treating it like security- adapter harness + prompt changes need significant testing and cannot be rolled into releases without check and balance.

[worldly tusk]

i've not seen any evidence that there something inherently wrong in injecting core persona files per-turn in turn developer instructions. they are not duplicating in the resulting API request and they are not increasing token count. maybe there is something i am missing?

[vocal night]

Might not be immediately breakage but is silent hard to trace bugs like long running agent timeouts, personality degradation etc.

I trust codex team for building efficient coding agents. OpenAI is not known for building personality and soul haha

[worldly tusk]

LLM does not care, it doesn't have any notion of threads. they are just codex abstractions

[shy rain]

What do these numbers represent? and what is P50 ?

[vocal night]

I’m not saying bad pr yet just saying prompt and harness level shouldn’t be last minute release adds. I need at least 3 hrs to review any pr that touches to patch the regressions codex team will add in

[final ocean]

50th percentile = average, 95th percentile is the spikes/upper number (not shown here)

[vocal night]

And make sure it doesn’t break a bunch of other systems

[shy rain]

what the.... nooooooo

[vocal night]

Pash strikes again 😮‍💨

As we like to say here

[shy rain]

oh cool thanks but what are you timing? Is this comms channel overhead only?

[vocal night]

Tbf tho it could be good PR just saying lets not merge in without testing lol i cant load agent on shitty plane wifi

Also it seems scoped only for discord and how discord handles agents, probably needs to be a discord fix not a universal @worldly tusk

Universal patches for one thing without considering everything it touched is what causes a lot of our issue in this area

[final ocean]

100-200 rounds of simulated messages and gateway restarts, the turns is what we are measuring and the average, we have a whole bunch of other stats. will share more when i have, still running

[worldly tusk]

no, the Nik's patch just moves soul, identity and user md files into thread-scoped developer instructions. everything else are just tests

[vocal night]

Can you require that on prompt and harness adapter pr’s too

[vocal night]

I forget didn’t i put them there first and he moved them out and now he’s moving back?

When @final ocean merges my thread fixes it technically will be okay for dev instructions

[final ocean]

its going to take another day to cleanup the prs and then given all this drama i will be taking it very carefully and understanding the issue. expect some delay on this one

[vocal night]

In dev instruction (inject once per session /new and once at compaction), in my testing personality and instruction degradation is about 50% at 20 turns and then post compaction 80%

Tldr is moving default core oc behavior from bootstrap to dev instructions once per session, more token efficient but personality and agent degradation is high

[final ocean]

this is designed to measure perf, we have a full task based eval suite which is what labs use and we have been developing and will be sharing soon for long running stuff to better understand config, models, harness etc

[vocal night]

It really should be an option config i kno we dont want more but users should have choice of default behavior oc for all time vs pash wants to save tokens

[final ocean]

i dont think there is a drive to save tokens

[vocal night]

Pash explains it that way every pr

Its token efficiency over personality lol

[worldly tusk]

i think he is wrong there

[shy rain]

The difference in times between Telegram and everything else is striking. Not a small difference at all.

[worldly tusk]

you moved them in another place 😄 he moved them back into thread developer instructions. i moved them into turn-level developer instructions.

the problem of thread-level vs turn-level developer instructions is that thread level is inherited by everything within that scope, including subagents. if we say that subagents should not have these files (everything they need is provided by parents), then thread developer instructions cannot be used. there is no way provided by codex for us to separate normal agents from subagents (maybe except for multi-agents-v2, but that is experimental)

codex however provides per-turn developer instructions. they are not inherited by subagents, so i moved soul, identity and user md files there. they are still developer instructions, so they don't lose priority in the hierarchy (that your patche degraded somewhet, btw). and also they are not increasing token usage, as the assembly puts them in LLM prompt only once, they are not duplicating.

so in reality, Pash is not saving tokens and not improving any other behavior - by this revert the only thing effectively that will be done, is the soul files will be inherited by subagents

[vocal night]

Good review. Only thru disagreement and testing does magic happen, I just wish codex team would use auto review and spend time architecting first before just sicking gpt on it.

You can be smartest person in world but if you don’t architect, try to understand how it works, or even spend half hour w agent doing so your code is gonna be subpar fix.

