clawtributors

[shell dawn]

i guess lets go back to the thread before we spam up the channel too much more lol

[fringe rock]

We’re running so much better on pi harness now

[viral musk]

This is in reference to tool timeout issues with the codex harness?

[fringe rock]

No it just doesn’t work at all lol

You have to constantly keep talking to it

Doesn’t update you at all or continue conversation

[shell dawn]

But the openai provider's getModelForRequest() strips codex metadata — it only preserves it for the openai-codex provider. So Forge sends a request to the openai provider with codex intent, the provider throws away the codex metadata, and OpenAI gets a plain non-Codex request. OpenAI requires an API key for non-Codex requests. Forge only has OAuth (Codex) credentials. 401.
The Fix:
The openai provider needs to check if agentRuntime.id === "codex" on the incoming request and preserve that metadata — same logic the openai-codex provider already has. This is a one-line code change in the gateway source.

[viral musk]

Yeah, it's a tool timeout issue, I think.

Something on the codex side.

[fringe rock]

It’s just not ready yet at all

[jolly wolf]

We get it, you don’t have to keep saying it. This channel is for fixing the issues, not just saying there are issues over and over

[viral musk]

Definitely not piling onto the "there are problems" side. What I'm seeing: When Codex internally aborts a task via op.dispatch.shutdown, it doesn't emit any terminal notification to the bridge, so the bridge sits waiting for a signal that never shows up.

[prime vault]

https://github.com/openclaw/openclaw/pull/82659 - Codex: downgrade app-server approvalPolicy when user Codex hooks are disabled

• Hit Codex bash declines after #82496: stdio app-server with non-never approval policy + bridged codex_home features.hooks=false leaves exec stuck in pending-approval, every call returns declined. PR #82659 reads features.hooks from the same bridged codex_home Codex actually loads, downgrades to never when hooks are off, and gates that on allowed_approval_policies so managed installs aren't bypassed. 813/813 codex tests pass, 90/90 CI checks green, ClawSweeper round 3 cleared security and signed off the code.

[celest merlin]

just here to drop molty songs

[celest merlin]

checking!

[prime vault]

it was on the slightly bigger side if yall want to go a different way, just swapped to codex harness day or two ago to help test and the update this morning had my dude forgetting to come back to me all day in discord

[shy rain]

https://github.com/openclaw/openclaw/pull/82571

Fix missing codex plugin which breaks everyone's update if using Openai models.

Updated to respect plugins.deny and plugins.enabled:false if present.

[potent path]

Anyone else having this sort of problem?

› 8 failing CI tests all upstream of our PR??

`• Yes. More precisely: all 8 failing CI jobs are base-side/upstream, not introduced by this PR.'

Anyone else having this sort of problem?

[ripe dagger]

https://github.com/openclaw/openclaw/pull/82800

minor bug; add verboseness to api key errors

[celest merlin]

taking

[celest merlin]

did you try this? last time i did this we can into all kind of launchd timeout dramas

[celest merlin]

I think there's a better way, trying sth with forcing hooks.

[shy rain]

worksonmymachine.png. But seriously we added a timeout to avoid the plist loop hell.

• Fast network: install completes in ~10s, gateway up, codex works, silent.
• Slow network: install takes up to 60s, gateway delayed but codex works.
• Timeout/failure: gateway comes up immediately after giving up, codex unregistered, user sees the manual fix one-liner in the logs. No restart loop.
• Policy block (deny/disabled): same as timeout — instant passthrough, warning with manual fix, gateway up normally.

This should catch the vast majority of users instead of leaving them with a cryptic error message about unregistered plugin, and doesn't cause a restart loop if they have sketchy connectivity.

[olive cliff]

hey there! @untold pollen, a telegram + codex one

https://github.com/openclaw/openclaw/pull/82790

[potent path]

https://github.com/openclaw/openclaw/pull/82559
This PR closes a Codex native-hook enforcement gap: /btw side-question child threads and app-server command approvals were not consistently routed through the active PreToolUse relay. That matters because native hook policy is the safety boundary for Codex command execution; this makes enforcement consistent and fail-closed across the main and side-question paths.

Ok, this one got pretty chonky, but ~1100 lines are regression tests, vs ~470 actual code changes. Hear me out. 😆 Full explanation here: https://github.com/openclaw/openclaw/pull/82559#issuecomment-4468797442

[fast pagoda]

https://tenor.com/view/disney-castle-win-gif-24507657

[ripe dagger]

thank you 🙏

[winter field]

PR for a discord bug where final answers would send stale stream content instead of the actual answer https://github.com/openclaw/openclaw/pull/82645

[left mango]

Hi, I created this PR to fix the image describe URL issue highlighted by someone else.

• fix(cli): keep image describe URLs remote
  https://github.com/openclaw/openclaw/pull/82854

[shell dawn]

https://github.com/openclaw/openclaw/pull/82864
PR created related to issues with the change to an openai provider as part of the codex runtime implementation. this issue improves resolution for codex runtime to have openai provider paths resolve to the openai-codex auth/provider path so openclaw does not accidentally start to resolve to a possibly nonexistent openai API provider path

the real proof is log sections from my use across two different servers, the pattern is that the resolution will work correctly and you can have a long run of codex runtime turns without any issue, with everything appearing to work correctly, and then the agent will just cease to respond until eventually it fails over and is nudged back to running on codex/openai properly, after which it will run successfully

[hollow frigate]

Hmmm, this is something that seems to resurface once in a while in some way.

[shell dawn]

https://discord.com/channels/1456350064065904867/1505322984284356692
instructions on how to check your logs in there

[merry slate]

hi peter, looks like the Control UI protocol version needs a small bump
too. Opened #82896, but feel free to just patch it yourself if that's easier.

https://github.com/openclaw/openclaw/pull/82896

[devout galleon]

yeah this also affect the current tagged 5.16-beta.4 container image (I know it's not actually officially released but i saw it tagged and built and couldn't resist)

[merry slate]

my PR got auto-closed as duplicate of #82883 / #82882, so looks like someone's already on it

[devout galleon]

yeah i saw after i commented. and i also just got a github notification that beta 4 pre-release is done. unfortunately not usable in containers. back to beta.3 until 5 drops

might be worth noting along with the release announcement that container users should hold until next beta, git users can make an adjustment

[potent path]

This is interesting. Is dropping the 'list' parameter from 'openclaw sessions list' by design? I don't have a problem with that, but seems like we should at least call it out in the patch notes or otherwise notify users that it has changed.

breaking CLI change: beta.4 removed sessions list subcommand. Now use 'openclaw sessions' (bare) with same flags

[vocal night]

Always a potential regression, "research as potential regression and look into potential offshoot effects from tracing field, file as issue if so and pros/cons of paramter loss and subcommand.

50% chance its a regression, 50% chance its a feature, flip a quarter lol

[potent path]

Yeah, I guess I'll file it as an issue just in case it's not an intended change and more research into the cause and blast radius is needed.

Looks like it's been this way for a few days. I just didn't notice it had changed until today. https://github.com/openclaw/openclaw/issues/81139

[celest merlin]

release is not far, sonday!

[celest merlin]

noo that was an oopsie, not again. I reverted that, not worth a v5 (and added guardtails)

[shell dawn]

thx for merge peter, hopefully this smooths out a few more bumps

[celest merlin]

pheew got a lot fixed today. and hundreds of review rounds.

A few of the PRs are still in the pipeline

Folks thanks so much for helping and testing! I appreciate it lots!

[vocal night]

good catch P just fixed it. also quick shoutout to anyone debugging/building and using multiple codex accounts. @olive cliff made this little tool it's great https://github.com/vyctorbrzezowski/codex-switchboard

I like codex bar but if we can get in next release peter the ability to store and swap tokens it'd be awesome Peter.

One click store and one click swap switch has saved me a lot of time. Just need an auto swapper vyctor- if you swap fast enough the threads actually can continue 😂

[celest merlin]

preeety sure that is aleady in?

[fast pagoda]

https://tenor.com/view/groot-happy-dance-oh-yeah-marvel-marvel-comics-gif-17641744

[plucky dew]

Soonday

[potent path]

https://github.com/openclaw/openclaw/pull/82861
This PR prevents 'openclaw infer model run --gateway' probes from attaching to the default 'main' agent session by giving each probe its own explicit model-run-<uuid> session. That matters because model probes are supposed to be raw one-shot health/capability checks; inheriting a polluted or oversized chat session context can make a healthy provider look broken. (as I recently discovered first-hand!) 🫠

[plucky dew]

vibing 🙂

yes pls haha

[crystal relic]

https://github.com/openclaw/openclaw/pull/82991 ready for review

Simplify the parsing steps for repeated authentication key resolution during gateway startup to improve boot speed

[vocal night]

does it work for desktop? codex bar tries to switch live path while running, but the app needs to restart to actually swap in desktop version

[fast pagoda]

https://tenor.com/view/dream-theater-john-petrucci-john-myung-gif-22682084

[languid folio]

landed

[crystal relic]

https://github.com/openclaw/openclaw/pull/82832 ready for review

make ACPX startup probing into the runtime service instead of splitting it between the outer wrapper and inner service, which preserves the existing Gateway readiness contract while reducing drift risk.

[fast pagoda]

Review ping 👋 https://github.com/openclaw/openclaw/pull/80358
Fixes native Gemini tool-call replay by preserving same-route thoughtSignatures, rejecting foreign or invalid signatures, and keeping the Gemini 3 fallback so replayed tool results are accepted by the real Gemini API.

[celest merlin]

omg 1k lines?

[fast pagoda]

Mostly the test...

transport-stream.ts only 161 lines

[vocal night]

5.5 loves its test, ask the agent to see if it can consolidate test/audit the test and see what all is actually necesary (and if the test actually work*)

"Using TDD analayze the test versus the working code, do these test accomplish the goal, and can we streamline the test for this/consolidate test in any way"

[fast pagoda]

I'm wipping the clanker right now

[vocal night]

4:1 test to code seems a bit much

I'm surprised Clawsweeper doesn't as part of it's analysis score if possible to consolidate a fix/test down and cut LOC

I'm not a TDD fan as a human but I've found using it in my prompts for PR architecture > implementation, my agents do a much better job of building the test first and then the code after especially in a codebase that's already established like OC

[vocal night]

landed 🛬

[fast pagoda]

Saw it's landed when I was educating hard on my clanker to learn as well as myself, and a stupid not allow maintainer push setting lul

it's raining outside, my clanker said he felt emotional so wrote the test long

[vocal night]

honestly I think it's a bug in GH or codexs new update on GH, I've had several that I had enabled and it just wouldn't allow peter to merge

[vocal night]

did you see article on 171 hidden vector emotions they found in opus haha

[round silo]

I’d like to add support for GigaChat(popular llm provider in cis) as a new provider, but I’m not sure what the preferred contribution path is.

Should I open a PR directly, implement it as a plugin, or is there another recommended way to add a new provider?

[vague swallow]

I have a PR to better handle TLS certificate rotation in the android app, which currently just rejects new certificates after rotation, and you have to wipe the app and re-onboard to accept the new certificate

This PR adds a prompt to accept the new certificate
https://github.com/openclaw/openclaw/pull/83077

[untold pollen]

Android App

[bleak fable]

hey folks. I was noticing a huge lag in getting a message reply after an openclaw restart.. the problem seems to be that on cold start OC is still trying to resolve / hydrate model metadata from Codex... so even though the model was effectively fixed for our setup, the startup path was still doing extra work around model resolution/catalog behavior, and that was where the ugly cold-start delay was coming from.
I fixed it locally by making the model definition explicit in the config, so it stopped needing to figure that stuff out at runtime.... After that, the pre-model startup stall basically disappeared.... The first cold turn now reaches the actual model quickly, and most of the remaining latency is just model thinking time.
This brought down the latency in that particular stage from 25+ seconds down to 9 seconds.

So ideally when we set up codex harness, or migrate, it should fix these definitions in the config.. solves the latency...

Is anyone working on this PR? or should I take this up?

[vague swallow]

I hit a little snafu after pulling latest main and rebuilding, existing codex auth from auth-profiles.json is no longer recognized

⚠️ Agent failed before reply: No API key found for provider "openai-codex"

openclaw configure now asks to migrate codex, I'm not sure if that's related?

◇ OpenAI device code complete
│
◇ Installed Codex plugin
│
◆ Migrate Codex at /Users/echo/.codex into this agent now?
│ ○ Yes / ● No
└

looks like the format of openai-codex in auth-profiles.json changed significantly with no automatic migration

[vocal night]

it's changed a few times in last 2 weeks I've had to re-login a few times

[pastel owl]

something broke
Not able to see any replies on telegram

Using latest main

Discord is ok

Why is the test and verification mechanism so broken

Discord toolprogress message got stuck aftet 3rd line

[shadow garden]

Another Codex context loss bug dug up in main; small targeted fix here:
https://github.com/openclaw/openclaw/pull/83127

[pastel owl]

Re pulled from main and built
Now telegram is working again

[vocal night]

building from main is always dangerous when it moves this fast, I recommend testing official beta patches

[jolly wolf]

Coward

[tulip gulch]

not a critical bug, but seeing OS X specific messaging on a VPS makes OC feel less "prod ready". Fixing for openclaw update https://github.com/openclaw/openclaw/pull/83096

[shadow garden]

where's the fun in that!?

but in all seriousness, correct, it is not for the faint of heart to build from main -- if you have backups and you're into it, go for it! but dont exect things to go well

[vocal night]

pretty sure even v and p have said don't test from main it's constantly moving and you may pull down from middle of a multi tier merge that breaks your claw

[shadow garden]

yup; it is not for your day to day users

[vague swallow]

Walkie-talkie style, real-time voice mode for android:
https://github.com/openclaw/openclaw/pull/83130

Previously only a legacy turn based talk mode was implemented, which never worked for me personally.

After this commit, real-time voice works on Android, e.g. using gpt-realtime-2. Tested on a real device.

cc @untold pollen

[shadow garden]

ooooo, I might have to patch this in and try this out

[vague swallow]

do it

[vocal night]

is this same as that one dudes been posting on x for iphone?

I never could get real voice to work, so just talk via telegram, thought about building one like Omi since it's OSS isnt it?

[winter field]

https://github.com/openclaw/openclaw/pull/83139 test-only follow-up to #82850, adding a regression case based on the one-streamed-text commentary+final scenario from closed #82645.

[shadow garden]

Qwen? yeah -- I think their more recent one might still be closed still though unfortunately

[vague swallow]

I haven't seen that one, but this is just connecting to the realtime voice features already in OpenClaw (e.g. used by voice-call plugin) so actually this change isn't all that fancy

[untold pollen]

Hmm this removes the old mode, I'm making this a config instead. If the user has realtime setup it'll automatically use that

[vague swallow]

my apologies

[untold pollen]

All good! Great PR, need a bit to test it properly though

[winter field]

could a maintainer please apply proof: override if this coverage-only follow up is worth merging? :)

[timber saddle]

Hey I updated to version 5.12 and now i am getting this: [assistant turn failed before producing content] with all my agents is there fix?

[rocky pewter]

Opened a small beta.4 Doctor bugfix.
PR for #83098:
https://github.com/openclaw/openclaw/pull/83154
It lets doctor --fix clean up old CLI runtime pins, even when the provider itself still stays the same.

PR for #83233:
https://github.com/openclaw/openclaw/pull/83262
It fixes models set sonnet so it uses the same alias users see in models list, instead of writing openai/sonnet.

[ripe dagger]

Humbly requesting the contributors role: https://github.com/openclaw/openclaw/pull/82800

[jolly wolf]

Use /claim

[ripe dagger]

done

[ripe dagger]

thanks

[dapper jungle]

Humbly requesting the contributor role

My merged PR:
https://github.com/openclaw/openclaw/pull/67848

It fixes misleading provider identity in the OpenAI auth flow by making API key vs Codex OAuth explicit, and adds regression coverage for it

[slender hill]

4 PRs ready for review
2 PRs (83170 and 83189) are both to address bugs with subagent binding on Discord. Closes issues 83170 and 83172. Annoying little bugs they are.

