@DimitrisPapail Good idea, although I’m sure you can implement this with a companion-writer skill and a sub agent that after each step invokes the skill to document things as you want

Happy to vibe something up for you and see if that doesn’t solve your use case

@reach_vb maybe but the point is many people, including myself, are lazy, and this is reflected as bad user experience. Probably an easier fix from your side :)

Also please add timestamps on messages :D

@reach_vb and yes if you have a quikc fix would love to try it out.

@DimitrisPapail hey dimitis, would you be open to trying a different autoresearch orchestrator ? it keeps a track of all experiemnts tried / hypothesis / specific diffs etc . is compatible with codex as well off the shelf.

would love your take and feedback! thanks http://github.com/evo-hq/evo

@DimitrisPapail codex is on it! if this works, we can ship this as an official plugin :)

@DimitrisPapail You could try developing a program using Codex that allows the model to progressively organize the story within a context window segment, while retaining a spare window for the model with accurate memory of the current segment to absorb information from the next segment

also another thing that's been bothering me and may be how i've setup the app or something: a git subprocess is always eating 100% of cpu, even when i work with remote gpus. I may have done something idiotic, but it's a codex app only issue (eg i dont have it with claude code cli etc). Any ideas on what this would be?

@DimitrisPapail you can just recover the sessions from e.g. ~/.codex and reconstruct a story, no?

@DimitrisPapail Curious if you can say more about your autoresearch harness. This got a lot of publicity when Karparthy suggested it, but haven’t seen the community move to a clear set of best practices around it (memory, logging, orchestration).

@DimitrisPapail Yeah we have a great case study.

Open agentic science achieved more than a 13x improvement in a Quantum algorithm. But its really hard to pull through what the models actually did.

There is a lot of data to study here on frontier scientific research.

@DimitrisPapail Curious what prompts did you use? I recently found that if you ask it to narrate what happened it can do it quite well

@DimitrisPapail "Commit everything with a detailed writeup of what you changed, what you expected, and what you actually found" and then "review the commits and tell me what moved the needle the most" works pretty well

@DimitrisPapail For some sessions, I have Codex turn raw session logs into a narrative, then generate graphics, audio, or even a video from it (using local tts and simple graphs). It’s worked surprisingly well. And now with imagegen tool, I can see this becoming an even better pipeline.

@DimitrisPapail I read the reasoning trace and the steps it takes to get there, and I give it guidance on that

@reach_vb @DimitrisPapail Why not show thinking summaries in codex app

@reach_vb @DimitrisPapail Created a skill for exactly the case.

@DimitrisPapail I mean you can just have a line in AGENTS.md to ask it to read the full trajectory on the disk and summarise, right?

@DimitrisPapail If you enable the memory feature you will find a good summary of your session in the .codex directory

@DimitrisPapail Hit this same wall. Model finds the answer but can't narrate how it got there. Companion writer makes sense - search and storytelling need different capabilities.

@DimitrisPapail @reach_vb thats gonna be spendy on the tokens for the context, even tying the chain together, better to have a subagent not your best brain put that mess together, and tracking logic for a24 hours run is gonna be some context...