OpenClaw 4.23 Update: GPT Image 2 via Codex OAuth and Context Forking

By Beau Johnson··8 min read

Quick answer

OpenClaw 4.23 lets GPT Image 2 run through your existing Codex OAuth login, so you no longer need a separate OpenAI API key. It also adds context forking for sub-agents, per-call timeouts on media tools, more than 15 security fixes, and a catalog fix so GPT-5.5 works.

In this video I go through the OpenClaw 4.23 release: what changed, which fixes matter most, and who each change is for. If you use OpenClaw to generate images, run multi-agent pipelines, or just want to know whether the project takes security seriously, this covers it. I recorded this in April 2026. Tools change fast, so check the current docs and changelog before you rely on any of the details here.

What I cover in this video

  • 0:46: GPT Image 2 now works through your Codex OAuth session, and OpenRouter keys now cover images too
  • 2:16: How I run my Wu-Tang-style multi-agent setup, and the problem with handing work from one agent to the next
  • 3:48: sessions_spawn now supports forking context, so sub-agents start with the parent's transcript
  • 5:19: Per-call timeouts on every media tool
  • 6:05: Quality and output format hints passed straight through the image generate tool
  • 6:50: More than 15 security hardening fixes in one release
  • 9:07: Smaller fixes: the memory doctor, dreaming, Windows Codex, and image attachments
  • 12:09: GPT-5.5 support, the Codex cleanup, and why this release matters

Image generation without a separate API key

The headline fix in OpenClaw 4.23 removes a big source of setup friction. Before this release, you could pay for ChatGPT Pro, sign in with Codex, and have your text models working fine. Then you'd try GPT Image 2 and OpenClaw would ask for a separate OPENAI_API_KEY. That meant another billing line, another key to rotate, and another place for things to break without telling you.

In 4.23, GPT Image 2 works through your Codex OAuth session. If you're already signed in, the image tool picks up that authentication. OpenRouter got the same treatment: one OpenRouter API key now covers both text and image generation, with no extra configuration. These changes fix issue 70703 on the OpenAI side and 55066 for OpenRouter.

For me this is a big quality-of-life win. Members of my community message me almost every week saying their image generation is broken, and nine times out of ten the answer was that they needed another key. Now I can just tell them to sign in with Codex. If you're new to Codex, I have more on it at OpenAI Codex.

Context forking for multi-agent pipelines

This one is for power users. I run a Wu-Tang-style team of agents. Ghostface gathers intel, now through the X API. The RZA writes scripts and outlines. U-God handles thumbnails, X posts, and cold outreach on X. Then work gets passed on to Method Man. Each agent hands off to the next.

The problem has been that when a parent agent spawns a sub-agent, the child starts blind. It gets no transcript and no shared context. You have to copy the relevant parts of the parent's conversation into the child's prompt every time. Either you give it too much, or you leave out something important and it goes off the rails.

In 4.23, sessions_spawn supports forking context. You set a flag on the spawn call and the child inherits the requester's transcript. Isolated sessions are still the default, so your existing flows won't break. In my setup, every handoff gets cleaner because the next agent already knows what the previous one was thinking. If you're building agent chains, my AI Agents page has more.

Media tool timeouts and image quality controls

Before 4.23, a long generation like a 4K image or a longer music clip could hit the provider's default timeout, and the whole call would fail. You lost the generation and had to start over, which hurts when you're paying per call. Now every media tool (image generation, video generation, TTS, and the rest) takes an optional timeout parameter. You can raise the timeout only on the call that needs it and keep the short defaults everywhere else.

Image generation agents can also pass provider-supported quality and output format hints straight through the image generate tool. For OpenAI, that means you can set background, moderation, and compression. These used to require going to the raw API. If you generate thumbnails, banner art, brochures, or product photos, this gives you much more control.

Security fixes, smaller fixes, and GPT-5.5

4.23 ships more than 15 security hardening fixes, which is above the weekly average. Here are some of them:

  • Teams bot framework tokens now verify app ID and azp, which blocks cross-bot token replay.
  • Android cleartext gateway connections fail closed unless TLS is present. Mobile pairing requires a private IP or loopback host, and .local cleartext is no longer allowed, so nobody on your Wi-Fi can hijack the pairing flow.
  • Webhook secret rotation takes effect on openclaw secrets reload instead of needing a full gateway restart.
  • A privilege escalation bug in the MCP bridge is closed. Before the fix, non-owner callers could invoke owner-only tools like cron jobs.
  • The Discord slash command channel policy no longer bypasses owner and member restrictions, and QQ bot approve now requires framework authentication.
  • Agent-driven gateway config apply now uses an allowlist of narrow, tunable paths instead of a deny list.
  • Secrets on Windows now strip the UTF-8 byte order mark and fail closed when ACL checks can't be performed.

