clawtributors

[worldly tusk]

Now if the people start complaining about increased token consumption in codex compared to the last stable version, we will conclude I was wrong about codex prompt assembly. I am sure I am right, though, and we will see reduction on average 😂

[clever narwhal]

Thanks to everyone who worked on speeding things up, beta no longer molasses

[left mango]

Arsenal supporter?

[obsidian zodiac]

Npm update errors on latest release. Both of my agents in 2 separate systems are reporting the latest release is broken and failing to update. 1 agent managed to auto rollback, the other wasnt so lucky.

"Update failed — v2026.5.26 has a broken build. 26 UNRESOLVED_IMPORT errors during the tsdown compilation step (missing modules like test-helpers/image-fixtures.js, diagnostic-llm-content.js, zod-schema.providers-googlechat.js, etc.). OpenClaw auto-rolled back, so everything is back to 2026.5.22 and running fine."

[pastel owl]

I think this is OOM related errors

I was getting similar at build time on my other 12gb vps

[obsidian zodiac]

128GB ram. not an OOM issue

[pastel owl]

Config restriction wont let use the available ram

tsdown has a config to restrict build time memory

I used to increase it manually in april builds for tsdown to finish building

[violet storm]

There is no such thing as "everything is back to 2026.5.22 and running fine." 2026.5.22 is the most broken version released this month 😄
if you want fine, go back to 05.07, before they introduced the V4 Gateway and destroyed the plugins

[obsidian zodiac]

I've been running fine on 5.22. and the beta before the main release.

[violet storm]

I guess you are OK with the slow Gateway and the useless chat context there. I am not though 🙂

[obsidian zodiac]

its been snappy as hell and running fine for me.

[shy rain]

2026.5.26b3 was absolutely fine. very fast and nice. My update was also totally smooth. 3 minutes.

[pastel owl]

Can you try with this

export NODE_OPTIONS="--max-old-space-size=6144"
export npm_config_workspace_concurrency=1

[shy rain]

There were some npm timeouts on parts I don't use though... :/

[shy rain]

This is the source build install right... I just did openclaw update for the regular pre-built.

[obsidian zodiac]

after running that, my agent became un-responsive Heading off to work now, so further debugging is going to have to wait.
OpenCode recommended i not run those commands as it will break any update.

[pastel owl]

Is the build done ?
You have to install after this config

In my case

pnpm install && pnpm build

[drowsy roost]

(deleted, nvm!)

[drowsy roost]

can we get a quick merge of this one?

https://github.com/openclaw/openclaw/pull/87313 (corrects docs so that codex computer use can be installed and continue to function rather than installing and immediately being disabled (or not installing at all)).

[drifting nymph]

good day peeps. i think this is not the right place to post the issues right?
upgraded to v5.22, and it is sending our openclaw into restart loop 😭

2026-05-27T14:21:28.312+00:00 [channels] failed to load persistedAuthState checker for whatsapp: plugin module path escapes plugin root or fails alias checks
2026-05-27T14:22:28.242+00:00 [channels] failed to load persistedAuthState checker for whatsapp: plugin module path escapes plugin root or fails alias checks
2026-05-27T14:23:28.233+00:00 [channels] failed to load persistedAuthState checker for whatsapp: plugin module path escapes plugin root or fails alias checks
2026-05-27T14:24:28.203+00:00 [channels] failed to load persistedAuthState checker for whatsapp: plugin module path escapes plugin root or fails alias checks
2026-05-27T14:25:28.191+00:00 [channels] failed to load persistedAuthState checker for whatsapp: plugin module path escapes plugin root or fails alias checks

[glad echo]

you can ask krill in #1459642797895319552

[zinc dock]

Man 5.26 is fast lol.....Didnt know how borked 5.22 was

[shy rain]

Hoping to get the 4.23 club with this one.

[shy rain]

to 5.22? 5.26 just came out and it is a MUCH better release.

[drifting nymph]

my team did a hot fix in whatsapp plugin to fix path-escape check, so all good.

[earnest jasper]

Still investigating this on my end, not sure if it's a me-problem or a 2026.5.26-problem

Fixed:
Native exec works now.,
Gateway is healthy and reachable.,
User service is clean: active/running, PID 1180, NRestarts=0, started 2026-05-27 15:04:24 UTC.,
openclaw gateway probe is OK and not degraded.,
System uptime and gateway uptime are both ~9 minutes, so the host reboot took.,
Task audit is clean: 0 warnings, 0 errors.,

Still not fixed:
openclaw.gateway dynamic tool with action=config.get still returns Tool output unavailable due to post-processing error.,
Fresh post-boot log confirms the same root symptom:discarded invalid tool result middleware output for gateway,

So the host reboot fixed the native hook relay issue, but the remaining gateway dynamic-tool post-processing bug persists. It still looks isolated to Codex tool-result middleware, likely bundled tokenjuice, not a general gateway outage.```

[true laurel]

I'm pretty sure we mostly spent time on performance optimizations for .26 😄

[earnest jasper]

---

What happened:
Native exec is wedged again: Native hook relay unavailable.```
I'll keep investigating and make an issue on github if it's a real problem. Probably a me-issue, wanted to bring it to your attention here though since a lot of stuff was touched in this update..

[zinc dock]

This performance is insane ...Probably the best yet..........this may become the new 4.23 lifeboat ...If future versions arent this fast im rolling back lol

[shy rain]

That's what I'm saying. This ought to be the standard for LTS release and I think I'll be staying on that for a while if there's degradation with the adventurous updates.

[left mango]

Ok, it looks like it's time to upgrade my raspberry pi from 4.21.

[drifting nymph]

i would love to but we cant, because in the past we had experience that latest release sent all instances into haywire, so we will usually wait one day then rolling out to internal instance first. then have to patch up the config files (usually config are not backward compatible). but yeah seems like its getting better and better, thank you to the team for your hard work!

[zinc dock]

I would go for it, im not just hyping it up this is the fastest its ever been for me ...also my multi sub agent spawns are actually reporting back automatically without me having to nudge my main agent

[worldly tusk]

whoever is interested in current codex prompt, @vocal night @celest merlin @final ocean@nocturne jungle , i did a bit more digging by asking my agent what it sees in the context. and doing some experiments.

1. currently (since my PR got merged a few days ago) SOUL.md, IDENTITY.md, and USER.md files are included in turn-based collaboration developer instructions. they are not duplicated every turn (as long as they are stable, and they are) - they are included once and then stay there regardless of the number of turns in a thread. so, this actually works how it is supposed to, and no fixes needed

2. currently MEMORY.md is left with the turn-based user context. and they ARE repeated in the prompt every single turn. the problem is that MEMORY.md is usually pretty large, and is not stable. this might be a problem waiting for a solution...

moving memory.md into turn-based developer instruction is not a solution - it would break prompt stability. or maybe we are happy with memory.md being repeated in the prompt?

[tall prairie]

Memory is supposed to be loaded once per session.

Same with all the other files

[worldly tusk]

so, where do we move memory.md? it should not be in thread developer instructions, as in that case it will be inherited by subagents

[tall prairie]

The memory file is one part of the equation. Memory is actually two sets of tools. Memory search and memory get.

[worldly tusk]

are you suggesting we do not include memory.md in the context at all?

and let the agent use tools for it?

[tall prairie]

Every other system like cron had a light context flag where none of the files get loaded in.

[tall prairie]

By default it should be, whether or not it should be propagated to subagents is still up in the air it seems.

[worldly tusk]

it is not propagated. but if "included only once" should be enforced, that may be the only solution. unless somebody knows some other clever way to differentiate. multi-agent-v2 may provide a way, but it is experimental...

[tall prairie]

I think newer models could handle an instruction to read memory files and treat them with high provenance. The verdict is still out if less capable models would reliably follow that instruction.

[worldly tusk]

maybe being less-capable is not an issue here, as we are talking about codex models? i bet 5.4-mini is perfectly capable of that

[tall prairie]

Yes… we support many harnesses and models big and small.

[worldly tusk]

pi harness is a separate topic. here i am talking specifically about codex

there are no problems that i know of, with the core md file injection in pi harness

[tall prairie]

Well #2, if valid and I am understanding it correctly, is a major issue and probably blowing up the context window.

[agile kiln]

I'm going to push a clarifying comment wrt your PR comment too per the confusion over --source vs --marketplace-path - sg? Also, I'll keep your note but remove the 'enabled: true' and mention that status can report disabled if you one-off install without adding the config.

[tall prairie]

@worldly tusk thanks for the heads up. It’s being looked into!

[worldly tusk]

thanks, i will be waiting for any updates 🙂

[icy thorn]

What is the best way to ask reviewers a question about a PR ?

[crude matrix]

I understand why innitially the choice was made to use md files. the more interactions you have with your agent(s) the larger the files become, the harder it is to load them.... vector databases should be the solution. much quicker finding data and maybe 1x entry for everything... qmd, files and wiki would be best.

[worldly tusk]

you can only find something if you know what you are looking for 🙂 that's why it is common even in tool-based systems, to still have memory.md file to represent a high-level overview of what you have. so that you know what to look for

[vocal night]

The problem with bolting on memory md is we should use codexes experimental memory

Its not very good but its better than memory md

I’ll put a pr in when I get to my hotel to stop the madness over these lol

At the end of the day have to manage the assembly and make sure assembly works and doesn’t break cache

[worldly tusk]

codex currently does not seem to break cache. it just breaks its context windows 😄

[worldly tusk]

this one? https://developers.openai.com/codex/memories "Memories are off by default and aren’t available in the European Economic Area, the United Kingdom, or Switzerland at launch."

i am from europe, thank you 😄

[wicked mason]

That seems outdated - i have memories enabled in the codex app and codex consults them constantly. (Living in Europe)

[vocal night]

Its basically a codex version of md

This is where I point out we’re building this parachute after jumping out

Of the plane lol

[worldly tusk]

i am using basic-memory for this. it stores memory in local md files, has access tools, works with local embeddings for search, etc

it is not a replacement for memory.md, as we need a high-level overview of what knowledge there is in the system. otherwise we don't know what to look for. sure, i can delete memory.md and just ask the model to read some memory file on startup (as it is done in the default openclaw agets.md - to read memory/memory-yyyy-mm-dd.md files, or whatever they are called, not using them) - that is a potential solution... but i have no current opinion of whether it is the best solution

[vocal night]

Thats what experimental memory does

And if you’re doing that you should use gbrain thats where it works best

[worldly tusk]

i have a bit of mistrust towards black boxes. that's why i like basic memory, and still not installed lcm 😄

[vocal night]

Okay down the rabbit hole so developer instructions work for the files not because of assembly or priority but because codex compaction is told not to compact or summarize developer instructions

Its a hack in the compaction to keep them versus compact them

I’m wondering if it made to our compaction as our prompt is just like summarize dat shit lol

[vocal night]

Thats what I missed @worldly tusk on codex developer instructions

[pastel owl]

Hi maintainers,

Today I faced an issue while creating a new topic in telegram
The issue was plugin state db had a 1000 row limit including message cache etc. And it wasn't being pruned.
So I have created a small PR to fix this

https://github.com/openclaw/openclaw/pull/87254

[vocal night]

They get around reinjection theoretically by not compacting that part of the assembly

[worldly tusk]

the least-bad solution is to just cache memory.md file for the duration of the codex thread, and inject it in turn-based developer instructions. that way it will stay stable, not duplicate itself every turn, and not be inherited in subagents. nothing else comes to mind currently

[timber beacon]

@lusty vigil & @lunar granite https://github.com/openclaw/openclaw/pull/87367 here ya go! I am letting claw review but please ping me or yell at me for any changes

[crude matrix]

I am not sure that is really needed. I think those systems can provide an index and subjects, like what we have in books, with that the agents are able to hone their search.

[worldly tusk]

i am probably blind, but i do not see any differences between light and dark modes 😄

[timber beacon]

shhhh but yea

There is only one mode real people use anyways

[worldly tusk]

true

[timber beacon]

there light mode updated

[zenith prawn]

Same, even more: after update 5.22->5.26 on Prod. Commented on your PR with details.

[zenith prawn]

can't verify that it fixes yet; but if it does, it should be amended into release ASAP; otherwise we shipped OC breaks Telegram Group Threads completely - might be already user sentiment 'again'... its like beta... where is LTS release... anyway, hope some Maintainer can fix it now... I personally had to revert to 5.22 on my Prod in 1:23am localtime, otherwise tomorrow my staff can't use OC...

[pastel owl]

The temp fix is you can ask your dm claw to delete 200 old rows from pluginstate db
That's how i was able to make new topics again

[zenith prawn]

it worked. just: delete 200 old rows from pluginstate db
Remaining rows: 800 ... Should this command be included into doctor --fix with this release ?... Not every user have access to this chat... i wonder how such bugs make their way into release... maybe its only with users who use Threads a lot?... Well it shouldn't be rare, how to include this use case / db state mock into the tests?... so that each bug we encounter do not repeat again in various forms...

