feat(lifecycle): close the self-improve loop end-to-end over the SWE Docker judge#544

Draft

drewstone wants to merge 45 commits into

mainfrom

feat/self-improve-e2e

Commits on Jul 9, 2026

feat(swe): judge-free patch emitter + code-surface meta-harness entrypoint

Drew Stone
and
claude
committed
feat(bench): honest-oracle structural-lever rig for HumanEval (best-of-k + self-repair on docstring examples)
drewstone
committed
feat(bench): CodeT-style generated-assert oracle (TESTGEN) for the structural-lever rig
drewstone
committed
feat(bench): DIVERSE mode — per-slot strategy prefixes to attack the all-k-samples-fail bucket
drewstone
committed
feat(bench): MBPP structural-lever rig — visible=shown assert (+TESTGEN), hidden=remaining asserts
drewstone
committed
docs(design): structural-rollout integration map — CheckSource is the only net-new seam; extend strategy family, verifier-environment, field routing
drewstone
committed
feat(runtime): structuralRollout strategy — k-sample + visible-check select + guarded repair (measured +8..+21pp)
drewstone
committed
chore(bench): track the prompt-evolution rigs behind the documented null result (hev-eval, hev-improve)
drewstone
committed
feat(improvement): rollout-policy surface — improve() tunes structuralRollout dials under the held-out gate
drewstone
committed
fix(runtime): carry the stopping content-only assistant reply in the tool-loop conversation
drewstone
committed
fix(runtime): fuse instruction+task into one user message for trajectory-less consults
drewstone
committed
feat(bench): structuralRollout runtime smoke — real model, real HumanEval, hidden grading script-side
drewstone
committed

Commits on Jul 10, 2026

Commits on Jul 12, 2026

Commits on Jul 13, 2026