20× less scaffolding before you type a word

Plugin scaffolding used to inject ~1,300 lines of routing tables, duplicate skill lists, and hook context before the user said anything. On-demand loading and deduplication cut session-start overhead from ~200 lines to ~10, and resume from ~200 to ~3.

20×

Leaner at start

67×

Leaner on resume

60×

Faster resume hook

50%

Fewer skill entries

Every line injected at startup is a line that can’t reason about the actual problem. The fix loads an 8-line stack note instead of a 200-line architecture router, and pulls full docs only when blueprint or app-architecture is explicitly invoked. Lifecycle-aware hooks inject ~10 lines at startup, ~3 on resume, ~1 on compact — not 200 every time.

Deduplication did the rest: 80+ registered skills collapsed to ~40 unique once redundant marketplace plugins were disabled, and a doubly-registered MCP server dropped to one. A 60-second agent-based resume check became a sub-second bash command — 60× faster.

Load what you need, when you need it — full docs only on explicit invoke
Lifecycle-aware hooks: ~10 lines at start, ~3 on resume, ~1 on compact
Resume check 60s → <1s by replacing an agent hook with a command

Read the full technical breakdown — timeline, token counts, and cost math

This is what Pro delivers.

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

All results