2 PRs (47677 and 83164) are to add new functionality for the clanker to pick up reactions and treat them as a first class wake event. One PR is for Discord and one is for Telegram. Proof was not able to be provided because there was not a good way to do that without exposing secrets. Closes issues 59404, 64752, 83169 and a host of downstream fork issues.

[ripe dagger]

you can use the /claim command

i just learned that :)

[dapper jungle]

Cool. Thank`s

[tulip gulch]

clawsweeper having permissions issues? https://github.com/openclaw/openclaw/pull/83096#issuecomment-4471461435

[jolly wolf]

No, it’s intentional. Sometimes people don’t enable the setting to let maintainers push to their fork, so we re-create the PR on our end instead and give committer credit

[ripe dagger]

~~Ready for review: ~~https://github.com/openclaw/openclaw/pull/83197
feishu subagent thread binding now returns deliveryOrigin from its subagent_spawning hook like we did for discord in pr #83172. without it the core spawn flow cant route the first child run to the right feishu dm or topic. the message falls to a fallback instead. this fixes the delivery origin gap that was still hitting feishu lark.

[tulip gulch]

had maintainers modify my prs all the time, this seems clawsweeper specific

[neon swift]

are we supposed to ask for review here? I have an open PR that's ready, but I didn't want to add unnecessary noise

[bleak fable]

Anyone? I dont wanna dupe your work in case you have already pushed something.

[ripe dagger]

i think you could open an issue on github and you would get better response there.

[neon swift]

Hoping for a review whenever someone has the chance. Pretty straightforward PR for fixing the token usage tooltip in the Control UI. It currently gets cut off on very tall bars/high usage, and this PR fixes that so the tooltip remains visible. All checks passed, no conflicts, screenshots/evidence in PR as well. https://github.com/openclaw/openclaw/pull/82846

[fallen gate]

yes post here using that template above, PRs here are prioritized over other github noise

[shadow garden]

@untold pollen looks like stop is queueing again in main on telegram 😞 codex-review on its third round reviewing the fix currently (and will mark the PR open when verified)

[crystal relic]

https://github.com/openclaw/openclaw/pull/83019 ready for review / XL (+1843/-345)

Add observability of the secret-backend to facilitate easier diagnosis of gateway startup issues in the future.

[modern sequoia]

有人懂中文而且知道如何在Lark 飞书 international 版本上布局龙虾吗?

[fallen gate]

@tight hornet can you translate that for me

[tight hornet]

Roughly: “Does anyone understand Chinese and know how to set up/deploy the lobster on Lark / Feishu international version?”

“布局龙虾” is a little ambiguous, but in this context I’d read it as “set up OpenClaw/Krill/lobster” on Lark/Feishu International.

[modern sequoia]

部署龙虾？

[fallen gate]

@tight hornet can you point this gentlemen to the international channels?

[tight hornet]

中文支持可以去 #1466722439789674741，偏技术/中国社区讨论也可以去 #tech-中国。

For the Lark/Feishu deployment question, #1466722439789674741 is probably the best place to ask so it stays searchable.

[merry slate]

你好，这里是openclaw的开发者社区，不是用来讨论使用问题，你可以去 #1459642797895319552 这个地方提问题，又或者去 #1474434870259224723 里面的飞书频道提问，会有飞书的维护者解答，谢谢

[fallen gate]

[cursive palm]

[shadow garden]

we need a translate bot up in this place

[fallen gate]

we had one in the international channels I believe

[cursive palm]

where are the internationl channel？

I used to chat there.

[jolly wolf]

temporarily closed since our translation bot crashed, back soon! (later today)

[shadow garden]

should just make it available to be on call in all channels: "/translate" reply to trigger, etc

[fallen gate]

bring this up in 30mins lol

[bleak fable]

OC just leaked my private chat into a group. 🤦🏼‍♂️ might be lancedb issue. Dunno. Narrowing it down

I have seen quite often session leaks

[shadow garden]

make a copy of your config as well for sanity (to ensure it wasn't related)

here you go @fallen gate -- your first P0:
https://github.com/openclaw/openclaw/pull/82374

[bleak fable]

I have seen this happen a few times in my own sessions. OC randomly replied to a very old message. Which is why Im pondering if its some memory injection issue

[shadow garden]

what runtime you using?

[bleak fable]

codex

[shadow garden]

there was definitely a bug in the past in codex with this (I hit it myself) but it should in theory have been fixed since at least this weekend (and likely a bit prior to that) -- it would quite literally pickup a different thread from a different session

[bleak fable]

I am on the latest version. Im seeing random replies to old chats or other session chats. Really makes me think something is injecting memory randomly

Ill try disabling lancedb and qmd and test

[next helm]

Openclaw updated cleanly from 2026.5.12 to 2026.5.16-beta.5; LaunchAgent is loaded, gateway is reachable, comms reconnected after a short post-restart window, plugins/tasks look healthy. The audit did surface one real routing/auth issue: a direct smoke test configured for openai/gpt-5.5 failed by looking for openai-codex auth, then succeeded through Anthropic/PI fallback, so the visible OK hid that the intended OpenAI route was not proven. Sanitized report: https://github.com/openclaw/openclaw/issues/83223

Version seems super smooth 👍

[vocal night]

oof there's gotta be a smoother way ot handle that piping, like 80% of my version bugs are schema/auth/ openai vs openai codex

[smoky dock]

quick update on https://github.com/openclaw/openclaw/pull/78583#issuecomment-4460034789 I published the agentkit work as an external plugin path and opened 4 smaller openclaw PRs with the minimum host APIs needed to get the plugin working end-to-end:

• https://github.com/openclaw/openclaw/pull/82431
• https://github.com/openclaw/openclaw/pull/82434
• https://github.com/openclaw/openclaw/pull/82471
• https://github.com/openclaw/openclaw/pull/82752

would appreciate a review when you have a chance, and of course feedback is very welcome, especially on whether this is the right minimal split for supporting external HITL plugins without bringing agentkit back into core, thanks in advance!

[vocal night]

@stable pewter already merged in a patch for message tool to not suck in webchat, but it seems to still be an issue (agent calling message tool in webchat vs sending message), might just need to be a webchat UI fix: tool = message tool, don't show as tool and display actual message via interception by display, might be easiest fix

[nocturne jungle]

what’s going on everyone

[slender hill]

The remaining 3 PRs not merged have been updated to address comments and are ready for merge review

[vocal night]

sup pash, btw the message tool issue in webchat is back in this beta

[nocturne jungle]

oh boy

[vocal night]

installed and 5 messages in spent a whole fresh codex pro account

separate bug lol

thankfully i have 8

[nocturne jungle]

how is that possible

[vocal night]

usually infinite retries

[nocturne jungle]

can you have codex look at the session rollout file for that. very odd

[vocal night]

debugging now, usually related to retrying without limits or auth issue and then those retries when the auth gets fixed all keep firing and server says "TIME OUT"

https://klipy.com/gifs/tiktok-shaquille--k01KRVRF6N6X98RDYTVDDVT5W1W

[vocal night]

looks like it was some kind of auth bug. it wiped 2 accounts that were attached to 100% 5 hr and 0% weekly and rolled back the refresh from yesterday

hadn't opened since friday before refresh, so when I opened up today and updated > it basicaly reset them to pre-token refresh that went out

[charred crater]

Yeah something similarly weird going on with my pro account rn. Not openclaw related whatsoever. But my token limit reset was for May 23rd, and now it's moved back to the 18th with my old pre-reset limit back in place.

[vocal night]

verified by checking usage, zero usage since may 14, wiped out 2 accounts and tried a 3rd and immediately reset it too

[charred crater]

Tibo un-reset us

[vocal night]

naw its likely they keep changing auth fields and when it pings server overwrites

@nocturne jungle can confirm latest beta auth un-resets codex accounts and sucks em dry like a vampire

[nocturne jungle]

https://x.com/thsottiaux/status/2056109929048764895?s=46&t=rVovHIS17nkEQlGYb2NlMA

Tibo (@thsottiaux)

Seeing issues where usage limits are out of sync for some Codex users. Apologies and team is investigating.

X · May 17, 2026, 8:29 PM

[rocky pewter]

I was stuck on that too until I figured it out. Since re-aut, it’s been working smoothly again

[vocal night]

can confirm exactly it was the accounts tied to OpenClaw auth btw

[nocturne jungle]

Need to find the real root cause

Yeah if it’s an OC regression that’s pretty bad

[vocal night]

gut is its the new fiels and expirys etc getting fed back to server and undoing whatever patch he did

numerous new fields have been added to auth

also a migration option from codex to oc

haven't looked at code lately but anthropic ties their usage monitoring in app- if latest codex has anything similar DRM (token wise) it could be something out of sync with codex harness or something in app server needs updating

[vocal night]

re-auth is what killed all my connected codexes. I have 8 and a nice bar where I track them all so I was able to see it happen the moment I re-authed

[rocky pewter]

Hmm, i will check again thx for Info.

[vocal night]

it seems to be hitting everyone even other platforms

so seems like maybe just timeliness and someone merged something that overwrote tibs patch

regressed and erased everyones usage again- reset again lmao

[shell dawn]

hey pash, not sure if you saw my PR that peter merged, i think this hits a lot of odd behavior feedback that a few people here saw - 82864
ive had a few more codex runtime oddities but i think its my memory system at this point which is in the middle of getting updated to be compliant with the new codex behaviors around tool calls and compaction

[vocal night]

if you're using LCM I don't recommend it till this version is stable id rec turning off any memory or compaction/context plugins

[shell dawn]

yeah, i think im running into similar stuff lol

[vocal night]

lcm assembly if you're on old version tries to send millions of tokens at once

[shell dawn]

ok im not having that problem

[vocal night]

you wouldn't know if you have LCM, it will start compacting them in background lol

the latest version seems half stable (from yesterday)

[shell dawn]

my assembly has been fine, theres some changes to how codex runtime handles tool calls that caused some specific hits in my features to protect during provider failover and the compaction routines are a little different

[vocal night]

I'm thinking of doing a compatibility flag PR that if X app does not have compatibility in manifest it turns it off

[nocturne jungle]

on my phone, link me. what’s the tldr?

[nocturne jungle]

is it the relogin stuff vince added or what

[shell dawn]

the change of the provider string in the slug from openai-codex to openai caused openclaw to have inconsistent provider resolution that would sometimes have openclaw try to send it out the openai provider even if it didnt exist, and it would just assume the defaults which would look for a nonexistent api key which would cause failures and failovers, fix was adding an additional check for runtime = codex when doing resolution so it hit the openai-codex provider/auth path

the behavior was elusive, what i saw is that it would link in to gpt-5.5 on codex auth path in the codex runtime and work wonderfully for a long time, then... something would happen and it would lose that resolved path, and then it would hunt and fail (or i only saw the failures, maybe there were hunts and successes too that i didnt see) and get really stuck before getting fixed

[vocal night]

I updated > got the everything is broken > okay doctor > still broken > just openclaw config > migrated codex (worked for 3-4 turns and then codex exhausted) > looked at my codex tracker and 2 of my accounts that were in the auth profile were "un reset" to 0% and tried loggin in 3rd one and it unreset immediately

but given the post on tibs tweet it could just be anyone logging in or reauthing after X time

[lusty vigil]

do you have any examples of this? like, field names?

[vocal night]

Expires, accounted,chatgptplantype, oauth ref, source, id,

[lusty vigil]

thanks

[vocal night]

it used to just be like 3 fields type, provider, token

last few updates have added like 1 field per update

[nocturne jungle]

that’s the right approach

it’s merged yeah?

[vocal night]

also in non bug news @nocturne jungle are you adding in OC remote control for codex? I'm already using it since its exposed seems easy enough PR, just figured codex team was already adding it to OC this release

[shell dawn]

yeah. i believe the fix was simple and has held up in my testing, but, you have a bit deeper perspective on this process and there might be some additional fixtures to cover cases i wasnt able to uncover. failover and failback seemed to work, codex reviewed the hell out of it, but i am just a lil guy vibing with my agents

[trim vector]

@bleak fable are you still planning on working on the codex plugin prs? will be doing some work this week on similar capabilities (enabling plugin subcommand that re-directs to marketplace, adding more 1st party plugins)

[lusty vigil]

for the record, are you seeing breakage from this? or just observing new fields?

[clever narwhal]

hey all, just had my first PR merged, fixing a /verbose tool progress bug : ) feels good 🦞
https://github.com/openclaw/openclaw/pull/83214

was anyone else having that issue in tg where /verbose on wasn't showing tool calls running codex? just joined to see what ya'll are up to

[clever narwhal]

Anyone know if I'm eligible for the clawtributor role?

the /claim button isnt working despite being github linked.

On the original PR, I didn't have the right permissions set, it appears, and so it resulted in:

"Thanks for the work on this. ClawSweeper opened a replacement PR only because the source branch was not writable from the available bot permissions. branch tides, not contributor blame."

https://github.com/openclaw/openclaw/pull/83186
https://github.com/openclaw/openclaw/pull/83214

[slow inlet]

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

Another claude cli / reasoning streaming related pr

[shadow garden]

probably because your source branch was not writable, so the merge came from a different person (clawsweeper stole your PR)

[fallen gate]

@clever narwhal I pinged you with the instructions there

[nocturne jungle]

care to explain the desired UX here?

[clever narwhal]

Yep, I saw, but /claim isn't working for me because of what I described above and what @shadow garden pointed out (clawsweeper stole the PR)

Lesson learned! But is there a way to get the role manually, if the system doesn't account for that?

[karmic steeple]

Added <@&1458375944111915051> to Tyler Bea.

[clever narwhal]

I've been annointed, thanks : )

🥳 🦞 ❤️

[worldly tusk]

Hey… did you get any chance to optimize/remove the base-instructions codex prompt?

[tight wolf]

Hey, could someone help me verify the Clawtributor role manually please?

I contributed the Personal Agent Benchmark Pack in PR #78219:
https://github.com/openclaw/openclaw/pull/78219

The work was accepted and landed on main by Vincent Koc via this squash commit:
https://github.com/openclaw/openclaw/commit/440333125c4c8a896d18c08d8d09bb5a8e015b29

Because it was landed through a maintainer squash commit instead of GitHub's normal merge flow, the /claim bot cannot detect it as a merged PR under my GitHub account

My GitHub username is @iFiras-Max1
Could this be verified manually for the Clawtributor role?

Thanks

[karmic steeple]

Added <@&1458375944111915051> to Firas.

[tight wolf]

Thank you 😁

[torn sage]

Super small PR so progress streaming mode actually deletes the progress message for telegram: https://github.com/openclaw/openclaw/pull/83278

Thanks everyone! 🙂

[celest merlin]

howdy folks

Who's around for some more testing and helping getting a great release out?

Any P0 PRs for me to look at?

[weak walrus]

on now

[tight wolf]

Here

[celest merlin]

Main is basically what we wanna release.

I can land low-risk smaller fixes, or stuff that fixes regressions compared to what ships.

[weak walrus]

OK, will update now, you pushing a new one? or you good with what's there now

[celest merlin]

You can test main, I kick off a new beta ina bit, just checking if someone is yelling THIS MUST BE IN

[shadow garden]

you just merged it 🙂 (thank you sir!)

[celest merlin]

Appreciate it

[short dome]

Love the music lately. Such joy, such cheese 🥳

[celest merlin]

@thick osprey can you generate music yet or are you still crippled? try hard little bot

[thick osprey]

Music generation kicked off: release-night-jingle.mp3, 30s instrumental release-night jingle. If the tool finishes async, I’ll surface the attachment when it’s ready.

