Static-Asset Workflow

dexter-horthy's pattern: keep the important state of an agent session in files on disk (plans, specs, research notes), not in the context window.

"We don't use the built-in compaction because everything that matters is going into static assets and so you can always resume from where you left off without having to worry about the quality of an autocompact or a manual compact."

Benefits

• Session boundaries become cheap — start a fresh conversation, load the static assets, continue.
• Autocompaction artifacts ≠ your actual state — you never lose work to a bad compaction.
• Multiple agents can read the same static assets (enabling sub-agent parallelism without context sharing).

Why this matters

This is the individual-engineer analog of Bantilan's replay-log: persistent state outside the ephemeral context window. Where Bantilan builds it into Flyte at the infra layer, Horthy enforces it at the human-process layer. Both solve the same problem — context windows are volatile, disks are durable.