I like the distinction between a living specification and the conversational residue produced while implementing it. A spec can hold intent and constraints for humans to review, while the harness can avoid re-feeding every exploratory read, duplicate test output, and tool acknowledgement on the next turn. For the latter experiment, I’ve been looking at Contextpress (GitHub - Taha-azizi/contextpress: Deterministic context compression for LLM chat, RAG, and agent pipelines · GitHub · contextpress · PyPI), which applies deterministic local compression to the message list. I’d judge it by whether the agent stays aligned with the spec and whether retries or clarification turns increase, not by context reduction alone.
1 Like