hillclimb chart
Live hillclimb chart: best score so far vs tested candidates.
Live hillclimb chart: best score so far vs tested candidates.
One chart per problem. With no argument and more than one problem (or study) in the folder, first a table of them, newest activity first; enter opens that chart, esc comes back. hillclimb watch reaches the same chart with c.
The climb: one staircase across every run of the problem, every scored candidate a dot (bright where it set a new best, dim where it missed). With several runs, v cycles three views: the plain climb (the default), the same climb with each run's dots in its own colour, named in the legend (#2 15:08: number and start time, or the name you gave the run), and a comparison — one line per run from its own first candidate, ending in a dot at its best. ] / [ focus one run (the others fade to grey) and step back to all; in the plain climb they open the run colours. A study gets one line per experiment. Problem-config baselines are dashed lines. Refreshes as candidates land. Keys: v=climb/runs/compare, ]/[=focus next/previous run, r=refresh, t=toggle improvement text, d=detail (the focused run's search: every scored candidate as a mark, parent edges, accepted lineage bold), h=toggle holdout/validation, c=cost overlay, p=switch problem, esc=back to the table, q=quit.
hillclimb chart [OPTIONS] [SEARCH]Arguments
| argument | what it is |
|---|---|
SEARCH | latest (default), <run-id>, or <run-id>/<search-id> optional |
Options
| option | what it does |
|---|---|
--detail, -d | One search only, with its exploration tree drawn on the curve. |
--holdout | Force the holdout view. Default: a holdout-scored problem opens on holdout (the split its reference baselines live on), others on validation; h toggles either way. |
--cost | Overlay cumulative coding agent tokens and verifier CPU-minutes on right-hand axes — what the climb cost as it climbed. |