OpenClaw 4.20 Update: Session Pruning, Completion Bias, Kimi K2.6

By Beau Johnson··8 min read

Quick answer

OpenClaw 4.20 is mostly a reliability release. It caps and prunes session stores when they load so the gateway stops crashing, pushes agents to finish and check their work, adds Kimi K2.6 with cost estimates, and moves cron runtime state out of jobs.json.

In this video I go through the OpenClaw 4.20 update and explain why it matters if you run agents every day. Most of the changes are about reliability: setups that used to fail slowly and quietly, with no clear error to tell you why. If you have cron jobs or agent pipelines running in OpenClaw, this one is for you.

I recorded this in April 2026. Tools like this change fast, so check the current OpenClaw release notes and docs before you rely on any detail here.

What I cover in this video

  • 0:00: How OpenClaw setups can break quietly: cron jobs fail without errors and the gateway dies
  • 1:32: Why session stores kept growing and how that can crash the gateway at startup
  • 3:03: The new default entry cap and age pruning for session stores
  • 4:34: Completion bias, or getting agents to finish the job before they report back
  • 6:05: Kimi K2.6 as a built-in model, with cost estimates in token usage reports
  • 7:35: Cron runtime state moves out of jobs.json so your job definitions stay clean in Git
  • 9:52: An exec fix for Python and Node scripts, plus onboarding and BlueBubbles group chat improvements
  • 11:23: Why the problems that build up quietly hurt the most

The session store fix keeps the gateway alive

OpenClaw saves session data every time an agent runs, a cron job fires, or a context window gets used. That's how it's supposed to work. The problem was that nothing cleaned those session stores up, so they kept growing.

My setup has 17 cron agents running on a schedule. RZA writes scripts and outlines, Ghostface does intel research, Method Man processes clips, and every one of them starts sessions that add to the store. The longer a setup runs, the bigger the store gets, and the gateway has to load all of it every time it boots. Eventually loading it can use up all the available memory. The gateway tries to start and dies, and you get no warning and no error you can act on.

In 4.20, OpenClaw applies a built-in entry cap and age pruning by default. When the session store loads and it's too big, it gets trimmed at startup. Old sessions go first, so you lose dead weight and keep what matters. It reads like boring cleanup in a changelog. But if you've been running agents for weeks or months, this fix stops your whole operation from falling over at the worst possible moment. For me this is real production for my business, not a toy setup.

Completion bias: agents that actually finish

You've probably seen this one. You give an agent a task and it stops partway through, or it does the first part and says it did everything, or it trails off and you have to prompt it again. 4.20 adds stronger defaults built around completion bias, meaning the agent is guided to finish the job before it wraps up. That includes:

  • Live state checking
  • Weak result recovery
  • Verification before the final response

So the agent can't decide it's done just because things seem fine. It has to check that the output is what it was supposed to produce. In my pipeline, I need the full script from RZA, the full intel report from Ghostface, and a complete newsletter from the agent that sends it, not a summary of what a newsletter would look like. Agents fall short of that more often than they should, and this is a real step toward fixing it. If you're new to how agents like these work, my AI Agents page is a good place to start.

Kimi K2.6 and seeing what each model costs

Kimi K2.6 is Moonshot AI's frontier model, and it's now built into OpenClaw. Token usage reports now include cost estimates for both K2.6 and K2.5. That means you get a capable, affordable model inside OpenClaw with no extra setup. It also supports thinking, and you can keep the full reasoning trace if you need it.

That matters if you're running agents on a budget. Claude Opus and GPT-5.4 are powerful, but at scale they cost real money. With Kimi K2.6 showing up in your cost reports, you can send some tasks to a cheaper model and still see what each one costs. I've been doing this kind of routing by hand for months: I use OpenAI for scripts and heavy generation, and Claude for the main session. I still use Opus 4.7 for the main brain, because for me the better output is worth more than the savings. OpenClaw now makes this routing easier and more transparent, because you can see costs broken down by model.

Cleaner cron files, script exec, and smaller fixes

Before 4.20, cron runtime execution state was stored in jobs.json next to your job definitions. Every time a job ran and updated its state, that file got rewritten. If you track your agent setup in Git (and you should), every commit picked up timestamps and in-between state you don't care about. Now job definitions stay in jobs.json and runtime state goes in a separate jobs state file. Your definitions stay clean and stable, so in version control you can tell a change in logic apart from a runtime leftover. When something breaks, that clean history can be the difference between a 10-minute fix and a 3-hour nightmare.

