#clawtributors

1 messages · Page 13 of 1

weak walrus
#

well, what I think is really interesting is how it has the plugin - and my first thought was "ok- so now what? what happens to skill" and it's like, well, actually they coexist and the plugin will bundle skill since it will reference the plugin tools, really exciting stuff, really expands capability

#

(base) heroic@Alien-Hunter:/mnt/c/Users/heroi$ openclaw plugins inspect ai-persona-os
│
◇

🦞 OpenClaw 2026.5.18 (50a2481) — I've read more man pages than any human should—so you don't have to.

AI Persona OS
id: ai-persona-os
The complete operating system for OpenClaw agents. 24 souls (JARVIS, Thanos, Mary Poppins, Rook...), SOUL.md Maker, memory tools, Discord routing fix, native heartbeat, never-forget context. By Jeff J Hunter.

Status: loaded
Format: openclaw
Source: ~/dev/ai-persona-os-plugin/dist/index.js
Origin: config
Version: 3.0.0-alpha.1
Shape: non-capability
Capability mode: none
Legacy before_agent_start: no

Install:
Source: path
Source path: ~/dev/ai-persona-os-plugin
Install path: ~/dev/ai-persona-os-plugin
Recorded version: 3.0.0-alpha.1
Installed at: 2026-05-19T09:54:26.498Z
(base) heroic@Alien-Hunter:/mnt/c/Users/heroi$

vocal night
#

right now my sdk upgrades are only being used in china, not really announced yet with stability focus so the more builders move from stability to feature (versus stuffing features into OC 🫪) the better

weak walrus
vocal night
#

scaffolding wise just switch to 5.4 for 1m context or opus 4.6/7, 5.5 is waste of tokens to scaffold and architect it preferes issues/epics and running to build

#

run 1 thread per plugin for sure

weak walrus
#

he uses sonnet 99%

vocal night
#

thats why you havent updated skills in months 😉

#

too much time fixing sonnet breaks haha

weak walrus
#

that's how he is lol 😛

vocal night
#

swap him out with an eva coding agent much cheaper and less opinionated lol

#

eva + 6 codex subscriptions (6 desktop apps running) and you'll cook like never before

weak walrus
#

sheesh

willow narwhal
#

First thing that isn't tts from Conan lol

devout galleon
#

fair, fair 😂

potent path
#

Running out the door and no time to figure out which I've already posted and which are new. Clanker-generated summaries cause I'm running late. Sorry.

#
devout galleon
#

what's hurting people in 5.19 beta? i was using 5.18 all day pretty heavily, fast, logs clean as a whistle. single thread, no subagents, manual "save this topic away, we're going to change the subject" though

#

someone mentioned subagents were iffy earlier?

#

how's their tts?

#

that's obviously the important thing

edgy plaza
#

Some folks are reporting 3x more token burn on codex runtime vs PI
any one see any truth to this
I asked for an issue and got this

devout galleon
#

that.. doesn't seem hugely convincing. the sweeper focused on the pi/codex transition rather than the claimed 3x token usage, which seemed to be source: that guy's ass

#

lemme see what my clanker says

#

🎵 you may be a king or the humble grim reaper
but sooner or later, we all meet clawsweeper

#

(I switched it up)

runic totem
#

Quick follow-up: I opened a PR for this here:
https://github.com/openclaw/openclaw/pull/83992

It covers refresh-visible history plus draft/queued-message restore for the same gateway/session. Checks are green and proof is passing now. Preview/title metadata is still out of scope, so #83344 should stay open for that follow-up.

devout galleon
#

clanker is 30 or so tools calls in, taking a look. in the meantime i'm wondering how much of this was pi, how much was a potential 5.4 -> 5.5 transition too. 5.4 loved nothing more than to stop fast and ask if you'd like to read a 0.3kb in it's own workspace. 5.5 will go tool hunting before it even bothers to respond

#

clanker says it smells like shite

odd anchorBOT
worldly tusk
shy rain
#

oh I looked at the issue raised... it's an anecdotal "it used 3% of my limit!" which isn't exactly reproducible

devout galleon
shy rain
#

My test data says it isn't. I didn't publish the actual numbers but as part of the doc, I note that context use was higher within oc compared to codex cli. didn't compare with pi

edgy plaza
devout galleon
#

oh for sure it's higher than cli. you're injecting the claw prompt, partially or wholly, cached or uncached. that wasn't the claim in that issue though. they said 3x between pi and codex on 5.4 or worse on 5.5

worldly tusk
#

code mode basically is saying that the agent should just code their own way out of any task. so, there won't be a tool for reading files - the agents needs to call shell or whatever to do that.

#

as the agents are pretty good at coding now, there is no need to give them specialized tools and use up tokens in the structured tool syntax definitions

devout galleon
#

i'm looking at this and wondering what's the comparison though? the same question? that's not an equivalence

shy rain
#

But then why are you even doing it...

devout galleon
#

how much extra does it cost me if I install LCM? it is still the same test?

vocal night
vocal night
shy rain
#

I'm giving it a go. I decided to go for it. LCM+Cognee

#

wrong chat but anyway...

astral topaz
#

heh i wish i had codex oc integration set up...

shy rain
devout galleon
vocal night
#

either way its common sense you add more files, an extra agent.md and a LOT more to the prompt and change how everything works (the DB upgrade coming soon tho) going to have some rough edges

#

while my message sounded gloomy 😂 I have no doubt LCM will make it better

astral topaz
devout galleon
astral topaz
#

i'm slow adopter of coding harness thanks to gemini kek

devout galleon
#

ok, seriously peeps, we're shitting up clawtributors, even if it's quiet at this time of the day

vocal night
#

TLDR the team knows about issues already and has plans to make them better, tokenjuice works okay, LCM stubs works okay, and new squlite db coming soon, fixes will be in place this week most likely. In all seriousness if there is an assembly bug point it to me and I'll put in a fix on it

devout galleon
#

we're over there

terse solstice
#

usually I would make the PR for this myself, but i am not sure what is the direction wanted for this kind of security questions: https://github.com/openclaw/openclaw/issues/84144 but essentially this makes it harder to automate provisioning of nodes/gateways

shrewd mortar
#

Would appreciate eyes from anyone who can triage this Codex oauth one:

I filed #84114 earlier today (Codex harness not honouring documented multi-agent auth-profile inheritance — OpenAI OAuth rotation not firing when a secondary agent has empty or populated auth-profiles.json; bootstraps from binary instead). clawsweeper closed it as "completed" within 13 minutes, citing PR #83824 and commit 4e60ad721209.

Post-closure verification evidence doesn't hold:
• PR #83824 is test(qa-lab): add personal no-fake-progress scenario; body says verbatim "does NOT change runtime behavior"; zero files in src/agents/auth-profiles/ or extensions/codex/
• Commit 4e60ad721209 is fix(media): decode remote URL fallback filenames (#84108); also zero auth files
• src/agents/auth-profiles/store.ts is SHA-256-identical between v2026.5.19-beta.1 (tested) and current main — so the repro applies to main on the same code

So the bug-class is actually "feature-present-yet-not-firing," not "missing-feature." Likely-suspect (one-function scope) is the early-return in ensureAuthProfileStoreWithoutExternalProfiles:

const runtimeStore = resolveRuntimeAuthProfileStore(agentDir, options);
if (runtimeStore) {
  return runtimeStore;  // merge-with-main happens AFTER this
}

If the gateway-auto-created empty auth-profiles.json (39 bytes, {"version": 1, "profiles": {}}) gets resolved as a non-null runtime-store, the merge-with-main path is skipped.

Full counter-comment: https://github.com/openclaw/openclaw/issues/84114#issuecomment-4488299468

left mango
#

Hi, I have created this PR to address a user reported bug. The Pr status is "diamond lobster" and the PR also hatched a common Pearl Lint Imp.

fix(discord): preserve streamed replies after tool warnings
https://github.com/openclaw/openclaw/pull/83844

agile kiln
#

merged! thanks for the issue. Added codex to OPENCLAW_EXTENSIONS, this does remove codex from Docker pruning, so all good there!

fervent sphinx
#

what？

hollow mural
agile kiln
merry slate
#

👋 Just opened a small fix PR (+20 lines, all tests pass):

Non-vision models + media-returning tools (browser screenshot etc.) on webchat can cause the agent run to silently terminate mid-execution — the user gets an incomplete response with no error. This fixes it.

https://github.com/openclaw/openclaw/pull/84165

Would appreciate a quick look when you have a moment. Thanks!

#

btw, ClawSweeper's egg thing is so fun lol 🦞🥚

fast pagoda
#

Ready for review, all 🐚 platinum hermit, even got a 💎 rare Mossy Crabkin:

  1. https://github.com/openclaw/openclaw/pull/84045, P2, fix/cron-diagnostic-classification, fixes cron diagnostic classification so recovered tool warnings don’t override successful cron output.
  2. https://github.com/openclaw/openclaw/pull/84056, P2, fix/embedded-runner-provider-error-preservation, preserves original provider errors when session takeover occurs during cleanup.
  3. https://github.com/openclaw/openclaw/pull/84067, P2, fix/error-token-classifier-structured-fields, replaces prose-based cron denial detection with structured error signal classification.
verbal heart
vague swallow
#

just finished setting up discord with realtime voice, what a great integration, thanks all contributors

weak walrus
#

Is there documenation for that somewhere?

timber beacon
#

Yea it is

#

with oauth!

#

Tool calling and steering from VC

empty nest
#

Friendly ping: anyone with a few minutes to review this PR ? Small docs-only (+36/-0 in docs/tools/browser.md) adding a third provider example under "Direct WebSocket CDP providers", paralleling the existing Browserbase entry. Sat for ~6 days, all CI green, Codex auto-review cleared the technical surface and punted to maintainer judgment. No rush — first-time contributor, just wanted to surface it. 🦞

timber beacon
vague swallow
#

So many people don't know you can use realtime voice with codex oauth

#

Even one of the maintainers lol

timber beacon
#

It was the last update for that

#

before this last one

weak walrus
vocal night
timber beacon
weak walrus
vague swallow
#

"talk" in openclaw.json

timber beacon
#

yea I have it set so config uses it in the Talk Realtime Model and Talk Realtime Provider

fallen gate
vague swallow
#

two maintainers, then kek

timber beacon
#

I don't think most people knew/figured it could be done with oauth

#

its pushed as part of the API

vague swallow
#

minimal conf

{
  "talk": {
    "realtime": {
      "mode": "realtime",
      "provider": "openai",
      "providers": {
        "openai": {
          "model": "gpt-realtime-2",
          "voice": "cedar"
        }
      }
    }
  }
}
#

use with codex auth

timber beacon
#

cedar is great voice

rocky pewter
#

That’s really interesting. Until now, I’ve always used the API for my glasses, but I’m definitely going to take a closer look at this directly now. Thanks for the tip!

vague swallow
#

the only thing I'm missing from other providers is custom voices 😅

vocal night
#

i just noticed it's missing and I only use elevenlabs voices

#

someone want to add it or should I?

rocky pewter
vague swallow
#

But I want to record my own voice so I can build a voicemail assistant with OpenAI

#

But it's paywalled

#

Well more than paywalled

#

It's "contact sales and explain your use case" walled

timber beacon
#

Is mantis something I could set up on my own to run and add to my PRs?

weak walrus
#

hehe wait until you guys see what I'm building to bring your OpenClaw agent to your calls with you (zoom, meet, and teams)

I'll be doing a launch next week hehe (this isn't connect yet, so you can't even send me your info so don't try) https://aimeet.live

vocal night
weak walrus
#

Already have the plugin / Skill built with full 2 way and MCP haha

vocal night
#

was going to work on this but have too many projects juggling projs

weak walrus
#

OK, have to run errands be back later!

weak walrus
#

I won't lie, I use way too many apis and it cost me to much i don't epxect to make a lot of money on this, margins will be small, but the value this brings to my own life is amazing already and every call my clients are blown away

untold pollen
timber beacon
hollow frigate
echo phoenix
timber beacon
#

Its so good

timber beacon
hollow frigate
#

Dang...work has got me by the balls. I'm missing so much. Lol

sinful anvil
#

🙂

vocal night
#

Found my first serious bug of the release

verbal heart
#

@runic totem seems starting a new conversion has made that issue stabilize...

#

after the upgrade to 5.18

astral topaz
shadow garden
rocky pewter
#

@timber beacon This is insanely cool that real time works through OAuth. That was exactly the missing piece in my glasses OC setup. You’re seriously awesome, thanks man.

timber beacon
rocky pewter
timber beacon
#

Nice!

rocky pewter
# timber beacon Nice!

