EcoSIM Climate Readiness

Run a deterministic input check against the actual namelist and launch directory. Do not infer complete coverage from filenames, year coordinates, or successful NetCDF opening. This skill checks climate inputs; a pass is not whole-case readiness or evidence of scientific correctness.

Run the checker

Requires Python 3, NumPy, and netCDF4. From the ecosim-agent root:

python3 .agents/skills/ecosim-climate-readiness/scripts/check_climate_readiness.py \
  /path/to/run/name.namelist

• Relative input paths and rpointer.esim are resolved from the namelist's directory. If launch differs, pass --run-cwd /actual/launch/directory.
• Default scope is the remaining simulation: start_date, forcing triplets and cycle counts, stop settings, and the continuation checkpoint date determine what must be available. A branch's finidat date must match start_date.
• Use --all-periods for a full forcing-schedule audit, including completed spinup. It intentionally ignores restart/start-date truncation and stop_n.
• Reports default to a new timestamped directory under result/climate_readiness/<case>_<UTC timestamp>/. --output-dir PATH selects a new destination. Existing report directories are never overwritten.
• Exit 0 means the implemented checks passed (possibly with warnings), 1 means invalid inputs/configuration, and 2 means CLI/report-writing failure. A shell workflow can use the exit status to gate its authorized run. The checker itself never launches EcoSIM or edits inputs.

Read report.md and report.json. Summarize the checked window, selected files, status, and failures with their variable, forcing date, one-based hour index, and affected-value count. Preserve the distinction between forcing year and model year, particularly during spinup. Report an invalid restart/stop window as a configuration issue rather than silently examining a different period.

What is checked

The helper follows the standalone reader contract documented in reader-contract.md. It covers ordinary scalar/list Fortran namelists, repetition counts, reverse forcing periods, zero cycles, Gregorian leap years, hourly stop boundaries, continuation pointers, and the daily-to-hourly transition. It deduplicates repeated forcing records and reads one annual variable slab at a time.

It checks the five required hourly meteorological fields, the corresponding six daily fields when selected, required annual roughness/solar-noon/rain-chemistry fields, and optional hourly pressure, longwave radiation, and dry deposition when present. Missing values, NaN/Inf, sentinels and physically implausible values are failures. A missing December 31 in a non-leap year is a failure; unused day-366 padding is not. Missing/ambiguous unit metadata is reported; explicitly incompatible units and unsupported packed data fail.

For selected hourly forcing years, it also recomputes annual/monthly/seasonal precipitation budgets using the shared Tools/precipitation_quality.py audit. Sourced site MAP stored by the forcing converter supports broad annual and multiyear plausibility screens. Missing climatology or measured precipitation validation produces explicit warnings; a valid hourly range is not a credible annual water budget. Partial years are not annualized. Stored tower comparison status is invalidated when the rainfall identity differs. These findings reach whole-case preflight without relying on a stale quality JSON. See precipitation-budget review for reference provenance, thresholds and a read-only audit of older files.

Also surface misspelled restart options and year-based stops capped by the forcing schedule. These findings do not imply that weather data caused every reported model failure.

Limits and next steps

• Inspect the selected model's reader if it differs from the documented standalone contract. Do not treat unsupported ASCII, coupled/prescribed forcing modes, non-Gregorian calendars, or unsupported namelist syntax as verified. Unsupported inputs fail explicitly; investigate them separately.
• Year/day/hour arrays do not establish timezone, interval labeling, solar-noon alignment, or registration against observations. The report retains this uncertainty even when numeric checks pass. Other readiness workflows must assess provenance and the other case inputs (including GHG forcing and climate-adjustment factors).
• The numeric bounds are broad error screens in reader input units, not ecological suitability criteria or universal calibration ranges. Review any flagged extremes against the source before changing data or thresholds.
• Recommend source-backed repairs or a shorter supported run window. Do not silently interpolate, repeat a neighboring day, download replacements, or relax checks. After an authorized repair, rerun against the final file.

Validation

python3 -m unittest discover -s .agents/skills/ecosim-climate-readiness/tests -v

The tests use generated NetCDF fixtures. They cover the SOAP missing-last-day failure, leap padding, NaN/Inf/sentinels, units/ranges, optional fields, mixed daily/hourly selection, forcing cycles, run/restart windows, relative paths, CLI exit codes, and preservation of input files.