Repository navigation
R regressions can now keep their current specification and recoverable history up to date as part of estimation.
What changed
- Group current results by outcome and result family, so main estimates and supporting specifications can remain current together.
- Fit and register with
run_fixest_and_register(): retain the specification, estimation data, model, and source code before advancing the matching current result. - Keep bookkeeping in two files: a compact current registry and a consolidated archive. Identical current reruns add no history; requested reversions can restore an earlier result.
- Preserve current results when estimation or saving fails. Experimental alternatives enter the archive without replacing current results.
- Document the workflow in the README and reduce the always-on instruction card.
- Align the Claude plugin, Codex plugin, and marketplace versions at 0.33.0.
Automatic registration covers the R helper workflow; direct estimator calls that bypass the helper are not intercepted. The helper requires fixest, yaml, digest, and jsonlite.
Validation
Regression lifecycle, historical recovery, failed-write handling, plugin manifest, installer, and stop-gate checks passed. Offline trigger evaluation introduced no regressions against its existing baseline.
The synthetic DiD prompt pilot is included as illustrative development evidence, not a general model benchmark. This release also includes the source changes documented for versions 0.26–0.32 since the previous published GitHub release.