That was exactly the Part i was missing. This is rly awsome. no api costs anymore. 🥳

vocal night
#

codex limits reset again

#

more debugging! more releases! more claws!

shadow garden
#

Has anyone seen Clownfish recently...? feels like it went on vacation or something. Either that or I live under a rock and just never seen it

weak walrus
#

@clear violet^^ here

weak walrus
# vocal night codex limits reset again

LOL right in time eh? lol - PS @summer quartz is my developer I was telling you about, I finally talked him into using OpenCLaw a few days ago, and really he "talked himself" into it when he thought of the use case of using it to audit repos like the clawdaddy does

clear violet
#

Yeah, making a plugin to do automated security/pii/guardrail/code quality and web audits of repos and sites. It's mostly working, just not happy enough with it yet to release.

zinc dock
vocal night
#

@nocturne jungle good news message tool works in webchat. bad news it disappears when switching tabs sometimes or refreshing chat. fix is here: https://github.com/openclaw/openclaw/pull/84268 only bug I've found so far in latest release (that wasn't 3rd party plugin or weird beta npm upgrades) so everyone should be proud so far

#

any maintainer can take over above if we're doing a hot fix update. I'm going to sleep, it's like p3 but p0 for ux given users can think agent never responded

odd anchorBOT
vocal night
clear violet
#

So I’m trying to do as much of it as I can without invoking the LLM - file level first pass, then deep scan with LLM.

#

storing reports in SQLite, with a dashboard running on a different port

#

I’m also doing a full scan initially or when prompted, then incremental scans when files change and it checks every 15

#

Also trying to do it all on local models so it can be even more powerful if someone is using a big public model

vocal night
clear violet
#

My $20/month codex sub isn't up to the task so i'm testing with models on gb10. I can add guardrails to the plugin to sanitize/filter/ignore patterns it sees on the code it's auditing to reduce that attack surface. It is definately resource intensive, which is why first scans are file level from within the plugin and not using the LLM, the more in-depth scan where it sends it file to the LLM is mostly optional. If it can't do it, it fails, but you still have the file level scans.

#

I planned to build it for me first, and publish it later. I figure if I build it to work from the weakest model to the strongest, it's not a broken plugin for the dude trying to run qwen on his mac mini.

low lark
#

OpenClaw currently uses sendMessageDraft only for tool-progress status lines during generation. The actual answer text still streams via the old sendMessage + editMessageText loop which produces choppy, non-animated updates.

Telegram's sendMessageDraft API (available since Bot API 9.5) provides ephemeral draft previews with smooth, client-side typing animations pretty much the same effect you see in bots like ChatGPT's Telegram bot. This PR extends OpenClaw's existing native draft infrastructure to cover answer text streaming in DMs not just tool progress.

What changes:

  • Answer preview streaming in DM chats uses sendMessageDraft instead of editMessageText
  • A new channels.telegram.streaming.nativeTransport config toggle controls the behavior (opt-in/out)
  • Group chats gracefully fall back to the existing editMessageText approach, since sendMessageDraft is private-chat only
  • Final message is persisted via sendMessage after streaming completes (drafts are ephemeral 30s previews)
  • Edge cases handled: draft expiry, tool interruptions, long text >4096 chars, media in finals, quote reply skips

Earlier attempts caused duplicate final messages and stale preview bugs, leading to a maintainer note against it.why
This implementation addresses those issues with proper finalization gating and draft lifecycle management. comfyblob

https://github.com/openclaw/openclaw/pull/84320

naive horizon
vocal night
#

not sure if I'd say "old" send message tool, it's a codex tool we're using

low lark
vocal night
#

also I dont believe codex works that way, the send message tool is not really a draft during the whole tool process, it's a final message that wraps the end, pash is more an expert than me at it as it's his child might want to ping him 6am for me too going to bed

#

I recommend having your agent install the patch yourself (just tell it to checkout the PR as a test patch) and test it and video record it and add the video after a long tool call of it

vocal night
low lark
vocal night
#

I just shipped 1k+ loc in fixes to telegram, so I'd recommend holding off on any more features till hotfix stability patch and then we can test in next beta new features

#

and no I'm not the telegram person, just had to fix another "feature" that regressed the lanes lol

low lark
#

But the maintainer note made me concerned and took extra time to make sure things are okay.

vocal night
low lark
vocal night
#

we're no longer in "faster=better" nobody wants to break 30M+ claws on telegram haha

#

confident tho if it works and you post a video it'll get priority reviewed

low lark
#

No pressure at all, huh? It's just 30M

#

The other day we had a fairly huge release, the QA person literally said we'll just let the users test it for us. ponder

#

I'm sure you know that means

#

Not talking about Open Claw BTW.

white pagoda
low lark
#

One thing tho, the build doesn't have an API key and I get the missing key error.
The goal was to check the style which already did.

#

I'm not lazy, you are. why

low lark
#

I REALLY gotta go to bed now.

potent path
#

https://github.com/openclaw/openclaw/pull/84269 size: XS, rating: 🐚 platinum hermit
With OpenClaw code mode enabled, the exec tool discovery process is B A D and the agent spends multiple turns flailing around trying to find a working command. This PR updates the tool-card text that the model sees to clear up what will work and what will not, resulting in faster, easier tool discovery.

potent path
#

https://github.com/openclaw/openclaw/pull/76091 size: L, rating: 🐚 platinum hermit
This is not my PR, but I've applied it as a local patch and it works nicely to re-enable the "Agent is typing..." notifications in Discord so that you know the LLM has received your prompt and is working on it, instead of just 30s+ of dead air. Please consider it for merge.

lone iris
potent path
#

Yes. I'm still on 5.18 and applied the PR as a local patch just now.

shy rain
lone iris
#

Excellent I have been working on stable releases for the most part before I get into testing in beta.

potent path
devout galleon
#

my typing indicator always worked (as in it says typing). it's the fact that after the agents takes it's turn I get a second indicator for a while again saying typing. it's harmless but weird and lacking polish

warped tide
#

https://github.com/openclaw/openclaw/pull/83875
run managed service update helper and updater processes from an existing stable cwd instead of a package root/current cwd that may have disappeared during self-update
make the early isMainModule(...) startup guard tolerate ENOENT/uv_cwd when the process cwd has disappeared
add regression coverage for both the handoff cwd and early startup cwd fallback paths

https://github.com/openclaw/openclaw/pull/83938
disable tar hardlink emission during backup creation so hardlinked files are archived as regular entries
include hardlink targets in backup verification and reject traversal/absolute targets
add regression coverage for hardlinked backup payloads and unsafe Link headers

https://github.com/openclaw/openclaw/pull/83947
ignore Kitty printable key-release events in the TUI custom editor so release events cannot duplicate text
decode German-layout AltGr Kitty CSI-u printable input when terminals report Ctrl+Alt plus a base-layout key
add regression coverage for AltGr @/€, ordinary Alt-modified input, and printable release events

karmic steepleBOT
warped tide
#

quick wins

midnight walrus
#

Sad to say, but we had a severe downgrade since it was decide to move to codex agent runtime, I been fighting my claw since .12. And I've tried a bunch of things to fix it. At .7 was working pretty fine.
Tested now with xai/grok-4.3 running on pi agent runtime back and it came back live as sharp as it used to be.

#

I'm here to help, and being on beta to test.

jolly wolf
midnight walrus
crystal relic
#

https://github.com/openclaw/openclaw/pull/84387 ready for review, size: XS, rating: 🦞 diamond lobster
It keeps slow node sockets from absorbing unbounded Gateway fanout by closing over-buffered connections before another event is queued.

shadow garden
#

Tak was nice enough to kick off automerge on these PRs, though they're both stalled now:
Fix for sent messages while agent is mid turn getting dropped:
https://github.com/openclaw/openclaw/pull/84219 (hung)
Fix for regression causing tool call failures to flood after chat turn ends
https://github.com/openclaw/openclaw/pull/84354 (requires human review)

And then a PR to fix realtime talk flow in WebUI and Android, it actually flows like a chat now similar to voice-call and does proper tool calling again!
https://github.com/openclaw/openclaw/pull/84181

shadow garden
midnight walrus
timber beacon
#

@shadow garden I’ve been digging into tool call use from VC I got interruption, steer, and cancel working excited to see what your improvements bring as well

#

Voice has been so good to use

shadow garden
timber beacon
#

Might have to try it out! I’ll make sure my PR does not conflict but supports yours as well as you might have some stronger set up.

midnight walrus
timber beacon
#

I had those codex issues before that last few updates. @midnight walrus what version you on? I had codex scrub my config with the docs on the harness to set it up and it’s so great right now

midnight walrus
#

stable and beta. both latest.
both the same.

#

stable on mac.
beta on linux vps.

midnight walrus
astral topaz
#

https://github.com/openclaw/openclaw/pull/82596
Feature/exec denylist
PR rating
Overall: 🐚 platinum hermit
Proof: 🦞 diamond lobster
Patch quality: 🐚 platinum hermit

Get explicit maintainer sign-off on hard-deny precedence, default managed curl/wget rules, malformed-rule fail-closed behavior, and prompt-on-match scope.
this Maintainers, please take a look

terse solstice
vocal night
#

@jolly wolf if anyone says 18 is breaking mid turn or stalling and failing to respond (its a bad p0 that perma stalls gateway until restart and you won’t know the cause unless you use debug mode) we should ask if they are on lossless claw.

#

18 else-wise I’ve thoroughly tested and can say it works flawlessly without any plugins and via telegram on codex (i haven’t tested every platform under sun) I found 1 small message bug thats fixed.

left mango
midnight walrus
vocal night
#

Its fixed now and works better than ever but re assembly + codex compaction so it matches native codex (needs to otherwise codex won’t continue turns etc)

#

If you are using it and need the beta patch PR’s (they are merged on main so just pull main or ping me for the 3 pr’s)

edgy plaza
midnight walrus
vocal night
#

We try to get one out asap once a final beta is out now

vocal night
midnight walrus
# vocal night FYI - If you are using lossless claw we found a regression that perma stalls age...

codex said: Confirmed and mitigated.

Lossless-claw itself was not active in the live config: contextEngine was already legacy, no lossless-claw entry, and the plugin list didn’t show it active. But the risky part was still enabled:

agents.defaults.compaction.midTurnPrecheck.enabled: true

That matches the fresh stall we saw at 01:00-01:03 Sao Paulo: Codex got through tool:bash:ended, then stopped progressing. I changed only that flag to false.

Validation passed:

  • openclaw config validate passed
  • gateway restarted successfully, new PID 9499
  • openclaw gateway status --json --require-rpc returned RPC ok: true
  • openclaw status --json --all shows active: 0, queued: 0, running: 0
  • post-restart log window: 5 strict Codex runs, 0 stalled sessions, 0 recoveries, 0 timeouts

I’d keep stable on contextEngine: "legacy" and midTurnPrecheck.enabled: false until the next OpenClaw stable release or an approved beta upgrade.

vocal night
#

Good it should work now, don’t re enable lcm until alpha 2 or beta 1 today.

#

Any maintainers see vague issue on stalling ask if they have lcm enabled or config leftover

low lark
crystal relic
#

https://github.com/openclaw/openclaw/pull/84408 ready for review size: L, rating: 🐚 platinum hermit

Moves the node systemd gateway token out of unit files and into a mode-600 env file, closing a real Linux credential exposure path while preserving upgrade and uninstall behavior.

#

hi maintianers! that's PR is ready to review, shoulw we continue?. I need this PR to enhanced observational capabilities are intended for subsequent ability optimization.

https://github.com/openclaw/openclaw/pull/83019 🐚 platinum hermit

astral topaz
low lark
#

Wait the 3rd one isn’t a bug tho

#

I need approval from TG owner

crystal relic
cinder bolt
pliant knot
#

I intake as much data as I can

#

U gotta

low lark
#

The review workflow is brutal!
It is walking me in circles!

vocal night
#

@worldly tusk just finished parity and regression testing on the prompt changes posting updated PR in a second

edgy plaza
vocal night
#

Turn LCM off until next release 100% it is P0 fubar and will break your claw, requires OC fix + LCM fixes on main but not in any release. I can give you the PR's to self patch if you want.

edgy plaza
#

Why are clankers soo unreliable in telling what’s going on inside it LOL!
A great skill we can make is one that helps everyone’s claws actually diagnose what’s going on.

edgy plaza
vocal night
#

There's always one more codex change that gets slipped in and breaks how LCM works. Trying to keep up with it is a few hours a week for me.

edgy plaza
#

I disabled active memory and my life has been PEACE

edgy plaza
vocal night
#

my next PR literally says "codex team vibe broke this again"

#

thats half my PR's historically

#

they know I have no love

low lark
#

Guys

#

Have you seen this?

