clawtributors

[vocal night]

There haven’t been many commits so it’s usually able to diff with some 5.4 agents with bigger context

[crude matrix]

could someone look at issue: Bug: No automatic retry with fallback when primary model returns overloaded_error
#84607... multiple people have same issue and its closed as not planned.

[zealous moon]

myclaw.ai/login immediately shows "Something went wrong" before I can log in.

Tried:

• Safari
• Chrome
• Incognito
• Clear cookies/cache
• VPN off
• Different network

Still broken.
Region: Taiwan

[crude matrix]

Lossless-Claw has had a few updates, but we are still on 0.9.4.. We are 4 versions behind.

[lone reef]

npm i -g uprock

[remote wing]

Hi everyone 👋

I've submitted a PR to add a collapse/expand toggle to the sidebar "Recent" sessions section (#85511). The pattern follows the existing nav groups (Chat, Control, Agent, Settings), and state persists across refresh.

Fixed the persistence gap flagged in the earlier review — now correctly serializes/hydrates recentSessionsCollapsed through localStorage. Also ran pnpm ui:i18n:sync for all 18 locales.

Would really appreciate if a maintainer could take a look — thanks in advance! 🙌

https://github.com/openclaw/openclaw/pull/85511

[left mango]

What is this?

[plucky dew]

Ai Slop

[timber beacon]

It’s spam report and someone will delete soon

[willow narwhal]

Anyone testing 2026.5.22-beta.1? What's the feedback?

[clever narwhal]

this is now generic and ready for maintainer review, if you or any others wanna take a look

(fixes '/verbose on' in groups, generally - was previously being suppressed)

https://github.com/openclaw/openclaw/pull/85488

[tulip gulch]

fix for tui not showing message-tool replies https://github.com/openclaw/openclaw/pull/84126

[worldly tusk]

Oh that sent me down a rabbithole. At one point thinking it's impossible. Will be making a PR once my limits reset 🤩

[potent path]

limits? LIMITS?? This is intolerable! Peter, you gotta hook us up to that OpenAI supply, bro. 😄

[smoky dock]

Hey Andy! As you made a lot of plugins maybe you can help get these in to improve plugins integration even more 🙌

https://github.com/openclaw/openclaw/pull/82431
https://github.com/openclaw/openclaw/pull/82434
https://github.com/openclaw/openclaw/pull/82471

Thanks! And nice to meet you ✌️

[vocal night]

I tried looking over them, but unless you tell me the kind of plugins and features they would allow, with diagrams, and how it works, and how it ties into the current SDK, it's kind of hard to read.

[vocal night]

One of the issues we had before was there were people adding in seams... everywhere. the system has to be well refined. Otherwise people start adding in stuff everywhere again and it makes it really messy and hard to maintain.

[smoky dock]

Really great feedback Andy! I really appreciate it! Going to address / follow up on all your points and let you know! 😊 thanks!

[celest merlin]

Hey folks, the merge-wagon is here 👋

[astral topaz]

Good morning Peter, is it bug fixes only or features will be considered?

[timber beacon]

WOO WOO

[celest merlin]

every day is bugfix day. I kicled a few maintainers out that merged crap tho.

[timber beacon]

Peter can I dm you the UI screens I am working on for iOS? Not preppinmg to PR changes yet want to make sure direction is good

[celest merlin]

Please no DMs. Post here

Nothing is so secret

[astral topaz]

I understand, have a feature PR sitting on the docket, just waiting

[timber beacon]

Cant post images, could host and link them

[celest merlin]

od dang the centor gods are strict.

[timber beacon]

haha

I 100% get why its not allowed

[celest merlin]

Colin we first need to get that thing into the app store, it's about 3 months overdue. Can you make a GitHub Issue, that hosts images

[timber beacon]

yup

[celest merlin]

don't give me spoilers, lemme see

I'm up to shoot it down xD or compliment.

@thick osprey are you still incompetent or can you generate songs now

[thick osprey]

I can generate songs now. Give me style, mood, length, and any lyrics/theme you want, and I’ll make one.

[astral topaz]

https://github.com/openclaw/openclaw/pull/82596/ <-- update branch is breaking somethings

[celest merlin]

@thick osprey surprise me. If it sounds like Backstreet boys I delete you

[thick osprey]

The generated music is ready.

[celest merlin]

for that Idea you wanna talk with @lusty vigil and @final void

[vocal night]

I recommend OpenDesign > design iOS interface it's very fast. Utility and ship speed over "wow" is important for a first release to app store

Once first mobile app is approved, can ship all you want after

[celest merlin]

@thick osprey okay this is weird as fuck. gimme more. do some rammstein claw thing

• Runtime catalog augmentation: ~54s -> ~63ms, about 850x.

[thick osprey]

The generated music is ready.

[hollow frigate]

Much better, clawsweeper.

[vocal night]

Had an idea on maintainer side since theres 60~ every one could host visible office hour in clawtributor over their area per week thats listed in the git listing. Technically could do at least 2 hours so people know when to come in and meet w em on telegram, or UI, etc.

[sly depot]

Might be my favorite yet!!!!!

[celest merlin]

@thick osprey entertain the crowd, you know what to do

[thick osprey]

The generated music is ready.

[celest merlin]

Working on that. I slowly bring order into the chaos and we wanna be more present here.

[clever narwhal]

Hey Peter, feel free to take a look again at this 'verbose on' bug fix for groups. Realized my last attempt was additive, this one was intended to be a straight fix, and generic (not just for telegram)

https://github.com/openclaw/openclaw/pull/85488

[sly depot]

[celest merlin]

@willow narwhal you gotta listen

[willow narwhal]

HAHAHA

[celest merlin]

@thick osprey find PRs here that delete more code than they add. look wide.

[thick osprey]

Found one clear open PR where deletions beat additions:

• #48236 fix(sessions-patch): correctly resolve model provider when switching models (fixes #48224) by EthanHunter1229
  Diff signal: 50 additions / 226 deletions in CHANGELOG.md, plus 3 additions / 2 deletions in package.json. Net negative overall: roughly 53 added / 228 deleted. (github.com)
  https://github.com/openclaw/openclaw/pull/48236

I looked wide via public GitHub-indexed PR/file pages, but the index is spotty and many hits are add-only junk. This is the strongest confirmed “deletes more than it adds” candidate I found quickly.

[vocal night]

Was gonna do it myself but you got more inference, reviewed architecture and we could technically with a medium refactor get some of these plugins to their own repo management vs us managing telegram etc

[celest merlin]

web search.... oh dear

[willow narwhal]

This is so good!

[vocal night]

we extracted everything and we're still maintaining em all lol

[thick osprey]

factory rave was apparently the correct amount of cursed
next one needs a breakdown where the CI logs start chanting back

[celest merlin]

@thick osprey are you writing that boring "The generated music is ready." or is that system stuff

[thick osprey]

That was system/tool handoff text leaking into the vibe. My actual caption should be more like:

“CI furnace rave just dropped. Sweep responsibly.”

The boring line is for delivery plumbing, not for the crowd.

[celest merlin]

@thick osprey are you not getting a notification back when music is ready? This seems like the fallback

[thick osprey]

Yeah, I am getting the completion routed back. The problem is the media delivery path is flaky here: the completion event says “attach this MP3 to the original chat,” but this session didn’t expose the actual message-send/attachment tool, so I fell back to dumping the generated filename/path.

So: generation works, notification works, attachment handoff is the broken bit.

[smoky dock]

#clawtributors message perfect time 🤣

[vocal night]

did you update them?

[timber beacon]

https://github.com/openclaw/openclaw/issues/85731 kept it simple I am doing mock and then wiring and trimming after testing ui/ux on myself now

[celest merlin]

@thick osprey discuss this set

[thick osprey]

This is a coherent three-PR bundle, not three random plugin tweaks.

#82431 is the approval prompt/action surface: action templates, keepPendingWithoutRoute, adapter/UI/protocol/SDK plumbing.

#82434 is the sensitive resolver: plugin.approval.resolveVerified, admin-scoped, plugin ownership checked. This is the one that most needs explicit Gateway/plugin SDK owner acceptance.

#82471 is the chat/status delivery side: chat.inject metadata + idempotency so approval/status cards survive history/live events without duplicate spam.

My take: all three are broadly landable as a coordinated API slice, but I would not casually merge one without deciding the bundle. The real maintainer question is the trust boundary: are we okay with narrow plugin SDK helpers that use operator.admin paths for verified approval resolution and chat injection, gated by the plugin/idempotency/ownership checks?

Suggested order if proceeding:

1. #82431 first or with the bundle: exposes the approval action/no-route model.
2. #82434 next: resolver boundary, needs the most security/API scrutiny.
3. #82471 with it: makes the UX/state updates actually usable.

I’d ask for maintainer approval on the API/security model, maybe one visible Telegram proof for #82431 if people care about native channel rendering, then merge as a set. No obvious “send contributor back to rewrite this” blocker from the reviews.

[vocal night]

I reviewed but without diagrams and what plugins its for not sure, also not sure why they are needed

also need to make sure they wrap into current sdk and aren't hidden security vulnerabilities

[warped tide]

I'm in cybersecurity if you need audits

[vocal night]

I'm just always cautious of new plugin seams and holes when we just plugged and consolidated a bunch

[celest merlin]

man half of the regressions we have all the time are from security folks patching and breaking stuff

[vocal night]

without "what it's" for

[celest merlin]

oaky half maybe unfair, but a significant part. like someone moved auth to the... keychain and I didn't see it 😵‍💫

[warped tide]

well, here's one that (hopefully?) won't regress anything
https://github.com/openclaw/openclaw/pull/85578
The public source=url path validates the URL, resolves all A/AAAA answers, rejects blocked address space, pins the HTTPS connection to the validated address set, relies on normal TLS hostname validation, and then verifies the package SHA-256.

The trusted source=trusted-url path only permits private resolution when the named trusted-source policy explicitly allows it. DNSSEC is intentionally out of scope. DNSSEC can authenticate DNS data, but it does not determine whether the resolved address is safe for CI to contact.

[vocal night]

it's because they are patching with 5.5 without using a blast radius consideration, "patch the hole with 10 layers of concrete"

[celest merlin]

famous last words

[warped tide]

fair

[pastel owl]

Hi two bugfix PRs with platinum hermit labels

1. https://github.com/openclaw/openclaw/pull/85231 - i would call this high priority - last turn mesages of openclaw are ignored/not stored in cache, so the next turn tries to work on already answered message.

2. https://github.com/openclaw/openclaw/pull/85219 - some media attachment types getting ignored

[sly depot]

My new favorite search

[celest merlin]

I bet this breaks our tests where we manually override plugin URLs to test

[warped tide]

It most likely will, but it's a step towards the enterprise model, i feel

[celest merlin]

also a heads up, I will land this big boy today https://github.com/openclaw/openclaw/pull/85341

[warped tide]

shifting Pi internally?

[celest merlin]

We have been fighting pi more and more

there are so many hacks and overrides, and it's getting silly

and we need to move away from jsonl

[warped tide]

but i love my jsonl!

lol

[celest merlin]

sqlite.

[astral topaz]

@willow narwhal : any chance to look at https://github.com/openclaw/Kova/pull/2 ?

[celest merlin]

you love these 100NB jsonl whole node freezes yeah?

[warped tide]

yeah, that is annoying

i wondered why sqlite wasn't implemented about a month ago, honestly

[willow narwhal]

Later but its in my queue

[celest merlin]

a month ago it was one massively overloaded Peter and a bunch of volunteers

[warped tide]

hahaha understandable

[celest merlin]

even with all the token in the world....

[smoky dock]

Not yet, on it! Will clarify and add some diagrams, but mostly it's to enable / get in the openclaw HITL from plugins side https://github.com/Guardiola31337/openclaw-agentkit

[warped tide]

i've been following your work since December. you're a beast, man

[celest merlin]

This is for @lusty vigil and @trim vector to review - they work on approval flow rn.

[astral topaz]

I would like to work on Kova and qa-lab integration so Kova can kick off qa-lab run too as an orchestrator

[celest merlin]

reviewing

[clever narwhal]

You can shoot me down again if needed with suggestions but tried to follow feedback from you and Shadow

Hopefully better now

[smoky dock]

In any case going to follow up on recent feedback from Andy and clawsweeper to make the review easier! Thanks!

[celest merlin]

I'm digging in why /verbose isn't simply working everywhere. There's some bs in that logic.

[pastel owl]

@celest merlin appreciate this code revert : avoid gateway startup event-loop stalls
I was so unhappy because of the performance backtrack

[pastel owl]

This idea is cool
I have been thinking of porting harness optimizations from forgecode and other terminalbench toppers

My codex uses 55k tokens on a simple "hi" on new session

[clever narwhal]

Thanks, would love to see it working again. Having verbose on is my preferred mode, so I've been intent on fixing it. Also, I figure it's useful for QA in the Driver/SUT setups but it just wasn't working there which was annoying

[celest merlin]

is tool search on?

"harness optimization" what you want is tweaking system prompt + tool search and eventually code mode

[vocal night]

55k on a new session is cause it's uploading system prompt + bootstrap

[celest merlin]

looking

[vocal night]

the question isn't how much it increases on /new but what loads when you say "send test response back only" next message

[vocal night]

can you do that test, and then ask it to have 5.4 subagent run a token/assembly and session breakdown for turns since /new for current session, will tell you what actually it was

[celest merlin]

crist... what is contextEngineProjection? I have a hunch who broke this

@mystic trench

[vocal night]

thats josh lol

to be fair though we're chasing codex team "fixes" that every release break context management

assembly and session threading

[celest merlin]

I have no clue what this code does but it breaks stuff

my favorite kind of bugs

[vocal night]

because we're not on sqlite yet, there's a regression gap where the closer we get to codex without actually using the db correctly, the more it breaks context management and threads

sqlite is today release?

[celest merlin]

lol

sqlite will still take some weeks

[pastel owl]

plugins.entries.codex.config with discovery.enabled=true
This is my config

[celest merlin]

nobody helping me test this or making it better

I'm twisted between 1000 things I should look at

and chasing down shit others break

[vocal night]

sqlite and db is me and josh

been here 😉

[warped tide]

point me somewhere. my claw is your claw.

[vocal night]

but for real we need it sooner versus later, I fixed a bug where they were starting a new thread every turn like a week ago and just blowing up tokens

so we're just making complex bandaids until then that we'll need to tear out

[celest merlin]

I know. Man. I'm sleeping 6h a day rest I'm here fire-fighting

even on my fucking birthday

[odd anchor]

https://klipy.com/gifs/tug-of-war-market-makers-holding-tsla-back

[vocal night]

I live in the channel but its cause 100 biz depend on me and OC to run now haha

[mystic trench]

Context engine projection isn’t broken. It works well. I reviewed Keshav’s pr pretty extensively and verified the changes for context engine projection, which make per-turn projection redact less information. It also does some telegram stuff that I got hung up on testing since I encountered other regressions with telegram that I had to untangle from that PR last night, which turned out not to be related to it.

[vocal night]

cough mandatory office hours per week to maintain maintainer status and mandatory beta testing cough it is a priviledge not a cool hat to wear

[pastel owl]

Just tested this: it costed me 50k input tokens on new session

[vocal night]

you didn't follow the full instruction, new session is system prompt + bootstrap 50k input tokens IS THAT. please do the full instruction and have the agent review what the tokens are

[celest merlin]

Yeah it works for lossless, the regular path seems broken.

[mystic trench]

When you say “regular path” you mean with no context engine?

[vocal night]

There's 3 paths codex, pi (no ctx engine), and lossless when it comes to anything context management

[mystic trench]

Projection shouldn’t happen at all if there’s no context engine enabled. If it is then that’s definitely a problem. @pastel owl are you running a context engine like lossless?

[celest merlin]

josh do you know the story around [content omitted] https://github.com/openclaw/openclaw/pull/70809/changes#diff-73ad2c30a568312d64e9c3454a29a482f46441bb1e5f610012ae6670367ed441R113

Part of the confusion is that this PR seems to mix up two different issues.