A few other fixes:

  • Exec fix: In security "full" mode with "ask" turned off, a bug in the preflight hardening path rejected some Python and Node script forms that should have run. Direct interpreter, stdin, and heredoc forms now work. If you've had unexplained script failures in some configurations, this may have been the cause.
  • Onboarding wizard: A better security disclaimer, a loading spinner while the model catalog loads, and cleaner API key prompts. Confusing setup is the main reason people quit before they run their first agent.
  • BlueBubbles group chats: You can now set a system prompt for each group, so each group's own behavior instructions get added automatically on every turn.

You can find more about running these setups on my OpenClaw page.

Key takeaways

  • Session stores now get an entry cap and age pruning by default, so they can't grow until they crash the gateway.
  • Completion bias defaults (live state checks, weak result recovery, verification) push agents to finish and check their work before they report back.
  • Kimi K2.6 is built in, and cost estimates in usage reports make it easier to see where your money goes when you route tasks across models.
  • Cron runtime state is now stored separately from jobs.json, so your job definitions stay clean and easy to audit in Git.
  • The problems that hurt most are the ones that build up quietly, and this update targets several of them.

FAQ

What is the biggest change in OpenClaw 4.20?

For heavy setups, the session store fix. OpenClaw now applies an entry cap and age pruning by default and trims old sessions at startup, so a session store that keeps growing can't use up all the memory and crash the gateway.

What does completion bias mean in OpenClaw?

It's a set of stronger defaults that guide the agent to finish a task before it wraps up. It includes live state checking, weak result recovery, and verification before the final response, so the agent checks its output instead of stopping early.

Why did OpenClaw move cron state out of jobs.json?

Runtime state used to rewrite jobs.json every time a job ran, which filled your Git history with timestamps and other noise. Now runtime state goes in a separate file and jobs.json only holds your job definitions.

Can I use Kimi K2.6 in OpenClaw?

Yes. In this update Kimi K2.6 is built into OpenClaw with no extra setup, and token usage reports include cost estimates for Kimi K2.6 and K2.5.

Transcript

Read the full transcript

0:00There is a specific kind of bad that just hits different. Not the kind of something that explodes and you know it, but the kind where everything looks fine on the surface. You know, your dashboard is green, the logs are quiet, and then one morning you wake up and your whole Open Claw setup is broken. Your agents stopped working, your cron jobs fire off, but they silently fail. Your gateway is dead. Have you been there before? I know I have. And basically you have no idea what actually broke your system because it was breaking slowly, quietly, a little bit at a time, and a little bit more every single day while you were sleeping. So, with this newest Open Claw 4.2 update, a lot of this has been fixed. A lot of Open Claw setups had failed silently in the past before.

0:45And if you're running a heavy pipeline, if you've got agents firing off every morning, or if your setup looks anything like mine, this update just saved you from basically your setup crashing all the time and you didn't even know why it was crashing. I'm going to break down exactly what changed in this 4.2 update. So, the main thing about session stores, every time your agent runs, every time your cron job fires off, every time a context windows get used, Open Claw is storing that session data. That's normal, and that is how it works. But what nobody was watching, what was just building up quietly in the background, is those session stores were not getting cleaned up. They just kept growing. And that could be a huge issue. So, if you think about my setup, I have 17 cron

1:32agents running 24/7. Every single morning they fire off. The RZA, who writes the scripts in the outline, Ghostface does the intel research, Method Man processes all the clips, and all of them running on a schedule. All of them spinning up sessions. All of them adding to a session store that nobody was pruning. And the longer a setup has been running, the bigger that store gets. The bigger that store gets, the more memory the gateway has to load every single time it boots up. At some point, and this is the part that actually scares me, you know, at some point the session store gets big enough that loading it takes all of the available memory. So, the gateway tries to start, it chokes on its own history, and then it just dies. There is no

2:17warning, there's no error you can actually act on, it just dies. Real quick, before we keep going, if you're watching this and you want to actually build with some of these tools, not just watch videos about them, then you're going to want to check out our community down below, Shipping School. We have a full Cloud Code course, a full Open Cloud course, and four live boot camps every single week, where we actually help you get set up from scratch. Like actually set this thing up, not just watch a tutorial and figure it out by yourself. And we also provide one-on-one coaching, so you could book a call with me, we could share screens, and I can help you get Cloud Code or Open Cloud running on your machine. That's it, no fluff. I built this community because watching YouTube's only gets you so far. We launched it just 3 days ago, and we

