86%
cheaper with prompt caching
A full CLI auth system, planned and built in 28 minutes
OAuth 2.1 PKCE, token management, shell hooks, and a curl installer — graph-scanned, blueprinted, and implemented in a single session. $12 in tokens instead of $88, because the graph reads files straight into a warm cache instead of spawning cold exploration agents.
A developer needed to add authentication to an existing plugin suite — OAuth flow, token storage, a unified command, and a one-line installer. Composure scanned a 4,155-node code graph in eight queries, asked four scoping questions, and wrote a 500-line per-file blueprint before a single line of code was implemented.
Every file matched the spec on the first attempt — no backtracking, no “actually, that should be different” rewrites. Because the graph fed files directly into the conversation, 97.7% of every API call's input was served from cache at a 90% discount: $12.17 of tokens instead of the $88 the same work would cost uncached.
- Graph-first reads stay cached; agent spawns start cold every time
- Planning catches design flaws before code — rework is the expensive part
- 8 graph queries replaced 20+ grep-and-read cycles
This is what Pro delivers.
Not features for their own sake — measurable leverage on every session.