You're giving OpenClaw access to your keys, files, gateways, and bots. A release with this many security fixes tells me the core team is reading and triaging issues.

Smaller fixes I cover:

  • Running OpenClaw doctor with the fix flag now merges duplicate root memory files into the memory markdown file and makes a backup first. This ends a silent data-loss problem. If your agent ever forgot things it should have known, this may have been why.
  • Dreaming, the background pattern extraction feature, is no longer tied to the heartbeat. It runs on its own managed cron job, so turning off the heartbeat no longer silently turns off dreaming too.
  • On Windows, codex.cmd shims now resolve through PATHEXT, so Codex models work without a manual workaround.
  • If your primary model is text-only, the gateway now keeps image attachments as media refs instead of dropping them, so downstream image tools can still read the original file.

OpenAI released GPT-5.5 the day before 4.23. When Codex catalog discovery leaves out the openai-codex/gpt-5.5 model row, OpenClaw now adds it itself, so cron jobs and sub-agent runs don't fail with an unknown model error. GPT-5.3 Codex Spark is no longer listed, and stale rows now point you to GPT-5.5. The Codex CLI auth path has been removed from onboarding and provider discovery, and Codex runtimes now use CODEX_HOME or ~/.codex directly. That leaves fewer places for auth state to get out of sync. There were 27 contributors on this release.

For more on the platform, see my OpenClaw page.

Key takeaways

  • GPT Image 2 now works through your Codex OAuth login, and one OpenRouter key covers both text and images.
  • Setting the fork flag on sessions_spawn lets a sub-agent start with its parent's transcript. Isolated sessions are still the default.
  • Media tools take an optional per-call timeout, and image generation can pass OpenAI's background, moderation, and compression settings.
  • More than 15 security fixes, including an MCP bridge privilege escalation fix, make this a strong release on security.
  • My advice: if you run OpenClaw, update. The image auth fix alone is worth it. (When I recorded this I was still on 4.21.)

FAQ

Do I still need a separate OpenAI API key for image generation in OpenClaw 4.23?

No. In 4.23, GPT Image 2 works through your Codex OAuth session, so if you're already signed in with Codex, the image tool uses that login. An OpenRouter API key now covers both text and image generation too.

What does context forking in sessions_spawn do?

It lets a sub-agent inherit the requester's transcript when you set a flag on the spawn call, so the child agent doesn't start with no context. Isolated sessions are still the default, so existing flows keep working.

Does OpenClaw 4.23 support GPT-5.5?

Yes. If Codex catalog discovery leaves out the openai-codex/gpt-5.5 model row, OpenClaw adds it itself, so cron jobs and sub-agent runs don't fail with an unknown model error. Stale GPT-5.3 Codex Spark rows now point you to GPT-5.5.

How do I fix duplicate memory files in OpenClaw?

Run OpenClaw doctor with the fix flag. In 4.23 it merges duplicate root memory files into the memory markdown file and makes a backup, which fixes a problem where data was being lost without any warning.

Transcript

Read the full transcript

0:00Openclaw 4.23 just came out and here is everything you need to know. I know it's just been a few days since the latest release, but this one is actually feature packed. So, this one literally just dropped an hour ago and the headline feature is something every single person watching this channel is going to love because it kills one of the most annoying bits of friction in the whole Openclaw setup in general. So, let me break it down. If you've ever tried to just get image generation working inside Openclaw, it's been a pain in the butt. You know, you already pay for ChatGPT Pro plan, you've already authenticated with Codex, and your text models work fine, but your coding agent might work fine. But, the moment you want to generate an image with the new GPT image 2, it slams the door in your

0:46face and it asks for a separate OpenAI_API key and that's a pain in the butt. So, it's just another billing line, another key to rotate, another spot where things can silently break. It's not stupid, but I'd say it's pretty dumb. Like, the cool thing is is the Openclaw team finally fixed this. So, in 4.23, the GPT image 2 works through your Codex OAuth session. So, it's no separate key. You already log in, you already paid for Pro, and now the image tool just works. And they did the same thing for Openrouter. Your Openrouter API key now unlocks text and image generation through the same key. One key, two capabilities, and then there's zero extra configuration that needs to go into it. This fixes number

1:3170703 for the OpenAI side and 55066 for Openrouter. So, if you want to go read the actual commits, I have it linked down below. For me, this is a huge quality of life win because I have had school members in the community ping me literally every week going, you know, my image gen is broken and nine times out of 10, the fix was just, you know, you have to go with another. But, you know, I do not have to say that anymore, which is really nice. You just log in with Codex and the image tool inherits the authentication, which is pretty awesome. I I love that. Just keep it simple. And it is the kind of cleanup that separates a tool that feels finished from a tool that feels like a side project. So, moving on to the second feature in this

