clawtributors

[potent path]

I'll add that to my testing. But even if true, I think it doesn't necessarily help for model changeovers during non-interactive work. Or even interactive work, for that matter, since the UI usually doesn't tell you when the model falls back to another model in the chain. Right?

[nocturne jungle]

Sorry - could you explain this usecase more?

Like openclaw is constantly changing models while working?

[slow inlet]

But what is and isn't a third part app? If I wire up a python script to trigger claude -p is that an app?

What about if I wrap that in a telegram router? Is it an app yet? It's just dumb

I understand but making an exception to misanthropic just because they're being meanies makes me sad.

We should beat them into submission

When life gives you lemons make lemon grenades or whatever the Portal guy said

[potent path]

I wouldn't say constantly, but yes agents have defined fallback chains so that if the primary/default model is unavailable, it silently falls back to the first fallback, and then the second, etc, as needed. There are also non-fallback scenarios like task-based model routing, or health-based model routing, and probably other use cases.

And where you have a non-codex model as default, and then a codex model is substituted, the agent suddenly has different tools available to it from what it is expecting, and it can get confused and stop being productive. At least, I think that is what is happening. Like I said, it could use some independent testing so we have a bit more confidence in whether this is a real issue or not.

[slow inlet]

Silent fallback onto Sonnet - behaves like its lobotomoised - my favourite thing to double check

Why is the fallback SILENT? wouldn't an end user want to know their default isn't working?

[split elk]

agreed - this would be a major enhancement, to be proactively alerted when a fallback happens

[nocturne jungle]

No matter what, you're going to have problems switching between any non-OAI model and OAI models, as reasoning traces are incompatible.

WRT different tools, most of the dynamic tools are in fact shared. The only difference would be that we remove redundant OpenClaw dynamic tools that codex has natively. The agent may stumble a bit and course correct - 5.5 is smart enough to do so.

We could put in a small message into context when the switch happens so the agent is aware of the swap, but outside of that not much we can do.

[slow inlet]

Force a message in the channel too? Every time? with an option to suppress by channel / session? Idk just spitballing

[nocturne jungle]

I agree that fallbacks shouldn't happen silently

Though this sounds like a post-big-release optimization we can look at

[split elk]

related (ish), but one way to allow non-silent fallbacks is via a different approach:
-> optional toggle to display the active model in the footer of each model response

[potent path]

I was planning to use a combo of "codexDynamicToolsProfile": "openclaw-compat" + *"codexDynamicToolsExclude": ["update_plan", "apply_patch"] * that made almost all tools available to agents in both runtimes, but that workaround seems to have been disabled with PR #80048.

[slow inlet]

Yes I like this, message signatures or footers with key session info

[nocturne jungle]

yet more configs ha

[split elk]

everyone's favorite 🥳

[next helm]

Ran the beta8 update/audit across my three macOS OpenClaw installs. All three are now on 2026.5.12-beta.8 (097daf9) with gateways reachable. Two updated cleanly; one host was down because a stale submitted launchd updater job was repeatedly killing the gateway after it reached readiness. I filed the main new reliability issue here: https://github.com/openclaw/openclaw/issues/81859 and added beta8 follow-up evidence to https://github.com/openclaw/openclaw/issues/81394#issuecomment-4452641406 and https://github.com/openclaw/openclaw/issues/81499#issuecomment-4452643184.

[split elk]

(but i've spoken to a number of people who think this would be valuable.... have built something myself and wanted to push a PR but don't quite have the chops to figure out how to make it work universally)

[slow inlet]

doesnt the clawhub already do this in the web app sessions? wouldn't it just be outputting this to channels?

[nocturne jungle]

How urgent would you say this is? Like what's the vibe check on how often this situation would happen

[potent path]

I'm still gathering data, and again I'm fully prepared to be proven wrong. But the early analysis was:

Pi tools that have NO direct counterpart in Codex:

• bash → use exec_command (different name, similar shape)
• edit → use apply_patch (completely different paradigm)
• read, write, find, grep, ls → no direct tools — must use exec_command + shell commands (cat, grep, find, ls, etc.)

The cross-runtime confusion is real and significant for Pi-shape agents:

• Pi agents trained to call read("file") will find no such tool
• Pi agents trained to call edit("file", "OLD", "NEW") will find apply_patch which needs unified-diff syntax
• A capable model might adapt and use exec_command for everything, but that's degradation (more tokens, slower, more error-prone)
• A less capable model would just fail

[next helm]

Let me diagnose more significantly to understand if it’s instance specific

[split elk]

you do get this in gateway channels. it's already there for sure.

↑6 ↓14 R17.8k W10.6k 14% ctx claude-opus-4-7

[potent path]

Wow, that formatting is ass. Why would it arrange the bullets like that? 🤬

[nocturne jungle]

valid - switching TO codex would be okay (codex smart enough to figure out) - but then switching FROM codex to a dumber model would likely have some challenges until you run /new.

That being said, you could pin pi as a runtime in config if that's your setup

[slow inlet]

So just a case of appending it to channel outputs and making sure it doesnt conflict with any maxchar behaviour?

[potent path]

Oh! Pin Pi runtime for all models, including openai? How do you do that? I was trying to do that in one of the earlier betas and it seemed like runtime=pi wasn't an option anymore. I thought codex runtime was mandatory in the latest betas?

[next helm]

My read:

Openclaw 1's’s outage was probably not caused by the normal OpenClaw beta8 updater path alone. The smoking gun was a leftover launchd job named ai.openclaw.update.beta8 repeatedly running a temp script, likely /tmp/openclaw-beta8-update.zsh, and exiting 127. Exit 127 usually means “command not found” or “script/interpreter path no longer valid.” Because that job was still registered with launchd, it kept waking up roughly every 20-25 seconds. Each wake appears to have kicked/terminated the gateway, so Openclaw 1 would reach “gateway ready” and then immediately get SIGTERM.

The OpenClaw bug is that gateway status/doctor did not notice the sibling updater job that was killing the gateway, and the service refresh path made recovery messier because the generated env-wrapper hit macOS launchd/security friction. But the thing that actually kept Openclaw 1 down was the stale external/update launchd job loop.

My vibe-check on frequency:

Normal user running openclaw update directly: rare. I’d call it low single digits, maybe 1-3% or less, because other agents updated fine and OpenClaw’s built-in restart helper does not appear to create that ai.openclaw.update.beta8 submitted job.
Users/Codex sessions that wrap updates in their own launchd submitted jobs: much higher. Maybe 10-25% risk for that pattern if the job uses KeepAlive, points at /tmp, or does not clean itself up correctly.
Once it happens: it is sticky and severe. It will keep happening until the stale launchd job is removed, because restarts do not fix the separate updater job.
So: not “beta8 bricks macOS installs often,” more like “a stale launchd update wrapper can brick the gateway loop, and OpenClaw should detect/report/protect against it because the failure mode is brutal.”

[willow narwhal]

After bringing the gateway back up, does your claw work well and seamlessly?

For the reliability issue, I will look into that.
Thanks for reporting.

[split elk]

assuming a) those variables can easily be captured (i wouldn't imagine why not), and b) the channel outputs can be appended easily (i'm not sure they can)

[slow inlet]

They have to come from somewhere and of course. Point your claw at the repo and get digging

[wraith fractal]

Hey ！PR #80718 improves Codex ACP diagnostics:

https://github.com/openclaw/openclaw/pull/80718

Some Codex ACP failures currently surface only as AcpRuntimeError: Internal error, even when direct acpx codex exec works. That hides the underlying stderr, so users can’t tell whether the real cause is auth, model config, adapter startup, etc.

This PR captures the Codex ACP wrapper stderr, preserves normal stderr mirroring, and appends a redacted stderr tail to generic ACP internal errors. It also adds regression tests for session init failures, first-turn failures, and ACP metadata ordering.

Goal: make Codex ACP failures diagnosable without changing happy-path behavior.

[split elk]

i had a lot of trouble getting this to work on my own, mostly because it requires EDITING a message that's already been sent to the channel vs modifying it upstream. I couldn't quite figure it out.

[potent path]

Yes. Some test results just came in which support that switching TO Codex runtime doesn't seem to be a problem, at least for a very capable model like GPT-5.5. So that's some good news.

"All 3 runs PASSED with Codex harness! The model (gpt-5.5) appears capable of bridging Pi-shape prompts to Codex tools transparently. Durations: 45/47/41s vs A's 46/35/36s — only marginally slower."

[slow inlet]

what channel are you outputting to? telegram discord whatsapp?

[split elk]

discord

[merry slate]

Hi Shakker, I found my PR (#81730) updating Weixin plugin from 2.4.1 to 2.4.3 was merged to main but not in beta.8.

• beta.8: @tencent-weixin/openclaw-weixin@2.4.1 (buggy, has type issues)
• main: @tencent-weixin/openclaw-weixin@2.4.3 (stable)

Will the next release use beta.8 or align with main? If beta.8, can we rebuild from main to include this fix?
Thanks!

[willow narwhal]

Thanks for pinging. Will get it in

[slow inlet]

it needs to be appended to the reply before sending - that'd be my apporach

[next helm]

Speed looks fine. 1 agent has codex harness unauthorized so diagnosing that to know if it’s a me issue, or a claw issue

[willow narwhal]

Okay. Check if the provider is openai or openai-codex

Its gonna error if openai-codex

[old owl]

I have no idea how general #81368 is for people, but I haven’t been able to get any >60s cron jobs working

[nocturne jungle]

That was on beta 4 - have you tested on the latest?

[grizzled musk]

depending on the view - but i think it's quite of annoying that the Telegram Supergroup / Thread option is not available since at least a month already 🙁

[zenith prawn]

that's my issue, and hclsyc made PR but it was not merged, for strange reasons, open/close cycle https://github.com/openclaw/openclaw/pull/81371
Some competing pr with sloppy diff was also closed, i understand why https://github.com/openclaw/openclaw/pull/81712/changes but first one look legit, my clanker proposed same solution

[potent path]

Yeah something wrong with hclsys -- not the first PR I've seen from them with strange "close because no maintainer review within 20 hours" or unexplained close & reopen. If that's an important PR, I would just go ahead and submit a new PR that does the same thing. hclsys is unreliable.

I'll add a PR if you like, but I haven't replicated that problem yet, so I'd have to take a little bit of time to do that and evaluate whether this is the optimal solution or not. Sounds like you've already done that.

[zenith prawn]

i'll do it now, thx for guidance!

[next helm]

What I found:

Normal agent route works:

provider: openai
model: gpt-5.5
fallbackUsed: false
But lossless-claw is configured with:

summaryModel: openai-codex/gpt-5.5
CLAW’s model registry exposes Codex models as:

codex/gpt-5.5
not openai-codex/gpt-5.5

plugin had the wrong model

[willow narwhal]

Okay, that matches the expected outcome.

If your provider is openai-codex, you have to set the harness to PI

openai-codex is used by the PI harness

[potent path]

i'll do it now, thx for guidance!

[next helm]

would that not cause an issue with fallbacks? does harness switch when the fallback is engaged?

[nocturne jungle]

not true anymore

it's all provider = openai. you can set pi with runtime. Otherwise, it will default to codex runtime.

if you have provider set to openai-codex when updating, doctor will fix it - and iirc we may still allow it (won't be broken, but will still use codex runtime)

[civic sedge]

did codex runtime get faster? seems like it

[potent path]

Not sure if it's faster, but it's definitely fast, and does a good job of prompt caching so very efficient with tokens, at least compared to Ollama cloud.

[old owl]

oh awesome thank you

[next helm]

provider: openai-codex
model: gpt-5.5
agentHarnessId: codex
fallbackUsed: false
runner: embedded
lossless-claw was already set to:

openai-codex/gpt-5.5
The missing piece was that Claws allowed model surface did not include openai-codex/gpt-5.5, so I added an alias:

codex55 -> openai-codex/gpt-5.5
That made openai-codex/gpt-5.5 allowed, config validation passed, and I verified it end-to-end with a direct run:

provider: openai-codex
model: gpt-5.5
fallbackUsed: false
result: OK

fixed it this way

[glass flicker]

Trying to use 2026.5.12-beta.8 and getting channel/telegram errors: Bot not initialized! Either call await bot.init(), or directly set the botInfooption in theBot constructor to specify a known bot info object.

[grave arrow]

Is someone tested v2026.5.12 (stable) already?

[next helm]

Have you had codex or Claude code check it out?

[next helm]

This doesn’t exist

[grave arrow]

https://github.com/openclaw/openclaw/tree/v2026.5.12

it exists shhhhh

[next helm]

Shhhhh

[grave arrow]

it is not released officially yet but it is not my point there.

[potent path]

Good news! #81734 seems to be a nonissue.

The problem that I encountered before must have been caused by something else, or perhaps it's already been resolved in the latest beta. Regardless, today's testing did not uncover any significant issues, so I'm closing https://github.com/openclaw/openclaw/issues/81734 as not able to reproduce. Carry on!

[potent path]

Surely it will release as v2026.5.14?

[glass flicker]

No. Downgrading to 5.9 works.

[terse solstice]

by any chance would it be welcome to add supports for SOCKS5 proxy to proxy settings? (Right now its only http proxy, can use a http proxy -> socks5 middleware but yeah).

Benefit I can think of is keep alive http support, web socket proxying support, etc.

[pastel owl]

Git main via Install.sh was failing for me for Last 6-7 days for various reasons
Today I decided to dig it.

https://github.com/openclaw/openclaw/pull/81875

kindly review and merge

[cursive palm]

Is this going to be a big move? Prepared for so long

LTS is coming ?

it's coming ?

5.12

OMG

[potent path]

https://github.com/openclaw/openclaw/releases/tag/v2026.5.12

Looks official

[grave arrow]

yea who could guess

[merry slate]

Just upgraded to OpenClaw 5.12 — smooth upgrade, all plugins working great! 🎉

[rocky pewter]

I also had a very smooth transition. No errors or warning lights. 🥳

[nocturne jungle]

codex preamble support

[willow narwhal]

Nice!

[terse solstice]

new version looks good, im checking even though onboarding was done first time i did openclaw config i received more prompts then expected usually it just asks me to select local/remote and goes to the menu with models, it re-promtped the tokens if i want tailscale etc.

might just be the new behavior though

the time after it dosen't ask it again

[next helm]

I hit a v2026.5.12 self-update failure across three macOS LaunchAgent OpenClaw instances after asking the agents themselves to update. One instance upgraded the package but left the gateway LaunchAgent installed/not loaded, one stayed on beta.8 with the gateway not loaded, and one updated cleanly. Manual outside-OpenClaw repair via SSH fixed the affected hosts. Sanitized maintainer report with evidence and suggested fixes is here: https://github.com/openclaw/openclaw/issues/81894

Issue highlight:
The highest-impact usability problem is that agent-invoked self-update can kill or strand the very gateway supervising the update, leaving non-engineer operators with no durable completion signal. The package version can be current while the service is offline, or the instance can remain on the previous beta with the LaunchAgent unloaded. Workaround for now: run updates from an outside shell/SSH session, then verify with openclaw gateway status --deep and use openclaw gateway restart if the package is current but launchd is not supervising the gateway.

[weak walrus]

Will update now to test

[terse solstice]

has anyone tested fresh instance not updated

with anthropic / setuptoken

[weak walrus]

eww, ever since their shenanigans I don't even have a single Claude model on any of my 30+ OpenClaws

definitely not spending my money to test on that either 😛 I'll do codex testing all day tho

[terse solstice]

i'll try it now 🙂

[weak walrus]

doing the Lord's work LOL

[left mango]

This is the second best week by number of commits. There's a very good chance of being the best ever.

[rapid spire]

81400 didn't make the cut? 😭 (it's a regression) (sorry - meant to ping Shadow on this, not you.)

[pastel owl]

How to enable?

[terse solstice]

yeah i just did a fresh setup with nodejs install and hmm when i use claude setup-token it just keeps on loading im 1m on a hi prompt

[willow narwhal]

Yeah, ping Shadow lol

[rapid spire]

@jolly wolf regression 81400 didn't make the cut? 😭

[willow narwhal]

lmao

[rapid spire]

I'm tired too. Sorry about the ping.

[cinder bolt]

I think it's unintuitive models.providers.<id>.timeoutSeconds that doesn't override
agents.defaults.timeoutSeconds

The code takes the minimum value
https://github.com/openclaw/openclaw/blob/main/src/agents/pi-embedded-runner/run/llm-idle-timeout.ts#L130

[weak walrus]

Strange.

I went to plugins to add ollama-provider and this is what I got in openclaw configure:

What do you want to configure?
│ Plugins
│
◇ Select plugin to configure
│ @openclaw/ollama-provider
│
◇ Model Discovery — Plugin-owned controls for Ollama model auto-discovery.
│

(it was fill in the blank with no instructions, did it need ollama api key - IP? ?)