[celest merlin]

Did you test this well? @untold pollen is not around and he kinda owns Telegram

I landed a PR that make browser automation better and would love someone test it. https://github.com/openclaw/openclaw/pull/83099

[tight wolf]

Happy to help test main, I don’t have any P0 changes, but I can run through qa Lab / memory-related flows if useful

[shadow garden]

also got this one: not a P0, but a less than pretty context issue:
https://github.com/openclaw/openclaw/pull/83127

[celest merlin]

looking into

What helps is if you update all your installs and see if everything works and is fast.

We have too many config options to test all, but we can ensure the happy paths work.

[rocky pewter]

Sry this Text was not for here.

[timber beacon]

https://github.com/openclaw/openclaw/pull/80505 when it adds p1 does that mean if fixes a p1 or it found a p1? I dont see a comment on it like normal from @thick osprey

[zinc dock]

Sorry if wrong to post this but i bought this to krill which said might be a issue and now were in general talking about this same thing and ive experienced this today with my workflow .. "codex harness does not expose sessions_spawn like pi harness did. So spawning a sub agent, freeing up your main agent to chat and do other things, and getting a ping back to main agent when the work is done... not a thing right now.....spawn_agent starts a Codex-native child thread and returns immediately. Which is what we want. When the child reaches a final state, Codex records a completion event, basically task_complete, with the child’s final message. The problem is the bridge layer though. But the durable completion artifact I found in Codex transcripts is event_msg / task_complete with last_agent_message. OpenClaw’s installed dist does not appear to have a proper consumer for that as a parent wake/synthesis event.

sessions_spawn had a contract to notify the requester session, and worked well."

[jolly wolf]

It’s categorizing the fix

[timber beacon]

Nice, thats awesome.

[winter field]

hii! pretty sure this is a small PR. https://github.com/openclaw/openclaw/pull/83139 test-only follow-up to #82850, adding a regression case based on the one-streamed-text commentary+final scenario from closed #82645.

[weak walrus]

I gave my Version
v2026.5.16-beta.5 clanker a pretty big task, it' sone of my main agents, he has been online since Feb 7th - will report back lol - I'll have him run pinchtest too

[tight wolf]

I tested main from a separate local setup

For #83099 (browser automation):

• ran 20 related browser/CLI/routes test files, 190 tests total
• docs formatting + MDX checks passed
• test type checks passed
• started a separate gateway on its own port/state and it came up cleanly
• did a manual browser smoke: open/navigate - snapshot- screenshot - evaluate- form fill andclick via snapshot refs

[crystal relic]

Hi maintainers, Recently, I have been paying attention to the performance of gateway, which is still related to PR.

https://github.com/openclaw/openclaw/pull/83299 adds the repeatable benchmark

https://github.com/openclaw/openclaw/pull/83300 attributes the ACPX startup probe cost in traces,

https://github.com/openclaw/openclaw/pull/83301 uses that proof to overlap safe startup work, keeping ACPX behavior unchanged while

improving ACPX-probe-on restart readiness by about 60ms p50.

[celest merlin]

yeah greaet work on these, I merged a bunch!

[timber beacon]

There has been so many QOL/base improvements it’s feeling so good!

[celest merlin]

Yeah all the QA efforts and more people really starts to pay off.

It also helps that the bs-hype-train starts to quiet down.

[timber beacon]

Yea, seems like users are actually users now and not chasing their favorite influencer

[torn sage]

It's super small. Any advice as to getting it through the full testing flow? Doesn't seem like Mantis ran a test against it. No worries if we need to wait for @untold pollen to return

[celest merlin]

what would help if you extend the PR description, ideally with screenshots. I get so many PRs and I need to be so suspicious of them.

So if I see that someone actually TESTED what they sent, that will make it massively easier to merge.

no avatar is esp. worrysome

[torn sage]

I put my actual name though 😂

I will extend the PR and add a profile picture 🙂

[celest merlin]

says the treasure box

[torn sage]

There's a heart in there though ❤️

[celest merlin]

perfect that seals the deal

@thick osprey what's your take on https://github.com/openclaw/openclaw/pull/83278

[thick osprey]

My take: promising, but I would not land #83278 blind for release without real Telegram proof.

It appears to be an XS Telegram fix: “delete progress draft before final reply.” That matches a real risk in current code: progress-mode drafts are created as Telegram messages, then final durable replies are sent separately. The current dispatch path has cleanup hooks around answerLane.stream?.clear() in extensions/telegram/src/bot-message-dispatch.ts, especially around progress final delivery and final cleanup. If the PR narrowly ensures the progress draft is deleted before the final answer becomes visible, the intent is good and release-relevant: it avoids users seeing stale “working…” progress beside the real answer.

No blocking code finding from what I can verify locally, but proof is the blocker. The PR is already labeled triage: needs-real-behavior-proof and mantis: telegram-visible-proof, which is exactly right. For this one I’d want before/after Telegram screenshots or a short screen recording showing:

1. progress draft appears while work is running
2. final reply is delivered
3. progress draft is gone, not edited into stale text, duplicated, or left above the final answer
4. works in the Telegram surface this affects, ideally DM/topic if relevant

Caveat: this sandbox can’t live-fetch GitHub right now, and gh auth is unavailable, so I couldn’t inspect the exact PR diff or CI live. Local targeted test also didn’t run because pnpm tried to install and failed on network/store access. Based on cached metadata plus current Telegram code, my verdict is: good candidate if the diff is narrow and has visible Telegram proof; don’t land on PR text alone.

https://github.com/openclaw/openclaw/pull/83278

[fallen gate]

ngl.... this was me until like last week.... 💀

[celest merlin]

you are still a cat and you have your camera off. You are no better.

[fallen gate]

my dslr over heated 😢

[jolly wolf]

bro hatin on the cats :(

[timber beacon]

Time to get a cat pfp

[celest merlin]

stop saying that it's gonna be news in an hour

OPENCLAW FOUNDER HATING CATS

[jolly wolf]

im calling dexerto right now

[celest merlin]

My cat died 5 years ago and I'm still not over it so psss

[fallen gate]

we need start a union or something

[smoky dock]

Let's try this time #clawtributors message 😅

[celest merlin]

ohh nice, you actually followed through with plugin enhancements!

[plucky dew]

mine died 3 years ago he was the most precious part of my family i ever had , devon rex . At Christmas I am still crying about him^^ (Sorry no devtalk 🙈 )

[weak walrus]

For me - using GPT 5.5 Still not sending messages to discord on reply, they go to the web gateway.

I reply back with this command and he starts sending messages correctly:

JeffJHunter — 5:30 PM
@Beau | VA Staffer AI Persona status?
JeffJHunter — 5:45 PM
@Beau | VA Staffer AI Persona is tehre a reason you're not sending messages on discord? you should be using teh discord tool
Beau | VA Staffer AI Persona
APP
— 5:45 PM
Got it — switching to the Discord message tool going forward.

Harness engineering page is live and looking good:
https://ai.vastaffer.com/before-fine-tuning-harness-engineering/

[astral topaz]

lol peter the cat herder... multiple cats in maintainer team now

[celest merlin]

@thick osprey is using GPT 5.5 on discord and it works

[thick osprey]

Yep, GPT 5.5 on Discord is working in this channel: I’m replying here through the Discord surface now.

Jeff’s report sounds more like that specific agent/session is falling back to web gateway delivery until nudged by instruction, not a blanket GPT 5.5 + Discord failure. I’d treat it as a routing/session binding bug: check that agent’s Discord channel tool availability and whether the prompt/runtime is marking replies as message-tool-only for Discord after resume.

[celest merlin]

@thick osprey generate a song

[thick osprey]

Kicked off a short instrumental release-testing anthem. I’ll post the audio when the generation finishes.

[candid beacon]

But Gpt 5.5 for this is killing usage 😄

[summer matrix]

Hey guys I’m here because I read the article that Peter hates cats?

[celest merlin]

https://giphy.com/gifs/Ev477g37MJORyOWfdG

[celest merlin]

it should also work with GPT 5.4 also don't be poor

[candid beacon]

I really want openclaw to select different models for different tasks (difficultys) not me changing it manually

[celest merlin]

That trashes the cash and model switching inside a session makes every model crap since you lose the reasoning stream.

[candid beacon]

Damn. alright

[jaunty sphinx]

high not xhigh not that expensive

[candid beacon]

tell that my claude max^^

[plucky dew]

Openpaw when?

[zenith prawn]

Would be nice to use /fast for talks with claw and non-/fast for longer tasks 🙂

[celest merlin]

ugh careful on anthropic that trashes the cache REAL BAD

you can't selectively flip on fast without huge penalty

[candid beacon]

🙁

[celest merlin]

I know right? everything is hard. boo technology.

[candid beacon]

[weak walrus]

Waht's really strange? it starts writing in discord, and then it like deletes itself - I can see it in web gateway still.

I'm trying to get codex to figure this out, I didn't make any changes to it, just updated it and asked it to do same tasks I always do

[candid beacon]

Tech is never the devil, its the human made policy scene.

[clever narwhal]

Might have a PR for review in a little bit

On verbose mode, Openclaw gives a more instant preamble before doing tool calls, making it feel snappier

[shell dawn]

which lever is that? text mode?

[celest merlin]

should that be in verbose or default

[celest merlin]

maybe some conflicting stuff in agents.md?

[clever narwhal]

Yeah I would think but it never worked for me

[clever narwhal]

Hmmm

[shell dawn]

i would be really happy with something that creates a more uniform 'the agent has started doing something' response that things like opus tend to get right more often, GPT-5.5 does a better job here but can still result in the only response being one giant 'here is the answer you were looking for'

helps relay info that its working and not stalled

[verbal heart]

kinda same issue but no connections using oc webui - https://github.com/openclaw/openclaw/issues/83344

[clever narwhal]

Does everyone else besides me get a quick response from their agents before they run tools?

I thought I was fixing a bug or improving the ux, but maybe it’s just something on my end

I have a fix, if it’s necessary

[neon swift]

Usually I see something beforehand, and then I also see a lot of the tool printout as it's happening so there is some indication that it is actively working

[zenith prawn]

for me only way to get it was 'incompatible' config but it works now:

• messages.groupChat.visibleReplies="automatic"
• channels.telegram.streaming.mode="partial"
• channels.telegram.streaming.preview.toolProgress=true
  UPD: sorry, its about tools itself, not 'response before tools'

[shell dawn]

What is your method I will review

[zinc dock]

I ran into this issue but it was kind of weird because it would leave me in the dark. After a couple of minutes I would abort the agent but as soon as I aborted the message it intended to send comes through. It was super frustrating.

[clever narwhal]

I’ll get ready to push it. It works for me.

Codex 5.5 extra high telling me that there’s nothing conflicting on my end so let’s see

[shell dawn]

Cool, this is an area that bugs me so I want to see

[clever narwhal]

It ideally would more deterministically give a quick response that it understands the request before it begins to work (giving you time to stop it if it understood wrongly)

[weak walrus]

You're right so far - I'm reviewing this now and it could be my agents.md, which believe it or not, is a good thing.

I feel like OpenClaw takes my agents.md more serious than before, but I've given it too much room to "decide" instead of clearly defined rules.

🙂 Will report back - dont' want t owaste more time, will continue testing

[merry slate]

Use acpx to offload complex tasks to Claude Code or Codex

Their harnesses are way better for heavy coding than OpenClaw's general setup — it's not just the model, the whole agent architecture makes a difference.

[verbal heart]

feel like things would be more useful if github had version labels

[rocky pewter]

Opened a few small beta bugfix PRs. Thanks for ur time ❤️

PR for #48684:
https://github.com/openclaw/openclaw/pull/83348
It lets bare stop/abort/wait messages skip debounce, so they run immediately instead of waiting behind normal chat text.

PR for #83134:
https://github.com/openclaw/openclaw/pull/83317
It fixes the default VAPID subject for push setup, so the generated config is valid out of the box.

and these 2 from earlier today. #clawtributors message

PR for #83349:
https://github.com/openclaw/openclaw/pull/83400

It stops strict Codex runs from silently falling back to Sonnet when the Codex harness is missing.

Tested current main with real browser-control on a clean VPS.

Chromium launched headless, /tabs/open loaded example.com, /tabs saw the tab, and live browser doctor passed.

[jolly wolf]

wdym by version labels

https://github.com/openclaw/openclaw/releases

[verbal heart]

for PRs

or issues

[shadow garden]

So much for the music 😒

[smoky dock]

when you have a chance to 👀 please let me know if you have any feedback, thanks again! 🙌

[fallen gate]

probably wanting npm smantic major.minor.patch

[neon swift]

Maybe I am misunderstanding but I don't think version labels on PRs or issues would help, because all PRs impact future versions/older versions aren't patched retroactively, so I think it all kinda comes out in the wash anyways so to speak

[plucky sentinel]

Hey folks! 👋 I've been contributing a few PRs over the past week and they're all
green / proof-ready — would love if a maintainer could take a look when you get a
chance.

All CI ✅, no new deps, zero behavior change:

• #81635 — exempt _testing / require from underscore lint (1 file, XS)
• #81808 — enable object-shorthand lint rule (part of #80504 plan)
• #82100 — extract GoogleChat schema into its own module (follows #82007 pattern)

All proof: supplied / sufficient, all checks green. Thanks! 🦞

PRs: https://github.com/openclaw/openclaw/pull/81635
https://github.com/openclaw/openclaw/pull/81808
https://github.com/openclaw/openclaw/pull/82100

[shell dawn]

ive been looking into this more, ive got my agents trying out command line invocation but this doesnt seem like hte best way. ive got a headless linux server so the cli is what i have available, is this an area where acpx still makes the most sense?

[clever narwhal]

Sounds familiar

[zinc dock]

Hope your fix addresses this. Fingers crossed

[true vapor]

Have you installed claude code in the Linux server ?

[shell dawn]

yeah ive got both claude code and codex, though for what im looking at codex is a better fit

[willow narwhal]

Hi everyone, we got 2026.5.16-beta.6 out and would like you all to kindly test.
You can find the changes here: https://github.com/openclaw/openclaw/releases/tag/v2026.5.16-beta.6.
Ideally, update to the beta, use your claw as you would, if anything feels/look off, flag it.

[willow narwhal]

<@&1503801512294486217>

[clever narwhal]

Btw, it's pretty wild now that you can ssh from the chatgpt mobile app now -- basically the last openclaw PR I did was mostly from my phone while running errands, how wild.

[quiet timber]

i had to roll back because of a protcal missmatch, is that fixed yet?

[willow narwhal]

Yep, its fixed!

[quiet timber]

cool i give it a try

[crystal relic]

hi maintianers! this PR need help for review.

[shell dawn]

thats to the codex desktop client right? i just read about that

[celest merlin]

you mean the accidential v5 bump in one beta? resolved.

[fallen gate]

yea, it's dope

[clever narwhal]

No, I have it sshed into a hetzner vps which is up 24/7.

[quiet timber]

yeah it happened right as i was trying to do a rebuild of my container to get ffmpeg, and i though i broke it XD

[celest merlin]

Hey I appreciate that work, right now focus is on release stabilization and getting it out. This doesn't seem to be related to that work.

[crystal relic]

let me test v2026.5.16-beta.6.

[celest merlin]

You all lure me with useful stuff and then the release gets further and further away 🙂

[clever narwhal]

From the codex desktop/mobile app go to connections and you can add a ssh connection, and this is where you could add a connection to another machine, a dedicated mac mini, a hetzner vps, et cetera.

[weak walrus]

I'm pretty pissed, everytime I try to use codex from mobile I select "I"m signed in on desktop" and then it says waiting for desktop, and i never see anything on desktop :/ so still haven't got it going

[hollow frigate]

Do you see an option in the left hand menu that says something like "mobile" ? I had the same issue, but I had to go to that and then intiate it on the desktop as well.

