Run bounded metric-improvement experiments with MiMoCode
Run MiMoCode’s opt-in experiment workflow against a fixed evaluation budget, then review kept changes, measured deltas and an independent audit.
npx skills add agentskillexchange/skills --skill run-bounded-metric-improvement-experiments-mimocode
Use this workflow when an operator has one mechanically measurable codebase objective, a fixed-budget evaluation command and an explicit editable-file scope. The agent records a baseline, tests small hypotheses and returns measured changes with an experiment report. Invoke it when ordinary coding conversation would make comparisons and stopping rules difficult to audit.
What this skill actually does
Start MiMoCode in a disposable checkout with MIMOCODE_EXPERIMENTAL_WORKFLOW_TOOL=true. Ask its agent to invoke the workflow tool with operation: run, name: research-experiment, an absolute workspace and an actual JSON args object. Required args are dir, goal, metric, evalCmd and editable; dir must equal workspace. Choose maxIters, lowerIsBetter, guardCmd and targetValue explicitly when applicable. The script keeps improvements only when its metric and guard gates pass, escalates failures and requests a separate metric-gaming audit.
Retain PLAN.md, artifacts/results.tsv, artifacts/AUDIT.md and REPORT.md. Review source changes and trace reported numbers to recorded evaluations. The scope boundary is one bounded experiment, including automatic trial commits and reverts inside the disposable checkout; exclude deployment, general research and evaluator modification. Custom Agents is appropriate because invocation depends on MiMoCode’s own runtime, even with different model providers. The cited development snapshot has an experimental opt-in gate; confirm the installed build exposes this same contract. An audit result does not guarantee correctness.