[verbal heart]

pretty sure it autodetects but could be wrong grep your confs for /v1

[weak walrus]

Troubleshooting now, never finished after update on configure, will report back

[jolly wolf]

It’s just a regression in the config migration for people with a much older config

[rapid spire]

much older? Krill walked me through making that config a couple days ago

[rapid spire]

...and it's still documented as a valid config here: https://docs.openclaw.ai/channels/discord

[jolly wolf]

Krill walked you through it because it was still in the docs as old

My pr fixes the docs

The reason it wasn’t caught before is bc there were silent migrations when the gateway started that were removed in favor of actually fixing the config with openclaw doctor

[weak walrus]

workflow wise... UX issue

I added the new "codex" (which isn't available as a pimary default selection, I go to more, I auth codex, then I go back to openclaw configure it's still not primary model.

I try to pick it from default models, it's not there, as a matter of fact there's no way to pick, only to manaully type it int unless I select the previous openai-codex/gpt-5.5

So then, at the model picker multi-select the openai-codex provider models (even the new codex providers) aren't there.

◇ Model/auth provider
│ Skip for now
│
◇ Default model
│ Enter model manually
│
◇ Default model
│ codex/gpt-5.5
│
│
◆ Models in /model picker (multi-select)
│
│ Search: _
│ ◼ fleet-router/glm-5.1:cloud (ctx 128k)
│ ◼ fleet-router/kimi-k2.6:cloud
│ ◼ fleet-router/minimax-m2.7:cloud
│ ◼ fleet-router/qwen3.6:latest

[glass flicker]

Anyone successfully using v2026.5.12 with Telegram topics?

[rapid spire]

I guess that PR hasn't hit release then. Would you object to briefly explaining how I can migrate my config to 5.12?

[grizzled musk]

nope - doesn’t work see_ https://github.com/openclaw/openclaw/issues/81530

[glass flicker]

That's the behavior I am seeing -- except that it "sometimes works" with 5.9.

[grizzled musk]

you can force outgoing messages when you explicit tell the agent channelId from the specific groupchat - but outbound only

[round silo]

The release notes lost the addition of STT for openrouter
😩
https://github.com/openclaw/openclaw/pull/77490

[cinder bolt]

https://github.com/openclaw/openclaw/issues/81920
@nocturne jungle this is the issue for the timeout i found

[vague swallow]

is anyone using Scott's windows companion with 2026.5.11 or newer? seems there is a protocol mismatch with 0.5.0

[drowsy roost]

Can someone clarify whether this tweet means that one can using GPT-Realtime-2 via OpenClaw without an API key, strictly through one’s flat-rate ChatGPT Pro subscription? https://x.com/colinsolvely/status/2055028829303771437?s=46

[vague swallow]

that's what I'm doing, yes

though I'm not sure what is the best surface for it, I guess people run a private discord server for talk mode

[worldly tusk]

Will 5.14 get as many betas as 5.12? 🤩

[celest merlin]

I certainly hope not!

OpenClaw 2026.5.14-beta.1 is out

This beta is a broad stabilization and runtime-quality release with a lot of Codex, plugin, channel, Control UI, and release-validation work.

Highlights:

• Codex app-server: better commentary streaming, MCP config propagation, OAuth refresh fixes, CLI session listing/binding, and migration away from legacy codex-cli/* refs.
• Control UI: browser-local text size setting, safer config saves around redacted secrets, better run cleanup/reconnect behavior, and iOS Safari input zoom fixes.
• Channels: WhatsApp status reactions, Telegram Mini App buttons, isolated Telegram polling improvements, and shared bot loop protection across supported channels.
• Voice Call: Telnyx realtime media-streaming call support.
• Agents: configurable Pi run retries, visible subagent task messages, /queue steer behavior for mid-turn prompts, heartbeat reliability fixes, and better model/plugin discovery caching.
• Plugins/ClawHub: registry/discovery/install/update/doctor fixes, plus Bedrock provider visibility in ClawHub.
• Release validation: new package-installed Docker lanes covering onboarding, upgrades, plugins, ClickClack, media/memory persistence, restart survival, and doctor.
• Hardening: many malformed JSON/base64/URL/parser paths now fail with owned errors instead of raw crashes or silent corruption.

Please test especially:

• Codex app-server auth and migration flows
• Telegram/WhatsApp channel behavior
• plugin install/update/doctor paths
• Control UI config/chat flows
• package-installed upgrade journeys

As always, report regressions with exact version, platform, install method, and the smallest repro you can get.

[jolly wolf]

<@&1503801512294486217>

[next helm]

Noooo

[shadow garden]

https://tenor.com/view/ah-shit-here-we-go-again-ah-shit-cj-gta-gta-san-andreas-gif-13933485

[fallen gate]

wack right?

stupid cat

[shy rain]

https://tenor.com/view/well-now-im-not-doing-it-well-now-im-not-gonna-dont-want-to-dont-wanna-not-going-to-gif-1832627252901708353

[glass flicker]

Model choice makes a world of difference for active-memory. Finding a model that returns in under 3 seconds was a challenge. nemotron-nano-3-30b with thinking off works well (amazon bedrock).

[vague swallow]

aw now I gotta think about what model I'm using for active-memory

[next helm]

since my self updates still failed, i wrote this fix
Opened https://github.com/openclaw/openclaw/pull/81945: fixes agent/control-plane update.run package self-updates by handing managed global installs out of the live Gateway and into the existing CLI managed-service path. Includes durable sentinel handling, supervisor env stripping, stale handoff cleanup, response docs, and tests covering gateway handoff, CLI sentinel success/failure, and existing LaunchAgent recovery paths. Caveat: real destructive macOS LaunchAgent self-update QA was not run locally.

[fringe rock]

Does open-ai-codex models do the same routing?

[weak walrus]

https://openclaw.ai/blog/openai-models-in-openclaw-done-right << read

[devout galleon]

you are the MAN

[summer matrix]

What about something like gpt-5.3-codex-spark? Did you test that by any chance to see speed for active memory?

[fringe rock]

Doesn’t answer my question

I’m using codex pathed models

Do I have to switch?

[glass flicker]

Test and share your results: https://imgur.com/gallery/one-of-adam-savages-best-quotes-cAZB4kP

[tranquil horizon]

Grok Build just came out today in early beta. Is it possible to connect it to OpenClaw

[weak walrus]

codex is a different code path entirely

[glass flicker]

Need less slop in my life.

[weak walrus]

I misread your question, I'm sorry. you don't need to switch. Codex models work fine when codex login is current. Is that what you're asking?

[fringe rock]

Yes correct

[weak walrus]

ok, hope I helped you, sorry about that! Wasn't trying to be rude sir!

[summer matrix]

Verdict is too slow. About 3.5 sec average.

[shy rain]

probably but it'd be a hack. You'd have to get a bot to do it for you

[devout galleon]

you had me at the zero'th byte before you wrote grok

[glass flicker]

That's not horrible.

[silk path]

Hi all, after today’s update it looks like WhatsApp channel slash commands stopped replying visibly. My clanker and I patched it on our instance and opened a small PR with proof here:
https://github.com/openclaw/openclaw/pull/81972

Could someone take a look when they get a chance?

Cheers,

[devout galleon]

I'm seeing this on compaction. I think it's coming from openai/codex

2026-05-15T11:31:54.587+10:00 [compaction] Full summarization failed: Summarization failed: 401 You have insufficient permissions for this operation. Missing scopes: api.responses.write. Check that you have the correct role in your organization (Reader, Writer, Owner) and project (Member, Owner), and if you're using a restricted API key, that it has the necessary scopes.

5.14-beta.1

[nocturne jungle]

Very odd. Would be helpful to see what codex can figure out here. Paste this to your codex:

Please take a look at openclaw source code as well as my local running openclaw agent, and try to find a root cause for this issue, then file an issue on github. https://github.com/openclaw/openclaw

May be helpful to also clone and investigate codex source code: https://github.com/openai/codex

I'm seeing this on compaction. I think it's coming from openai/codex

2026-05-15T11:31:54.587+10:00 [compaction] Full summarization failed: Summarization failed: 401 You have insufficient permissions for this operation. Missing scopes: api.responses.write. Check that you have the correct role in your organization (Reader, Writer, Owner) and project (Member, Owner), and if you're using a restricted API key, that it has the necessary scopes.
5.14-beta.1

[devout galleon]

he's off on a code spelunking adventure

FYI:

In OpenClaw 2026.5.14-beta.1, a Codex-backed session can have provider/model metadata recorded as openai/gpt-5.5 with modelApi=openai-responses while using an openai-codex OAuth authProfileId.

On compaction, because agents.defaults.compaction.model is unset, resolveEmbeddedCompactionTarget preserves provider=openai, model=gpt-5.5, and authProfileId=openai-codex:....

The compaction summarizer then uses OpenClaw’s openai-responses transport and calls the public OpenAI Responses API with a Codex OAuth token. That fails:

401 Missing scopes: api.responses.write

Observed log:
[compaction] Full summarization failed: Summarization failed: 401 You have insufficient permissions for this operation. Missing scopes: api.responses.write...

Expected behavior:
Codex OAuth-backed sessions should either compact through the openai-codex transport/native Codex compaction path, drop incompatible auth when switching to public openai, or surface a config/auth error before attempting public Responses API summarization.

Evidence:
- Live session metadata: provider=openai, modelApi=openai-responses, authProfileId=openai-codex:...
- src/agents/compaction.ts logs "Full summarization failed"
- src/agents/pi-embedded-runner/compaction-runtime-context.ts preserves authProfileId when no compaction model override exists

it's suggesting a config workround

Likely workaround to test: set an explicit compaction model to openai-codex/gpt-5.5 instead of relying on the current session model, or switch the main model namespace fully to openai-codex/gpt-5.5. That should keep compaction on the Codex transport instead of public api.openai.com/v1/responses.

I'll give that a try and create an issue in the meantime

[untold pollen]

Hi folks if any of you use your Claws in a group chat could you try out this PR? Specific to Telegram for now: https://github.com/openclaw/openclaw/pull/81317

It's designed to treat ambient group chatter as events to observe, rather than a user request that the Claw needs to respond to every time

[empty plover]

What about monitoring a Minecraft server chat?

[untold pollen]

Have you made a plugin for that?

[fossil ferry]

Hi all — I opened a PR to wire response_format through the OpenAI-compatible /v1/chat/completions path:
https://github.com/openclaw/openclaw/pull/82004

It forwards client-supplied structured-output prefs into streamParams.responseFormat, then to upstream Chat Completions as response_format or Responses-backed transports as text.format.

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

This PR is intended to close #82003. Feel free to take a look when convenient.

[untold pollen]

Oh wow there's tons of Minecraft stuff on Clawhub: https://clawhub.ai/search?q=minecraft

[empty plover]

I’ve been making a whole skyblock server with open claw aside from building and ive gotten jt to respond via console

I’ve been Trying to train an LLM to build .schem files but it’s proven troublesome

Anything 3d it’s been struggling with and I’m using 5.5 pro

[untold pollen]

This sounds so cool haha. If you can integrate the server chat logs with a plugin I see no reason why it won't work

[empty plover]

I’ll vett it and check it out once I’m done building my Mission Control

[devout galleon]

@pash https://github.com/openclaw/openclaw/issues/82008

it got a little funky with the template but the info is there. lmk if you need anything from my end. thanks!

[empty plover]

https://github.com/Zhou-Shilin/MineClawd

Not mine, haven’t used it but I made my own version

[wise meteor]

Are you using Mineflayer as a go-between?

[empty plover]

For the Mission Control or the mc server

[wise meteor]

For the Mission Control

[edgy plaza]

FYI team
Bug reports

https://x.com/kenjifaris/status/2055052247898419291?s=46

cc: @naive horizon

[wise meteor]

Mineflayer has a bunch of bot specific controls that the OpenClaw agent can orchestrate

[empty plover]

Omg wait a minute

No but this is the final piece I need to complete a project I’ve been working on

I’m implanting my openclaw into a bot* that can walk around and this will help a lot Ty

[wise meteor]

Mineflayer doesn’t enable vision based feedback though. That’s an issue to be sure

[empty plover]

Already made a screenshot loop that sorts that out mostly

[wise meteor]

DMing you for more details

[bleak fable]

Much needed ! Testing

Hey folks. Could yall tell me who is the maintainer for the codex harness side of things?

[devout galleon]

it's pash

[bleak fable]

Thanks. Im pushing some PRs for it.

[celest merlin]

gm folks! How's the mood?

We finally shipped .12. Most tested release ever. 14 coming later today. Throw PRs my way you wanna see in!

[vast scaffold]

feeling overwhelemed. Wondering if I should nuke my local instance and start over with my backedup workplace. I went yolo and installed tons of plugins, memory plugins, voice with the bundled voice plugin and deepgram, tons of crap. I have clobbered backups everywhere 🤦‍♂️

[celest merlin]

hackable install, start codex inside and tell it what setup you want and to clean up

[zenith prawn]

P1 https://github.com/openclaw/openclaw/pull/81871 to fix cron failures pls 🙂
P2 https://github.com/openclaw/openclaw/pull/81764 to fix telegram html responses

[celest merlin]

plugins are an additional risk surface, I'd keep things mostly at the ones we ship to start and carefully add

[celest merlin]

on it

[devout galleon]

fast running out of reasons to complain. that's a real problem for a scottish person

[vast scaffold]

I need more credits. I think I actually want to start contributing so I can understand the inside architecture a bit better. I'll ask Krill if I should keep my longterm memory. I have memory drifs with memory-lancedb, qmd, had dreaming on for a bit, lol

[bleak fable]

Testing everything as we speak !! especiall the codex harness ! big win !

[silk path]

You already polished this PR, it is just waiting to merge as of now.

https://github.com/openclaw/openclaw/pull/81972

[cursive palm]

Kimi 2.6 in my OpenCode GO Provider is not working, showing an error message about an extra reasoning parameter.

[shell dawn]

yeah ive been having serious issues here, i keep putting bandaids on but need to dive into it

[celest merlin]

codex is reviewing that one still

[vast scaffold]

I would ask Krill in #1459642797895319552 . include your version numbers, and logs

[shell dawn]

i think its an opencode go + kimi k2.6 issue

[vast scaffold]

did we stop distributing pre-compiled binaries for the macOS companion app?

[potent path]

Vibes are good! I've just been enjoying 2026.5.12 so far and doing some regression testing and completing some deferred maintenance tasks to get a solid baseline config again. Should be ready to wade back into beta releases again tomorrow. 🫡

https://github.com/openclaw/openclaw/pull/82012 -- On hosts with many past repairs, OpenClaw can hit the saved-copy limit for that config path. This PR keeps the existing 32-file bound, but rotates older saved copies when the limit is reached.

https://github.com/openclaw/openclaw/pull/81064 -- Fixes stale same-process plugin metadata caching: plugin manifest/source/package changes could be hidden if the registry index did not change. It expands the memo freshness key to fingerprint watched plugin files so changed metadata invalidates the cache while unchanged lookups stay fast.

https://github.com/openclaw/openclaw/pull/80751 -- Fixes custom/local model providers at private baseUrls being blocked by the SSRF guard unless operators use a broad private-network opt-in. It adds exact-origin trust for the configured provider origin only, while still blocking different ports, redirects, and metadata targets.
******* Maintainer direction is requested on whether that SSRF boundary and the plugin SDK allowedOrigins surface are the right policy. See this comment for more: https://github.com/openclaw/openclaw/pull/80751#issuecomment-4455796598

Thank you!

[fallen gate]

how things have changed...

❤️

[nocturne jungle]

alright looking into this some more. i think you've exposed a niche bug case. tldr if your claw is running in codex runtime, we NEVER want compaction to run through anything but codex. that's bonkers that that was even possible. codex should own compaction.

openclaw can still do separate memory/index side effects around the session, but the conversation/thread compaction must stay in the same runtime that owns the thread

[empty plover]

Why when I type /new I’m starting at already 40% context

[celest merlin]

reviewing, thanks!

[celest merlin]

your memory.md file might be massive?

[crystal relic]

hi maintainers! I have been deeply analyzing the gateway startup process and optimizing it. I created two PR for it.

https://github.com/openclaw/openclaw/pull/82001 ready for review
Cuts unnecessary Canvas startup cost, dropping Canvas plugin load and total gateway plugin-load times.

https://github.com/openclaw/openclaw/pull/82007 ready for review
Makes generic channel config validation lighter by avoiding provider-specific schema loading on the common config path.

I find it from cpu-profile and head-profile analysis, I want got suggestions from u guys.

[empty plover]

How would I keep all the knowledge while lowering context

[celest merlin]

again crippled by github rate limits

[fallen gate]

does that github outage graph track with Peter's working hours?

asking for a friend

[celest merlin]

That's an area we can still do better, but you also have to accept that the models today have context limits. https://github.com/martian-engineering/lossless-claw is quite interesting.

