菜单
Measure whether surfaced memory hints were later cited in Codex transcripts. Use for memory hint follow-through and bypass rates. Not for rule-hit governance or memory editing; read-only.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Diagnose an agentsmd installation (安装故障诊断): prerequisites, hook registration/executability, config flags, and spec drift. Use after install/update or when hooks fail. Not for telemetry analysis; read-only.
Show agentsmd install state (安装状态清单), registered hooks, preserved tenants, config flags, and telemetry rows. Use for a quick inventory. Not for health diagnosis or rule analysis.
Extract CSS :root and Tailwind v4 @theme tokens into DESIGN.md plus an AGENTS.md pointer. Use for frontend design-system facts. Not for coding conventions, non-frontend repos, or Tailwind v3 config themes.
Distill coding conventions from source into AGENTS.md (提炼代码约定). Use for naming/import/error-handling analysis after agentsmd-init. Not for stack detection or design tokens.
Aggregate agentsmd rule-hit telemetry by section (遥测命中统计). Use for enforcement activity and raw hit counts. Not for promotion/demotion decisions or transcript violation rates; read-only.
Generate or refresh project AGENTS.md stack facts, commands, and structure (生成项目指令). Use before convention analysis. Not for inferring coding conventions or design tokens.
| name | agentsmd-lesson-bypass-audit |
| description | Measure whether surfaced memory hints were later cited in Codex transcripts. Use for memory hint follow-through and bypass rates. Not for rule-hit governance or memory editing; read-only. |
Resolve the script root first. Set SKILL_MD to the selected SKILL.md absolute path from the live skills list; never infer it from the process cwd.
SKILL_MD="<selected SKILL.md absolute path from the live skills list>"
CANDIDATE_ROOT="$(cd "$(dirname "$SKILL_MD")/../.." && pwd)"
if [ -f "$CANDIDATE_ROOT/scripts/lesson-bypass-audit.js" ]; then AGENTSMD_ROOT="$CANDIDATE_ROOT"; else AGENTSMD_ROOT="${CODEX_HOME:-$HOME/.codex}/agentsmd"; fi
memory-prompt-hint.sh records a suggest event with the memory files it surfaced. That the hint fired says nothing about whether the agent read the file — §7 "read the suggested memory" is HARD but leaves no direct telemetry. This joins each suggest row to its session transcript and asks: after the hint, did a non-user row name a suggested file?
Run:
node "$AGENTSMD_ROOT/scripts/lesson-bypass-audit.js" --days=30
Report cite-recall = applied / (applied + bypassed) plus the unmeasurable slice. A low recall means the hint is firing but being bypassed — uncited memory decays, so this is the signal that the memory layer needs tuning (better index keywords, fewer/stronger hints). Advisory only. From the repo instead of an install: node scripts/lesson-bypass-audit.js.