2:16release, this is for power users, the people who are actually building multi-agent pipelines on top of Openclaw. So, I know, if you're new to the channel, I run a Wu-Tang style setup where Ghostface, my Intel agent, gathers intel, specifically now through the X API, and the RZA writes the scripts and the outlines, and U-God handles the thumbnails, post for X and doing cold outreach on X, and they hand off work to each other. So, if you have ever tried to build something like that, 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 [music] have a full Claude code course, a full Openclaw

3:02course, and four live bootcamps 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 Claude code or Openclaw 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 have 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. You have ran into this wall where when a parent agent spawns a

3:48sub-agent, the child starts completely blind. It has no idea what the parent was working on, no transcript, no shared context, literally nothing. You have to manually stuff the relevant pieces of the parent conversation into the child prompt every single time. So, that is tedious and it makes the handoffs messy because you are either overstuffing the child with the stuff it did not need, or you are forgetting a critical detail and the child goes off the rails. But, in 4.23, sessions now spawn supports fork the context inheritance. So, basically, you flip a flag on the spawn call and now the child inherits the requestor's transcript. And isolated sessions still stay on the default, so you're not

4:33breaking existing flows. This is the feature serious multi-agent builders have been asking for. For a setup just like this, Ghostface, for example, hands off the intel to the RZA, and then RZA hands off all of that information to U-God to generate the thumbnails. And then U-God hands all of that off to Method Man. Every handoff gets cleaner because the child already knows what the parent was thinking. There's no more jamming the whole conversation into a new prompt, no more agents that feel like they are just starting from scratch every 5 minutes. So, if you are running any kind of agent chain on Openclaw, you're going to feel this one instantly. The third big thing in this release is the quality of life fix that probably should have been there from day one. It

5:19is the per-call timeouts on every media tool. So, image generation, video generation, TTS, basically all of it. Before 4.23, if you kicked off a long generation, say a 4K image or longer music clip, and it hit the default provider timeout, the whole call just died. You lost your work, you lost the generation, and you had to try again. This was very frustrating, especially when you're paying per call. Now, every media tool takes an optional timeout parameter, so you extend the timeout only on the specific call that needs it, and you keep the short defaults everywhere else. So, if you know this one image is going to be expensive or slow, you bump the

6:05timeout just for that call, and then you leave the rest of your setup snappy. And then you stop losing the work to arbitrary provider limits. On top of that, image generation agents can now pass provider-supported quality and output format hints straight through the image generate tool. So, for OpenAI, you can now pass background, moderation, and compression parameters. All the stuff that used to require you to drop down to the raw API, now it is exposed at the tool level, which means your agents can actually take advantage of it. If you are using Openclaw to generate thumbnails, banner art, brochures, product photography, basically anything where you care about the output quality

6:50and format, you're going to get much more control out of this release than the previous one. Now, the fourth thing I want to talk about, and this is the one that tells you something real about the what the Openclaw core team is thinking about. 4.23 ships more than 15 security hardening fixes in this single release. That is above the weekly average and, you know, I want you to pay attention to this pattern because this is how you tell a serious project from just a weekend experiment. So, you know, some of what landed is Teams bot framework tokens now verify app ID and AZP, which basically blocks cross-bot token replay attacks. Android connections cleartext gateway connections now fail closed unless TLS is present. So, basically, mobile

7:36pairing requires private IP or loopback host. No more .local cleartext, so nobody can sit on your Wi-Fi and hijack the pairing flow. This is huge. Webhook secret also got a huge no-rotation pass. Now, it takes effect on Openclaw secrets reload instead of waiting for a full gateway restart, which means you can rotate faster and not take downtime. The MCP bridge had a privilege escalation bug, completely closed now, where non-owner callers can invoke owner-only tools like cron jobs. That is the kind of thing that, you know, if you had shipped in production and gone unpatched, you could have bitten somebody off hard. Discord {slash} command channel policy no longer

8:22bypasses owner and member restrictions. The QQ bot approve now requires framework authentication. So, agent-driven gateway config now apply is on the allow list. So, with a narrow set of turnable paths instead of a deny list, which is the correct way to design that sort of thing because deny lists have always had had holes, and secrets on Windows now strip the UTF-8 byte order, which marks and failed closed when the ACL checks cannot be performed. So, I know it's a lot of jargon and, you know, I'm listing all of this because you need to go study it yourself. You do not need to like, you know, gloss this over because when a project drops 15-plus security fixes in a single release, that tells you the core team is

