hill-climb

Council members propose improvements and judge versions without knowing which is newer. Each round makes one candidate, checks it, and keeps it only if it wins under the agreed decision rule. Failed or losing edits are restored from a saved copy.

Use it after a change works and you want tighter copy, spacing, or code. You can also start from an open-ended goal across named paths, such as improving the settings pages. Check correctness first, for example with sol-loop; judge preferences do not prove correctness or quality.

Install hill-climb and council together: replace -s hill-climb in the command above with -s hill-climb council. A missing council skill stops the run.

Invoke it explicitly with /hill-climb in Claude or $hill-climb in Codex. Name the goal and paths, for example /hill-climb tighter spacing on the settings page. You can also specify a line limit, such as /hill-climb up to 60 lines, tighten the copy, or cap member calls.

For guidance such as a skill, the evidence is work produced under each version, compared with a no-guidance baseline. A shorter version can win when that work is no worse. Holdout tasks, which no proposer or round judge sees, show whether the gain carries beyond the tasks the climb learned from.

A fixed climb card records the goal, three to five criteria, scope, checks, decision rule, behavior, and budget. For an open-ended goal, or unclear scope, criteria, decision rule, or budget, the agent asks up to four questions with suggested answers. It freezes the card after your answers. Baseline checks run before edits; a failure stops the climb.

Choose judges, a deterministic score, or both. A score comes from a fixed command that prints one number. A guidance climb uses an answer-key tally instead. With both, an edit must improve the score and win a majority of judges. Judges see code diffs, UI screenshots, or text in reading order, as appropriate. Judges are told to ignore any claim inside the work about which version is better.

Before a score counts, the agent checks that it stays steady across runs, that a deliberate regression makes it worse, and that a deliberate cheat is caught. A noisy score, such as a build time, uses the median of several runs. The files that grade the climb, such as the score script, its fixtures, and the gate tests, are locked: if one changes, the climb stops.

In round one, each member proposes up to three improvements. The agent applies the best-supported ideas as a small edit and runs the checks. If only a false claim in text fails, it gets one repair and recheck. Another failure or a loss restores the saved version. If outside work changed the same files, the agent stops without restoring over it. It never uses Git stash or git clean.

The default is three rounds, with ten at most. The default size limit is the larger of 30% of the original change's lines and 20 lines. Behavior stays the same except for interactions you explicitly allow on the card. A ten-round climb with the four default members can use up to 48 council calls, plus retries. A guidance climb adds calls for its readers, baseline, and holdout tasks.

After accepted edits, the final comparison checks the starting and final versions with judges, scores, or both. If judges prefer the start, the report says so first and recommends dropping the climb. A score that fails to beat the start is also reported first. Blind comparisons reduce judge noise; they do not remove it.

SKILL.md has the exact budgets, restoration rules, and guidance-evaluation procedure.

• The final working-tree version. Commits require prior authorization.
• The climb card, round edits, check results, votes and reasons or scores, and the start-to-final comparison.
• For UI and guidance, a 0–10 before-and-after scorecard per screen or task, with its biggest remaining issue. UI evidence includes paired screenshots.
• Diff statistics, skipped follow-ups, and lessons the original work missed.
• A run record under ~/.hill-climb/runs/<id>/ and a report built from it. If you keep a runs page and the host can publish it, the agent offers to republish it.