effortmining

Decompose a request into subtasks, classify each by difficulty (T1 mechanical through T4 hard reasoning, plus R-research and C-coding), and dispatch each to the cheapest calibrated reasoning-effort tier a blind grader still accepts. Use when you want a multi-part job done at the right effort per part instead of burning one uniform (usually too-high) effort on everything. The per-subtask effort is applied by dispatching to a tier-pinned worker agent (miner-low through miner-max), because Claude Code has no per-spawn effort parameter.

Drive the effortmining A/B benchmark harness, the deterministic Python core (bench/effort.py) that measures pass-rate and token cost per (task-class, effort-tier) and fits the calibration table the effortmine skill dispatches from. Use to validate the instrument, run the matrix, grade, analyze, report, or refit. This skill shells out to the harness; it does not compute results itself.