3. Operator policy

The operator policy (π_op) gets what the selector policy chose and names the operator to apply to it. The base class already does the plain mapping:

So a policy of your own only has to say where it differs:

# drafts_only.py
from hillclimb.sdk import Action, OperatorPolicy


class DraftsOnly(OperatorPolicy):
    """Never improves: whatever was chosen, draft (but repair a failing candidate first)."""

    def propose(self, state, selection):
        node = state.journal.candidates[selection.target_id] if selection is not None else None
        if node is not None and node.status in ("failing", "buggy"):
            return Action(operator="debug", target_id=node.candidate_id)
        return self.draft_action(state)

hillclimb climber check --climber drafts_only.py
hillclimb run circle-packing --climber drafts_only.py

propose(state, selection) returns one Action (an operator and the candidate it starts from), or None to wait for a result in flight. selection is None for a root step. state is everything the policy sees: the history (with holdout scores hidden), what is in flight, the budget left, and which direction is better. draft_action and expand_action(state, selection, operator=...) build the usual actions, and observe(state, candidate) is called for every result.

The built-in one

greedy is the plain mapping plus one decision of its own: before improving a chosen candidate that declared tunable parameters, it spends a few tune trials on it. Its settings go in params:

Three rules

• A policy only ever reads. It must not write to the history or to a candidate's folder.
• Every decision must follow from the history: compute it from state, or rebuild your caches in observe. That is what makes hillclimb resume work.
• A setting the block passes must be one the class declares in DEFAULTS, so a typo is an error before anything is spent.