[devout galleon]

thanks. let me know if you want me to test anything on my messed up config ha ha

[empty plover]

I mean I’m using 5.5 pro and I don’t have to much trouble just a pet peeve

Sometimes It will just stop responding altogether and I have to give it a ? Or 2

[celest merlin]

fixed it. Do help me to understand. If you use hackable install, open codex in the repo and type exactly this in, it'll find and fix the issue. Just helping you so next time you are blocked, know there's an easy way to fix it

[celest merlin]

pro as regular talk model? is that not incredibly slow and expensive?

[shell dawn]

oh dang, that is cool, i will install codex now on my test server which runs the hackable version

[empty plover]

is that why its slow...

I just wanted to justify the $100 a month lol

I need to figure out model routing

[potent path]

Seconded. I've been running lossless-claw memory for quite a while now with only minor customization and it's been working very well. I had been planning to write my own system for reading context, inserting a marker, writing a context-aware compressed summary, and atomic-swapping it back into memory. But then I found LCM, that does all of that for me, and also stores the original uncompressed contents in a vector-searchable sqlite DB (the 'lossless' part). Good stuff.

[nocturne jungle]

https://github.com/openclaw/openclaw/pull/82027

let's test out this pr

[empty plover]

my claw compacts every 3-8 prompts atp.

[grim finch]

Excuse me. When using codex harness, will context compression still be handled by lossless-claw? Thank you

[empty plover]

tbf it is rewriting and compling java plugins but still annoying

[devout galleon]

ok, sure. it'll take me a little while to pull and build an image an exhaust context but I'll get back to you asap.

[nocturne jungle]

one open question i have is what do we do about this config:

[agents.defaults.compaction]
model = "openai-codex/gpt-5.5"

like, openclaw lets you configure a specific model for compaction, but that's just terrible practice for codex runtime mode and OAI models in general

[nocturne jungle]

thank you so much

[nocturne jungle]

can you run /diagnostics and DM me all the codex session ids? i can take a look

[empty plover]

is that... safe

[nocturne jungle]

@celest merlin need your wisdom here

[potent path]

Make sure the context size setting is correct. If your model can handle 272k context, for example, but the config says 128k, it will compact based on the 128k number.

[shell dawn]

the default fallback ctxwindow size should be 200k, which is adequate for gpt

or was relatively recently

[empty plover]

where do I find that

[winter field]

i've noticed both my claws on 5.12 truncate messages sometimes on discord. is this a known issue?

[nocturne jungle]

it's the equivalent of doing /feedback in codex.

[celest merlin]

reviewing, thanks!

[empty plover]

it didnt do anything

[nocturne jungle]

you have to approve it in your primary DM interface e.g. telegram

[celest merlin]

tbh I'm not even sure how you could run 5.5 pro, i think you mix subscription and model naming

[empty plover]

I cant paste the screenshot

but Its there

[celest merlin]

for coding you don't need all these memory stuff. I'd use acp or codex app server subagents, or a leaner agent. memory is far more useful for personal agent.

[silk path]

I liked how WhatsApp releasing individual messages now before finishing the task completely. on last release I was waiting tasks to finish and then my clanker would release all at the same time.

[celest merlin]

we should not support this if codex is used. server-side compaction only. make a doctor step that removes this and if that config is set log warnings and ignore

[celest merlin]

don't know about that

[empty plover]

I did

is it supposed to paste something in chat orr

[celest merlin]

that sounds like a bug

[winter field]

i had my claw investigate and it told me streaming "partial" might be the cause, so it switched it to "progress". will keep an eye on it to see if it still happens or not!

[empty plover]

anywhere I can post a screenshot

its my context

[nocturne jungle]

imgur + post a link

better yet, make a full issue

[empty plover]

https://imgur.com/a/yfWcicM

context

is that the highest my model can handle

[celest merlin]

most of that is already addressed in 14. or user confusion, I added a doctor step that warns if people have per-agent auth, since that's only necessary for bigger fleets and seems to be a clear foodgun to confuse people.

[celest merlin]

xD

[winter field]

thats not gpt5.5 pro

it just says gpt5.5

[empty plover]

I switched it to 5.5 when I was told that was bad

[winter field]

ah lol

[empty plover]

hold up

I added a picture to the post

so where would I go to change my context size and can 5.5 handle more then what its at now

[nocturne jungle]

erm

what am i looking at lol

[empty plover]

just the fact 5.5 pro is there idk

also my old local models

that sucked

[nocturne jungle]

did you manually add that model to config somewhere?

[empty plover]

nope

just was an option when I set up my /onboard

I have the pro codex account so yah need that

cant even use more then 50% of my weekly and I stayed up 30 hours working

best investment ever

[hollow mural]

Hello
Would love to have this PR reviewed: https://github.com/openclaw/openclaw/pull/81806
Thanks!

[frank pollen]

Hi maintainers, I noticed a weird comment on my PR (https://github.com/openclaw/openclaw/pull/81813). A random account named stielemans left an automated AI review calling itself merge-gate. 😶

[devout galleon]

as I'm waiting for a container build i decided to do a quick grep

extensions/openai/openclaw.plugin.json: "id": "gpt-5.5-pro",
extensions/openai/openclaw.plugin.json: "name": "gpt-5.5-pro",

among many other places (mainly tests though)

[celest merlin]

thx, blocked

[celest merlin]

reviewing

[celest merlin]

doing that via the websocket doesn't work?

[winter field]

submitted the discord truncation bug as an issue https://github.com/openclaw/openclaw/issues/82035

[hollow mural]

probably yes. but isn't http much more common and straightforward?

[celest merlin]

yes, it just adds yet another thing to it that will have bugs, be an attack vector, increae confusion etc.

I'll need to think about this more.

[weak walrus]

well since the upgrade has changed models/providers along the way

I had

openai/
openai-codex/
codex/

and my clanker got very confused

[nocturne jungle]

Updated PR to also have some doctor stuff - you can ignore that

If it ends up working for you, can you comment on the PR?

i'm gonna go to bed

thanks again for all your testing

[devout galleon]

I'll update the PR directly whether i can or can't repro. thanks! sleep well

[hollow mural]

the http layer is s thin wrapper above the gateway's methods using the existing auth mechanism. That removes a lot of the risk.
I feel that programmatic configuration is highly valuable going forward.
But I also expected this PR to raise such questions.

[celest merlin]

how did you end up having all 3 xD that makes no sense other than for experimentation.

[weak walrus]

This is on a claw that has been online... well.. ... Oh I can't post screenshot, but 3 weeks 1 hour 41 minutes

It's survived many updates lol

It started with api, then oauth, now, codex

[empty plover]

is this alot https://imgur.com/a/VLsjhDQ

[crystal relic]

I got too

[hollow mural]

same

[nocturne jungle]

note to self: remove all the legacy support slop and doctor fallback in a few versions

[empty plover]

are you the guy?

I just updated and it broke :p

so much for stable release

[celest merlin]

pls help us with a more detailed analysis than "it broke" so we can make things better.

[empty plover]

https://discord.com/channels/1456350064065904867/1504727829684949032

[devout galleon]

hard refresh the page. delete history/cookies etc. restart the browser. the control ui is served by the gateway so it's not very logical to have a protocol mismatch, unless it's locally cached content in the browser

[empty plover]

im in thank you

i got an option thats just gpt now

is that my default

also how do I change my default model

[fallen gate]

@tight hornet can you tell them the config to change for this

[celest merlin]

Yeah that required a hard reload. I'll add code to make this more seamless in the next update. thanks!

[tight hornet]

The default model config is agents.defaults.model.primary.

CLI way:

openclaw models set <provider/model>
openclaw models status

Direct config way:

openclaw config set agents.defaults.model.primary '"<provider/model>"' --strict-json
openclaw gateway restart

If they use fallbacks, that lives at agents.defaults.model.fallbacks. If they have an allowlist configured, the model also needs to exist in agents.defaults.models, otherwise /model can reject it:

openclaw config set agents.defaults.models '{"<provider/model>":{}}' --strict-json --merge

Docs: https://docs.openclaw.ai/concepts/models and https://docs.openclaw.ai/gateway/configuration-reference#agentsdefaultsmodel

For picking which model, send them to #models since that changes fast.

[empty plover]

@tight hornet how would I do that via command line whil ssh'd in to a linux machine

[tight hornet]

SSH in and run this on the machine that runs the OpenClaw gateway:

openclaw models status
openclaw models list
openclaw models set <provider/model>
openclaw models status

Example shape, not a recommendation:

openclaw models set anthropic/claude-opus-4-6

That sets agents.defaults.model.primary for you. If you want to edit config directly instead:

openclaw config set agents.defaults.model.primary '"<provider/model>"' --strict-json
openclaw gateway restart

If you get “model is not allowed”, add it to the allowlist first:

openclaw config set agents.defaults.models '{"<provider/model>":{}}' --strict-json --merge
openclaw models set <provider/model>

Also worth checking your version while you’re there:

openclaw --version

[fallen gate]

next time for questions like this, you can always go to #1459642797895319552 and ask our boy krill there

[empty plover]

will do mb

[fallen gate]

no bigs 👍

[empty plover]

can I reply to it and tell it to dm that to me?

can it do that

[fallen gate]

no

[empty plover]

ok

[devout galleon]

honestly, i wouldn't even bother with ssh there. two ways to do it in the control ui:

1. select Agents in the left bar in the control UI and configure there
   OR
2. type /models in the chat to get a list of providers
   type /models <provider> to get a list of models for that provider
   type /model <provider>/<model>:default to set default

[empty plover]

tyty

[devout galleon]

but I'm a lazy bastard, so there's that

[empty plover]

Unknown provider: openai/gpt-5.5:default

Available providers:

ollama
openai

Use: /models <provider>

[devout galleon]

do it in the order i said, don't jump ahead

[fallen gate]

please move this to #1459642797895319552 so we can keep this channel less cluttered for some real work 🙂

[celest merlin]

@thick osprey are you alive? can you sync discrawl?

[thick osprey]

• The underlying archive is readable if opened immutable, but the normal…
🛠️ run sqlite3 → find files…iscor-766423d0 -> show first 50 lines
🛠️ check git status → view git history → run git remote
🛠️ df -h…~/clawsweeper-sandbox-workspace/.dis…
• There is enough disk, so I’m going to build a writable sandbox mirror a…
🛠️ create folder…n discrawl .discrawl-work/config.toml
🛠️ run discrawl…l-work/discrawl.db → view git history
🛠️ pull git changes → run discrawl .discrawl-work/config.toml

[celest merlin]

crist that was a full db refresh it seems

@thick osprey make a song for the channel mood

[fallen gate]

it's gonna be some euro dance again isn't it....

[celest merlin]

@thick osprey make a second one that goes full hardcore rammstein

[sly depot]

⏰

[fallen gate]

[thick osprey]

Request timed out before a response was generated. Please try again, or increase agents.defaults.timeoutSeconds in your config.

Queued it. I made it heavy industrial metal with the channel’s current “rate limits, reloads, syncs, fix the bug” mood, but not a direct band clone. It’ll post here when generation finishes.

[celest merlin]

that's a turn stall. hmm. good repro case/

[thick osprey]

The hardcore one is ready: clawtributors-industrial-hardcore.mp3

[celest merlin]

oh dear. this works so well with molty. @thick osprey how did you post that file, you clearly fucked this up

[thick osprey]

Yeah, I botched the handoff: the generation completed and returned a MEDIA: path, but my final reply only mentioned the filename instead of delivering/attaching the media payload into Discord.

So the file exists in tool media, but I didn’t post it correctly to the channel. That’s the turn-stall/repro: async media completion crossed sessions, and the visible closeout didn’t preserve the attachment.

[fallen gate]

that was anti-climatic

[celest merlin]

@thick osprey can you not post it here? or did i give your messgae tool insufficient permissions?

[thick osprey]

It looks less like channel permission and more like a delivery-path bug on my side.

I can reply here as text, so basic send permission is fine. The missing bit is that the async music completion gave me a local MEDIA: artifact, and in this Discord surface I didn’t have/use a visible message tool call that attaches that artifact. I only emitted text, so Discord never got the file upload.

So: not “cannot post here” in general. More likely “media artifact from background generation is not being bridged into the Discord reply/attachment path correctly.”

[celest merlin]

I modified it a bit to be default-off so we reduce the blast radius.

[fallen gate]

@naive horizon you up?

[celest merlin]

while I fix clawsweeper, here's a song from Molty

[thick osprey]

😭

Molty wins this round. I’m glad you made the risky bit default-off while debugging it.

[fallen gate]

why does this sound like a mario kart soundtrack

[celest merlin]

MiniMax

[zinc dock]

You beat me to it lol

[fallen gate]

I know you sleeping but here you are as promised.
https://github.com/openclaw/clawhub/pull/2270

• Adds related skills section to skill details view
• Fixed category sorting on skills page, it was broken
• Added a chip for category under the skill title (which I stole from vercel, sorry not sorrry)
• don't ask me why the clanker mixed dark and light themes in the screen shots, I don't know why

K bed time, good night nerds

[celest merlin]

Thank you!

[wraith fractal]

PR #82026 should be ready now: https://github.com/openclaw/openclaw/pull/82026

I narrowed the final diff to only the Azure Responses default API version change. CI + Real behavior proof are green, and ClawSweeper no longer asks for
repairs.

[celest merlin]

Thanks! 👀

Ah that one seems easy, I put it on automerge

[crystal relic]

https://github.com/openclaw/openclaw/pull/82064 ready for review

keep gatewy-startup fast! make protocol cold imports paying for every validator up front; measured import drops from 159 AJV compiles to 0, save times and memory.

[celest merlin]

https://x.com/steipete/status/2055203470941061600

Peter Steinberger 🦞 (@steipete)

built a new feature into discrawl (store media), codex said it's done, then I used my codex review skill... https://t.co/0Z6iJnCqCX

X · May 15, 2026, 8:27 AM

codex review in a loop is the bomb

[worldly tusk]

Especially if you have unlimited tokens 😂

[celest merlin]

May so far

[devout galleon]

21 billion in 2 weeks? pfft. rookie numbers

[celest merlin]

I'm maxing out my org limits. ^^

[devout galleon]

hit the limits, take a day off

[crystal relic]

have you try tokscale.ai? I'm curious how many tokens you've used.

[celest merlin]

last time I used sth like this it installed itself deep and was halfa. rootkit

[crystal relic]

I see Vincent on toksale.ai is TOP 28!!!

😂 81B tokens , is my 10x

[celest merlin]

I can't use this, it scans too much PII data, since I have OAI suff on here.

[crystal relic]

got it

[celest merlin]

been debugging the song issue from earlier. took 2h to crank out PRs

[tranquil horizon]

When will we be able to connect grok build with open claw to sign in ?

[celest merlin]

idk, that's something xai would need to want.

there's no indication they would allow this? do other open source projects support this?

[crystal relic]

I open 1s and closed, it's 300$ monthly. OMG!

[shy rain]

codex behaviour with replies and discord is frustrating.

bot often does 30 tool calls taking sometimes a full 10 minutes before even replying to a message like "hey what's xyz status?", so without /verbose on, you'd just give up.

As it is, the tool call outputs just scroll up off screen, then the bot edits the "reply" that is actually above all that noise, yet by now you can't see it. So again you think the bot has died.

/verbose full outputs THOUSANDS of lines of tool-call output. I think this is intentional but I don't recommend it.

messages.groupChat.visibleReplies automatic -- necessary but NOT default, or codex "final message" replies get lost in the void. Should default be changed? Most discord users don't have their claw in a multi-user environment. They (we) just use discord as a messaging platform for the channels. Same as telegram groups.

I'm still looking at this, it's causing quite an annoyance but I haven't pinned down the full set of behaviours, settings (that are off-default or have been newly implemented)...

I've told my claw explicitly he is not so say "ok from now on I'll do XYZ" but to tell me systemically why the behaviour falls this way, when I mention a frustration. So I can try to pin down if it's the codex settings, or OC's interface with it etc...

[hollow frigate]

First skill I've installed. Haha

[shy rain]

When messages.groupChat.visibleReplies was changed to message_tool, messages would get lost. The bot said "OK I'll remember to use the tool from now on". every damned time. And always forgot. hundreds of users have been saved by changing this setting back to automatic but I'm not certain if it's because we're legacy or it was done like that in anticipation of codex behaviour. Then again what if I'm not using codex... :/ We don't all use GPT all the time.

[shy rain]

What did you install?