[vocal night]

Got it. Hmm I can’t load git on my phone but I align w this.

I did a whole breakdown chart on my PR so this sounds right.

Now I’m on token efficiency and no personality wagon 🤣 BUT for subagents.

Main agent = need personality
Subagents = wasted tokens

[vocal night]

About to lose internet over ocean but fundamentally at end of day maintainers have to ask on vision wise (like when vincent said 6b tokens and I said yeah thats for coding agent work not conversational/assistant/personality/human like memory etc what LCM does better):

1. Are we building a less good codex (ie a coding agent, because we’re never gonna be as good as codex native as a harness riding a harness, we’ll just become more codex till it becomes “less good codex” and people just go to codex)

2. Or is what makes us different our customizability and personality or something else that gave oc its magic over claude code and codex months ago.

If it’s one, well maintaining is just a long interview to work at openai or making something that will never be “as good” as it.

If its two the question should always be how do we keep OC magic while improving it and thats the north star- to me thats personality and why I’m contrarian here.

I’d love to see molty consider a scoring on PR’s on prompt or harness - “are we chasing codex here or making OC magic” haha

[worldly tusk]

yeah, we aren't going to out-codex codex. ever. people can just use codex

openclaw in my humble opinion should stay on the personal assistant with personality lane. it must do good coding, but it should not be a codex replacement

[vocal night]

Also if its 1 I have a plugin that lets oc control codex desktop and cli with real organization and deep layers, 100x better than all our harness and finagling bandaids 😮‍💨. Eva on oc runs 5 codex desktop apps and keeps it all organized. Her subagents are awful in oc.

Thats the 1-1 and honestly where I think this goes in six months anyways I love eva controlling claude code and codex desktops i never run oc subs anymore

Lmao my only subs are claude code plugging into oc to get image gen and makem into game asset factory

[worldly tusk]

i still think grok is better at images, but that is probably subjective

[vocal night]

Doesn’t have the workflow ability I need for game

The option to me is manage 5 subpar token expensive subagents or manage 30-50 agents on long running task and projects with database level management + clean.

[shy rain]

I just have any of my apps use each other via tmux.

[vocal night]

Openclaw running all my coding agent platforms and being layer between them has been a gamechanger

[worldly tusk]

codex has computer use now, so it can run that as well

[vocal night]

I haven’t used it, I built a middle task management board system between all the sqlites and use lcm on claude code and codex for one central brain of everything going on always

[shy rain]

I just posted in general my token usage from 1st April. I've used claude code and codex-cli way more than openclaw during that time. Mostly because openclaw kept breaking and I spent most of my time trying to fix it with those tools.

[vocal night]

So one lcm grep or gbrain call and know everything

[vocal night]

They bought sky or something. Managed to reverse engineer it since its just silly wrapper on mac accessibility

Peekaboo also good I was gonna post report on how codex does computer use

[vocal night]

You've hit on something really important here. No agent today has been able to balance agentic coding and personality — it simply doesn't exist yet. You will always end up with either one or the other, at least for the foreseeable future. Codex made that decision deliberately — they realized it too, which is why they added that short, one-paragraph friendly prompt to make it seem like it has personality. But it really doesn't. You either choose one or the other or are mid at both.

GPT-5.5 is actually better than Anthropic's models as a coding agent largely because Anthropic's 100+ emotional vectors make their models susceptible to internal pressure, breaking, and long-term stress (emotional vectors + creative temperatur over stable they can build a website + are more creative/architect well....and gpt 5 builds 2000 myspace creative wise).

Internalized pressure from emotional vectors ultimately prevents them from effectively running and executing goal-oriented activities /goal like gpt 5.5, they will eventually degrade significantly whereas 5.5 can keep on chugging, it may never paint the ai mona lisa but it can code lol

[cinder bolt]

I've been a developer for like 20 years so I can code, I just want an assistant to do grunt work and have a more intellectual conversation than a yes-bot

[crude matrix]

Typing the the name does not work, workaround for me is go to the session page and click the session I want from there.

[cinder bolt]

yeah, the webchat is a bit brittle. I've been using the session-list to get to the correct session myself even on the old webchat.