#

I am losing my mind rn

vocal night
edgy plaza
low lark
#

This is just... a bomb

worldly tusk
edgy plaza
vocal night
#

I've settled into grumbling here and just being the slop cleaner

odd anchorBOT
vocal night
#

if next release is hotfixes and stability and fast, we're in good territory

worldly tusk
edgy plaza
vocal night
#

as you said, codex needs many things to be optional or configurable not default flip overs

vocal night
# edgy plaza I love codex!

It's also extremely complex system now so I have to have waves of agents to make decisions, read 10k+ loc sometimes to make sure a decision is right and run all kinds of adversarial reviews, AND then test.

cinder bolt
#

I'm getting this now on latest beta.

⚠️ Cron job "OpenClaw Beta Watcher" failed: EmbeddedAttemptSessionTakeoverError: session file changed while embedded prompt lock was released: /home/ping/.openclaw/agents/main/sessions/ac6a3bcd-648f-4bf4-b143-3bc493b9f0d9.jsonl

vocal night
worldly tusk
#

Oh wow!

#

Nice

vocal night
#

TLDR is simply that moving the Soul File to developer instructions and out of every turn causes personality degradation.

Ultimately, the user should be able to choose their personality settings and keep their current OpenClaw experience running the way they like it vs token "savings" that degrades personality.

@nocturne jungle https://github.com/openclaw/openclaw/pull/84331 under your new PR when you make changes, they don't carry through until new thread or session.

The complexity of this new system, combined with the current session system and threads, is genuinely difficult for anyone to navigate, so I'm not here to just criticize — I just want to explain that this is the right fix.

I've run multiple agents, waves, and tests over hours, and it ultimately comes down to this: users should have the choice.

Giving them that choice doesn't break anything. If Codex users want a coding agent with a personality that essentially regresses to the point where you might as well not have added it in the first place, then that's their call — but everyone else deserves better.

nocturne jungle
vocal night
#

I spent like 50m tokens testing

nocturne jungle
#

There is absolutely no reason to duplicate the same thing to the model multiple times in its context window

vocal night
#

That being said a first run agent will come to your conclusion

#

has to look over all of the code and how it actually works and how threads resume/start etc

vocal night
#

Thats not how context works

#

Context window contents are just the current request payload assembled for that turn. If a long-running native thread, resume path, compaction, or prompt rebuild stops carrying the SOUL/profile material, the model does not magically retain it as authoritative instruction.

#

So the tradeoff is not “duplicate or don’t duplicate.” The tradeoff is:

per_turn_context: quality/freshness first; user-edited files are actively present each turn.

thread_developer: token-efficiency/priority first; stable files move to the native thread instruction surface and require thread rotation/fingerprints for freshness.

worldly tusk
#

Aren't base_instructions and developer_instructions both constructed and injected every single turn? I don't get it, how they can disappear. User instructions perhaps may be compacted, but not developer instructions

nocturne jungle
#

Incorrect framing. Anything that should only be present once should be in session-level context. If you are injecting a file into context, you should not re-inject the same thing multiple times. Multiple turns can happen in a single session.

The better fix is to fingerprint the assembled developer instructions/SOUL block, and when it changes, force Codex onto a fresh/cold thread path so the new developer instructions actually bind.

#

That being said, I don't expect SOUL.md and similar docs to change often.

vocal night
#

I did add that fix for your PR

nocturne jungle
#

Additionally, it would break the prompt cache every time one of those files changed.

#

That's why I explicitly left MEMORY.md out of it, as it changes more often

vocal night
# worldly tusk Aren't base_instructions and developer_instructions both constructed and injecte...

The actual risk cases are:

-If thread/resume.developerInstructions is not guaranteed by Codex to replace/reassert the developer layer on every resumed turn, then thread-level SOUL can stale.

-If OpenClaw’s fingerprint misses a file or mode change, native threads can continue with stale thread-level instructions.

-If users expect SOUL/USER/TOOLS edits to behave like legacy OpenClaw current-turn context, thread_developer changes that mental model.

-If we place SOUL in both turn context and developer instructions, that is wasteful duplication and should be avoided.

The real distinction is authority and lifecycle. thread_developer makes SOUL/USER/TOOLS part of the Codex thread developer lane, with freshness mediated by binding fingerprints and thread rotation. per_turn_context makes those files visible as current-turn OpenClaw context every turn, matching legacy behavior and user expectations for frequently edited workspace files. Both are valid; users should choose.

vocal night
shell dawn
#

andy will your patch let me get rid of the 'use message tool to reply' extra system prompt that codex runtime adds. my agents got a little confused with that at first and now some of them still throw an ack in every once and a while at the start of their response

vocal night
#

I run this same setup on my codex, and personality regression happens within 20 turns at session state versus turn basis. At the end of the day would 50% or more users prefer personality adherence or save a few tokens? That's why it's really a toggle

worldly tusk
#

If what's included in developer instructions degrade, then md files are not the only problem… there are some other things as well…

vocal night
shell dawn
#

sorry i havent scrolled back far enough 🙂

vocal night
odd anchorBOT
remote wing
#

Hi! A few PRs are ready to merge and could use a review 👀

  • #84444 — fix(memory): stop recall tracking when dreaming is disabled
    🦞 diamond lobster | size S | memory-core
    Fixes dreaming.enabled=false still writing .dreams recall artifacts during normal memory_search.

  • #84448 — fix(memory): strip invalid thinking signatures for signed-thinking providers
    🐚 platinum hermit | size S | agents
    Fixes incomplete thinking blocks (missing signature) persisted to session JSONL → Invalid signature in thinking block on next request.

  • #83480 — fix(minimax): normalize OAuth token expiry to absolute millisecond timestamp
    🐚 platinum hermit | size S | minimax
    Fixes MiniMax Portal OAuth expired_in being stored as relative seconds instead of absolute ms, causing every request to stall 30–50s in auth.

  • #84133 — [codex] Guard published dist against deprecated pi runtime imports
    🐚 platinum hermit | size S | scripts
    Adds a postpublish verifier guard to catch stale @mariozechner/pi-* imports in root dist before they ship again.

Thanks！

plucky sentinel
#

PR: fix: context-aware reserveTokensFloor in overflow recovery hint https://github.com/openclaw/openclaw/pull/84399

Fixes #65839 — The context overflow recovery hint hardcodes reserveTokensFloor at 20000, which is too low for models with larger context windows (200k/1M). Users following the advice still hit overflow.

This PR replaces the hardcoded value with dynamic tiers based on the effective runtime model's context window: 1M→100k, 200k→50k, 100k→35k, default→20k. It also accounts for agent contextTokens caps and fallback runtime models. 22 tests passing, ClawSweeper rating 🐚 platinum hermit, status: ready for maintainer look.

cinder bolt
shy rain
shell dawn
cinder bolt
shell dawn
#

i looked into it when one of my claws basically told me 'the system prompt is telling me to use the message tool to reply which caused the message to go to the wrong place instead of just replying in the webchat'

cinder bolt
vocal night
shell dawn
vocal night
#

I’m actually a fan of the tool it is pro personality it just has some bugs

vocal night
cinder bolt
#

I wasn't aware it have changed. Looked at my claw and this was at 17th may so

shell dawn
vocal night
cinder bolt
#

It's hard to keep up with the development speed. I've created a Cron job to help me out but it seems crons are failing att the moment for me with the EmbeddedAttemptSessionTakeoverError. Is this a known issue or should I investigate?

#

Going off for a bit, real life calls 😀

vocal night
edgy plaza
low lark
silver birch
#

please check this and let me know if its a good suggestion

left mango
drowsy roost
#

Is there a particular channel / thread / room in this Discord server where people are discussing use of multiple agents in channel rooms (as in, more than one agent in a group, multiple or all of which can respond to the user and/or each other)? I’ve worked a lot on this and it was suggested to contribute here, but I see no dedicated group for it. If there isn’t one, that’s fine, but it sounded like there was. See, e.g. now-somewhat-outdated-but-still-very-relevant PR 73455 https://github.com/openclaw/openclaw/pull/73455 to enable AI-gated contextual participation gating and new-room-message-awareness/freshness checks prior to delivery of agents’ messages in Matrix rooms. This has relevance to all channels in which multiple agents can be added to a room/group.

silver birch
drowsy roost
#

Separately PR 84111 https://github.com/openclaw/openclaw/pull/84111 is designed to stop the rather user-unfriendly and functionality-breaking absence of an interactive prompt in “openclaw doctor” (or, better yet, in “openclaw update”) offering users who use multiple agents the option of conforming all non-primary agents’ auth-profiles.json to main’s after running openclaw update (or whenever they diverge). This would prevent the post-update “could not decrypt legacy OAuth sidecar” issue that arose after updating to 5.18 and also help users to get their multi-agent setup fully functional again anytime that a legacy/outdated/rate-limited/different auth-profile is lingering unintentionally in a non-main agents’ auth-profiles.json.

I was unsure how to proceed in terms of implementation / testing, but I do think I’ve identified a problem needing to be solved; I encounter auth-profile drift across agents far too often and it is a real blocker.

verbal temple
#

Hi, is there any restriction to contribute adding cursor cli to the openclaw configure command?

low lark
#

I'm still waiting to know the lead maintainer for Telegram. why

#

My time is extremely limited and I know for a fact my PR will soon have conflicts and don't want to let this PR go to waste.

#

The review agent is pretty much moving in circles so the PR is fully tested and proof is added too but since there was a comment saying not to do it, I have to get to the maintainer in charge of TG.

silver birch
low lark
#

Unless you mean CLI?

silver birch
# low lark That's not practical BTW.

it kind of is, it allows you to purchase a monthly subscription that dosent run on credits so it cant run out unless you dont pay for the next month

#

basically for gemeni ai

#

its like 30 per month for pro

low lark
#

I meant adding that, technical wise.

silver birch
#

so if you log in with the account you have gfemeni pro with onto your open claw then you can connect your gemeni to the open claw

astral topaz
#

that's just an issue, no code/pr attached

low lark
silver birch
silver birch
# low lark That's not how it works tho, there has to be either an API or CLI

When a provider officially supports subscription-backed or OAuth-backed API usage, OpenClaw should let users connect that provider account through the normal model/auth setup flow and use it as a model provider.

The UI/CLI should clearly show whether the connected account is using API credits, an OAuth profile, or an officially supported subscription-backed plan. If a provider does not officially support subscription-backed API usage, OpenClaw should not try to bypass or scrape the provider’s web app.

astral topaz
silver birch
low lark
silver birch
astral topaz
#

heh, adding feature is bad for blood pressure. not gonna do more until current PR gets merged

low lark
low lark
silver birch
astral topaz
low lark
low lark
silver birch
#

but still because obviousl for api credits, each message you send costs money, for a monthly subscription, theres no limit, so you can kinda have it for unlimited use

low lark
silver birch
low lark
#

But I understand your point, API token is Pay as you go, basically token in and out.

low lark
silver birch
low lark
silver birch
#

learnt from freecodecamp

low lark
silver birch
low lark
silver birch
#

i just need to learn in a way that engages me otherwise im bored

low lark
#

Are you a book person? If so I'll introduce a cool book you can start with.

silver birch
#

nah not really i have adhd so i cant do that

low lark
#

I feel like we're violating thia channel's purpose, I'll write in general.

left mango
tulip gulch
pale crane
runic totem
# silver birch **if anyone is looking to make openclaw cheaper check this out - https://github....

I looked into #84528 a bit. My read is that this is not really a missing generic feature: OpenClaw already has some OAuth/subscription-backed paths like OpenAI Codex OAuth and xAI OAuth, plus Gemini CLI OAuth reuse.

I think the right framing is probably docs/clarification rather than implementing a generic subscription layer. OpenClaw should not scrape or bypass web app subscriptions. Any remaining gaps seem provider-specific.