9:07actually reading the issues and they're actually triaging it and then actually taking it seriously. And that's a trust signal. You know, you're running Openclaw on your machine, you are giving it access to your keys, your files, your gateways, your bots. You're giving it a lot of confidence, right? And you want it to be on a project where the team ships security hardening like groceries, right? Not like a fire drill. So, 4.23 is that. You know, a few secondary things I want to touch on real quick is the memory doctor got a split brain fix. So, if you run Openclaw doctor with the fix flag, it now merges duplicate root memory files into the memory markdown file with a backup. This ends a silent data loss mode when, you know, that was hitting people who had multiple memory files floating around. If you have ever

9:52noticed your agent seem to forget stuff it should know, this might have been why. Dreaming, which is a background pattern extraction thing, got decoupled from the heartbeat. If you're struggling to keep up with content, what I'm about to save you about 40 days worth of work. I [music] built something called content machine. It's 10 AI agents that run on the open claw orchestration and they handle everything. Scripts, thumbnails, X posts, blogs, outreach, clips, newsletters, all of it. So, I went from 1,000 subscribers to 4,000 subscribers on YouTube in 7 days using this exact system. Every single morning I wake up and the content's already done. I spend maybe 15, 20 minutes reviewing and

10:37approving them and I move on with my day. It works [music] for any niche, fitness, finance, real estate, marketing, whatever you're building and it is 100% completely customizable to your use [music] 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 [music] so it doesn't sound like AI slop. $97 one time, it's not a subscription. I'll put the link down below and you'll thank me later. So, if you had heartbeat disabled for any reason, dreaming was also silently not running. And now dreaming runs on its own managed cron job. The codex harness got a Windows fix where

11:23codex.cmd shims resolve through path text. So, now codex/starmodels work on Windows without a manual.exception. So, that one was a pain. It was basically for the Windows users in the entire community, that was the problem. The gateway now preserves image attachments when your primary model is text only. Before, if you had attached an image to a text first model, the gateway would just drop the image and your image tools cannot see it. Now the gateway offloads those attachments as media refs instead of dropping them. So, your downstream image tools can still inspect the original file. I think that's a pretty good fix. Here's the one that is very timely as well. Yesterday open AI obviously dropped GPT 5.5. I

12:09made a video about it. You should go watch it if you missed it. With two 4.23 open claw, the open AI codex/GPT 5.5 model row gets synthesized into the catalog when codex catalog discovery leaves it out, which means your cron jobs and your sub agent runs that want actual GPT 5.5 do not blow up with an unknown model error while your account is authentic. That is the team responding to a model launch inside of 24 hours with a real fix. So, that is shipping live in on hand, right? On the remove side, the GPT 5.3 codex spark and codex model is no longer advertised. Stale rows get suppressed with GPT 5.5. Recovery hint, so, you know, if you were pinned to your spark, your agent gets

12:55pointed to a working upgrade instead of a giant just dying. The codex CLI off path is gone from onboarding and provider discovery. The ACPX no longer materializes off JSON bridge files for codex runs. Codex runtimes now use the normal codex_home or tilde/ .codex off path directly. So, cleaner, less surface area, fewer places for off state to drift out of sync. So, to wrap this whole thing up, why does this release matter in the bigger picture? 4.23 lands one day after open AI's 5.5 launch and the model row synthesis is a deliberate response so so that you can use GPT 5.5 through your existing codex oauth without touching

13:40configuration. The image gen off cleanup is a direct reaction to how much people have been using the new image two model from open AI since it launched. And instead of making everyone pay for a separate key, the team just wired it through your off you already have. And 27 contributors on a single release tells you that this project has real community momentum, not just one guy shipping in a garage eating crayons in the corner. So, my take is if you're running open claw, just update it today. The image gen off fix alone is worth it. For anyone who has ever fought with open AI keys, do it. If you are running multi-agent pipelines, you know, your sessions_spawn with context inheritance is going to change how you build. And if you have been on the fence about open claw

14:26because you were not sure if this project was actually serious, 15 plus security fixes in one release, 27 contribu- contri- contributions in the same day is a response to a major model launch. All the signals you would want are there. You should go check it out. I put the link in the description down below. Check out the change log, run the update and let me know what your comments are. I want to know. Feedback. What model are you on still? You know, I have yet to update. I'm still on 4.21, but this is a type of thing where it just the agent workflow gets a huge upgrade. You know, if you want to see more open claw builds and you want to be around other people who are actually building and shipping instead of just talking about AI, we run a community called Shipping School. it about 32 days ago. We have 210 members in there and we do nine boot camp calls a week. So,

15:12check it out. I'll put the link in the description. If you haven't subscribed to the channel already, please do because we release videos like this every single day. 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