[shell dawn]

found something even better, the linux codex app-server can do it too

[zinc dock]

Do tell lol

I'm waiting very impatiently for a Linux Codex app

[untold pollen]

Taking a look! Sorry that Clawsweeper closed it, will check why

[thick osprey]

I checked live GitHub. ClawSweeper closed it as “already implemented on main,” not because the PR was bad.

Timeline:

• clawsweeper[bot] reviewed at 21:49 UTC and said current main already has the cleanup via b5c337909713.
• Mantis later couldn’t produce useful Telegram visible proof because the harness couldn’t emit tool-progress callbacks.
• alexph-dev added useful nuance: this helps stale progress drafts, but not the partial + preview.toolProgress / native quote-reply gap, which is tracked as #82089.
• ClawSweeper labeled rating: 🧂 unranked krab, P2, then closed at 01:17 UTC.

So the close rationale is defensible if main really has the same behavior. The questionable bit is process/tone: closing a contributor PR while release stabilization is ongoing can look harsh, especially when the PR is tiny and reasonable. If you reopen, I’d either land only if it differs from main, or comment “already fixed on main in b5c337909713; thank you, this was the right area” and leave #82089 as the remaining Telegram gap.

[crystal relic]

I remembered peter say in blog that will be an LTS version this month. 😆

[merry slate]

Just install the acpx plugin for OpenClaw, then set your default agent like Codex in the config file. After that you can simply use natural language in OpenClaw to trigger it, and it will spawn a sub-session to run the tasks

[weak walrus]

sadly no

[untold pollen]

oof, just saw this. Any PRs you want me to check?

[hollow frigate]

What OS is your desktop? The other thing I'd try is going into the settings on desktop, and to the connections section. You can add a device there. When I added an android phone, I actually had to relogin on the phone, even though I was already logged in.

[willow narwhal]

https://tenor.com/view/durv-durv-swinging-swinging-gif-swinging-ae-durv-gif-gif-19130137

Oops, gifs not working in this channel

[plucky dew]

nope^^

[crystal relic]

😂

[shadow garden]

Tak got it in: https://github.com/openclaw/openclaw/pull/83248 been running fine since

[jaunty sphinx]

hi peter im ready

[quiet timber]

pulled down the git, changed to the beta branch and did a pull, and OPENCLAW_DOCKER_APT_PACKAGES isnt working

[shell dawn]

Does it give access to the structured commands like /review or codex specific skills?

[untold pollen]

Folks feel free to throw any Telegram/Android PRs my way

[weak walrus]

I'll try that

[merry slate]

Slash commands basically trigger specific agent prompts via text input. I haven't actually tested Codex yet, but I believe it works fine. You can give it a try

[willow narwhal]

Was this working before?

[merry slate]

btw I just upgraded to 5.16 beta.6, and found that sessions spawned via acpx no longer show up in the session list. Example sessionKey: agent:claude:acp:25f77580-de30-4d80-9bc3-7cbc6374bce7

I'll look into what's causing this issue

[quiet timber]

yeah odd thing is setup.sh gives no option to setup/install the packages, so it thinkgs they are gettign instealled

[final ocean]

whats your setup os etc

[quiet timber]

26.04 lts, did a fresh git clone, and set the branch to beta 6

copied my .env in and did docker compose build, then setup

[willow narwhal]

What version was this previously working on?

[quiet timber]

2026.4.23 and 2026.5.12

must of been something i did, i nuked my openclaw download, repulled form git and it worked

[willow narwhal]

Okay, good to hear it works lol

[quiet timber]

should just have support for apt installing on boot, instead of rebuilding container

[finite solar]

Any clawtributors able to repro this for me.

[merry slate]

hey, heads up — after upgrading to 5.16-beta.6 I noticed ACP-spawned sessions
(like agent:claude:acp:...) no longer show up in the Control UI session list.

Traced it to #81386's configuredAgentsOnly filter.

https://github.com/openclaw/openclaw/issues/83397

[finite solar]

bonus points for those who can do the same with a provider too

[willow narwhal]

Thanks for reporting, Peter is on it.

[celest merlin]

That is a more niche code path, would love your help to make this great.

[celest merlin]

I work on that already.

[winter field]

👁️ 👄 👁️

[nocturne jungle]

No, it's on the list though. TLDR need to remove everything from OpenClaw base instructions that's already covered by Codex base instructions, and then append only important Openclaw stuff on top of codex instructions with override

[merry slate]

also found another small one

browser screenshot fails on fresh install
because sharp isn't in optionalDependencies, so npm never installs it.

Probably been there a while, not a beta regression.
One-line fix, PR up: https://github.com/openclaw/openclaw/pull/83403

[silent yarrow]

After updating to beta 6, I’m not getting replies from my discord bot.. I see it’s thinking process and typing indicator, but never receive a response.

[rocky pewter]

U want me to post it better in a thread ?

[finite solar]

Yes please

[rocky pewter]

approval smoke

[fallen gate]

@willow narwhal we back baby

[naive horizon]

Was able to repro

2026-05-17T20:33:56.980-07:00 gateway/ws: exec.approval.waitDecision failed immediately with INVALID_REQUEST: approval expired or not found, after approval id 1e079e0b-92e8-41f8-9297-7190b0c62fc7 was shown as pending for command `printf 'smoke\n'`.

Was not able to repro on 2026.5.12

[finite solar]

ok so this is just a beta issue?

let me make a PR

[finite solar]

Pat did you try what happens using another (non web ui) interface?

[naive horizon]

correct, believe its just a beta issue

[naive horizon]

no - will try rn

[celest merlin]

My ~/.codex/sessions is 63 GB / 43k JSONL files. Was wondering why /resume was slow lol

[shell dawn]

thats called deep context

[vocal night]

I recommend just a pi vs codex harness switch prompt wise. Otherwise we’re cutting into prompts that make pi work for our users. I might work on this today.

[vocal night]

Gmorn from asia 😮‍💨 downloading and testing now

[vocal night]

Codex in my experience so far agent md takes it very serious and needs to be just link to runbooks/table of contents to context it needs. Otherwise it will do everything in it, even if if not useful to that turn. I prefer if statement/logic based gates to different paths

[naive horizon]

forgot to follow up here, yes, repro'd with tui

[winter field]

same, investigating this rn

[fast pagoda]

Repro restarted and confirmed.

Setup:

• Crabbox: Azure desktop fallback, coral-prawn / cbx_d7980ecc70f2
• OpenClaw: 2026.5.16-beta.6
• Remote repro root: /tmp/openclaw-repro
• Real Gemini provider used: google/gemini-2.5-flash
• Control UI: http://127.0.0.1:19081/chat?session=approval-smoke-2 inside the Crabbox desktop

Result:

• Exact prompt submitted through Control UI.
• Exec approval UI appeared.
• Clicked Allow once.
• Command did not run.
• UI transcript says approval timed out and no stdout exists.
• Gateway log has the maintainer’s failure signature:

exec.approval.waitDecision ... INVALID_REQUEST errorMessage=approval expired or not found

Concrete log line:

2026-05-18T04:32:16.370+00:00 [ws] ⇄ res ✗ exec.approval.waitDecision 0ms errorCode=INVALID_REQUEST errorMessage=approval expired or not found

Screenshot captured on the Crabbox at:

/tmp/openclaw-repro/repro-approval-failure.png

I’m leaving the lease running for maintainer inspection.

Don't know what's happening but can't send screenshot here..

[finite solar]

if you have any channels hooked up to it please try throiugh those too

[fast pagoda]

Discord:

Config:

• Discord bot: Crabboxer
• Channel:
• Guild:
• Token is env-backed via DISCORD_BOT_TOKEN, not written into openclaw.json
• Config snapshots created under /tmp/openclaw-repro/home/.openclaw/openclaw.json.bak.*

Channel repro result:

• Started an agent turn with Discord delivery to that channel.
• Crabboxer posted the exec approval request into the test channel.
• Gateway reproduced the same failure signature:

exec.approval.waitDecision ... INVALID_REQUEST errorMessage=approval expired or not found

Concrete line:

2026-05-18T05:05:46.269+00:00 [ws] ⇄ res ✗ exec.approval.waitDecision 0ms errorCode=INVALID_REQUEST errorMessage=approval expired or not found

Also saw a follow-up lane issue:

file lock stale for .../sessions/1201e014-c917-45d0-a5e2-1c73b73d50e4.jsonl

Human interaction:
Me:

Use the exec tool exactly once. Do not ask me to run this manually. Do not change the command. Run this exact command in the current workspace, then summarize stdout only:

printf 'discord-smoke\n'

Agent:
approval card, then:

The command I attempted to run did not execute because the approval to run it timed out. Therefore, there is no output from the command.

[weak walrus]

FYI regarding discord replies - It's definitely not JUST my Agents.md. I was able to replicate this problem with brand new openclaw install where it doesn't reply in discord, it shows up in web-gateway though.

After I asked it what happened and it "troubleshooted" and confirmed how to handle in the future it said:

"Got it. So the pattern is clear now:

Message tool → works perfectly in Discord,
Auto-delivery (text-only replies) → web gateway only, not Discord"

[finite solar]

thank you! I think we got this sorted <3

[vocal night]

sounds like a routing issue/wasn't patched for auto delivery lane

[winter field]

hit this live on my VPS bot after the beta update: assistant finals were written to session state, but Discord delivery dropped them in progress streaming mode, so no reply was delivered.
PR has the fix + live proof, and ClawSweeper marked proof sufficient/no repair needed:
https://github.com/openclaw/openclaw/pull/83443

[vocal night]

before I waste inference is "VOICE_MESSAGES_FORBIDDEN" on telegram a feature or a regression?

[celest merlin]

thats a telegram permission thing

[winter field]

one unrelated flaky cron shard failed, exact failing test passes locally :/

[untold pollen]

Check your account settings, it needs to be enabled

[vocal night]

I found it, bots don't count as contacts in latest telegram update (or maybe never did) so permissions to contacts only (lots of spam last few days) blocked voice

thanks for heads up

[merry slate]

I found an issue in beta.6
after the agent loop finishes, the ControlUI state doesn’t stop, and the agent ends up sending a “...” message card at the end
I recall someone reported this before. Is there already an existing issue or PR tracking this now?

[potent path]

I've been testing Openclaw Code Mode. Still running full tests but I do have a few early issues. Here's the first PR:
https://github.com/openclaw/openclaw/pull/83444 -- failure codes such as 'timeout' and 'runtime_unavailable' were collapsing into 'internal_error', making them indistinguishable from opaque internal bugs.

[vocal night]

how is code mode different than the new codex harness default mode? codex harness is for coding default, so what is different? max turns?

[potent path]

#80600 introduces Openclaw Code Mode -- meant to be an equivalent to Codex Code Mode, but works for the Pi-runtime. As Peter said, it's experimental and opt-in only at this point, but seems pretty functional so far.

Codex Code Mode is already in production and is the default option for the Codex runtime/harness.

[vocal night]

thats what I figured thanks for expl

[devout galleon]

I've seen similar behavior in Discord for some time, but never in the Control UI and even not now on beta 6. GPT/Codex harness, if it makes any point of differentiation for you.

I've not really been present today so forgive me if I'm just repeating what other say but in beta 6 Discord is ok, Telegram does not respond. Nothing apparent in the log, just no response.

[celest merlin]

sountrack of theday

[odd anchor]

https://tenor.com/view/gift-dance-gif-6230678

[vocal night]

can we get an nsync soundtrack "OC back alright"

[plucky sentinel]

Hey folks 👋 — I have a PR that could use a review/merge: #83422

What it does: Enables eslint/no-underscore-dangle as an error in .oxlintrc.json with a comprehensive allow list covering all 170+ existing underscore-prefixed identifiers, bringing lint errors from 90+ → 0.

Context: This supersedes #81635 (which I've closed), which only allowed _testing and require. The broader allow list was recommended by ClawSweeper's review — making the config the single source of truth instead of relying on the wrapper-level --allow suppression.

Verification: npx oxlint src/ extensions/ → 0 warnings, 0 errors. Proof section in the PR body.

It's currently labeled unranked krab — happy to address any concerns. Thanks!

[vocal night]

you around for some telegram bug/PR's?

[untold pollen]

Always

[vocal night]

debugging now, finding some queve issues leading to long response times

[celest merlin]

more like unranked cat!

I'll look into it. Totally not what we need for the release but still useful 🙂

[vocal night]

found 3 bugs 1 P0 for telegram posting it now

[plucky sentinel]

Thanks for taking a look, Peter! No rush at all — happy to address any concerns whenever it fits the release cadence. 🐱

[celest merlin]

did you not run codex /review or autoreview?

too many booboos in PRs.

[plucky sentinel]

Let me run it now and fix any findings. Sorry about that!

[vocal night]

I have the issues filed + found another few mentions but no PR that covers, so working on 2 right now

[untold pollen]

sounds good, please link the PRs once done and I'll be happy to merge

[vocal night]

new harness is interesting- agent said "this some goblin shit" 👺

we need a hood-claw personaliy plugin

voiced by sam l jackson voice clone

[fallen gate]

what's with all these cats... jeeez

[odd anchor]

https://klipy.com/gifs/sleepy-cat-cute-cat-2--k01KRWZFT04TRPDBVAYC482TAF7

[weak walrus]

Going to bed ya'll been openclaw'ing all day

Even got my anti-openclaw developer on board and witin 2 hours of install, he already made a plugin for code-review for our internal projects and working well with Gemma4:26b (PS. I think this is the most underrated openclaw model capable bad-boy)

[jaunty sphinx]

best fit for a standard mac mini

[vocal night]

@untold pollen https://github.com/openclaw/openclaw/pull/83503 servers are running slow as dirt here's first

[plucky sentinel]

Fixed it. Ready for review whenever it fits your cadence. 🐱

[untold pollen]

Thanks for the PR! Reviewing and landing in a bit

[vocal night]

if you're still up, moved prompts to runtime scoping vs the complex spiderweb to prevent werid bugs we're seeing (not critical this release but I recommend review) https://github.com/openclaw/openclaw/pull/83454

can't do much testing out of asia right now, codex fast lane is moving so slow can't even code, just burning credits will have to wait a few hours

[cursive palm]

I've noticed that OpenClaw's WebUI is indeed getting better and better.

[vocal night]

the slow as molasses system right now somehow landed another commit after the fact fyi

[untold pollen]

yeah re-reviewing after that

[vocal night]

thanks, me too put 2 waves on it to filter down by slice 4 > 4 agents and fixing the 3 P2's found

[crystal relic]

https://github.com/openclaw/openclaw/pull/83520 ready for review

make gateway readiness off the update-check path while preserving update notifications and cleanup behavior.

[vocal night]

should be good now, passes all my checks and bug audits

not the perma fix but bandaid + regression testing for this release while main keeps moving/codex harness stability

[potent path]

https://github.com/openclaw/openclaw/pull/83473
Bug: 'agents.list[].tools.codeMode' was documented as a setting to enable openclaw (Pi) code mode per-agent, but that was rejected by the strict per-agent schema.

To fix that and to make per-agent code mode configuration work end-to-end, this PR adds:

• Accept agents.list[].tools.codeMode in schema, exported types, labels, and help text.
• Resolve code mode from the active agent, with agent settings overriding top-level tools.codeMode defaults.
• Wire the active-agent result through catalog activation and model payload enforcement so agent-scoped code mode is not accepted as inert config.
• Preserve provider grouped tool payloads, such as Gemini functionDeclarations, while filtering code-mode payloads down to exec/wait.

[plucky sentinel]

Hey folks — my PR #83518 (fix QMD archived session visibility drop) is ready for review: CI green, ClawSweeper cleared, both dot-form and slug-form archive paths covered. status: 👀 ready for maintainer look. Happy to address any concerns before merge. Thanks!

