Autonomous experiment loop with predict-attribute cycle and multi-dimensional evaluation
Run autonomous optimization experiments with predict-attribute cycles and multi-dimensional rubric evaluation.
.vibe/experiment.yaml must exist in the project root.vibe/experiment.yaml for domain definition (objective, scope, evaluator, constraints)..experiment/beliefs.md for current assumptions (if exists).vibe experiment start to create worktree, results.tsv, and beliefs.md.LOOP (until constraints.max_iterations reached OR no improvement for constraints.stale_threshold consecutive iterations):
scope.modifiable only. NEVER touch scope.readonly files.evaluator.type:
command: Run shell command, extract scores from stdout via extract_pattern.agent_judge: You are the judge. Re-read the modified files with fresh eyes. Score each rubric dimension 0-10 with justification. Be critical, not generous.behavioral: Run the target skill on fixtures, measure step count.constraints.must_pass_tests: true): log as crash. Attempt trivial fix. Otherwise discard and continue.vibe experiment results or direct append.keep (advance branch).discard (git reset).stale_threshold consecutive iterations, try:
END LOOP
vibe experiment apply to merge best result.scope.modifiable.scope.readonly files are sacred. Never touch them.constraints.must_pass_tests: true, tests MUST pass or iteration is discarded.