Repository navigation
projdump v2.0.0 — the fast one
The fast one
⚡ Performance — the headline
- Batch scanner: 7–25× faster dumps on large repos. v1 forked ~6 helper processes per file (~40 s for a 2000-file repository); v2 scans everything in ONE python3 process over a NUL-delimited stream — 2000-file repo: md 40.6 s → 5.5 s, xml → 2.4 s, json → 1.6 s; the scanner alone takes 0.07 s
- Pure-bash tree renderer (
tree(1)no longer a dependency) +--tree-depth N - Zero subshells per file: pure-bash
human_size, cached language tags
✨ Added
.projdumprcconfig files —--config FILE,--no-config--split N[k|m]— chunked output for long chats (_dump.part1.md,part2.md, …)--model NAME/--list-models— token presets for 8 popular models- Weighted token estimates (minified · data · code · prose)
Full Changelog: v1.1.0...v2.0.0
Docs: https://cheginisoroush.github.io/projdump/
Install: curl -fsSL https://raw.githubusercontent.com/CheginiSoroush/projdump/main/install.sh | bash