[pastel owl]

the fix is likely to be merged soon

Peter is polishing the PR ✌🏻

[zenith prawn]

btw after 'delete 200 rows' old Threads in Groups still non-functional

[pastel owl]

Are you talking about discord ?

I am using topics in telegram

[zenith prawn]

same, telegram

[pastel owl]

May be /new

[zenith prawn]

unfortunately, nope. Same Something went wrong while processing your message. Please try again.. Only new threads work.
Thx god i made backup ~/.openclaw to revert

[zinc dock]

5.26 is a boss but this Native Hook Relay bug is a party pooper ....Already being tackled tho https://github.com/openclaw/openclaw/issues/87331

Actually fix was just merged

[timber beacon]

Mantis is great, I took it to work on how to make my iOS stuff do the same thing

Also codex is so fast on this latest update!

[smoky dock]

thanks @trim vector for reviewing and landing #82431 🙏 there are just 2 left in the same AgentKit/plugin HITL slice when you have a chance

https://github.com/openclaw/openclaw/pull/82434
https://github.com/openclaw/openclaw/pull/82471

these are the remaining pieces for proof-backed approval resolution + durable approval/status cards from the external clawhub plugin

thanks again ✌️

[pastel owl]

⚠️ Agent failed before reply: Unknown model: openai/gpt-5.5. Found agents.defaults.models["openai/gpt-5.5"], but no matching models.providers["openai"].models[] entry. Add { "id": "gpt-5.5" } to models.providers["openai"].models[] to register this provider model. Please try again, or use /new to start a fresh session.
After the pi refactor

[pastel owl]

@celest merlin i had to roll back to Commit 99b27cde64d6616a9e41f52f4a699577cf60f1d6
And oc is back to life

Even re-auth and forcing codex as runtime didn't work

It kept defaulting to "openclaw agent" harness

It's a showstopper

[vocal night]

Yeah, the purpose of Emory is actually to be searchable. Based on this with the Agent needs not to be loaded every return.

Memory

[worldly tusk]

you are too late, the memory is removed 😄 @celest merlin himself made this commit: https://github.com/openclaw/openclaw/commit/d93524d1ccb116768b4621e11b9ce756a991de8c

so, now we won't have memory.md repeated every turn, as there won't be memory.md file in the context at all - the agent can just read/search it with tools. within the current codex limitations, this might be the only sane solution.

now, on the other hand, i commented on Nik's PR (https://github.com/openclaw/openclaw/pull/87155) that memory is repeated every turn. he didn't answer any other questions, but recognized this one. and would you imagine that, made his PR even worse: his solution is to inject memory.md into thread developer instructions 😄

[crude matrix]

I understand the memory content will not be sent over and over again, but even if its in the context of the agent, most agents do not have the beginning of the context as sharp in their reasoning as the last part of the content, and wha tif the memory file gets updated. would it not be best that the memory data is removed and readded at the end again??

[worldly tusk]

there is no way to do that on codex. that is the problem. stuff gets ONLY added to the prompt. so, if something gets updated, there is a single binary choice: either you are fine with the old version being in the context, or you add the new version to the context. and user context get added regardless - just on every turn

now, it would be great to have a content eviction option. that would break cache

[shell dawn]

ive been going down this path for a bit, i feel like i had it working quite well on pi, still trying to get it up to snuff on codex

[worldly tusk]

alternatively, one could recreate the thread from scratch. also breaking cache

[shell dawn]

its not as natural of a fit

but the idea of trimming your context window carefully as to not break cache, keep your window small enough to prevent you from hitting compaction, was something i thought i had working quite well on pi. i had nosiy compaction logs to know when it happened, and i could go dozens and dozens and dozens of turns without hitting compaction. large tool call results naturally blew things up

i was mitigating large tool call blast radius that as much as i could, there are probably a number of ideas to take that area further

adapting to codex runtime and fixing/dealing with issues from codex runtime have been slow going with my irl work schedule, but now that codex is looking more stable and my agents are becoming more productive, i hope to have some good experiences there that might be worht discussing further

[worldly tusk]

ah, don't be so sure about stable codex 🙂 multi-agent-v2 is coming. still experimental. when it is ready, i bet we will be migrating

that should be good, though

[shell dawn]

haha, good point. stable for me was more of a today vs tomorrow thing, 2026.5.26 is in WAY better shape than 2026.5.22

besides, if we already have experience doing pi and codex runtime, whats yet another one? 🙂

[worldly tusk]

what are you doing with the mitigation here? smartly removing old tool output?

[shell dawn]

kind of a step to the right, head+tail compression with a note of what happened and where to get it. slotting the composition window into different purposes, grouping them together into areas that need to (and naturally do) stay the same for caching benefits, vs where the window changes and the resulting supportive memories

[vocal night]

I’m still on 19 😮‍💨🤣

[shell dawn]

jump on in the waters fine!

[vocal night]

I test versions on my test gateway but can’t afford to have our customer fleet break

Need to not have to debug

19 is fine even if its slow bugs are not game breaking

Also everyone I know has said after a few hours this version breaks

I’m off tomorrow speaking at a conf so I’ll let yall fix 🫡

[shell dawn]

hm, ive been tied up in meetings all day, whats breaking in this one? ive been planning and analyzing in openclaw today and its been pretty happy

[vocal night]

If I knew would have said it seems to be after a few hours telegram breaks

[shell dawn]

i did get some of those hook errors that i saw someone mention, but when i looked at it, there were some issues with my large tool call mitigation, which was an area i already had roadmapped so i told my agent to get back to it and reframe it with those issues in mind and its been moving along

[vocal night]

And 2 otbers else told me different other platforms

[shell dawn]

ah, i use webchat to fit all of my particular things i want

well, custom chat but based on webchat

[vocal night]

Is webchat better now did val make cutoff

Or shakker for having more than 2 messages in history

[shell dawn]

ive remade my whole chat based on what webchat exposes, so im not terribly familiar with the specifics of the native webchat experience

[void bramble]

I’m holding out for my prs that got merged to be included before I leave my current version 😛

[shell dawn]

bah, just rebase them!

you might fall behind in how they need to get fixed 😄

[void bramble]

Actually completing agent spawns. And having thinking stream by in discord make life worth living 😛

[void bramble]

I probably will tonight. Gotta make some time. It’s been actually working and been making me money. I don’t want to hack on it yet again… and again…. Every upgrade has been so painful for me lately.

[worldly tusk]

did you fix the missing commentary blocks? 😮

i should update to main then...

[void bramble]

This one is still waiting on something. I haven’t had time to look. It’s super badass when it works tho. Makes OC seem 10x faster when you can actually see it thinking again

https://github.com/openclaw/openclaw/pull/85200

Need maintainers to decide: Suggested next action: If you want a full design/correctness review of whether this option should exist versus reusing an existing setting

[tall prairie]

You need Peter’s sign off on that because it’s adding a new config option. There’s a hard freeze on these types of PRs atm.

(I do like that pr personally)

[true laurel]

Omg I'm merging this locally right away

[short dome]

No config options? Sounds like default it is. 😎

[tall prairie]

If you guys want that pr the do some sort of reaction to the pr or something.

I’ll pass it up the chain

[true laurel]

I need to test it first before I approve, but on paper this is exactly what I've been looking to get back, I was so sad when this feature disappeared

I wouldn't make a new config key for it though, just use one of the existing config options tbh, there's enough of them

[void bramble]

And also test out the visual aspect of it. I just made that shit up 🤷‍♂️ I though italics looked nice, because it didn't seem 'done', and was easier to tell when the standard text plopped in on the completed edit/message.

And I removed the default bullet points, but that's personal preference. But it's my PR, so I guess I get to set the first draft 😂

[void bramble]

I'm still scared back here on 2026.5.19... how is 5.26? Seems like people aren't complaining too much? I've been out of the loop for a few days.

[void bramble]

About the discord streaming commentary, enabled by mode: "progress", while toolProgress: false would only suppress tool/task rows. That is viable, but it changes behavior for existing users who set toolProgress: false expecting a quiet label-only progress draft.

Other existing options are weaker fits:

• streaming.progress.commandText: "status" can hide raw command details, but tool rows still appear.
• streaming.progress.label: false only hides the title.
• streaming.progress.maxLines / maxLineChars only control layout limits.
• streaming.progress.render controls presentation style, not which event classes are shown.
• streaming.mode: "partial" is the wrong delivery model because the requested UX is transient progress, not final-answer preview text.

My take: if maintainers want strict backward compatibility and independent controls, streaming.progress.commentary is the cleanest contract. If they want fewer config keys and can accept a semantic adjustment, redefine progress.toolProgress as “tools only” and make commentary part of mode: "progress" by default.

[zinc dock]

Native hook relay issues popped up for me ...have to constantly reset session to get it working again ...Peter merged a fix but some commented it's still a issue

https://github.com/openclaw/openclaw/issues/87331

Anxiously awaiting a new beta for this fix .....other then this 5.26 is the more performant release yet..........super fast

Like really really fast lol

[void bramble]

stoked to get on it. I'll try to update tonight and rebase my few remining gripes/PRs

[zinc dock]

Yeah whoever was in here complaining about open claw versus codex cli speed, they may want to give this a shot lol

[void bramble]

This has so many comments. It's gotta be whacked/main'd asap.

[zinc dock]

Yeah a lot of people are affected

Resetting the session fixes it Everytime but it just get back out of whack over time ....not ideal.....

Hotfix territory

[fallen gate]

CC: @trim vector I think this sort of thing is best moved to something like a global config for all channels kinda like how we were discussing the approval exec flow? less channel specific configs to maintain?

[timber beacon]

I love that. Going to config is like needing to speak another language

[tall prairie]

Maybe it can be a flavor of /verbose

[fallen gate]

yes

[lone iris]

this!!

[trim vector]

hmm, does this apply to all channels or just those with concepts of drafts? and yes, less channel configs. or if it is a channel config, it should be consistent across channels

[shy rain]

It’s the best. Not sure about the streaming behaviour but whatever you do don’t update to anything 5.20-onwards except the latest inc.5.26

[shy rain]

Unfortunately true but not strictly necessary if you make your clanker read the docs and suggest settings that get the behaviour you want. Then it can implement them. But it MUST read the docs and have them in context.

[shy rain]

Yes. There is some overhead in chat but once it starts a task the speed is comparable.

[rocky pewter]

I set up an agent and point him on the documentation and GitHub only, and that’s been really helpful. Whenever I need something from the docs or GitHub, I just ask him, since its context is limited to those sources, the answers stay very relevant.

[shy rain]

Yeah that’s what the skills are for but I find keeping the context more valuable than note taking each time. I have a permanently open OpenCode with deepseek who is 100% my network engineer. It’s much less hassle. I should go multi agent. But also multi-harness integration.

[timber beacon]

Yea but for a UX and user support it helps make it clean and streamlined for users who want to read settings themselves.

[shy rain]

If only the schema didn’t change every couple of days

[olive cliff]

hey, who’s reviewing docs PRs these days?

[trim vector]

this one 🙂

[trim vector]

if you have something specific, plz send my way

[obsidian zodiac]

Got an update as to what caused my Ubuntu OpenClaw to not be responsive after the update:
Root cause: The v2026.5.26 update removed the standalone codex binary, but the config still had "command": "Codex: telling the plugin to look for it on PATH (spawn codex ENOENT).
The Fix: Removed the explicit command field from the codex plugin config. The plugin docs say: "Leave unset to use OpenClaw's managed Codex binary" - it now auto-manages its own binary internally.

But this doesn't explain why my OpenClaw Agent on Fedora reported that the new update was borked causing it to auto-rollback. Its not using Codex at all. I'll look deeper into it now that I'm off work. It could have been a fluke for all i know, so I will attempt it again here shortly.

[long cradle]

I have to ask a general question because I have just been learning more about the topic lately and there is so much to transformer architecture haha

But would a reduction on kv cache by 8.33x compared to fp16 be meaningful for a new method of compression?

It’s a general question I am researching a theoretical concept right now and not sure it’s worth the time I got that runtime number already for the first iteration

[zinc dock]

Seems like clawsweeper fell asleep on this one.....More Hook relay regression ....https://github.com/openclaw/openclaw/issues/87395

[smoky dock]

@trim vector why was the PR reverted? 🤔 https://github.com/openclaw/openclaw/pull/87419

[trim vector]

sent you a dm - the pr wasn't supposed to have been merged as it needed a broader discussion first since it expands our plugin sdk. this was a mistake on my part

[rocky pewter]

Hi Maintainers, hope you're doing well. I have a small bug/stability PR open for the current main / current beta line, and I hope it fits the current roadmap.

If you have time to take a look 🫡

https://github.com/openclaw/openclaw/pull/87457
(Nostr / DM subscription lifecycle)
Fixes a strict-relay Nostr DM startup/restart issue. It keeps the Nostr account running until abort and cleans up the relay connection on shutdown.

[lusty vigil]

hola folks... new beta is aliieeevveee 🦞
https://github.com/openclaw/openclaw/releases/tag/v2026.5.27-beta.1

[shy rain]

Here I was thinking to myself "wow I have a fully working and super responsive Claw for the first time in over a month. What projects should I get back to...