I’ll leave a comment on the GitHub issue with that framing unless maintainers prefer a small docs PR.

untold pollen
#

I've had to close 10+ PRs that implemented sendMessageDraft so far

#

We had this for a couple of releases when it was launched and it caused a ton of bug reports. It's meant to be ephemeral streams

low lark
#

I've added video proof.

left mango
untold pollen
low lark
#

Additionally there is a graceful fallback as well. cat_sweat

untold pollen
shadow garden
low lark
smoky dock
vocal night
#

Shoutout to peter and his repos on mac apps. My agent deployed our mac app today no issues and holy shit for some reason the web ui dashboard works 10x faster in a local app remote from vm vs local dashboard. Blazing fast difference.

weak walrus
#

I prefer it over the TUI now first time ever

odd anchorBOT
vocal night
#

GIT out of here, Jeff — go build those plugins! I can't wait to test them! 😄

midnight walrus
#

wanna thank all the team here and clawtributors, i've being debugging, fixing, trying thins in stable and beta since yesterday to make codex agent runtime run smoothly and sharp as pi was. being having some improvements, all findings are being commented in existing issues on GitHub to help maintainers drive the the solutions and other users as well.

slow inlet
smoky dock
#

current main is red, looks like failures are from mainline drift around OpenRouter lint/package guard/fs-saf is someone 👀

karmic steepleBOT
astral topaz
#

the thing is the posture change is justified, so going back weakens standing

weak walrus
fallen gate
smoky dock
rocky pewter
#

was just messing around with this in real time using voice, and honestly, I really like the new personality feature. It gives the real time experience a bit of life again, which it was definitely missing. Shout out to @fallen gate

odd anchorBOT
vocal night
#

The cat is back

astral topaz
#

does that mean his retirement failed?

fallen gate
astral topaz
left mango
warped tide
#

https://github.com/openclaw/openclaw/pull/84652

Mirror the accepted Codex app-server user prompt as soon as the turn has a stable id.
Reuse the existing turnId:prompt mirror identity and thread-scoped idempotency so the final transcript mirror dedupes the prompt.
Add a regression test that checks the session transcript contains the user message before turn completion and still only has one user row after final mirroring.
Fixes #83528

weak walrus
zinc dock
rocky pewter
willow perch
rocky pewter
#

v2026.5.19 is out, time to test 🥳

vocal night
#

it ties into the context engine fixes in 2026.5.19

zinc dock
void bramble
zinc dock
void bramble
zinc dock
#

but i felt this difference before this landed

zinc dock
void bramble
zinc dock
void bramble
# zinc dock Sorry this one https://github.com/openclaw/openclaw/pull/83945

My clanker says this in response 😂:

AGENTS.md can change agent behavior, not the underlying runtime delivery path. It can make the parent agent choose a different workflow, for example:

  • use wait_agent after spawn_agent
  • avoid background delegation
  • ask the user to ping later
  • use OpenClaw sessions_spawn instead of Codex native spawn_agent
  • tell the subagent to post somewhere directly if it has a message tool
  • keep the parent turn open until the child finishes

Those can make the problem appear fixed, because the parent agent is now polling/waiting or avoiding the broken async path.

But AGENTS.md cannot by itself add:

  • durable child completion detection
  • transcript/event reconciliation
  • parent wake/enqueue
  • visible Discord delivery
  • retry/idempotency if delivery fails

That has to exist in OpenClaw/Codex harness runtime code.

zinc dock
zinc dock
clear violet
void bramble
# zinc dock Dm'ed you with the prompt

If you're up for it, try my PR, let me know. You could wipe out that entire agents.md section 😛 "Just use spawn_agent, don't wait for it, let me know when it's done".

That, plus the work clanker and I did to get reasoning streaming working in discord (like claude or codex app), it's glorious.

rocky pewter
#

Hi Maintainers, hope you're doing well. I opened a small bug/stability PR:
If you have time to take a look 🫡
https://github.com/openclaw/openclaw/pull/84628

This fixes a startup issue where restrictive plugin configs could still make OpenClaw scan all installed plugin manifests. It keeps the narrow path safe, and refreshes old indexes instead of letting them stay on the slower fallback path.

tulip gulch
terse solstice
#

noticed i've been getting those LLM request rejected: messages.87.content.69: Invalid signature in thinking block (with v18, is it something addressed in v19?)

🦞 OpenClaw 2026.5.18 (50a2481)
🧠 Model: anthropic/claude-sonnet-4-6 · 🔑 token (anthropic:default)
⚙️ Execution: direct · Runtime: OpenClaw Pi Default · Think: adaptive

pale crane
astral topaz
rocky pewter
astral topaz
# rocky pewter The team will definitely get to them. I regularly see older PRs of mine getting ...

that is the thing, my PR can't get to "status: 👀 ready for maintainer look" step because and i quote

Next step before merge
The remaining action is human maintainer/security acceptance of the new exec policy contract, not a narrow automated repair.

so literally in order to get it looked at, i first need to get it looked at... if maintainers were to filter by status, it would never surface lol

#

the logic error bothers me, a lot

rocky pewter
timber beacon
#

And certainly areas of change get harder to get a review due to impact it can have

astral topaz
#

I don't think i mind waiting, i'm just seeing a glaring bug and it makes my hands PR itchy

#

alright, back to work on Kova, i think it's lovely and i haven't worked on it for 2 days

timber beacon
midnight walrus
rocky pewter
#

~~Hi Maintainers, hope you're doing well. I opened a small Codex stability PR for the current version:

If you have time to take a look 🫡
https://github.com/openclaw/openclaw/pull/84736

It fixes a Codex app-server crash when a bootstrap/context file comes in with path/content but no name. The report path now keeps the path-based lookup, derives a safe display name from the path, and no longer calls the basename helper with undefined.~~

tulip gulch
left mango
remote wing
#

Hi maintainers, could someone take a quick look at this PR?
It's already ready to merge with comprehensive real behavior proof included. Current rating is diamond lobster 🦞. Appreciate the help!
https://github.com/openclaw/openclaw/pull/84444

tight wolf
potent path
vocal night
#

How’s 19 any major p0 bugs so far?

obsidian zodiac
#

Havent put in a PR/Issue for it, cause I'm almost certain theres already one for it, but openclaw [plugin] --help returns OpenClaw's help menu, rather than the plugins help menu, despite metadata being injected/present.

jolly wolf
thick ospreyBOT
#

⚠️ 🛠️ gh api repos/openclaw/openclaw/pulls/63336 --jq '{number,title,state,html_url,user:.user.login,body,merged_at,updated_a… failed

jolly wolf
#

so true

obsidian zodiac
#

From my Codex(shortened haha):
PR #83841: “Fix memory plugin CLI help dispatch”
Merged on May 19, 2026. This is the closest match. It says it makes root CLI help config-sensitive for plugin command descriptors and specifically fixes memory-plugin help dispatch.

Issue #81077 plus PR #81083
This is the more generic bug: openclaw <unknown-command> --help incorrectly shows root help. PR #81083 is still open from what I found.

I'm still on 5.18, so i havent checked to see if that merge in 5.19 helps or not

#

Quick temp install test:
Result:

2026.5.18: openclaw helptest --help incorrectly showed root OpenClaw help.
2026.5.19: openclaw helptest --help correctly showed plugin help
So yea it looks like it got fixed in the latest update. I probably should have tested the new update before mentioning it, my bad.

zinc dock
#

5.19 Pretty solid for me all day ...even tested Grok Oauth and x search all good

silver birch
#

I don’t see an exact open PR for “known plugin root openclaw <plugin> --help falls back to top-level help.”

Closest matches:

  • #81083 fixes unknown command roots with --help, not known plugin command help.
  • #83841 merged a memory-specific plugin help dispatch/config-sensitive root help fix.
  • #84786 is help fast-path perf and explicitly leaves plugin-sensitive/nested help on the live path.

So if nit3stalkr has a concrete plugin id/repro, I’d treat it as probably still worth filing unless it’s specifically the memory/LanceDB case fixed by #83841.

https://github.com/openclaw/openclaw/pull/81083

zinc dock
#

Kinda crazy ...just logged in with my x account that has a premium sub and it was able to use all the tools

shell dawn
#

everyone is clawing

rocky pewter
obsidian zodiac
rocky pewter
fossil ferry
#

Hi maintainers, I opened a PR to forward more sampling params through the OpenAI-compatible /v1/chat/completions path:

https://github.com/openclaw/openclaw/pull/84094

It threads client-supplied frequency_penalty, presence_penalty, and seed through gateway streamParams, runtime stream wrapper, then upstream OpenAI Chat Completions transport.

It also adds gateway validation for invalid penalty ranges and non-integer seed before provider dispatch.

This PR is intended to close https://github.com/openclaw/openclaw/issues/32496. CI is green and ClawSweeper marked proof sufficient / ready for maintainer look. Main question is whether we accept unconditional forwarding for these standard Chat Completions fields, or want provider-gated forwarding instead. Feel free to take a look when convenient.

karmic steepleBOT
vocal night
#

anyone else getting telegram regression

#

was fine pre 19 and now immediately after my agents repeating messages

#

looks like messages sent while down can still pile up in queve and then loop

#

loopback later*

vocal night
final ocean
final ocean
final ocean
vocal night
# final ocean which version and which model you running?

latest version, 5.5 model, I thought I squished this bug the other day. messages got stuck in queve from when agent was still down, and then fired back later. I fixed it going into 18 and didn't see it again, but just got 2 messages back from when it was down a little later

#

not p0, and only related to telegram queve refires

final ocean
vocal night
#

I consider it p0 for UX because it leads to users complaining that the agent never responded

final ocean
vocal night
#

separate issue related to send message tool and webchat (ive PR'd and we've merged a few fixes on it but it really is gonna need UI dashboard fixes longterm) these are bandaids

#

telegram has just been having minor regressions here and there from the improvements made, I personally think it should never queve and refire messages from when it was down (99% of the time it's hey hey ??) if anything if it's down I'd combine all messages into one post reconnect message but tbh I'm not the telegram guy)

plucky sentinel
#

Hi everyone! 👋 I have 4 fix/refactor PRs currently waiting for maintainer review. Just wanted to check in and see if anyone has time to take a look:

#84852 fix(config): append numeric bound hints to ceiling/floor validation errors — Size S, platinum hermit, proof: sufficient, no merge-risk. Fixes #52500. Three previous PRs (#62091, #73220, #83916) were all closed due to author-side issues, not code quality.
#84399 fix: use context-aware reserveTokensFloor in overflow recovery hint — Size M, platinum hermit, no merge-risk. Fixes #65839. All 3 P2 review findings resolved.
#82100 refactor(config): extract GoogleChat schema — Size S, platinum hermit, proof: supplied. Follows samzong's #82007 channel schema split pattern.
#83955 fix(diffs): eliminate toolbar icon XSS sink via ToolbarIconName union type — Size XL, platinum hermit, merge-risk: availability. Screenshot verification uploaded in comments.

vocal night
#

honestly #84399 is a good bandaid to a really poorly design config patch

#

I'd rip out the overflow recovery config entirely, conext management systems manage this already, and so does codex native compaction, it's legacy bandaid someone added in that we forgot to take off

#

at best it "overly" protects, at it's worst it messes with how context engines calculate tokens (LCM had to patch around it), it was added when runtime was half/half, now that codex owns it's own compaction it's just conflicting code

#

I'd love to see PR's taking out code 😂

remote wing
#

Hi everyone 👋

I've submitted a few PRs that need review:

Each includes solid real-behavior proof. Happy to iterate quickly based on your feedback!

Thanks 🙏

plucky sentinel
#

Thanks for the context! That makes sense — if the overflow recovery config itself is legacy that conflicts with Codex's native compaction, patching it further doesn't make much sense. Happy to close #84399 if you'd prefer to rip it out entirely instead. I'll focus on other fix PRs in the meantime. @vocal night

vocal night
#

Not saying close it, it might be some maintainers baby haha. I just have spent hours working around it. It's there because of token approximation rather than accurate token calculation. Because it's an estimate, it builds in a bit of buffer room to account for the difference but it's a bit wide and codex handles it's own now- and pi never needed it.

#

Taking the config feature out would be good, but also just putting it to 0 nulls it out, it's not a critical issue, just technical debt (taking it out might break something else) and fixing it isn't really necessary.

It was more LCM related when we were doing every turn compaction and deferred compaction cache windows. We now do 90%>30% (industry is 20% final compaction) but we've found that to be too much.

Legacy compaction doesn't need the overflow, and codex compaction has it built in now, so it was a patch for LCM that LCM didn't need.