Is the webchat still integrated in the core? Or is it a separate module nowadays, it's hard to keep up with the changes 😄

I think it be a separate plugin like the other parts like discord, telegram, etc..

That would unlock releasing UI polish separately and at a different cadence.

The ClickClack UI looks neat
https://github.com/openclaw/clickclack
I haven't tried it out so much but my claw needs direction to use it and it will not automatically respond to messages

I guess in the future we get the WebUI to just contain config of the claw and a separate app for webchat

[shell dawn]

i think your problem is a PR that i got in, there were lots of refactors from the beta so my specific PR wasnt committed but peter took the ide and i believe used the fix

[willow narwhal]

Did so much work there but been busy on openclaw these days.

I'm currently focused on the webchat and it should be much better soon. Did a rewrite that fixed chat rendering on the webchat

Before vs after

Even your channel related chats gets rendered correctly now

[crude matrix]

a periode I also had issues with Lossless-claw.... and I did miss it.. without it my agents performed so much worse after compaction. Same goes for Active Memory and wiki... Openclaw is really forgetfull without these external plugins..... I believe similar functionality should be standard build in... a agent without memory is nothing. I do find the current options way to complex for most people, maybe it should work with profiles, off, low specs, medium specs, high specs and advanced (manual config)... Personally I think thinks it would be nice to replace the onboarding configuration with a Agent intake talk.. agents should ask questions and based on this choose best configuration for user and system.

[cinder bolt]

Yeah, I like the new webchat. One thing that would be neat would be to support getting the streaming response from the claw so I can see the message in real time for longer agentic tasks.

Has it changed how to inline a image? For example discord or telegram it inlines a image by just adding the full path of image, but on webchat it's different, I don't remember the details though 🙂

[final ocean]

Agent intake talk is already in the works

[willow narwhal]

Yeah, gonna work on the stream.

I want it too

[crude matrix]

Yes tried beta 1 and openclaw 2026.5.26-beta.2... issue is still there, workaround for me is accessing sessions trough session page. and clicking on the session there.

[cinder bolt]

This was like two-three weeks ago, I have been moving the claw to a new machine and haven't set up the reverse-proxy yet so I haven't been able to use the latest webUI.

I will in the next days try to get the web up again and will report back if I still have any issues

[willow narwhal]

What issue is this?

[crude matrix]

Main reason is that my agents stop mid work many times, I mentioned a few times the cases with links. so far I still have the issue.

[cinder bolt]

I prefer the web as it doesn't have the same limitations as discord for example. I like the idea that the content doesn't need to leave my server and not jump to discord servers in the US

[shell dawn]

hmm, i thought it was the message tool issue, theres a number of things that could fit that description though. peter took the idea i posted in this beacuse this was not quite right

Codex
"webchatDirectHint": "You are in a WebChat direct conversation. Your replies are automatically sent to this conversation.",
"currentUserRequest": "Delivery: to send a message, use the message tool.\n\n[operator request text]",
"perTurnDeliveryHint": "Delivery: to send a message, use the message tool."

i saw that cause all kinds of issues

prompts just got lost

[cinder bolt]

Yeah, I have that issue too, I'm getting this sometimes:

⚠️ Cron job "daily-job-search" failed: ⚠️ Agent couldn't generate a response. Note: some tool actions may have already been executed — please verify before retrying.

and sometimes I get that the agent stops, I'm trying to get a reproducing case so I can pinpoint reliably why it happens

[crude matrix]

Cool, this will help with best config for ones system.... You where busy with LCM stuff, could you look at this feature request? https://github.com/openclaw/openclaw/issues/84571 .... Some of my agents have been complaining about compaction and my dev wrote this feature request. .

[cinder bolt]

I'm running OmniRoute between OpenClaw and my LLM, it has this request log that makes it easier to track the conversations.

https://github.com/diegosouzapw/OmniRoute

[crude matrix]

issue that I am not able to choose a session in chat... workaround for me is choosing session via the sessions page.

[cinder bolt]

Real life calls, going away for a while, catch ya'll later

[shy rain]

What version ?

[shy rain]

Nowadays Openclaw starts reporting it as an error, whereas before it would just silently fail and we'd never know why.

