hillclimb

Invoke the skill with the metric, its direction, and a stop predicate pairing a target with an attempt floor. It freezes a benchmark harness, measures a baseline, then builds one hypothesis per attempt in its own worktree, keeping only changes whose median beats the spread with checks green.

Hillclimbing iteratively makes one measured change at a time, keeping improvements and reverting regressions until a target is reached. This skill formalizes it with a frozen harness, median-of-five measurements, per-attempt commits, and a decision log recording every verdict.

Do not use it for a single already-decided change, which a normal implementation step handles, or for a failure with an unproven cause, which debugging must diagnose first. It also does not choose what to improve; candidate ranking happens beforehand.

Freezing prevents the run from editing its own yardstick, so every attempt is measured against the same standard. The harness must separate two known-different workloads beyond its spread before use, and any harness flaw ends the run with a fresh baseline next time.

Each attempt is one change built in its own git worktree and measured one at a time to avoid machine contention. It is kept only when the median moves past the spread in the right direction with regression checks green; everything else is fully reverted and logged.

Three consecutive rejects count as a plateau, not a stop. The run logs a pivot row and changes category, combines near-misses, or rereads the hot path, stopping only when the target predicate holds or the named budget is spent.