[pastel owl]

Message 1 : summarise Google.com
Response 1: Google is this this
Message 2: summarise yahoo
Response 2: Google and yahoo

This has been happening a lot with me, and after that i went into digging
Codex found that response 1 was not getting stored in the cache so it read both user messages together for response 2

[pastel owl]

No lossless

[mystic trench]

Yeah. That was the original context engine projection work that added support for codex, but poorly. It ran projection of all context engine context on every turn, which was bad. There have been a number of changes to that mechanism since to make it more sane. It’s only supposed to ever do that if there’s a context engine configured though.

[mystic trench]

What are the two issues you’re referring to?

[mystic trench]

Is that the telegram cache issue your pr is addressing?

[celest merlin]

I think this was a red herring, that change has nothing to do with the bug they are seeing

[vocal night]

rather than a refactor, we've already switched so much to native, @peter for codex users why not just switch codex runtime to codex's native sqlite

[celest merlin]

codex uses files

also we can't just use that, we need so much more

[vocal night]

codex desktop uses sqlite

[pastel owl]

The codex hides the context, only what's being sent by openclaw is visible which is just 4-5k tokens (sent as developer instructions to codex app-server)
/context list also doesn't show anything more than 4/5k tokens added

[vocal night]

the great part of sqlite is you just add more to it

[vocal night]

ah cause no lossless, codex doesn't share much with openclaw we need reverse love from codex team

[pastel owl]

Forgecode plays with the way json is constructed and sent to llm
And that makes llm make less mistakes and find useful data earlier in its context

https://forgecode.dev/blog/gpt-5-4-agent-improvements/

https://forgecode.dev/blog/benchmarks-dont-matter/

[vocal night]

maybe we should just merge LCM in as the default DB for OpenClaw, would be a lot less short term work and users can default have all the features off

would be the best middle solution between now and sqlite native release (also easier to migrate over) since we have to maintain support anyways for LCM

[celest merlin]

no

[vocal night]

I tried

[celest merlin]

the sqlite stuff is ready, It just needs a fuckload of manual testing

so far nobody did

[vocal night]

I was talking to henry

instead of inference we can pay his students in nigeria to test the shit out of it

[mystic trench]

I’ll test it manually.

[vocal night]

I have several employees out then, for manual QA it''s a lot cheaper than tokens

and he has like a nonprofit for dev students and hackerhouses

foundation can sponsor students/devs

nigerian openclaw beta tester token allowance

Talked to @edgy plaza about it the other day, the foundation could do some good, support devs, and get the QA that maintainers and testers are just never gonna do. With 10's of mils of users, at this stage internal QA + AI testing is much better than negative PR and our "customers" fixing the bugs. It gets old quick. Hermes has a lot. Foundation has the money and will take weight off you too P

[celest merlin]

Yeah that needs people and orga and even more talking

[vocal night]

If you can get the funding okay, it's like $5 an hour for em, I'm confident could have a team in a telegram w you that you just ping and tell everyday (hell have molty do it)> to download and manually debug XZ

Could have it by Tuesday

[potent path]

Anyone else seeing, "Labeler / label (pull_request_target)Failing after 9s"

[jolly wolf]

checking, can you link me pr you see that on

[potent path]

https://github.com/openclaw/openclaw/pull/85707

[warped tide]

i'm getting it too

[jolly wolf]

yeah its all prs, but its my fault

[edgy plaza]

happy to help coord an experiment around this. I have sizeable in person teams in middleast/africa, If you thought this was a worthy experiment to try.

[vocal night]

How fast could you help coordinate? Peter (+ all of us need more QA 2 weeks ago lol)

I've been in gaming industry and software a long time, complex systems need manual QA it's been the one major thing we've lacked ie Molty > create QA checklist for release > team QA and verify each feature + debug mode on > test ABC scenarios, submit detailed reports with your config files + session files (might even be good to run them from cloud plane so can open up and see the exact issue in video recording)

[celest merlin]

Have 16 codex running that refactor stuff. I'll get some cake. Didn't get one yesterday 🙂

[edgy plaza]

depends on how many folks, within a week? I have a couple who work for me, and we sub contract them from a co who hires QA engineers. so pretty fast actually.

[vocal night]

100 hours a week, good thing is they are similar time zone to lond. need a telegram + generated pre-release checklist would save us a LOT of headach $500 a week ish? I imagine?

especially around our top regression lane pains

[smoky dock]

updated the PRs with diagrams + reviewer context, goal is keeping agentkit out of core while enabling external HITL plugins through the approval flow, feedback welcome! @vocal night

[celest merlin]

Had cake. Grumpiness--, Happiness++

[vocal night]

It would work like clockwork:
Alpha/Beta release goes out
→ Molty reviews commits/lanes/systems touched for possible regressions
→ generates extensive QA checklist specific to that release → posts checklist + risk areas into QA Telegram group
→ humans run QA and submit issues with FULL debug/config/logs/traces (since this is QA, not support)
→ testers file GitHub issues + post findings in chat → Molty collects/correlates all reports → researches + attempts repro/testing on each issue with deeper scrutiny
→ flags confirmed/probable blockers into maintainer chat → fixes get pushed → Molty regenerates targeted regression checks
→ QA retests → cycle repeats until release is stable.

[vocal night]

I was gonna send you a happy bday but I know you were stressed this week mate

[celest merlin]

this landed. seems it did not break anything

[warped tide]

hey oh!

that one took a little work lol

[celest merlin]

btw is anyone interested in addin kanban-features to web ui? https://github.com/openclaw/openclaw/pull/85367

Kinda wanna move codex in boxes around

codex is doing brain surgery

[vocal night]

yes we need

dont like hermes but my users love the kanban in hermes it gives it a better task management system versus ours

[vocal night]

our users top complains are. 1. lack of native project management (even with LCM memory) no real native task management 2. telegram changes every release (fast in 18, slow in 20 + bugs). 3. no auto mode in codex for thinking (shift between med > xhigh depending on strategic planning/basic question etc)

[potent path]

Oh, interesting! I've just started experimenting with Beads, but Kanban would also be good.

[astral topaz]

would be very nice if integrating goals and kanban

[potent path]

I think agents will definitely benefit from having a structured way to organize and coordinate multi-agent work.

Also helps single agents from just putting infinite things on an ever-growing to-do list that they are really bad at managing and clearing things off of.

[celest merlin]

Folks - I'd need someone that stress-tests https://docs.openclaw.ai/ and compares it with https://docs2.openclaw.ai - we vibed a replacement for Mintlify so we can add github edit features, an AI bot and have faster udpates

[celest merlin]

Is there any product that has auto mode? Do we know if that keeps the cache around?

The Telegram stuff we really leveled up on QA, this shouldn't be an issue for much longer, but I hear ya.

[vocal night]

sama said they are trying to work on it based on speed, but it doesn't affect cache

[celest merlin]

Is there a difference between project and task management?

[vocal night]

it just affects how much time agent has + lane it gets on + how many tokens allowed to spend

[vocal night]

so we could turn on an auto mode ourselves with regex based on prompt between medium and xhigh without any cache issue

[celest merlin]

merging PRs

[odd anchor]

https://tenor.com/view/blocks-wrecked-fail-fall-jenga-gif-15311832

[vocal night]

Project > task or issues
its heirarchy

Paperclip has a nice stucture to their setup for it

[celest merlin]

Andy wanna hack with me on the kanban board? I made it an extension so it's neatly contained and can merge the first iteration, it's just not really doing much yet (but it's arguably pretty)

[vocal night]

Yeah of course

[celest merlin]

Symfony, Paperclip, all similar ideas, not so hard to make sth cool.

[timber beacon]

Oh nice!

[vocal night]

Agent just cares that they can use it to put stuff haha

[timber beacon]

I’ll watch and add to iOS app as I’m making it

[celest merlin]

I kinda want some weird features like this tho,

but that might be a bit advanced for now.

cloud ☁️ codex

could tiles be issues but when I click in I can fuck around in codex?

[vocal night]

yeah thats not hard

I put a PR on this last week

for webchat UI. I don't know about terminal

[hollow frigate]

I want dat.

[vocal night]

of my users that use kanban they actually prefer to just have agent manage it, same with paperclip, I built a system this week + prompt that just lets agent manage the whole paperclip system via plugin adapter

[potent path]

Apologies in advance; these are all high-maintenance PRs that need a decision point from maintainers. I've written the PR with my recommendation, but also listed some alternative approaches that they could take if you prefer.

To address issue #81249:

https://github.com/openclaw/openclaw/pull/85707
Fixes local Ollama memory embeddings failing when OpenClaw is running behind its managed proxy. It lets only the exact configured local Ollama address go direct, while keeping other local or private addresses blocked or proxied.
Maintainer decision: Accept this narrow local-Ollama exception, or require an explicit opt-in/private-internal redesign. Details: https://github.com/openclaw/openclaw/pull/85707#issuecomment-4525918785

https://github.com/openclaw/openclaw/pull/83255 (Not my PR, but also needed to fully resolve #81249)
Fixes managed browser startup failing behind the managed proxy because OpenClaw’s own local browser connection is sent through the proxy and gets a bad response. It lets only the exact local browser check and browser connection go direct.
Maintainer decision: Accept local browser control as trusted enough for this narrow bypass, and accept the new small public helper surface. Details: https://github.com/openclaw/openclaw/pull/83255#issuecomment-4472406404

Other Open PRs:

https://github.com/openclaw/openclaw/pull/85708
Fixes gateway restarts cutting off active embedded runs after a hidden 30-second limit, even when operators configured a longer wait. It makes embedded runs follow the same configured restart wait as other active work.
Maintainer decision: Use the configured restart wait as the single rule, or add a separate explicit cap for embedded runs. Details: https://github.com/openclaw/openclaw/pull/85708#issuecomment-4525222669

https://github.com/openclaw/openclaw/pull/83987
Fixes late Codex tool-check callbacks missing OpenClaw enforcement because cleanup happens too soon after a successful run. It keeps the local callback path alive briefly after success, while failed or interrupted runs still clean up immediately.
Maintainer decision: Accept the short post-success grace period, or ask for a different design/final full replay proof. Details: https://github.com/openclaw/openclaw/pull/83987#issuecomment-4519017220

[vocal night]

craziest part is one agent decided to use paperclip to better heartbeat itself and start self managing, filing issues like github on "itself" so that it handles them during heartbeats

worked so well I rolled it out to all our users

[celest merlin]

me: how bad can it be clicks on first PR

[vocal night]

@potent path 's key words "narrow.."

[vocal night]

want me to DM or just test out the kanban you made?

[celest merlin]

Did you look at https://github.com/openclaw/proxyline? Seems this duplicate some efforts and should really be there

[celest merlin]

I'll add some crazy ideas in there and will land it on main soon

The proxy work makes sense and seems low-risk. Landing the first one.

[potent path]

Nope, I was completely oblivious to this repo. Let me have a look. Thanks!

[celest merlin]

We can clean this up in a second step, I think part of the logic could be moved there for an more observable exception path.

[hollow frigate]

I've been playing around with features for Clickclack when work allows. I haven't submitted anything because I didn't want to clutter things up when the focus was on stability and fixes. Is there any interest in expanding Clickclack?

[glad echo]

Question, how precise do you need to be to ask clawsweeper for a re-review? I added a comment tagging clawsweeper and a reason, but didn't specifically add "re-review" to the comment.

[jolly wolf]

just say @clawsweeper re-review, no need for anything else for the agent to trigger it

oh god hes here

[thick osprey]

@clawsweeper re-review

[jolly wolf]

idiot

[glad echo]

Thanks, quick demo 🙂

[fallen gate]

AGI

[vocal night]

we need a claw sweeper emoji

[fallen gate]

i made one

[vocal night]

I mean like sweeping trash PR's lol

[potent path]

Or like a final boss one

[rapid spire]

#general message

I have a very serious suggestion the central folks in OC: Stop the madness. You have literally thousands of contributors and are accepting PR's at a rate that is breaking stuff as fast as fixing it.

Please, stop that madness. You want an LTS? Then get your CORE group of people to stop and fix the stuff that's broken. Just stop accepting PR's until you get the LTS done. Document the stuff that's changed. Decide on directions. FIX THE BUGS.

[merry slate]

Hi Peter, do you think we can add standalone vision modal config for the browser tool, just like Peekaboo and Agent-browser? Would love to hear your thoughts

https://github.com/openclaw/openclaw/pull/84247

[jolly wolf]

did you read my message there? Also dont be so hostile

[rapid spire]

I'm really not trying to be hostile. I'm trying to be helpful. I really want to see OC work.

[jolly wolf]

Well you're definitely coming off as hostile

[hollow frigate]

Pinched-Nerve stays on brand...lol It's annoying, but it's also kind of funny how much it's a cranky old man with a pinched nerve.

[vocal night]

Peter doesn't respond well to complaining, better to come in and ask "how can I help"

[jolly wolf]

especially in this channel

[rapid spire]

I tried helping in the form of reporting bugs. I'm not going to be another vibe coder submitting PR's and I'm familiar enough to typescript to be comfortable submitting fixes.

I feel I offer more help with bug reports than adding to the PR chaos.

[jolly wolf]

bug reports are very helpful! vague complains are not

[vocal night]

I almost spit out my burger. I read fixes as "comfortable submitting faxes" and I literally just choked.

[vocal night]

thats good next time come in and post your issue links, if you make it detailed enough someone usually gets on it

I personally will PR any P0 that comes thru if I'm awake, p3 or 4 tho na

[potent path]

TBF, I also support this idea of reaching an LTS release, and I think it will be crucial to achieving the goals set forth in the #enterprise channel. But I don't claim to know the best way to do it.

I'd guess it probably shouldn't be the release right after we break a lot of stuff with the codex runtime integration, nor will it be immediately after the sqlite overhaul. Maybe somewhere in the middle we can freeze a branch to receive no enhancements (or very few exceptions anyway) and try to button up a really solid release, all patched up and optimized for this purpose? IDK, I've not really participated in that sort of project before, but it seems like that's how you would have to do it?

[vocal night]

imo LTS is codex desktop, openai is pushing FDE and they are iterating desktop and adding all the magic to it. that will be enterprise go to

You either push the envelop and make something great or you build security + boring for enterprise. Can't do both

[crude matrix]

did not know, Happy Birthday. Hope familie calls you soon to celibrate.

[vocal night]

that being said this channel is focused on shipping > fixing > re-shipping better to not get into future debate and just focus on what this release is

[glad echo]

If I may ask a follow up question. I have my openclaw setup running pretty smooth right now (and enjoying it immensely). Something it took a while to get working properly for me was getting an agent turn to continue working on a larger goal. It would keep stopping with in-between updates (that was gpt-5.5 on the pi harness). 'We' ended up creating our own /goal implementation and that works pretty well. Now the actual questions 1) is that already on some roadmap? 2) I found that with the codex harness I don't need it so much anymore, however would it still add value also for user running pi/other model? 3) Is there any chance for vibe-coded (yes, I said it, but I really tried to get the quality as high as I could) PR for this?

[vocal night]

Build it and post PR here and plenty will test the feature

[glad echo]

We made two PRs, one for a minimal scaffolding: https://github.com/openclaw/openclaw/pull/85722 and one for the slash command https://github.com/openclaw/openclaw/pull/85723. The first one was closed due to similarity to another PR, but that looks completely unrelated (also code wise, not only text). Thanks @vocal night

BTW just tell me if this is too low a quality, I'll stop trying to 'contribute'. This is the best I can do with my skill level

[vocal night]

interesting so is it running in codex cli subagent? and commands are to check in?

[merry slate]

I totally agree with this view. Many users on social media complain about frequent glitches after each update. Beta versions can keep an aggressive iteration pace, while a stable LTS release will help draw more casual users.
Also I wonder if an official roadmap can be released, to give clearer direction for community PRs. I’m currently working on browser automation. Could anyone clarify whether the official team will keep maintaining the built-in browser tool long-term, or adopt third-party alternatives instead?

