cli / commands / similarityhillclimb

hillclimb similarity

Live 3D similarity views of a search's candidates: `map` (pairwise, default) and `reference`.

hillclimb similarity COMMAND

hillclimb similarity map

Candidates embedded by pairwise distance: nearby means alike.

One 3D graph: lineage edges join parent to child, the best-so-far sequence is a gold trail, colour is score rank (cold to hot; the champion a gold diamond, the origin a white open diamond), shape the tree's fate. An experiment arm opens its whole run instead, coloured by arm like chart (--single for the one-search view). Keys: m cycles behavioral/structural/blend, v opens the reference view, space replays the search growing, s toggles the idle spin; hover reads a candidate's distances to the selected one, click dims everything outside its lineage. Drag rotates, scroll zooms, q quits.

hillclimb similarity map [OPTIONS] [SEARCH]

Arguments

argumentwhat it is
SEARCHlatest (default), <run-id>, or <run-id>/<search-id> optional

Options

optionwhat it does
--singleOne search only, even when it is an experiment arm.
--metric, -m TEXTPairwise distance to lay out by: behavioral, structural, blend. Default: behavioral.

hillclimb similarity reference

Candidates at their distance from one reference candidate.

Every candidate sits at (behavioral, structural, lineage) distance from a reference candidate — the origin the search grew from (its seed, else its baseline) by default, c toggles to the current champion — coloured by score rank (cold to hot; the champion is gold, the reference white). A search that is an experiment arm opens the whole run instead: every search of that problem in one cube, measured from the shared seed, coloured by arm like chart, n/p stepping through the run's problems (--single for the one-search view). v opens the map. Drag rotates, scroll zooms, q quits.

hillclimb similarity reference [OPTIONS] [SEARCH]

Arguments

argumentwhat it is
SEARCHlatest (default), <run-id>, or <run-id>/<search-id> optional

Options

optionwhat it does
--singleOne search only, even when it is an experiment arm.

hillclimb similarity scores

Pairwise similarity matrices between solutions, one per score.

Scores are pluggable: solution-card (an LLM writes a method card per solution, cards are embedded, cosine between them — needs OPENROUTER_API_KEY), api-calls (imports + library calls, no LLM), code-tokens (token overlap), or your own SimilarityScore subclass in a .py file. 1.0 = the same; representations are cached per file content in ~/.cache/hillclimb/similarity/, nothing is written into the run.

hillclimb similarity scores [OPTIONS] [SEARCH]

Arguments

argumentwhat it is
SEARCHlatest (default), <run-id>, or <run-id>/<search-id> optional

Options

optionwhat it does
--score, -s TEXTScore to compute (repeatable): a registry name, a .py file, or module:Class. Default: similarity.scores from config.
--candidates, -c TEXTComma-separated candidate ids (default: all).
--file, -f PATHCompare these solution files instead of a search (repeatable).
--explainPrint each solution's representation text (e.g. its solution card).
--jsonEmit the matrices as JSON.
--listList the registered scores and exit.