@__tosh Yep, I think truncation’s job would be not to help the model choose the most important line in the output, but just to safeguard the model from accidentally spamming its context with 10 MB of random program/webservice output
@__tosh Yeah. I think truncating the bottom is also very natural to how humans work. When I open a large file in a viewer/editor, I only see the top X lines that fit on my screen — to give me a better idea if it’s the file I need, and whether I need to scroll down or search for something
I've been wondering: why would any harness ever have an unstable prefix (apart from person going back to a previous message to edit it and thus fork a conversation in another direction, still retaining part of the stable prefix up to that fork point)?
Do other harnesses routinely change something in history for some reason?
To be fair, sh is so effective, most of the time shell-level parallelism is strictly better and much simpler than running subagents for most practical tasks :D
I just wanted to try making it self-recursive and asynchronous with parallel goroutines, and also to model the user so that it's "kind of" another agent that talks to the top-level agent :)
@__tosh Since you rewrote it in Go, I had to scratch an itch to make it support parallel hierarchical subagents with supervision trees just for fun https://t.co/dLWdiih7Xl
@__tosh I love your "smol" harness by the way. Being an experienced engineer, I was still treating Claude Code/Codex like some kind of "magic" that I don't need to think about. Your Python script opened my eyes :D
@__tosh You could do it in English like OpenAI did with Symphony: https://t.co/J5oAFrJZTk — the source code is the spec, "ask your favorite harness to build it in your favorite language".
Would be cool if your minimal harness could reliably build itself from its own English spec! :D
@vertinski@JoshuaSuasponte@sammyhill99@skdh@StuartHameroff Good questions. I don’t have a positive definition either. I could approximately describe it via negativa as “something that most humans and maybe some animals have, but other living organisms don’t” — though I am not sure it’s something important.
@sammyhill99@skdh@StuartHameroff Then any living cell of any organism is “conscious” in that sense, and we lose the meaning of word “consciousness” by playing with words too much
@ctjlewis I follow just 300 accounts, if even 10% of them post 20 times a day, no way I can scroll through all of their posts to see them. There inevitably should be some banger threshold
@the_dt1@NousResearch Yes, Anthropic turned that off long time ago. You have to use Anthropic’s Agent SDK to use subscription limits, but Hermes Agent does not use it (allegedly for architectural reasons, but I do not know Hermes’s architectural details well enough to gauge how justified that is)