ompsure
Cache economics

98.8%

cache hit across a 9.5-hour session

The session that wouldn’t fit in one context — or one budget

One campaign drove a build for 9h30m active across nearly 13 wall-clock hours, ran out of context twice, and resumed both times from the persistence engine’s state file — closing with a versioned release shipped to every install. $1,061 actual against $7,764 uncached.

98.8%
Cache hit rate
$1,061
vs $7,764 uncached
149
Commits, one session
9h 30m
Active, 2 compactions

Twice the conversation ran out of context and continued from a compact summary. Each time the persistence engine re-fired the campaign loop and work resumed where the state file said it was — not where anyone’s memory did. The state outlived the context window.

By close: 31 subagents, 30 tasks completed, 149 commits, 153 files touched, 3,019 assistant turns, 1,405 tool calls — ending with a plugin version released to every installed user. The 98.8% cache hit rate is the difference between a $1,061 session and a $7,764 one.

  • Two context compactions, zero lost work — the state file carries the campaign
  • 149 commits and a shipped release in a single continuous run
  • $6,700 saved by caching on one session — the budget proof, not a benchmark

From the session record

One campaign invocation drove the control-panel detail workbench for 9h30m of active time across nearly 13 wall-clock hours. Twice the conversation ran out of context and was continued from a compact summary; each time, the persistence engine re-fired the campaign loop and work resumed where the state file said it was — not where anyone's memory did.

By close: 31 subagents spawned, 30 tasks completed, 149 commits, 153 files touched, 3,019 assistant turns, 1,405 tool calls — ending with a plugin version released and delivered to all installed users.

98.8% cache hit rate: $1,061 actual versus $7,764 uncached. The session that wouldn't fit in one context window also wouldn't have fit in one budget without it.

Session record, Composure plugin repo. Figures from the session's own stats.

This is what Pro delivers.

Not features for their own sake — measurable leverage on every session.

All results