jcode: a Rust harness that minimizes per-agent RAM and fast-starts headless coding agents

What is actually new: jcode is a Rust-written agent harness that targets high-density multi-agent workflows by sharing a single daemon for sessions and keeping per-session RAM very small. The project’s README includes reproducible scripts and measured comparisons that show substantially lower resident set sizes (PSS) and much faster time-to-first-frame/input than several competing CLI agents tested on Linux. Key evidence from the README Benchmarks were measured on Linux (noted in the repo) with jcode v0.89.19-dev (default build, local embeddings not compiled in) and Claude Code 2.1.267. Representative results the project reports: Headless sessions (PSS): 1 concurrent session — jcode 32.6 MB vs Claude Code 261.0 MB (about 8× less); 10 sessions — jcode 66.7 MB vs Claude Code 1749.7 MB (≈26× less). The project reports an incremental cost of ~3.1 MB per additional jcode session versus ~164 MB per additional Claude Code session (~54× difference). Interactive startup: time-to-first-frame ~14.0 ms and time-to-first-input ~48.7 ms for jcode; the README shows substantially larger values for comparison tools (hundreds to thousands of ms). The repo includes a script (python3 scripts/bench_headless_memory.py) and a stated measurement date to aid reproduction. What the README does not show: the published numbers focus on PSS (a Linux process metric) and interactive boot