#

hope that makes sense trying to educate at same time

#

TLDR- codex does long tool calls and could overflow because LCM wouldn't let it compact mid turn. All 3 compaction methods (legacy+lcm+codex) all do mid turn compaction and align now. TBF tho ask your agent @plucky sentinel to check compaction defaults and mid turn legacy wise (for pi) I might not have updated it when we put these fixes in place.

plucky sentinel
#

Will do! I'll check it.

vocal night
#

the best feedback today would be anyone running long tool runs and if the main agent can go 100+ tool calls > compact > keep going would be signal for stability

plucky sentinel
#

That's a great stability signal. I can set up a long tool-run test locally🐱

vocal night
#

best to test it real time in building something and telling it to loop a few times and give a final report after enough to blow its context window for a compaction so like 30-45 mins of work (best to use tokens for something useful, you seem smart, maybe build a plugin for yourself with your agent- give it these new docs for SDK and talk with it on which plugin to build- tell it to check the recipes https://github.com/openclaw/openclaw/pull/74853) "check these recipes and see which plugin recipe or sdk seams would help us most in what we do" and have it build the plugin @plucky sentinel that would be amazing test

#

also would help you make something cool on new sdk updates

shy rain
#

this was on 5.16beta7 maybe?

vocal night
shy rain
#

ah I don't have such a long job for it to do today

#

plus I'm already using LCM

vocal night
#

that's how I normally find bugs, I assign a watcher in claude code > tell openclaw go build something and when claude code finds a bug it tells my agent who then fires off a codex desktop thread on it

vocal night
rocky pewter
#

Hi Maintainers, hope you're doing well. I opened a small Ollama stability PR for the current version:

If you have time to take a look 🫡
https://github.com/openclaw/openclaw/pull/84855

It preserves native Ollama tool-call IDs during replay, which fixes the Gemini/Ollama Cloud follow-up turn failing with the missing thought_signature error.

remote wing
#

The PR(https://github.com/openclaw/openclaw/pull/84444) has been labeled with automerge, but the process failed due to the Git RPC/network error mentioned above. Could a maintainer please look into this or trigger a manual retry to resolve the sync issue?

plucky sentinel
shrewd lichen
#

Small idea related to the Contributor / Maintainer dashboard:
https://github.com/openclaw/openclaw/issues/84599

Would a contributor recognition / quality-signal view make sense there?

I do not mean a raw PR-count leaderboard (so called "EGOBOARD"), because that could create bad incentives and encourage low-effort PR volume. I’m thinking more about an impact-weighted view based on existing labels and accepted maintainer signals.

For example, it could highlight contributors who consistently help with:

  • merged fixes linked to P0/P1/P2 issues
  • PRs with tests, regression coverage, or clear proof
  • high-quality reproductions that make issues actionable
  • useful triage that reduces maintainer load
  • reviews that catch real problems before merge
  • docs fixes that reduce repeated support-like questions
  • small but clean fixes that move queueable work out of the backlog

I would avoid publicly ranking maintainers. Maintainer work is harder to score fairly because it includes judgment, review, prioritization, risk management, saying no, and protecting project quality.

For contributors, though, a rolling 30/90-day “high-signal contributors” view or tiered recognition section could make high-quality, low-noise contribution patterns more visible without turning the repo into a PR-count competition.

Important guardrails:

  • no points for merely opening PRs
  • no raw volume-based ranking
  • only merged, accepted, or maintainer-labeled contributions count
  • caps or lower weight for tiny/cosmetic changes
  • extra weight for proof, tests, linked issues, and higher-priority work
  • maintainership remains a human trust decision, not a score threshold

The goal: make high-quality contribution patterns more visible.

Just a small idea to make contributor life a bit more interesting and enjoyable. 🙂

vocal night
final ocean
final ocean
final ocean
vocal night
final ocean
vocal night
#

wasn't josh who added overflow config to OC.

drifting nymph
versed kindle
vocal night
#

they weren't in 5.18

#

@untold pollen ^ is this from merging anyting new in 5.19? did you add williams pr lol

#

basically telegram is unusable for me now

cursive palm
#

I don't know if it's a problem with my model or something else. I feel like sometimes my agent forgets what it just said, for example, just now I asked it to plan something, and I thought it was good, so I let it execute, but it replanned and then executed again.

untold pollen
versed kindle
#

@rocky pewter curious, what is the motivation behind using gemini through ollama?

#

and not directly?

vocal night
rocky pewter
vocal night
# untold pollen No, haven’t merged anything yet

2026.5.19 changed the progress-preview / runtime recovery surface enough that the older partial-preview overflow behavior became visible and nasty during fallback/session-rotation churn.

Most suspect commits:
d60ab48511 / PR #83847, Add Telegram progress preview flows

Highest suspect for the visible Telegram preview spam. It changes extensions/telegram/src/bot-message-dispatch.ts (line 603) and the channel-streaming helpers.

c49d909b60, Slack-only persistent inbound dedupe
Not a culprit, but a smoking contrast: Slack got durable inbound delivery dedupe in this release range, Telegram did not. That supports #84886.

f526d96c98 / PR #83829 and a559ccc084 / PR #83827
Likely amplifiers for Telegram topic/followup concurrency, not the direct root of duplicate DM preview sends.

a059309a9f
Relevant because the live logs show LCM compaction/replay/rotation and EmbeddedAttemptSessionTakeoverError right as Telegram previews kept sending.

versed kindle
rocky pewter
# versed kindle so ollama gives free gemini access?

Well, it’s not exactly free access since you still pay for a subscription there. But to answer your question, yes, if you have a subscription, you can pull the model from there.
I think a lot of people in the community use Ollama and would really appreciate being able to use Gemini through it, especially because of the 1 million context window.

#

@versed kindle Thats great thank u ❤️

versed kindle
#

if possible, limited

rocky pewter
# versed kindle not sure if I ask on huggingface side for an ollama sub, not sure how many peopl...

I unfortunately can’t really answer how many people would use it, but once it’s fixed and word gets around, I can definitely imagine a lot of people using it as their secondary or third LLM.

I tested it overnight on my Smart glasses, and I can tell you wow, it’s incredibly fast with tool calls.

And regarding your other question, unfortunately I don’t have a limited key. But since you’ve got that nice golden badge, check your DMs — and let me know once you’ve done so I can delete it again. (nvm u have to dm me for that) 😂

shrewd lichen
#

I know you've probably heard a lot about codex-app-serv's poor performance and you're fed up with it... It's actually impossible to use Openclaw with Openai sensibly... There's probably some deeper problem with the logic - will there be a major refactor? Or should I file more issues/PRs?

worldly tusk
vocal night
shrewd lichen
shrewd lichen
final ocean
#

the ministry is at your service

willow narwhal
worldly tusk
midnight walrus
shrewd lichen
worldly tusk
#

are you sure it's openclaw, and not openai who's crashing?

vocal night
#

ive officially given up on 19

#

after 2 hours of debugging and doctor destroying config, rolled everything back

willow narwhal
shrewd lichen
vocal night
#

I was like hmm doctor might fix this, nope somehow ended up on anthropic and back to auth issues again after running it second time 😂 I haven't used anthropic in months

#

18 was smooth no issues, but today had issues updating, then telegram regressions (issues posted) tried fixing those and once I got those fixed, had to disable plugins and just like chasing one regression after another

midnight walrus
vocal night
#

going thru config and doctor added new fields and left empty etc

vocal night
willow narwhal
vocal night
#

why does it keep changing each version

#

there's always weird schema issues between them and doctor flips em around

willow narwhal
#

That should not change except something going on with doctor

#

Do you wanna DM me your config before running doctor so I can repro on my end?

final ocean
shrewd lichen
#

yes

#

#83200 #83476 too

#

I'm telling you, codex-app-server is unstable and crashes 🙁

final ocean
shrewd lichen
# final ocean what os and setup (git, npm/pnpm, docker?) 
Host: PenClaw
User: probo
OS: Kali GNU/Linux Rolling                 <----- MAN, DON'T EVEN ASK PLS, it's VM :D
Kernel: 6.19.14+kali-amd64
Arch: x86_64

=== HARDWARE ===
CPU: AMD Ryzen 5 3600 6-Core Processor
Cores: 4
RAM: 7.8Gi total, 5.7Gi available
Disk /: 51G used / 79G total (69%)

=== DEV TOOLS ===
git:       OK      /usr/bin/git | git version 2.53.0
node:      OK      /usr/bin/node | v22.22.0
npm:       OK      /usr/bin/npm | 9.2.0
pnpm:      OK      /home/probo/.local/bin/pnpm | 11.1.0
python3:   OK      /usr/bin/python3 | Python 3.13.12
pip3:      OK      /usr/bin/pip3 | pip 26.0.1 from /usr/lib/python3/dist-packages/pip (python 3.13)
make:      OK      /usr/bin/make | GNU Make 4.4.1
gcc:       OK      /usr/bin/gcc | gcc (Debian 15.2.0-17) 15.2.0
g++:       OK      /usr/bin/g++ | g++ (Debian 15.2.0-17) 15.2.0

=== DOCKER ===
docker: OK      /usr/bin/docker | Docker version 28.5.2+dfsg4, build 9cc6dea35e9a963f281434761c656fba4ac43aed
compose:        MISSING/ERROR
service:        active
access:         OK - current user can use Docker

=== PACKAGE MANAGERS ===
apt:       OK      /usr/bin/apt```
final ocean
shrewd lichen
#

nah, not main...

final ocean
#

haha. all good, some of your changes overlaps with some other changes but i will look into it

final ocean
shrewd lichen
final ocean
vocal night
#

got my agent responsive, doctor messed up auth and wiped auth profiles. had to relog in again

#

it had moved it over to codex app server auth?

#

anyways I logged in again and it's working now, but the auth situation has been dirty a few versions, I should have just relogged in at the start. never had a version wipe my authprofile json though and profiles in config

#

my debugging agent was like its' fine you're app server still has auth even if doctor wiped authprofile.json

vocal night
astral topaz
#

anyone want v2026.5.20-beta.1 automated testing results?

Actual OpenClaw Issues Still Standing

Large-memory fresh install: slow status and status-cli RSS over threshold.
Dirty plugin states: doctor-cli RSS around 1.17-1.23 GB across dirty plugin variants.
Release update recovery: rollback did not preserve the symlink dev plugin path, plus runtime-management RSS over threshold.
Agent/provider paths: several pre-provider dominance failures remain real product signals.
Gateway session send: gateway auth/token mismatch after agent turn.
TUI message turn: no provider request captured.
MCP tool call: RSS over threshold.
Browser automation: gateway closed abnormally during browser snapshot/start.
Media understanding timeout: no provider request captured.
Soak large-memory session: command p95 over threshold.

Note I did not ask for RCA yet, so take RSS "issues" less seriously

astral topaz
tulip gulch
smoky dock
astral topaz
odd anchorBOT
shy rain
#

quite a lot of people in chat use ollama cloud. I have used it myself in the past. It's a nice sub to have a wide range of models to choose from

potent path
#

Yep. Using it as primary for my claw setup. Ollama cloud subscription is cheap and GLM-5.1 is underrated for agentic AI. Kimi-k2.6 and Minimax-m2.7 also do good work, in the right circumstances.

willow narwhal
astral topaz
potent path
#

Haven't seen Barnacle Bot lately. Did it retire? I'm gonna miss these bot battles: 🤖 ⚔️

<@&1499243542898409495> clawsweeper Bot added the proof: sufficient label last week
@openclaw-barnacle openclaw-barnacle Bot removed the proof: sufficient label last week
<@&1499243542898409495> clawsweeper Bot added the proof: sufficient label last week
@openclaw-barnacle openclaw-barnacle Bot removed the proof: sufficient label last week

left mango
#

Hi maintainers, I created this PR to fix a user reported bug.
It is p1, platinum hermit and ready for maintainer look. Thanks.

fix(codex): beta blocker - keep context engine on canonical session key
https://github.com/openclaw/openclaw/pull/84954

worldly tusk
vocal night
mystic trench
#

For all of you having LCM issues, please tag me directly if/when you encounter more. Some updates for stalling issues went out in 5.19 and lossless 0.11.2 (yesterday) that should fix the latest batch of stalling issues.

tight wolf
#

Hi maintainers, I submitted this PR to address a reported Telegram/session context issue affecting gateway behavior
The fix has been validated with real interaction behavior and should be ready for maintainer review

fix(gateway): mirror source message sends into transcript
https://github.com/openclaw/openclaw/pull/84837

vocal night
#

@mystic trench the context one was a real headache + db cleanup

mystic trench
#

Was aware of the second one, reviewing it now. I’ll take a look at the first.

vocal night
left mango
final ocean
left mango
# final ocean checking

On a side note, how do you manage your email notifications from GitHub.

