stevennitesh Hillclimb
Improve one measurable result through a bounded sequence of comparable experiments. Keep only changes that produce a material measured win without breaking accepted behavior.
The invocation authorizes repeated local source attempts inside the named target. It does not authorize staging, commits, push, live or external mutation, or unrelated cleanup. Preserve unrelated work. Apply repository instructions and the engineering contract.
1. Frame
Settle the accepted behavior, representative real workload, metric and unit, improvement direction, target, practical improvement threshold, attempt or time/cost ceiling, and nearest useful correctness checks. The target and campaign ceiling must come from the request or an authoritative repository contract; ask when either is missing. Infer only mechanical measurement details.
If the workload does not reproduce the property or no stable meaningful metric is available, return that measurement gap before mutating production code.
2. Baseline
Prefer an existing repository measurement path. Create the smallest disposable probe only when needed, and exercise the real caller or artifact whenever practical.
Hold material environment, configuration, workload, cache, warmup, and ordering conditions constant. Take enough observations to expose relevant variation; do not select one best run. Confirm that the ruler distinguishes a meaningful change, define the keep threshold before candidate results, record the baseline, then freeze the workload and comparison method.
If the workload or ruler must change after attempts begin, stop this campaign as incomparable. Continue only under a separately bounded campaign.
3. Climb
Trace the current owner, callers, and data path far enough to form one mechanism-specific hypothesis. Test one coherent hypothesis per attempt; it may require several coordinated edits, but keep it isolated from unrelated work. Inspect existing target changes and capture the invocation baseline. Do not try an edit that cannot later be removed without touching pre-existing work or an accepted attempt; if exact removal becomes uncertain, preserve state and stop.
Measure under the frozen conditions. Prove that baseline and candidate perform equivalent work inside identical measured boundaries; otherwise reject the attempt. If the improvement clears the practical threshold, run the nearest correctness check and keep it only when accepted behavior remains intact. Otherwise remove only that attempt. Code inspection or a plausible explanation is not a measured win.
Run decisive measurements serially so competing work does not distort the result.
4. Learn
Retain only enough attempt history to avoid cycling and choose the next
distinct supported mechanism. Keep it in the conversation, or under
.tmp/hillclimb/<run>/ only when resumption or comparison needs it; no formal
ledger or schema is required.
Do not stack unmeasured changes, weaken the target after seeing results, or add permanent harness machinery when an existing command or disposable probe is enough. Continue only while a worthwhile authorized hypothesis remains.
Stop when the target is met, the declared ceiling is reached, no worthwhile supported hypothesis remains, or measurement, proof, or authority fails.
5. Finish
Remeasure the integrated candidate against the original baseline under the same conditions. When observed variation could change the keep decision, confirm the result against a fresh baseline in alternating or counterbalanced order. Inspect the final diff and real artifact. Remove rejected attempt residue, debug scaffolding, and complexity that no accepted win needs; when two candidates perform equivalently, keep the simpler one.
Return the workload and metric, baseline and final result, delta, retained changes, correctness evidence, and either target attainment or the bounded stopping reason. Do not automatically stage, commit, publish, review, document, or start another workflow.
Complete when the final candidate has a trustworthy baseline-to-final comparison, relevant behavior passes, rejected attempts are absent, the final diff contains only accepted changes, and the target result or bounded stop is reported honestly.