[cinder bolt]

OpenClaw 2026.5.24-beta.2 (abb43c9)

[willow narwhal]

Have you seen this?

I think it only shows active session by webchat and channels though (need to confirm)

[shy rain]

That's the worst. I recommend you update to 5.26b2 or whatever is new by the time you can get round to it.

[cinder bolt]

alright, I'll try it out and report back. Thanks mate

[celest merlin]

openclaw@2026.5.26-beta.2 is out

[final ocean]

@tender loom new beta!

[shy rain]

And it's GOOD!

[shell dawn]

checking

[willow narwhal]

<@&1503801512294486217>

[shy rain]

openai models completing tool-heavy test bench in roughly the same time as codex cli...

only one hang. filed an issue. Just make sure your entire config uses openai-codex/ and not openai/ to avoid the bug. PR in the works to fix it anyway

We are so BACK

[willow narwhal]

Yes!

[vocal night]

Is there any use case to keep openai-codex still in refs or is it just a backward compatiability alias?

[vocal night]

Wait maybe I’m high, I thought we use openai NOT openai-codex @willow narwhal

[shy rain]

That's what I'm trynna figure out. It just failed because it saw openai-codex/ and openai/ as different providers. But openai-codex/ wasn't even in my config.

[shell dawn]

well if we are back to using openai-codex to use codex auth, that clears the openai = api except when using codex runtime then it means codex oauth

[crude matrix]

I think in the cases when it fails for me a automatic retry would work, if that does not work use failover llm and try again.

[shy rain]

Yes they homogenized the openai endpoint whether you're on oauth or PAYG

[vocal night]

Doctor has a bad habit of putting config back to openai-codex and breaking it but my understanding is it was deprecated a few versions ago

[shy rain]

It was doctor who put it to openai/ and then somehow switched it back.. and then back again..? idk what's going on I'm just trying to homogenise my own system and find out why I hit this bug, i.e. because of config or because of a real bug. I suspect real bug because my config was previously openai/ entirely

The config was openai/ everywhere. The gateway promotes that to openai-codex at runtime when it starts the codex binary. So the running session reports as openai-codex/gpt-5.5. Then the live-switch check resolves the persisted/config selection as openai/gpt-5.5 (because that's what's in config), compares against the running openai-codex/gpt-5.5, and fires.

  So: the runtime promotes openai → openai-codex, but the live-switch resolver doesn't know about that promotion. It compares the promoted runtime provider against the unpromoted config provider and sees a mismatch.

  The bug is still in hasDifferentLiveSessionModelSelection not using areRuntimeModelRefsEquivalent — but the root cause is clearer now.
  It's not a config problem at all. Your config was correct. The runtime legitimately changes the provider to openai-codex when it spawns the codex binary, and the live-switch comparison doesn't account for that.```

[crude matrix]

I think so, the issue is not not seeing correct session but opening it... so in your screenshot main session is on screen. If I would click on the session below (telegram:....) it does not load, main stays on screen.

[shy rain]

It's complicated. runtime "promotes" it, and then it fails.

[vocal night]

I call him doctor lector

[worldly tusk]

Why is there still "openai-codex" at all in the code?

[shy rain]

because runtime promotes it to that, for some reason.

[crude matrix]

These are the 2 issues I have identified. https://github.com/openclaw/openclaw/issues/84536#issuecomment-4515450826 and https://github.com/openclaw/openclaw/issues/84607#issuecomment-4524660698 ...

[shy rain]

but yeah... it's never been settled.

[worldly tusk]

But that is wrong. There have been so much effort to make it work as "openai" instead of the misleading "openai-codex"

[shy rain]

This is lulz...

  But ripping out the promotion touches auth profile resolution, compaction routing, context config, and the auth order lookup — all of which key off openai-codex as a provider ID. It's load-bearing in a lot of places. 

Basically it's a huge search/replace throughout the code, so instead they've just been switching between them behind the scenes. Hoping to settle on openai/ at some point.

[worldly tusk]

Yes this should be done with a global search and replace

[shy rain]

replace "openai" with "openai-codex" ...
... NO WAIT....

/ model openai-codex-codex/gpt-5.5

[worldly tusk]

The other way around

[willow narwhal]

It loads up well for me. Have you tried the latest beta?

