rebase hillclimb

Start, watch, stop, and promote agentic model searches.

Synopsis

rebase hillclimb start <target> --budget 2h
rebase hillclimb list
rebase hillclimb status <run-id>
rebase hillclimb stop <run-id>
rebase hillclimb promote <run-id>

Subcommands

SubcommandDescription
startStart a search (hosted by default; --local runs it here).
listList hosted hillclimb search runs.
statusLive search state from synced GCS state.
stopGracefully stop a hosted search.
promoteFetch the selected model(s) into the workspace repo.

Requires the hillclimb extra: pip install "rebase-toolkit[hillclimb]".

start

Start a hillclimb search.

Synopsis

rebase hillclimb start <target> [--budget 2h] [--name NAME] [--model MODEL]
                       [--backend BACKEND] [--project PROJECT] [--local]

Arguments

ArgumentDescription
targetProblem target, e.g. emflow://gefcom2014:solar.

Options

OptionDescription
--budgetWall-clock budget, e.g. 2h, 30m (default 2h).
--nameSearch name.
--modelAgent model, e.g. sonnet.
--backendOperator backend: claude-code (default) or dummy (smoke tests).
--projectProject for the platform run (default hillclimb).
--localRun on this machine instead of the platform.

list

List hosted hillclimb search runs in the active workspace.

rebase hillclimb list [--limit N] [--json]

status

Live search state (candidates, best score, budget) read from the synced GCS state; also shows the platform run status.

rebase hillclimb status <run-id> [--bucket BUCKET]

stop

Queue a graceful stop: the search parks after the current operator finishes and can be resumed later. Delivered within one sync interval (~30 s).

rebase hillclimb stop <run-id> [--bucket BUCKET]

promote

Download every search's selected best/solution.py into the workspace repo (default models/<problem_id>.py). Review, commit, and open a PR; protected environments deploy through gitops.

rebase hillclimb promote <run-id> [--dest models] [--bucket BUCKET]

On this page