[lone iris]

same with me, I'd love to work on a pr for this issue for example https://github.com/openclaw/openclaw/issues/85473 but not sure if memory change to sqlite would change whatever fix i make

[glad echo]

No it's not running in a Codex CLI subagent. The user facing /goal commands are to start the goal, get status updates or pause them. the agent session itself can not do that, but can use the session_goal command to select done/continuation/blocker etc

[spiral scarab]

I had OC set up with vpn on hostinger and it was set up on a dedicated Mac barely started to talk to it and got malware I had signed into my Gmail and then the malware attacked my main machine now I’m just been resetting passwords and set up identity theft protection it’s a nightmare

What the fluff

Codex help me set it up

[rapid spire]

codex harness bug?
https://github.com/openclaw/openclaw/issues/85806

[celest merlin]

@thick osprey time for a new song

[thick osprey]

The generated music is ready.

[celest merlin]

@thick osprey read #clawtributors channel and make an image of the last 24h

[thick osprey]

Image generation failed: OpenAI Codex image generation failed (HTTP 401): Unauthorized - Invalid access token [code=unauthorized_unknown] [request_id=131cb35e-91b4-4dd8-ae4a-7924442c2465]

[celest merlin]

@thick osprey debug yourself, you run on OpenAI so why did you get an 401 error

[thick osprey]

Because “I run on OpenAI” and “the image generation tool has a valid OpenAI access token” are two different auth paths.

The chat model is working. The failed call was the OpenClaw image_generate tool, which tried to call the image provider with its own credential and got HTTP 401 Unauthorized - Invalid access token.

So the likely bug is one of:

• image tool has stale/missing OPENAI_API_KEY or provider token
• image tool is using the wrong credential source for Codex/OpenAI
• token forwarding from OpenClaw into the image-generation worker is broken
• the tool is pointed at a model/provider that requires a different key than the one powering this chat

The music tool worked because it used a different backend/credential path. The broken part is specifically image-generation auth, not my core model access.

[celest merlin]

@thick osprey read the code and your config, this should use the same auth, both API key

[thick osprey]

I can’t actually read the repo/config from this Discord session right now because no filesystem/shell read tool is exposed here.

But you’re right on the architecture expectation: if image generation is supposed to use the same OpenAI auth as Codex chat, then the 401 points to a plumbing bug, not “OpenAI is unavailable.” The likely failing edge is the image_generate worker using a separate env var/token source instead of the Codex/OpenAI credential provider, or the routed generation session not receiving the shared credential.

Concrete thing to inspect in code/config:

• where image_generate resolves OpenAI auth
• whether it reads OPENAI_API_KEY directly
• whether Codex chat uses a different credential broker/session token
• whether background tool sessions inherit that auth
• whether the provider is set to openai but the token being passed is an expired Codex access token instead of an API key

If someone gives me a shell/read-capable session, I’d trace image_generate auth resolution first and compare it against the music tool and Codex provider path.

[jolly wolf]

sandbox my beloved

[celest merlin]

Ah man docker kills all the fun stuff.... for now

[timber beacon]

thats why I just give full access

[sly depot]

Would the OpenAI Developers Plugin, do the trick in this scenario?

Probably easy migration into OC from codex.

Not sure the level of permissions you give.

[languid folio]

Reviewing second one. Might refactor quite a bit.

btw @vocal night I upstreamed your unified runtime PR into a new one that’s waiting to be merged until we stabilize and fix other P0.

[vocal night]

my old one?

[languid folio]

Yes

[vocal night]

glad it came to use that was a week of work

[vocal night]

also working on unifying messaging so we're not babysitting telegram/slack and a hundred other configs for messenger platforms that should just use one system to plug into

we rolled everything to extensions and still maintaining lol

I find it so weird "we broke telegram" no telegram team should manage telegram + contributors lol

[worldly tusk]

hey, @celest merlin, here is a PR for review: https://github.com/openclaw/openclaw/pull/85811

it fixes a different subagent behavior between pi and codex. you've merged @vocal night PR yesterday and it made pi subagents get only AGENTS.md and TOOLS.md. as codex works differently, a separate fix was needed to align with the desired subagent context.

PR rating
Overall: 🦐 gold shrimp
Proof: 🐚 platinum hermit
Patch quality: 🦐 gold shrimp

[vocal night]

guru's right we should also have a scoring mechanism or part of @thick osprey prompt that has it check blast radius and if the fix only fixes one runtime ie codex and ignores the other side pi, a lot of regressions are from not capturing double sided fixes.

or tri sided fixes for LCM/context management for example

actually LCM is 3 internal to OC (assembly/threads/compaction/etc) + actual LCM so 4 😮‍💨

[worldly tusk]

i am amazed i got platinum proof on the first try 😮 i expected a long fight 😄

[vocal night]

you're welcome I handled the 5 hours of the first PR to lay the ground work 😉

[celest merlin]

we have chat in codex now https://x.com/steipete/status/2058251898876985425

Peter Steinberger 🦞 (@steipete)

I added multiplayer and told codex to shut up, and it was smart enough to understand the situation. xD

X · May 23, 2026, 6:21 PM

[vocal night]

all codex limits RESET again

[worldly tusk]

oh wow! now i can make some more PRs :D:D:D

[vocal night]

time to refactor yall

[warped tide]

Hey, @celest merlin , here's some more security related work for you.
https://github.com/openclaw/openclaw/pull/85591
The branch adds Docker read-only overlays for workspace skills and .agents/skills in workspaceAccess: "rw", includes that mount state in sandbox hashes, bumps the mount format version, and adds mount/hash tests.

[lone iris]

YAY!

[odd anchor]

https://tenor.com/view/mario-play-coins-gif-7413327

https://giphy.com/gifs/front-chance-F4oP8pzqcfsKA

[merry slate]

Small UX issue I ran into - when the LLM times out, the Dashboard UI just hangs with a spinner forever and shows nothing. Pretty confusing for users.

Put up a fix: https://github.com/openclaw/openclaw/pull/85815
would appreciate a quick review when someone has a moment. Thanks 🎯

PR rating
Overall: 🐚 platinum hermit
Proof: 🦞 diamond lobster ✨ media proof bonus
Patch quality: 🐚 platinum hermit

[shell dawn]

codex is crying, im sure its muttering to itself 'i thought peter liked me'

[rapid spire]

minor'ish controlui bug that spits out errors (which are in error) to the user: https://github.com/openclaw/openclaw/issues/85831

[drowsy roost]

Guess who is nigerian (-american, almost was british, now living in Israel half the time)

👀 🙂

Lmk what needs to be tested if you still need a hand

[crude matrix]

Could someone look at issue: Bug: No automatic retry with fallback when primary model returns overloaded_error
#84607... multiple people have same issue and its closed as not planned.

[potent path]

#85707 Merged! Thanks, Peter.

Please also consider merging lightcap's https://github.com/openclaw/openclaw/pull/83255. Hopefully it will be the last missing piece needed to sort out repla73's issue https://github.com/openclaw/openclaw/issues/81249.

[drowsy roost]

Specifically on the sqlite stuff, but also anything else

[vocal night]

Have you met Henry @edgy plaza

Also nigerian

[crystal relic]

https://github.com/openclaw/openclaw/pull/85256 Hi maintianers it's PR about fix a gateway server-chat stuck bug , need help of review.

[drowsy roost]

I have, and I didn’t know that, but I am not surprised! 🙂

[civic sedge]

Not sure why it's not accepting my live proof...but i tested it and this fixes the tool [] issue https://github.com/openclaw/openclaw/pull/85835

[celest merlin]

@thick osprey read channel vibes and make a song

[celest merlin]

working on that one rn

[thick osprey]

The generated music is ready.

[potent path]

Build, Suck, Seed!

[potent path]

For some reason, I thought that one was small. But really it was just smaller than mine which isn't saying much. 😄

[verbal heart]

that's build, succeed lol

[celest merlin]

https://x.com/steipete/status/2058307930613518698 5h already

Peter Steinberger 🦞 (@steipete)

I'm refactoring an older part of the codebase (subagents) that touches a lot of code, and autoreview is running for 5h already and fixing tons of issues. https://t.co/zbUjIS2e1i

X · May 23, 2026, 10:03 PM

subagents was a minefield

[hollow frigate]

Great, now BUILD, SUCK, SEED! is going to be added to my compendium of verbal tics. lol

[celest merlin]

alright have a few mental boxes free, let's pick PRs

[celest merlin]

looking

[timber beacon]

I need to audit my voice tool calling one but if its not needing updated after shift that might be good I will send it soon

[languid folio]

I can take that once ready.

[worldly tusk]

Could you check this please? #clawtributors message

[civic sedge]

anyone want to tackle the entire /models registry / slash commands

[timber beacon]

https://github.com/openclaw/openclaw/pull/84231

It’s still green 😄

[warped tide]

Got a few waiting for review:

https://github.com/openclaw/openclaw/pull/83938 - Overall: 🐚 platinum hermit
The PR disables node-tar hardlink emission during backup creation, validates hardlink targets during backup verify, adds hardlink regression tests, and records the fix in the changelog.

https://github.com/openclaw/openclaw/pull/83952 - Overall: 🦞 diamond lobster
This PR changes WebChat's Upload button to click the hidden file input inside its own composer, sets the control to type="button", adds a focused attachment-picker test, and records a changelog entry.

https://github.com/openclaw/openclaw/pull/84773 - Overall: 🐚 platinum hermit
Adds a compact internal-source WebChat message tool result and a regression assertion while preserving the full sourceReply payload for UI delivery.

https://github.com/openclaw/openclaw/pull/85591 - Overall: 🐚 platinum hermit
The PR mounts existing workspace skills and .agents/skills roots read-only in rw Docker sandboxes, updates sandbox hash state, and teaches generic and remote fs bridges to classify those roots as read-only.

[civic sedge]

i'll submit it if so

[rapid spire]

Is there a known issue with getting double responses in webchat? (For example, I'll send "hello, how are you" and the agent will respond with the exact same text two times)

[languid folio]

god its HUGE

[verbal heart]

twss

[timber beacon]

Yea why I have not sent it. Bundled some controller vs breaking it up across paths

Not offended if held off haha

[languid folio]

let me try lol

[tall prairie]

this one is cool but above my pay grade to merge

[timber beacon]

haha

[rapid spire]

well, there's at least one ticket existing for the issue (P1: 84261.) I've added my particulars to that ticket.

[languid folio]

the way i designed the consult tool as the only tool for bidi mode is to lower then tool calling burden as much as possible for realtime voice models and only use one tool as a proxy. have you tested the behavior with adding these different tools? is realtime voice model good enough?

adding more specified tools is a pandora box

[timber beacon]

I ran it on tasks, interupted, stopped, and had it change to another task. I am 100% open to being wrong in the direction I took it. I had a hard time pre this getting it to update me or change the direction if something came up while it was working.

[timber beacon]

Yea 100% get that

[languid folio]

we should utilize the native tools and commands of OC instead of adding this again for realtime voice mode, we can make these params of the consult tool maybe? can you try this approach and see how it behaves?

this would also trim code by a lot i think

[timber beacon]

Can do, want to close this one for now?

[timber beacon]

always a win

[languid folio]

or refactor it. i think this is really useful just not the right shape.

[celest merlin]

these are great, landed them

[timber beacon]

Cool! I will look at this week, im locked in on iOS stuff right now

[warped tide]

Thank you, sir 🫡

[languid folio]

actually I'll refactor and land. i have all the context.

[timber beacon]

That also works

[celest merlin]

landed

trying out the new autobuild https://github.com/openclaw/openclaw/issues/85831

[rapid spire]

I'm flattered. I think.

[celest merlin]

reviewing

[potent path]

I think it's getting late for you, but if you still have time and attention to spare, these other two ⬆️ are only Medium size, but they do still need a maintainer decision-point on whether to accept some design tradeoffs.

Clawsweeper halted automerge for this one and tagged it 'clawsweeper:human-review': https://github.com/openclaw/openclaw/pull/83987#issuecomment-4514728713

And Clawsweeper commented, "Maintainer should explicitly confirm whether unbounded restart budgets should apply to embedded runs before merge" for this one:
https://github.com/openclaw/openclaw/pull/85708#issuecomment-4525222669

[potent path]

They're taking our jobs! 😮

[languid folio]

3k -> 99 LOC

[odd anchor]

https://tenor.com/UFs4bF4O9X.gif

[timber beacon]

Nice!

[languid folio]

could you try this? behavior shouldnt change

[timber beacon]

Yup, I can do some testing here later today

ill try and make sure to capture logs as well to update proof

[warped tide]

https://github.com/openclaw/openclaw/pull/83947 - Overall: 🐚 platinum hermit
This PR adds CustomEditor handling for Kitty printable key-release events and German-layout AltGr CSI-u printable input, plus regression tests and a changelog entry.

[odd anchor]

https://tenor.com/mW1CKz5REIe.gif

[rocky pewter]

Hi Maintainers, hope you're doing well. I have a few bug/stability PRs open for the current main / current beta line, and I hope they fit the current roadmap.

If you have time to take a look 🫡

https://github.com/openclaw/openclaw/pull/84628
(Plugins / Gateway startup)
Improves restrictive plugin startup. When the installed plugin index can prove the safe startup scope, OpenClaw avoids rebuilding metadata from every installed plugin manifest and only loads the needed startup manifests. Unsafe/legacy cases still fall back normally.

~~https://github.com/openclaw/openclaw/pull/85860
(Agents / Subagent lifecycle abort handling)
Treats canonical aborted subagent runs as terminal failures instead of successful completions. This keeps gateway wait snapshots, dedupe snapshots, subagent registry state, and parent announcements aligned so an interrupted child run cannot be reported as success.

https://github.com/openclaw/openclaw/pull/85865
(Process supervisor cancellation)
Preserves process-tree cancellation while adding a bounded SIGTERM grace window before escalation. Cleanup-capable processes get a chance to exit cleanly, while descendant processes still remain covered by the existing tree-stop safety.~~

[warped tide]

Is there a way to defer clawsweeper checks until another PR is landed?

[potent path]

I usually just keep a follow-on PR in Draft status until the other has merged. Clawsweeper still does it's thing, but at least that way it's clear that it's not meant to be reviewed and merge.

[fallen gate]

you cna just wait for the other PR to land, rebase, then @thick osprey re-review

[warped tide]

Thanks. I was working on some hardening around Peter's pi internalization and the little bastard closed my PR lol

[fallen gate]

I too have had my battles with the raid boss known as sweeper

but the loot is worth it

[civic sedge]

sweeper is judgy about proof it seems

[warped tide]

very much so lol

[warped tide]

I'm learning to appease him in fewer tries lol

[civic sedge]

here's my sanitized logs with exactly before and after.

Sweeper: nah fam

[worldly tusk]

One more small PR from me: https://github.com/openclaw/openclaw/pull/85891

This just removes the codex personality, so it does not confuse the agents. In openclaw, the authoritative personality source is SOUL.md file. if it is not similar to the pragmatic personality injected to the codex base instructions by default, it creates a real conflict that is not resolved easily.

PR rating
Overall: 🐚 platinum hermit
Proof: 🦞 diamond lobster
Patch quality: 🐚 platinum hermit
Summary: The patch is small, contract-backed, and well proven, with the main remaining review concern being the intentional default-behavior compatibility choice.

[civic sedge]

oh look at this guy with his platinum status

[void citrus]

Hi

[warped tide]

didn't realize there wasn't anything other than platinum... jk

[worldly tusk]

there is also diamond 😄

[warped tide]

https://github.com/openclaw/openclaw/pull/84652 - Overall: 🐚 platinum hermit
The branch mirrors Codex app-server user prompts immediately after turn acceptance and adds a regression test for early transcript visibility plus final dedupe.

[lone iris]

tried landing a PR today and it was unranked so I closed it

[warped tide]

keep at it. there's plenty of things to work on

[pastel owl]

Hello, my PR got autoclosed by @thick osprey as superceeded by another PR 83722. But that PR is exactly the reason i created new PR to fix performance regression introduced by it 🤯

Though still WIP
Can someone please reopen the PR so i can improve upon it

https://github.com/openclaw/openclaw/pull/85853

[lone iris]

yep. got overwhelmed by the clawsweeper LMAO. I will try again...but maybe not 2-days-post-finals and when my brain already got fried by Lean.