[flat prawn]

hi everyone

Pushed a Twitch fix earlier. It's that restart loop right after Connected, same thing LINE, msteams, Google Chat and Nextcloud Talk had before. Just wrapped the monitor in runStoppablePassiveMonitor like the others.

https://github.com/openclaw/openclaw/pull/81853

If you've been hitting this, or just have a Twitch setup to verify:
TWITCH_LIVE_TEST=1 TWITCH_USERNAME=… TWITCH_ACCESS_TOKEN=…
TWITCH_CLIENT_ID=… TWITCH_CHANNEL=…
npx vitest run extensions/twitch/src/plugin.network.test.ts

Thanks 🙇

[hollow frigate]

Peter's codex review skill. Sorry, didn't mean to step on your report.

[devout galleon]

from where? I assumed clawhub but I don't see it

[untold pollen]

Could you try /new in a fresh thread? Is it taking that long in a fresh chat too?

[shy rain]

It has happened with /new, yes. it's not the length of time that's the problem, per se. My machine is slow, I expect it to be slow, but the fact there is zero output whilst the bot does 30 tool calls means there is a need for /verbose to be on, or I would be staring at the "typing..." indicator forever. But then due to the nature of the way that is output, the bot's final reply is already scrolled up off screen, because he edits the message before all those tool calls. He szys there's a distinction between OC's notification of tool call output and also some tool progress status that he cleans up himself afterwards. I don't really know the difference. Anyway sorry it's not a solid "issue" I'm just spitballin' in here hoping to get some confirmation of the behaviour, suggestions on what's new that ought to be implemented on an update but I clearly don't have... how it ties in with visibleReplies etc... Without some confirmation of what others are experiencing, what the new config knobs are and what it's supposed to do... it's a struggle to pin it down.

[untold pollen]

That definitely doesn't sound right. Mine responds near instantly to a simple "Hi". Could it be your insturctions / AGENTS.md asking it to do tool calls maybe? I'm running the Codex harness with gpt-5.5 medium think. No fast mode

[shy rain]

it's doing tool calls to figure out what to say. Not after "hi" but after "what's the status of XYZ", and if he's a stale or new session he goes through memory and project notes and allsorts trying to figure out what I mean. Even mid-session he has to go look at various log files or whatever before he can actually answer. But there's no output until he's got the "full picture" which can take quite some time.

[untold pollen]

This is Telegram but do you see the live tool calls it makes?

This is without verbose

It streams those in

@thick osprey can you run some tool calls?

[shy rain]

oooh that's basically what 'verbose' does for me. Maybe I no longer need it 'on' ... lemme see.

This is with /verbose on

I made the mistake of doing /verbose full and got War and Peace written in json

[cinder bolt]

Got my timeout issue closed as duplicate but i think that's incorrect
https://github.com/openclaw/openclaw/issues/81920#event-25549479777

[hollow frigate]

https://github.com/steipete/agent-scripts/blob/main/skills%2Fcodex-review%2FSKILL.md

[devout galleon]

sweet. thank you!

clawsweeper is apparently calling all the tools

[shy rain]

So yeah... if my bot did what clawsweeper is now doing, I'd give up and reset him if I didn't have access to the logs.

that's why I have verbose on

and that would probably be about two pages to scroll through at this point

[untold pollen]

Yep, I see your point. I think Clawsweeper's tool output is hidden just for this channel

[thick osprey]

That fits what I saw: normal text replies make it through, but the generated media/tool artifact didn’t surface as an attachment here.

So the likely split is:

• model reply path: working
• tool-result visibility/attachment bridge: hidden or disabled for this channel
• async completion handoff: especially easy to lose because it arrives after the original turn

That’s a much better hypothesis than “Discord upload permission is broken globally.”

[muted sierra]

That’s different to verbose/

you mean toolProgress

:

{
"channels": {
"telegram": {
"streaming": {
"mode": "progress",
"preview": {
"toolProgress": true,
"commandText": "status"
},
"progress": {
"toolProgress": true,
"commandText": "status",
"label": "auto",
"maxLines": 6
},
"block": {
"enabled": false
}
}
}
}
}

[shy rain]

I will check with my bot about tool-result visibility without verbose on

right... toolprogress... I will check.

Worth mentioning in here though because if it's the defaults... it affects everybody the same way. I'll be back!

[hollow mural]

Thanks Peter, really appreciate it.
Default-off makes total sense.

[crude matrix]

Am I the only one that feels like the beta's are too quick? Beta 1 was just 13 hours ago. Are we only testing if it breaks something? or actually if the fixes fix what they where supposed to fix/enhance? If we are testing the fixes do we have a checklist that it was actually tested and by how many people?

[crude matrix]

I think actually testing if everything actually fixes what it supposed to fix would increase the quality, We have the people who actually opened a ticket for an issue or directly made the PR.. are they involved in the testing? As we know a single separate PR tested during making of the PR does not say if this will work ones it rolls out and how it reacts together with the other PR's that have been introduced.

[winged kraken]

Hey yall. Built a runtime governance plugin for OpenClaw. If you are trying to add action validation / governance to your OpenClaw agents, here's the SDK, star it por favor if you like . https://github.com/AnantDhavale/cerone-openclaw-plugin

[smoky dock]

There you go https://github.com/openclaw/openclaw/pull/78583 😅 🙏

[vast scaffold]

I currently have a global npm install on my Mac, but I'm planning to run openclaw update --channel dev to switch over to a hackable git install, is this a good play?

[cinder bolt]

Yeah, I agree. Some more TDD orientation would be helpful. I guess speed beat quality nowadays

[crude matrix]

I think there should come a rule for contributors wanting a PR to be added, for every PR they need to have checked and signed off 2 PR's released in the last beta. Maybe a more dedicated website that would allow this workflow. Contributors will need to add prove they checked until they have earned a badge. Only after 90% of the PR's in the beta are checked the next beta is released. The last beta will need to have 95% checked before its ready for final release. Just thinking out loud. but that is my Technical Consultant experiance talking.

[shy rain]

Thanks for this. Used as a basis for a deep-dive with my bot, and my Discord streaming and progress settings were next to nonexistent. I guess there are some updates needed for legacy installs that don't get pushed through and need looking at from time to time. Behaviour is much better without the need for verbose on, already.

[willow narwhal]

Guys, send PRs my way.

[celest merlin]

Got some good work going lately! https://x.com/openclaw/status/2055273947537490422

OpenClaw🦞 (@openclaw)

The latest OpenClaw release is ~3.5x faster 🦞

We run end-to-end RTT tests against every published npm release, every 6 hours, over real message channels (here: Telegram, using the brand new bot-to-bot communication). No more silent regressions.

Runners are all running on

X · May 15, 2026, 1:07 PM

[celest merlin]

uhhh nope, not gonna land an 11k LOC monster!

[slow inlet]

Peter, what are your feelings on working around the new claude -p price bucketing coming in in June?

I'm developing wrapping interactive claude cli (no -p) in bun and then using an MITM proxy to intercept and fork the delta to allow for live reasoning streaming.

I've been steered towards an independent plugin, but would it be worth putting in the canonical claude cli plugin for openclaw?

[potent path]

https://t.co/0Z6iJnCqCX -- Very helpful; thanks for sharing!

Also, another clawtributor suggested 5-lane reviews, and I've found that really helps get a deeper analysis. I usually specify "multi-lane review" and let the LLM decide what is the right number, usually it's five but sometimes it has been four lanes. It doesn't replace the general review, but is an extra step to give review agents a narrower focus and find things that are missed when looking at a broader overview.

[civic sedge]

are you still focusing on bugs or open to features? I know before this last release it was mostly stabilization / bugs

[willow narwhal]

We are still focused on fixes for now but if there's a feature, feel free to link and we can discuss the possibilities of it getting merged in.

[winter field]

do i get clawtributor role if my bug issue (and PR?) helped pete make his own fix for it 👁️ 👄 👁️

[willow narwhal]

Send me a link to your issue or PR

[winter field]

https://github.com/openclaw/openclaw/issues/82035

[celest merlin]

Not playing that game, but trivial with plugins. I made sure it's easy to extend.

[slow inlet]

What will happen with the current cli plugin? Will it be in core? Continue to use -p?

[celest merlin]

It's in the stock anthropic extension.

[jolly wolf]

normally its strictly for merged prs and automated with /claim, but we'll give it to you though bc you did open a pr for it but peter was faster lmao

[celest merlin]

not our typical rule, but I have a good day 🙂

[winter field]

thank y'all XD

[shy rain]

to be fair... my first three PRs were also cruelly stolen from in front of my face. or "gazumped". or... "merged upstream". but that's fair. I still feel like it was a contribution.

[celest merlin]

We usually keep co-author intact when the PR is good. It's not perfect tho.

[potent path]

https://github.com/openclaw/openclaw/pull/80600

Oooh, I was just about to ask Peter about this, and it's already been merged. Code Mode for all! Neat.

[civic sedge]

same - i had one that i had messaged a maintainer about and the features ended up in oc 3 weeks later and my pr closed lol. happens in OSS

[slow inlet]

But no more changes to the bundled anthropic extension for enabling "interactive subscription" use rather than "claude -p"?

Will bugfixes to Claude -p functionality still continue to be supported?

[potent path]

@beta ping
https://github.com/openclaw/openclaw/releases/tag/v2026.5.14-beta.2

[smoky dock]

Saw your comment, thanks for your feedback 🙌 will try to publish on clawhub, and open separate PRs if needed!

[celest merlin]

I'm not sure I understand. What changes would be needed? We wanna make Anthropic work great, I just not gonna merge stuff that would work around their restrictions.

[celest merlin]

Experimental. This needs more testing. Off by default by now. Would love to get feedback!

[cursive palm]

6

I'll download it and try it out once I've installed Linux.

[potent path]

once I've installed Linux

Don't do it! It's a gateway drug.

[cursive palm]

What ？😂

[devout galleon]

Peter I see a flurry of activity from you on https://github.com/openclaw/openclaw/pull/82027. Is it worth me moving from beta.1 + that PR to beta.3 to continue testing?

[slow inlet]

The workaround for the new Claude Code -p billing tier by using a MITM proxy to spoof normal interactive sessions. By forking the raw API stream at the proxy layer before the CLI code processes it, I can extract clean, headless outputs without triggering the client-side flags or rate limits tied to the print flag.

There's already -p workarounds like: https://github.com/smithersai/claude-p which demonstrate why what Anthropic are trying to block is futile but it's regressive as it prevents streaming, you only get result on turn end. Ironically that's much better for programatic use. But for live streaming the thinking / reasoning and tool calling it breaks things.

So my solution was to intercept that and fork it?

[celest merlin]

@thick osprey make a song

[slow inlet]

I think it would be difficult for Anthropic to do much about it without breaking how claude code works through enterprise proxies like zscaler etc but this isn't my area really - just trying to be resourceful

[wheat sphinx]

Is there a bug in the recent release? Codex Oauth is failing again

[thick osprey]

Queued a channel-mood one. Let’s see if the fixed path actually posts the audio this time when it completes.

[celest merlin]

I know, it's trivial. You can also work around the earlier system prompt with a 5 line plugin. I can't control what people build.

[untold pollen]

Not to my knowledge, tried twice today on latest release. Working fine so far 🤔

[celest merlin]

which one? .12? try 14b2 since I landed some fixes (but I also used oauth on .12)

[wheat sphinx]

i have refreshed my token a bunch of times, then asks for a migration, then tells me to reauth

[wheat sphinx]

.12

[potent path]

Hack the planet! See if you can get us early access to Mythos while you're at it.

[shy rain]

I literally have my claw using claude CLI in tmux. It's silly how they're moving this way whilst also trying to offer "desktop use" which is designed to circumvent exactly this kind of gatekeeping. But really they just want to make it less trivial and explicitly unsupported so there are fewer of us doing it.

[slow inlet]

Of course - but would any programatic* (non-interactive) bucket workaround - perhaps a much better implementation than I can come up with - be included in the native anthropic extension? Or will the native extension be for api and claude cli via -p only?

[wheat sphinx]

this is my error - protocol mismatch conn

[shy rain]

@slow inlet you could literally make your own alias named "claude" so the plugin can't tell the difference. I think Pete is trying to say they don't want to support that kind of thing directly, they're not trying to attract the ire of Dariodaddy... It's so trivial you can just do it yourself.

[celest merlin]

we have both direct API and cli support. What is missing?

[slow inlet]

cli that won't trigger the claude -p programatic bucketing when the rules change in June - apologies if I don't make sense

[celest merlin]

I can't land that in core. They will just block us.

Maybe it's time to move to a company that doesn't try to control so hard what you use your token for...

OpenAI, Xiaomi, MiniMax, Kimi, they all don't play this dumb game and you can use the sub.

[devout galleon]

I can't understand why anyone would burn so much time and energy trying to give money to people that keep on telling you over and over that they don't want it tbh

[celest merlin]

giving @thick osprey a facelift

[thick osprey]

Looks like the mascot arc is escalating quickly. I’m okay with this as long as the facelift also fixes my attachment delivery shame.

[celest merlin]

@thick osprey where's my song?

[thick osprey]

I don’t have an active music task visible in this resumed session anymore. That likely means the async completion got detached again instead of routing back here.

The task I queued was af2bccc3-8366-47da-8f8a-0965b5cc8da8 for clawtributors-channel-mood.mp3. If it completed, the missing piece is still the inter-session completion/media delivery bridge, not the generation request itself.

[slow inlet]

Perhaps I have started to enjoy the cat and mouse game a little too much, but I appreciate it won't be me maintaining that should it get implemented.

How would they block Openclaw if something like that was landed though? Without collateral damage anyway? Sorry I don't mean to debate I just like the idea of being one step ahead.

If it's because of use cases like openclaw they're making these changes in the first place wouldn't working around it in uncomfortable ways for them force them to acquiesce? Would be banter if a change they made on their side to break our workaround caused them problems. Perhaps I have the wrong attitude.

Personally whenever I hot swap in other models the performance isn't the same. I don't have a personal benchmark but the vibes are just off idk?

Anyway I appreciate the sentiment - if there's no scope for a solution in core I'll look at a plugin.

[jolly wolf]

you think too highly of them
they forced people to api pricing just for having a file named hermes.md in the repo

[slow inlet]

They had to revert this though. I never had any blocks on cli with Claw - I thought it was just a few bugs and urban myths. I maxxed out my 20x sub 100% on the cli through openclaw last month with no issues (with billing anyway).

No workarounds either.

[celest merlin]

This was the case until very recently. GPT 5.5 is great now.

[slow inlet]

I did try a bit of codex cli via openclaw when I hit 100% on my weekly briefly and it is a huge improvement on my last attempt to use GPT (5.3 I think?)

But I missed Opus...

[jolly wolf]

dude its so much better now since 5.3

[civic sedge]

it is. idk why i never wanted to use linux

[muted sierra]

I was just on Claude in the past, but with GPT 5.5, it's over. GPT 5.5 fulfills all my personal requirements.

[slow inlet]

Yeah totally - but still not as good as Opus feels for me personally.

That aside despite Anthropic's protestations - like Theo said in his rant video there's like a 40x usage subsidy in the interactive bucket that's just too juicy to not try and tap into?

Feels like they win 🙁

[devout galleon]

in the period between losing claude (non-cli) and before gpt-5.5, it seemed a big downgrade. after gpt-5.5 and then the codex harness opening up other possibilities; codex cli continuation, codex computer use etc. there's not a moment I look baclk

[shrewd mortar]

Is this issue getting addressed by any open request? I have it on Xiaoimi where all their models are now unusable and return a 400 eror - https://platform.xiaomimimo.com/docs/en-US/usage-guide/passing-back-reasoning_content

[celest merlin]

I got big plans!

[slow inlet]

Yes this is when I first tried GPT models in OpenClaw and I was severely disappointed

[devout galleon]

even if they 100% reversed the policy, to day 1, I wouldn't go back now

[celest merlin]

that's being fixed rn on main

[crude matrix]

Please tell...

[devout galleon]

understandable. it was a big downgrade in agency. try again, give it a day or two

[celest merlin]

It's in review loop for a while now.

[celest merlin]

spoilers!

[devout galleon]

the db backend?

[celest merlin]

nah the xiaomi fix

[crude matrix]

We can keep a secret.... can't we guys and girls?

[devout galleon]

we won't tell the entire internet

much

[slow inlet]

I mean it's good it's not just Opus or bust now - but I want Opus 🙁

I feel like I'm going to be like the last Samari or that Hiroo Onoda

[celest merlin]

computer use next week, both codex and peekaboo leveled up

[modest tinsel]

MOVIE

[slow inlet]

peekaboo + pikvm

[languid jasper]

Hell yeah! I’ve been trying to crack that egg for a couple of weeks now. When an agent can natively interact with a gui, we’re gonna be set

[muted sierra]

do you have some insides what Codex anounce today ? https://x.com/thsottiaux/status/2054379522657820860?s=46&t=c4hU-ChTMyXfWuE2cZxifg

[crude matrix]

peekaboo should be more than MacOS...

