okr¶

Reproducible R source context for coding agents.

okr retrieves exact R package sources and arbitrary reference repositories, organizes them into a greppable source tree, and records enough provenance to verify every byte later.

Why¶

An installed R library is poor context for a coding agent. Package R code is stored in binary lazy-load databases, and compiled src/ trees are removed during installation. Metadata remains visible, but much of the implementation the agent needs to inspect does not.

okr creates a synthetic monorepo alongside the project. It can contain exact CRAN package versions, packages from Git forges, and non-package repositories that provide useful standards, protocols, or examples. A lockfile and aggregate tree digests make the result suitable for reproducible, offline evaluation as well as day-to-day agent work.

Start here¶

cargo install okr
okr init
okr add ggsci
okr sync

The result includes:

• okr.toml, the declarations you maintain;
• okr.lock, the resolved versions, commits, acquisition methods, and digests;
• deps-src/, the pruned source trees;
• deps-src/_manifest.md and _manifest.json, indexes for people, agents, and evaluation harnesses; and
• a managed marker block in AGENTS.md that tells coding agents where to look.

Continue with the quick start, then see source declarations for CRAN, GitHub, GitLab, Bitbucket, arbitrary Git hosts, direct tarballs, and reference repositories.

Less is more¶

okr installs nothing. It does not install R, mutate an R library, or solve package dependencies. Use rig for R toolchains and pak, renv, rv, or install.packages() for package installation. okr can inspect the installed library read-only and report whether its versions agree with the vendored sources.