What problem does it solve? Audit scans of agent sessions surface recurring anti-patterns (polling loops, retry clusters, token hotspots, compaction storms), but without a controlled process these findings stay as raw inbox entries or get fixed ad hoc. This Skill turns those anomalies into evidence-backed, reviewable change proposals so improvements are analyzed systematically and never applied without approval. ## Core Features & Use Cases - Systematic Problem Analysis: Quantifies baselines from the evolution-inbox JSONL file, aggregates patterns by frequency and severity, classifies root causes (missing constraint, unenforced constraint, unenforceable constraint), and requires web research on known solutions before proposing anything. - Draft-Review-Approve Pipeline: Drafts minimal patches with a cheap model, runs local static prechecks (validate_repo, quality_report, dsh-config-sync), gets frontier-model review, then hands off to a human for approval. - Auditable Status Flow: Tracks inbox entries through new → processing → applied/rejected states and produces both an analysis summary and a proposal document with verification commands and rollback plans. - Use Case: After evolution_scan.js reports 26 polling anti-pattern entries out of 390 inbox items, use this Skill to aggregate the evidence, research event-driven alternatives, and produce a proposal to add a hard long-polling constraint to the supervision skill. ## Quick Start Process the new anomaly entries in my evolution-inbox and produce a reviewable improvement proposal with quantified evidence and root-cause analysis.
