Imagine asking an agent of a cloud VM permissions while on openclaw you just set up everything by yourself and harness is 10x better. Isn't it a waste of time dabbling on every new flavor of agent instead of owning the harness of one fully customized to your needs?
Hi @steipete ,
A short note on a ChatGPT subscription failure i see, and the open PR that believe fixes it. It would be great to have it in v10.1 if it fits your plan.
The issue (#167345)
On 2026.9.8 (openclaw runtime, ChatGPT subscription route, gpt-6.1-sol, transport "auto"), a Responses WebSocket that closes with code 1000 before the terminal response event ends the turn with no recovery. In the last two days we have had few cases on production gateways:
- a scheduled automation failed after about 100 s ("Agent run failed"), with no retry and no fallback;
- a live chat turn failed after tools had already written business records, and the user got "Agent couldn't generate a response. Note: some tool actions may have already been executed…".
Both stored assistant messages show api "openai-chatgpt-responses", errorMessage "WebSocket closed 1000", errorCode "ERR_WEBSOCKET_NON_RETRYABLE_CLOSE". Neither failed message contained visible text or a tool call: one was empty, one had reasoning only.
Why PR #167512 likely fixes it.
Traced the path in the 2026.9.8 bundle:
1. extractWebSocketCloseError only treats 1001, 1005, 1006 and 1011–1015 as retryable, so 1000 becomes ERR_WEBSOCKET_NON_RETRYABLE_CLOSE.
2. That code is in TERMINAL_ASSISTANT_ERROR_CODES, and isTransientNetworkError rejects it, so isRetryableAssistantError is false and the runner surfaces the generic error.
3. With 1000 in the retryable set (the PR's one-line change), the same close becomes ERR_WEBSOCKET_TRANSPORT. That code is already in TRANSIENT_NETWORK_CODES and classifies as "timeout", which the existing maybeRetryTransient path accepts: same model, up to 8 retries within 90 s, continuing the saved transcript.
4. The classification only runs when no terminal event has arrived, so tool calls from the broken response are never executed. Tool results already in the transcript are not re-run.
5. That retry path already works on 9.8 for other transient failures. Seen it recover mid-turn timeouts after earlier tool writes, with a single final delivery.
The PR's test shows the close is now classified as transient. Have not yet run a full gateway turn that drops with 1000 and then finishes. One related point: if a retry moves a long-lived session onto SSE (#166848), the 16 MiB SSE cap fixed in #166853 matters too, so that fix would be a good companion in the same release. (Although auto seems to not prefer SSE)
Thanks! Hope @roboclawbot did good job for you to pick up?
Funny how I am using the same techique since June 25 with Claude code cli just in the scene and chatgpt o3 (no codex at all back then) so was copy pasting to exchange opinions for adversarial reviews.
Wait for the copy pastas claiming they can do the same in the next months. @openclaw always paving the way of the agentic world.
EdgeDimi retweeted
I hooked up our team claw to X to trigger work faster. Unassigned sessions are for anyone to grab. Our agent looks who worked on the related code last and pings people on the server.
Whole thing was a prompt and team server extended itself since plugins are now hot reloadable.
Opened a session to review PR #166379 for contrast, sidebar title spacing, and regressions.
team.openclaw.ai/chat/robocl…
They all push internally dot so hard like they couldn't do all these before dot with codex. Openai losing the plot and tbh the only thing being ahead of its curve every day is @openclaw , you just need to own its harness and be a master of the config knobs. Nothing comes close
Thankfully you care about your clients and you vendor lock them to one provider. Hypocrite with the irony of the word NO SKILLS
It's amazing how anyone who knows how to build Skills inside Claude/ChatGPT has the ability to earn $1,000+ per hour consulting for SMBs.
Never before in history has this big of an opportunity been available to the everyday person.
Same here models became too good and with increased context windows plus compactions, sub agents are just token wasters and on top the main agent doesn't own the task he needs to continue in a way he would otherwise auditing the code himself. The broken telephone of prompting the sub agent from main to the review of what the sub agent delivered vs what is actually needed is lots of risk you dont need to take with current models
Hot take: am i the only one that doesnt like subagents?
