/hillclimb

name: hillclimb
description: "Sustained single-metric improvement loop — hypothesis, before/after measurement against a fixed baseline, decision log, one accepted commit per verified win. For targets climbed over multiple rounds; not one-off fixes."
aliases: []
metadata:
  credits:
    hillclimb-loop: pstack
    author: Lauren Tan
    url: "https://github.com/cursor/plugins/tree/main/pstack"
    port: pstack-claude
    port-author: Michael Denyer
    port-url: "https://github.com/michael-denyer/pstack-claude"

Hillclimb

Climb one metric scientifically: each round is a hypothesis, a measurement against a fixed baseline, and a decision — keep the win, revert the loss. One accepted commit per verified win; the log is the climb's memory.

When to use

• a measurable target will take multiple rounds ("cut cold-start to <2s",

"get flaky suite under 1%")

• the user names a metric to improve, not a feature to build

When not to use

• a one-off fix with a known cause — `debug` or `execute`
• no measurable baseline exists yet — build the measurement first (that is

round zero, not skippable)

The loop

1. **Fix the baseline**: the metric's definition, the exact command that measures it, and the current value — measured, not estimated (`benchmark-checklist` discipline). The baseline command never changes mid-climb; if it must, the climb restarts. 2. **Log the target and the climb** in `.omc/state/decision-trail-{task-slug}.tsv` via the `show-me-your-work` trail: one row per round. 3. **One hypothesis per round**: what you believe will move the metric and why. State it before touching code. 4. **Implement on a branch**, measure with the baseline command, and apply `benchmark-checklist` before believing any number. 5. **Decide**: a verified win is committed alone (one commit per win, so any regression bisects to a single round); a loss or wash is reverted and logged with what was learned. Neither counts as failure — both are data. 6. **Repeat** until the target is reached or the hypothesis space is exhausted — then report the plateau honestly instead of grinding.

Rules

• Never stack two hypotheses in one round; a stacked win cannot be

attributed.

• Never move the baseline, relax the check, or cherry-pick the favorable

run — a climb that cheats its measurement is worse than no climb.

• Every decision row names its measurement evidence; every claim is

labeled measured, inferred, or guess (`principles` skill, Explain the number).

• The final reply reports: rounds run, wins kept, the measured delta from

the original baseline, and what was tried and rejected.