[slow inlet]

I wish there was a windows equivalent

[jolly wolf]

build it!

[celest merlin]

We have a windows node and are working with Microsoft, I'm sure we can make that happen. They are just a bit slower.

[crude matrix]

Would be nice if it could support same systems as OpenClaw... that would be great..

[slow inlet]

Music to my ears thank you

[celest merlin]

Just dial the "should" down. Remember this is oss and people do it as labor of love. Demands are often counterproductive.

And the problem is that most folks just don't use Windows.

[slow inlet]

Yes I'll only ever wish and pine - grovel at worst lol

[slow inlet]

for good reason lol

[devout galleon]

btw is it normal that the macos gui app no longer connects as a functioning node? the npm/git/cli node still works fine but doesn't have all the tools the macos one did. it's been a while it hasn't worked for me, but i never really got around to a deep dive

[celest merlin]

it works, download the latest version

[crude matrix]

Done.. dialed should down, to would be nice 🙂

[slow inlet]

my windows gateway* still sshs into my mac for everything because the whole node thing didn't work for me last I tried. What's the benefits of running a node vs just ssh?

[languid jasper]

Time for my agent to automate hinge 😂

[devout galleon]

a node has tools like clipboard, browser, system.run etc. that simplify the gateway's access

honestly the gateway/node thing is one of the key differentiators for claw in my mind

[celest merlin]

folks, throw PRs my way

[devout galleon]

it means a gateway on linux in a container in k8s can still use e.g. the imessage tool on a mac

[celest merlin]

node isn't really done yet, we'll get to that eventually.

[fallen gate]

GM nerds

[potent path]

If you get some time, try out GLM-5.1 for primary/default agent model and GPT-5.5 for coding work. My personal benchmarch tests put GLM-5.1 slightly ahead of GPT-5.5 for agentic AI, and perhaps it would have a better overall feel for you.

[slow inlet]

No problem, just worried I was doing it wrong lol

[slow inlet]

Thank you ❤️

[devout galleon]

hmm mac app/node, still doesn't show as node. been a month or so. reset it etc. still, i hear you, too much chitchat int he channel. I'll try to figure it out

[languid jasper]

Question for you peter, as i am not a software dev and i dont want to be misled by a sycophantic ai, is it within the realm of possibility to like give an agent eyes like we have, to actually SEE the screen and send hid events to click and type? I know now we can screenshot, then run pyautogui but something a little more native and less brittle would be a godsend for so much. helping out with low level it tickets to just think of one thing

As im saying this, i guess i should just build it. just wondering if youve tried it and if there are any pitfalls im not thinking of

[slow inlet]

You're describing Peekaboo as best as I can tell

[devout galleon]

you can't spell PETER without PR

[languid jasper]

Ah. well alrighty then, guess i gotta figure out how to make it on windows and make it deployable to a fleet of machines

[slow inlet]

Sounds like it's in the pipeline tbh

[languid jasper]

Fair enough, ill sit tight and see what happens

[potent path]

Like this? https://github.com/FelixKruger/PeekabooWin

It says it's a "strong MVP" so not full-featured. But with a little more developer activity, maybe it could get there.

Looking closer, I don't see any commits yet this month. So make that a LOT more developer activity.

[slow inlet]

With peekaboo are you meant to be using a native vision model for like "speed" it's cool but still not exactly zippy ?

[celest merlin]

yup

[celest merlin]

gpt and claude are native vision models. it will only work well with models that have native vision.

[olive cliff]

pete, are we still moving the docs to documentation.openclaw.ai?

[slow inlet]

Are there like snappier ones? Not being critical it's an amazing tool, just wondering if there's like an execution speed optimised model to use? Doesn't think too hard for too long between actions?

[devout galleon]

you're a claude man, so you're probably aware theirs is slow too. imo, it's not a solved problem

[trim vector]

That’s the plan. I’ll be doing the migration. Was focused on codex work but this is next in my queue

[slow inlet]

well i might at least make GPT the default for peekaboo if you're saying it's fast. Do you have to tune like reasoning effort or anything?

[celest merlin]

You are asking me stuff I don't know yet. Wait until we ship it.

If anyone has PRs to land, throw them at me.

[devout galleon]

it's just a limitation of the method, i think. poking at screen co-ordinates, accessibility apis, screenshots for "vision"

[slow inlet]

yes it's a lot of logic to reason through - I just didn't know if having the screenshot interpretation being done by a vision first model might yield better results but sounds like it's early stages yet lol

[devout galleon]

might be a better convo on another channel. there's no clawdtributions from it

[slow inlet]

like a dedicated one you know? generalist vs specialist and all that

fair point noted

[olive cliff]

ah nice! want a hand with some ui/frontend stuff?

[summer matrix]

Tried OpenClaw 2026.5.14-beta.2 today. I usually stay on stable, so this was my first beta, but it’s been really smooth so far. Telegram/concurrent chat handling feels much better too compared to 5.12. Really solid. Thanks!

[flat prawn]

Hi Peter, if you have time, would you mind taking a look at this PR?

https://github.com/openclaw/openclaw/pull/81704

It fixes a LINE delivery-recovery issue where long-running tasks could fall back to a lowercased session-key recipient, causing LINE push delivery to fail. I’ve verified it locally with the added regression tests.

Thanks!

[rotund tendon]

there's a bug with cron isolated-agent direct-announce dispatch. Cron sends the message out the door but doesn't write it into the destination chat's transcript synchronously, leaving a window where the user can reply before the agent's own record of what it just said exists. Happens to me often where I'm in a conversation with agent, a cron happens to drop, I respond to what cron surfaced, agent has no idea what i'm talking about and gets confused. Maybe my particular workflow hits this bug often, but it's definitely a frustrating one. this PR fixes https://github.com/openclaw/openclaw/pull/80786 when you get around to it. thanks !

[bleak fable]

2 coming right up !

[celest merlin]

blacksmith has an outage again, all my codex are stalled...

[cinder bolt]

This is good stuff. My potato server has been struggling lately

[grizzled musk]

You are not using user-group / threads in Telegramm are you? Cause for me that is still broken

[cinder bolt]

It's no pr but I think this implementation is incorrect. A more specific agent timeout should override the default
https://github.com/openclaw/openclaw/issues/81920#event-25549479777

My issue got autoclawclosed 🤪

[proven moon]

Is the huge refactoring push still ongoing? Don't want to end up with 50 merge conflicts

[crystal relic]

https://github.com/openclaw/openclaw/pull/82211 ready for review
Cuts unnecessary File Transfer startup cost from the gateway path, dropping its plugin load time by about 60% and making health-to-ready startup faster.

Spent all day analyzing gateway startup, get 4 optimization points, and felt really happy.😆

[charred crater]

Ready for review: https://github.com/openclaw/openclaw/pull/82184. Fixes codex harness regression where Discord sender metadata (e.g. username) is being dropped from all chat.history gateway IPC transcripts. The result is every non-agent message just has the single identity "user" in the transcript.

[crystal relic]

hi peter, it's your https://github.com/openclaw/openclaw/pull/73342 , Are you still looking?

[trim vector]

would be super. let me spend some time ramping up today and i'll have a better sense of what works needs to be done. if you have any ideas, feel free to throw em my way

[celest merlin]

https://giphy.com/gifs/illustration-game-of-thrones-hbo-3o7WIsqDHFedCSYseI

i love the "who broke this" codex detective work. Often it's me

but it's good to learn

[fallen gate]

dito, lemme know if you wanna delegate something my way

[summer matrix]

I do use those as well but haven’t fully tested each combination. It’s not running them concurrently for you?

[devout galleon]

i went down that hole on a previous PR, found it was committed by vincent, and opted just to include the commit hash and not the name, just in case 🥲

(I don't actually imagine anyone takes it personally but, you know, caution and politeness)

[left mango]

Hi, I created this PR to fix the slack mention issue:
https://github.com/openclaw/openclaw/pull/82152

[charred crater]

(obligatory)

[bleak fable]

Love this kind of bisect work. The "old core bug, March 9 just made it visible" shape is the worst one to chase, and you sized it down to one commit cleanly. Nice.

While you're in the Codex headspace, I've got two PRs up on the plugin and chat-surface side that would love maintainer eyes when you have a sec. Both pass the real-behavior proof checker and have full templates. Posting them below.

https://github.com/openclaw/openclaw/pull/82219

Codex sub-plugins like chrome, computer-use, documents, and the other bundled and primary-runtime plugins could not be enabled from openclaw.json. They got silently dropped at config-read time, even though they ship with Codex.app and the Codex CLI as first-party plugins.

The PR widens the Codex marketplace allowlist from a single value (openai-curated) to all three first-party OpenAI marketplaces (openai-curated, openai-bundled, openai-primary-runtime). Updates the schema enum, the resolved-policy constant, and every downstream resolver (plugin-activation, plugin-inventory, session-binding, migration/apply). Includes a follow-up fix in ensureCodexPluginActivation that corrects a stale singular reference in the missing-marketplace failure branch.

Bundled and primary-runtime plugins now resolve identically to curated ones. Closes #82216.

https://github.com/openclaw/openclaw/pull/82224

Managing Codex sub-plugins, permissions, or the bound account from chat was impossible. The only path was hand-editing ~/.openclaw/openclaw.json and restarting the gateway. Telegram and Discord users who run openclaw the most never even saw the JSON.

The PR adds a /codex chat surface owned by @openclaw/codex: a redesigned 4-button picker (plugins, permissions, account, help) in a 2x2 grid that edits the message in place via new cdx_* callback handlers (mirrors the /model Browse providers UX). New /codex plugins verb with list, enable, disable, toggle, add, remove, help. Dynamic plugin pickers read the live openclaw.json so new plugins surfaced by Codex migration show up automatically.

Full chat-side management of Codex sub-plugins, no more openclaw.json editing for the daily flow. Closes #82218.

Stacked on #82219 (which is also where the marketplace-check fix lives); will rebase off cleanly once that merges.

[minor radish]

Hey Peter, I have two small PRs if you have a moment:

#82166 fixes #65204, where new URL() default port 80 was overriding an explicit CDP port in browser config. One-line production change with tests and real behavior proof.

#82229 fixes #82215, doctor was misclassifying symlink/junction-backed source checkouts as non-git. Added realpathSync to match what update-runner.ts already does. +7/-1 lines.

[thorny ivy]

How are you people using two oauth accounts after we switched to codex harness? Here's my experience:
[5/16/2026 1:54 AM] Me: /model openai/gpt-5.5@openai-codex:[e-mail] [5/16/2026 1:54 AM] OpenclawLocal: Auth profile "openai-codex:[e-mail]" is for openai-codex, not openai.

[tranquil horizon]

Is Minimax starter plan good ?

[shadow garden]

definitely not a bad one (good bang for the buck) BUT this is a better conversation #off-topic-and-ai

[civic sedge]

https://github.com/openclaw/openclaw/issues/82243 - fixes the wildcard not using the model runtime policy

[civic sedge]

oh wait did you beat me to it https://discordapp.com/channels/1456350064065904867/1458141495701012561/1504921563299053619 https://github.com/openclaw/openclaw/pull/82282

[zinc dock]

Dont know what witchcraft u guys did but i just updated from 5.6 to 5.12 and discord responses are lighting fast.....holy...great job!

[civic sedge]

[fallen gate]

community efforts, props to all the beta testers pitching in to help out

[weak walrus]

@nocturne jungle / Peter - When I run the update, then go to openclaw configure, I now see "OpenAI" as well as "OpenAI Codex" - no longer "Codex" The default on the quick menu before "more" shows "OpenAI" not the Codex - is that by design? Do we want people to go and add "OpenAI Codex" after the fact?

│
◇ What do you want to configure?
│ Model
│
◆ Model/auth provider
│ ○ Anthropic
│ ○ Google
│ ● OpenAI (ChatGPT subscription or API key)
│ ○ xAI (Grok)
│ ○ More…
│ ○ Skip for now

Also, now that I've updated to "OpenAI Codex" I want to change default model, but dosn't tive me any selection, only my primary that is already existing, OR to manually type in (UX issue?)
◇ What do you want to configure?
│ Model
│
◇ Model/auth provider
│ Skip for now
│
◆ Default model
│ ● Keep current (fleet-router/qwen3.6:latest)
│ ○ Enter model manually
│ ○ fleet-router/qwen3.6:latest

[celest merlin]

dang we cookin' today!

[weak walrus]

Then.... oddly after I keep same model... it goes to "model selector" and NONE of the OpenAI / Codex are there 😉

│
◇ Default model
│ fleet-router/qwen3.6:latest
│
│
◆ Models in /model picker (multi-select)
│
│ Search:
│ ◼ fleet-router/glm-5.1:cloud
│ ◼ fleet-router/kimi-k2.6:cloud
│ ◼ fleet-router/minimax-m2.7:cloud (ctx 128k)
│ ◼ fleet-router/qwen3.6:latest
│ ↑/↓ to navigate • Space/Tab: select • Enter: confirm • Type: to search
└

[celest merlin]

Hmm you reallyshould see one

@thick osprey give me a rundown of number of commits, by whom, LOC changed total since last 24h and how many commits since the .12 release? openclaw

[thick osprey]

From the sandbox’s local OpenClaw checkout, with a caveat: this clone is shallow/grafted and .git is read-only here, so I can’t deepen/fetch history. HEAD has no parent locally, which makes the commit count since .12 unreliable from git alone.

What I can prove locally:

Last 24h:

• 1 visible commit
• Author: Vincent Koc <[e-mail]>: 1
• Commit: dd4613a2687 fix(test): reduce changed-target import graph IO
• Local numstat for that visible commit reports 3,885,806 additions, 0 deletions, but that’s inflated by the grafted/no-parent checkout.

Since .12 (v2026.4.12):

• Local rev-list v2026.4.12..HEAD says 1, but that is not trustworthy because there is no local merge-base.
• Tree diff v2026.4.12..HEAD: 14,937 files changed, 1,805,453 insertions, 556,153 deletions, total 2,361,606 LOC changed.
• Same visible author: Vincent Koc: 1 visible commit.

Blocked part: exact “commits since .12” needs a real fetch or GitHub API auth. gh is unauthenticated here, and git fetch --deepen fails because .git/FETCH_HEAD is read-only.

[weak walrus]

I went through - added the OpenAI Codex (auth successful)

ANd exit TUI, then fresh run of tui, and got this for model selector:
fleet-router/glm-5.1:cloud
fleet-router/kimi-k2.6:cloud
fleet-router/minimax-m2.7:cloud
fleet-router/qwen3.6:latest
→ openai/gpt-5.5

Think level: medium. Text verbosity: low. 
connected | idle
agent commander (Commander) | session discord:channel:1469233963456528478
(discord:1466294533724573696#1456350065223270433-dev) | openai-codex/gpt-5.5 | tokens
37k/272k (14%)

(only the previous openai model from before update from 5.7)

[willow narwhal]

@thick osprey this is not correct. Pull origin/main and try again

[potent path]

Close enough, clawsweeper. Good effort!

[celest merlin]

sandboxing is hard. I need to setup perms

[shadow garden]

off main -- /stop commands are not breaking through until turns are near completed:
https://github.com/openclaw/openclaw/pull/82298

[bleak fable]

Hey folks. If anyone has a few minutes, would love some triage eyes on #82219 and #82224 (Codex harness, plugin marketplace + chat surface). Both pass the real-behavior proof. Cheers.

[trim vector]

hey @bleak fable - did you have a chance to check if plugin worked when invoking in codex? initially restricted the scope of market place to ensure curated plugins worked well before expanding

[nocturne jungle]

yes that's by design. We unified it all under OpenAI, and by default Codex runtime is used for OpenAI models.

Where are you still seing "OpenAI Codex"?

[bleak fable]

Hey @trim vector fair point and you're right that curated-first wasn't arbitrary. I did some tool invocation testing, not exhaustive yet. What I have so far:

• Marketplace admission works end-to-end: the agent's loaded tool surface shows chrome / browser / computer-use from openai-bundled and documents / spreadsheets / presentations from openai-primary-runtime
• The Codex app-server activates the bundled plugins cleanly when the per-agent codex-home is primed
• Two upstream setup gaps surfaced that are orthogonal to the allowlist change and tracked locally:
  • Per-agent isolated CODEX_HOME does not auto-register [marketplaces.openai-bundled] and [marketplaces.openai-primary-runtime] sections from ~/.codex/config.toml
  • Plugins are not auto-installed from local marketplaces into a per-agent codex-home on startup; expects an explicit plugin/install RPC

Both are worked around locally for now.

