proposing-improvementslisted

# Proposing-Improvements — Logging harness improvement proposals to kaizen.md **Announce at the start (in pt-br):** "Usando proposing-improvements para detectar padrões sistêmicos e propor melhorias ao harness em kaizen.md." **Input:** the current transient `findings.md` (written by `recording-findings`, deleted at harvest end) plus the durable `kaizen.md` (cross-run memory). Since `findings.md` does not survive the run, `kaizen.md` is where cross-run precedent lives — a second occurrence of a signal is detected by matching against existing `kaizen.md` entries, not against an old `findings.md`. --- ## GOLDEN RULE — read this first **This skill NEVER edits agents or skills in `~/.claude/agents/` or `~/.claude/skills/`.** It only appends proposals to `kaizen.md`. The human reads `kaizen.md` periodically and decides what to apply. Auto-applying would cause unreviewed drift in the harness itself — that is the failure mode this rule prevents. --- ## Pipeline 1. **Read findings.md** — scan all entries for the current feature run. 2. **Detect systemic signals** — a finding is systemic if it matches one or more of: - Same **category** of issue appears in 2+ separate tasks (e.g., two adversary `edge-case` findings on different tasks). - Same `NEEDS_CONTEXT` key requested by executor across 2+ tasks (missing resolved_judgment type). - Compliance or adversary repeatedly flagging the same global-rule violation. - A structural friction point: the loop stalled, needed e