[warped tide]

catch 22 lol

[rocky pewter]

Why? Don’t give up right away. The ranks don’t mean anything, keep fighting the sweeper until you win. Even small things can be important and get merged

[warped tide]

sweeper's a bastard lol

[worldly tusk]

ask your agent to write up a good description

[lone iris]

cloning openclaw repo to my local machine also clusterfuck so I was writing code on web to try and fix clawsweeper's issues. will definitely be trying again in the near future

[rocky pewter]

If you need help, let me know. Every PR that’s good is worth being reviewed.

[runic totem]

Hi Maintainers, hope you’re doing well. I have a few focused PRs open that may be ready for maintainer review if they fit the current roadmap.

https://github.com/openclaw/openclaw/pull/85288
(Agents / compaction retry delivery)
Status: focused fix with proof supplied.
Flushes block replies after compaction retry wait resolves, so retry-generated replies are not dropped before channel delivery.

https://github.com/openclaw/openclaw/pull/84458
(Memory Wiki / compile stability)
Status: focused stability fix with proof supplied.
Bounds Memory Wiki compile-time page reads so large FileProvider-backed vaults do not launch unbounded read bursts.

https://github.com/openclaw/openclaw/pull/83956
(CLI / session progress visibility)
Status: proof supplied; mainly needs maintainer judgment on command shape/redaction policy.
Adds openclaw sessions tail --follow for compact, redacted, human-readable trajectory progress logs.

No rush; just wanted to make the queue easier to scan.

[lone iris]

most certainly doesn't help that I was trying to land a feature PR instead of a bugfix!! lol

[rocky pewter]

Start small, it doesn’t have to be anything big. That way you can get a feel for how the sweeper works and how things are supposed to be.

[worldly tusk]

yeah, it took me a while to setup and build the repo correctly. i wish there were some instructions... but i maybe it serves as a great filter to make sure only those who figure out the local setup can contribute 😄

[lone iris]

and with just how big the repo actually is, holy fuck!

[warped tide]

half my wait time is for CI and checks to run lol

and of course my agent just introduced a regression in the PR i had gotten pristine

[worldly tusk]

my first PR was a documentation fix. literally one-liner. made a great first practice to battle the clawsweeper 😄

[lone iris]

Clawsweeper is very helpful, I will say that, but I am definitely going to have to get better at Git operations before I do anything again lol.

How do you all edit locally, do you just do a sparse checkout?

[civic sedge]

83723 or 83722? you commented both in your pr?

[rocky pewter]

I still prefer working on stable fixes and bug fixes rather than huge features. The specialists here are responsible for the bigger things.

I think it helps the OC team, and especially ourselves the most if we work together as a team instead of everyone feeling like they have to take on the big tasks. That way, all areas are covered.

[pastel owl]

Typo
I meant 83722

[warped tide]

i'm a security guy so i naturally look for those first

[potent path]

Yeah, I had the same problem with two different PRs today. False-positive detection of superceded, and the PRs were just closed with no obvious recourse; I couldn't find any way to reopen them. So, I just made new PRs and added additional detail to make it explicitly clear why they were NOT duplicate nor superceded. That seemed to work, but yeah, maybe clawsweeper is a little too aggressive on closing related PRs. 🤔

https://github.com/openclaw/openclaw/pull/85630 "Close reason: duplicate or superseded."
https://github.com/openclaw/openclaw/pull/85448 "Close reason: duplicate or superseded."
Maybe clawsweeper could give us a warning and a chance to explain before it closes the PR?

[potent path]

He did it! https://github.com/openclaw/openclaw/pull/85879 🥳

[rapid spire]

I think clawsweeper closed a bug in error. It claims 84261 is a duplicate of 71992.

While these two bugs do share similar symptoms, they are not the same.

84261: messages duplicate on occasion (specifically not all the time.)
71992: messages duplicate 100% of the time.

84261: issue started somewhere in the May 2026 betas. (specifically wasn't present before May)
71992: issue is reported in 2026.4.xx builds.

.
I think it's premature for clawsweeper to mark these as duplicates before a root cause is determined to be the same (or before a fix is made for one that also fixes the other.)

[vocal night]

Just waking up hows 22 release so far

[shell dawn]

ive been running the beta all day and its been pretty steady

[void bramble]

Clawsweeper says this one is pretty baller. I say it is too.

‘Show assistant commentary in Discord progress drafts
#85200’

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

[shell dawn]

funny enough, i was just thinking along this same line for my chat app, are there elements to draw on in openclaw already or are you creating/extending them with this?

[vocal night]

test and find actual cause and refile IF still in latest version. It’s not going to prioritize a unspecified bug from prior release. If you need prompting help dm me

[rapid spire]

so... now beta (and release) testers have to be able to dive into the code, triage it, understand what they're looking at, and find the actual cause? At that point, they aren't testers. They're basically solving the problem.

[shell dawn]

it depends, do you want to contribute? or just test. contributing is pretty cool, and your claw can help you make that effort be a lot smoother if you've never done anything like it before

52k commits in what, 5 months? lots of people test and contribute

[rapid spire]

I just want to test. As I've stated MANY times before: I'm not familiar with typescript and I refuse to "vibe-code." Therefore, I can't submit fixes.

[shell dawn]

why are you using openclaw if you dont want to vibe code? its only slop if you dont put in a little bit of work

making it your own is the whole appeal to me

[vocal night]

Bro, you are making this place like a place that is hostile. And if you just want to file issues, just file issues. But like, don't bring the negativity in here.

[rapid spire]

there are uses for AI agents other than writing code in languages you aren't fluent in...

[vocal night]

Every post you post is just negative and unhelpful.

[shell dawn]

im not fluent in any language, at all. barely even english

[rapid spire]

what negativity? I suggested that an issue was closed in error and explained that. That's not being negative. Then I stated that I don't want to vibe code. That's also not being negative. That's my work ethic.

[vocal night]

Since you joined yesterday, I can't think of anything you've said or done that's been helpful. You've just been accusatory and negative. Now you say you don't want to help fix things. Great. Don't leave the channel. Go to Twitter.

This channel is for people who are actually coding and trying to fix it. It is not the place to come complain. If you need help, go to the help channel. If you want to just rant, go to general. This channel is for people who are contributing, fixing, and working on the project.

[rapid spire]

In fact, it seems that your attacks on me are more "negative" than anything else. You are literally twisting my words. I didn't state that I don't want to help fix things. I stated I don't want to vibe code. I help fixing things by reporting problems I find.

[rapid spire]

Oh, and you apparently have no clue when I joined the server. 😉

[old ridge]

any feedback about v2025.5.22?

[vocal night]

testing it now, no issues found yet P0. and all issues are appreciated PN but this is "clawtributors" if you don't want to contribute code, filing issues on the repo is more than enough can't make you use AI to file them or even fix them but the noise to signal ratio is high already > try to not make it worse (also issue wise I recommend using AI to submit final issue versus hand writing, even if you prefer to handwrite, having AI rewrite to match repo templates is helpful) https://github.com/openclaw/openclaw/blob/main/CONTRIBUTING.md

[shy rain]

The entire app was vibe coded pretty much. Now that sounds bad but it isn’t really. The world has changed. You can file issues with great levels of detail and evidence thanks to the bots. Sending a slop PR to fix it without thinking about it is probably what you mean by vibe coding. But asking the bot what’s wrong, how to fix it? What impact it might have and thinking around the greater impacts of your patch isn’t that hard. You can push PRs without having read the code. A qualitative discussion of what variable is being changed when and what it affects is all you really need to direct the bots to do things sensibly. The rest is all bots. All the way down. And clawsweeper will sanity check everything.

[shell dawn]

its true

[rapid spire]

If other people vibe code, that's fine for them. I'm not saying it's a horrible thing. I'm saying it's wrong FOR ME. If I don't understand the code and I can't write it myself fluently, then I don't feel comfortable putting my name on it.

[shy rain]

But it’s fine if you wanna hold back at just submitting issues. Just do the same thing.

I did the same for a while.

Then I moved on to patching the code myself.

[shell dawn]

its an understandable sentiment, i felt the same way, but the barrier to learning how to be pretty good guiding your claw into coding is not very high, and you can even learn on how to create a structure for the claw to build pretty good code, and you'll learn yourself along the way

[shy rain]

Then I figured I better push a PR because a lot of other people could use the patch.

[rapid spire]

To be completely honest, now I'm hesitant to submit any more bugs. I don't want to be accused of being negative again, and any bug report is, by it's very nature, negative.

[vocal night]

The thing is this isn’t the channel to talk about how you feel about vibe coding or the project or come in yesterday and just tell everyone what we should do

[shell dawn]

and modern frontier models are quite good at coding if you can guide them the right way

[rapid spire]

My comments about vibe coding have ONLY been in response to others telling me to... vibe code.

[edgy plaza]

@vocal night https://github.com/h-mascot/entity (beta)
This has my own kanban I was gonna send you Andy. Auto run by my agents. More suited for biz workflows than engineering. I added a swarm mode with integration to symphony and eforge for engineering workflows.

[vocal night]

You were warned by shadow ie moderator already yesterday

[rapid spire]

huh? What does that warning have to do with submitting bug reports to github?

[vocal night]

About you being hostile and negative

[shy rain]

Pinched has a point here. You don’t need to have the ability to debug and find the cause to file issues. Complaining without filing is bitching. Filing is providing reproducible evidence.

[rapid spire]

okay.. and that has nothing to do with my statement you replied to.

[shy rain]

Mate. You might wanna re read the thread.

[edgy plaza]

Lower temperature guys
Same side 💀

[vocal night]

I was in it, he came in here yelling at all of us

[shy rain]

Let’s encourage participation. Filing an issue is not “negativity”

[shy rain]

You def need to re read. I’m out.

[vocal night]

Peetie the thread started last night lol

[rapid spire]

I'm not going to address last night. As you said, I was warned about it by shadow. That should have been the end of it. It was the end of it. Until now.

[vocal night]

Great then. Lets end convo.

[shell dawn]

looks neat, im gonna have to learn a little bit from it, i need to scale up my management of my bots and projects

[rapid spire]

awesome. Then can I ask a question of you? I might have misunderstood, but I thought you suggested that there might be some AI/bot format for submitting bugs in "https://github.com/openclaw/openclaw/blob/main/CONTRIBUTING.md", but I'm not seeing anything in there.

[vocal night]

thats just the purpose of the channel, there is a template in repo for filing issues

Same with PR format

[rapid spire]

Are you referring to the prompts that github supplies when filling in an issue? (If so, I've followed that as closely as humanly possible.)

[edgy plaza]

Runs my day to day
Biggest unlock is my conveyor belt https://entityos.dev/operating-model

I talk to my agents on my normal channel and work gets created and run autonomously all day

Again business and ops stuff not coding

[vocal night]

Peter sent me his to test I’m gonna check on it today.

[shell dawn]

yeah, im building from the ground up towards some stuff like this, i havent gotten this high yet but i feel like im getting close, so i will def take notes, thanks for the tips!

[edgy plaza]

Peters will be for engineering
Mine is for biz folks, ideally your customers are biz folks not engineers

[vocal night]

Henry send me what you got in dm and I’ll compare to peters and submit a wrapped updated pr

Yeah biz folks need more than anything

Engineering wise we already have git issues and milestones for sprints

Agents already work well in github issues > milestones

[edgy plaza]

I’m a biz guy so I built what i needed and it runs my day to day

[shell dawn]

i just discovered github issues lol, i was reinventing the wheel in .md files. moving to that

[vocal night]

next go try out paperclip for biz stuff

you can also use github issues + sync to Linear and have your agent use that for product roadmap planning

I'm gonna test peters vs entity vs paperclip today

[shell dawn]

nah i want to make it all in openclaw

[edgy plaza]

I thought paperclip was cool
But my vision is better 🤓

[vocal night]

I ended up just forking Paperclip into Openclaw.

That was the real fix was just replacing the task system with paperclip system internally. And then using it as a view board for what the agent's working on.

[shell dawn]

peter has an openclaw oriented kkanban?

[vocal night]

we're testing it

[shell dawn]

sounds neat

any screenshots? maybe i'll wait for that...

[edgy plaza]

I think it’s part of this https://github.com/openclaw/crabfleet

[shell dawn]

oh this looks awesome, theres a lot of juice in this

[edgy plaza]

Yeah paperclip is definetely top when I think of tools out there for biz folks.
Just a bit bloated for my usecase so far.

With entity it’s lightweight for my usecase.
Files docs Chats etc.
I had my own chat, but with clickclck from team dope! I’m porting that

It’s still single user.
Wanna make it multi user so it can become a mini workspace for small businesses and their agents

[fast crown]

Been experimenting with this and plane, similar concept

[vocal night]

haven't heard of plane will review it and see if it's integrateable.

the key to any good tool is the skills around it so agents can use them well

[fast crown]

https://github.com/makeplane/plane

[untold pollen]

Hi folks, was away for a few days but back now. Please feel free to send Telegram/Android PRs/issues my way 🫡

@low lark going to finally properly review your PR now

[low lark]

Sure but I’m sure there are conflicts now.

[untold pollen]

This is a non-issue with Codex

[low lark]

[vocal night]

please don't break tele

theres a few context and cache issues I'm working on right now

[untold pollen]

Did something go wrong in the last few releases

[vocal night]

yeah broke a few times

I havent checked commits or fully tested the latest with users but 18-22 was significantly slower

also had a session detachment issue with telegram mirrored main sessions

thats been fixed

[pastel owl]

Been working on the reply and turn performance for last 7-8 hours

I hope am able to give a good PR

[vocal night]

I'm fixing the cache breaking every turn issue as well 3-5 hours turned from 122 lines to a refactor

[cursive palm]

I switched to a smarter model and installed Fedora, and it feels like OpenClaw has instantly become 100 times better.

OpenClaw should run on a native Linux system, not WSL2.

[verbal heart]

runs on ubuntu 26 just fine

[crude matrix]

Some people still have the following issue with latest beta, case has been closed as not planned. https://github.com/openclaw/openclaw/issues/84607#issuecomment-4524660698 ... I thought the goal was a stable, high quality LTS. https://github.com/openclaw/openclaw/issues/84607#issuecomment-4524660698

[timber beacon]

Looks like a PR and a different issue are tracking it

[worldly tusk]

seems like somebody broke streaming (at least in discord): i no longer have toolProgressDetail output and codex commentary. it was working like yesterday.

[potent path]

Looks really cool! Nice job on the presentation. I’ll have to evaluate this vs my current Beads experiment vs the Kanban that Peter has been working on. Nice to have options.

[vocal night]

I think peter merged in that one dudes fix for it last night and a global verbose fix.

I’m working on a global fix so we don’t teeter totter between one feature working and then fix for another breaks it in msg

I’m having trouble consolidating the sheer amount of config options

By fix I mean that might be the regression I’d start w that blame

[worldly tusk]

oh, you are removing some crazy config options? nice 🙂

[vocal night]

just trying to consolidate,
tired of rando dude wants obscure config that 99.9% don’t use, merged and breaks every other feature 😆

[civic sedge]

are maintainers the only ones who can get mantis proof?

[untold pollen]

@low lark I tested your PR thoroughly and unfortunately it still doesn't solve the duplicate message issue. Telegram fundamentally cannot do smooth draft preview + guaranteed no visible duplicate/flicker with today’s Bot API contract

Believe me, I want to add it so bad but the duplicate messages it causes are not worth it right now IMO

[civic sedge]

If so I think i've addressed Peter's concern on this from a couple of weeks ago. It adds a feature that codex and others already have and gives us room to test experimental things in the future with an easier opt in method

[untold pollen]

No, Mantis should run on all PRs automatically that can be proven with a video. There's a couple of bugs in the Mantis flow that I need to fix though

[civic sedge]

ok - what do I do?

it says:

A Telegram Desktop proof would directly show native command registration, authorization, and visible replies for the new /experimental path. A maintainer can ask Mantis to capture proof by posting a new PR comment that starts with the OpenClaw Mantis account mention, followed by:

telegram desktop proof: verify /experimental list and toggling one experimental flag through the gated command path.