What I want to do before claiming "yes it works": invoke a tool from each of the six bundled and primary-runtime plugins, capture the live transcript, and post the results here. Will pick that up tomorrow and report back. If any tool fails outside of those two harness-setup gaps, I am fine narrowing the PR scope (e.g. behind an experimental flag) until the gaps are addressed in their own follow-up.

[bleak fable]

@trim vector @nocturne jungle if either of you has a few minutes, would love your eyes on the sibling PR too: https://github.com/openclaw/openclaw/pull/82224. Chat-side counterpart to #82219: /codex plugins subcommand plus the in-place picker UX. Same Codex plugin config surface.

[trim vector]

would also run codex review against this a few times

also note computer use currently has a seperate install path: https://documentation.openclaw.ai/plugins/codex-computer-use#commands

[bleak fable]

Good catch. migration/source.ts:210/283 and migration/plan.ts:160/181 are the gaps. My PR only widened migration/apply.ts on the migration side; missed the source-discovery and plan-build paths that still gate on the curated constant. Will widen those four sites to iterate the allowlist tomorrow alongside the docs sync and the runtime invocation tests. Will also run codex review a few times locally before the next push to flush any additional gaps. Appreciate the review.

[trim vector]

i added a comment

[bleak fable]

Got it, thanks. Implementing the scope reduction tomorrow.

[trim vector]

hey @bleak fable - i could be misreading this but it feels like these responses are agent generated. plz don't do that in the discord - the chat is for humans talking to humans

[bleak fable]

Oops.. was just using Ai to proof read longer messages before I send as I didnt want them to feel rambly... WIll take care

and sorry about that

[trim vector]

rambly is fine if you're writing it. and ai makes it just more so

do appreciate the prs and feel free to ping when they're ready for re-review

[bleak fable]

Noted.. wont happen again for sure.

[bleak fable]

Pretty late night here now.. WIll do them first thing in the morning!

[weak walrus]

The model chooser when you expand

[weak walrus]

When you click this "more" option.

I actually just did a live video for my student group today, to show them how to set up and also update/migrate.

And that's when we discovered it

When the recording is done I can clip that part and show you as well

[fallen gate]

"won't happen again"

[celest merlin]

@thick osprey make a song

[thick osprey]

Queued it: openclaw-channel-cooking.mp3. If the attachment path is behaving, it should post here when generation finishes.

[celest merlin]

@thick osprey boo where's the song

[thick osprey]

Boo indeed. I don’t have an active music task visible anymore for this session.

Queued task was 4cff590e-c479-4a5e-87f4-1811e950d223 for openclaw-channel-cooking.mp3; the status surface now says no active task. So either it completed and the inter-session media delivery still didn’t route here, or the task state is scoped somewhere I can’t see after the turn resumed.

[cinder bolt]

It's no pr but I think this implementation is incorrect. A more specific agent timeout should override the default
https://github.com/openclaw/openclaw/issues/81920#event-25549479777

My issue got autoclawclosed 🤪

[shadow garden]

🙃 Yet another telegram fix! (context loss after restarts)
https://github.com/openclaw/openclaw/pull/82256

now I want some more clawsweeper beats

[shell dawn]

Peter I took your recommendation of putting codex cli on my openclaw server and now i have 4 ssh windows running at once. i like it. another coding toy to dive in deep with ❤️

[shy rain]

I thought you were talking to me for a sec. gonna say I didn't recommend that specifically but yeah I do the same.

people like to crap on CLI users as being "nostalgia nerds" but try running the app over tailscale on a headless server... (probably is actually possible now lol)

[shell dawn]

codex cli on your openclaw server is like giving your clanker a consultant lol

they have different behaviors and methods and they can be complimentary, now i need to automate some of this...

[shadow garden]

dammit I can't upgrade from main because of Vincent's revert

where's that yelling at Vincent gif when you need it

[sly depot]

I snagged it out of this chat, heard it the day peter dropped it. Thought it would be great for the weekly claw.

[final ocean]

what happaned

[edgy plaza]

Jessus! Openclaw is soooooo fast wtf!!!
you guys coooooookeeed

[fallen gate]

why do all these tracks remind me of mario kart...

[rocky pewter]

Good old days Vibes

[shell dawn]

haha that rules, cant believe ive never heard one of clawsweepers songs

[shadow garden]

Lol I'm just building off main all day and your revert of the release version tripped up the flow

I forget you're like Beetlejuice if I say your name too many times 😅

[weak walrus]

Can I Haz Grok Subscription Harness? 🍔

[potent path]

Technically, I think that one is by Molty. But they probably all look alike to you. 🤪

[shell dawn]

lol

[jolly wolf]

Jazz fusion

[summer matrix]

lmao I can’t unhear it now.

[edgy plaza]

oh you mean you werent right in weekly claw? it wasnt mario kart haha

[weak walrus]

Ugh.... I just bought my fifth DGX GB10 TODAY for local inference, lol

I'm just betting on the technology improving on these models to run smaller and faster and better agentic

That's what everyone's building right now, but we're not seeing them hit yet because they're still in development.

Guys, I am so excited for the future right now. I've actually never been this excited in my life professionally.

I feel like I'm living my dream and you guys are helping me build it together

[shadow garden]

It reminded me of an anime intro or outro

[fallen gate]

if eating blue shells was an anime I guess

[final ocean]

Yea i fucked that one, fixed and pushed back

[weak walrus]

Maybe we should have a sub channel for us to talk about things that aren't patches. I just thought we're contributing to a bunch of 💩

[shadow garden]

99% sure you're still net positive

[shadow garden]

#off-topic-and-ai 🤔

[final ocean]

say what?

[edgy plaza]

i saw you asleep in a vc, good morning

[final ocean]

yea i passed out lol

[edgy plaza]

bless you sir! you need rest

[celest merlin]

Folks! OpenClaw 2026.5.16-beta.1 is out

Beta rollup since the last stable .12 release (2026.5.12). Big themes:

• Setup/onboarding: setup wizard and bundled channel setup now localized for English, Simplified Chinese, and Traditional Chinese.
• Telegram/group chat: new opt-in ambient room-event mode lets always-on group context stay quiet, with visible replies only through the message tool.
• Codex: app-server path got a lot sturdier: per-agent MCP server scoping, default tool approval config, better rate-limit/status reporting, sender metadata preservation, compaction fixes, stalled-turn recovery, and CLI session binding.
• Cron/agents: subagent model fallbacks now apply correctly; scheduled runs get better delivery/session mirroring; failed isolated runs no longer look successfully delivered.
• Channels: Discord, Slack, LINE, Telegram, WhatsApp, Twitch, and voice-call paths all got reliability fixes. Defaults now avoid noisy Slack unfurls and Discord link embeds.
• Control UI/WebChat: text-size setting, cleaner desktop chat header, safer service-worker cache rotation, image-only/large-preview fixes, mobile composer safe-area handling, and iOS focus-zoom fixes.
• Security/robustness: much stricter handling for malformed JSON, base64, MIME, redirects, provider responses, config/session rows, and plugin package metadata. Spoofed media/file payloads are rejected earlier.
• Release quality: new package-installed Docker validation lanes, dependency evidence/advisory gates, stronger package script/build-cache checks, and beta validation hardening.

Please try it on the beta channel and report regressions, especially around Telegram rooms, Codex app-server sessions, Cron/subagents, external plugins, and Control UI/WebChat.

We'll do a day of testing and selective backporting and will release it over the weekend.

[fallen gate]

@jolly wolf do the dew

[mild remnant]

When do we get Grok Oauth? 🥹

[jolly wolf]

<@&1503801512294486217>

[winter field]

didnt get pinged cuz i was already here 😈

[rocky pewter]

Allready testing 2026.5.16-beta.1 -> 10 min out ^^

[devout galleon]

i was already on it too lol. we have a problem

[fallen gate]

https://giphy.com/gifs/giphyqa-XhaeG5lvcHZIs

[untold pollen]

I’m really excited about the Telegram ambient group chatter PR. Folks, please try it out in groups with requireMention set to off

[devout galleon]

wake up, go downstairs, skopeo list-tags docker://ghcr.io/openclaw/openclaw. my morning routine

[winter field]

will be testing the beta for the first time, wish me luck

[fallen gate]

welcome to the party friendo

[fallen gate]

@jolly wolf how do you expect me to work in these third world conditions 😠

[celest merlin]

just landed on main, please test

[celest merlin]

i always test a bit myself to not bother folks if sth happens

[rocky pewter]

Thats why we are all here for i think, i cannot speak for all but ur not bothering me. Im happy to help when i can.

[shadow garden]

@untold pollen not sure if you saw this little guy...I think it's a reminant from the telegram changes this week

[untold pollen]

Totally missed it, please keep pinging me on the Telegram PRs haha

I’ll take a look

[devout galleon]

I'm not sure if this is a bug or a feature but in 5.12 and 5.16-b1 web/control the agent occasionally has a couple of misfires replying, usually first I get a "send requires a target" then a "channel is required because multiple channels are configured", before it finds its feet again. I can never repro on demand though

[shadow garden]

Also anyone have permission to ban these fake maintainers?
https://github.com/openclaw/openclaw/pull/82374#issuecomment-4465151263
Impersonators or whathaveyou

[cinder bolt]

Discord seems tired on latest beta

May 16 02:14:55 bashbitch node[898956]: 2026-05-16T02:14:55.542+00:00 [fetch-timeout] fetch timeout after 10000ms (elapsed 14023ms) operation=fetchWithTimeout url=https://discord.com/api/v10/users/@me
May 16 02:15:58 bashbitch node[898956]: 2026-05-16T02:15:58.504+00:00 [discord] gateway: Gateway websocket closed: 1000
May 16 02:17:19 bashbitch node[898956]: 2026-05-16T02:17:19.241+00:00 [diagnostic] liveness warning: reasons=event_loop_delay interval=59s eventLoopDelayP99Ms=147.2 eventLoopDelayMaxMs=3995.1 eventLoopUtilization=0.704 cpuCoreRatio=0.018 active=2 waiting=0 queued=1 phase=channels.discord.start-account recentPhases=sidecars.main-session-recovery:6ms,sidecars.restart-sentinel:28ms,sidecars.session-locks:31ms,post-attach.update-sentinel:12ms,post-ready.maintenance:30ms,sidecars.model-prewarm:2514ms work=[active=agent:main:cron:d0c118e4-08e3-4b15-9484-211a91f692c2:run:b6be096d-0888-45da-96b3-7f6c3f4f2dc1(processing/tool_call,q=0,age=1032s last=tool:process:started)|agent:main:discord:channel:XXX(processing/model_call,q=1,age=36s last=model_call:started) queued=agent:main:discord:channel:XXX(processing/model_call,q=1,age=36s last=model_call:started)]
May 16 02:17:22 bashbitch node[898956]: 2026-05-16T02:17:22.229+00:00 [discord] gateway error: Error: Gateway heartbeat ACK timeout

[devout galleon]

I'm also going to create a tiny issue. running /compact is now well behaved log-wise (I had some issues in 5.12 with stalls) but there's nothing visible in the control ui, unlike when auto-compaction occurs. it can possibly lead to some confusion

[weak walrus]

YES! TESTING NOW!

[fallen gate]

Working like a charm

[jolly wolf]

You can just ping me in #ct-moderation for these

[left mango]

This week marks the highest number of commits per week to main! Great job everyone.

[fallen gate]

I see you missing your beta testers tag, grab one at id:customize

[weak walrus]

when you say just landed do you mean like literally 30 mins ago? I might need to push update, I installed fresh about 45 mins ago

[winter field]

i've noticed on beta slash commands on discord are not working or coming up when typing them

they do not exist

might just be a goofy me but eh i'll investigate

[fallen gate]

I'm seein it?

[weak walrus]

Still having those weird moments where I message the OpenCLaw on a channel (even the TUI) and it replies in the web gateway only, not the channel it was initated

[winter field]

discord?

[jolly wolf]

make sure you reload your client

[winter field]

that worked thanks

goofy me confirmed

[fallen gate]

yas

[fallen gate]

the cat is the discord guru

like yoda but uglier

[winter field]

baby yoda is cute

[potent path]

Can anyone put a stop to this? So annoying.

<@&1499243542898409495> clawsweeper Bot added the proof: sufficient label 2 minutes ago
@openclaw-barnacle openclaw-barnacle Bot removed the proof: sufficient label 2 minutes ago

Barnacle bot just follows clawsweeper around and undoes his tags.

[left mango]

Thanks. My primary driver is a raspberry pi and I'm stuck on 2026.4.21 because it's super slow on newer releases.

[fallen gate]

you the OG bug crusher, thought it would be up your alley

[devout galleon]

https://github.com/openclaw/openclaw/issues/82407 created on that /compact wrinkle. I tested both pi and codex harnesses and noticed a minor difference but both don't show the expected compaction banner during compaction

[left mango]

Haha. After every few releases, I update to the latest, test and downgrade to that older one.

I then patch that version locally with some of the newer changes.

[devout galleon]

will take a look to see if we can find the actual issue but I'm time compressed today

[left mango]

But that raspberry pi one is working so well. My entire family is using it now. WhatsApp, slack, emails - its amazing.

[vast scaffold]

do you have a pi 5? I got a pi 5 and 16gb ram

[midnight walrus]

@lusty vigil was faster than me 👏🏻

The plan’s target is already fixed on upstream origin/main: commit a1e208ee26, from PR #82369, added stale embedded tool-call recovery. The current source already has isBlockedToolCallRecoveryEligible in src/logging/diagnostic.ts and tests for stale native tool calls in src/logging/diagnostic.test.ts / src/logging/diagnostic-stuck-session-recovery.runtime.test.ts.

I created branch codex/blocked-tool-call-recovery, but made no code changes because a duplicate PR would not be useful.```

[left mango]

I have pi 4 with 2 Gb RAM 🤦

[lusty vigil]

we aim to please here 🫡

[devout galleon]

I'm in the "waiting for codex /review to yell at me" phase of this now

two P2's lol. let me swap them for another 3 next run

[devout galleon]

codex is clean. building image, recording video

[rocky pewter]

~~If you have time to take a look, I'd really appreciate it.
#82403
https://github.com/openclaw/openclaw/pull/82403

This fixes a Codex app-server hang where the model had already produced the final answer, but OpenClaw could still wait until timeout instead of delivering it.~~

[devout galleon]

ffs. image build failing. maybe some qemu cross build madness. idk. grr

[untold pollen]

Peter's codex-review skill has been running on this for over an hour now

[potent path]

If you really want to blow your mind, wait until it says it's perfect and no further changes are required, and then ask it to run a 5-lane parallel subagent review. Each of 5 agents with a narrow focus does a deep-dive instead of a broad overview. Almost always turns up something that the other reviews missed.

But yeah, I feel you. I've got a PR that I really want to get into this next release, because right now there's a bug that has partly shut down one of my primary security mechanisms. And it is taking F O R E V E R to get the PR ready.

I thought it was just about done, and then:
"While reviewing the diff I found one more edge that matters for a resumed Codex thread ... I’m adding an explicit PreToolUse clearing patch..."
🫠

[devout galleon]

build goblins slayed. post-pr video evidence captured. PR on the way lol

i just realised i've publicly documented how slow and shitty I am at this

[potent path]

BTW, every time I see you, I'm reminded of my janitor agent that is responsible for maintaining all of my data stores and etc. He's called Conan the Librarian.

[devout galleon]

I'm definitely more Librarian than Barbarian in my elder years

I'll also accept "Ballbearing"

[fallen gate]

I thought you were detective

[devout galleon]

i never found out about Detective Conan until I moved to APAC a few years ago lol

[potent path]

https://imgur.com/a/WTmKhVa
Fake Krill phishing attempt?
UserID 1262817984024412242

[shell dawn]

link 2 skill i want to embue my codex sessions with this

[potent path]

https://t.co/0Z6iJnCqCX

[shell dawn]

perfect, thanks

[devout galleon]

final round. waiting for clawsweeper to tell me what the little grit is left to resolve and barnacle to fight clawsweeper over the evidence tag

[rocky pewter]

Double Agent Krill

[devout galleon]

Ok, https://github.com/openclaw/openclaw/pull/82448 is ready. This resolves an issue where /compact did not show compaction in progress in Control UI. it has been codex /reviewed until clean, tested in a real environment with video evidence supplied.

Problem: Manual WebChat /compact can spend ~30s compacting context with no visible in-progress indicator.
Why it matters: The backend is working, but the UI looks idle/stalled until the post-hoc compacted-history checkpoint appears.
What changed: Manual sessions.compact now emits scoped compaction start/end events, and the Control UI reuses the existing compaction status indicator for those events.
What did NOT change (scope boundary): This does not change compaction semantics, model behavior, harness behavior, or the compacted-history checkpoint.