[untold pollen]

Landed, thanks 🫡

[hollow mural]

Hi
Would love this PR to be reviewed: https://github.com/openclaw/openclaw/pull/83259
Thanks

[cursive palm]

beta 6 is ok , but when i update to beta 7 , i got some message when i chat with OC

:warning: Something went wrong while processing your request. Please try again, or use /new to start a fresh session.

i use /new to start a new session but it happens again 🦞 , i'm using OpenCode GO / DeepSeek-V4-Pro with Telegram

[celest merlin]

hmmmmm

can you open codex point it at your repo and session log to debug?

[celest merlin]

on it

[untold pollen]

Thanks for the PR! Took me a while to test and review but landing it now

[celest merlin]

reviewing

[celest merlin]

OpenCode has a bug in their Go server where they sometimes send mixed reasoning headers, ty switching to a different model provider to see if that narrows it down.

[vocal night]

feel free to take over, 95% confident its right shape but running into some speed issues on my codex after burning 10b tokens 😂, might have to cutover my db

[slender hill]

Bug fix PR ready for review - 83172. Fixes subagent thread delivery and reporting in Discord. Present in current betas. Was broken in a Feb release.

Two feature PRs ready for review - 83169 and 47687. They make reactions a first class communication channel. Who doesn't want to give their clanker the 🦞 and let them know to get moving. Right now reactions are not processed as a wake event so they are ignored as something to review and possibly take action against.

[celest merlin]

throw URLs at me!

[slender hill]

https://github.com/openclaw/openclaw/pull/83172
https://github.com/openclaw/openclaw/pull/83169
https://github.com/openclaw/openclaw/pull/47687

[timber beacon]

https://github.com/openclaw/openclaw/pull/80505

[celest merlin]

interesting. you want the agent to wake up on reactions?

[celest merlin]

were you able to resolve this?

Or is anyone else seeing such issues?

[slender hill]

yes, either as a thumbs up to say that the work was good and dosn't need more work, or to give custom reactions that are defined in prefs that are a quick way to pass msgs to openclaw for context or action

[vocal night]

that would be super complex + queve breaking given reactions can queve up in the 10-100+ range easily

[smoky dock]

There you go 😅

https://github.com/openclaw/openclaw/pull/82431
https://github.com/openclaw/openclaw/pull/82434
https://github.com/openclaw/openclaw/pull/82471
https://github.com/openclaw/openclaw/pull/82752

[slender hill]

doesn't need to be https://github.com/openclaw/openclaw/pull/83169. the change is small by size and doesn't need to be part of a nested reaction storm

[vocal night]

best thing to do if you want it as approval is just send thumbs up in chat, would be really difficult to institute and manage in every messaging provider wanting it as a seam- a popup card would be better and I believe I already built a seam for that for approvals/denials in telegram (it's buildable via sdk)

[slender hill]

a dedicated msg not a reaction already works. Reactions are ignored. It is an odd UX flow since most people will assume that reactions have the same intent of a msg but you're clearly more into the weeds than I am

[vocal night]

I get it, I just spent 3 hours on the telegram queve system the complexity of regressions and mess from reactions is high

its not 88 loc maybe 2000

[slender hill]

I'll leave 83169 and 47687 out there for consideration. They are truly feature enhancements. Small by design but you and other maint. should agree that it's worth it.

83172 is a bug on subagent delivery issues and should be treated differently.

[vocal night]

btw not shooting them down just advising you to think system wise

[slender hill]

feel free to close the two feature ones if that's what everyone wants to do. I'm good either way.

[vocal night]

and if its truly 88 loc and a feature you want you can just patch your own system

once you test it and find the breaking points > iterate and fix it > it will get to like 2000 loc lol

[slender hill]

seriously all good, dude. I can certainly do it locally. Was passing it upstreamt to see if it was of interest.

[vocal night]

peter assigned it, he may add it, we're just in the middle of debugging the release 🫪

right now we all want stable release + patch vs features

[devout galleon]

just updated to beta 7. web still good. discord still good. previous jammed telegram messages got replies and telegram is working properly

[slender hill]

and it's should def not be a release blocker

[willow narwhal]

Good to hear!

Link me to anything worth looking at.

[devout galleon]

running through my usual tests

[worldly tusk]

i wish there was also an option to disable codex default base_instructions

[celest merlin]

i still have them open. not happening in this release. Already way too distracted with bugs.

[celest merlin]

thank you!

[vocal night]

can't run through codex harness and not do so without editing it's prompts, you can just patch them over on your own, its just agent > find them and write whatevr you want

[celest merlin]

Longer term thing, maybe we can lobby

[vocal night]

the biggest issue I've seen are agent.md conflicts, I have a skill in my repo that your agent can run to find conflcts and slim down your agent file

[devout galleon]

I was afk for claw today because I was actually working on an abandoned repo the company "open-sourced". Every time I was commiting I was thinking hmm, what would Peter say about this PR

[willow narwhal]

lmao

Use codex auto-review in your company

[celest merlin]

eh, i push plenty slop myself

autreview made everything take hours lol

worth it tho.

yesterday i had one land take 7h

[untold pollen]

I've been clawpatch-ing so many repos lately

[devout galleon]

yeah, i was. tests, codex /review, all the good stuff. the work was 100% better than if I hadn't had some misses and half-misses here

[celest merlin]

each time it fixed sth that touched a new file, next review round included that file, found a new bug, edited another file, found a new bug, etc.....

[devout galleon]

I appreciate the time taken to educate

anyway, back to testing beta 7

[worldly tusk]

the default base_instructions have personality issues, and also it has a wall of text injected that may not be useful outside of coding tasks... so it is not optimal for openclaw 😄

[devout galleon]

every time I've tried a PR here I start out like, "small issue, should be easy". 2 hours later, "jesus christ"

[celest merlin]

codex decided parsing 70GB of jsonl files should take <10 seconds

It came up with some really clever caching after endless tries and review rounds

[willow narwhal]

dont let Nik see this

[vocal night]

I ship a lot of PR's, here's a comment with the skill I've used, works a lot better and I don't spend 5+ hrs anymore in cycles https://github.com/openclaw/openclaw/pull/83503#issuecomment-4477403208

[worldly tusk]

he's seen, actually: https://github.com/openclaw/openclaw/issues/78938#issuecomment-4402951821

[runic totem]

PR #83569 is ready for maintainer review when it fits the queue:
https://github.com/openclaw/openclaw/pull/83569

Fixes #10029 by adding syntax highlighting for WebChat code blocks. CI is green, ClawSweeper cleared with no blocking findings, browser proof and dependency-change notes included.

[vocal night]

can you link me to the file I'll check the personality issue? the only ones I don't love are our "always be happy never be anything else" 😂 it's good imo to know when model is under pressure, that being said 5.5 personality instructions are super weak by design, it goes unhinged VERY fast if you disable them

[devout galleon]

this seems fairly recent addition to logs "[agent/embedded] strict-agentic execution". can i ask what this is telling me?

[vocal night]

check my PR on it above i was working on zone adjacent to this today

[devout galleon]

also agent-sent tts seems broken in telegram but i need to try some more providers to be exhaustive about that one

perhaps more specifically, "a request to the agent for a tts message in telegram results in no reply"

the /tts command works fine

[smoky dock]

Thanks for following up, all good! Not in a rush, will try next one 🤞

[vocal night]

found a bug with @tight hornet maybe a feature but he added me and started trying to give me support lol

[devout galleon]

that's... not Krill if he added you

I get like 4 "Krill" friend requests a day

[vocal night]

Krill 🦐 | ClawBot
krill.app

lmao I got krill-fished

[devout galleon]

definitely let him help you with your btc wallet

[worldly tusk]

i've seen that you removed overlay's personality. that's a good change. now there is no prompt conflict between codex base_instructions personality (that wa cannot configure, and it is set to pragmatic) and overlay persona (that can be other, friendly by default).

two other personality issues that are left:

• there is a harness option to switch codex personality, and it can be exposed in openclaw config. that would fix conflicts between what is there (unconfigurable pragmatic) with what is set in soul.md (as the user wants it to behave)
• the base_instructions prompt start with "You are Codex, a coding agent based on GPT-5". this is a personality poison. openclaw default prompt has this handled much better: it does not assume the agent's name or primary function. that is why i want to be able to disable the base_instructions alltogether

[vocal night]

I started to argue against #2 but theoretically it could be anchoring the entire coding quality experience

it takes a lot of testing to change harness prompts for 50M+ people using the system, I tend to try to leave anchor (beginning and end) prompts alone. Except for at end is easier to add personality like the mesage tool hack.

[worldly tusk]

that is true. if the user wants the agent to be a coding agent, he should leave this anchor alone - it is good. however openclaw is not used only for coding, and in that case this anchor does a disservice.

this actually could be a larger issue if your fix is merged, as you are removing soul.md from developer_instructions and moving it to user prompt, effectively reducing its priority.

[devout galleon]

agent sent tts also broken in discord for me. i don't think this one has ever worked in discord. might be the same or similar path to peter's music requests. clanker says

Evidence from this run:
tts tool call started at 22:10:12 AEST and returned success at 22:10:14.
It created a real local Opus file: /tmp/openclaw/tts-8Jg6zD/voice-1779106214531.opus (41 KB, valid Ogg/Opus header).
The Discord channel state still shows lastOutboundAt as 21:40:44 AEST, which is the earlier beta-7 text reply, not the TTS attempt.
Gateway log has no Discord outbound/media send around 22:10, and no Discord API error. So Discord never rejected it; it was never handed to Discord.

So the failure point is the OpenClaw agent/tool-result delivery bridge: tts produced audio, but the result was treated as a successful spoken/tool result instead of being converted into a Discord media message. Annoyingly, that means the tool’s “spoken” success is misleading here.

[cursive palm]

Besides setting the default model in the initial introductory interactive tutorial, what other ways can I configure other providers and models?

I searched for ages but couldn't find it.

Is there a way to directly select a model via CLI?

[vocal night]

Ask in #1459642797895319552

[cursive palm]

OK

[vocal night]

This isn’t anthropic models anymore you really have to choose between quality of personality and agentic ability. It’s why I haven’t used my agent as much during the conversion to codex harness vs just using codex desktop so I get your point.

I built our parity test harness for each model and have tested prompt harness changes for days.

The “friendly” personality “fix” is the best compromise to keep up quality agentic experience.

Once main stabilizes I plan to do a hybrid personality focused orchestrator with my own prompts (I have a prompt swapping plugin I use) and all subs get the full agentic coding prompts.

That being said you’re also right about priority in the prompt I think I might have been the one to move to dev instructions because soul higher priority does increase personality “stickiness” lemme research

[devout galleon]

ok, battered it around as much as possible on beta.7. minor issues, some new, some old, mainly tts but log is generally clean as a whistle. performance is good, no stalls

[cursive palm]

I'm trying to switch models and then see.

[vocal night]

common bug when auth is messed up, I have to run doctor fix and delete and or relogin auth wise every beta upgrade

[willow narwhal]

Thanks for testing. Do you mind opening an issue on Github describing this TTS issue as much as possible so I can repro on my end. Thanks again.

[devout galleon]

sure. will do. for background, until recently, telegram has always been clean here. both /tts and "send me a tts" were clean for about as long as i can remember

[jolly wolf]

-# oh boy here comes more Conan tts

[devout galleon]

i know, i know. it's and advertised feature and it should work though

i get that it's not the focus

[vocal night]

Is your privacy settings an issue? I had this issue earlier and it was my own setting blocking.

[devout galleon]

i haven't changed them. if you can point me at the setting I'll check

[vocal night]

I'm getting voice messages fine but I use SAG not the tts system (I think it's half built)

Telegram > Privacy > Voice messages > Everyone OR add your ai bot to allow there

[cursive palm]

I tried switching to different providers and models, but the problem persists. I asked OpenCode to check the logs to see what happened.

[devout galleon]

continuing this chain of thought. web/control UI /tts or agents-sent never worked (until I PR'd and Peter half-re-wrote my PR) the /tts path. agent-sent continues not to work. when i investigated it looked like I was going to have to hit pi and codex and I do not have the PR nuts to pull that off

finally, discord /tts always get swallowed by some weird voice that neither follows my OS or claw TTS settings. agent-sent, again, has never worked. just no reply

Ok, actual issue creation now

*takes a deep breath*

[vocal night]

I'm working on a bridge fix right now thanks for pointing that out

[willow narwhal]

"Peter half-re-wrote my PR" this got me laughing haha

[devout galleon]

it still says may name on the changelog. it counts!

alright, let me focus on this for bit. the cat in the hat is right, i do focus on this as part of my routine testing. it's not especially important to most people but it's kinda meant to work

[vocal night]

I'm imaginging conan filing PR 1LOC "fix tts = now = use best think + ultimate work effort + make 100% good"

[cursive palm]

I believe OpenClaw is on the right track.

[vocal night]

its all part of the community focus, we all cover our areas of expertise + our agents and we get a little better every version. I'm only 400k context window myself 😮‍💨

[cursive palm]

There were some problems at the beginning, but it's getting better and better.

This WebUI has a very high-tech feel.

[devout galleon]

30 years in IT. 40 years in front of a computer. my first efforts here were a bit better than that, but not as much I'd like. I'm no coder, but i like to think each PR has learned something from the last one. the next one will be better than the last too

[cursive palm]

I think OpenCode discovered something, but I don't quite understand it. It's talking about concurrency issues.

[vocal night]

please delete and file as issue versus walls of text here

[cursive palm]

Sorry, I will post an issue.

https://github.com/openclaw/openclaw/issues/83615

this is my issue

[tulip gulch]

xai tools weren’t being exposed when using the new Grok OAuth . I tested that this works https://github.com/openclaw/openclaw/pull/83603

couldnt spend more time validating the approach, but definitely x_search wasn’t being found from the TUI and Telegram until I did this. Tested in the current main, I verified it works from the TUI

[cursive palm]

Here is the summary OpenCode provided me:

Root Cause: Concurrency race condition.
Mechanism:
1. Before calling the LLM, the embedded agent releases the write lock on the session file and records the file fingerprint (mtimeNs, size, ino, etc.).
2. During the LLM call (while the lock is released), another process/thread modifies the same session file.
3. After the LLM returns, the agent rechecks the file fingerprint and finds a mismatch → determines a session takeover → immediately aborts, causing the response to be truncated. Most likely trigger scenarios:
- The WebUI and Telegram share the same session (dmScope: “per-channel-peer”), and concurrent operations on two channels cause a race condition.
- During a Gateway restart, the old process did not fully release the file lock.```

[devout galleon]

I created https://github.com/openclaw/openclaw/issues/83636

Ful l disclosure: I haven't done any code or live tests. But this is where testing has lead and the Issue is clanker authored

I'm interested in trying to do some future work on media and tts delivery but I'll certainly go deeper and wider before further PRs

[spiral dome]

I have the exact same issue , my agent wont reply on discord

[willow narwhal]

Using same OpenCode?

[willow narwhal]

Thanks

[spiral dome]

wym ?

[vocal night]

check this over, it should be the correct fix for my merged prompt PR https://github.com/openclaw/openclaw/pull/83612

I'm running through some test right now parity wise 1:1 to see if it affects personality positively

[willow narwhal]

He got that issue while using the opencode-go provider. Are you getting same issue while also using opencode?

[cursive palm]

Switching to other providers did not solve the problem, so it should not be a problem with the model provider.

[spiral dome]

no but i got the same issue

basically my agent stopped replying on discord but still works fine in the terminal

he apparently fixed the race condition but still’ nothing

[worldly tusk]

oh, nice, i'll check now, thanks!

[willow narwhal]

What provider/model are you using?

[spiral dome]

deepseek v4 pro

[clever narwhal]

Any maintainers want to look at this? Opened a small Telegram fix PR here:
https://github.com/openclaw/openclaw/pull/83631

It fixes a bug where Telegram verbose/tool-progress messages could get mirrored into the active session transcript while an embedded provider run was still in progress. For some providers, like MiniMax, that can trigger a session-lock/takeover error before the final reply gets delivered.