I've been getting so many lately! Many where I'm mentioned but clawsweeper on some PR and get subsequent emails.

final ocean
warped tide
#

here's a quick win
https://github.com/openclaw/openclaw/pull/84773

Keep internal webchat source-reply delivery payloads intact for the UI sink.
Return a compact message tool result for message_tool_only internal webchat sends so tool cards do not duplicate and truncate the full visible reply.
Cover the internal-source path with a regression assertion that the tool result omits the sent message body while preserving delivery metadata.

left mango
merry slate
#

Hi maintainers — three PRs open against main, all green on CI with proof: supplied + proof: sufficient. Would love a look when you have a moment:

  1. https://github.com/openclaw/openclaw/pull/84165 (rating: 🦐 gold shrimp)
    Most important. A race during an active run can hang/pause OpenClaw. Tiny fix (P1, size XS, 1 file, +23 LOC). Screen recording of the before/after is in the PR. (btw this one's already assigned to Peter, but I imagine he's swamped 🙏)

  2. https://github.com/openclaw/openclaw/pull/84932 (rating: 🐚 platinum hermit)
    Reasoning-capable vision models currently fail image describe with "Image model returned no text" because the default token budget is too small. Small alignment fix to match the other image path (P2, size S, +89/-2). Before/after CLI capture in the PR.

  3. https://github.com/openclaw/openclaw/pull/84247 (rating: 🐚 platinum hermit)
    Bigger diff but adds capability: the browser plugin describes its own screenshots via a configured vision model (like peekaboo / agent-browser), so text-only main models can reason about pages. Video walk-through included (P2, size XL).

Happy to split, rebase, or narrow scope on any of them. All three came up while I was building a browser-automation setup on OpenClaw. Thanks!

drifting nymph
karmic steepleBOT
drifting nymph
#

yes @fallen gate - it has been tested locally and we also applied it on our own openclaw and it works. everything has been mentioned in the issue page. and in the PR you can check it has the full details. hope it helps!

fallen gate
warped tide
#

lol

drifting nymph
#

ops thanks man. it didnt show in the conversation tab hence i missed it

left mango
drifting nymph
#

lol this happens. not ashamed of that, happy to contribute. its true we did test it locally.

drifting nymph
potent path
#

This is how we learn to do better. 😄

fallen gate
#

we just have ALOT of PRs and we just want to increase the quality of them before we take a look, hence the clawsweeper final boss there etc.

drifting nymph
#

Thank you @fallen gate - appreciated that. Will take note of this, thanks for pointing out again!
we will try to actively contribute to the openclaw 🙂

timber beacon
shadow garden
#

Reminder for those that have Proxmox at home (or elsewhere) -- Crabbox runs on it natively! Use it to get those PRs fully validated 🦞 (and if you're interested in Proxmox, it's free to self host!)

image.png
slow inlet
drifting nymph
#

this is kinda cool i have to admit 😄

ClawSweeper PR egg

✨ Hatched: 🥚 common Clockwork Merge Sprite

icy thorn
#

Hi maintainers (and thanks for all you do!). I wanted to check in on a PR for the new reply_payload_sending hook:

https://github.com/openclaw/openclaw/pull/82823 (rating: 🦐 gold shrimp | proof: 🦞 diamond lobster ✨)
This adds a Plugin SDK hook for mutating normalized outbound replies before delivery. The PR includes a screenshots of it working with Telegram inline buttons. It’s currently blocked on a P2 from ClawSweeper to "Regenerate the Plugin SDK API baseline," which I'm ready to do as soon as there's maintainer buy-in on the global hook composition order.

drifting nymph
# fallen gate we just have ALOT of PRs and we just want to increase the quality of them before...

looking at all the requests i understood where are you coming from. I was a maintainer for open source firmware community for years before moving to AI field a few months ago - I understand the workload of a maintainer and we are basically doing it for free and out of our own passion, still there are unrealistic expectations sometimes from others. also during my time I never seen 5000+ PRs or issues before..

tall prairie
#

I think we peaked around 9k open PRs. Those were the days.

naive horizon
#

how many issues at peak?

jolly wolf
tall prairie
icy thorn
left mango
#

I don't know what a proxmox is, so I don't know if I'm the intended audience

tall prairie
shadow garden
#

this way when you're done testing, the isolated VM/container is just disposed of

left mango
#

I often spend quite some time on clawsweeper review and re-review. So I'm trying to understand if I can somehow use crabbox to accelerate that process.

shadow garden
# left mango Ok. But as a contributor, can I use it somehow (after setting it up on my local ...

depends on how serious you are -- I am a bit on the crazier side so that comment was more for people crazy like me (as I know there are a few of them lurking in here lol) -- that said, if you do wanna dip a toe you can work with your Claw or Codex, etc to get it setup and use one of those providers

BUT the easiest way to try to minimize ClawSweeper giving you a stink eye is using codex on at least high thinking and ensuring you're doing autoreview or using Peters codex-review skill (just beware it's not light on usage)

left mango
shadow garden
left mango
#

I love crazy. I bet I'm one of the 3 people globally running openclaw on a 2 Gb raspberry pi 4 😂

warped tide
#

i'm one of the other 2

celest merlin
#

20b2 is out, gonna release latest soon

#

we back to daily releases

timber beacon
#

Nice!

left mango
celest merlin
left mango
odd anchorBOT
left mango
odd anchorBOT
drifting nymph
#

ok fine let me upload my profile pic. just dont hunt me down

shadow garden
odd anchorBOT
vocal night
#

I highly recommend turning off the friendly personality filter. Now the codex is more integrated and the soul file is better prioritized, the friendly personality is just noise.

It's written really poorly and it's written very corporate, HR-like. Don't be melodramatic, always be friendly. It really kind of takes away the soul. It's very openAI/chatgpt tone.

I ran a parity test, one to one, same session previously with one on, one off to see how the personality would respond. One sounds like chat GPT, one sounds like a person. Immediate difference.

{
agents: {
defaults: {
promptOverlays: {
gpt5: {
personality: "friendly" // friendly | on | off
}
}
}
}
}

merry slate
#

I didn't even know this setting existed😂

shadow garden
vocal night
#

Its actually good if you have no soul file or brand new user but anyone in here likely has a customized one.

lone iris
#

I still need to customize my soul file. It’s pretty practical for now.

jolly wolf
#

Here’s your reminder that if y’all are just talking about OpenClaw and not #clawtributors specific stuff like PR’s, use #general :)

vocal night
#

Thanks shadow, I would have posted in gen, but they need the prompt haha

#

Technically my post is about our goal of token efficiency 235 tokens every turn on the friendly filter 🫡

slender hill
#

A few PRs that are waiting on final review.

  • #84190 Log Discord component registry error details
    Adds structured error/cause metadata when Discord persistent component registry state falls back, without changing fallback behavior. https://github.com/openclaw/openclaw/pull/84190

  • #84189 Allow Discord component registry TTL override
    Adds bounded agentComponents.ttlMs config for Discord component callbacks, including schema/docs/tests and account override handling. https://github.com/openclaw/openclaw/pull/84189

  • #84187 Preserve reusable Discord presentation buttons
    Carries reusable: true through portable presentation buttons into Discord component registration so repeatable actions are not consumed once. https://github.com/openclaw/openclaw/pull/84187

  • #83774 test(release): guard Discord delivery origin packaging
    Adds release/package verification so stale compiled openclaw@2026.5.18+ artifacts missing Discord deliveryOrigin are rejected. https://github.com/openclaw/openclaw/pull/83774

shell dawn
#

is there a skill for generating songs that way? its remarkable

tall prairie
shell dawn
#

its the real AGI that peter built making openclaw and all it does is make bops

tall prairie
#

jk - the prompt is Molty make me a song

vocal night
shell dawn
vocal night
#

When the channel is empty after a release, it's because Peter and maintainers are all recording their EP, the whole thing is a cover for their Kband launch

shell dawn
vocal night
#

Greatest psyop of all time, k-slop daemon hunters

shell dawn
#

omg

vocal night
#

that's the album name

shell dawn
#

k slop daemon hunters, im dead

vocal night
#

honestly peter put em in a spotify album under that name it will be a hit

shell dawn
vocal night
#

gotta go to bed, finally got 19 to work for me with some patches, we'll see tomorrows release

halcyon wing
# tall prairie it's a secret

They still don’t seem to get it: it’s ClawFather himself, singing in person, and he’s telling everyone it’s Molty just to test their reaction. Has he been busy lately..? YES, with singing lessons, and this is the result 😄

celest merlin
grave arrow
#

Pete you still focus on stable releases or new features?

karmic steepleBOT
celest merlin
#

There's still a never-ending stream of issues tho

#

And I have some bigger refactors I wanna get ready and move the codebase towards as well

grave arrow
warped tide
#

i'd rather not concentrate on something that will be superseded in a few days

vocal night
vocal night
warped tide
vocal night
#

probably not same but same class otherwise clawsweeper would kill it, post it. theres so many bugs around message tool it needs its own sprint

vocal night
# warped tide I think i've got one in for this, too

not a fan of this one, the best decision here is simply to webchat render tool results as messages it's a UI fix and it solves half the bugs and bandaids everyone keeps doing, I just think only @stable pewter manages the UI and havent seen her around

#

would kill like 80% of the bugs/bandaids around it

celest merlin
#

we duplicate so much of pi already and esp with sqlite we'd fight it too hard

vocal night
#

internalizing or forking? Its a headache for me to consider 3 compaction systems and multiple runtimes when I code, I always forget one

celest merlin
#

internalize

#

and dedupe, bigger refactor

#

basically

Screenshot_2026-05-21_at_21.49.26.png
warped tide
vocal night
# celest merlin internalize

If peter will sign off on it I'll do a quick UI patch. TLDR send message is tool and in webchat creates a ton of issues + extra prompts to tell agents to ALSO send message. Its token wasteful, UI should just convert tool into message. Allows us to cut code actually vs add on.

celest merlin
#

Wait we did change the default for that back to last turn unless codex harness is used, since most models seem to be too dumb to use the mesage tool correctly.

slender hill
celest merlin
#

meanwhile

Screenshot_2026-05-21_at_21.53.03.png
vocal night
celest merlin
#

Wanted more codex. Figured I have to move it to cloud as my mac doesn't scale

celest merlin
vocal night
#

Peter most people are having issues not because of tool but their session refresh

celest merlin
#

oh because they don't do a /new after updating?

vocal night
#

I posted PR on this with a whole table on what gets refreshed mid session and what requires new threading and fingerprinting its a bit complex

#

correct

celest merlin
#

sigh

vocal night
#

a LOT of bugs come from no new

celest merlin
#

Yeah 100MB jsonl

vocal night
#

my system shit bricks today till I new'd on 19 and refreshed auth

celest merlin
#

pi was not designed for never ending conversations.

vocal night
#

I ranted half the day

shell dawn
#

oh is that why the jsonls are big? i had a slow first turn after a beta update yesterday and i looked into it and my claw was like '360mb jsonl for a first turn is not right'

celest merlin
#

Good thing tho once we interalize it's way easier to add reload

#

yeap.

shell dawn
#

how many tokens does that take lol

celest merlin
#

good thing is I have a big pipe to the source

tall prairie
slender hill
celest merlin
#

clanker said no when I fed it

#

to be fair I steered it a bit. not quite sure what happened in the .18 in the first place tho.

drifting nymph
# vocal night Peter most people are having issues not because of tool but their session refres...

what we experienced is a lot of issues shows up after update is there is no backward compatibility in configuration so we have to spent hours just to chase down which configs mess it up. a simple example is that gpt 5.3 codex was deprecated suddenly with new update (i think it was 5.12?) and it sent all our 20+ instances with those configs into boot loop:
2026-05-15T18:08:06.707+00:00 Gateway failed to start: Invalid config at /config/instance/config.json5. agents.defaults.model.fallbacks.1: Unknown model: openai-codex/gpt-5.3-codex. gpt-5.3-codex is no longer supported for ChatGPT/Codex OAuth accounts. Use openai/gpt-5.5 through the Codex runtime. agents.defaults.models.openai-codex/gpt-5.3-codex: Unknown model: openai-codex/gpt-5.3-codex. gpt-5.3-codex is no longer supported for ChatGPT/Codex OAuth accounts. Use openai/gpt-5.5 through the Codex runtime.

celest merlin
#

autoreview is killing me but I also love it

Screenshot_2026-05-21_at_22.03.44.png
timber beacon
#

Prepping a PR for iOS to align with the realtime-2 updates from other platforms

celest merlin
warped tide
#

I had similar issues. with the back-compat

timber beacon
#

Goal is to make voice amazing (and also move the iOS app forward selfishly)

warped tide
#

i was sure to new it, and it still threw all kinds of tantrums

#

re-authing didn't seem to help either

celest merlin
#

Getting all these states correct on updates is so hard.

tall prairie
timber beacon
astral topaz
#

Peter, any chance exec deny list is gonna get any traction? Asking for a friend ponder

drifting nymph
vocal night
# celest merlin Getting all these states correct on updates is so hard.

