Nothing here yet.
Yeah, good catch! I鈥檝e actually taken care of this in Sadhan. I鈥檓 merging "stderr" into "stdout" using "asyncio.subprocess.STDOUT", so both streams are read together before checking for the sentinel marker. But yeah, I can definitely see how this could become a nasty bug with a persistent shell if the streams were handled separately. Thanks for pointing it out! 馃檶
Thanks, Kartik. Yeah, I completely agree with this. Once the harness starts handling things like retries, context trimming, and step limits, those can definitely become failure points themselves. I think the tricky part is figuring out what actually caused a long run to fail, whether it was the model, a tool call, or something the harness did along the way. Logging the model outputs, tool calls, retries, context changes, and step-by-step results would probably make that much easier to trace. I鈥檓 still exploring this part of harnesses, but I think measuring and observing the harness itself is a really interesting next layer. Thanks for bringing this up.