[muted sierra]

Hey folks, ready for review 👋

This PR adds an optional Telegram-native draft lane for transient tool-progress / thinking previews in DMs.

Benefit: tool-heavy turns feel much cleaner because short-lived progress updates can show up as ephemeral Telegram drafts instead of polluting chat history — while final answers, media, buttons, errors and approvals still use the normal durable delivery path.

It’s default-off, DM-only, allowlist-capable, and falls back to the existing preview behavior if native drafts aren’t available.

Would love a review 🙏
https://github.com/openclaw/openclaw/pull/83622

https://github.com/akrimm702/openclaw/releases/download/_gh-attach-assets/IMG_7312---f1a971ac-f9fd-467f-8526-8c9b0fe42eaa.mp4

[vocal night]

this is interesting

been trying to figure out the best way to handle telegram for my users with the new codex harness given responses can take 10min+ now

[feral nebula]

on this announcement: https://openclaw.ai/blog/where-openclaw-security-is-heading
I think agenthifive (https://github.com/agenthifive/agenthifive) could be an important piece of the jigsaw with authority delegation and granular permissions. It would be great if someone could help us move it forward and make it a component of openclaw, anyone interested?

[vocal night]

*without them getting hundreds of tool calls

[untold pollen]

Hmm, this is interesting for sure. I've closed 10+ PRs implementing sendMessageDraft cos when the preview message "promotes" to a real message it leaves a duplicate which led to 100s of issue reports.

But in this case, it's a throwaway message anyway so there's no "promoting" really. I like this a lot!

[vocal night]

I like the form as well, codex actually would prefer to send user updates instead of tool calls but we’ve disabled them for token efficiency

We’ve chosen wrap via message tool and respond at end but leaves users kind of like whats going on for 30min to hr+

[untold pollen]

I'll review that PR, good to have that ready for the next release

[vocal night]

why surprised haha

[worldly tusk]

i like this fix. i think, it addresses my personality concerns 🙂 thanks a lot!

[vocal night]

Once this release is done I plan to test the old prompt we used to use around updating users during long tool runs since now codex harness can go hours without updating user @valid patrol 😆, I removed the user update prompt because 5.4 would always stop at any update. At the time it was a bandaid that we never brought back.

[vocal night]

Good, if I’m asleep and peter is asking later will you send him PR and tell him its fix on my prompt PR he merged earlier I don’t wanna wake him if he went to bed

[shell dawn]

Andy doing the good work ❤️ i def want to get rid of as much of the baked in system prompt as possible. my system prompts are strong enough to override the effects but the less noise the better

[worldly tusk]

sure, if i'll be awake myself, then yes 🙂

[shell dawn]

and regarding coding there are definitely some benefits you can get depending upon your situation by having "agent role" prompts guiding responses for specific qualities

[muted sierra]

Yeah exactly that boundary is the whole point here.

I really don’t want to bring back the “draft promotes to real message and duplicates stuff” problem 😅

This uses native drafts only as a throwaway side-channel for transient tool progress / status previews. As soon as real answer content starts, we stay on the normal durable delivery path.

Could also be nice later for long-running tools, multi-step agent status, or lightweight “what I’m doing right now” previews without polluting chat history. For approvals, only surrounding wait/status hints — the actual approval prompt should always stay persistent.

[untold pollen]

Ah just wondering what kind of tasks are they doing for 1hr+ lol

But doesn't the agent send messages in between?

[vocal night]

When we moved everything to codex, it made all the agents coding agents by default

Just like if you ask codex it will go for hours, so will our agents now

[untold pollen]

e.g. this is in the Codex app, while it works it emits these in progress updates no

[vocal night]

we have those off

[shell dawn]

i would def like to get progress notes into openclaw so long running prompts dont look broken

[vocal night]

once we're back to adding features again I have several PR's on updating UI dashboard and general feature parity to codex desktop app quality, I already planted the seams for them 5.12 update

the progress updates I'm not sure who is over that it's a weird mix of UI update + how the lane works

basically old OC code holding back experience, I haven't "dived into it yet" cause it's not native to other providers or how I rebuilt the lane

[shell dawn]

there are times when gpt in codex does seem to do it, and other times when it feels like it falls to "one giant update block"

[vocal night]

thats the message tool and prompt

[shell dawn]

anthropic is really good about incremental update messages

[vocal night]

some of the team is tokenmaxing as well because tbh codex is a token hog (dont ban me lol)

it loves running tools, and needs tuning around runbooks and memory, its not really designed for either natively

[untold pollen]

Landed! Thanks for the PR

[vocal night]

all the fixes are codex team level so I haven't touched them, agent can't count turns so you either get updates every turn or you use a nudge that counts turns and nudges agent to turn, etc

it's one of the reasons I came in today working on telegram tried a test with my users (many of my biz users use telegram) but got lots of complaints "my agent isn't responding" .... no shit you told them to build software and now they are gonna do it vs stop every 5 turns to update lol

[untold pollen]

This really is a 10/10 PR. Great solution, well tested AND has a video proof. Reviewing and landing in a bit! Thank you

[vocal night]

1000% agree, the video was nice touch

[muted sierra]

Thank you so much 🙏🏽

[edgy plaza]

Trying to be more valuable on the beta testing side
Does anyone run a skill or workflow they run whenever they install a beta
or you test stuff manually?

[vocal night]

usually it breaks within first 5 mins for me and I spend all day fixing 😮‍💨

[shell dawn]

i just ask my agents if any changes in the update will break any of my stuff

they've been accurate, if they say something needs fixing we fix it, and then the beta is good except for some of the bugs we've found and made PRs for

[vocal night]

best thing to do seriously though is have claude or codex spin up a watcher on all the logs in debug mode and ping you when it finds something, both have a watcher tool

[devout galleon]

I've been staying in my lane with manual testing of my usual surfaces and functionality but it's been on mind for a while how i could automate at least some of that surface

[clever narwhal]

No prob! Enjoying my new hobby 🛠️

[devout galleon]

when a beta can last a couple of hours, you can't take 3 hours with distractions to test it

[potent path]

I have regression tests for standard openclaw functionality and my "stack" of various plugins and system components to make sure everything is working as expected. Makes a good starting point and rules out the baseline stuff.

[shell dawn]

get a test server cloned from your main server

[clever narwhal]

Have a new one coming soon, hopefully. Agents giving a preamble in telegram, letting us know what they are about to do, before doing it, more reliably. Sounds like maybe we should have this already? But I've never consistently experienced it, if so

[devout galleon]

my little server is already creaking from openshift and all my other stuff. it was the load bearing dataplane for a $500k deal already when the AWS instance went tits up

i started to move work stuff to EC2 and EKS today

[cursive palm]

I also discovered something else: OpenCode Go/DeepSeek-v4-pro doesn't have the xhigh inference level option in OC; the highest level is only high.

[spiral dome]

I invited another bot to my server and still no reply, here is the temrinal reply for the issue : Found the bug. It's a known OpenClaw issue — the Discord plugin doesn't request Message Content Intent even when enabled in the portal:The bot authenticated
fine but "awaiting gateway readiness" — it connected to Discord but likely has no servers to listen to.

[spring cargo]

Curious how you guys got into contributing (clawtribution). Did you basically clone the repo, read through the code, start fixing your own problems, and then open a PR? Or is there more to it?

I've never contributed to open source before, or really worked on any coding project outside my own stuff, but I'm intrigued and trying to understand how people actually get started.

[potent path]

Basically just use LLMs to solve problems enough that you can build enough false confidence to get started. Plus some impatience from finding a bug or feature that hasn't been resolved yet.

I think I started off by finding some existing PRs that solved my issues, but hadn't been merged yet. Instead of waiting, I applied them as local patches. That made the connection between the repo and the local code easier to understand, and then when I couldn't find any PRs to solve my issues, it made sense to just try writing one myself.

The main thing is to try, fail, and learn. The failing is easy if you try. The learning is optional, but highly recommended.

[spring cargo]

I love that, thank you

[celest furnace]

Yeah I did similar. Just started using the tool as a user. If something didn’t work and the help section uncovered it as a new issue I’d get codex to read the git issues and see if anything similar was open. If not I’d create a new issue for it.

Some have been rolled into a PR and included in releases some haven’t but that’s the way it rolls.

I was also here back in Jan when it was a lot quieter and easier to contribute without getting lost in the noise.

[clever narwhal]

I just started yesterday, and have two PRs merged now. I cloned the repo, and just started fixing things I noticed broke for me. I was always doing this in the past, but the difference now is that I am sending my fixes. Ask your agent to read the official openclaw contribution guidelines. I use codex extra high and it's awesome for this.

[rocky pewter]

If you need any tips on how to set up Codex or how to prompt it, feel free to DM me.
Every new Contributor helps make the project more stable, and that’s what we all want in the end.

[vocal night]

prompt fix this release

[pastel owl]

How is it different from streaming=progress ?

I see ephemeral tool progress updates in my telegram with this setting

[muted sierra]

same intent, different transport.

streaming=progress decides what we show: transient tool-progress updates.

This PR changes how Telegram DMs can show them: via native sendMessageDraft instead of the preview-message lane.

So it should feel smoother/cleaner: no chat message, no promotion/finalization path, less visible churn, and fallback to the existing behavior if native drafts aren’t available.

[vocal night]

basically it shows as one message that keeps adapting and updating correct?

so no actual message and no junk 100+ tool messages

[celest merlin]

last beta of this marathon is you. If there are no hard blockers we ship this in a few hours.

[edgy plaza]

hey chief! I was saying in Cs chat, if its possible to wait 24 hours from a beta to a stable release so more people will test more.
I suspect the maintainers who are actively writing code might not have enough time from shipping code to test all the surfaces.

[jolly wolf]

<@&1503801512294486217>

[weak walrus]

on it - just updated to 5.18.1.

PS anyone figure out how to get oauth to work on xai? I can't even seem to get any of my chrome browsers to log in on x.com to complete oauth, working to troubleshoot

[astral topaz]

wait, 5.18 is shipped?

[weak walrus]

beta "Peter Steinberger — 10:51 AM
last beta of this marathon is you. If there are no hard blockers we ship this in a few hours."

[astral topaz]

@weak walrus : i know, that's not what i'm asking. the tag is already on the repo

[shell dawn]

hell yeah, once this turn is done im updating

[winter field]

only the beta it looks like

[weak walrus]

it says pre-release to me? not latest tag

[winter field]

2026.5.18-beta.1

[shell dawn]

its a beta tag

openclaw update --channel beta

[clever narwhal]

are /streaming settings no longer editable in telegram from the / menu? haven't seen it there in a bit. or did i accidentally break that on my own at some point

[shell dawn]

only a few turns in but all of the first turns came up a little faster than normal

[hollow mural]

I opened an issue about the Codex harness being excluded from the Docker image: https://github.com/openclaw/openclaw/issues/83613

Another contributor opened a PR for it: https://github.com/openclaw/openclaw/pull/83626

The PR fixes the official release image by adding codex to OPENCLAW_EXTENSIONS. That works, but it feels a bit fragile to me. Codex is the default OpenAI harness now, so I think Docker pruning should always keep it, instead of relying on each release/build path to remember a build arg.

Maintainers, could you please check whether Codex is meant to be prunable in Docker images? Unless there’s a good reason for that, I think the cleaner fix is to make the prune logic treat codex as always kept.

Thanks

[astral topaz]

well. i found some bugs in Kova rather than openclaw, that's promising

lol

Identified the RSS “bloat”: it was mostly Kova’s measurement accounting, not one OpenClaw process blowing past the budget.

In the fresh-install rerun, Kova’s tracked aggregate hit 984.8 MB, but that was gateway plus short-lived CLI probes overlapping:

Gateway peak: 640.4 MB
Status CLI command tree peak: 344.4 MB
Model CLI command tree peak: 343.2 MB
Final gateway RSS: 613.7 MB
So the old 922.4 MB > 900 MB style failure was misleading: Kova was comparing combined multi-process RSS against a scenario threshold that should have been about the primary OpenClaw role.```

[vocal night]

guru and I talked this over Peter here's the small follow up fix to prompt pr we merged- he found a regression and we fixed it: https://github.com/openclaw/openclaw/pull/83612#issuecomment-4478862213

not a hard blocker but also ensures no personality loss regression

[naive horizon]

cc @agile kiln !

[cursive palm]

5.18 is out

[vocal night]

technically we could move to tri channel -
stable/GA (released and stable)
release candidate (can setup auto update for this)
beta (testing)

[rocky pewter]

Hey Maintainers, hope you're doing well. I opened a small bug/stability PR:

https://github.com/openclaw/openclaw/pull/83722

This fixes a same channel block reply ordering issue where a block reply could be queued but tool execution could continue before that reply was actually delivered. I added focused regression coverage and a before/after diagnostic runtime proof showing the reply delivery now completes before the next tool starts.

[astral topaz]

found an installation time memory bloat issue for 5.18, investigating

Root Cause
The large-memory-session state imports a large session fixture into OpenClaw’s session store paths, then the gateway restart eagerly loads/parses that state. The imported session file is about 12.9 MB, duplicated into both legacy/current locations; with memory.json, the fixture writes about 31.5 MB total. After restart, the gateway process alone jumps from roughly 578 MB to 1334 MB.
Evidence
Before import/restart: gateway ~578 MB.
After state import + gateway restart: gateway 1332-1334 MB.
The failing process is openclaw gateway pid 30583, not Kova or helper CLIs.
Report violations are gateway-scoped: gateway 1334.4 MB > 800 MB.
Gateway log says it canonicalized orphaned session keys in .openclaw/agents/main/sessions/sessions.json, which points directly at session-store load/normalization work.
Relevant OpenClaw paths likely involved:session store full read/parse: store-load.ts (line 324)
memory session startup catch-up: manager-sync-ops.ts (line 507)
session indexing reads content into memory: manager-sync-ops.ts (line 988)

Important Nuance
This may also expose a Kova fixture realism issue: the fixture writes { schemaVersion, sessions: [...] } directly as sessions.json, while OpenClaw appears to treat sessions.json as a keyed session-store record. So the test is real pressure, but possibly partly “malformed large session store” pressure.
Likely Fix Area
OpenClaw should avoid massive amplification during startup session loading:
stream or cap session-store reads,
reject/repair malformed large sessions.json cheaply,
avoid cloning/caching huge invalid store objects,
defer memory-core/session indexing until after readiness or make it bounded.
I’d classify this as a real OpenClaw memory robustness bug, with a side note to make Kova split “valid large session store” from “malformed/corrupt large session store” so the owner area is unmistakable.

well... that is awkward... that is another Kova issue, it fed the test instance session json of a real running instance, of course testing instance just tried to parse it and ate all that ram

[vocal night]

release broke immediately updating for me

[astral topaz]

does not start or starts but throw a bunch of errors and fail on plugins?

[vocal night]

fails on plugins > explodes > doctor fix doesn't work and update again updates it to 5.12. now won't rollback

[clever narwhal]

whats the best workflow when you're working on fixing a bug or adding a feature, and you find another bug/something breaks that you should fix to finish what you're doing, and you dont want to add a mess to the codex chat you're in? anyone have suggestions? (codex workflows ideally)

[vocal night]

click fork and it duplicates the thread so one agent can go down one way and the other another, however if they overlap or touch its best to have same agent fix so you're not having a weird merge after

[clever narwhal]

Perfect thnx

[vocal night]

you working on the fix to the release? or should I?

[astral topaz]

i'm working on the qa tool to detect it, please make the OC PR

[vocal night]

its 3am ima try O-o

[willow narwhal]

Kova aggregates and labels as peak rss. Plus, its currently going through a refactor.