3:03have over 55 members. You need people around you who are actually building, people who hold you accountable, and coaches who can help you when you get stuck. I'll put the link in the description down below. Get in now before the price goes up. Open Cloud 4.2 enforces a built-in entry cap and age prune by default. So, when the session store loads, or if it's too big, it gets trimmed right there at startup. Old sessions go first, you keep what matters, and you lose what was just dead weight anyway. And the gateway stays alive. This is the kind of fix that looks completely boring on a change log, but session maintenance is super important for reliability. There's an entry cap, there's age pruning, and it sounds like a lot of database janitor work, but if you're running a real

3:49pipeline, if you have been running agents for weeks or months, this is the fix that keeps your whole operation from going sideways on you at the worst possible moment. We are talking about 216 people in our community waiting on content at $14,000 of monthly recurring revenue that depends on the pipeline running. 17 agents that need to fire off every morning. This isn't a toy setup for me. This is an actual production element to my business. And it was just sitting on a time bomb that nobody told us about. want to talk about something else in this update that I think gets really underrated is the completion bias. So, you have probably seen this. You run an agent, you give it a task, and it gets most of the way there, and then it just stops. Or it does the first

4:34part and reports back that it did the whole thing, or it kind of trails off and you have to re-prompt it to actually finish. This has been a persistent thing with AI agents. They start strong, they lose the thread, they declare victory early, but 4.2 puts a stronger default system in place. And it specifically bakes in completion bias. So, what that means is the agent is guided to actually finish the job before it wraps up. There are live state checking built in, weak result recovery, verification before the final response. So, instead of the agent deciding it is done based on just the good vibes, it has to actually check its work and make sure the output is what it was supposed to be. So, for a pipeline like mine, this matters a ton. So, we'll

5:20just take for example, I have an agent called the Rizza. He writes the scripts, the outlines. I need the full script. When Ghostface pulls the intel, I need the full report. When Jizza sends the newsletter, it needs to be a complete newsletter, not most of it, not just a summary of what a newsletter would look like. The entire thing. And right now, the agents fall short of that more than they should. The completion bias changes 4.2 are real, and that's a real step towards fixing all that. So, if you've ever had to chase down an agent and actually ask it to finish what it started, you know exactly what I'm talking about. Want to talk about another cool feature here. Kimmy K2.6 just came out the other day, and if you're not familiar with Kimmy, it is a moonshot AI model, their frontier model that scored really well on benchmarks.

6:05And it is now the default surface in Open Claw. And the token usage reports now include bundled cost estimates for both 2.6 and 2.5. So, what this means practically is now you have a capable affordable model available right inside of Open Claw without having to do any extra setup. Kimmy K2.6 supports thinking. You can keep all, which means you can basically get the full reasoning trace if you need it. For anyone watching who is trying to run agents on a budget, this is a big deal. Claude Opus and GPT-5.4 are powerful, but they cost real money at scale. Having Kimmy K2.6 as a bundled option that shows up right cost reports means you can start routing certain tasks to a cheaper model without you

6:50flying blind on what is actually costing. I've been doing this manually for months, running scripts and heavy generation on Open AI, while keeping Claude for the main session on the brain. I use Opus 4.7 still because I think the output is just way worth more than me cutting costs and having to fix problems with other models. You know, Open Claw is now making that kind of model routing easier and more transparent for everyone. You can see your costs broken down by model, figure out where your money is actually going, and make smarter decisions about what runs where. There is also a cron infrastructure improvement that I want to highlight because I think this matters for anyone who's serious about building this stuff long term. So, before this latest update of 4.2, when your cron jobs ran, the runtime

7:35execution state was mixed in with your job definitions in the jobs.json file. That means every time your job fired off and updated its state, your jobs.json file was getting written to. And if you are tracking your agent setup in Git, which you absolutely should be, that means that your job definitions were constantly getting noise from runtime state. Every commit was polluted with execution timestamps and intermediate state you did not actually care about. 4.2 splits that out. Job definition state in jobs.json, runtime execution state goes into jobs state.json. So, now your jobs.json stays clean, stays stable, and stays Git trackable. You can look at your job definitions in version