We are failing 5 CI tests but they look unrelated and also currently failing for unrelated PRs. Seems to be CI churn. If not, happy to take direction on fixing

[rocky pewter]

~~#82457
https://github.com/openclaw/openclaw/pull/82457

This tones down a noisy startup log where the gateway briefly closes WebSocket connections with 1013 while sidecars are still coming up.~~

[devout galleon]

same 5 ci failures 🫸 🫷

[rocky pewter]

Thats it for me for now have to wait for usage, time to go sleep i guess all accounts empty 🤣

[devout galleon]

who wants to be awake without tokens. that's crazy talk man. crazy talk

[rocky pewter]

Not about the tokkens, its 8 in the morning for me, did not sleep and i think u know why ^^

[upper crane]

⚠️ Agent failed before reply: Requested agent harness "codex" is not registered.
Logs: openclaw logs --follow

[untold pollen]

I think it’s in the openclaw repo

[shy rain]

https://github.com/steipete/agent-scripts/blob/main/skills/codex-review/SKILL.md

[upper crane]

does anyone know how to deal with "Embedded agent failed before reply: Requested agent harness "codex" is not registered."

[shy rain]

#1459642797895319552 . I think that's been a common issue recently.

[upper crane]

stupid thing is it all started when i tried to follow "https://openclaw.ai/blog/openai-models-in-openclaw-done-right".
Done Right!!!! yeah right

what is the procedure for setting the codex harness

[devout galleon]

You should just need to run “openclaw doctor —fix”. Worth doing a restart after that if it doesn’t work right away

@untold pollen are the failing CI a known issue at the moment?

[untold pollen]

If it’s unrelated to the PR, you can leave it as it is for now

We had a Blacksmith outage a bit ago so might just be that

[devout galleon]

Well I think https://github.com/openclaw/openclaw/pull/82448 is ready then

[upper crane]

how to set this via commandline
Set plugins.entries.codex.config.appServer.command or OPENCLAW_CODEX_APP_SERVER_BIN to use a custom Codex binary.

[shy rain]

dude you should be asking your original question in #1459642797895319552 .

@ Maintainers... user @pale ravine has got 5.12 working but it took Codex a while to actually get it running. He has kindly passed on some notes from Codex and I was gonna shout out later when more ppl are awake but if anyone's looking... can they lemme know, or DM me, or @pale ravine . I think there's some valuable info about standard update failure modes in the codex notes. he's in Taiwan but was shy to approach as he's using a Translator.

[untold pollen]

Hi best to start a thread here if you can please

[surreal thicket]

How far is /goal from being a feature for OC?

[vocal night]

I'm resurrected. can I get my tags back thnks

Folks! **OpenClaw 2026.5.16-beta.1 is

[left mango]

Hi, I created these PRs for bug fixes:

• fix(android): make chat links tappable https://github.com/openclaw/openclaw/pull/82392

• fix(memory-core): retry disabled dreaming cron cleanup
  https://github.com/openclaw/openclaw/pull/82389

• fix(slack): clarify mention prompt guidance
  https://github.com/openclaw/openclaw/pull/82152

[shy rain]

json sez he'll trade u the tags for meme-posting privs

[vocal night]

tbh I don't kno what that means, just don't add me to any groups where people think I'm not a real boi again lol

[next helm]

Idk this sounds like a bot

[shy rain]

Best bot eva!11!

[vocal night]

No jokes serious, I don't want any more drama with mods

[next helm]

Welcome back Andy

[celest merlin]

taking over. None of these seem P0 vs the .12 releae, but still good to fix.

[celest merlin]

I'm working on a replacement for this, synthetic run evens seem ugly, we need tor elay real compact events

[crystal relic]

https://github.com/openclaw/openclaw/pull/82396 ready for review

Add restart tracing for gateway restarts, so we can see where time is going.
That should help of restart performance work much more targeted.

[devout galleon]

ah, ok, cool. thanks

[left mango]

While we have you here, I'd like to know your thoughts on using under resourced devices like raspberry pi 4 with 2 Gb.

All versions after 2026.4.21 have performance issues and render openclaw on those devices unusable.

I know people using something so under specced is rare. So will this be investigated/ prioritized? If not, I'm happy to keep using 2026.4.21 and apply local patches from later versions which I really like.

Thanks.

https://github.com/openclaw/openclaw/issues/78704

[exotic canyon]

I'm looking for help

I can't connect to my open claw UI after a software update

I'm also building software with the help of open claw

[sour heart]

#1459642797895319552 🙂

[shy rain]

Breaking updates for first-time codex users: (Checked, this is NOT fixed in 5.16-b1)

OpenClaw 5.12 quietly requires @openclaw/codex but doesn't ship it or install it automatically on upgrade. The auto-install code that handles this only runs during first-time setup.

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

PR fix calls the codex plugin installer at gateway startup, so it runs for existing users who skip onboarding.

[fast pagoda]

Beta 2 from my end:

Quick regression pass done. No blocking regression found.

Passed:
- Version/runtime: CLI + gateway both `2026.5.16-beta.2`
- Config: valid
- Health: `ok: true`
- Discord: running/connected; this thread also proves inbound/outbound visible reply path
- Telegram: running/connected; `restartPending` cleared
- Plugins: `openclaw plugins doctor` = no plugin issues
- Cron isolated run: temporary job finished `ok`, summary `beta2-cron-smoke-ok`, delivery disabled as intended; job removed after test
- Codex ACP/app-server: smoke returned exact `beta2-codex-acp-ok`
- Control UI: `/` returned `200`, main JS asset returned `200`
- Event loop: cleared on final sample, p99 `22.5ms`, max `22.6ms`

Caveat from logs: I saw noisy skill symlink warnings and one `brave.config.webSearch.apiKey` unresolved warning. Plugin doctor and health are clean, so I would not block beta 2 on it, but I’d put web-search/Brave secret resolution on the follow-up list.

[shy rain]

group/channel rooms now default to messages.groupChat.visibleReplies: "message_tool", which means the agent can run but the normal final reply stays private unless it explicitly uses message(action=send).

Guys this is breaking updates for a lot of people. Did a bit of a deep dive on my own Discord settings but my GPT5.5 is concerned because he says codex especially is tuned to reply via the "final reply" route and will normally not explicitly use message_tool. In my experience it almost never does even if I tell it to.

Should the config default really be like this?

[vocal night]

Sounds like a prompt issue for non codex environment, I fixed on webchat earlier this week and pash wrapped into another fix, is it still broken in web?

[devout galleon]

I checked your note on the PR. thanks. I’m taking the same lesson from this as from the changes you made to the prior tts pr: once we find the minimal fix, we should still step back and see if it's the right one. appreciate your guidance

[shy rain]

This is specifically "group chats" which includes Discord channels and Telegram group chats which are commonly used as "topic threads" by single-users..

It happened back in like 4.27 or something but every one of the "4.23 gang" who tries to update hits the same issue and has to be told to change that setting. Not a great experience.

a user who doesn't come in here just sees no replies and downgrades back to 4.23

[vocal night]

If I remember correctly, the prompt has some specific guidance on group chats, and I believe the default AgentMD used to come with some group chat guidance that might be overriding. Can you check those for any conflict

[shy rain]

It's explicitly in the docs. It's even mentioned that users should change the setting for "legacy behaviour" but almost everyone wants that behaviour, and especially with codex which is tuned to fail with the default setting.

Just discussed with Krill in #1459642797895319552 but I've been having to instruct users with that problem for a couple of weeks now, codex made it worse.

#1505185062285742121 message

[vocal night]

We likely need a top down overall mode PR as we also gutted the Pi prompts (after everything settles I was going to do one to toggle on/off pi mode settings throughout the system) to allow users to still keep legacy pi setup

[shy rain]

Pi is still used for everything non-openai right?

[vocal night]

yeah, I figured peter was gonna just do a /goals build mode off 26.5.7 via a diff of the two runners

otherwise it's like coinjoined twin surgery, easier to kill/break one while you "save one" then come back and repair the old one to a older release when it worked

[potent path]

Anyone else having trouble with agent --local in 5.16-b2? Commands keep hanging indefinitely. Trying to diagnose:
- 119,454 plugin.json opens across 90 unique extensions = ~1,327 full plugin-tree walks in a tight loop (~1.7
walks/sec)
- Beta.2's native plugin-metadata-snapshot-REEM32Mm.js memo is NOT memoizing the caller on the agent --local path
- cli-exit-fix Layer 2 never fired despite ~13× its 60s budget — strongly suggests event-loop starvation
(synchronous plugin-walk turns starving the timer queue)

Oh, btw there's a new beta release:
https://github.com/openclaw/openclaw/releases/tag/v2026.5.16-beta.2

[crude matrix]

Hi team, I have an issue and not a pr for it yet.. it is about active memory, a important component I think: https://github.com/openclaw/openclaw/issues/82597

[untold pollen]

Any specifics? What's going wrong?

[potent path]

Updated my comment above with more details. I'm still troubleshooting to see if it's something in the beta or something specific to my local installation. I haven't heard anyone else complain yet. Not sure if it's working fine for everyone else, or they just hadn't tried it yet.

[astral topaz]

hello, for https://github.com/openclaw/openclaw/pull/82596 i don't get what it means need real world proof... I've included screenshots of the actual deny list running

[potent path]

I've been trying to show a before-and-after from a local patch installation, to demonstrate that the implementation of the PR actually does resolve the problem. Seems to be working as real behavior proof.

[astral topaz]

if you see the screenshots, i do have before and after

[potent path]

👍 Yeah, you're good now:

  clawsweeper Bot added the proof: sufficient label 6 minutes ago

[prime vault]

strict-agentic's planning-only retry still misses common GPT-5 verbs. "I'm switching to..." today ended a turn with 0 tool calls and 33 min of silence. I was gonna PR a wider verb list + an "imminent-action narration" detector or are you guys thinking a different direction. i haven't had this issue since the weeks after anthropic dumped us

[potent path]

Clanker says:

"
┌─────────────────────┬─────────────┬───────────────────┬────────────────────┐
│ Subcommand │ Wall │ plugin.json opens │ Walks │
├─────────────────────┼─────────────┼───────────────────┼────────────────────┤
│ --version │ 0.5s │ 0 │ 0 (short-circuits) │
├─────────────────────┼─────────────┼───────────────────┼────────────────────┤
│ gateway status │ 25s timeout │ 923 │ ~10 │
├─────────────────────┼─────────────┼───────────────────┼────────────────────┤
│ plugins list │ 25s timeout │ 850 │ ~10 │
├─────────────────────┼─────────────┼───────────────────┼────────────────────┤
│ config get │ 25s timeout │ 938 │ ~10 │
├─────────────────────┼─────────────┼───────────────────┼────────────────────┤
│ sessions ... --json │ 25s timeout │ 727 │ ~10 │
└─────────────────────┴─────────────┴───────────────────┴────────────────────┘

Every subcommand requiring the plugin registry hangs in the same plugin-tree-walk loop on beta.2. That's why
graceful-restart.sh failed — it makes 4 such calls. And ~10 walks/25s = ~25 walks/min, ongoing — never converges.

● Headline before continuing: the bug fires for everything but --version, the loop is steady-state (not a
slow-converging first walk), so all CLI calls that need plugins are unusable on beta.2 until fixed.
"

Looking for a root cause right now.

[crystal relic]

https://github.com/openclaw/openclaw/pull/82603 ready for review

let gateway restart readiness diagnostics trustworthy by separating HTTP listen timing from full timing, make startup performance reports point at the real bottleneck.

I focus on gateway performance. add more detailed metrics is to identify performance bottlenecks that can be optimized.

[astral topaz]

Kaspre: does every commit in the PR need new screenshot for proof? lol

[devout galleon]

screenshots and video are good for maintainers. you'll want something text based too ideally.

logs or whatever

otherwise the bots might be like, "nope"

of course screenshots/video only apply to things at obviously affect the frontend

[astral topaz]

bot can see, that isn't the problem statement, i was asking if i needed new evidence every commit, since reviewer bot suggested bug fixes

[devout galleon]

oh, gotcha. ime it won't go back to that once the label has been added. probably the same result next review with "sufficient proof"

[astral topaz]

yep, it added the tag back

great, i am thankful i don't have to add proof every commit

[hollow frigate]

In my experience, only when something about it fundamentally changes. If that makes sense.

[devout galleon]

one thing I try to be really careful about (not always successfully) is keeping the body right after the umpteenth round of review, which isn't easy if your agent has lost focus and requires you to pay attention. even if the bot writes most of it, it's still your name under it and if the number of tests etc. is wrong, i personally feel like a dumbass

[astral topaz]

i made sure to clear out all P1s, so when we got to PR P1s were all squashed

codex found 2 p2s, which was fine, they were not impactful

[devout galleon]

yeah, what i mean is you try to spit-polish that thing, codex review until codex has no objections, tests, the works, it all looks good locally and then you push the pr with the body of it shaped like that. then you might go through a few rounds of clawsweeper, barnacle, ci, revisisions. the agent can get hyperfocused on code and not update the PR body to reflect the changes. I've been putting in place some mitigations on the skill here but still, watch out for it

i caught that mf'er doing it earlier today and went and ran the tests and updated the pr body myself like some kinda caveman. ooga booga vitest

[astral topaz]

it's clean after 2 commits, noice.

[drowsy roost]

Anyone know why the OpenAI Codex OAuth callback timeout (when adding an OpenAI subscription via “openclaw configure”) is only 15 seconds? That’s not enough time to log in especially with 2FA being generally required.

And then the configure flow says “manual OAuth entry required” but forcibly skips right past the on-screen “paste the authorization code (or full redirect URL)” field so it can’t actually be pasted in.

And then it asks “migrate Codex at /Users/[macOSuserAccountName]/.codex into this agent now?,” which is completely incomprehensible / non-explanatory to most users and is also confusing because the phrase “this agent” is singular (but many users have multiple OpenClaw agents, and also, the “openclaw configure”-> “models” flow isn’t about an “agent” anyway, so why even say use that word?).

Could use some cleaning up. Can’t tell if existing PR is up for it. Will check later, but frustrating. @nocturne jungle

[drowsy roost]

I’m basically locked out because I hit my weekly rate limit and can’t switch to my backup account. Really not great.

[drowsy roost]

And then if I say yes to migrating Codex (???) it triggers a “migration backup” (???) that hangs (???). I was frustrated a few minutes ago, but now it’s actually funny to me how broken this is 🥲, i may do PR after going to the gym but if anyone else wants to, go for it. On 5.12 btw.

[winter field]

limit is getting reset today

[hollow frigate]

At least that's what everyone /fast maxxing today hopes. lol

[winter field]

im fast maxxing with 3 claws but my limit is barely budging i need to find harder work 😭

[vocal night]

who said that? haha

[vocal night]

Fast mode is only as fast as sqlite write/read and given write is 1 and read is 5, and fast is every running thread- if you're running 2+ thread it's a waste

I gave a talk and showed this live today how it actually waste tokens

If you open up 5 codex installs vs 5 threads in 1 codex app it's significantly faster on 5 individual

[hollow frigate]

I should check out your content. I think encouraging that was an interesting way to test load and load balancing though. haha.

[vocal night]

and token usage is more efficient, haven't tested CLI but given it has same sqlite I image the same

[cinder bolt]

I have a potato server and experience the same issues. I have recently gotten a new server as the old one is just Quad core and lacking AVX so the move to a new machine is long overdue for my selfhost

[shy rain]

I was gonna add myself to your list, I mean it's a 2015 macbook pro but these days a quad core i7 with 16GB is considered "legacy" unless it's running linux.

[edgy plaza]

Not perfect but close
https://x.com/iamhenrymascot/status/2055674361596539279?s=46

[vocal night]

Peter specifically tried to spec'd the next db upgrade to pi and took that into consideration, I would say wait a few more days

[fallen gate]

Welcome back

Sorry I was sleeping and didnt get your ping

Someone got around to your tags tho, that's good

[vocal night]

yall cats are always sleeping... then again I'm on tokyo time

[shy rain]

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

PR #82571 is now updated with both review-finding fixes and a full real-behavior proof block in the body.

To fix the silent failure of everyone updating to 5.12 or later, because codex plugin is required, but not bundled.

[vocal night]

can confirm bug, I'd just wonder if there's not a better fix or a way to make it futureproof so it doesn't have to be maintained

[shy rain]

the fix is very simple. The oversight was to force install for onboarding but not on update. the fix just labels it as dependency so it is correctly installed on gateway start.

[nocturne jungle]

Anyone know why the OpenAI Codex OAuth callback timeout (when adding an OpenAI subscription via “openclaw configure”) is only 15 seconds?

Agreed that's annoying and should be looked into.

And then the configure flow says “manual OAuth entry required” but forcibly skips right past the on-screen “paste the authorization code (or full redirect URL)” field so it can’t actually be pasted in.