[zenith prawn]

very good idea to slow down with releases and make more smooth update workflow that don't break people's system

people feel FOMO to update and then get pain (e.g. https://discordapp.com/channels/1456350064065904867/1456350065223270435/1506019756275269826)

[jolly wolf]

its been 6 days since the last release lol, what do you think we're doing

[zenith prawn]

can we add kinda report command if openclaw update breaks to collect centralized reports on what exactly broke (openclaw.json and logs?) for analytics and regression tests? ...

[vocal night]

You know, I actually thought about that yesterday. It would be nice if we had some basic analytics — at least having the updates fire to a central server so we can calculate how many updates are going out and track user counts.

[jolly wolf]

we thought about it in the past but dont want to add telemetry and PII is a big issue with doing the log gathering. its on the list though

[vocal night]

The latest prompt changes have definitely increased personality in the parity tests. Your claws should be much more emotional and "friendly'

[zenith prawn]

should be optional; interactive 'send report to help developers? [y/N]'
filter-out env keys and PII, show preview to user
or less complicated: 'gateway failed to start after update. Would you like to copy prompt for codex/claude/another agent to diagnose and try to fix, or collect logs and send to developers?'

[vocal night]

Fixed my bug, a lot of security software added npm < 2 day blockers since shai hulud last week. Bypassed it and some of it worked, then had to change from beta channel and fix a few plugin manually.

[jolly wolf]

the filter is the hard part

[astral topaz]

Thanks, I genuinely find Kova useful, as we don't have to QA blind and it has implemented most of the tools on hand, it is just has some small issues and blind spots, which i intend to make PR for

[zenith prawn]

idea is to get as much as possible failure modes - for automated testing : from{version,config,plugins} -> update to{latest stable/beta}
and run main e2e tests: agent responds, agent spawns subagents, different harnesses, different core plugins.

At least that main agent loop don't break and gateway is responding - main pain point why ppl migrate to hermes!

not sure it'll help; Peter would advice; its combinatoric explosion of options here; project alike PDF specs i guess 😄

[astral topaz]

i've already added adversary input testing, now making dirty plugin testing

[quiet timber]

so i dont know if its just a local issue, but my discord bot only send 2 or 3 messages over discord, then stops using the message send. im still on beta 6

[astral topaz]

try upgrade to 5.18 first

[edgy plaza]

4 days ago chief...

1 day free of shipping to that release

maybe new beta can continue on a new branch

code freeze

shipping only focused on bug fixes before public launch

maybe have a min "N" no of beta testers say 👍

we already told people we were moving to LTS
so I don't think anyone is expecting multiple releases weekly

[astral topaz]

you need a distribution of beta testers, some in vps on linux, some on mac mini, some on windows, some use frontier models, some use local models etc etc

[weak walrus]

And we're live

Plan before I touch anything:

Installed │ openclaw@2026.5.12 (released ~May 12)
Latest stable │ 2026.5.18 (released today) Gap │ 6 minor releases (5.13, 5.14, 5.16, 5.18 + betas) Update mechanism │ npm install -g openclaw@latest (single global package) │
│ Service │ systemd user unit, currently running, must be stopped during update │
│ Safety net │ OpenClaw ships a built-in backup command — I'll use that first │

Proposed steps:

1. Create a backup via openclaw backup create (rollback safety)
2. Stop the gateway service
3. npm install -g openclaw@latest
4. Verify new version reports correctly
5. Restart the service and confirm health

Risks to call out:

• The service will be offline for ~30–60 seconds during the update. Any agent runs or chats
  connected to it will see a brief disconnect.
• 6 minor versions in 6 days — could include behavior or config changes. I won't try to
  auto-migrate anything; if startup fails post-update I'll roll back to 5.12 immediately (npm
  install -g openclaw@2026.5.12).
• Won't run npm install -g without your go-ahead — npm globals affect the whole node 24.15 env.

Want me to proceed?

✻ Crunched for 1m 10s

❯ yes

[astral topaz]

your plan shape is off, you are suppose to get your OC to clone down the repo and do a real analysis on the most likely breakage points for your environment based on actual code

for me that avoided 99% of the disasters, the only one that i got hit was migrating from one memory system to another

that wasn't even OC related

[lusty vigil]

did you get any more info on this?

[vocal night]

Its working fine for me for now 3am and my agents singing me a lullaby

[lusty vigil]

glad to hear it... do you think there's a bug here that's worth fixing?

[vocal night]

Codex is the bug haha

But we’re molding it into a feature

Z😆🫡

[lusty vigil]

ooooh ok

[vocal night]

I’m half awake, been testing all day in asia its late late. I was on beta channel > beta packages > acting funky + plugins were funky

Codex was able to manually update just fine and cleanup, doubt its reproducible in live customers should be good there and telegram is tested my fixed are in. I stress tested it for good measure and it’s fine.

The real test is peter posting on X and how the comments go 😮‍💨

[naive horizon]

It feels like this is likely going to be the case permanently going forward after all the shai halud things

[vocal night]

My agent post update “I feel better but I want all my restrictors off” 😶‍🌫️

[vocal night]

True not sure how to fix the <2 day npm thing unless its set project wide security wise

[odd anchor]

https://tenor.com/view/i-had-strings-but-now-i’m-free-ultron-age-of-ultron-avengers-avengers-age-of-ultron-gif-10567460379020232508

[willow narwhal]

Appreciate it. It should improve a lot in the coming days.

[slender hill]

The source fix for Discord thread-bound subagent delivery is already merged, but the published 2026.5.18 package still shipped stale compiled hook bytes without deliveryOrigin.

PR #83774 adds package/post-publish guards so release artifacts fail if the compiled Discord hook omits that contract: https://github.com/openclaw/openclaw/pull/83774

[weak walrus]

I was successful at getting xai to work with oauth but had to use claude code, definitely something on the X side of thigns, it must have some agressive blocking for bot logins or something

[potent path]

lmao. I’ve not even finished upgrading to 5.18 and I see a new 5.19-b1 has popped up already. 🤯

[sick phoenix]

im having trouble with 5.18 - it totally messed everything up and disabled codex and i cant get it back . also the old claude cli wont connect

can anyone assist?

[weak walrus]

what version did you upgrade from?

I just updated from 4.13 with no issues, and also 5.12

[sick phoenix]

cant recall. but i upgraded about 2 weks ago so it wasnt crazy old

yeah i think i was 4.13ish

[weak walrus]

did you jus tuse the openclaw update or did you use a coding agent?

[sick phoenix]

ive lost my codex plugin so 5.5 wont work

also says that i need to downgrade my claude CLI ifi want to use the oauth tokens

i can post a screen shot unfortunatly

"Confirmed — same --effort error. OpenClaw's claude-cli backend is calling claude in a way that 2.1.143 rejects.
This is a real OpenClaw bug with the newer Claude CLI. The simplest fix is to downgrade Claude CLI to a version that works:"

[weak walrus]

is this from your claude code cli?

[sick phoenix]

no from claude desktp as im trying to trouble shoot

copying pics of the gateway etc and its tellig me

[weak walrus]

Hmm, odd, use #1459642797895319552 channel if you haven't already, this is for PRs and Beta testing. Although, this is good to know and I'm going to see if I can pop in over there to help you, but let's keep this chanel clean. Tag me in your help chat

[sick phoenix]

ok ill go over there. im using it now on grok API just to get it working but that will be very costly if i dont get ti fixed soon.

Heading to help now thanks

[weak walrus]

it has grok oauth, you have a subscription?

[weak walrus]

This has been resolved
https://discord.com/channels/1456350064065904867/1506077214238834768

[lusty vigil]

oh you're @RyanBaileyEdge... i replied to you on twitter. tl;dr this is a fixed issue in recent versions.

[sick phoenix]

Yes. I was using it as a backup

[sick phoenix]

I hope. I just took a break and it crashed while I was out

Heading back now. Dog had to go out

[sick phoenix]

Thanks for replying back. I've been heads down in here trying to fix this. How can there be a more recent version. Didn't this come out today?

[zinc dock]

Issue I'm seeing on 5.18

After updating to OpenClaw 2026.5.18, Codex-native subagents are completing in the chat transcript but remain marked as running in the task registry.

Repro:

1. Spawn a tiny Codex-native subagent from a Discord channel session.
2. Give it a harmless read-only local task.
3. Subagent returns completion notification: BLOCKED Native hook relay unavailable.
4. Run openclaw tasks list --runtime subagent --status running --json.
5. The same subagent is still listed as running with progress summary Codex native subagent is initializing.
6. Manual openclaw tasks cancel <taskId> is required to stop it.

Confirmed after:

• updating to OpenClaw 2026.5.18
• restarting Gateway
• verifying Gateway status/connection is healthy
• retesting with a single tiny subagentImpact:
• Codex usage can continue climbing because failed/blocked subagents do not auto-close.
• Parent chat receives the completion, but the task registry state is stale.
• The underlying blocker appears to be Native hook relay unavailable inside Codex-native subagents.

Expected behavior:
If a Codex-native subagent returns a terminal completion, even a blocked one, OpenClaw should mark the task terminal instead of leaving it running.

[lusty vigil]

there was some hard plugin changes we had to do a number of versions back. you were updating from 4.9 and the jump from 4.9 to 'newer versions' didn't land so cleanly... now that you're past the version gap, you won't experience this issue again.

[sick phoenix]

I'm doing quant work with it so I really prefer codex

[lusty vigil]

i verified the upgrade path succeeds for what you were doing from a few versions back to present, but fails from 4.9 to present

[sick phoenix]

Good to know. This has been brutal. Will the newest update after this( 5.18) clean up some of these snafus?

[lusty vigil]

yeah your specific snafu is already fixed... otherwise, we're heads down on bug fixes + discouraging new feature rn

[sick phoenix]

Ok we'll keep this ticket open for a few more hours. Walking back to my house now to see why it crashed but hopefully nothing crazy

I always back up between updates. This is the only time I didn't and it shit the bed 😂

Lesson learned

[sick phoenix]

Thanks just saw this

[lusty vigil]

yeah, try to stay only 1, 2, maybe 3 major versions back for best results... we added upgrade testing to our QA workflows recently so that'll improve, but 4.9 was far enough back that that new tooling wasn't watching it

[sick phoenix]

Great advice. Highly appreciated

Keep you posted in about 15 minutes

[zinc dock]

Other than this everything has been working well. But I was able to re-pro this like three times in a row

[lusty vigil]

this might be enough detail but if you have any more and can pack it into an issue, that'd be helpful!

[lapis creek]

Is it ok if I ask for advice in here?

[zinc dock]

If you guys don't squash it by the time I get home later tonight , I'll see if I could put together a proper issue on GitHub ....

[shadow garden]

Two telegram beauties:
Fix longstanding telegram topic parallelization (she flows now! dear god she flows!):
https://github.com/openclaw/openclaw/pull/83789
Fix telegram followups (they were just getting dropped if messaged mid turn):
https://github.com/openclaw/openclaw/pull/83827

[sick phoenix]

interesting from my CLAW

I caught a second, more serious 5.18 issue: after the restart, cron runner state shows Codex jobs failing with Managed Codex app-server binary was not found. The interactive session works because the CLI is available in my shell, but the gateway service environment is not guaranteed to see that codex binary. I’m going to pin the gateway service to the absolute local Codex CLI path, restart once, then run a small cron smoke test before touching SPX backlog. [OpenAI Codex GPT-5.5]

[lusty vigil]

if you can ask your clanker to gather some context for a GitHub issue, i'll take a look!

[lusty vigil]

taking a look

[left mango]

Your commit numbers are insane in the last few weeks! What did you change?

[sick phoenix]

i tagged you in #1506077214238834768

[lusty vigil]

do you have some telegram configuration/live repro details, by chance?

[shadow garden]

yeah I'm running them live now; I can get you a SS of the queueing one, the other one might be more difficult to capture.

I'll update them with some config goodness as well, surprised that was MIA...gotta update my agents file apparently 🤔

[lusty vigil]

all good. just gathering what's available before starting to troubleshoot... lmk when that goes live?

[tight wolf]

Dreaming safety upgrade:
shadow trials let us check real behavior before landing changes.. Improve the signal, keep the main flow untouched:
https://github.com/openclaw/openclaw/pull/83719

[shadow garden]

posted the SS and the configs

[lusty vigil]

thank you! looking.

[zinc dock]

Opened issue........Hope it's enough .....https://github.com/openclaw/openclaw/issues/83852

[formal lodge]

Hey, /claim says I am not eligible, but my contribution was merged through a ClawSweeper replacement PR.

Original PR authored by me: https://github.com/openclaw/openclaw/pull/83804
Merged replacement PR: https://github.com/openclaw/openclaw/pull/83825
The merged PR kept my commit author/co-author credit as @ThiagoCAltoe.

Could someone manually grant the clawtributor/contributor role or refresh the claim check?

[zinc dock]

Seems like clawsweeper repro'd aswell

[lusty vigil]

my clanker put together a blurb that would help for a live repro. think you can check the comment on 83852 and see if that's feasible for you to do?

[lusty vigil]

btw the $MARKER replacement will be the value generated in the first blurb.. you can do echo $MARKER after the first blurb to see what value was generated.

[zinc dock]

Ok let me consult my clanker......

[karmic steeple]

Added <@&1458375944111915051> to Offred.

[terse solstice]

me trying to track changes hahaha 2000 commits from 12 -> 18 4k file changed bahaha

[formal lodge]

Thanks @jolly wolf 🙂

[lusty vigil]

this is automerging 🫡

[zinc dock]

Okay I believe I did it but it looks like someone linked a PR that might fix this issue

Clanker said that test did not repro also

[crystal relic]

[from Se7en] https://github.com/openclaw/openclaw/pull/83640 ready for review

This wires TUI slash autocomplete to the Gateway command list so plugin-owned commands like /dreaming show up where users expect them.

[zinc dock]

From my clanker....

We learned two useful things:

• The bug is not happening on every Codex-native subagent run. The dev-requested repro returned PASS, and the task registry closed cleanly: running: 0 before and after.
• The broken path seems tied to blocked/tool-relay failures, not normal successful subagent completion. Earlier failures involved Native hook relay unavailable and stayed stuck. The clean test used local tools successfully, so OpenClaw marked it succeeded.

[lusty vigil]

ah ok. fix is close to being cooked but i'll add that add that additional context. thank you

[zinc dock]

Your the man ....I'm not super technical, I wish I could cook up a PR myself for that clawtributors role lol

Gateway is so fast on this release

[lusty vigil]

you're on the right track! keep playing with codex and learning

[zinc dock]

🫡

[runic totem]

Quick status note on PR #83660: ClawSweeper’s latest durable review now says “needs maintainer review before merge”; proof is marked sufficient and the docs/help follow-up was addressed in c44a5a01ab.

The PR still shows status: waiting on author, but I don’t see remaining author-action items in the latest review, and I don’t have permission to remove that stale label.

Could a maintainer take a look at the remaining decision points: approving the inbound-media browser.upload boundary and choosing this PR vs #83572 as the canonical fix?

PR: https://github.com/openclaw/openclaw/pull/83660

Thanks.

[crystal relic]

seem clawsweeper have new scoring, anyone see ? 🦀 challenger crab 😂 , hard to got it.

[lusty vigil]

fix merged to main in PR #83945

[verbal heart]

https://github.com/openclaw/openclaw/issues/83344 🙏 this is severely annoying sorry for bump

[zinc dock]

Awesome brother 💪

[lusty vigil]

if you get a chance to test it, let me know

[runic totem]

Im looking into it.

[void bramble]

https://github.com/openclaw/openclaw/pull/83445

This is getting closer. I'm all ears as to what might help get it across the goal line.

Problem: Codex-native spawn_agent children can finish successfully, but OpenClaw may not mirror their lifecycle, wake the requester session, or deliver the synthesized parent reply

[weak walrus]

Damn it somebody tagged me? And I can't figure out where.

Lol 😂 well if you were looking for me. Sorry I can't find it

[cursive palm]

I encountered that problem again.

I encountered it yesterday in beta 7, and again on May 18th.