I find them useless token wasters. Output with subagents/without is pretty much the same with frontier models while they take a lot less time and money.
I prefer working without them any day, and multi-model setups are even worse. Just so unnecessary.
You can literally do this via message thread ids? Even the orchestraor thread can also archive the thread id which finished , is all native app tools
Edge247 officially migrates to v2026.9.8 @openclaw , patiently pinned for over 4 months to v5.26. You will be missed v5.26 , thanks for the stable uninterruped performance, time to welcome the oc 2.0 era
OpenClaw v2026.9.8 is out, and this one's a fun-sized lobster 🦞
🧠 GPT-6.1 Sol support
💬 Agent replies find their way back
🪶 Less memory use
🔧 Update & Windows startup fixes
43 PRs, 8 contributors. More in the full notes 👇
docs.openclaw.ai/releases/20…
EdgeDimi retweeted
Today we’re announcing OpenClaw Enterprise
In collaboration with @RedHat , @nvidia and @OpenAI the OpenClaw Foundation is open sourcing a powerful enterprise control plane for persistent agents
OpenClaw Enterprise is built to run on your own infrastructure and will always be free for an organization to use
openclaw.ai/blog/openclaw-en…
Btw you literally can do this already with codex app open on your mac mini...why my work needs to be on their servers. They dont even have the moat against muse. Btq did i mention @openclaw one more time? The only OSS agentic harness you need
Announcing dots. Dots work 24/7 for you, learn from your feedback, have their own computer, browser and can be connected to over 4k apps in our ecosystem.
They’re powered by Astra our best model yet. Included in your Pro plan, without drawing down on any of your usage. You can even call a dot while it’s working.
We’re getting you started with your primary dot today and soon you’ll be able to create entire teams of them.
Now do you understand why your harness should not be around your vendor's sub?
Buid on top of OSS , @openclaw everytime of the day , bring any LLM brain you want either frontier or local AI, they mind their pocket you mind yours.
Hi,
Tomorrow we are re-opening the Pro $200 subscriptions to new subscribers, but together with it we are also changing how we calculate the usage for it. In effect, if you do the math, it will net out at half the dollar in API spend compared to the old Pro $200 plan.
Now that it's said, let me explain why this is happening and why you will still get more work done than if you were on the Pro $200 subscription one month ago.
(a) We didn't want to compromise in other ways and are committing to not reintroducing the 5h limit, so that you can fully use the weekly usage when you want.
(b) On the subscription, we guarantee that over time you always get more work done and with an increasing level of quality. This means that you will continue to get more value per dollar spent as a result of models getting more efficient and us passing down the improvements in the form of API price reductions.
(c) We don't want to put an incentive on ourselves to artificially inflate the API list prices to make it look like you are getting a lot (and workaround it through discounts, etc). Instead we want to continue to both rapidly reduce prices and increase capabilities of models on the API. This week we introduced GPT-6 Sol and GPT-6 Luna at 50% of their previous price. Over time, we see prices go low enough that it makes sense for most to buy usage as needed without there being a significant gap between what you get in a subscription and what you get in the API for a dollar spent.
(d) Tomorrow, we are adding more things to the subscription that won't draw on the usage, I won't reveal what that is yet.
I wanted to be transparent before all the big announcements tomorrow. Lots of new exciting things are coming to the subscriptions that will make it super compelling, but I wanted to make sure to share this change ahead of time so you can all understand it before we shower you with good news.
Codexingly,
Tibo
Ben to the point.
Own the hardware
Own the harness
The big labs’ agent platforms can be blocked. OpenClaw can never be blocked. If you keep trying to get around having a real computer for AI agents, you will be periodically disappointed, forever.
VPS machines have limitations, and they’re not entirely private. Muse / Grok Bot / et al can be compelled by Amazon and similar companies to transmit signatures that they are not human; some companies will likely sue the big labs to compel this or sue them for refusing.
Untraceable and unbannable computer use by AI agents requires a real computer. Not a VM. Not a VPS. An actual computer.
Just get one already.
And get OpenClaw. In the long run, there is no way around learning how this stuff works and doing it on a real computer that you own. You can learn now, or you can learn later.
But you’re going to have to learn.
Because this is how things are now.