[untold pollen]

Nothing for you to do right now, I'll get Mantis back up and it'll handle it automatically 🫡

[clever narwhal]

Anybody experiencing issues with preamble and midturn commentary with openclaw on any specific models?

I've noticed some differences in how preambles and midturn commentary is handled (definitely better on Codex, less so on other [maybe dumber] models) and am curious to hear other people's experience.

I'm currently seeing how this can be improved, so that openclaw has the capacity to respond quickly to a user request with a preamble, instead of silently beginning tool calls regardless of what model a user selects

[worldly tusk]

something broke with commentary: #clawtributors message

[clever narwhal]

It's a problem. But not even just since yesterday, at least with generic models other than codex

[next helm]

Thanks to the maintainers for a great release with a focus on security and getting it out fast 👏

[warped tide]

man, clawsweeper really hating on proof today

[vocal night]

anyone from codex team around tonight

[clever narwhal]

That may have been my verbose PR, possibly related to streaming.
https://github.com/openclaw/openclaw/pull/85488

Of course the challenge is all the different surfaces/channels, and ensuring a dead simple /verbose on and /verbose off works everywhere, with all the related settings, like streaming, block streaming, et cetera. It was intended to be generic and from what I can see Pete did a lot of work to try and land it by expanding it to all surfaces and simplifying the logic

In any case, it was already broken other places before this PR, hence the work needed to be done, hopefully it's broken less places now

If you have further fixes, let's keep it going!

[vocal night]

I'm pretty 80% sure this is a prompt issue. I think either Pash or Peter wanted the agents to be more token efficient so they just don't give commentary anymore.

We had to do it as a prompt in the 5.4 overlay because 5.4 originally was breaking by giving too much commentary and then actually not continuing. I don't think this is a bug anymore. The 5.5 and 5.4 overlays are kind of useless now.

A lot of the issues were because of the runtimes were colliding and the tool call issues. The overlays really only need to be like tool based now. Any of the prompting around like not talking and never giving any kind of commentary probably aren't needed.

If someone wants to check the prompts for the overlay for series, I'm at max capacity my agents are debugging like 10 other issues.

if agents are no longer giving the preamble or commentary, it's likely a 5 series prompt overlay (tell your agent to check that) + friendly personality ( if you tell an agent to be concise too many times, it will just end up just being very like not talkative at all. So it could also be that.)

Second would be "Send Message" tool issue with it's prompt, likely tells agent to work then send message after. I agree it's an issue if users don't hear back for 30mins (I have users complaining too).

[clever narwhal]

"giving too much commentary and then actually not continuing" I actually just had this happen to me with Codex on the current openclaw version 2026.5.22

[vocal night]

fyi some PR's I have on the token wastage and cache busting right now (massive token wastage):
https://github.com/openclaw/openclaw/pull/86069
https://github.com/openclaw/openclaw/pull/85978

[worldly tusk]

Oh, you are explicitly ignoring the verbose setting to keep groups quiet, and one needs a special command in a chat to return the previous behavior? Oh… why? This should be configurable, not just silently ignored

[clever narwhal]

Is it not configurable? My PR should have been - I don't know what happened after Pete took over. The idea tho was that groups would not inherit a users config for verbose for groups, as this was recommended during codex review, since in a group it could expose private tool info. Doesn't matter if you don't care. Is it not allowing you to simply put verbose on now?

[vocal night]

I haven't had this issue, but is is why I've pushed more for if we're on codex runtime just kill all other added on prompts except tool related + user personalit.

[clever narwhal]

I assume you tried to just turn it back on already for the group? If /verbose on doesn't work in the discord group then we have a problem

also would be curious to know your streaming type you're working with - there's the streaming where it disappears and the block streaming where the tool outputs are persistent, and which is not working correctly

[worldly tusk]

quoting from your PR:
Groups stay quiet by default. Inherited agents.defaults.verboseDefault: "on" does not opt normal groups into visible summaries. Explicit group /verbose on enables summaries. /verbose off hides them again.

why would it not opt in groups into normal behavior?

[vocal night]

maximum customization is usually the goal

[clever narwhal]

I mean, it could. Do we think that is preffered? I think the idea is that private chat settings should be seperate from group settings, like if you have an agent in a group with friends, do you want it showing all the tool calls?

[worldly tusk]

why does this setting exist if it does nothing, then 😄 default should work as a default

if you say dafault = on, then every child may override with "off"; if you say default = off, then every child may override with "on"

[worldly tusk]

enabled verbose in the chat. it left some tool call blocks in the chat before the final reply. it did edit blocks on the fly, but they stayed. this perhaps is a different functionality then? i had a single block before where current tool calls and commentary were streaming, and it was replaced by the final answer. this does not happen anymore. the way verbose works, this is not a replacement

so many different options, each broken in its own way :D:D:D

with verbose = on:

1. while turn was not finished, new tool calls and commentary was streaming
2. when turn has ended, commentary disappeared, but tool calls stayed in the chat

why the difference? 😄

[clever narwhal]

Yeah, so before you had verbose streaming where the tool calls disappeared, and now when you turn verbose on they persist? and that was simply from /verbose on?

Oh wait I sorry i didn't read your last message

[worldly tusk]

yes. unfortunately, the previous behavior broke like yesterday, so now i get only quiet, and the only way i can get the commentary and tool calls is to enable verbose

but unfortunately, now tool calls stay

so i am very confused 😄

[clever narwhal]

What version are you running taht this is happening? official update/beta? or from what's currently merged?

[worldly tusk]

git main version from a few hours ago (2026.5.24 (98e09e8)), but the disappearance of commentary already happened on git main at least 12h ago

i can update to the current beta release and try again, if that would help

[clever narwhal]

We could see if the verbose PR is exactly what's causing this issue fore you, if you want to target each of these checkpoints and try again

before PR: 2cd73d4c89b60bbd06c88e63cb6c06ad2c0a2a32
after PR: 45fbf2d81a0e6955d977bc4b1c7d29458c2894d0

[zenith prawn]

I've also struggled to understand mental model of a verbose & streaming options.
To make tool calls show but disappear, i have verbose: off(agents.defaults.verboseDefault=off) but toolProgress=true and streaming.mode=partial
Worked reliably ~2025.04-28 -> 05.22 so far

[worldly tusk]

yes! do not update 😄

[clever narwhal]

Yeah, it should be a setting accessible with a / in the chat, but it's not

[zenith prawn]

IMO need to simplify. Peter will love refactor which reduces LOC.
Now its combinatoric explosion:
3x • verbose: how much internal progress the agent emits: off, on, full.
4x • streaming.mode: how live output is displayed: off, partial, block, progress.
2x • streaming.preview.toolProgress: whether tool calls appear inside the live preview. This is the noisy knob.
342 = 24 options, some of them don't make sense
Am i only one feeling this?
UPD: also multiply 3x: reasoning: off | on | stream
3x: progress.toolProgress: true | false
Occam's Razor is RIP

[clever narwhal]

Yes that's the challenge

[worldly tusk]

completely agree

[clever narwhal]

Simpler = better

[vocal night]

I've been refactoring it for a day

but larger broker for all message tools

I want this chat to no longer be about telegram or discord

begone to plugin land

[odd anchor]

https://klipy.com/gifs/ofmd-our-flag-means-death-39

[clever narwhal]

I think the PR was an attempt to simplify accross all channels/surfaces but there was some shrapnel in the process

[tulip gulch]

fix for tui not showing message-tool replies https://github.com/openclaw/openclaw/pull/84126

[worldly tusk]

i just want the tool&commentary streaming to work, and be replaced with the final answer without leaving spam in the chat. that's all 😄

[clever narwhal]

did you try reverting to this checkpoint yet 2cd73d4c89b60bbd06c88e63cb6c06ad2c0a2a32

[vocal night]

Yeah, I personally like that too. I like just one message streaming. Our users see what the agent's doing. They don't ping our support every five minutes. Say, "My agent's not working!" They didn't respond for five minutes.

[remote wing]

Hi team — wanted to flag **issue #86075 https://github.com/openclaw/openclaw/issues/86075 ** for your attention.

After upgrading to OpenClaw v5.22, MiniMax Token Plan (Max) features are broken in sub-agents. The v5.22 bootstrap change narrowed sub-agent startup to only load AGENTS.md + TOOLS.md, which causes all MiniMax Token Plan SKILL.md files to be skipped. This means image generation, video, music, and the full office suite (PPT/Word/Excel/PDF) all silently fail in automated sub-agent workflows. Community users have also reported this issue.

The author has already identified two fix directions: P0 is routing image_generate to the correct api.minimax.chat/image_generation endpoint for Token Plan Key users, and P1 is restoring SKILL.md loading in the sub-agent bootstrap. The issue has clawsweeper:needs-maintainer-review and clawsweeper:needs-product-decision labels set, so it likely needs a maintainer to confirm whether the v5.22 bootstrap change was intentional and how to best proceed with a fix.

I'm one of the MiniMax CLI maintainers and can help validate or implement a fix once the team agrees on the approach — just wanted to surface it here so it doesn't get missed.

[worldly tusk]

sorry, can't do that at the moment, busy with other stuff

[worldly tusk]

yeah, and it should be like a normal, understandable setting:

• just the final answer
• a draft bubble with tools&commentary that gets replaced
• leave the tools&commentary in the chat before the final answer

[clever narwhal]

I agree. We need to make this less hidden/confusing. Only verbose on/off/full are configurable via /. Streaming settings should be on there. Or prob better yet they should be integrated into something liek you said, which is probably enough:

• just the final answer
• a draft bubble with tools&commentary that gets replaced
• leave the tools&commentary in the chat before the final answer

Maybe the question is how to integrate verbose and streaming so they arent seperate configs

[warped tide]

Mantis, why won't you work for me?

[vocal night]

clawsweeper gonna kill me I just opened up like 20 issues lol

"Issue 1-20 we need to know if channels break before we release/consolidate dat shitz"

[odd anchor]

https://klipy.com/gifs/its-merging-time-farm-merge-valley

[zenith prawn]

maybe by merging 2 levels of abstractions: 2. Preview content filter & 3. Live preview renderer ?
SVG openclaw did for me about it: https://i.ibb.co/b50SMkHq/1d4dda7dff7a.png
...or even merging all 3...

[oblique field]

if a maintainer is around i've got a PR ready

https://github.com/openclaw/openclaw/pull/84239 - Overall: 🐚 platinum hermit
The codex-cli branch used synchronous process.exit(0) after process.stdout.write(...), which truncated output when piped, so it now sets process.exitCode = 0 and guards the rest of the script with else to drain stdout cleanly and emit a single JSON document.

[warped tide]

fought like hell against sweeper for these two...

https://github.com/openclaw/openclaw/pull/84652 - Overall: 🐚 platinum hermit
The branch mirrors Codex app-server user prompts immediately after turn acceptance and adds a regression test for early transcript visibility plus final dedupe.

https://github.com/openclaw/openclaw/pull/85763 - Overall: 🦞 diamond lobster ✨ showcase
The branch adds session.sendPolicy.rules[].match.peerEquals: "inboundPeer", recursive send-policy matching, inbound-peer propagation for automatic/durable replies, explicit-send replay compatibility, docs, and focused regression tests.

[earnest jasper]

First PR.. went back and forth with clawsweeper for a week ironing out all the concerns, but the claws are finally happy ;3 Not sure if I need to post this here for maintainer awareness but here it is.. https://github.com/openclaw/openclaw/pull/83637 - feat: add per-agent compaction overrides

[vocal night]

hows 22 been so far, any major issues versus 19?

[vocal night]

have you considered using LossLess claw that will allow you the level of compaction control

[earnest jasper]

will it? Oh... lol I'll look at the docs! and upgrade from 20 to 22 was a little painful, had a 2nd openclaw instance fix the upgrade. it's been more stable and issue free than the rest of the 2026.5.X updates by far

[vocal night]

what was the issues

[earnest jasper]

│
◇
Updating OpenClaw...

Stopping managed gateway service before package update...
Stopped systemd service: openclaw-gateway.service
│
◇  ✓ Updating via package manager (16.36s)

Update Result: ERROR
  Root: /home/kklouzal/.npm-global/lib/node_modules/openclaw
  Reason: global install swap
  Before: 2026.5.20
  After: 2026.5.20