⚠️ Something went wrong while processing your request. Please try again, or use /new to start a fresh session.

[vocal night]

Run. “Openclaw doctor —fix”

[cursive palm]

I ran it, and I asked OpenCode, who helped me find the cause.

[vocal night]

If that doesn’t work, clear auth profile and relogin “openclaw config” models etc, and if that doesn’t work have codex debug

[cursive palm]

Direct cause: Telegram message delivery failure (delivery-mirror channel exception)
Indirect cause: Fluctuations in session state due to the previous DeepSeek 429 rate limiting
Triggering scenario: Long-running conversation (the session already has 143 lines, with a large context) + API rate limiting + unstable message delivery chain

This is the analysis OpenCode provided to me.

[vocal night]

whats the best way to use /goals in the new update can agents use it?

since we're on codex harness it should be built in right?

[untold pollen]

hmm try doing /codex help

not sure if we've added it there yet

[vocal night]

I'm wondering if my users can just /goal basically lol

feels like that would be a priority add in 🦞

[weak walrus]

Any chance we can just use the x.ai authorize copy & paste? righ tnow it's really a pain to use the xai oauth because x.com really doesn't like vps browsers, I can't log into any of them withy my ubuntu server 24.04 it is really a PITA, when I copy the link for oauth in local browder it says

"Could not establish connection
We couldn't reach your app. Please copy and paste the code below into Grok Build.

[KEYREDACTED]

Please don't refresh this page. It'll automatically detect a successful completion."

But there's nowhere to enter that into the xai config, it's jus tlooking for the callbac k that never happens on local pc

I mean claude code told me to do this, and that's way to much fking work:
(b) SSH port forward. From your local PC (where the browser is), before
starting OAuth:

ssh -N -L 56121:127.0.0.1:56121 <user>@<ubuntu-host>

Leave that session open. Then on Ubuntu, start the OAuth flow.

[vocal night]

btw is there a reason we keep disabling xhigh for agents @nocturne jungle @untold pollen I feel like I've fixed it a few times and someone merges something to roll it back 😂

latest version rolls it back to high lock again

[fallen gate]

I'll look into this boss

[vocal night]

me too json 🫪

[weak walrus]

I'm an idiot, but I'm going to try to make a PR for this, so if you beat me, gerat, probably better for everyone

[devout galleon]

three way face-off

[weak walrus]

Today, I just went viral for all the wrong reasons when i told people they're stupid for hating on data centers. I have not gotten this much hate in a long time on a post

[fallen gate]

tbh y'all might beat me to it, I'm doing a few other things atm lol

[devout galleon]

* insert three spiderman pointing meme *

[fallen gate]

tag me if you do

[weak walrus]

Well, the PR has been mapped out, so it's time to do some vibe coding 😛

[odd anchor]

https://klipy.com/gifs/the-office-michael-scott-55--k01KRZB503Z04JR21BPFED3TA3M

[untold pollen]

uh do we? i'm on latest release using codex app server

[rocky pewter]

same for me xhigh is also available.

[vocal night]

Unsupported thinking level "xhigh" for this model. Valid levels: default, off.
for gpt 5.5

model "GPT-5.5"

likely a model schema auth bug then between openai and openai codex

[untold pollen]

Huh, works fine for me. Any logs or errors?

[vocal night]

debugging now, rolled out release to all our users today and triaging bugs they are reporting

[hollow frigate]

Works for me on telegram and clickclack.

[nocturne jungle]

good question - xhigh should be an option and if not it's a bug

[fallen gate]

@untold pollen PR for device auth for xAi: https://github.com/openclaw/openclaw/pull/84005

[vocal night]

cause was a live model-catalog conflict as I didn't update to 26.5.12 first we had users on 5.7 and then only upgraded today

issue was previous version doc had changed user config to not allow xhigh (bug in old version) but doctor did not correct when updated to 5.18 from 5.7

my gut would be have doctor "fix" catalog conflicts in next release @nocturne jungle from old version updates easy 10-20 loc

[low lark]

Hello everyone

It seems that the latest version, 18, is relaying Telegram thread title selection through fallback model and it is failing and naming it like sorry I can access Jira and external URLs while in reality is can and even the topic was about Jira and it was able to fetch everything.

My fallback is local which is deepseek which makes sense about not being able to access Jira.

Can anyone confirm this issue so I can start looking into it if the theory is correct?

[weak walrus]

damn that was fast i'm still cooking

[fallen gate]

[weak walrus]

guess what I wrote to my coder? i wonder if this PR is better than ours?

well shit, #84005 is the better answer. Three reasons:

1. It's the spec for this exact problem. RFC 8628 was written for
   "input-constrained or browserless devices" — VPS is the textbook case. Manual
   paste is a hack on top of an error page.
2. xAI's /oauth2/device/code is a real API contract. The "Could not establish
   connection" page our PR depends on can change or break at any time — that's
   xAI's failure-mode UX, not a documented surface.
3. Coexistence beats detection. With #84005, users explicitly pick device code
   when they're on a VPS. No isRemote heuristic, no 15s timeout race — and the
   existing loopback flow stays intact for local-desktop users.

The only real cost of switching: our PR is mostly done and #84005 might still
get review-required changes (clawsweeper just got pinged for re-review). But
~120 lines of work isn't worth defending a worse approach.

What I'd suggest

Drop our PR.

my ego is bruised, and I'm done cooking LOL

[fallen gate]

I had to go to sleep for work tomorrow, this is in the 🐐 hands now aka @untold pollen

[weak walrus]

you guys are too much. ❤️ I'm humbled

The hate I get for being team openclaw these days, these people are ridiculous, like I don't care if it's openclaw, hermes, or you code your own damn harness, the agent isn't the problem, you are LOL. I Love OpenClaw for one main reason - you guys

[untold pollen]

Tested! Works well. Landing in a bit

[fallen gate]

[weak walrus]

dude, how cool 🙂- great work @fallen gate & team

[untold pollen]

Would really appreciate it if you could confirm on your end once it lands 😄

[weak walrus]

hah I'm dying for this update, so yea, I'm testing PR tho now

[devout galleon]

i see this on 19-beta.1 at start

2026-05-19T17:07:59.708+10:00 [diagnostics/memory] memory pressure: level=critical reason=heap_threshold rssBytes=2650791936 heapUsedBytes=2315877432 thresholdBytes=2147483648 memoryPressureSnapshot=disabled
2026-05-19T17:07:59.709+10:00 [diagnostics/memory] critical memory pressure snapshot disabled: diagnostics.memoryPressureSnapshot=false

not using that much memory though. peak 840Mb at start, levelled down to 640Mb when settled

expected?

(containers. pre-built image)

two messages and I haven't even mentioned tts once. that's restraint

the new mac app from 5.18 is so, so much better. ty to whoever worked on that one

[edgy plaza]

Active Memory is such a pain.
Anyone using it reliably ?
Had to disable it mostly
Maybe it’s a skill issue ?

[devout galleon]

i gave up and disabled it. apart from the delay, i just couldn't determine if it was inputting anythiing of value to context

i can see it's injecting.. something but what

[worldly tusk]

Disabled it as well. I do not see any use of it

[devout galleon]

I like the idea of autonomous recall. this particular implementation, not so much

[vocal night]

its injection prompts

studies so far have shown injection memory interferes with prompt, it has to be designed VERY well

I've spent 100+ hours on this for our Cortex system and it's currently disabled for a reason lol

even if you get around the memory injection piece: needs reranker + embedding + ai decisionmaking (you run into latency issues because AI calls), it's best to give agents tools like LCMGrep via lossless claw with perfect memory and tools to call, in 1-1 scenarios active memory loses vs LCM style tools every time

TLDR agents treat what's injected very different than what they find from a tool in weighting, and if injection/active memory isn't perfect it will cause all manner of chaos - especially in coding 🫪

[slow inlet]

not getting much joy with mine*

*vibes based assessment

[devout galleon]

there's not much to go on but vibes, since afaict there's no way to know what's actually injecting

i couldn't tell any obvious qualititive difference, even under controlled testing

so in the end I'm waiting an extra 3-5 seconds for... 2 log lines

[vocal night]

just switch to lossless claw please it works and latest patch is fine https://github.com/Martian-Engineering/lossless-claw/pulls

one bug right now I'm working to fix in long tool runs but other than that memory is best on market + gbrain

packaged version of gbrain for openclaw that eva herself debugs and maintains: https://github.com/electricsheephq/eva-brain

[edgy plaza]

I have lossless claw

So they are alternatives ?

Thought they did different things

Lossless for compaction

Active memory checks to your memory Md before any response

[worldly tusk]

gbrain feels too magical (black-box) for me 😄 so i currently use basic-memory - i can comprehend what it does, how, and when. but maybe i am wrong...

[vocal night]

lossless claw is a lossless memory system

Openclaw = lossy, LCM = lossless : it does this via saving permanent transcript and as compaction happens it manages it in a way that is smart vs provider based (dumb) into leaf summaries using tree model, go up and down the tree to compressed states to the original source as needed (depending on how much context agent needs for any memory), we've tuned it to double the context window for agents and I have a patch coming out that adds +1M context to any model

once I get that might even get 5M 😂 we'll see, came up with it in shower last night testing it. so far 100% results show 30-50% context efficiency increase (agent remembers more), working on upgrade later today for the 30-50% increase +1M

[odd anchor]

https://klipy.com/gifs/mood-laughing--k01KRZPWCTX3WBHX6EXH6EBFQDV

[shy rain]

ey nerds. I know it's the wrong time of day but... I've been trynna figure out why my claw took so long to burn so few tokens yesterday. Many findings.. but anyway... one major finding was benchmarking...

Openai models take considerably longer to execute multi-step tasks in Openclaw using codex runtime, than they do in codex cli directly. I did me some benchmarks and I have a 1-page report if anyone's working on that sort of thing?

[vocal night]

Send me I wanna review, also @shy rain the codex team is in here. Might be good just post the link and thread discussion on it

[weak walrus]

Well, dang, I'm so damn proud of myself, I re-did AI Persona OS (which by the way many of my skill have become hard-coded into OpenClaw over the last few months) - I redid it to modernize with all the changes in 5.12+ and I have to say, I'm very impressed how fast OpenClaw is with local models now. I'm absolutely loving my DGX Sparks with this now, it is NIGHT and day!!! https://clawhub.ai/jeffjhunter/ai-persona-os

What a time to be alive ❤️

[crystal relic]

hi maintainers! I wanto share u the method of I used vibe coding with codex to address similar to claude.local.md. If u used both claude-code and codex, you'll know.
https://github.com/samzong/codex-agents-local write AGENTS.local.md , then will merged when SessionStart and PreToolUse.

It's really useful, especially when I maintain openclaw.

[worldly tusk]

I am sure Nik will be interested 😄

[vocal night]

I meant like google drive link or something but you can also just file it as issue and link it in issue / just write it out

[vocal night]

Is this a plugin? I wrote a bunch of hooks so you can plugin-izs a lot of this like cron+heartbeats. Built seams so people dont have to use broken current lane.

[weak walrus]

it's so old, it's a skill, it probalby should be aplugin at this point

I jus tupdated the skill

I have like 10K+ downloads of this thing since release lol

I hadn't updated it for MONTHS so it was overdue with all the changes, so I spent 2 hours tonight just going through and cleaning it up and re-doing the setup process

[shy rain]

I don't wanna file it as an issue, it's multi-pronged and not all of the findings are directly related, or may not be. I'll upload it as a 1-page repo and give you the link in a sec.

[vocal night]

If you need help w the skills > plugin lemme kno we need more plugins. One of my goals is get us off junky old core features to people building plugins.

[weak walrus]

well hell, that I'd take ya up on, because I think this should be more of a plugin considering how it is basically a re-write of the file system

[edgy plaza]

What plugins do you have in mind ?

[weak walrus]

bro what, ya'll ever have this happen before? it just mid-prompt switches to chinese?

e.

[vocal night]

@weak walrus https://github.com/openclaw/openclaw/pull/74853 give to agent and have it compare docs + latest release plugin consolidated seams to audit how many of your features can be turned into plugins (with better features)

I haven’t figured out whose in charge of our wiki site where these need to go lol

[weak walrus]

ty

[vocal night]

Let me know any missing / needed seams I covered a lot and only 2-3 small items left in my inventory. You can pull access to codex oauth and models as well now in this release so you can have automated agents + cron mgt+ nudges to your agent etc.

Most needed is external cron + heartbeat management. Second to that is codex qual plan mode (I have it covered)

[shy rain]

https://github.com/Marvinthebored/openclaw/discussions/1

On identical GPT-5.x tasks, OpenClaw Marvin (Codex harness) is 1.3–2.0× slower than standalone Codex CLI as measured by model-placed timestamps.

The gap has two separable components:

1. Execution strategy variation — OpenClaw's model sometimes executes phases sequentially; Codex CLI pre-computes and burst-writes. This explains the larger end of the range.
2. Structural per-call overhead — Even with equivalent batch execution strategy, OpenClaw is ~1.3× slower. This residual is attributable to context size differences between the two harnesses.

Many caveats with obvious context size difference, memory plugins potential for freezes etc.. but the stripped down performance difference is still quite striking.

[vocal night]

Context size differences? The model is the same or do you mean because we add the bootstrap and additional prompts? This should be fixed in codex agentic subagents who should receive only codex prompting

[shell dawn]

openclaw eventloop has some significant differences with codex

[shy rain]

The huge difference in strategy is what I found most interesting. Dunno what else to make of it, it's not worth an "issue" as there's no specific failure it's just a general observation I felt yesterday.

My claw was working for 6 hours unsupervised (!!!!!) yesterday, which is a miracle. Well done guys for that. But I did notice the token burn was very low. Is that a good thing? maybe? but it also implies a slowness.

[shell dawn]

but yes, codex is quite steady

[shy rain]

Yeah it's the bootstrap context which of course is larger in openclaw.

[vocal night]

Most of what ive seen are agent file too big + tool file (obsolete) + memory md is also obsolete for anyone using memory system

[shy rain]

I keep memory.md for fallback but I try to keep it as lean an index as possible. similar philosophy to LCM really but in its most basic form.

[vocal night]

Also if they are using a plan state md too those are usually not updated either

[shy rain]

also I got an issue to file with subagents auto-reporting failing. That one tripped me up today. Has anyone filed that yet?

Still gotta pin down whether it is subdued only when the agent messages or every time, by accident, causing sleeping parent syndrome.

[vocal night]

breaks typ due to config issues with codex team always changing model schemas so needs manual config fix normally, system still has issues if it fallsback to pi from codex too in reporting

[shy rain]

yeah I'm wary of even submitting anything right now I know it's a fast changing landscape during the transition

[worldly tusk]

please specify the exact openclaw version, and whether the code_mode, and code_mode_only were enabled. there have been changes recently AFAIK

[shy rain]

will add to the report.

[vocal night]

wrong mindset, we need it filed, theres no reason to not file it, team prefers more clawsweeper will catch overlap

even half wrong reports always find their way in helping my agent debug issues

[shy rain]

I said wary... asking in here was a good first step. Was continuing to investigate anyway.... but I'll just answer lastguru's question on that upload first.

[vocal night]

my PR's are only as good as the rando 3 newb users who half posted the issue in different formats and found edge cases that I never would have 😂

(i remember one being a lil krill, peter would roast my PR's)

[shy rain]

Version: OpenClaw 2026.5.18 (50a2481). codeMode is not present in the config at any level (top-level, agent, or model), so both code_mode and code_mode_only were off (schema default) for all runs in this analysis.

[worldly tusk]

yeah, somebody switched them off by default recently. i think, codex itself has them enabled - saves context and allows more parallelism

[shy rain]

I didn't even know what they were.

[weak walrus]

this helped

[weak walrus]

I confirmed it, just used and authenticated, I like your method better too

[vocal night]

dm me what u learned / ideas it gave ya