NEWS.md
In CodeCarbonR: Track Energy Consumption and Carbon Emissions of R Code

CodeCarbonR 0.1.1

Corrected author metadata: fixed the spelling of Rachel Peterson's surname and the capitalization of Dylan Van Bramer's, and added Rachel's ORCID iD.
citation("CodeCarbonR") now also lists the accompanying preprint (Bock, Van Bramer & Peterson, 2026, ).
No functional changes to the package.

CodeCarbonR 0.1.0

setup_carbon_tracker() installs codecarbon into a dedicated r-codecarbon conda environment, with confirmation.
carbon_tracker() wraps a codecarbon OfflineEmissionsTracker with $start()/$stop()/$flush() methods.
with_emissions_tracked() measures a single block of code in one call.
list_carbon_tracker_countries() lists supported country_iso_code values.
carbon_tracker_ready() checks whether codecarbon is installed and importable.

carbon_tracker()/with_emissions_tracked() now create output_dir if it doesn't already exist, instead of crashing at $stop() after the tracked code has already run.

setup_carbon_tracker() installs codecarbon >= 2.2.2 -- a floor, not an exact pin (see ?setup_carbon_tracker for why). Versions actually validated against so far:

2.2.2 -- full local validation, Windows (comparison cases 01-06 rendered end to end; tests/testthat/test-tracker.R's tracker-restart regression test confirmed against this version directly).
3.3.0 -- validated via GitHub Actions CI (Windows/Mac/Linux, R CMD check + testthat), and directly against the tracker-restart behavior in an isolated environment. Tracker-restart behavior differs from 2.2.2 (see Known limitations below) but CodeCarbonR's own tests pass against both.

If you validate against a different version (e.g. a Monsoon HPC run), add it to this list rather than replacing what's here -- the point is an honest record of what's actually been checked, not just the most recent.

Restarting a single carbon_tracker() instance ($start() after a prior $stop()) does not produce an independent measurement for the new phase -- codecarbon's underlying tracker never resets its internal clock on stop(). What that produces downstream is codecarbon-version- dependent (frozen stale reading on 2.2.2, inflated cumulative reading on 3.3.0 -- see the "codecarbon version compatibility" section above), but it's never a clean isolated measurement either way. Use a separate carbon_tracker() per phase instead. See ?carbon_tracker.
GPU tracking (comparison/07_gpu_neural_net) is implemented but not yet validated on real GPU hardware.
macOS/Apple Silicon and Linux: the package itself installs and passes R CMD check/testthat on both via GitHub Actions CI, but the comparison suite (cases 01-06) has only been rendered end to end on Windows so far. CI passing is a package-level signal, not evidence that the comparison suite's accuracy validation holds on those platforms -- don't cite this as cross-platform accuracy validation without an actual comparison-suite render on Mac/Linux.