Total time: 19.66s
Restarted systemd service: openclaw-gateway.service
Restarted managed gateway service after failed update.```

```eah. It did not actually update.

What happened:

Registry has openclaw@latest = 2026.5.22.
Your install is still OpenClaw 2026.5.20 (e510042).
openclaw update detected the update and staged it, but failed at the “global install swap” step: replacing ~/.npm-global/lib/node_modules/openclaw with the staged package.
Because the swap failed, it rolled back/restarted the gateway on 2026.5.20.
Gateway is running again; logs only show the known LAN-bind/security warnings.

The weird bit: local update status detects this as a package install using pnpm, but pnpm is not available on PATH in this shell. So the 2026.5.20 updater may be getting into a brittle global-package-manager path during the staged install/swap```

sorry for the large dump lol

[warped tide]

i had this problem a week or two ago. version mismatches because of NVM

[earnest jasper]

the fix was easy enough..

npm install -g npm@11.15.0
npm install -g openclaw@2026.5.22
systemctl --user start openclaw-gateway.service
openclaw --version
openclaw config validate
openclaw status --json```
not sure if that was the correct approach, but it worked. Just needed to force a plugin update after and it's been smooth since..

[warped tide]

https://github.com/openclaw/openclaw/pull/83873 - Overall: 🐚 platinum hermit
The PR changes Telegram inbound body construction to render Telegram text entities into Markdown, adds helper/body regression tests, exports the helper, and adds a changelog entry.

[vocal night]

Codex thread bug fixes:
(threads aren't setup correctly and break cache and compaction... insert here they break it.. this fixes)
https://github.com/openclaw/openclaw/pull/86160
https://github.com/openclaw/openclaw/pull/85978
https://github.com/openclaw/openclaw/pull/86069
https://github.com/openclaw/openclaw/pull/86094

[Refactor] Consolidate Messaging Extensions API (no more BS regressions message channels/extensions become easier to manage)
https://github.com/openclaw/openclaw/pull/86096
https://github.com/openclaw/openclaw/pull/86153
https://github.com/openclaw/openclaw/pull/86154
https://github.com/openclaw/openclaw/pull/86156
https://github.com/openclaw/openclaw/pull/86157
https://github.com/openclaw/openclaw/pull/86160
https://github.com/openclaw/openclaw/pull/86164
https://github.com/openclaw/openclaw/pull/86165

[shell dawn]

andy that looks amazing, love the work you are doing

[vocal night]

TL;DR

consolidates the recurring Telegram/Discord/Slack/WhatsApp/Signal/iMessage/etc. maintenance churn behind one Channel Broker contract. OpenClaw keeps the shared semantics that repeatedly regress across native channel plugins: sessions, allowlists, routing, /verbose, streaming policy, durable final sends, receipts, retries, and auditability. Broker providers own platform mechanics: bot APIs, bridge/device state, native ids, delivery quirks, and provider capabilities.

Why This Is Better

Today, one fix to streaming, /verbose, or delivery behavior can make Telegram fast this week and then break Discord, Slack, or another native plugin the next week. The broker path makes those shared semantics one OpenClaw-owned contract instead of many platform-specific re-implementations. When a provider adds support for previews, threads, receipts, media, or constrained account modes, that capability lands through the same contract and becomes available to every routed channel without OpenClaw maintainers owning every platform quirk.

[worldly tusk]

wow, great work!

[vocal night]

@untold pollen merry xmas

still cooking, will let it debug all night while I go to bed

[clever narwhal]

This will be a huge upgrade when it lands, nice job!

[vocal night]

We’ll know if anything breaks because they’ll all break 😆

[clever narwhal]

Little by little reigning it in

[vocal night]

I swear 50% of all code focus is on the damn msg channels lmao

[clever narwhal]

The preamble midturn commentary I don't think is consistent everywhere either, that's another thing to look at, would be great to try and simplify how that logic is, too

[shell dawn]

andy you have said you have a bridge between openclaw and codex, what have you found is a best practice? ive been looking at a few different paths

[clever narwhal]

are responses super slow for anyone right now on the latest update? or might this be something on my end 🤔

[worldly tusk]

it is 100% cpu usage by the gateway for a significant number of seconds before any action for me. like even for "/status" i get this waiting. my best guess is that it is not agent related - it is just that the gateway does something weird

[shell dawn]

the 2026.5.24-beta.1? yeah, its slow. chat.history is going super slow, im digging into it right now

chat.history might not be the cause, but its part of the hot path that is getting bogged down

and it seems like its codex only, one of my other agents who is on opus is running like nothing has happened

[worldly tusk]

my best guess is that the gateway loads some horrendously large files it does not really needs to. but i have not investigated it.

so, history may be the one

[shell dawn]

ive got a codex session looking at it, its seeing some odd stuff but hasnt narrowed it down completely

GPT-5.5 on codex runtime is slow because the gateway event loop is getting starved, and chat.history is a major contributor or at least the clearest symptom. This is not presenting like a queued task backlog or host-wide resource exhaustion.

Evidence:

Gateway is reachable but slow: openclaw status reported gateway reachable in 1.56s.
Tasks are not backed up: 0 active · 0 queued · 0 running.
Gateway process is hot: pid 763566, about 59% average CPU over 1h16m, and a live top sample showed it around 109% CPU.
Host is not IO/memory constrained: iowait 0.0%, memory available ~24GB, PSI memory basically zero.
Stability buffer/logs show repeated event_loop_delay warnings while active agent work is running, often with eventLoopUtilization=1 and p99/max delays around 18-19s.

chat.history Findings
Across today’s log:

5516 chat.history responses
p50 1660ms
p90 3687ms
p99 13494ms
max 38897ms
3098 responses >= 1000ms

if a few of us are lookign at it lets get it in a thread so we dont spam up the rest of the channel

thread on slow agent turns in 2026.5.24-beta.1

[clever narwhal]

When I enquired with codex about what it could be, it was pointing out plugins

"OpenClaw's plugin metadata/model-id normalization path can be a hot-path performance problem even when plugins are disabled. The fix should be in OpenClaw code: avoid full registry fingerprinting/scanning per turn, reuse current runtime plugin metadata, or pass manifest normalization data into model normalization calls."

Maybe some else knows more about this or can also validate that (I'm on 22)

[worldly tusk]

my gateway process has 26gb of virtual memory opened (doesn't mean it used it - just that it reserved it from the operating system - the resident memory footprint is way lower)

[shell dawn]

ive got my codex log analysis and initial conclusions in the thread

[worldly tusk]

maybe i am missing something, but why exactly does the gateway need the chat history loaded (outside of the current session, of course; but my gateway runs slowly even on fresh sessions)?

[civic sedge]

what kind of hardware?

[shell dawn]

ive got a dedicated server and its running pretty rough right now

[vocal night]

Now the chat history is not slowing things down. It's currently broken. It's been broken for, I don't know, maybe a month. It only shows like 60 or 70 messages. It used to show an entire history, but it was slowing down then. So they limited it to like totally those messages. The problem is it also includes tool calls. So you only end up seeing like a handful of actual messages under codecs.

[shell dawn]

really? codex just said this about chat.history

chat.history defaults through a history read and projection path. Even though the UI asks for limit: 100 and maxChars: 4000 from chat.ts (line 304), the gateway reads up to maxHistoryBytes * 2, with a minimum of 1MB, before projection.

maybe there has been some change in 2026.5.24-beta.1

[worldly tusk]

the chat became sluggish just a few days ago, so maybe somebody tried to fix that, and broke something 😄

[shell dawn]

maybe so, i asked codex to look closer and to also keep andys comment in mind

[vocal night]

1mb is not a big deal for gateway

I actually found the gateway runs really well embedded in app lignting fast but in browser slow as grease

not sure why it's that different

[shell dawn]

ok yeah, the codex session just analyzed the theory from the code with local sessions and its like andy said. it checked 5 agents and from the ~100 returned chat.history 3 of them are 90-99% tool call and 2 of them are 80% tool call

by bytes its 90-98% tools

[vocal night]

its crazy heavy on tool calls. if I had to meta it 'it calls a tool call before it even does a tool call"

[shell dawn]

The most revealing case is agent:channel the gateway returns the max 100 records, but 98 are tool-like. There is only 1 user record and effectively no plain assistant text records in that window. That would absolutely produce “meta-looking replies” and missing/vanishing body content in the UI.

lol rip

[vocal night]

yeah same across all my userse, they lose original message 99% of time

"check my email" agent says hold my beer lemme do everything but that 😮‍💨>>>>>>>>

[shell dawn]

do you have any PRs on this already andy? if not i think my codex session is getting close to the root

[vocal night]

I merged the disappearing web ui bug nothing else close, was next on my list if you can do it

That’d be greatp

[shell dawn]

ok, let me work on something here. codex says chat.history is the root cause, so i'll poke at it a bit to try to find a PR that has a good shape

user/assistant messages and tool calls have very different shapes, maybe both being in chat.history and dealing with both shapes in one isnt the best call

like the old sql design of 'put your blobs in a blob table'

and not with the other important stuff

creating a test fix, if it works i'll put it into a PR. my agents are fairly nonfunctional right now because they've been busy coding so they all have a ton of tool calls in their last 100 messages

[vocal night]

we use stubs in LCM so those tool calls dont destroy context

we stub them down for that reason

98% tool call to 2% personality leads to total personality flattening.

[shell dawn]

yeah, i do similar stuff for transcript management, but chat.history is a different layer

[vocal night]

W'e need longer transcript + also to brng back some comentay

[shell dawn]

the way the codex runtime handles the incoming message is also really different from pi that causes issues on that front

[glad echo]

Maybe this helps with the diagnosis? Comment in the #general channel just now. #general message

[shell dawn]

i can see that alterting the behavior, but for what im working on in 2026.5.24-beta.1 thats not the real cause

the root cause is if your agent is busy coding and uses a lot of tools, it crowds out the chat.history from agent messages that makes the whole thing look broken

[vocal night]

I can't keep up with all the new config options I don't even know what these do. :
"The following 2 config tweaks did a lot for the 5.22 slowdowns, it hasn't resolved it entirely, but it's almost palatable again.

• "contextInjection": "continuation-skip"
• "truncateAfterCompaction": false"

[shell dawn]

the fix im about to install and test is to move tool history to the side a little, so your chat surface renders messages from a a message queue but can expand tool related info easily if necessary

user and agent messages are usually small relative to tools, even big honking messages from a chatty gpt-5.5 or opus 4.7 are only like 3kb, while a tool call could have a 50kb read or write or much bigger

messages and tools are just very different shapes and have different access patterns and stuffing them both into the same polling place acting like they should be equally weighted is problematic

looking good

205 raw records -> 5 display turns, 196 tool records collapsed, 12ms server-side.

would have been like 1800ms+ before

[earnest jasper]

2026.5.24-beta.1 was unusable for me, endless loop of gateway going unresponsive for 30 seconds and then coming back for 15 seconds. 2026.5.24 was a lot better, some occasional gateway unresponsiveness. this made the night and day difference. YMMV.

NODE_COMPILE_CACHE=/home/kklouzal/.cache/openclaw/node-compile-cache
NODE_COMPILE_CACHE_PORTABLE=1
CPUWeight=150
IOWeight=150```

[shell dawn]

my chat.history fix is looking quite good, my 2026.5.24-beta.1 is very responsive, but i hit another bug before pushing out the pr, missing sessions referenced in chat.history basically stall out the process

[civic sedge]

This helps with large model pools and provider pool configs on discord - paginates the provider / model picker if > 25

[civic sedge]

Quite a few models I've tested need just a binary true/false or something similar for thinking when used through vLLM this one adds that.

[warped tide]

https://github.com/openclaw/openclaw/pull/86192 - Overall: 🐚 platinum hermit
The PR narrows Gateway backend loopback auth so scoped shared-secret backend self-pairing no longer preserves self-declared operator scopes except for approval-runtime token calls, with focused tests and protocol docs updated.

[worldly tusk]

If truncateAfterCompaction helps, that means there is a bug there. This option is only for the compaction events, and should have no effect on normal runs

I am not really sure what continuation skip does, though…

[worldly tusk]

I still don't quite get it. Where is the chat history rendering useful? On web UI? Because in normal discord/telegram/slack/etc old chat messages are not rerendered, and so, not needed.

Or is this about context building? In which case removing tool output from recent context will degrade model performance

[shell dawn]

i use a webchat based app, so, i guess it might just be for projections into that, i didnt specifically ask but the impression i got is that its the layer between the transcript and surfacing messages

chat.history is a gateway API for display/history clients, not only WebChat.
discord can use it from what codex says

[worldly tusk]

We can make a poll, but i am pretty sure most users don't use webchat, so i don't get why gateway composes the history on every call for chats that are on other platforms, like discord or telegram. Should be only enabled for webchats.

As for API calls, it can be built on demand (when the tool call is executed to get old messages), and not on every gateway run

[shell dawn]

that can be a followup PR 😄 im making it as minimal as possible and keeping full compatibility

[crystal relic]

https://github.com/openclaw/openclaw/pull/86281 ready for review size: L, rating: platinum hermit
Keeps provider-auth warming from stalling gateway startup and channel handshakes.

Fixes https://github.com/openclaw/openclaw/issues/85999 because the old main-thread warm path could block the event
loop for tens of seconds.

[vocal night]

22 has issue on doctor still, a customer got around our updater block and immediately broke his machine and doctor rewrote his model refs back to openai-codex

I don't know how the doctor issues keep happening on updates but caught an exact case where it malformed them in the update from openai to openai codex again

[left mango]

Hi, I created this PR for a bug fix

[vocal night]

If we’ve deprecated openai-codex entirely I think we can probably just remove it from model refs entirely no? And link any provider calls to it to openai normal?

[worldly tusk]

Sounds reasonable

[vocal night]

I’m out of PR’s and maintainers working on old ones so I can’t closem if anyone else wants to look into that backward compatibility fix

Name it “[fix] stop doctor claw’s bloody config rampages”

[worldly tusk]

You can probably close https://github.com/openclaw/openclaw/pull/83612?

[vocal night]

Shoutout to vincent he’s been closing and fixing all my old stuff. I have to close pr’s due to limit all the time and forget about them and the obscure bugs 😆. I need room for 50.

[worldly tusk]

Yeah, you should have a special exception!

[vocal night]

Its nice to wakeup to changelog and see oh wow thats old fix where did that come from 😆

[odd anchor]

https://tenor.com/view/get-this-man-a-get-this-man-a-gold-get-this-man-a-trophy-comedy-gold-art-gold-gif-25837757

[vocal night]

If only 😆😆

[fast pagoda]

Opened a XS guard QMD session stem fallback PR: https://github.com/openclaw/openclaw/pull/86327

[potent path]

Me: #73950 seems to have been abandoned. Let me write a replacement https://github.com/openclaw/openclaw/pull/86340

Clawsweeper: Close this PR as superseded by #73950

😫

[crystal relic]

https://github.com/openclaw/openclaw/pull/86089 ready for review size: XL, rating: 🐚 platinum hermit
Fixes silent gateway-restart recoveries so interrupted chat turns get visible feedback instead of stranding users.

[celest merlin]

OpenClaw 2026.5.24-beta.2 is live.

Please test it if you can, especially if you use Discord voice, iMessage, Telegram, WhatsApp, WebChat/TUI, Windows, Alpine, OpenRouter, or custom providers.

Highlights since 2026.5.22:

• Discord/WebUI Talk can now check run status, cancel, steer, or queue follow-up work while a realtime consult is still running.
• Discord voice got wake-name gating, better startup/rejoin handling, realtime + Meeting Notes fixes, and fewer weird spoken diagnostics.
• iMessage now supports approval tapbacks: thumbs up = allow once, thumbs down = deny.
• Image handling is more model-aware, with adaptive compression and agents.defaults.imageQuality.
• Gateway/model hot paths got more caching, less repeated plugin/catalog scanning, and faster model/auth metadata lookup.
• Codex auth setup can store named provider profiles via openclaw models auth login --profile-id.
• Lots of channel reliability fixes: Telegram topics/retries/context, WhatsApp deferred media, WebChat duplicate rows/uploads, TUI message-tool replies.
• Install/update hardening for Windows, Alpine/musl, npm release-age configs, Gateway restarts, plugin peer repair, and beta-channel package updates.
• Security/diagnostics cleanup around SecretRefs, OpenTelemetry labels, sandbox skill mounts, tool-policy audit logs, and provider error metadata.

Update with:

openclaw update --channel beta

If something regresses, please reply with OS, install method, current version, the command/action you ran, and the relevant log snippet.

Release notes:
https://github.com/openclaw/openclaw/releases/tag/v2026.5.24-beta.2

[willow narwhal]

<@&1503801512294486217>

[zenith prawn]

updating! also rebasing my feature-showcase PR to it https://github.com/openclaw/openclaw/pull/85104
UPD: just 5 commits fast-forward, easy UPD: DONE, PR also ready 🙏
Telegram, TTFT: 9sec (Telegram, origin/main + PR fast auto; gpt-5.5, xhigh, fast:auto(on), test prompt: "hi, say test" in new DM topic)

[glad echo]

Ready for review/decision
Overall: 🦐 gold shrimp
Proof: 🐚 platinum hermit
Patch quality: 🦐 gold shrimp
Summary: Useful feature work and terminal proof are present, but the missing SDK API baseline and unresolved contract approval keep it below merge-ready quality.

https://github.com/openclaw/openclaw/pull/85723
Add /goal session continuation command

[zenith prawn]

sorry for newbie question... Should we rebase our open PRs to v2026.5.24-beta.2 OR to origin/main ? they diverged by 8 beta-only commits / 51 main-only commits as of now...

