context a session moves grows with roughly the square of its length

Readme doesn't mention that LLM providers do context caching — your costs aren't quadratic unless you frequently invalidate the cache.