[odd anchor]

https://tenor.com/view/thats-the-joke-ranier-wolfcastle-mc-bain-the-simpsons-stand-up-comedy-gif-4372917

[worldly tusk]

"openai-codex" should be replaced everywhere with "openai"

[vocal night]

That’s what I thought they were doing I don’t think openai-codex does anything anymore

[worldly tusk]

The oauth vs api distinction is irrelevant - other providers happily support both, so openai can too

[celest merlin]

Yeah we first have to remove pi to do that clean. In the works.

[shy rain]

it already does... the point is that lots of the internal code still uses openai-codex so it's a big refactor.

[celest merlin]

@thick osprey need a song to wake up

[shy rain]

instead it's aliased... and at one point the comparison saw them as different providers and failed.

PR cookin' right now

[thick osprey]

The generated music is ready.

[shy rain]

I've been using 26b2 for a couple of hours already. long before it was up in beta channel, I just happened to refresh the page. I'm so happy!

[vocal night]

Bro isn’t it 3am in sf! Go to bed code will still be here in the morning

We’ll corral these bugs 🤠

I’ll be landing in 3 hrs will test

[celest merlin]

I'm in London.

[crude matrix]

Yes, I am using 26.beta 2. I can share the console error I get...do I have your permission to send it in private message?

[cinder bolt]

Oh, the Queen's English, how's London treating you? Raining I ponder 🤔

[shy rain]

Issue: https://github.com/openclaw/openclaw/issues/87226

Live model switch falsely fires when the codex runtime promotes openai → openai-codex, killing active turns mid-execution with no user action. Root cause: provider identity is mutated to encode runtime choice.

Fix: https://github.com/openclaw/openclaw/pull/87252

Uses areRuntimeModelRefsEquivalent (which already knows openai-codex ↔ openai) instead of raw string comparison in hasDifferentLiveSessionModelSelection. Targeted fix — the deeper provider/runtime conflation is flagged for future refactor.

This is the only issue I hit today. (on 5.26b2)

[shy rain]

apparenly it's been a "scorcher" of a weekend, and still sunny.

[cinder bolt]

Strange times now indeed. We had sunny here in Sweden and then within 30 minutes, thunderstorms and hail rain. And back to sunny again after 30 mins

[willow narwhal]

Yes please

[shy rain]

That sounds like a normal British day in May. except potentially with snow as well.

[shell dawn]

2026.5.26-beta.2 is smoking fast

[shy rain]

yup. it's great.

[shell dawn]

and i even found a few more small speedups

rebasing the same one again lol with new surfaces

[vocal night]

Everyones been saying you moved to SF this weekend…. Well in that case lets get a release out today 🔥🫡

[shy rain]

genpop is begging for a stable release.

[odd anchor]

https://tenor.com/view/rtk-gif-23307679

[shy rain]

https://isitstable.com/openclaw

Tracks... THE PEOPLE NEED 5.26 (or 27)

[glad echo]

Could I bother a maintainer to have a look at this PR? I'm looking for a "might be interesting, keep it mergeable" or "not a chance, stop wasting time". The PR is about adding a /goal session continuation command. My version of "nice job, keep going please" https://github.com/openclaw/openclaw/pull/85723
it's "ready for maintainer look"

[final ocean]

The ministry summons the immediate release of the newly minted lobstar

[shadow garden]

speaking of unstable...github is being a a bit more than flaky right now -_-

[dusk vine]

Made the GitHub issue here: #87270

[vocal night]

“The Crab Lord sayeth! No more bugs! We shall no longer bow to crashes and brokenness. He cometh to heal all, disable all broken plugins, speed up the servers and redeem the lost claws”

[odd anchor]

https://giphy.com/gifs/BikeDiscount-karneval-bike-discount-bikediscount-65wfzcLCmxtSQAZORw

[final ocean]

the sea bearing custeans will bless you and will be there to heal and bring the strength from crayfish to lobsters, giving us a happy and prosperous release. Amen

[exotic tree]

Hello, for those of you experiencing connection issues with the gateway due to a service worker that is incompatible with a reverse proxy, here is my PR #87077