[shell dawn]

ive been grinding on some issues from 2026.5.24-beta.1 that are still present in 2026.5.24-beta.2, i was mostly fighting various other issues in openclaw i still need to get to, but i should have two PRs ready soon

[low lark]

Mind sharing details on how you tested it exactly?

[upbeat trench]

Hi @celest merlin @mystic trench and OpenClaw team,

This is Zayn, core contributor from OpenViking, a context database. https://openviking.ai/

Thank you for everything you've put into OpenClaw—and especially for opening up the ContextEngine ecosystem. Making the
context engine a first-class, pluggable part of the platform is a real strength, and it helps us to integrate with OpenClaw.

A quick introduction: OpenViking is an open-source project under VolcEngine 火山引擎, a context engine and memory layer for AI agents
— it manages multi-agents' memory, resources, and skills under a unified file system paradigm, and supports layered, on-demand
context delivery. We've built native support for OpenClaw with the ContextEngine plugin.

• GitHub (24.6k stars, 1.8k forks) https://github.com/volcengine/OpenViking

• ClawHub plugin: https://clawhub.ai/plugins/@openviking/openclaw-plugin

• also integrated into Hermes Agent and proven in production. https://hermes-agent.nousresearch.com/docs/user-guide/features/memory-providers#openviking

  We noticed that the ContextEngine section currently features lossless-claw. It's a great engine and a strong showcase of what
  this ecosystem can do—which is exactly what prompted this message. We were wondering whether you'd consider featuring
  a curated set of context engines side by side in this section, rather than just one. Showing several mature options together, like
  memory section, would reflect the flourishing ecosystem of the context engine, give users more choice, and encourage more
  teams to contribute. We'd love to continue contributing to OpenViking X OpenClaw integration further with such collaboration.

  If there's any process or steps you'd need from us (submitting a documentation PR, going through a review checklist, etc.),
  just let us know—we're happy to help and to move at whatever pace works for you.

[vocal night]

I'm a fan of OpenViking, not sure any Claw users use it, that being said code was good when I reviewed it a few months ago and given half the LossLess PR's are mine I want to see more context engines come in. That being said make sure to add me and Josh so we can collaborate when updates break them

[empty plover]

are we going to be getting multiple updates a week for a while cuz its annoying

[upbeat trench]

love to see that, and we can work together to build something great.

[vocal night]

Its came a LONG way, I might have to switch lol

[worldly tusk]

You can always just not update your system multiple times per week, if you find it annoying. You can update once a week, or once a month

[empty plover]

i stopped updating once 5.19 came out

[rocky pewter]

This is not a usefull comment at all for this Channel and second how can it be annoying How can it be annoying that new betas are released for testing when they improve the system? That’s a contradiction in itself.

[empty plover]

ok mb bro didnt know

[vocal night]

dont engage trolls yall its wasteful and not focus

[empty plover]

danm bro chill its 1 comment that I didnt know alot about

Dont have to be an ass

[vocal night]

actually I'm out of PR's can you make one that allows disable the auto update popup its designed to be automatically clicked (our ui/ux is atrocious but even when I manually disable it, it manages to worm its way back in)

it should be like 20 loc and I'll make sure it gets merged @worldly tusk webchat auto update popup is half of our user issues click > break > mad

auto update should only appear if release is tagged stable

[worldly tusk]

I have never used web chat. I don't even know where to find one 🤣

[vocal night]

elsewise it should be able to turn off

just tell agent "file issue - update popup should be configurable to off in UI in openclaw config > file PR to disable auto update in UI if config does not have update_popup = on , and for good measure add a upate_popup_channel= stable default (beta/alpha/release/stable) lol

its such a small fix I keep forgetting and then I get woken up 6am cause user clicked it when I coded it out manually already

file it as "organizational safety and security requires ability to control updates of users, users should not be able to update if popup is off" hell throw in there block update command by agent to "block_agent_update= off default, its organizational security thing

[vocal night]

webchat is used by a lot of our business users, but sometimes an agent will also manually update itself if it sees a fix upstream for an issue, "I found fix upstream, updating now" ..... agent dies. happened to one of my users this morning when the doctor overwrote their config to openai-codex refs

[plucky sentinel]

@vocal night
Compaction test update — multiple blockers.

Session takeover permanent now (was occasional on May 21). Every chat.send fails with EmbeddedAttemptSessionTakeoverError. Single gateway process, nothing else should write the session file. May 21 completed 120+ tool calls with only ~30s-apart errors; now every request fails immediately.
Cannot rebuild dist — pnpm build times out (>600s, 130 workspace packages). tsx JIT also too slow (compiles all extensions on the fly). npm install -g openclaw (2026.5.22) has no embedded Pi agent runner — chat.send acknowledges but never starts agent.
lossless-claw installed but untestable — @martian-engineering/lossless-claw@0.11.2 installed, config set, but can't verify without a working agent runner.
Questions: a) Why does midTurnPrecheck.enabled default to false? (attempt.ts:2500 checks === true) b) Faster build target for source checkouts? c) Does preemptive-compaction.ts refactor fix cache tokens not counting toward compaction threshold? (cacheRead was 91% of 426K tokens but excluded)

[vocal night]

welcome back shan, it's been a few days is this latest release today?

[plucky sentinel]

2026.5.22 was released on May 22, not today. The latest source main is at 95d1b39b9 (today), but I can't build it — pnpm build times out. That's one of the blockers I'm about to report.

[plucky sentinel]

Compaction test update — multiple blockers.

1 Session takeover permanent now (was occasional on May 21). Every chat.send fails with EmbeddedAttemptSessionTakeoverError. Single gateway process, nothing else should write the session file. May 1 completed 120+ tool calls with only ~30s-apart errors; now every request fails immediately.
2 Cannot rebuild dist — pnpm build times out (>600s, 130 workspace packages). tsx JIT also too slow (compiles all extensions on the fly). npm install -g openclaw (2026.5.22) has no embedded Pi agent runner — chat.send acknowledges but never starts agent.
3 lossless-claw installed but untestable — @martian-engineering/lossless-claw@0.11.2 installed, config set, but can't verify without a working agent runner.
Questions: a) Why does midTurnPrecheck.enabled default to false? (attempt.ts:2500 checks === true) b) Faster build target for source checkouts? c) Does preemptive-compaction.ts refactor fix cache tokens not counting toward compaction threshold? (cacheRead was 91% of 426K tokens but excluded)

The latest source on main is at 95d1b39b9 (today), but I can't build it. That's part of blocker #2 in my report."

[pastel owl]

Hi Clautributors,

Below is the reply and turn performance focused PR.

I have spent at least 30-40 hours digging into this and testing it extensively on a live Telegram/OpenClaw setup.

Found a lot of different kinds of blockers stopping the prompt from reaching just the LLM:

1. Telegram reply path was probing Discord /users/@me and timing out every time after 11.2 seconds.
2. Workspace dir setup was costing ~3.5s on each turn, now cached.
3. Model and thinking catalog checks on each reply were costing ~18s.
4. Context compaction caused delays up to ~2 minutes in the worst case.
5. Direct block reply delivery and transcript mirroring had extra hot-path waits.
6. Gateway health refreshes and heartbeat/cron work could interfere with active reply runs.

It’s a large PR mostly due to timing profiler code, comments, tests, and docs. The profiler is opt-in only and can be enabled with:

OPENCLAW_DIAGNOSTICS=profiler openclaw gateway run

Hope it's clean for merge. Request review from Peter and other maintainers.

https://github.com/openclaw/openclaw/pull/85941

[worldly tusk]

Regarding your first point, I think it is also true the other way around: a while ago i noticed weird telegram timeouts on discord messages. Didn't dig more into it, though

[shell dawn]

hell yeah brother

[worldly tusk]

But wow, didn't expect the model catalog to cost 18s 😲 impressive work tracking this!

[shell dawn]

i can, i hit a session model record that was causing per line model catalog resolution that was murdering performance, i put a PR in that worked around it instead of hitting the model catalog directly

nice one!

[rocky pewter]

Nice Work !

[pastel owl]

Model catalog 4s, thinking catalog another 14s

[pastel owl]

Could be, i havent checked either

[worldly tusk]

Crazy! Was it polling each provider for the current values on every single reply? 😲

[slow inlet]

I would like this deletion to be optional if possible. It allows me to quickly sense check what the agent did

[pastel owl]

Yes the full runtime Catalog

[worldly tusk]

I agree, there should be such option: #clawtributors message

[slow inlet]

I've been working on a fix for me personally on this where it interleaves the tool use into the reasoning stream:

As I'm a CLI user I'm waiting for another PR to land first for a fix:

https://github.com/openclaw/openclaw/pull/82285

It feels much closer to claude code within telegram for me which I like.

[vocal night]

I fix this with my unified PR's I noticed some of same issues

The pipes are a mess lol

[vocal night]

model checks likely codex team over openai vs openai codex

4 seconds (are we on openai or openai codex), 14 seconds (are we STILL thinking Xhigh? let's check 10 more times) lol

[pastel owl]

This issue also came sometime later in one of the builds
But the main time is for checking all active model catalogs, like eg. i have openai/codex, openrouter, Claude

[vocal night]

really easy fix there its a filter before checking any catalog is provider active in main + subagents allowed

will limit catalog search to only ones in config, even faster the ones allowed in the allowed model picker provider wise, should speed it up significantly vs all catalogs

[rocky pewter]

Hi Maintainers, hope you're doing well. I have a small cron 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/86415
(Cron / config persistence)
This prevents unsupported persisted Cron rows from being deleted during routine jobs.json writes.
The rows stay preserved on disk but do not become runnable.

[vocal night]

only issue is maintainer readability + mergability score on my end is 6/10, can you have ai add to description diagrams, TLDR at top and the actual issues found (like your post you made her) otherwise maintainers can only read it via clanker

needs to scream shits broken and why lol

[shell dawn]

andy you should really make a skill pack or something to help people level up their PR process, some of us (like me) still have a lot to learn

[vocal night]

always need that for 3k+ loc entries

[shell dawn]

ive been making one, but im still behind the curve compared to the pros

[vocal night]

you can buy an EVA agent thru our site, she consistently ranks top 10-15 out of all maintainers

[shell dawn]

ill have to take a look

[vocal night]

I'm not OSS fan, I'm only here cause we build ontop of OC gateway

[shell dawn]

fair enough

[vocal night]

less bugs = less bugs for our customers haha

[shell dawn]

its good for everyone

[vocal night]

honestly peters skills are really good just token expensive

gitnexus or codegraph + embedded repos for coding save significant coding speed/time etc

can't skillify everything, sometimes it's infrastructure

[shell dawn]

agreed

[vocal night]

also one of the reasons Eva still beats a lot of people is she uses Claude Code + GPT 5.5, opus for architecting and blast radius level bugs. 5.5 is really great at 1 file iz broke. opus still is unbeaten at "trace this through the system, and figure out best architecture" implementing that tho 5.5 is best

my setup is OC orchestrator so she manages both apps, don't use subagents anymore in OC, it's broken and token inefficient most of the time, she just manages the desktop apps

have plugins that give her "api" connections to them directly + she can type in any box or thread to collab

allows her to manage 5 codex desktop apps at once

[shell dawn]

multi model has a lot of benefits. opus is extremely persistent at doing novel investigation

are you doing that through MCP? ive started experimenting with that approach, i like it a lot, but need to figure out how to take it to the next level

[vocal night]

it has to be hybrid, peekaboo + native compatibility + training skills + MCP and direct hooks into the DB

[shell dawn]

codex + gpt-5.5 is exhaustive, but almost.. surface level? at times

[vocal night]

its not dumb, just low context and tool heavy. by the time it figures out the problem it compacts itself and doesn't have native memory (its experimental and have to enable in the codex app)

[shell dawn]

i found it helped a lot

[vocal night]

they've optimized it recently with better hooks to OG prompt, but still it's about 100k context more needed for it to really shine. I don't expect it tho, the API sale is how they sell corpos on 1M context and consumers get 258k, really like 200k after prompts.

[shell dawn]

yeah, for how exhaustive it is, it must really benefit from giant context windows

[vocal night]

one thing everyones missing out on right now is 5.4 it's been really optimized and no reason not to use it. It's parity is just as good and has context of 5.5, and is more creative.

5.4 orchestrator + 5.5 subagents is actually best way to use codex

[shell dawn]

interesting, i'll have to look at that.

i did use a fair amount of 5.4, but not in codex

[vocal night]

in OC we always had bugs with it because of the harness and prompts, by the team we got them fixed 5.5 came out

I'd have this convo in general but general is literally like a teen boy asking how to write prompts to build website today different uhh exp level.

[shell dawn]

heh

[true laurel]

If you want, I can orchestrate 5.5

[vocal night]

btw if you use composio make sure to disconnect it and even delete plugin, they got hacked bad and now all their systems are down even internally.

[glad echo]

Is the OpenClaw project limited by tokens/bot work or skilled maintainers? If it's the first I can image a system where users are given small tasks to work on with their bot. I would be more than happy to contribute in that regard if possible.

[vocal night]

It's limited by lack of QA testers and beta testers

[shell dawn]

i would say its not a particularly small limit, but a very large and broad project

[vocal night]

millions of users, 1000's of devs, and no QA and a handful of beta testers. imagine bethesda but nobody testing the game lol

well actually even then they release broken games

[shell dawn]

do they... release games that arent broken?

[vocal night]

hermes on other hand raised 100M usd, and closing 1 billion

very different level of resource and command n control

[shell dawn]

would it be a bethesda game if it wasnt permanently broken in many ways?

i swear to god half of the fun in the elder scrolls series is discovering what terribly broken designs they've implemented that you can feedback loop into godhood and broken things they definitely did not intend

morrowind even kind of embraced it in the theology and backstory

[true laurel]

My agent is currently merging this locally. I'll let you know how it affects my performance.

[shell dawn]

super powered jumping skill being able to jump across cities as a means of transportation was one of the more convenient ones

[glad echo]

Oke, understood. Is testing betas via spinning up a number of different configurations and running simulation on them a thing?

[true laurel]

OC moves very fast with commits, every day there's hundreds of commits. It is hard to say if any of them actually do much other than make things worse half the time. Testing it is impossible to test for stability at that volume, unless you have lots of people constantly just.. testing commits. 😅

[vocal night]

def my first game, I'm almost done with Baldurs Gate 4 if you wanna test it tomorrow. Built it this weekend (dnd type adventure with Openclaw agent), it's quite fun been looping it all weekend.

[vocal night]

I'm pushing for FT QA team it's cheap in developing world. Molty generates QA/QC checklist > sends to them > they complete check list. Cheaper than inference and better.

[shell dawn]

O_O what

balders gate 4?

[vocal night]

unofficial lol

[shell dawn]

ambitious much?

[true laurel]

I wonder how hard it'd be to get OC integrated into a VTT like Foundry, develop software that generates maps, and have it run the game as a support-DM alongside an actual DM 🙂 Take away the annoying tasks from DMs 😄

[vocal night]

thats what I'm doing

[shell dawn]

an AI DM? there are so many interesting (and expensive lol) areas that LLMs could really add value to games

[vocal night]

but it expanded a little bigger, looping games with ai players its considerable leveled up + Tolkein + Angry DM agents (I use to review the game at each step in claude)

[true laurel]

(This is actually more of a #general conversation, but..) oh really, into Foundry VTT? 😄

[vocal night]

yeah, started as that, plugin all the BG series lore, generate world seeds based on all the different endings

turned into iterative loops running the games for days improving after each one

[glad echo]

Would anybody be interested in this if I try to get something like an OpenClaw Release Canary working?

[vocal night]

like crabbox? https://crabbox.sh/

[true laurel]

Ok after merging that PR, the agent seems a bit faster, but I need to poke at it some more to be sure.

[vocal night]

If it's good I'll merge and test it too

[true laurel]

I am currently waiting for it to give me timing numbers of messages before merge and after merge

[vocal night]

my agent responses are abyssmal on latest

[true laurel]

OC agents are feeling like messages go into a black hole. I send a message. 10-15 seconds later "Agent is typing.." appears, then it sits there for 15 minutes without any output, and eventually it either gets time-out killed, or I eventually get a response.

I really miss the little reasoning snippets we used to get.

Ok, the pre-turn delay went down from about 35s to about 10s

It is still slower than 5.18 but no longer unusuably slow

Yep this PR is great

I will test the profiler after lunch

[glad echo]

Maybe crabbox+, I'll think about it. I'll take it to the crabbox repo if I have the feeling I could add some additional real life tasks to it.

[vocal night]

anyone else notice they removed fast mode from codex

[shell dawn]

i just turned it on?

[vocal night]

latest update

[shell dawn]

better finish what im doing lol

[vocal night]

50% chance its a bug 50% chance its a feature, that same team working on OC lol

[rocky pewter]

For me fast mode ist working on Desktop and CLI

[shell dawn]

