During the next few days, I will share some of my private security research work that I have done in the last year. ​ All those projects are @aave related, and I feel very proud to have been chosen as one of the security partners to review them. I'm pretty sure that the dedication and knowledge that I have demonstrated on their codebase has been a pretty good investment 😁 ​ Those projects were very challenging, but at the same time it was fun and a pleasure because you always feel motivated and engaged when the quality of documentation, specifications and code is high quality. ​ When @avara or @bgdlabs contacts you, you know that you are going to enjoy it. They are both a top-notch team to work with. ​ The first one that I want to share is a.DI (Aave Delivery Infrastructure) It's a cross-chain communication abstraction layer for decentralised systems like the Aave DAO to communicate across networks, minimising the risk of underlying individual bridge provider failures, via consensus rules.

is there a reason why the default Codex behavior to "spawn" agents is with `fork_turns=all` that makes the subagents inherit the whole parent history inherited? Isn't the whole point of subagents to start with a "fresh" context? Am I missing something @OpenAI ?

I was thinking about the blog post "Building agents: When to fold subagents into your main agent loop" from @AnthropicAI this morning. I think that we have settled on the best practice (does it really exist one in this field?) that you have to structure your "agent workflow" depending on the context and the needs and the eval result. But being able to easily test it, evaluate it and maintain it has a cost and the same capability can't be easily embedded and provided as a tool, skill or subagent. I feel like we are at a point that we need a new standard that enables us to define capabilities and then "distribute them" to the agents that can consume them as skills or subagents depending on the context/need we have defined for that agent. claude.com/it/resources/arti…

Do I have any connection with someone from the Codex Cyber Program? When will you enable Daybreak Blue for GPT-6 Astra and GPT-6.1 Sol?

Codex has started to provide data illustrated as YAML instead of JSON during a discussion 🤔 It's the first time I've seen it.

Is it just me or the "suggested next prompt" feature in Codex macOS has stopped working across any chat? I've also tried to restart the app but it seems not to work anyway?

Here we go again... "Error running remote compact task: Selected model is at capacity. Please try a different model." @thsottiaux 😭

TIL that you can ask Codex to "create a new chat" to further discuss something and it will spawn a new chat. I don't know tbh the side effects or what happens under the hood. For example, in a chat where I had a long architecture discussion I've said "I agree with the overall architecture. Record it as agreed, then create a new chat to discuss and design P01, with the architecture, agreed rules, dependencies and relevant research as context." The original chat (which spawned the new chat) will in some way record it visually in the chat. Any guidance on the usage and which are the sideeffects or implications of adopting this behavior @Dimillian ?

I can't test OpenAI Dots because I live in europe but as far as I get you only get "one dot" for the whole account? How does it work internally? I mean it can't be a whole huge infinite context right? They "mask it" as a single conversation I assume (from the UI pov) but what happens under the hood?

How many of the big companies have finally capitulated to make a proper iPad app just to be ready for the iPhone Duo?

I've asked codex a UI/UX suggestion... it has linked a GOV UK guidance page 🤦 Now I understand why Codex is the WORST at design/UI/UX.

Now Codex suggests in the textbox the "following prompt" after that it has finished work. After implementing a request it has suggested "The filters should apply to all results, not just the current page. Implement server-side filtering through core, server, client and webapp." The question is: why hasn't it been done already during the implementation?

Do I know anyone from the @OpenAI Daybreak program or someone that can connect me with them? I'm part of the Daybreak program but my agents are failing. I'm using `gpt-6.1-sol` with my ChatGPT subscription. It appears that the error from the traces is "This content was flagged for possible cybersecurity risk. If this seems wrong, try rephrasing your request. Otherwise, confirm that the access_programs.cyber parameter is set to the appropriate tier, and note that some cybersecurity requests are still limited, even when Daybreak is on."?

In your experience, even if you provide "guidance" via zod schema descriptions on "how to use" a tool and which are the input schemas/rules, how much is the model failing at calling the tool correctly? Is it "expected" that it fails and learns from the experience or should it always one shot it?

Can we get some Dots in the EU for the w/e?

MCP seems to be like Matter. It evolves with new versions, but no one (or at least the big ones like codex, opencode, pi and so on) does not adopt it. Aren't those actors in the MCP org?