agent-interaction-audit

# Agent Interaction Audit Use this skill to audit recent human-agent collaboration for a repository across Codex and Claude Code traces, automation logs, and layered prompt/rule context. The goal is not a status summary. The goal is a private, evidence-backed interaction audit: what the user has been asking for, how agent work has actually progressed, where the collaboration is getting stuck, whether tool-call decisions were reasonable, which prompt or project rules are stale or misleading, which scripts and tool SOPs should be preserved, which repeated agent mistakes should become guardrails, which project facts are repeatedly rediscovered from scratch, and which few changes would improve the next iterations. If this skill is loaded, the answer must contain actual insights and recommendations, not just a timeline. A compliant report explains "so what": repeated request patterns, mismatch between user expectations and agent behavior, tool or Skill-selection decisions that created friction, and the highest-leverage changes for the next session. Use layered evidence and confidence levels. Strong insights require confirmed, non-low-confidence interaction evidence with exact or strong repository affinity. Candidate traces, ordinary repo state, low-confidence evidence, or sparse samples can support only weak leads, unknowns, or next-instrumentation recommendations. Do not fabricate patterns to fill the report shape. Default scope: last 30 days, current repository, Codex and C...