He Learn

Hard Eng Learn

Contract

• Input = affected repository + explicit he learn evidence OR verified trigger from any route.
• Any route = Direct + Diagnose + Feature Loop + Build + Ship; active Feature Brief = not required.
• Output = <repo>/.agents/learning/<learning-id>.json + assigned prevention destination, or exact non-candidate.
• Lifecycle = unchanged; overlay only.
• Load workflow.md before classification or repair.

Ownership

• Failure classification + prevention owner + follow-through = this skill.
• Recurrence/root-cause proof = repeated-failure-learning.
• Prevention implementation = destination owner under its normal direct/he route.
• Product implementation = current stage owner; learning does not seize it.
• Repository learning state + runtime wiring = learning_state.py.

Invariants

• Trigger = verified repository recurrence OR engineering correction OR false passing check OR systemic protected-boundary gap OR repeated manual waste.
• One-off implementation finding = current build loop; learning candidate = forbidden.
• Record fact once in affected repository = failure + evidence + root cause + occurrences + prevention owner + required proof.
• Global ~/.agents = learning engine only; repository-specific record/prevention promotion = forbidden.
• Narrowest durable owner = remove cause → reuse/repair owner → invariant/type → regression test → scanner/hook → CI → script/tool.
• Mechanically detectable prevention = executable rule/tool/fixture at the closest owner; prose-only prevention = incomplete.
• Skill fallback = same root proven ≥2 by repeated-failure-learning + explicit deterministic limit + canonical <repo>/.agents/skills/<name>/.
• Skill discovery = Codex + Copilot read canonical .agents/skills directly; Claude uses setup-owned .claude/skills/<name> symlink.
• Prevention placement = before the expensive/failure boundary + cheaper than recurrence; independent checks parallel.
• Trigger record = learning_state.py start; spawn one depth-1 he-learn helper only when output says helper=he-learn.
• Repair = tracked + non-blocking; duplicate helper selection for one record = forbidden.
• Block/pause only when continued product work risks security/privacy/data loss/irreversible action or another protected boundary.
• Cross-repository repair = destination-owned normal flow; routine source PLAN pause/writer lease/nested lifecycle = forbidden.
• Open learning blocks task closure; resolved, non-candidate, or deferred with an owner + real next action permits closure.

Complete

• Candidate validity = evidence-backed.
• Prevention owner + proof = explicit.
• Applied prevention = violation fixture + valid fixture + actual-seam proof.
• Repository check = ~/.agents/setup.sh repo-check <repo> PASS.
• Deferred prevention = repository-owned next action without blocking unrelated delivery.