okr

Reproducible R source context for coding agents.

Installed R packages are poor context for coding agents. R code is packed into binary lazy load databases (.rdb/.rdx), while compiled package src/ trees are stripped at installation time. okr vendors exact source trees into a greppable synthetic monorepo with records hashes.

Install

cargo install okr

Quickstart

okr init
okr add ggsci
okr sync
# Point your coding agent at deps-src/

okr init selects and writes the latest available exact dated CRAN snapshot, so a bare CRAN package can be added immediately. It checks from the current UTC date backward and keeps the successful package index in the download cache.

okr sync writes okr.lock, deps-src/_manifest.json, and deps-src/_manifest.md. Repeating it with an intact tree is a no-op.

Evaluation workflow

Build the source context online, then use the integrity gate in an evaluation sandbox:

okr sync
okr verify --strict --json

verify always treats vendor-tree drift as an exit-4 failure. --strict also requires the read-only installed-library coherence check when R is available. Use okr sync --offline to rebuild entirely from the local content-addressed cache. The compact lock records an acquisition-artifact hash and one aggregate digest for each post-pruning source tree; verification recomputes the latter from every vendored file.

GitHub authentication

okr uses the gh CLI when present and authenticated; otherwise it uses GITHUB_TOKEN; otherwise it uses anonymous GitHub API access with low rate limits. Private GitHub repositories use the same tiers before falling back to git. GitHub Enterprise and self-hosted GitLab, Codeberg, SSH, and other git:: remotes use your existing git credential helpers and SSH configuration; okr never stores credentials.

Scope

okr installs nothing: rig, pak, renv, and rv do that.

License

MIT