8:21control and actually see what changed in the logic versus what was just a runtime artifact. This is infrastructure maturity. This is the kind of thing that matters when you're running 17 agents and you need to be able to audit and roll back your setup without waiting through the noise. That sounds like a small thing. It doesn't, you know, seem like a big headline, right? But when something breaks and you need to debug it, if you're struggling to keep up with content, well, I'm about to save you about 40 days worth of work. I built something called content machine. It's [music] 10 AI agents that run on the OpenClaw orchestration, and they handle everything. Scripts, thumbnails, X posts, blogs, outreach, clips, newsletters, all of it. So, I went from

9:071,000 subscribers to 4,000 subscribers on YouTube in 7 days using this exact system. Every single morning, I wake [music] up and the content's already done. I spend maybe 15, 20 minutes reviewing and approving them, and I move on with my day. It works for any niche, fitness, [music] finance, real estate, marketing, whatever you are building, and it is 100% completely customizable to your use case. So, you get the mission control dashboard, all of the cron jobs, everything I've built over the last 40 days helping me gain more and more people to subscribe and join the community. So, you plug in your own thing and it molds it to you, it learns how you talk, and it writes so it doesn't sound like AI slop. $97 one

9:52time, it's not a subscription, I'll put the link down below and you'll thank me later. Having a clean history of what your job definitions look like is the difference between a 10-minute fix and a 3-hour nightmare. One more thing worth mentioning is the yellow exact fix, let's go. If you have ever had open claw refuse to run a Python or node script through exec in a mode where it definitely should have worked, that was just a bug in the pre-flight hardening path. In security full mode with ask off, certain script forms were getting rejected even when they should have sailed right through. 4.2 fixes that. Direct interceptor stdin and heredoc forms now work correctly. If your agents are doing any kind of script execution and you had mysterious failures in

10:37certain configurations, that might have been the whole problem this whole time. Also, the onboarding wizard got cleaned up in this release, you know? Better security disclaimer, a loading spinner when pulling the model catalog, cleaner API key prompts. That stuff matters for anyone setting up open claw for the first time. Confusion at startup is the number one people bounce before they ever get to run their first agent. Every improvement there means more builders actually get through the setup and build something real. There's also a fix for blue bubbles group that chats that let's you basically set the system prompt per group. So, if you are using open claw to power agents in different group contexts, each group can have its own behavior extractions injected automatically on every turn. That is a small thing for most people, but huge if you're using it in that way. So, to pull

11:23it all together, if you are just casually curious about AI, 4.2 is a nice update. A few improvements, a few fixes, and then move on with your day, you know? If you are actually building with this, if you have agents running every day, if your business depends on the pipeline staying alive, 4.2 is more than a maintenance release. The session store fix alone changes the risk profile of running a heavy setup. The completion bias improvement changes the reliability of what your agents actually produce. The cron split state changes how your maintainable of your infrastructure and how reliable it becomes more over time. You know, I've been building in public for a while now for these 90 days showing you my open class setup, 17 agents. We have 215 members in our school community I launched 30 days ago.

12:10We crossed about 14,500 in monthly recurring revenue and the single biggest lesson is that problems that hurt you the most are not the ones that blow up in your face. They are the ones that build up quietly and you just don't know it and they take you down when you're not looking. You know, a session store that keeps growing, an agent that half finishes a job, a cron configured that becomes impossible to audit. That's a nightmare. 4.2 addresses all that and I think that deserves more attention than it's getting right now. If you want to build something like this for yourself, you know, if you want to run real agents, real pipelines and actually get to the point where AI is doing meaningful work in your business and you want it to do it every single day, come join us in the community. I'll put the link down below. We actually have nine boot camp calls a week and one

12:55of the calls I'm excited about is setting up local models. We've had messages of how do I set up I got a DGX Sprout, I got a Mac Studio, but I don't know how to install local models. Well, every Sunday now we have a class for that where we actually go through the basics, the concepts and the principles of getting local models running on your computer. So, like I said, I'll put it down below, nine other calls throughout the week. We show like my actual pipeline, how I'm running things, keeping you up to date, what breaks, how to fix it, how to build something that runs while you sleep and we'd love to have you there. And if you haven't subscribed to the channel already, please do cuz I release videos just like this keeping you up-to-date. Hit the notification bell. We'll see you in the next one. Have a blessed day.

Ready to start building with AI?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

Keep reading