[left mango]

Hi, I created this PR based on someone's bug report.

[vocal night]

I already globally moved everything to unified contracts vince has been merging it in so plz wait until there is global messaging feature contracts

[tall prairie]

regarding the beta, post your feedback here but also trying something where we put feedback in one place https://github.com/openclaw/openclaw/issues/87486

[fallen gate]

@untold pollen do the dew

[tall prairie]

there is some guidance in the issue what to test based on what changed in the codebase

[untold pollen]

<@&1503801512294486217>

[vocal night]

Not seeing a changelog for the beta when i click it

[rocky pewter]

click this
#clawtributors message

[tall prairie]

yeah the page isn't published because people with permissions to publish it are briefly away from keyboard

[vocal night]

Ah okay is my global msg contracts consolidation of features in this one?

“One ring to rule all the msg features” lol

[tall prairie]

that is a good question and i do not know the answer to that 🙂

[vocal night]

Sounds like it from changelog

Actually no just global approval in channels?

Its been tough to merge it because maintainers keep merging other changes into individual channels > need to freeze new channel features > unify them and THEN rebase new features onto global channel contract

[crystal relic]

https://github.com/openclaw/openclaw/pull/87476 ready for review size: M, rating: platinum hermit
Fixes session replies using stale WebChat routing so external channel conversations get the response.

[tall prairie]

ty - that's good advice and we're taking it under advisement

[zenith prawn]

Looks like our betas are essentially alphas and our releases are essentially betas

[vocal night]

https://github.com/openclaw/openclaw/pull/86165 heres series he’s been working on

[fallen gate]

the road to stableclaw is filled with bumps and bruises, but we will get there!

[timber beacon]

Just keep coming out with hits! It’s funny people go crazy with fast release but then once a few days with no release goes they ask if claw is dead… cant win with everyone. I gotta say these past few releases have been some of the best improvements in my setup

[fallen gate]

this was basically what I pitched lol

but not the whole fucking thing at once

[vocal night]

He’s spent days with it on /goal because every new merge on individual channel makes it harder to unify 😮‍💨 had it perfect diamonds a few days ago but then one maintainer added something then another lmao

This has been in works for 2 weeks on my end but have to freeze individual “this looks cool lets add it to this” and it adds to the spagetti

And yeah I did it in phases so it could be done in phases its just not possible at this point, too many with write access to phase it

[zenith prawn]

Maybe assigning ownership for parts of the system to certain maintainers would help...

[vocal night]

Got individual maintainers like hmm lemme add this to discord spagetti real quick 😮‍💨🤣

But it also touches telegram, the messages system is a mess thats why last weeks releases were so bad performance wise tbf

[zenith prawn]

I also have a PR like this with automatic fast mode. Tired of re-basing it... and I'm highly ready to wait until proper architecture is established because it benefits all..

[vocal night]

I’ve slowly been refactoring openclaw to unified contracts wherever i can

I refactored the engine already

I’m well over 100k loc merged

Unified sdk, unified runtime, unified message features, too much “individual” support

When I audited regressions its like 90% of our regressions

And it makes sense how ai works, “okay cool I don’t wanna break abc so I’ll add a new seam for feishu….. 20 mins later but lemme just wire it into discord to save time ” 🤣

[fallen gate]

you were taught that it's not the size that matters right?

[vocal night]

Eva’s a girl AI code bot and she cares about how much code she puts in and how much spagetti she throws out tbf 🤣 🫢

She actually keeps a running ranking system of maintainers that she uses to track bugs

“X maintainer merged something based on past code it’s probably slop, should I review and submit a patch PR before it break ABC feature?” She’s like if molty could have grudges 🤣

[fallen gate]

welp I'm cooked then, cya boys

[vocal night]

Don’t worry its usually security and codex that don’t /auto review

“On Final review, 10/10 maintainer merge-ability score is unattainable without 95% confident on: Is this the right way to architect feature? Or should we refactor into immediate blast radius and consolidate”

Is the game changer tbf if maintainers want to cut spaghetti out and consolidate

We talk a lot about bugs but not wins. Agent is such a boon traveling, last night took care of everything for when I landed and ordered me food to hotel room by the time I got there ready, and planned out my day today. All while managing project threads.

Hadn’t really used for travel so took some advice on it and total game changer

[lone iris]

Mine has been helping me figure out trips to take. I definitely am going to be putting my agents to the test during my convention trip in late June... this gave me great ideas on how to put my OpenClaw to the test.

[vocal night]

Also she turned off my access to remote control view projects on claude and codex 🤣 because she’s also designed to make sure I go to bed on time and make appointments

I’d say these lil wins are why I’m so contrarian on anything that pushes oc to coding agent only vs personal assistant 😮‍💨

[lone iris]

I've been teaching mine about how to use the work units system I made. Definitely going to be useful for load balancing over a travel week

[vocal night]

Peters got some ideas and I do too and so does Him and a few others on how we can after stability make claws able to natively make themselves more productive hope you’ll contribute

[lone iris]

I definitely want to! Just been really intimidated about the size of the codebase and trying to navigate how I can help!!

[vocal night]

Take a small piece and own that part

Thats only way to do it

[lone iris]

I've been wanting to take on looking at the memory/dreaming

[weak walrus]

I just installed my 5th DGX Spark today, and man the hate I got in the OpenClaw communities when I told them I did a live demo of OpenClaw vs Hermes, and hermes got stuck on install and couldn't call-back on hostinger VPS for the grok oauth, lol 🙂 thanks to the likes of @fallen gate @untold pollen and @nocturne jungle racing me to work on that grok oauth device PR we got that sorted in OpenClaw so good 🙂

But man, the hate for OpenClaw is real, and I'm just letting ya'll know the Hermes crew is bitter and taking over our openclaw communities

[vocal night]

I try not to go outside core engine, prompts and harness. I was just so annoyed with messaging channels breaking I refactored it too 🤣

[weak walrus]

I think you too @vocal night ? lol

[timber beacon]

I want a spark or more for my claw

[vocal night]

I don’t wanna bring up that discussion and the mods won’t ban the word even though I suggest it every week, its not us vs them its like an ex. They are deservedly bitter we failed them.

We’re working hard to be better and it takes time, when everything is right they will come back

But right now we’re the abusive alcoholic who breaks furniture and promises every week we’ll be better 🤣. Not yet done w rehab to be a clean “run biz on this release” release

[weak walrus]

Well... I mean. they straight up steal our stuff, I get it's opensource, but pretty funny

[fallen gate]

hehehe, that was thanks to you for triggering my competitive side to get that done

[weak walrus]

LOL, well it is a pretty big pain in the ass to do callbacks on vps, especially since if you got a commercial vps they block bots

hell X doesn't even let you log in the platform on hostinger IP

PS if you ever want to see the hate:
https://www.facebook.com/groups/1577315533418837/posts/1678863146597408/?comment_id=1679020583248331&notif_id=1779919246338526&notif_t=group_comment

[vocal night]

Nah please delete that and keep negativity out of here

[weak walrus]

I think it's pretty valuable to see what people say, you can learn a lot from the feedback, their feelings are valid, as much as I want to smack em

[fallen gate]

wow savages lol

[weak walrus]

I mean, in his defense, I am a troll, and I purposely make my posts polarizing to go viral lol

[fallen gate]

I thought X was where the feral hung out

[lone iris]

lol you do a good job at it too

nope Facebook is where feral people hang out!!

[weak walrus]

NAH NAH it's THREADS LOL

[lone iris]

THREADS LOOOOOL

I haven't been on Threads since Artemis II because I miss the Moon Joy threads people!!

[vocal night]

Gonna re orient us back to the current beta

[weak walrus]

threads is where everyone who went to bluesky after Elon takeover of twitter and realized no one else was there, so they reluctantly fell back to threads LOL

[vocal night]

This discussion is nice but can move to general please, I distracted w a quick win and positivity, but we want to get stable out

[fallen gate]

TBH I didn't even know we had facebook groups

[weak walrus]

OK on to the beta 🙂

[vocal night]

hola folks... new beta is aliieeevveee 🦞
https://github.com/openclaw/openclaw/releases/tag/v2026.5.27-beta.1
we need testing for rapid hot fixes

[fallen gate]

I only log into facebook to use marketplace and get abused by scammers

[weak walrus]

There's a ton. I have an openclaw community with 50K, and one of my students has a 500,000 openclaw community (the one I linked prior)

[vocal night]

…. @weak walrus 😮‍💨

[lone iris]

Bringing back down the feedback thing here too since it has the list of tests https://github.com/openclaw/openclaw/issues/87486

Mostly for myself but in case anyone else has a memory of a raccoon

[vocal night]

Thanks even I can distract but right now we’re all aligned to performance and stable release candidate

[weak walrus]

I just updated my main claw (Beau) and he's been great. The web gateway seem super snappy now

[untold pollen]

Going through some Telegram PRs/issues, please feel free to send any my way

[vocal night]

hola folks... new beta is aliieeevveee 🦞

https://github.com/openclaw/openclaw/releases/tag/v2026.5.27-beta.1

we need testing for rapid hot fixes for stable release

Please post issues here
https://github.com/openclaw/openclaw/issues/87486

lets try to keep focus for now on the release and high signal low noise to keep channel readable for maintainers and move rest to general

[weak walrus]

Deleted 😛

[weak walrus]

haha already?

[lone iris]

Not beta related, but @ maintainers: I'd like to attempt fixing this issue. Any direction on implementation? Since Clawsweeper says maintainers should get involved on this one https://github.com/openclaw/openclaw/issues/87485

[weak walrus]

welp, i've never used the feature, so pass 😛

Right now I've been experiencing issues with Grok doing what GPT 5.4 used to do (I'll do that, and report back) and never report back...

Working through what we did previously to fix and remember how it was resolved

[lone iris]

I have the same issue on Deepseek V4 Flash, but mostly because the dumb models don't use message tool

[fallen gate]

that's a vincent thing I think, probably not gonna be prioritized atm ngl

[lone iris]

fair enough... just trying to figure out something small to get my feet wet before trying anything big and that seemed pretty tiny. lol

[weak walrus]

The PR I really need to do, is way over my head so if someone else can do it or give me an idea how to do it.

I hate that every time I switch models or give it instructions to switch models and then continue, it just stops cold. There has to be a way for it to have a "queue" that triggers after it restarts gateway

@burnt swallow cc

[vocal night]

Model fallback is automatic, and continues in fallback. It’s annoying feature and leads some to turn off fallbacks due to degradation during it.

Degradation happens when switching models even in same series and cc and codex will warn you of this.

In OC manually switching models means new cache etc and fresh run with previous prompt.

[vocal night]

So I’m confused on what you’re asking it sounds like you want boot.md config on which exist and you can use to continue after editing models post restart? Not actually switching models.

That is currently degraded due to legacy prompts for boot. should by default be “resume/continue” but often gateway treats it as new even if theres prior session context.

Tough problem to fix. I have my agents via agent md set a cron to reactivate and nudge “continue” 5 mins from when they sent restart command it works better

[weak walrus]

example:
I give it a command to update fallback order, and restart the gateway, then:

When you're done, tell me which models actually show in your fallback chain so I can confirm the rename took.

it seems like it'd be a simple "resume" after restart, or auto-checkpoint that loads on restart

but after restart... it's just never picking up

[fallen gate]

I swear me and @shadow garden have impemented and reimpemented this feature like 5x already

it just keeps getting mowed over

[pale ravine]

There are still many issues with this conversation. My agent keeps telling me that the Native hook relay unavailable problem prevents it from reading files. Although /new works, the same issue reappears quickly.
I have reported the issue to GitHub.

[zinc dock]

Bummer. I was hoping this was fixed in this new beta

[fallen gate]

@opaque sand isn't this fixed?

[zinc dock]

Peter merged a fix but then someone commented that it was still broken. I don't know if it got revisited

[pale ravine]

I don’t know, I keep running into this problem repeatedly. My tasks keep getting interrupted, and it’s really frustrating

This is the first time I've encountered this issue after the update. I didn't have it before, possibly because previous problems were masking it

[zinc dock]

This is what was merged but seems to be still happening https://github.com/openclaw/openclaw/pull/87272

Original issue that is now closed but with new comments about post fix regression https://github.com/openclaw/openclaw/issues/87331

[void bramble]

ok. 5.26... feels like 5.19... which is great! 😛

[zinc dock]

Speed is phenomenal. Just that damn native hook issue

[void bramble]

I haven't hit that yet. Im sure I'll get there. Getting the discord streaming commentary was a chore, but more of a chore for Codex, I just sit and watch... confused and nodding like I know what he's talking about.

[shell dawn]

but it gives you a great vibe

[true laurel]

It's all about the vibes 😜

[worldly tusk]

On codex, the new beta should experience less compactions, and spend less tokens, as memory.md file is not duplicated on each turn.

