Week 0x2B

Back to the work lifecycle. I can’t believe this is taking so long, it felt like a simple problem to start. There seems to be very little variability across well-known harnesses, which could mean it’s obviously simple and I’m trying too hard. I prefer to think there’s room for some new ideas.

As ever, dogfooding has exposed a lot of weaknesses and so the server/client protocol and handling of tool calls needed attention. The latter had gotten way too complex over time, so we burned down the “obligations” system and build a new one based on a specific session-bound leases.

While I was going, I noticed Builder Bear wasn’t remembering enough, so now the whole memory system is more auditable.