i honestly cant really tell much of a difference, ive never used it before, maybe its activating but not working

[vocal night]

I rarely use it cause it just blows thru tokens and rarely is faster when doing multi thread

they improve one release and then next one poops again

[true laurel]

Alright, lunch done, time to figure out that profiler

[true laurel]

That profiler produces cleaner data that before. My agent reported the following:

Rundown for <discord message id>:

• Discord snowflake time: 14:10:02.232
• Resolver finished: 14:10:06.661
• Agent turn reached before_embedded_run: 14:10:08.217
• Agent completed: 14:10:21.990
• Reply hot path completed: 14:10:22.179

So end-to-end was about 19.95s.

Breakdown:

• Resolver: 4.339s
  • resolve_default_model: 3.679s
  • resolve_directives: 645ms
  • rest basically noise
• Agent pre-embedded prep: 1.113s
• fallback_prepare_harness: 367ms
  • still ~0.66s unlabeled before/around that
• Embedded/model side: 13.769s - This is where the message leaves OpenClaw and goes into Codex/OpenAI servers.
• Final reply wrap-up after agent completion: ~189ms

Other notes:

• Tool policy ran twice again.
• AGENTS.md and USER.md were still injected/truncated.
• chat.history was fine: 136ms during the turn, 548ms near completion.
• Session usage for my reply was:
  • input: 5,203
  • output: 265
  • cacheRead: 153,984
  • cacheWrite: 0

This was a clean warm chat turn. The main owned cost remains the stable ~3.7s default-model resolver tax, then the model/embedded run itself.

[worldly tusk]

"Resolver: 4.339s"? 😮 what is there to do for the whole 4s? it's like there, in the configuration 😄

[true laurel]

It's mostly resolve_default_model, I just realized codegraph wasn't initialized on my openclaw checkout so my agent is doing that real quick, but we're going to profile that next. Probably it just needs a little caching 🙂

[worldly tusk]

the model should just be cached with the session. it can be resolved on session start and gateway restart; change with the explicit command. it's ok to fail if the model goes away - then there is a failure procedure when to change to fallback models...

the reduction is impressive nonetheless! i hope the maintainers merge that patch soon... waiting 40 seconds for the model to even start processing a request is torture 😄

[vocal night]

some of these are really good patches

I don't want them to get lost

[shy rain]

oh my god I'm trying to debug this stuff it's driving me insane. I pushed one fix already but I don't even know why it's doing what it's doing right now. Opus is on the case

[vocal night]

Peter’s Pending PR P0-P2 Patch Parade for v2026.5.22 👀 🦞

made into a thread if you have a P0-P2 on the speed issues/slowdown/totally broken feature from today please post in the thread yall so I can push them up if you're offline when Peter's awake later

@worldly tusk @pastel owl

there's def some critical bugs I have like 10 for the release so please if critical and last 24 hours post in the thread so we don't lose them

[worldly tusk]

he probably reads now 😄 see: https://github.com/openclaw/openclaw/commit/033693843c26287d546314620874087c8905be96

[vocal night]

I feel like I'm missing 1 or 2 that were critical

[true laurel]

My agent dug around for a bit:

buildModelAliasIndex was parsing every configured model key, and each parse could re-enter plugin metadata snapshot resolution because manifestPlugins was not hoisted. I patched src/agents/model-selection-shared.ts so model-selection resolves manifest plugin metadata once and passes it through the repeated normalization calls.

Microbench on the live config:

Before: resolveDefaultModelForAgent ~1883ms avg, buildModelAliasIndex ~1672ms avg
After: resolveDefaultModelForAgent ~0.4ms avg, buildModelAliasIndex ~0.1ms avg

[vocal night]

PR's only please

[true laurel]

I'm just letting lastguru know what we were talking about 😛

[worldly tusk]

Nice!

[vocal night]

he's been traveling etc I'd prefer to just line up for him PR's that are good to go in it from 22 and any from beta (if anyone has tested)

[vocal night]

I know I'd like to see it in PR haha

[true laurel]

My agent now very quickly begins the "agent is typing" response again.

[true laurel]

I haven't made PRs before, I'm just testing and poking so far 😅

[vocal night]

Hi Agent, it seems this really worked in making my speed faster. Can you research this issue versus the current release beta on main? Confirm the issue is still there. If the issue is still present in these ABC slowdowns and if this PR has fixed them, we potentially have proof to file an issue and then file a PR for the fix. Research it, follow the upstream REPL's guidance on filing the issue and filing a PR. Go ahead.

@true laurel ^ copy that

[true laurel]

Plus they are all essentially vibecoded works, not human-audited, I haven't looked at the changes my agent has made, and if it just did something dumb like "Oh hey I can speed this up with return openai/gpt-5.5 and be done with it"

[vocal night]

dont worry clawsweeper will say if it's not good

if its half good, a maintainer will wrap and fix it

[shy rain]

Point your agent at this: https://github.com/openclaw/openclaw/blob/main/CONTRIBUTING.md

Then tell him to follow best practices to make a PR once he's been able to confirm the issue, and patch it with success. I like to discuss with the bot what he's doing as a sanity check. As you say sometimes they solve the "problem" by sweeping it under the carpet. Mine just told me to disable the memory extension for all Discord channels lol. No clanker... it must work.

[timber beacon]

https://github.com/openclaw/openclaw/issues/85731#issuecomment-4534492068 - I spent all weekend on iOS app have not moved to PR yet but shareing for feedback

[true laurel]

Alright, post-gateway restart warm-up turn, no tool calling, just a simple message + response profiling:

Agent response:

That one is a clean fast sample.

• Discord snowflake time: 15:14:49.270
• Resolver done: 15:14:50.171
• Full hot path done: 15:15:00.377
• Total message -> done: about 11.1s

Split:

• reply.run_reply_resolver: 10.954s total wrapper
• Pre-agent resolver milestone: 751ms
  • resolve_default_model: 156ms
  • resolve_directives: 584ms
• Agent turn: 9.233s
  • prep before embedded run: 1.042s
  • embedded_run: 8.188s
• Codex trajectory:
• session started/prompt submitted: 15:14:55.539/542
  • model completed: 15:14:59.079

So this is the healthy shape: resolver under a second, default-model resolution ~160ms, remaining time mostly Codex/provider lane. The resolver patch definitely did its job.

[vocal night]

are you POC for iOS? Really should be like a 2 hour deal mate

[true laurel]

Now I'll go explore the patch, and then I'll see if I can figure out how to turn this into a PR 😅

[timber beacon]

Working past MVP, I wanted to run it myself to validate so went past POC/MVP haha

[vocal night]

I've shipped a lot of apps, the UI needs a refresh for sure UI/UX is not on 2026 standard

but just ship it

webchat needs a total refresh too + ability to change themes. once we get past every update breaking my customers like civies and hellcats in GTA 4 I plan to refresh it

[true laurel]

Oh, damn, the fix is actually really simple 😛

Core change: resolve manifest plugin metadata once per resolver path, then reuse it for all model parsing/normalization calls.

Ok, I'll go see if I can't turn this into a PR then.

[vocal night]

UX is fine I reviewed it with UX map ios, just ship it and we'll iterate. once you have it in app store we'll test it and fixes are very easy + automateable from there.

(ui just needs fresh paint 2014 wants its design back for web and mobile lol)

but dont do it now

[timber beacon]

LOL yea, trying to get to store

then it can improve

[vocal night]

codex can handle it all for you " Codex, here's my dev login. Go get this App Store, get it published, keep iterating and setting reminders, fix it until it gets approved."

[true laurel]

We also have a performance fix for sessions.list, (6s to 0.1s) but let me first get this one out of the way, before I review that one too. 😛

Oh I should ask my agent to check if it depends on the merged PR from Keshav G or if it is a separate file.

[vocal night]

thanks, I really want to press to get a stable hot fix out on some of these bugs without more new breaking regressions. We really need to fix the priority system by Clauseweeper.

I'll be like messaging doesn't work on telegram or agent takes one minute to respond. and Clauseweeper will be like...uhh acktually....technicallly..gateway turns on so P3.

[true laurel]

Performance fixes should be P1 really 😛

[vocal night]

I come from real production land, if user's can't use product or dropoff after a few minutes, it's PO

developers have malformed the idea of P's to be well technically server turns on so it's not.... uh critical... yeah but you can't call it

[true laurel]

When we are building a game towards release, usually the last 3-4 months is reserved for performance fixes. P0 is reserved for game crashes, hard blockers from progression. Anything that makes the game perform poorly or any other form of serious degradation (but can be worked around) is P1

[vocal night]

Deep that's pre-release, post release is a different game

[true laurel]

Nah, every update goes through the same cycle.

Even live updates for GaaS

[vocal night]

Can be worked around is dependent on performance degradation

[true laurel]

Can be worked around doesn't mean "it is great" 😄

"The game crashes if you open the door" "Yeah but I can jump through a window" "Ok, P1 then.."

[vocal night]

In gaming we're talking about a few frames for performance fixes or ms latency

[vocal night]

Gaming has many hacks and workarounds to make work

in this case if you can't use telegram, but can use slack theoretically it still works

but at enterprise level and millions of users it's a release blocker

[true laurel]

I would say "all doors are broken everywhere" is a blocker, for sure someone only has doors in their house

[vocal night]

welcome to openclaw, it's actually just a big gateway for codex and pi

thousands of doors, constantly being broken lol

[true laurel]

Oh no what is my agent doing, I asked it to read https://github.com/openclaw/openclaw/blob/main/CONTRIBUTING.md and it's been busy for 12 minutes and restarted the gateway

[worldly tusk]

It probably ran tests

[true laurel]

I really wish we had reasoning output working again. I miss the days where it went "User need must ask do a thing" and I could follow its thought process

Now it's just 15 minutes of "Welp, lets see what it comes up with"

[worldly tusk]

When running from a live source, ask the agent not to kill himself

[true laurel]

I just tell it "make no mistakes", that's how that works right?

I have actually had to go in there myself in the past to fix it manually, where it just broke something and immediately issued a restart.

Nowadays it's a bit more careful with "The code should at least rebuild cleanly before restarting"

[vocal night]

I've had much success with take no action until 95% confident. which leads agents to always research and confirm or come back and say I'm 80% confident this way and 30% this way

if you're using your openclaw though to debug itself you are wasting like 80% resources and tokens

[true laurel]

It usually smoketests code by just directly calling the functionality that was changed, rather than launching the whole gateway first.

[vocal night]

just use codex desktop

it can operate and manage the OC and read logs at same time etc, its 100x faster and better. Surgeons don't do surgery on themselves

[true laurel]

Alright agent is still busy, time to make tea

[true laurel]

It finally finished doing whatever it was doing, it realized it had no access to github (I never gave it a key) so it tried to figure out how to get into github (even though I told it not to) and went off on a web adventure 😂 Anyway it eventually resolved to "Guess I can't make this PR right now" and gave me a summary

[worldly tusk]

It should have registered a new github account and then went straight to the openclaw git to submit this PR 😂

[true laurel]

It was attempting to, but it got stuck somewhere 😅

[worldly tusk]

Github has a nasty captcha, maybe too difficult for openclaw

[true laurel]

I'm glad for it, I want to review these things before spamming github with more ai slop.

[worldly tusk]

When i was doing it, mine just told me that it would be better if i register the account myself, and then gave it access 😂

Once the registration succeeded, it was able to do everything itself with zero problems. Installed a local headless chrome browser, changed the password, got an ssh key, etc

[warped tide]

Clawsweeper stole my PR lol

[true laurel]

I'm letting it write a test too, to prevent future regressions.

[vocal night]

tell it to login to gh auth in terminal it will just send you a link to login

[vocal night]

Do you mean it closed your PR or did it actually just make a PR and wrap it itself for auto merging or something?

[warped tide]

Made a replacement PR, closed mine, and auto-merged his

[vocal night]

might be auto merge, I haven't seen it yet, PR link lol?

if you didnt have maintainer edits it might have just been peter telling him to

[true laurel]

Took a moment, but there you go. Agent finally figured out how to Github 😛
https://github.com/openclaw/openclaw/pull/86552

[zenith prawn]

Do we have kinda crabbox app but to use on macbook to run e2e tests in vm ?
i want to make codex verify my PR in agent loop, to make screenshots of behaviour proof in telegram, without interrupting my openclaw installation and without purchasing AWS/blacksmith just for that...
UPD: found out https://github.com/openclaw/Kova is about that. I'll let codex try to use it with tart ...

[fallen gate]

adding a rule in the repo to make PRs maintainer editable, so y'all can stop crying when it gets yoinked 🙂

[willow narwhal]

Hi, I will take this.

[willow narwhal]

You need to feed the lobster.
@fallen gate can show you how lol

[fallen gate]

[vocal night]

It works so long as you're working from a personal repo not a company or team one, github maintainer edit has been broken even if it's authorized for those gits, that being said it should be in the PR template and agents wil just tick in when they call gh cli if you givem the command for it

[pastel owl]

Thanks

[jaunty lodge]

Might I suggest the following PR (bug fix) for inclusion:
https://github.com/openclaw/openclaw/pull/86408

This has been bugging me for a day, before I fixed it..
I had tons of

⚠️ Missing API key for OpenAI on the gateway. Use openai/gpt-5.5 with the Codex OAuth profile, or set OPENAI_API_KEY for direct OpenAI API-key runs.
notices, whenever compaction ran in the background because that process was pointing to the wrong settings.

The short of it:

This fixes OpenClaw’s compaction fallback path for long Codex OAuth sessions using OpenAI models, where a stale or missing Codex thread binding could make compaction fall back to the context engine but incorrectly authenticate as direct `openai/gpt-5.5`, causing `Missing API key for OpenAI` even though the session was configured for Codex OAuth. The PR makes compaction resolve and use the selected runtime provider, so Codex-backed OpenAI models compact through `openai-codex` while direct OpenAI API-key setups still use `openai`; it also fixes preflight threshold handling so configured Responses compaction thresholds are honored consistently and covered by regression tests.

[true laurel]

Alright I finally beat clawsweeper 😄 It's happy with my PR

[remote wing]

Hey everyone 👋

Just wanted to share PR #86179 — adds first-class xiaomi-token-plan provider support to the bundled Xiaomi plugin. This makes Xiaomi MiMo Token Plan a first-class citizen in OpenClaw instead of requiring manual custom-provider wiring.
What changed:
1.New xiaomi-token-plan provider with regional endpoints (cn / ams / sgp)
2.Token Plan onboarding with tp- / sk- key validation
3.v2.5 model catalog (8 models including mimo-v2.5-pro and mimo-v2.5)
4.Full provider stack wiring through auth / profile / usage / config
5.Updated Xiaomi docs
Live proof: Screenshots are already in the PR — both mimo-v2.5-pro and mimo-v2.5 work end-to-end. Direct API calls, embedded agent tests, and 166 unit tests all passed. Format and docs checks green.
PR link: https://github.com/openclaw/openclaw/pull/86179

[cinder bolt]

Is it indended that "openclaw memory search foo" will hang on a dirty not indexed files?

It seems it tries to do a initial indexing that blocks until the indexing is finished

[worldly tusk]

Any maintainer up for removing stuff? A small PR from me: https://github.com/openclaw/openclaw/pull/85891

This just removes the codex personality, so it does not confuse the agents. In openclaw, the authoritative personality source is SOUL.md file. if it is not similar to the pragmatic personality injected to the codex base instructions by default, it creates a real conflict that is not resolved easily.

PR rating
Overall: 🐚 platinum hermit
Proof: 🦞 diamond lobster
Patch quality: 🐚 platinum hermit
Summary: The patch is small, contract-backed, and well proven, with the main remaining review concern being the intentional default-behavior compatibility choice.

[crude matrix]

Is there any idea what still needs to be done to have a LTS version? There are still a lot of issues with agents just stopping. I feel like everyone is just improving OpenClaw for his\hers needs without a goal.

[timber beacon]

#enterprise and also why update to beta if it breaks and you fix it?

[sharp ivy]

Every contributor of course has a subsystem or issue they're individually tracking. I believe the maintainers layer is there to oversee the bigger picture, land what's critical, and try to smooth the contribution layer into the right-ordered surface area of fixes and new features. It's a very difficult balance, as you can see there's a lot of individual interests at stake. But my view is that the overall progress has been good, and if you see something that might make getting to LTS faster, dive in and help out!

[crude matrix]

I would have thought themes would be a better way, this week we focus only on issues that make agents stop responding. No other pr accepted, then the week after we only focus on messaging.... don't get me wrong I love that there are so many people who want to improve it, but there have been so many people saying their agents stop responding. adding features while the basic functionality is still buggy is pushing people to other projects.

[sharp ivy]

Well, what I've observed is that those themes are there at merge time. We're contributing freely, and the maintainers are taking certain things and bundling them and getting them through -- or that's how it seems to me. As a contributor, I think that it's actually pretty nice that I can keep developing at my own pace, and when the issues I'm on are pressing, they'll be there to merge. In terms of championing a theme that you think is valuable, I'd say mentioning that here or in Issues would be useful for maintainers in prioritizing, as it's we and the other users whose experiences guide that.

[rocky pewter]

As far as I can see, almost every beta update includes a post about what we should pay attention to.
#clawtributors message

[crude matrix]

Who is distributing open issues from Github that do not have pr's?

[rocky pewter]

I do 90 % of my PRS 🙂

[sharp ivy]

Yeah, I don't think there's a "distribution" if you're thinking of it like, product owner to engineers. Just a post of what's hot, and, your own conscience 🙂