hillclimb

Use Claude Code and Codex to autonomously optimize a score you define.

You describe one problem as a single script: verifier.sh runs a candidate solution.py and writes back a number. hillclimb then spends a compute budget having headless coding agents write, debug and improve that solution.py — scoring every version through your verifier and keeping the one that scores best.

QuickstartRead the docsGitHubhillclimb.sh

three designs, one set of pages

01TerminalThe CLI on a page: a mono sidebar, hairline panels, justified prose./docs102InstrumentThe watch TUI as a docs site: section tabs, a status line, a scrubber for contents./docs203BlueprintA technical drawing: numbered sections, a measuring rail, crosshairs at every corner./docs304ConsoleA heads-up display: translucent panels over the page, a banner, contents that fold away./docs05BroadsheetReading mode: no sidebar, one wide column, a progress rail and big plates at the end./docs5

the climb

hillclimb chart
best score so far per search policy, candidate by candidate. greedy is a real hillclimb run on heilbronn-convex-13; the other two curves are mock data, for the shape of the comparison rather than a measurement

start here

01QuickstartInstall the CLI, get a bundled problem, start climbing.02What is hillclimb for?The problem shapes it suits, the ones it does not, and the rule of thumb.03WalkthroughThe Heilbronn triangle problem end to end, with the charts the run produced.04Defining problemsThe verifier contract, problem.yaml, holdouts, tunable parameters.05Search policiesgreedy, openevolve, the GEPA engine, and writing your own.06CLI referenceEvery command, and how runs and searches are laid out on disk.