Peter. PR on the prompting side. We don't have to merge all of it, but Pash and I both agree that the fingerprint fixes need to go in place because of his changes when he took out my code on the prompts. Because he...moved back to developer instructions without the fingerprint fixes. Changes to soul file only happen and are recognized when there's a new session. Also, people who don't do a new session don't get the file. There's always little nuances like that when we move from every turn to trigger and hook-based..https://github.com/openclaw/openclaw/pull/83612
i.

#

My keyboard just died and I need to go get a replacement from the store in the morning. So I am typing with text to speech and trying to edit with the cut and paste.cat_sweat

drifting nymph
#

and one issue we just saw recently: we saw there are both anthropic and claude-cli in our /models and we have no idea where claude-cli came from.
when we did
cat config.json5 | grep "claude-cli"
its empty result - pretty sure it hidden somewhere else, but not sure where it hides. it seems like another source of truth we are not aware of

stable pewter
astral topaz
vocal night
# stable pewter I'm here everyday, how may I assist?

hey Val, TLDR. We wouldn't have to do a bunch of little weird fixes and things around it if we just have the UI intercept send tool messages and parse the JSON into a message.

It's also more token efficient because right now the codecs isn't really designed for web UI or any of these other tools. And so it's always easier to have it translated or picked up and deciphered than it is to have an agent write twice or have to prompt an agent to use a tool in a certain way for every single message platform.

drifting nymph
celest merlin
#

huh that worked

Screenshot_2026-05-21_at_22.37.44.png
warped tide
#

noice

celest merlin
#

sometimes I envision codex to be a rapper

Screenshot_2026-05-21_at_22.44.22.png
warped tide
#

Claw Thugz 'n Harmony

shell dawn
drowsy roost
warped tide
#

?

shadow garden
shadow garden
drowsy roost
# warped tide You're meaning multi-agent routing: https://docs.openclaw.ai/concepts/multi-agen...

Sort of. I already have that set up, but I’m very interested in two specific things:

  1. Making multi-agent rooms work so that agents respond intelligently rather than having to pick between mention-gating or not using multiple agents in a room at all. I did it in Nextcloud Talk and then in Matrix but I don’t know who else if anyone is working on it and the Matrix approach needs to be redone in light of many changes to how OpenClaw works since April.

  2. Voice features generally. I built VoiceClaw Realtime for it — and have been sharing on X for a while about how I set it up even before realtime-2, going back to march. I would like to actively participate and contribute. Doesn’t feel great to see it worked on and not even who is working on it or how to contact them.

#

What is frustrating is that I solved #1 TWICE — first in NextCloud Talk and then in Matrix, and it went nowhere and never became a feature. Now Telegram is promoting that they can do it. I did it in TWO channels months ago.

lone iris
warped tide
timber beacon
#

Hey @drowsy roost !

shadow garden
drowsy roost
celest merlin
celest merlin
tight wolf
clear violet
#

The DAVE stuff with discord was a pain when I added AIMeet to it the other day @weak walrus. Nice to see it direct in OC now.

tall prairie
#

(preamble is a built in thing, usually something that has to be prompted out of the model though).

shadow garden
worldly tusk
weak walrus
crystal relic
tulip gulch
timber beacon
#

rating: 🐚 platinum hermit feature: ✨ showcase
This and the eggs make me want to come up with more improvements/PRs

rocky pewter
potent path
#

https://github.com/openclaw/openclaw/pull/85121
Can anyone tell me what's happening here? CI check is failing for "Real Behavior Proof" but:

  • it has the "proof: sufficient" tag and
  • clawsweeper says, "Good normal PR with stronger-than-usual proof for a CLI routing change" and
  • "Real behavior proof: Sufficient (live_output): The PR body includes after-patch local CLI output and remote Crabbox validation showing the new help flag, invalid-key rejection, mismatch rejection, and focused routing tests on the pushed head."

Edit: Found it! The CI check has different criteria than clawsweeper. I added a screenshot and now it is happy again.

left mango
void bramble
#

Any Discord / OC folks in here? I've got this PR that gets the reasoning commentary back into the discord flow. It's really, really nice! Let me know if it works well and what you'd change.

https://github.com/openclaw/openclaw/pull/85164

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/83722
(Reply delivery / Tool ordering)
Fixes same-channel block reply ordering. Accepted block replies now wait for dispatcher delivery before the next tool starts, so the user-facing block message is not queued behind tool execution.

https://github.com/openclaw/openclaw/pull/83348
(Channels / Inbound debounce)
Fixes bare abort/control words during inbound debounce. Standalone messages like stop, abort, and wait can interrupt immediately instead of being delayed behind the debounce window.

https://github.com/openclaw/openclaw/pull/83262
(Commands / Models)
Fixes models set alias handling. If models list shows an alias like sonnet from the effective runtime config, models set now resolves it to the real provider/model id instead of saving a bogus alias.

https://github.com/openclaw/openclaw/pull/85110
(Codex / Child task state)
Fixes Codex child task completion state. A child task no longer looks successful when it only produced progress text, missed the final output, or failed to deliver the required completion back to the requester.~~.

potent path
#

Interesting .. Installing Openclaw 2026.5.20 downgraded my LCM from 0.11.2 to 0.10.0. Is that intended behavior? Doesn't seem like downgrading should happen by default, especially if I had 0.11.2 pinned. Right?

  • openclaw CLI: 2026.5.19 → 2026.5.20 (commit e510042)
  • @openclaw/codex: 2026.5.19 → 2026.5.20
  • @openclaw/discord: 2026.5.19 → 2026.5.20
    ⚠️ - @martian-engineering/lossless-claw: 0.11.2 → 0.10.0
  • The 5.20 openclaw update rewrote ~/.openclaw/npm/package.json to pin LCM at 0.10.0.
  • pnpm-lock still has the 0.11.2 entry cached, and 0.11.2's peer is openclaw: '>=2026.5.12' — compatible with 5.20.

Filed https://github.com/openclaw/openclaw/issues/85184 for reference.

left mango
timber beacon
#

Always good to provide more if it helps

#

Its also going to get auto merged so you are good

left mango
#

I created this PR based on a user reported bug.
size S, proof:sufficient, platinum hermit, ready for maintainer look.

I got my first glimmer PR egg. All of my previous ones were common.

fix: preserve Google Gemini 3 cron thinking
https://github.com/openclaw/openclaw/pull/85185

potent path
#

Please consider:

https://github.com/openclaw/openclaw/pull/85022
Fixes agent --local message-tool sends failing with “Channel is unavailable” even when the channel is configured. It
lets that caller bootstrap the outbound channel adapter when needed, while gateway-loaded channels stay on the normal
fast path.

https://github.com/openclaw/openclaw/pull/85007
Fixes one-shot local agent runs exiting before async agent_end hooks finish. Local CLI and no-channel Codex app-server
runs now wait for terminal hook work, while channel-backed runs remain non-blocking.

https://github.com/openclaw/openclaw/pull/84381
Fixes Gateway-backed agent runs surviving after the CLI receives SIGINT or SIGTERM. The CLI now asks the Gateway to
cancel the accepted run and preserves normal signal exit behavior after cleanup.

https://github.com/openclaw/openclaw/pull/83987
Fixes late Codex hook callbacks missing OpenClaw enforcement after the native relay unregisters too early. Successful
runs keep the relay alive for a bounded grace window, while failed or interrupted runs still clean up immediately.

https://github.com/openclaw/openclaw/pull/83743
Fixes exec and cron completion events getting routed to an old direct-DM queue when session.dmScope=main. Shared event
routing now sends those wakes back to the drained agent main session.

https://github.com/openclaw/openclaw/pull/83483
Improves hook visibility for outer code-mode exec. Hooks now see code-mode source under both code and command, plus
toolKind metadata so policies can distinguish code-mode cells from shell exec.

https://github.com/openclaw/openclaw/pull/83444
Fixes known code-mode worker failures being reported as vague internal errors. Timeout and runtime-unavailable cases
now return the documented structured error codes without changing successful execution behavior.

void bramble
fallen gate
left mango
void bramble
crystal relic
celest merlin
#

morning folks!

zinc dock
#

Morning sir 🫡

crystal relic
vocal night
pastel owl
vocal night
# pastel owl https://github.com/openclaw/openclaw/pull/85231 telegram message cache issue wa...

how does Telegram have like so many regressive bugs? Is it because we we add a new feature to it and then it just breaks the server? I feel like the whole thing needs a refactor if we're fixing this many bugs with Telegram every release. Even I've been fixing bugs with Telegram. I think I've merged like 2,000 lines of code in the last week. Peter, do we just need like a... you to run a slash goals on the telegram server? And capture everything in the blast radius?🤣

#

I'd say it's definitely our number one used tool. At least out here in Asia and all my customers in the US, probably 90% uses OpenClaw via Telegram.

crystal relic
worldly tusk
vocal night
crystal relic
celest merlin
#

@thick osprey generate a song

thick ospreyBOT
#

Generating now. I picked a short upbeat indie/electronic “clawtributors” anthem; the audio will post here when it’s ready.

vocal night
# celest merlin thank you! working on these

A new PR for review. Token efficiency and improved security on subagents. Configuration flag optional. it doesn't force anything on users. This is one area that nobody talks about, but this significant token savings.

Sub-agents receive the context they need from the spawning main. anyone that runs an orchestrator agent, this is massive token efficiency. this is especially true of Codex agents.

Codex subagents run best off of their own prompt and information they're given, everything else is just wasteful tokens spending. (https://github.com/openclaw/openclaw/pull/85283)

celest merlin
midnight walrus
vocal night
fossil ferry
#

Hi maintainers, I opened a PR to forward more sampling params through the OpenAI-compatible /v1/chat/completions path:

https://github.com/openclaw/openclaw/pull/84094

It threads client-supplied frequency_penalty, presence_penalty, and seed through gateway streamParams, runtime stream wrapper, then upstream OpenAI Chat Completions transport.

It also adds gateway validation for invalid penalty ranges and non-integer seed before provider dispatch.

This PR is intended to close https://github.com/openclaw/openclaw/issues/32496. CI is green and ClawSweeper marked proof sufficient / ready for maintainer look. Main question is whether we accept unconditional forwarding for these standard Chat Completions fields, or want provider-gated forwarding instead. Feel free to take a look when convenient.

midnight walrus
worldly tusk
edgy plaza
midnight walrus
vocal night
#

And for your birthday, what do you get? More PRs.cat_sweat

astral topaz
vocal night
astral topaz
vocal night
astral topaz
#

the Kova project is for automated local testing, it pulls a tag of the openclaw, start an instance, and run all sorts of scenario tests against it

#

this is different from integration testing or unit testing, basically you throw a bunch cases against a running instance and see what breaks

#

as for my PR, i added a bunch of breaking scenarios

#

and tests against tools

vocal night
#

I'm wondering how different that is versus my QA lab parity runs that we do run them the same way. If you're like super on the testing front, I would actually be messaging Vincent because he's been working on testing a lot. He's been breaking down over the last three or four versions some of my QA lab scenario testing as well. I'm just wondering if there's any overlap.\

astral topaz
#

the reason i am using Kova is because it is the only full featured testing tool from openclaw, I'm sure the maintainers have their personal tools, but Kova is the only one shared

vocal night
astral topaz
#

the other two repos multipass and plugin-inspector are feature specific

vocal night
#

That's why eva in the changelog like 20 times every release last few releases.

astral topaz
#

i would love to have vincent's testing suite, if he decides to share it

vocal night
# astral topaz i would love to have vincent's testing suite, if he decides to share it

My QA harnesses are built into the code. Most of them are merged in. If you look in my GitHub and find like merged QA, you can just tag those harnesses that are merged in. Several of them, like they have like automated modes, but they have manual mode. You can just tell your agent to run. There's not like skills for anything. You can just, you'll have to just like tag the harness and say, okay, I want to run this. And you can run 100 line turns. You can run token an efficiency test. There's a bunch of stuff.

astral topaz
#

gone through it... seems need isolation... it's cross channeling into reading configs from running OC instance instead of using test specific config locations

celest merlin
vocal night
#

Hell, I don't even like ToolMD. It's deprecated prompt wise i.E. nothing mentions it anywhere. It's only if the user decides to use it and the user custom tells clanker to use it

Myusers and people I've talked to when they check it it's actually old deprecated stuff from Claude Anthropic days.