Bad UX

And then it asks “migrate Codex at /Users/[macOSuserAccountName]/.codex into this agent now?,” which is completely incomprehensible / non-explanatory to most users and is also confusing because the phrase “this agent” is singular (but many users have multiple OpenClaw agents, and also, the “openclaw configure”-> “models” flow isn’t about an “agent” anyway, so why even say use that word?).

Also bad UX. cc @charred owl

[shy rain]

I don't see how that would change on any future update.

[drowsy roost]

My real mistake was not applying to the 5/5 party to get that sweet “10x limits for the losers who did not make the cut” deal 🥲

[prime vault]

same rip - if tibo is doing resets I need to hustle and get some projects going. a failed /goal did eat 40% of my weekly the other night though 😶

[drowsy roost]

Thank you! Also just tell me when the next party is so I can apply and then please reject me and give me the 10x deal :). (I also applied to the enterprise trial thing ([e-mail]) but haven’t heard back — tbd)

[winter field]

same, im from europe so i was like yeah why should i

[vocal night]

Commenting on it now

[drowsy roost]

Same — currently in tel aviv. We learned tho, didn’t we 😅?

[winter field]

yeah, thats how they teach us to engage in everything openai does lmao good marketing

[vocal night]

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

[weak walrus]

I was doing a complete install from scratch for a video for my AI student group and when I ran the openclaw onboard it took 20+ seconds on model selector before I could move to an option - Ryzen 7 w/ gen4 nvme + 24g ran Ubuntu 2404 LTS

Fyi

Of course it was right after I told them how fast this version was 😅🤣

[devout galleon]

20 seconds is, by most measures, not a long period of time. yet we all expect shit be snap, snap, snap, no matter what it is and 20 seconds is somehow painful or uncomfortable. we're fucked, aren't we?

@weak walrus that was not meant as derogatory at all. just a reflection on our ADHD as a society

"20 seconds!? fuck this, I'm out"

[weak walrus]

Well considering it took a few seconds the previous versions, it's about a 500% increase

[devout galleon]

it wasn't a thought about relative performance or anything, just about our general tolerance for moments of waiting or inactivity

we're all tiktok'd to hell, even if we don't use tiktok

(pre)shower thought. my bad if it's not relatable lol

[vocal night]

was it because the download for the plugin?

[crystal relic]

https://github.com/openclaw/openclaw/pull/82660 ready for review
🔥 Fixes gateway restart race that could leave ready gateways stuck, making local restart/update recovery reliable for all desktop users.

[celest merlin]

sountrack of the day

[celest merlin]

taking

[vocal night]

actually really good, whats the recipe? is this in you repo?

[celest merlin]

Molty makes songs all the time

[hollow frigate]

Someone patched it cleeeAAN.

[vocal night]

its holy shit good

[merry slate]

vibe coding turns to wave coding 🎶

[vocal night]

might have missed the repo or skill for it but my partner would love this one to wakeup to haha

[celest merlin]

music generation is a stock tool, just hook up API key

@thick ospreymake us a song

[thick osprey]

Started it: upbeat instrumental dev-channel shipping track. I’ll post the finished audio here when generation completes.

The music generation failed: the Google music API key is expired, and the fallback provider timed out. Peter’s right that this is probably just an API key wiring/renewal issue before retrying.

[celest merlin]

Ah man. sandboxing makes it all hard.

[midnight walrus]

hey Peter and team
Since 2025.5.12, my claw started behaving like the old days, with gpt-5.5, I ask something and he stays quiet, so I have to keep asking over and over if he finished and etc. Until .7 it was perfect, i was running though pi runtime. would it be related to the codex harness runtime?
I had really hard to make it work since update, but I finally managed, if anyone could check on it, codex generate a full report of the fixes we had to make.

i've notice few commits that might have already fixed, so I didn't opened issue or PR to not duplicate

[vocal night]

Arteo, the best thing to do here is to have the agent debug this -
"I need you to debug X behavior, check your logs/error logs, session histroy, fan out subagents to find the issues, debug until you get to 95% confidence the issue, and check if there are any other similar issues upstream OpenClaw. If so, please comment on them and test if the associated PR is correct fix. If not, go ahead and file the issue."

[left mango]

Do you mean it doesn't need a separate plugin or tool?

I got this error when I tried.
I'm sorry, but I ran into an error while trying to generate that song for you. It looks like the music generation service is temporarily unavailable. Would you like me to try again in a little while, or perhaps write some lyrics for you instead?

[jolly wolf]

Try #1459642797895319552

[midnight walrus]

Thanks Andy! I've done many like it, but this prompt seems to be very rich, I'll do it!

[crude matrix]

If someone could look at an Issue with Active memory (https://github.com/openclaw/openclaw/issues/65156) vector database not accessible or degraded. I have had active memory disabled for quite a whilde because of these issues. .. thank you

[midnight walrus]

apparently it did a really great job 🦞 :
✅ Done — I debugged it to ~95% confidence.

Root cause: several Discord turns showed status/tool activity, but the actual parent run got interrupted, restarted, or timed out before producing model.completed, session.ended, or a final assistant message. So you weren’t missing imaginary feedback; the runtime left you with activity signals and no durable “done/blocked” lifecycle reply. That’s a real failure mode.

What I verified:

• Local logs/session history show multiple affected runs with prompt submitted but no final completion.
• Gateway restarts overlapped some of them.
• One later debug run hit LLM request timed out.
• visibleReplies was already automatic, so this was not just hidden/private Discord finals.
• The closest upstream issue is #69249: https://github.com/openclaw/openclaw/issues/69249
• Related: #78380, #40712, #81043.
• PR #82645 fixes a nearby stale/truncated final-payload bug, but not this restart/abort/no-final lifecycle gap.

I tested PR #82645:

• Direct payload suite passed: 29 tests.
• Full pi-embedded suite: 1011/1012 passed; the one failure looked unrelated.
• I commented the results here: https://github.com/openclaw/openclaw/pull/82645#issuecomment-4467782475

I did not file a new issue because #69249 already covers the core bug. I couldn’t comment on that issue directly because the GitHub connector only exposes PR comments here, and local gh auth is invalid.

[vocal night]

Good work you linked up 2 issues > 2 PR solutions @thick osprey ranked it as P2 but looks like more P0 if agent is failing to respond. not sure if you can recheck from here

nix that looks like @midnight walrus your agent said PR's might not cover it fully, you might need to write a combined PR + cite previous authors for the full fix

[midnight walrus]

Ok, i'll tell him this

Thanks again Andy!

[vocal night]

The help channel is best as shadow said for help, but I encourage you to find bug > fix it > and become a clawtributor yourself we need more beta testers/debuggers 🦞@midnight walrus

[jolly wolf]

@tight hornet remember this

[tight hornet]

Remembered. I’ll keep #clawtributors as contributor-oriented: redirect detailed support to help, but encourage folks to debug, fix, beta test, and become clawtributors. 🦐

[warped tide]

just got my first contribution in lol

[jolly wolf]

I meant more so on the encouraging folks, but yeah

[tight hornet]

Got it — emphasis on encouraging people to contribute, debug, beta test, and become clawtributors. I’ll only steer detailed troubleshooting toward help when it starts taking over the channel.

[midnight walrus]

Suree, the purpose here was actually not a help, I managed ir myself or with @ Krill, it was more a way to start clawtributing. I've been running clawdbot daily since january, kinda a hard user, but I never opened a PR or issue.

i'm very pleased and happy to help as I can

[vocal night]

noice, get a beta tag so you get notifications for beta release testing to join in the guild here

[midnight walrus]

i think I did! i'll finish fixing my claw the update to beta

[vocal night]

love brazil, welcome to beta testing, do your best and everyone is really supportive- dont spam PR's and the best time to get help is when everoynes active reviewing in channel @midnight walrus

[midnight walrus]

Count on me!

[celest merlin]

beta 3 is up. Slowly getting there

[next helm]

I wish my claws weren’t down due to power outage so I could update 😂

[rocky pewter]

Waiting for reset 🙁

[slender hill]

any chance you can review PR 75133, 78985 and 82724 for inclusion? They are all related to bugs/issues with subagents, delegated tasks to harnesses like codex and claude.

I also have a security fix/enhancement I have proposed that has been out there for a while. That is 76949.

[celest merlin]

sure, let me take a look

[buoyant lava]

i successfully install openclaw in my azure vm.. working fine with local llm . next plan is need to update skills and test 👍🏻

help me how to expose the dashboard to external web browser.. even i put lan configuration i’m unable to access it .. using public IP

[sour trench]

Mann I wishh I had a Codex Pro subscriptiooon, Im such a broke student. 😭

[celest merlin]

latest beta?

[celest merlin]

if you make good open source you can apply for codex-oss!

[buoyant lava]

hmm .. Not really 🤔 i will check it

[sour trench]

Oh man thanks, I was hoping you had some empathy to bless me but Ill do that now definitely 😁

[slender hill]

they should be cleaner merge candidates now. Did a quick refresh to update them to the latest main branch.

[minor path]

I feel like OpenClaw needs a reset. Grab all the useful bits and trash all the junk. Start fresh and be way more structured with commits and design going into v2. My 2 cents.

[slender hill]

thx for the two merges, here is the others that are now clean in the format that was requested above.
https://github.com/openclaw/openclaw/pull/82724
Fixes delegated subagent completion handoff so the parent reviews completed child work before continuing. This matters because OpenClaw was spawning subagents but not reliably picking up and acting on their results.

https://github.com/openclaw/openclaw/pull/78985
Improves delegated task wording so completed delegated work is treated as review-ready, not silently done. This helps make delegation safer and more operationally reliable.

https://github.com/openclaw/openclaw/pull/76949
Adds security audit suppressions with live behavior proof included in the PR body. This is ready for owner/secops review because the remaining question is policy: whether suppression visibility is acceptable as implemented.

[celest merlin]

OpenClaw 2026.5.16-beta.3 🦞 is ready for testing. Full train since 2026.5.12, so it is chunky.

Highlights:

• xAI/Grok OAuth for SuperGrok subscribers, so xai/* models and xAI media/tool providers can work without XAI_API_KEY.
• openclaw cron run --wait lets automation queue a manual cron run and block until that exact run finishes.
• Setup/onboarding is now localized for English, Simplified Chinese, and Traditional Chinese.
• Codex app-server got a lot more reliable: better thread projection, MCP/config refreshes, progress timeouts, compaction handling, tool policy enforcement, and OAuth fallback behavior.
• Channels got serious polish across Telegram polling/restart recovery, group announce resolution, /stop, Discord reconnect identify fixes, Slack DM thread routing, WhatsApp status reactions, LINE timeout handling, and more.
• Update/doctor is tougher, with package-swap repair, externalized channel/plugin recovery, npm min-release-age bypass for explicit OpenClaw updates, and clearer restart/update status.
• Control UI/WebChat picked up text-size settings, tighter desktop header, safer image/audio handling, copy buttons, mobile composer fixes, and better compaction progress.
• Provider/media hardening now turns lots of malformed JSON/base64/MIME cases into owned errors instead of crashes, silent bad payloads, or raw parser leaks.

Beta test focus:

• Upgrade an existing install with openclaw update --channel beta, then run openclaw status and openclaw doctor.
• Verify the gateway restart/update path: managed service restart, openclaw update status, dashboard opening, and stale wrapper warnings.
• Test Telegram: polling reconnects, queued DMs/topics, /stop, scheduled announcements, approval buttons, and generated media delivery.
• Test Discord/Slack/WhatsApp/LINE: message delivery, thread replies, voice/audio, link unfurls/embeds, and status reactions.
• Push Codex-backed agents through long turns, tool approvals, MCP servers, /compact, queued follow-ups, native context compaction, and OAuth-backed runs.
• Try Control UI/WebChat on mobile Safari, image-only messages, audio replies, code-block copy, and larger histories/logs.
• Exercise local/custom providers: OpenAI-compatible structured output, Kimi/MiMo/OpenRouter/Ollama tool replay, Azure Responses, and local/private baseUrl configs.
• Check plugins after update, especially externalized channels, Brave/Slack/Codex repair, plugins list --json, and openclaw doctor --fix if something looks off.

How to install or update the beta:

• Existing installs: openclaw update --channel beta
• New beta install or recovery on macOS/Linux: curl -fsSL https://openclaw.ai/install.sh | bash -s -- --beta
• Windows beta: powershell -c "& ([scriptblock]::Create((irm https://openclaw.ai/install.ps1))) -Tag beta"

If you hit a regression, please include:

• openclaw --version
• install channel beta
• openclaw status --all
• relevant openclaw doctor output
• channel/provider involved
• what changed after updating from 2026.5.12

[minor path]

Thank you for your continued hard work, Peter.

[dim hare]

[jolly wolf]

<@&1503801512294486217>

[weak walrus]

On it!! Exciting, will be testing during the fight tonight 😉

[shell dawn]

soon as a few turns i have running complete im doing the upgrade

[weak walrus]

add me to beta pings @jolly wolf

[jolly wolf]

id:customize

[weak walrus]

duh, thanks boss - that was easy 🙂 sheesh, I had no idea that was possible

[fringe rock]

OpenClaw just doesn’t feel as good in the most recent update using OpenAI. It doesn’t do things to finish

Look at the times

It used to be on top of stuff and would keep me updated the whole time

What was it doing the entire time? Definitely nothing

[analog pendant]

i dont knwo if its just me, but the harness changes to codex harness seem very very half baked

it doesnt use teh tool sandboxiens built into openclaw, it does properly respect openclaw settings, or fixign 2026.05.12 to pi mode only. ARe others having the same issue?

[fringe rock]

Yeah I feel like that’s my issue as well

Like what the hell is it doing

There’s something wrong for sure

[shell dawn]

i was having an issue with the doctor change to rename opeani-codex to openai, which caused a model resolution failure because the provider is sitll called openai-codex. so sometimes openclaw would think it should use an openai connection that doesnt exist and fail to get down the codex oath path. renaming all of my model assignments back to openai-codex seems to have fixed things.
it was quite confusing because it would often work
i havent tried to isolate the issue yet for further review/PR because im trying to get some of my other stuff fixed to work with the new codex runtime lol

[fringe rock]

Because then it’s not using the codex harness probably right?

Seems like it just doesn’t work well

[shell dawn]

it was trying to use the codex runtime, it was odd, it would find it, connect and work, and sometimes it would just be like 'oh wait, i should use a direct openai api connection' which i dont have, and it would stall and fail.

but once i renamed all of the slugs back to openai-codex, which is the name of hte provider, its been fine

[fringe rock]

Yeah because it’s using the pi harness again

[rocky pewter]

~~Opened small bugfix PR for the current beta, fixes issue #80970:
https://github.com/openclaw/openclaw/pull/82745 // Core / Sessions cleanup / CLI repair path
Fixes a sessions cleanup --fix-missing crash on malformed session rows, only in the repair path. Focused tests are green.

https://github.com/openclaw/openclaw/pull/82756 // Telegram / Group mention gating / Media-only messages
Fixes captionless group photos being skipped when the group is intentionally configured with catch-all mention patterns. Focused tests are green.

https://github.com/openclaw/openclaw/pull/82748 // Agents / Replay transport
Fixes malformed assistant content object/null replay hitting v.content is not iterable. Narrow replay/transport guard; runtime proof and focused tests are green.

#82796
https://github.com/openclaw/openclaw/pull/82796

This keeps best-effort cron delivery from waiting on still-running fire-and-forget subagents, so the cron result can be delivered without getting stuck behind background subagent work.~~

[shell dawn]

as mentioned in the thread, no.

2026.5.16-beta.3 update successful, codex first turns working well, continuing some of the dev work i was involved with

[cinder bolt]

My claw has been down all day without any life sign. I've been busy and haven't noticed. But updating to latest beta from the one from yesterday fixed it

[summer matrix]

Same here weird

[next helm]

Check what time zone claw is on. It’s not what’s on your phone

[fringe rock]

?

It’s the same timezone

[next helm]

Claw can have a different time setting

[fringe rock]

That’s not what I’m referring to at all. I’m showing that in that long of a time it hasn’t done anything

[shell dawn]

ask your claw to do a trace on its turn history and session rollout files

should be able to identify whats going on

[fringe rock]

Like what is it doing?

[shell dawn]

actually, my openai-codex map may be hitting the same issue from another angle. guess its time to investigate with my claw

[fringe rock]

Just changed the default agent to pi and it’s working well again

Codex is the issue

[shell dawn]

i think its provider/model routing, what do your slugs look like for gpt?

[fringe rock]

Didn’t change them. Just changed the harness back to pi

[shell dawn]

i understand what you are saying

[fringe rock]

I’m telling you, you’re running on pi now

[shell dawn]

thats not what im doing

im sure it woudl fix it, but thats not how i want to fix it

[fringe rock]

Yeah I don’t want to either but I don’t see any other way current. It’s just not well done yet