Experiment: 90c5493
rad experiment show 90c5493d7557a74ac79545556df038a89d124413
by hdhhdh@did:key:z6MkwQkGie444ZRwKmrNXSH8Y6dngFAJQ8zRGJRDReKBHwKR · Apr 16 19:27 2026
Tighten chaining rule, add unstage example — REGRESSED
Measurements
MetricBaselineThe benchmark measurement of the unmodified codeCandidateThe code with the proposed optimization appliedCandidate/Parent ΔThe performance change from baseline to candidate
avg_score primary0.943 (n=1)0.918 (n=1)-2.65% REGRESSED
failed_pct secondary2.100 (n=1)3.100 (n=1)+47.61%
partial_pct secondary7.200 (n=1)10.300 (n=1)+43.05%
perfect_pct secondary90.700 (n=1)86.600 (n=1)-4.52% REGRESSED
scored secondary97.000 (n=1)97.000 (n=1)0.00%
Annotations
hypothesisTighten chaining rule to only 'and', add unstage example
improvement0.9433 -> 0.9175 (-0.03) WORSE
next_action_hintRevert to the 0.9433 skill file. The exact wording of the chaining rule matters a lot. The version that said 'Only chain with && when user asks to do two things' worked better than 'explicitly asks with and'.
regressions#1 still over-chains, #68 stopped chaining (was fixed), #10 became no-tool-call, #67 git checkout instead of switch, #54 git fetch lost --remote
rollback_reasonSignificant regression from 0.9433
Base CommitThe starting commit before the optimization was appliedd1bd1c3f56300ec011d71b6697b91700a3c85f62
Candidate CommitThe code with the proposed optimization appliedno new code recorded — metadata only
Bench scriptba1951bb3138
Schemav5
Environment
Archx86_64
OSUbuntu 24.04
CPUAMD Ryzen 7 5700X3D 8-Core Processor
Agentpi-autoresearch / pi
Files