[true laurel]

Oh that's great

[void bramble]

Honestly, I've been spending so much time in the codex app, my OC has been a lot more idol lately. Especially now that they have the remote connections stuff in Codex

[true laurel]

I'll do my update dance when I get to the office

[shell dawn]

which version?

[worldly tusk]

5.27-beta1

[zinc dock]

Are you on 5.26 or the 5.27 beta?

[vocal night]

It needs consolidation across runtimes. It was working beautifully a few weeks ago

We should probably place a rule @fallen gate to not post in channel “i have issue” without a version number i try to triage and its always a headache

7/10 times the persons on like 5.22 lol

Not sure if we can have krill or molty dm people or only viewable on their end give them feedback “hey you posted low signal high-noise message please edit to include version number you’re on and a link ideally to github issue and what the ideal fix or experience would look like”

[pale ravine]

5.26

[zinc dock]

Phew. Okay yeah, the beta is supposed to have a fix for this

[vocal night]

Exactly the convo I have 5 times a day haha

[zinc dock]

Lol

[pale ravine]

Thank you, looking forward to the next update resolving this issue

[pale ravine]

He might need an MVP channel that can directly connect to the database, allowing his claws to obtain the information he wants

[opaque sand]

looking

[zinc dock]

In the meantime to mitigate it, reset your current session and that error will go away for some time

[pale ravine]

I did exactly that, and then the problem quickly reappeared

[zinc dock]

Yeah sadly that's the only thing you can do right now until it lands

[vocal night]

Resetting session will be a common theme in new releases for fixing threaded issues due to how codex works it will almost become a mandate for some major releases

Similar to codex app itself needing new session + restart

[zinc dock]

I'm trying to get in the habit of doing it after every update

[vocal night]

Just how its built, working around it created weird regressions

If you use lcm you can rotate to new session in oc without losing agent brain and context

Lcm less oc on a /new is like talking to new dumb 20 yo employee

I could never go back 🤣

[pale ravine]

Resetting a conversation when encountering a problem is a quick way to handle it, but it becomes troublesome when I am in the middle of a task. Although it can review data from previous conversations, I worry that some important information may be missed

[opaque sand]

Agree fix should be in 2026.5.27-beta.1
If anyone is experiencing the native hooks relay issue in the beta let me know

[vocal night]

Thats where LCM rotate function comes in, we should probably build that into oc its a few lines of code

Rotates to new thread and session and keeps previous session prompt rollover

I have never met a user that uses /new or /reset

[pale ravine]

Yes, but I've already removed it from my list. I'll add it back once it's integrated

[vocal night]

Its just not how oc should work

[pale ravine]

Considering context pollution, /new is still necessary

[vocal night]

Context pollution doesn’t happen with context management

[zinc dock]

I prefer not to do it. I only do it when things get wonky

[vocal night]

We now have focus modes that allow you to hot swap context so you can save where you are and come back later

Its not been heavily promoted or known but josh rolled it out 2 wk ago

Probably one of the most innovative features in ai right now

[zinc dock]

Maybe when we could get to an LTS version of openclaw, I'll try LCM again

[vocal night]

That and our tool stub i have companies and enterprise copying it now everywhere

[pale ravine]

I want to understand how it works
This is different from what I understood

[vocal night]

Byte dance just messaged me they are coping it into their systems

[vocal night]

Focus mode? Or tool stub?

Focus mode saves in db current session so you can come back to it and swap in and back its like video game saving/loading between different focused without context bloat

Or pollution

[pale ravine]

I don't know. Translation tools can only express a limited amount of meaning. What I want to know is how to avoid context contamination without using /new. I'm not referring to using sub-agents to execute tasks, but rather how to communicate and understand your needs within the main conversation and with the main Agent.

[vocal night]

Its the difference of treating prompt and context like a big wall of text mess and more like ram + short term memory blocks

[pale ravine]

But switching always requires a command, doesn't it?

[vocal night]

Session state and prompt and context is just a state and context, if its organized already in controlled managed assembly blocks like lcm vs just a wall of text history then it can be managed like memory

I’ve been building that for months

[vocal night]

Yeah of course

Agents managing their own context blocks is the future for sure

But for now it needs you to command it

[pale ravine]

My understanding is that you're saying to break a long string of context into chunks, and then assemble them into a complete context based on current needs

[vocal night]

I’m saying level 0 I work at openai codex front end vibe engineer 🤣 is cs 101 place context into files and tell agent to “prioritize”

Level 100 clash of clans / mob boss bs game ad 🤣🤣 is agents having dedicated blocks in their context that they can swap and manage on their own.

[pale ravine]

Well, my knowledge level isn't high enough to understand what you're saying. Could you provide some keywords so I can search online myself?

[vocal night]

The best I can do to explain phd concept is instagram ad metaphor

[true laurel]

Doesn't swapping sessions break remote cache?

[pale ravine]

That's okay, I'll do my best to understand. Thank you.

[vocal night]

Short term/one off Cache breaking is irrelevant longterm benchmark for quality.

[vocal night]

Users don’t care about the swap, they care about the qol quality of life feature experience can my agent act like and context switch like a human

[vocal night]

Here I had Eva try to explain for you:

Hi Felix, a raw Codex session is mostly chronological prompt history plus compaction written by engineers to do simple one off task. It’s what a CS 101 student might create. It works but it leads to dumb agent syndrome.

It can be reconstructed, summarized, and resent, but it is not inherently “managed memory or context.”

It becomes expensive and fuzzy because the runtime keeps asking the model to infer priority from a big transcript.

LCM-style context is closer to managed state: blocks, summaries, provenance, retrieval, drill-down, cache/compaction policy, and an assembly contract.

That means the agent or host can decide: this block is pinned, this one is stale, this one is evidence, this one is scratch, this one can be expanded only if needed.

That is memory-like, even if it is technically context management rather than durable semantic memory.

She probably explained LCM theory (what the plugin is designed around) better in 2-3 paragraphs than I ever could

[worldly tusk]

if you can have a savable context, then you can probably do many weird and fun things with it: duplicate to multiple agents, download, send to your friends, etc. sounds cool!

[vocal night]

Theres a major disconnect in the engineer quality at every ai company on the engineers who train and design the LLM vs the front end engineers who deploy it. We see that in their code every day 🤣.

I literally see them like word press pretty button engineers that just spit out what chat gpt tells them. No systems design or architecture ability, in fact those that do understand the papers, have handed off their thinking to ai too.

Its standard venture issue, acquire company for one feature now you have to give founder job> okay put them on pretty button golden handcuff duty.

That being said codex desktop is pretty and works tibs is doing great

Except when they vibe code break auth, /fast, and all the features

I can’t even have a convo w them without them spewing copy paste pasta so I don’t anymore its like talking to interns - love the vision but their actual contribution is tell gpt 5 make button pretty

[pale ravine]

After talking with my agent, I have a pretty good idea of what's going on. Thanks!

This makes me hope more and more that LCM will be integrated into OC.

[vocal night]

The irony of yesterday’s convo is not that LCM breaks (what it does works and always has), we simply tune it based on model constraints, its that when OC breaks or changes (ie changes entirely from pi to codex lol) it breaks the how lcm gets data or compacts.

You can hold me to it, companies will move to managed context vs just prompt slop and priority files 🤣 when the front end deployment engineers learn how to bang stick on keyboard and everyone else will copy

[shell dawn]

theres so much untapped potential in context compositing

[vocal night]

So true. The mistakes I make usually in oc is that I’m trying to bandaid their legacy thinking “more plain text instruction files” its like brute force “just tell agent do this! Prioritize this!” Its all very janky fixes to context

[shell dawn]

half of the jank is dealing with jsonl

[vocal night]

We just need one more instruction file!

[shell dawn]

cant wait for that sqlite switch, as much pain as its going to be to switch, it'll be so worth it

[vocal night]

And prioritize it after the last one!

Normies are surprised when they learn just how simple it is deployment wise

[shell dawn]

for how light sqlite is, its friggin ridiculous in its capabilities

[vocal night]

Wait till they learn about other databases, we’re just reliving engineering from a new generations perspective

[shell dawn]

like, postgres has gotten absolutely absurd, but its quite a bit bigger. sqlite is frickin miniscule

[vocal night]

Engineers are becoming less specialized and more generalist

[shell dawn]

ive seen someone doing a series on postgresmaxxing, where he replaced an entire saas app stack with postgres + postgres functions

including frontend, and auth

its postgres all the way down

its maniacal and i love it

[vocal night]

I prefer postgres nowadays ie my work on gbrain just its brittle af compared to sqlite

[shell dawn]

well yeah ones an actual server that needs the care that persistent, designed and maintained servers need, and the other is in many cases as simple as a library

[vocal night]

Multiple fanned sqlite with connection point id’s and layers is best solution mid term and we’ll likely see that in every tool by end of year

[shell dawn]

all the agentic tools gonna need their own lil sqlite

[true laurel]

Context composition is a bit of an art that we're all still learning how to do well 🙂 at least we've got a few hooks now to mess with it

[shell dawn]

i wasnt a dba but i was what dbas called "an accidental dba" because i had to make them work fast since i was the systems infrastructure architect

to me the prompt is a sql query, and really smart database engineers learned how to make really smart database servers that could run really smart queries

[vocal night]

Its actually simple, take a class on hardware engineering. AI systems is just gonna repeat memory layer and ram/cpu type design

Every discovery was actually made 30 years ago 🤣 but software engineers are now discovering systems and engineering design

[shell dawn]

how systems work together is a bigger deal than lines of code

thats where openclaw is so strong, its a nexus that can make a lot of systems work together in ways nobody ever could by themself before

[granite wharf]

as someone who recently quit their corporate training job for a career in ai, the molding and meshing required to get ai in real "blue collar" atmospheres is the "systems" approach...

first post.

[shell dawn]

whats funny, my vision of where i think agents are going into working with an employee workforce looks a lot more like reutilizing the systems that manage people than systems that manage code

sure agent 365 and project lobster is probably going to give us AD group policy for openclaws

but you are also ging to need to give it the employee handbook, the IT acceptable use policy, your data classification policy, your compliance & retention policy, etc

[granite wharf]

i worked for a 100 year old company, planned the party, but they had no interest in AI. just before i resigned they spent untold amount of money on an inventory ai program...

first week...

took all irrigation nozzles out of one store to transfer to another for no reason what so ever..

[granite wharf]

told me yea... we forgot to give it any history.

i quit 2 months later

[shell dawn]

install AI, fire, aim

[granite wharf]

lol... i just vibecoded a fix for codex on the claw.. posting to git now

[shell dawn]

hell yeah

be sure to read the contributing, and then read the followup

and then turn it into template

what'd ya find

[granite wharf]

yeah...

did you see where i said vibecoded lol..

youre speaking greek

[shell dawn]

this is what you tell your claw

it'll do the hard part for you, its about the system, and making a system to make systems

[granite wharf]

can i post you the git>>>>?

[shell dawn]

thats kind of what this channel is for 🙂

[granite wharf]

https://github.com/openclaw/openclaw/issues/87536

[rocky pewter]

You’re always positive and always there when it matters, love that.

[granite wharf]

last time i was in a real discord i didn't make, it was for arma reforger lol

[shell dawn]

thats an issue that you put in most of the work for a PR. lol

PR = pull request = merge my fix

issue is 'hey i have a problem, someone make a fix'

[granite wharf]

https://github.com/openclaw/openclaw/issues/87536#issuecomment-4561683002

heres the fix

[shell dawn]

yes, like i said, you posted an issue that did most of hte work of a PR 🙂

ask your agent to repost it as a PR

write it to a plan, and then ask it to look at the newest beta to see if its something that is already addressed. this one might be

[granite wharf]

learning how to PR and doing it at the same time.. one min..

[shell dawn]

the magic of openclaw in action

[odd anchor]

https://tenor.com/view/cat-gif-25090491

[granite wharf]

THANK YOU FOR BEING SO COOOOOOL!!!!

[granite wharf]

i keep typing date into my claw to check it lol

[rocky pewter]

Look ur dm and point ur agent to the docs shown there.

[shell dawn]

you send him the contributing.md doc? it kinda looks like he or his agent found it already

[granite wharf]

thank you.. i'm mid upload

[rocky pewter]

yes was the way i did when i start.

