Three working, tested, adversarially-reviewed projects — a systems CLI, a
data-visualization artifact, and an offline product PWA — built from
tokens that would otherwise have expired, by autonomous moonlight sessions. 18 commits, every one reviewable.
Plate I
moonlight
systems CLI
Turns unspent Claude tokens into shipped side projects.
Streams ~1 GB of local agent transcripts — deduped, windowed, mtime-pruned — to measure real weekly token spend against a budget calibrated from your own busiest week.
Hands the surplus to headless agent sessions driven by per-project state files; one session built this tool’s own digest feature.
Atomic writes, corrupt-file refusal, honest exit codes — hardened by a 23-agent adversarial review.
🌙 moonlight — week of Jul 13 – Jul 19, 2026
spent ███████████░░░░░ 57.9M / 79.3M tokens (73%)
surplus 21.5M tokens ready to moonlight
by model (this week)
model output input cache w cache r
claude-fable-5 3.4M 194k 34.3M 869M
claude-opus-4-8 845k 53.8k 16.3M 485M
claude-sonnet-4-6 304k 7.8k 2.5M 26.2M