. Most people, in my experience nowadays, use an AgentMD w links to runbooks for workflows

#

FYI, I am not having a stroke. My keyboard died and I'm in the middle of the mountains this weekend. . So I'm using text-to-speech plus the accessibility. keyboard.😂

worldly tusk
#

It might be beneficial to even have an override for subagents: should there be subagents.md present, they get that file instead of agents.md. For example, my agents.md if full of memory management stuff that is useless for subagents

vocal night
# celest merlin I don't really wanna introduce yet another config option - I do agree that subag...

Even then though, people give their agent file. So the thing is, the security issue is the agentmd file, Peter. So the problem is that if your agentmd has certain accesses or links, it says, "Oh, here's these processes and runbooks."

Subagents, if derailed or prompt injected, could give out that information. That's why I really believe there should be a subagentmd instead as an option regardless because it's one of our largest security flaws in the system is a main agent that might give a subagent task,

its actual prompt and its accesses to its workflows and how systems work, etc., are all shared with subagents that don't need that context.

vocal night
# worldly tusk It might be beneficial to even have an override for subagents: should there be s...

Correct. That's why it's also a waste of every turn for them to get that file. It's massively wasteful. Everyone talks about, "Oh, token efficiency for OpenClaw." I've ran the test. As a percentage, over 50% of our inefficient token usage. Is the damn. sub-agents getting tons of context running every turn

When you're telling them to go do a basic task, they have no need for any of that context. They literally just gets ignored and wasteful.

#

I've used what I call a context swapper that forces the system. It's a custom patch plugin that swaps sub-agent MD for agent MD and I've used that for months now. And in my token efficiency, I've done one-to-one tests.

It's crazy difference. It's over 50% savings in tokens. And that sub-agents never ever really need the context because the main agent gives them the context no matter what. It's just what it does.

I don't like the idea of another file to manage, but that's why I purposefully think it should be just the Codex Agent MD versus the User Agent MD.

The main Agent MD is very different, and what users actually want, UX-wise, is the main agent to have that Agent MD. When they want sub-agents, they want the Codex Agent MD, really, because the Codex Agent MD is better.

#

Again, forgive my six-year-old typing skills right now.

#

Holy shit, they just launched Goal Mode in Codex Desktop. Has anybody even saw that yet?

astral topaz
#

i tried, it failed to set goals

#

still brittle?

celest merlin
#

GOAT MODE

astral topaz
#

/goal watch reviewer comment on PR and fix issues until Overall rating is 🐚 platinum hermit

low lark
odd anchorBOT
vocal night
#

Peter semi-fixed this, maybe on accident. The Claw Sweeper basically prompts Codex to keep reviewing and try to get the high scores.

vocal night
astral topaz
#

lol forever loops, now with AI

astral topaz
#

actually... now that i think about it, you don't need to set a reminder, just watch email for when someone comments on your PR, it's a poor man's webhook for when the agent should take a peek at review message

#

although, it is kinda spam punishment if you subscribed to notifications for openclaw lol RIP mailbox

final ocean
vocal night
#

Best of both worlds. Personality without any bugs 😮‍💨

astral topaz
#

that is my architecture as well

willow narwhal
#

Kindly send critical PRs my way.

vocal night
astral topaz
willow narwhal
willow narwhal
astral topaz
willow narwhal
astral topaz
#

i used a real local ollama provider

#

make it look more realish

prime vault
# willow narwhal Kindly send critical PRs my way.

https://github.com/openclaw/openclaw/pull/85205 not mine but most my crons broke after update this morning and I went to write a PR and this was already done earlier today.

openclaw gateway run (default config, openai/gpt-5.5) currently fails every embedded turn on a fresh session: the dashboard webchat call lands as [ws] ⇄ res ✗ agent errorCode=UNAVAILABLE errorMessage=EmbeddedAttemptSessionTakeoverError, and openclaw agent propagates GatewayClientRequestError: EmbeddedAttemptSessionTakeoverError. The model call itself succeeds the full assistant message is persisted to the session JSONL but the runner aborts on the fence check that fires after Pi's own write. The user-visible effect is "every chat fails" on a clean install.

willow narwhal
potent path
left mango
astral topaz
#

lol the XXXL PR has grown to 20k lines yolo

left mango
astral topaz
warped tide
#

I want a kaiju...

astral topaz
#

ask Peter to update the clawsweeper PR game to include those

vocal night
#

All I get are weird looking humans. Like these like funky looking weird Japanese anime looking people head things.

warped tide
#

i'v only gotten common so far

#

I'd love to see what the others look like

timber beacon
pastel owl
vocal night
potent path
#

Or just write it up as an issue and someone else will grab it and do the PR

#

I would, but I don't use telegram and so replicating and providing proof would be a pita.

pastel owl
#

Basically in one instance last night the telegram topic was making every tool progress message as final, so i had like 150 messages from toolprogress. But on the next turn it started working fine, without any restart etc.

#

So i have let it slide

#

Using current main

vocal night
#

This is a really easy debug, maybe 200 lines of code. You just tell your agent, "This is what happened. It happened with the Telegram session." Tell it if you're using LCM, that that might be a broken issue. If you're using LCM, there's been some weird cross-session key stuff going on. If not, just say, "This is what happened. This is the time." Give it text. Give it one of the messages so we can search and find it with the time, and tell it when it started happening. The range, like at the time it happened at this time, at this time, the agent will fix it really quick.

potent path
pastel owl
#

I currently have three open PRs
Will take it up after these are merged

vocal night
pastel owl
vocal night
#

Nothing like PR blackmail. If you don't merge my PRs, I won't post the bug.😂

odd anchorBOT
astral topaz
#

getting back to 🐚 platinum hermit after conflict resolved merge is hard mode

runic totem
shadow garden
#

not sure what happened in main but there's a nasty delay in responses in all mediums I have tested. went from nearly immediate typing indicators to 10+ second latency

codex is having difficulty targeting the root, I'm starting to dive in myself with it

#

if it's OC core related it happened in the last 20 hours or so

crystal relic
pastel owl
#

Is anyone else also feeling the latest main has made telegram very slow. After sending a message, its taking 1-2min for it to react

#

My claw is saying something related to provider auth pre-warm exhausting the event loop

cursive palm
#

It seems there are some.

tall prairie
#

I’m out of pocket today but post here

#

Appreciate you all 🫶

#

Also we want to know the root cause please. Not to blame but so we can improve our systems to prevent reoccurrences 👊

pastel owl
rigid whale
celest merlin
celest merlin
#

/verbose doesn't do the trick? that's again a lot of added complexity

#

For once I'd love a PR that removes stuff.

clever narwhal
celest merlin
#

but that's what /verbose is for

#

so you see all tool calls

#

all these extra features are unmanagable, and then we are the bad guys when stuff breaks with every release.

vocal night
vocal night
clever narwhal
#

Yeah I understand. I wonder why tool calls persisting in a telegram group hasn't been the case for me. I was trying to make it happen.

vocal night
# celest merlin How often did this happen to you? I dislike adding yet another random file based...

Yeah, Telegram has way too many config, every time someone fixes one of them, it breaks the other features. half our Telegram problems this week come from adding in one feature where obvious was like yeah this is gonna break all a telegram and the dude was like nah it's not and then yeah it literally broke like all of our other features and we've been having regressions ever since but it looks bettercat_sweat cat_sweat

clever narwhal
#

I'm looking closer if verbose simply isnt applying to these telegram groups by default, for whatever reason.

celest merlin
#

very possible that /verbose isn't working correct. I just think most users don't want to see all tool calls in telegram mode.

#

sorry for shooting you down on this. Let

#

Let's fix existing concepts before adding new ones.

timber beacon
#

I love not having tool calls, just not seeing a drop/fail sucks sometimes. But the noise is more annoying

#

I'd assume most users want less noise and more function/success on their chats. If the work is hidden and it just works most non power users will be happy

clever narwhal
clever narwhal
timber beacon
#

Yea, more robust log system could cover that.

#

Not sure you get that outside of webui

clever narwhal
#

I personally like verbose, which is why I'm into getting it to work properly

worldly tusk
astral topaz
jolly wolf
clever narwhal
# jolly wolf Yes fix

Great -- think I got the fix at the correct place, not adding anything. Will put up a PR soon

rocky pewter
astral topaz
#

it did give me showcase tag, so that was like epic drop after battle

rocky pewter
vocal night
left mango
serene laurel
raven kernel
#

Please join us in the Stage for this week's WEEKLY CLAW!

zenith prawn
#

Hi friends, need maintainer review, someone available?
Worked whole day on new PR with feature no competitors have (yet): Fast Talks - Cheap Execution
https://github.com/openclaw/openclaw/pull/85104
No more switch /fast on&off. Set to auto, and it will auto-switch 'off' after 60s of tool calls (configurable), and auto-switch back 'on' after done - for fast response to the user.
Spent a day making sure it works and no slop. Its continuation of idea #1504047490956267570 message

smoky dock
clever narwhal
jolly wolf
clever narwhal
timber beacon
#

Doing some work on UI/UX for iOS app

#

Trying to get it more aligned with standard and not have voice overlap with main screen weirdly haha

warped tide
#

Security hardening PR
https://github.com/openclaw/openclaw/pull/85512

What it does:
  • Hardens source=url package downloads in scripts/resolve-openclaw-package-candidate.mjs
  • Enforces HTTPS-only, no URL credentials, default HTTPS port only
  • Blocks localhost/private/link-local/metadata/special-use targets, including redirects
  • Adds bounded download size, timeout, manual redirect handling, and atomic temp-file cleanup
  • Adds tests for unsafe host rejection, private DNS rejection, redirect-to-metadata rejection, oversize cleanup, and successful atomic write
timber beacon
#

Dang can’t share images here haha would love to show what I got so far

rotund tendon
#

Hopefully can take a look at this: A newer same-target message with no visible reply could wrongly cancel an older in-progress final reply. So if you sent a pic and then wrote something about right after it could cancel.

https://github.com/openclaw/openclaw/pull/85517

rigid whale
warped tide
astral topaz
shell dawn
wary garnet
warped tide
# astral topaz does that mean trusted dns all the way up?

More like validated and pinned resolution at every hop. DNS is still treated as untrusted. For each URL and redirect, we resolve the host, reject unsafe IP ranges unless the named trusted-source policy explicitly allows them, then pin the actual connection to the validated address set. That prevents DNS rebinding and keeps enterprise/private mirrors scoped to repo-controlled policy instead of arbitrary workflow input.

astral topaz
#

oh i thought you are going to do some fancy dnssec validation all the way from root to zone, get the ip, then ssl validate it

warped tide
astral topaz
#

sure, i just assumed you would filter any resolution from rcf1918 space

warped tide
astral topaz
#

ah

cursive palm
warped tide
# cursive palm Can you please resolve the SSRF blocking issue of Fake-IP in Clash Verge? I've d...

That sounds like a Clash Verge / Fake-IP compatibility issue, not something I would solve by weakening SSRF checks.
Fake-IP works by returning synthetic addresses instead of the real destination IP. If that synthetic address lands in a reserved or special-use range, a crawler or plugin with SSRF protection is expected to block it. Firecrawl blocking that segment is probably intentional, not a bug in the SSRF logic.
The safer fix is to change the proxy/DNS behavior for that workflow, such as using a different DNS mode, adding the affected domains to the Fake-IP filter, or routing the crawler so it sees the real resolved public address instead of the synthetic Fake-IP address.
I would not recommend globally allowing that range inside SSRF protection. At most, it should be a local trusted-network override with explicit scope, not something enabled for arbitrary user-supplied URLs.

left mango
lusty vigil
warped tide
left mango
fallen gate
willow narwhal
odd anchorBOT
lone iris
#

👀

void bramble
winged kraken
#

Are OpenClaw plugin promotional posts allowed here ?

fallen gate
shell dawn
pastel owl
#

My claw is extremely slow to respond since yesterday's main install

Is this regression being addressed ? I also checked openclaw-rtt repo that's also showing 2x-3x rtt times for telegram

void bramble
shy rain
void bramble
pastel owl
shy rain
#

hmm 5.22? I haven't updated from 5.20 yet. Working on some memory stuff...

pastel owl
#

Codex is not able to find the regression
May be i have not prompted enough

#

Yes 5.22

#

Whatever has happened is in the last 36 hours of commits

vocal night