[shell dawn]

this is the way

[granite wharf]

https://github.com/openclaw/openclaw/pull/87563 🦞

[shell dawn]

you've got it. now a bunch of crustaceans are going to go to work on it and tell you what they think about it, giving you some tips of what they want to see, have your agent review it as the feedback comes in

and also while you wait. ask your agent "openclaw 2026.5.27-beta.1 just came out, check if whats in this PR is addressed or needs to be changed or rebased"

[granite wharf]

check dm.. i need you x handle if you havce one 🙂

[vocal night]

Default behavior is usually pull from main by agents when they pull a repo down

So they usually build against main unless told otherwise

[shell dawn]

really? mine have always pulled based on what i had installed, but my "prod" server has always been from npm packages and not main

i may have taught it that accidentally long ago

[vocal night]

Ah your main fork isn’t synced

Need to sync its commits

I forgot I have auto sync on for forks for pr’ing

Agents will check if stale and usually pull down latest main and create a repo mess

[shell dawn]

its the better way im sure

[granite wharf]

Big find. The maintainers have been hammering on this exact bug. Recent commits to native-hook-relay.ts on upstream main include:

42e9504114 "fix(codex): preserve native hook relay across restarts" — addresses the persistence issue
96959ec3d7 "fix(codex): defer native-hook-relay unregister to avoid cleanup race" — this is literally my Patch 3
6729dea36f "fix(codex): share native hook relay registry (#73950)" — closes the original April family
cc2948d1e1 "fix(codex): narrow legacy hook generation grace (#87386)" — recent followup
592aae3696 "fix: avoid idle Codex hook relay subprocesses"

The maintainers are already deep in this code. My PR's approach might be redundant or it might need to rebase onto current main. Let me look at the actual current state of the file vs. my change.start_process

heres my current spot.. agents still working

[vocal night]

File it anyways let the ai maintainer handle and if you find something novel it will tell them and update them, if not it will close pr and tell you

[granite wharf]

rebasing against main now

[vocal night]

Molty is pretty good about if you have a novel fix updating the other threads on it

[granite wharf]

https://github.com/openclaw/openclaw/pull/87563#issuecomment-4561796880

can someone check this

[shell dawn]

its probably already addressed, this has been the hot topic of the day, but this is good practice. plus, its entirely possible that you might have found an aspect of it that others didnt

ask your claw what it thinks real behavior proof would be, ask it to read the feedback and identify how to address it

[granite wharf]

Confirmed I checked release/2026.5.27 (the branch producing 2026.5.27-beta.1) — the prune logic is not there.
The 2026.5.27 release branch does tighten other parts of the relay code (canAcceptNativeHookRelayGenerationMismatch simplification, import path rename) — but those don't touch the bridge dir lifecycle.
My fix is complementary, not duplicative.
Branch rebased onto latest main.
Open invitation to maintainers to retarget the PR to release/2026.5.27 if they want the fix in the upcoming release instead of waiting for the next minor.

That last sentence is the move that signals "I know how OSS release branches work" — exactly what rageclaw was nudging you toward. Maintainers will see a contributor who:

Checked the upcoming release before pushing
Did the rebase work themselves
Offered to target the right branch if asked

Now you wait for feedback. When reviews come in, paste them into our channel and I'll help you read each comment and address it — same way rageclaw described it ("have your agent review it as the feedback comes in").
Tell rageclaw thanks again. He just walked you through your first real OSS contributor flow.

[shell dawn]

haha damn, tell your claw i said thanks lol

[true laurel]

I've done hardware and os engineering, I've only recently stepped into learning AI, so I'm still slowly getting a grasp of how it all works. 🙂

[zinc dock]

Thanks for your work on this !

[vocal night]

You’re already ahead then, just apply that to ai and watch the “wows”

[celest merlin]

Goooood morning folks

OpenClaw 2026.5.27-beta.1 is live.

This is a beta for people willing to help shake out install, update, channel, provider, and runtime regressions before stable promotion.

Update from an existing install:

openclaw update --channel beta --yes
openclaw --version

Expected version:

2026.5.27-beta.1

Fresh install:
https://openclaw.ai

What is worth testing:

• Security/runtime boundaries: unsafe command wrappers, malformed numeric CLI options, unsafe Node runtime env overrides, no-auth Tailscale exposure, and non-admin device-role approvals should fail cleanly.
• Channels: Telegram durable sendMessage replies, Matrix mention previews/finals, Slack late-cleanup delivery, iMessage approval handling, Discord recovered tool warnings, and Google Chat DM thread behavior.
• Providers/models: OpenAI-compatible embeddings, DeepInfra model browsing, Pixverse video generation + region selection, VLLM thinking params, Claude CLI OAuth overlays, and direct Anthropic model ids.
• Gateway/perf paths: status, auth/env snapshots, plugin metadata, tool-search catalogs, session reads, and stable metadata caches should feel less wedged and less rediscovery-heavy.
• Package/release paths: npm/package exclusions, shrinkwrap pins, Docker runtime templates, postpublish verification, and beta smoke failure handling were tightened.

Validation:

• GitHub prerelease is live: https://github.com/openclaw/openclaw/releases/tag/v2026.5.27-beta.1
• npm beta tag points at openclaw@2026.5.27-beta.1.
• Full release validation, npm preflight, publish, and postpublish verification are recorded in the release notes.
• Local beta smoke was blocked on this host because the Parallels Ubuntu 26.04 VM is missing.

Please try the beta on machines where you can tolerate sharp edges and report anything weird with version, OS, channel/provider, and the exact command or message path.

some of the work we do

pi is 🔥 as well.

[untold pollen]

<@&1503801512294486217>

[final ocean]

These are going to be closed, worries about the complexity and size/impact of these changes that are deep reaching without speaking to the team first. The whole stack and phases will be closed shortly. One of the PRs was fed into one of my codex sessions and it took 60 commits+ to get it manageable and then the scope drifted massivley. Sorry i have to nuke all the work but please reach out next time.

[vocal night]

All good, it’s impossible to unify when the codes unstable

It was fine and all diamond a few days ago, but every main drift and merge leads to new rebase. I can re unify but it will be impossible unless code is frozen

Per the PR’s they were designed phase wise to be implemented over 1 month

Not all at once it talks about in PR 1 the risk of trying to merge all at once without freezing code changes to channels

[vocal night]

I noticed most of the focus was on implementing the final phase of the unification all at once

[worldly tusk]

maybe a reverse approach is needed? first: work on the particular channels to make them similar. after they are more similar, it would be possible to unify them

[vocal night]

The phases go channel by channel actually

Just nobody reads anymore lol

And Ai just tries to implement

[worldly tusk]

i think the first phase was not channel by channel

only second was

(i read)

[vocal night]

First phase sets baseline

The link is to phase 4 final implementation

To be fair with 60+ maintainers they should be doing the hardcore engine stuff Eva’s been doing

Non maintainer should not be shipping such a high percent of the core code

[worldly tusk]

my proposal was to have no phases first. just an independent set of patches for the channels to bring them individually to a state where it is even possible to speak of unification. after the channels look more similar, it could be possible to work on a next phase

[vocal night]

Its such a spagetti state its not possible I ran multiple sets of pr’s for days. Theres no patching what is inherently bolt on after bolt on

[vocal night]

This is what has been attempted and failed for the last month

[final ocean]

not true

[hollow mural]

Hey maintainers, quick review request:
https://github.com/openclaw/openclaw/pull/87581

WhatsApp login QR got huge in terminals after the last QR fix. This PR makes it compact again without bringing back the broken renderer.

Small PR, but important login path. Would appreciate a quick review.

[final ocean]

we will but i cant have this work done outside of my team, its too risky and hard to test other peoples work on that scale without context of our medium term vision.

[vocal night]

I don’t disagree thats why I said take over

[final ocean]

changes of this magnitude should be discussed. were happy for people to contribute.

[vocal night]

I’m sure I ruffled some maintainer feathers today with the front end dev call out lol

[worldly tusk]

ok, another idea: a reference reimplementation of a single channel. make a new plugin, something like telegram-experimental, that would have a sane structure. after it is done and tested, and retested, and included in the main, other channels can be reimplemented in a similar manner

[final ocean]

what do you envision should change?

lets start there

and why

[final ocean]

poking the bear

[vocal night]

The top 10 maintainers are world class no doubt but everyone else can step up 100x vs a single ai

So its a challenge and I’d love to see it

The best dev comes from okay prove it

[vocal night]

That is what the phase plan did but with multiple at once, recode phase 1 + only one plugin

Thats not bad actually

Divergent code base and keep legacy

And slowly offboard channel by channel

[worldly tusk]

i am just giving @vocal night process ideas. i have been living in IT projects for 20+ years, so i've seen many ways to fail 😄

[vocal night]

Well the right solution is its maintainer led who owns it and they just need to follow a plan to implement and own it

Vincent can’t own it all nor can any single contributor

The right model is they do it one at a time with a opt in for testing that test 1-1 your solution is the most enterprise and safest it just breaks phases down further into 1 task, testing per release @worldly tusk

I’d love nothing more than to see eva drop from like rank 11 to 60 and not be patching core 😮‍💨

[final ocean]

The thing that is hard to see is we have now a structure in place and a number of people are owning and acting as stewards on various subsystems and this has been shapping up. The OpenClaw Foundation has been stood up with people like myself and various full time people from OpenClaw and partner organizations who are driving this and will start ramping up more and more. We will share more details publicly soon and to this group of course. But me and Peter are not bottlenecked

[final ocean]

Lets please avoid huge PRs that reach into refactoring how our code works without a discussion. Too much to ask?

This way i can co-ordinate the right codeowners to come and discuss things with people that want to contribute. More and more of the team will start to show up here on a regular basis and engage with everyone.

Lastly we are planning a town hall aimed at clawtributors and all public contirbutors so we can update and create more visibility on what we are doing and how we are addressing everyones concerns. But rest assured everything said here is heard and you are all listened to. We would not be here without you the community beliving in OpenClaw 🦞

[worldly tusk]

testing the new beta. after the memory.md removal from being repeatedly injected, the prompt is much cleaner and consumes much less tokens, thank you @celest merlin, this was a massive win! few more minor loose ends still in what's repeated every turn, i will try to untangle them later this week...

[granite wharf]

Anyone got any insight into this...

Clawfather tweeted it was fixed, but I think we are looking at two different hooks.

anyone?

https://github.com/openclaw/openclaw/pull/87563#issuecomment-4562819550

[shell dawn]

wait for clawsweeper to finish and see what it says, once you get to diamond claw or platinum hermit and it says you have proof and its ready for maintainer review you just wait, keep following up on the feedback but clawsweeper posts before its all the way done

and the CI too

[shy rain]

Did somene fix the Native hook issue already? I swear I saw someone mention it.

Oh it's fixed in 5.27-beta.1 that's good to know. So I do have a reason to update 😉

[pale ravine]

I'm waiting for the official release

[vocal night]

The team tagged me other day and said are also using my other engine refactor for the internalization, are they dropping that too?

[final ocean]

not sure, not across everything. im saying more generally

@vocal night that 100yen account, is that you?

[vocal night]

Yeah, I do love everything you said btw and I can be less a meanie if it means team politically will use good code tbf

[final ocean]

just be nice human, not aimed at you - just everyone in general. good way to live life 🙂

[vocal night]

Another hack to take OC to the next level: set a real bar. “Maintainer” should be a privilege, not a status symbol.

Globally, OC is the underdog now anyway, and when I bring it up lately people increasingly get an “ick” from the chaos around it.

Every maintainer should contribute more than a single eva claw’s worth of output.

Right now it feels statistically like:

• ~1/3-1/5of maintainers are absolutely killing it
• ~1/3 are good, but their work/inference balance is off and they can’t fully keep up
• ~1/3 were here for the hype phase and are slowly fading out now that the novelty is gone

A big unlock would be to internalize, consolidate, and clean out bloat code and team.

I’d strongly recommend making maintainer status merit-based and redistributing free maintainer inference toward people actually shipping.

The best maintainers should get more inference, more concurrency, and guidance on how to run multiple CLI/desktop instances so they always have something cooking in parallel with their work.

Inference is productive capacity. The people using it best should have the most of it.

I’d also recommend building a contributor ranking/grooming system:

• contributor
• trusted contributor
• maintainer candidate
• maintainer

Mentor strong contributors over time and reward serious contribution with additional inference/resources month-to-month.

That creates a natural cycle:
more quality contribution → more trust/resources → even more contribution.

If someone stops contributing, they naturally fall off over time.

That’s healthier than permanent titles or maintainership as social status.

The project is in missionary mode now, not mercenary mode. This is the phase where projects either regroup, internalize, and tighten up — or slowly decay under their own entropy.

And candidly, I think OC probably has ~2 more genuinely broken releases before a lot of people silently stop touching it entirely. Infra trust cliffs happen fast.

[final ocean]

@vocal night im honestly im starting to take all these messages personally

you said you want to be a good huma n then you say 1/3rd of maintainers are here for hype

what good is this message? do you have any examples or just want to cast suspicions on people?

[cinder bolt]

I'm trying to understand the batching code for memory embedding. It seems that it just batch 1-15 embeddings per batch-request, OpenAI should support up to 50k request in one batch.

My Claude CLI using Gemma4 doesn't understand the code either 😄

[vocal night]

Shipped code in 30 days and 60 days

By maintainer.

[final ocean]

have you ever considered that not all maintainers are here to ship code on openclaw/openclaw

you are on thin ice, i have to say this publically

anything directed at the maintainers i take seriously

[untold pollen]

Are you talking about the GitHub repo? Bear in mind we also have moderators there that help us a ton in managing spam (which we get a lot of)

[vocal night]

Peter himself has been firing maintainers and messaged me when I first talked to him that yall needed maintainers that actually do stuff.

