zerd in numbers
Size, test count, built-in surface and before-and-after measurements, taken from this repo.
Measured from this repo unless it says otherwise.
| Codebase | about 25k lines of Rust, plus 8k lines of tests |
| Tests | about 500, all offline (a scripted model, no network), around 20 seconds |
| Surface | 21 built-in tools, 19 skills, 7 model roles, 10 hook points, 7 approval modes |
| System prompt | about 3k tokens for a small-window model, 4.1k with the skill list |
| History | 480+ commits, one person |
Before and after, each taken from the commit that fixed it (measured on the Python version; the Rust port keeps the same logic):
| What | Before | After |
|---|---|---|
| Prompt token budget per task | 1.5M | 300k |
| Context replay without clearing old results | 3.4x the input tokens | baseline (clearing turned back on) |
| System prompt | ~1,990 tokens | ~1,140, then held under a cap |
| Sub-agent in a review spawning its own challenger | 80k tokens, 4 minutes | none, 150s limit per sub-agent |
| Search index of a large repo (24k files) | 13.7s cold | 1.2s warm |
Broad grep on the same repo | 5.1s | 1.1s |
| Model calls that only updated a todo list | 31 of 113 in the benchmark | rides along with the next real call |