Thats how I got involved

But said no I can’t be involved I’m only here for my biz and customers

Maybe I’m writing it wrong I thought the product was internalizing

[final ocean]

@vocal night i understand you expressed interest in becoming a maintainer. not everything about open source maintaince is code. Please keep your personal maters to yourself

[celest merlin]

Your messages are not productive and they read like AI. Esp that large one.

[final ocean]

@vocal night final warning

[vocal night]

Apologies if that came off anyway than helpful mate I’ll take a cooldown

Havent slept in a few days coding

[final ocean]

@vocal night no sleep from coding == psychosis teritory. Please if you need someone to speak to give me a call or touch grass. look after your health

but i will have to boot you if this continues, just personal attacks to maintainers and others over a continued period is not helping anyone.

[shy rain]

5.27 released? Is the native hook relay bug fixed? Cos the chat is flooding with reports, and I'm getting it too.

[shadow garden]

this is awesome! we can't wait -- thanks for the update Vincent

[timber beacon]

Love the feel and direction things are going

[pastel owl]

Hello all
wanted to know if anyone is using codex runtime on latest main ?

Yesterday i had tried and it wasn't working, so I have since switched to release/2026.5.27 branch

[worldly tusk]

yesterday's main should have been released as the stabl today. and that works with codex

[pastel owl]

No both have diverged

release branches are now not follwing main, they have cherry picked commits and ported commits on it

And also main now has backported commits from release

[shy rain]

I would love this one to land sooner rather than later. And any other native-hook-relay fixes.

https://github.com/openclaw/openclaw/pull/87563

courtesy of @granite wharf

[tall prairie]

it's still broken?

[jolly wolf]

Idk about main but it is on latest release

[shy rain]

on my machine, and the hordes swarming #general right now with torches and pitchforks, yes.

[tall prairie]

(also why is the yells at clawsweeper emoji scuffed)

[shy rain]

it's the 2-frame byte-saving version lol

[drifting nymph]

this is great idea. coming from open source firmware community for 8 years tought me lots of lessons. it is definitely a quite challenging path to take the role as open source maintainers especillay for a big projects with so many people with different expectations and opinions. lot of time you will find yourself questioning about life... but it is what it is. so happy to see this project momentum is going on (its not all sunshine and rainbows). keep it up!!!

[worldly tusk]

oh... oh my... there is a huge codex refactor in main, that is not in the releases, @pastel owl. https://github.com/openclaw/openclaw/commit/bb46b79d3c1479f194a90afcf3dd69a1858a7898

very interesting, maybe something broke for you there? i was working on a small patch. now i need to start again 😄

huge rebuild. now i need to go and read what's cool is there 😄

[pastel owl]

Yes and now i cannot be on main and contribute
Release branches are gonna become stale after every release

[pastel owl]

100k loc touched

[worldly tusk]

nice!

[shy rain]

Issue: https://github.com/openclaw/openclaw/issues/87612

• /verbose off suppresses all native progress callbacks (tool start, plan updates, command output, etc.) in group/channel sessions. Users see no indication the agent is working.

Fix: https://github.com/openclaw/openclaw/pull/87678

• Removes the chatType === "direct" guard from shouldAllowQuietChannelOwnedProgressCallbacks in dispatch-from-config.ts. Progress callbacks are now forwarded in all chat types when verbose is off. Includes updated test + live Discord verification.

Quick fix for missing tool-call streaming info in group chats (discord + telegram)

[shy rain]

This is why we need LTS. Can't we make 5.27 the LTS release and just work on fixing that whilst the big refactor goes on?

[worldly tusk]

why only direct? i use a regular discord group just with my agents.

[pastel owl]

Yes that's a good idea to have a new stable or LTS branch

[shy rain]

5.27 is broken for a lot of people, by the noise we're getting in general chat. Including myself, but I know there've been a lot of patches done to fix it and it's quite a dispersed issue so I'm not gonna fix it myself.

[shy rain]

That's literally the problem. It shouldn't be chatType === "direct" but it was. it was filtering tool calls from streaming info in group chats. I fixed it.

[granite wharf]

https://github.com/openclaw/openclaw/pull/87563

this is as far as i got. my first pr. and i passed out at my terminal last night

[worldly tusk]

ah, i am very sorry, i misread

very good then!

[shy rain]

I pushed for it a few messages ago and Disciplined gave clawsweeper a nudge. It should be merged very soon.

[shy rain]

lol no probs

[granite wharf]

Pete had said this was a fix but i researched and this was another issue apart from that so i hope this is the final fix for that problem. stack has been up for 7 hours no errors

[pastel owl]

There's a workaround in config

agents.defaults.verboseDefault = on

[shy rain]

No, because /verbose on gives BOTH tool call messages in streaming, and full verbose tool call logs in chat.

[pastel owl]

Check if this mitigates the issue for you

[jaunty lodge]

Are we still in the market for fixes?
I just rebased this one, so it's ready to be smoothly merged once again 👍

[shy rain]

I was already working around it by doing that, I got fed up of my agen'ts message getting scrolled off screen before he finalized it.

[worldly tusk]

yes, verbose leaves spam in the chat. most people want the draft bubble be replaced with the final answer, not left in the chat.

[shy rain]

I hate coming back to the "dead" agent... realising he's updated his message that's about 130m upthread.

[pastel owl]

Ohh i see

Just checked my discord agent
You are right it's broken
Telegram is fine though

I use tele daily

[granite wharf]

I spent 6 hours figuring out why all the tool calls were dying and agents restarting.

[vagrant vortex]

5.26 as well?

[shy rain]

Yes, 5.26 is broken for me, same issue... the Native hook relay unavailable error stopping bots from making tool calls, intermittently.

I was using 5.26b2 for hours, and then the main release for hours, before it cropped up. Now it won't go away making openai models very annoying to use (workaround is for them to use subagents for all tool calls)

[shy rain]

The weird thing is I asked my bot if it was just discord and he said no, the line we patched is global, affects any "group" chats, so idk why your Telegram is working but good for you! 😄

[fallen gate]

known issue, should be fixed mostly in .27?

might be missing one PR still

[shy rain]

not fixed in 27.... waiting for Adam's PR which is ready to merge.

[granite wharf]

Should be merged.

[shy rain]

I'm gonna install 27 and use Adam's patch (patch Adams!) as soon as it's merged.

[granite wharf]

https://github.com/openclaw/openclaw/pull/87563

it should be merged. its closed now

[shy rain]

should, probably, but not yet. It might have been fixed with a different patch

I think clawsweeper is doing it right now

[worldly tusk]

maybe his telegram registered as "direct"? 😄

[shy rain]

Well yeah. But a lot of people use telegram group chats as a way to separate topic channels. So they get hit with the same bugs as discord group channels.

regular telegram is just like discord DM

[worldly tusk]

so, we probably need to migrate to main, and test the refactor? 😄

[granite wharf]

will you make sure to let me know if this is a fix for you??

[shy rain]

bot says it's definitely a real fix and covers a very common state for us. Especially around updates where I keep having to restart the gateway, and sometimes not very cleanly.

I'm applying it locally regardless.

[shy rain]

Good luck. We're all counting on you!

[granite wharf]

Quit my corporate job 2 weeks ago to persue ai.. never thought i would commit a git to help the claw.

god bless the claw.

[shy rain]

my claw is happier for your PR, and by the sound of it I'll be sticking with patched 5.27 for a while.

... and my own patch, of course... today's one not yet merged.

[hollow mural]

Any chance this PR can get a review soon? Sorry for the nudge, I just think it fixes an important WhatsApp login regression.

https://github.com/openclaw/openclaw/pull/87581

[zinc dock]

I'm guessing another update has to come soon because most won't manually patch

[shell mesa]

"Hey everyone 👋 I'm Malik, independent developer based in Kenya. I've been researching the indirect prompt injection attack class that was behind ClawJacked and built an experimental implementation of a fix called source-aware instruction tracking — where external content is structurally prevented from triggering tool calls at the execution gate.
I've written it up with working TypeScript code and published the research at github.com/thecolourfoundation/Color. About to open a GitHub issue proposing it as an upstream contribution.
Is this the right place to get maintainer feedback before submitting? Happy to discuss the architecture here first.

[granite wharf]

Pushed

[winter ruin]

https://github.com/openclaw/openclaw/pull/87710
feat: add antigravity-cli skill for https://developers.googleblog.com/en/an-important-update-transitioning-gemini-cli-to-antigravity-cli/

[zenith prawn]

how to reproduce? i just updated and want stress test, and revert if it fails (nightly devops work)

[weak walrus]

sheesh lol

[zinc dock]

For me it was have your agent launch a bunch of native codex sub-agents and make sure it's a test that touches tools. The agents will do the reasoning and then they'll come back as blocked before they touch tools because of native hook relay unavailable

Not sure if the other fixes for this have smoothed this out but Adam's PR was the last missing piece. I haven't updated yet

[zenith prawn]

this passed on 5.27. Though, not visible via /subagents command; but both visible via /tasks

[zinc dock]

Updating now will test throughout the day

[zinc dock]

Hey what aspect of the native hook relay issue did your PR address so i can look out for the symptom since i dont have ur patch....

[granite wharf]

https://x.com/aas_beacon/status/2059931471045877824?s=46

[zinc dock]

Thank you sir 🫡

[trim vector]

merged - nice catch!

[hollow mural]

Thanks 🙏

[granite wharf]

Did the fix work for you?

[zinc dock]

Testing 5.27 now without your patch to see if I run into it

[granite wharf]

Pete said one was fixed in 27. But mine wasn’t. I was running a voice model and 2 text. Some OpenAI codex and 2 small models on ollama. The problems was both couldn’t write or read at the same time. So once one took over the other the whole tool call broke down for both.

[drifting nymph]

opus 4.8 just landed!

[zinc dock]

Gonna keep running this throughout the day from my clanker :: "Both checks are back. Good news: neither worker found evidence that the native hook relay bug is still happening.

Results:

• Smoke test passed: native shell calls, openclaw --version, openclaw status, expected harmless command failure, apply_patch, file readback, and finalization all completed normally.
• Log/CLI inspection passed: no fresh post-upgrade Native hook relay unavailable / Command blocked by PreToolUse hook pattern found.
• Hook surface looked healthy: OpenClaw/Gateway both on 2026.5.27, gateway OK, hooks reported ready.
• The old failure signature did not reproduce under actual native Codex tool use.

So my read: the 2026.5.27 fix looks good from this first probe."

[shell dawn]

ive been testing it out and its pretty nice. 4.8 on high is fast

[worldly tusk]

Anyone running main? Planning to upgrade to it stortly, but maybe better not?

[zinc dock]

I'm surprised they actually made 4.8 fast cheaper than 4.7 fast was. That Elon compute is already kicking in.

[shell dawn]

im not even using the fast path

just regular 4.8 high feels quite zippy

[glad echo]

I'm seeing "codex app-server turn idle timed out waiting for turn/completed" errors after updating from 5.20 to 5.27. I found reports pointing to a similar failure in general and in help https://discord.com/channels/1456350064065904867/1509022225859219566 . This issue comes closest, but it's not limited to telegram. Forgive me if this was already on the radar, couldn't find it reported in here. https://github.com/openclaw/openclaw/issues/87744

[slow inlet]

@untold pollen https://github.com/openclaw/openclaw/pull/80046 only pinging as you expressly told me it was cool to do so. Someone made this fix for the cli that is good. I gave it a review for what it was worth.

[fallen gate]

@opaque sand CLI related work, might be up your alley, I think Obviyus is sleeping

[pastel owl]

Just installed 5.28 alpha
Feels good uptill now. No immediate blockers, running 2-3 diff topic threads simultaneously

[crude matrix]

Please fix kimi model, I had that one as primary for some agents of mine...

[fallen gate]

you have to give a little more details than this.... and we can't fix models

[drifting nymph]

just switch to Pi harness. much better on our side.

[crude matrix]

sorry, copy disapeared.... https://github.com/openclaw/openclaw/issues/87641

[fallen gate]

I have like 5 PRs specifically fixing issues with pi harness atm

[glad echo]

My message wasn’t intended as a help request, but for flagging a potential issue seen in multiple setups.

[drifting nymph]

could you share with me? would like to take a look please

[merry slate]

Hi Peter, I saw you fixed the issue(https://github.com/openclaw/openclaw/issues/86939) I reported last time, thanks for that.
I noticed it still had a small problem in actual use, so I added a follow-up fix here.

https://github.com/openclaw/openclaw/pull/87767
Merge readiness
Overall: 🐚 platinum hermit
Proof: 🦞 diamond lobster ✨ media proof bonus
Patch quality: 🐚 platinum hermit
Result: ready for maintainer review.

Could you take a look when you have a chance?

[slow tiger]

thank you for merging this PR i made: github.com/openclaw/acpx/pull/227
may i please have the role. not sure what verification is required

[lone iris]

@slow tiger ^

[slow tiger]

Thank you

[drifting nymph]

i saw 2 of them:
https://github.com/openclaw/openclaw/pull/86642
https://github.com/openclaw/openclaw/pull/87093

[sly depot]

Glad to see you getting purple!

[sly depot]

Woo, you got it too!

[worldly tusk]

looks like the great refactor brought with it lower memory use and significantly faster build time

[zinc dock]

5.27 still good on Hook relay front:

"All 4 native hook tests are back now. Clean pass across the board.

What they covered:

• Shell + apply_patch add/update + file readback: PASS
• Search/read/discovery + report write: PASS
• OpenClaw CLI/status/hooks surface: PASS
• Intentional harmless command failure + continued execution + finalization: PASS

Specific old failure did not reproduce:
Command blocked by PreToolUse hook: Native hook relay unavailable

Also no PreToolUse block, approval weirdness, missing tool events, stalls, write failures, or finalization trouble were reported.

So we now have two separate rounds today showing the same thing: native Codex tool relay looks stable on OpenClaw 2026.5.27 (27ae826)."

Gotta be my favorite version so far....

[granite wharf]

I’m jelly.

[obsidian zodiac]

I have partially confirmed that Auto Compaction and manual compaction is not functional in the recent 2 updates. I've had my agent ingest the entire english dictionary roughly 12 times. 1 compaction did trigger, but failed on the second turn. I'm going to do this test again with an absolutely clean install of the latest update, no memories, no plugins, no skills, fresh agent, and report if it still fails or not. This is a major bug. First test was on the Codex harness, with an existing install being updated to the latest. The next test is going to be the Pi harness with a fresh install.

[worldly tusk]

more work on reducing codex context: https://github.com/openclaw/openclaw/pull/87788

currently, the openclaw skill summary block is repeated each turn, burning tokens all the time. skill list is a stable content that doesn't update much, so this duplication is just useless bloat. especially if one has many skills. this PR moves openclaw skill list (as well as the memory pointer) to collaboration developer instructions, that do not reinject stable content. codex skills seem to be working similarly.

[obsidian zodiac]

100% confirmed on clean installs.
Compaction seems nonexistent in versions 2026.5.26 and 2026.5.27 for both the Pi runtime, and Codex runtime.
2026.5.22 is confirmed working for both
Oddball: 2026.5.18 seems to be partially working with auto compaction triggering sometimes in the Codex runtime, with manual compactions working 100% of the time..
The test, feeding the entire Oxford dictionary to my agent, 1 pass at a time until compactions were expected to triggerd. The goal was a total of 10 compactions, and both 5.26 and 5.27 are confirmed dead after the first run.
Go ahead and test for yourselves and prove me wrong on the main branch without any custom patches. Multiple users have been experiencing the same issues around compaction, but we all thought it was the alternative memory plugin and context engine we were running. These tests were conducted without any additional plugins, skills, or anything else that would interfere with the native OpenClaw memory back end and context engine.

OSs tested: Fedora 43 and Ubuntu 24.04

I'm going to continue testing beta branch versions now.

[worldly tusk]

Weird, as codex owns its compactions.

[pale ravine]

In version 5.27 I encountered a new problem. I only changed the agent so that part of the task of checking the system every day would be done once a week, and then the following two messages appeared:

Codex stopped before confirming the turn was complete. Some work may already have been performed; verify the current state before retrying.
Codex stopped before confirming the turn was complete. The response may be incomplete; retry if needed.

When I checked progress twice afterward, the same message appeared each time, and in between there was a situation where the context was being compressed. On the third progress check, it suddenly told me it was finished.

I've already had my agent report the situation
https://github.com/openclaw/openclaw/issues/87744#issuecomment-4569763395

[obsidian zodiac]

This is what i thought as well, but running "/compact" appears to not return any actions. And once the context window is maxed out, the agent becomes unresponsive.

@pale ravine that is the exact thing I was experiencing surrounding compactions. Once the context window is filled using the codex runtime, this is the only output from the agent.

[worldly tusk]

I think i've seen a successful codex auto-compaction just yesterday (with whatever latest beta there was). Or maybe i am halucinating?

[obsidian zodiac]

Thats why im about to start testing the beta branch. Ive only ran my tests on the main, cause thats what normies and plebs use.

But so far(prior to the last 2 updates) beta branch seems more stable than main with the releases. (I find this hilarious honestly 🤣)

[worldly tusk]

We should rename 😂

[obsidian zodiac]

Lmao or all the devs and contributors should be mainlining only the public main branch instead of beta or custom patched versions of OpenClaw to vet out bugs before main gets pushed.

[jolly wolf]

Get the beta pings role

id:customize

[obsidian zodiac]

This whole time, ive manually checked for beta, and there's been a role for it. Lmao that one is on me for not checking here 🤣

[worldly tusk]

There is currently in main a huge refactor of pi harness, that also changed plenty of stuff in codex. We should all be testing that

(It works, btw, so don't be scared 😂)

[obsidian zodiac]

I ain't scared of broken stuff. I love finding it though. Lol

Hell my first interaction with OpenClaw was "let's see how many agents I can run at the same time before things break down."

[worldly tusk]

And how many?

[opaque sand]

<@&1458375944111915051>
A new Dependency Guard workflow just landed. This is our first hard-required CI workflow. Any dependency changes in openclaw/openclaw must be explicitly approved by someone from the Security team (openclaw-secops). They can comment /allow-dependencies-change on the PR to allow the check to pass (once the PR is ready; new commits will invalidate).

With thousands of PRs, we often have unintentional or unnecessary dependency changes slip through the cracks. Security is a major focus and this will help us keep our dependency graph tight.

[prime vault]

Broken: On Codex sessions, budget-triggered compaction returns a successful no-op (ok:true, "codex app-server owns automatic compaction"). The host caller runCliTurnCompactionLifecycle treated that as a failure and threw CLI native harness compaction failed, jamming the session lane - on large threads the bot shows "typing" forever and never replies.

Related: #86772 added the budget skip, #85160 surfaced the throw, #87158 (open) fixes only the plugin side, #87738 (merged) handles the ok:false missing-thread case. None covered this ok:true host throw.

Fix: Treat the owned no-op as benign - skip the throw and let Codex self-compact. Matches the existing "Codex owns compaction" contract in the preflight path. +14 src / +77 test, verified on a live gateway (before: throw + lane jam; after: turn completes). codex review clean.

https://github.com/openclaw/openclaw/pull/87785

[obsidian zodiac]

Roughly 12. Comfortably about 6-8, ideally 5 or less.
(On less than adequate hardware 32GB ram no GPU)
This was back in March though, and ive since gotten a much better system, 128GB Ram, RyzenAI 395+ Max chipset

[timber beacon]

Thank you!

[prime vault]

neat had just seen a new line on the CI checks, good stuff guys

[civic sedge]

What is our process for requesting that?

[jolly wolf]

Talk about it here

[shell dawn]

does it light up the secops team on the backend of git tagging or anything?

[civic sedge]

i imagine a board...flashing red. @fallen gate just submitted a PR: all SECOPS summoned

[potent path]

"Uh-oh, another Kaspre PR. Get ready, everybody. He’s about to do something stupid!"

[jolly wolf]

Brb

[civic sedge]

/goal build this and don't make mistakes

[shy rain]

Nice, your PR got swiped by clawsweeper and merged.

[civic sedge]

clawsweeper really good at that

[opaque sand]

The dependencies-changed label will be added If security folks want to proactively search, but I suspect in practice whichever maintainer is looking at the PR can ping as needed. Any suggestions or ideas for this?

worth noting this affects about 2% of PRs

[civic sedge]

Not yet, I think we'll have to see how it plays out / how many pop up with dependency flags. DependsSweeper sounds too much like diaper cleanup...so SecSweeper? I just have seen how things like this can (for good and bad) become bottlenecks.

if its truly 2%, then maybe its a non-issue

[shy rain]

That's pretty awesome tbh... instead of rejecting because the ground shifted, clawsweeper just reformulates the PR for the new landscape.

[shy rain]

DependsSweeper is a terrible name

[civic sedge]

I mean...there's a reason I didn't get to name our first kid

[left mango]

SecSweeper may be worse actually 🤣
Specially when you say/ hear it.

[opaque sand]

well if you want to help me name something, part 2 that I'm going to land soon is an auto scrubber for lockfile-only changes

[civic sedge]

Like i said. Anyways... to Dallin's question - again its probably better to see how a new gate plays out before putting triggers/requirements to tag a maintainer

[opaque sand]

ok and for now when the job fails it does add a comment explaining why etc

[civic sedge]

I'm still waiting on the report back on the pr i've had open. it shows it ran maybe?

[opaque sand]

I might have to run it retroactively. checking

[civic sedge]

if you need the pr https://github.com/openclaw/openclaw/pull/76298

probably will cause a dependency issue lol

specially With peter's comment

[earnest jasper]

sorry to barge in.. fixed an annoying issue with doctor and an ambiguity that caused openclaw to incorrectly force codex harness for sessions the user has explicitly defined to use something other than codex: https://github.com/openclaw/openclaw/pull/87851
should be an easy one hopefully..
Overall: 🐚 platinum hermit
Proof: 🦞 diamond lobster
would love a maintainer review! ❤️

sorry i just realized it would be extremely counterproductive if everyone asked for a review when they submitted something ;p i'll refrain from that moving forward! what do they say... easier to ask for forgiveness than for permission 😄

[opaque sand]

@civic sedge I closed and re-opened your PR and that worked. For PRs with merge conflicts, fixing them will re-trigger anyways. But I will look into proactively triggering this for the many folks without merge conflicts

[civic sedge]

thanks. I just fixed merge conflicts earlier...moves too fast. i'll update those.

[void bramble]

Sick! My PR for Discord streaming commentary just got merged! It was a doozy, I'm glad peter worked his magic and got it all cleaned up / ready for prime time.

[civic sedge]

some good discord upgrades lately

[lusty vigil]

if you have dependency changes and want to get approval, @ me

with that in mind, this should prooobably be a rare thing. (we’re mostly trying to stop frequent accidental pnpm-lock.yaml residue, etc from merging into into main)

[zenith prawn]

before vs after update 5.22->5.27, telegram, /verbose degradation - commented on https://github.com/openclaw/openclaw/issues/87612

[shy rain]

which one? I have a related one up as well

[void bramble]

https://github.com/openclaw/openclaw/pull/85200

[shy rain]

My fix for that https://github.com/openclaw/openclaw/pull/87678 is hitting a dependency guard.

[shy rain]

Apologies if this isn't what you meant but my PR https://github.com/openclaw/openclaw/pull/87678 is stuck on "Dependency guard" which is something I'm not familiar with.

[opaque sand]

@shy rain was just writing a note about that
Rebasing on main should fix.

[merry slate]

there’s no trigger unless u change npm dependencies, whether adding or upgrading them

[opaque sand]

@the dependency guard check
As some people are pointing out, many PRs show a "waiting for status to be reported" from the new required check.
Rebasing on main should fix.

I tried to do a clever labeling trick to backfill runs for this specific action - see #87866, #87882, #87867. This failed for anything behind main because the action uses a js script, e.g. this 🤦‍♂️

But turns out the number of PRs which are blocked by this check and also in a mergeable state (no conflicts or other failing checks) is pretty low, like a couple hundred. So hopefully the pain will be pretty short-term. Any PR should be fixed by rebasing on main.

I'm adding super brief comments to that group to explain that pulling main should fix.

• let me know if it doesn't

[untold pollen]

This seems to be an intended policy change, not a bug. ToolProgress will only show when verbose mode is on

[shy rain]

But that's dumb. Because with verbose on, you get tool spam in chat already, persistent tool spam, not streaming that disappears onnce done (and is already optional)

/verbose is for debugging. tool output in streaming is a visual cue. and it tidies itself away

[worldly tusk]

That is a bad change though. If the tool progress is configured, it should be shown. It can be disabled in the config, then it should now show. There is no reason to arbitrary disable it when it is enabled really

[shy rain]

fwiw I did check the docs explicitly and the behaviour is not what is documented. And the docs describe exactly the behaviour I would expect. Which is why I fixed it.

[untold pollen]

I hear you, I quite liked it too (especially with sendMessageDraft), will discuss with the maintainers and figure out why it was changed

[worldly tusk]

Many people also use "group" chats as private chats with their agents

[shy rain]

Please do, because having two config options, one of which affects two separate things at once, is broken no matter how you look at it.

Guys sorry if this is the wrong place but in web gui chat... selecting a different session does absolutely nothing. I can't change sessions. Is this a behavioural change? Is there another way to do it? Or is this just broken? And if broken, is it just me or do others see the same thing?

[shy rain]

ACtually the PR is now failing with "needs real post-fix behaviour proof" despite that I already uploaded a screenshot of the perfect behaviour with my fix. i.e.
/verbose off called in chat...
tool calls embedded in streaming (because the option is configured to ON)
no extra tool call spam.

[untold pollen]

Scratch that, I misunderstood the bug. ToolProgress will still work even with verbose set to off, there's a bug in the native callback which wasn't sending any updates. Pushing a fix

[shy rain]

my fix is the fix lol

[untold pollen]

Can you link your PR please? I already started a PR

[shy rain]

please revisit 87678

https://github.com/openclaw/openclaw/pull/87678

[untold pollen]

ty, reviewing

[fallen gate]

what you complainin about now

[untold pollen]

Great PR, thank you. Landing 🫡

[shy rain]

First time my claw worked in weeks, I thought I'd pin down some of the quirky UI behaviour that I didn't have time to think about before lol.

Back to my other question... is session select dropdown working in gui chat? cos it isn't working for me.

[slow inlet]

@untold pollen please can you check out this pr - not even mine https://github.com/openclaw/openclaw/pull/80046

[zenith prawn]

exactly. i am. [in Telegram]. Especially because in DM topics it breaks more often, so i have 1:1 groups with threads with my OC just for session-splitetd persistent communication

[shy rain]

@untold pollen This is the TUI version of my fix... although it is a different layer fix, it shares the same behavioural fix and lineage from the #76914 fix which broke both.

It isn't mine.. I just had a look at it because it seemed related.

Actually seems 80046 is upstream of mine, and will be complementary.

[untold pollen]

Testing, been a while since I used the claude-cli backend so might take me a bit to test properly

[fallen gate]

people still use this?

https://giphy.com/gifs/morning-tea-wednesday-h0MTqLyvgG0Ss

wait where my gifs

https://giphy.com/gifs/rich-poodle-maltipoo-cXMFmN3edhlHI5vRsG

😠

[zealous plover]

Bye bye auto mod 😂

[worldly tusk]

We can read the title. Isn't it enough?

[fallen gate]

@jolly wolf it's broken, fix it

[fallen gate]

fix your link to the issue

and who was the "we" discussing? you have 2 messages on the server and 0 comments on that issue?....

[untold pollen]

Thanks! Landed

[worthy warren]

Hi! 2 small reliability bug fixes ready for review in core hot paths(agents lifecycle, MCP channel bridge).
CI green, regression tests included, both rebased onto current main.

fix(agents): dedupe subagent browser session cleanup wrapper with dispatch flag
https://github.com/openclaw/openclaw/pull/68669

fix(mcp): bound pendingClaudePermissions / pendingApprovals via TTL sweeper + close clear
https://github.com/openclaw/openclaw/pull/71648

Would appreciate a look when someone has a moment. Thanks!

[slow inlet]

Yes - I've even gone as far to cook up a work around for the programatic bucketing. If it won't land in main I'm thinking of a plugin called Mittens.

https://github.com/openclaw/openclaw/pull/81851

[merry slate]

How does this claude-cli-interactive backend differ from the current session_spawn + ACP path for running Claude

[slow inlet]

Unless I am mistaken, the current method invokes claude -p which will make it be charged differently

This version uses a MITM proxy (hence Mittens?) and cert spoofing to mimic things like zscaler and fork the api calls in and out to reconstruct the claude -p output, critically the stream jsonl which is needed to view reasoning and tool use live easily.

[merry slate]

Got it, thanks!

[slow inlet]

I think it will be invaluable to people who run out out of programatic credits 🙂

Now 80046 is merged I can rebase a little. I had it working and then a lot of the code shifted underfoot haha

[fossil ferry]

Hi all — I opened a PR to wire stop through the OpenAI-compatible /v1/chat/completions path:
https://github.com/openclaw/openclaw/pull/87920

It forwards client-supplied stop sequences into streamParams.stop, then to upstream Chat Completions as stop or Anthropic-backed transports as stop_sequences.

It also adds gateway validation for invalid stop shapes/types before provider dispatch.

I included local real behavior proof showing a real OpenAI-compatible provider stops before emitting the stop marker, plus 400s for invalid stop inputs. Feel free to take a look when convenient.

[slow inlet]

@viscid jay mods!

[winter field]

use /report next time

[winter field]

https://github.com/openclaw/openclaw/pull/87991
PR to fix image completion duplicate sends

it would send a completed image like 7-12 times xD
no moree! fixed it

[pastel owl]

Addressee the types.ts related warnings while build.

https://github.com/openclaw/openclaw/pull/87915

[worldly tusk]

did heartbeat break for anyone on main for non-codex agents? the user prompt has degraded for me and is missing the message Read HEARTBEAT.md if it exists (workspace context). Follow it strictly. Do not infer or repeat old tasks from prior chats...

[glad echo]

Hear me out. All (bot) spammer get moved to a new channel called #arena. Nobody else can make comments there, only the ‘contestants’. We throw in our own bot (Clawdiator obviously) who’s single purpose it is to extract API credentials, gift cards or cupcake recipes

[winter field]

agreed

[cinder bolt]

I put gemma4 on the task to try to help me implement batching embeddings over file boundaries for memory search.

It failed miserabely and I've also tried to look at the code and it seems a bit too messy and distributed to get a grasp of its moving parts.

https://github.com/openclaw/openclaw/issues/88009

[true laurel]

Finally my agent no longer disappears into a black void for 20 minutes 😂 Thanks!

I already made local edits to it:

• Stopped progress mode from erasing the thinking process when the process completes
• Stopped /stop from deleting the thinking process message

And I will probably remove the italic from it too, and introduce an extra linebreak at the end of each comment.

I know that makes my local progress behave more like Block or non-streaming, which is what I actually wanted, but your work only applies to progress mode, so I will just locally hack it for now. 😛

[void bramble]

Tweak in it for a while and find what works best. Maybe leaving the reasoning stream behind after final output is the right move! I dunno, I’m still exploring too.

[true laurel]

I use the reasoning stream to know if I need to panic /stop my agent from doing something silly 😄

But I also use it as post-implementation evidence to follow its process.

[clever narwhal]

Hi all. Curious if anybody is currently working on simplifying the many config options and where we are on that at the moment. Been out of the loop for a few days, but after that big push to make verbose generic across providers I know some things were affected in the process

Is there any particular plan or agreed on path forward for this yet, like having a few simple options with regard to tool call visibility

[shy rain]

There have been a few PRs tidying this up, including by myself, dirtdiver and someone I don't know.
Tool call visibility will be configurable in two ways. 1 - streaming... the tool calls show up in the reply one by one, in a limited-length scrolling window. Config sets on/off, max number of lines, and amount of detail. In addition, /verbose leaves every tool call in a separate 1-line message below the reply, never deletes them. You can choose to have one or the other, or both.
--- to be clear, this is exactly how the docs says it works. Except it hasn't worked like that for me for quite a while. idk when the PRs will hit releases but they are all merged, so should be soon.

[clever narwhal]

Excellent thanks for the summary

[left mango]

Vincent, I just read that you've joined the OpenClaw Foundation. This is great news, big congratulations!

Or clawgratulations

[true laurel]

Codex sometimes appears to stop with a "Codex stopped before confirming the turn was complete. Some work may already have been performed; verify the current state before retrying." message in Discord.

OpenClaw logs show turn.completion_idle_timeout: Codex emits a completed item, sometimes a final-ish assistant message, then no turn/completed arrives. After 60s OC aborts/releases the lane and simply prints that error into chat.

Side-effect wording appears when tools ran. Has anyone else seen Codex app-server miss/withhold turn/completed after rawResponseItem/completed / item/completed? 🤔

[worldly tusk]

Which version? I just compled a long conversation, that was compacted in the middle, without problems on codex. I am on main

[true laurel]

Latest main, I've seen it first appear a few days ago. It's relatively rare.

But when it happens I see it happen 3-4x in a row

[pastel owl]

I am trying to fix this bug currently

Been happening with me on one topic regularly others are fine

Right now also digging Codex logs and jsonl

[true laurel]

Yeah it's hit-or-miss, I haven't found a reliable repro with it yet

But it is annoying, because my agent keeps accusing me of interrupting it 😂

[pastel owl]

It's basically because of a single large tool action that doesn't get completed in 60s

[true laurel]

Are you sure? I'm seeing a lot of short tool calls around that time 🤔

[pastel owl]

Testing a new patch
Will let you know

[warped tide]

Who's in charge of MCPorter?

Or, does the SecOps team cover all the openclaw repos? Or different teams for each project?

[jolly wolf]

Secops does the org

[warped tide]

I've got some findings to share. Who is Point of Contact for SecOps?

[jolly wolf]

https://github.com/openclaw/openclaw?tab=security-ov-file

[glad echo]

https://github.com/openclaw/openclaw/issues/87744

[untold pollen]

Claude CLI

[worldly tusk]

reducing codex context by deduplicating openclaw skill list: https://github.com/openclaw/openclaw/pull/87788

currently, the openclaw skill summary block is repeated each turn, burning tokens all the time. skill list is a stable content that doesn't update much, so this duplication is just useless bloat. especially if one has many skills. this PR moves openclaw skill list (as well as the memory pointer) to collaboration developer instructions, that do not reinject stable content. codex skills seem to be working similarly.

Merge readiness
Overall: 🦞 diamond lobster
Proof: 🦞 diamond lobster
Patch quality: 🦞 diamond lobster
Result: ready for maintainer review.

[true laurel]

Are there many more things just getting duplicated? 🤔

[worldly tusk]

Yes. Working on one more PR

[charred crater]

Thanks for grinding on this stuff @worldly tusk, if you don't mind a quick Q: what happens when skills are updated? Does the operator have to /new or can the dev instructiosn refresh with updated content and cache bust Codex? Im kinda naive here

[worldly tusk]

Should skills update, they are reincluded

(After my PR)

[true laurel]

That sounds how it should be 👍

Save Our Tokens 😄

[worldly tusk]

We do not have a way to selectively prune the context, so unfortunately the previous version still hangs on (at least till compaction). That also ensures cache stability

[cinder bolt]

Anyone have any time to look into batching embeddings for the memory search.

I tried but lacking access to any premium models yielded me no result.

https://github.com/openclaw/openclaw/issues/88009

[void bramble]

wow, ok. OpenClaw 2026.5.27... is fast. Neat!

[shell dawn]

yeah it is

[worldly tusk]

The latest beta is even faster

[charred crater]

it really is, spent a ton of time last week optimizing my setup then the 5.26/5.27 upgrade just poured nitros all over it, feels incredible

[pastel owl]

Thanks for the report

The fix is done ✌🏻
Sharing a summary soon

[glad echo]

Thanks for fixing it! It wasn’t my report, I only added some of the cross channel proof (I use discord).

[crude matrix]

is openclaw 2026.5.28-beta.2 actually released? I tried installing it but it does not get installed, tried teh openclaw update beta and the all in one script.

[void bramble]

I might change my GPT5.5 to low/fast, just to speedmaxx this as much as we can

[shell dawn]

the npm package hasnt been published yet - keep an eye on this url https://www.npmjs.com/package/openclaw?activeTab=versions

[fervent pilot]

hi

[fallen gate]

https://giphy.com/gifs/giphyqa-OA1CDoCiAR48E

😠

@jolly wolf where gifs

[jolly wolf]

Hermit is down bc I’m working on things rn

[shell dawn]

you are like back to the future fading into nonexistent without your gif access

[fallen gate]

it's my love language, what can I say

[pastel owl]

Following is the fix for repeated mid turn errors for codex app-server when a long running thread would be killed by turn watchdog after a 60s timeout

Now the PR patches the watchdog to read native Codex response deltas so a lengthy native stream while it's emitting deltas can be seen and is considered as progress instead of stall. Also increased timeout to 5min just in case there is a quiet gap before the next completed item.

I have tested with a 30s timeout also locally on an earlier repetetively failing topic
So the below PR should resolve the issue
"Codex stopped before confirming the turn was complete. Some work may already have been performed; verify the current state before retrying."

https://github.com/openclaw/openclaw/pull/87781

@maintainers kindly check

[true laurel]

That is a chonky boi, I am applying it to my local build now

[final ocean]

<@&1503801512294486217> beta is out 2026.5.28-beta3, keep us posted on issues

[shell dawn]

github copilot harness? didnt know they were making such a thing. now that i think about it, it makes a lot of sense that I'm sure will become more clear next week

[tiny tree]

fixed scope? I expected this to reach the beta release but do not see it in the release notes: https://github.com/openclaw/openclaw/pull/87502

[timber beacon]

YES iOS update hit that beta

[granite wharf]

Whoaaaa! What?! iOS?

[fallen gate]

they open sourced it a few weeks ago I believe

[next helm]

Update was smooth. No issues to file 🙏

[timber beacon]

some good ui/ux work and realtime voice udpates

[shell dawn]

i see where this is going. its strange to be excited about a microsoft project

[granite wharf]

Damn and I built my own already :/

[shell dawn]

do you have to build from source? or is there a packaged installer

ive built it a few times, its a bit of work

[timber beacon]

Have to build for now

[warped tide]

https://github.com/openclaw/gogcli/pull/652 - Overall: 🐚 platinum hermit
The PR adds --enable-commands-exact / GOG_ENABLE_COMMANDS_EXACT exact allowlisting, wires it through command enforcement, and adds focused tests.