sdsrss/agentsmd/agentsmd-lesson-bypass-audit

agentsmd-lesson-bypass-audit

Use the selected SKILL.md absolute path from the live skills list; never infer it from the process cwd. Define and call the adjacent launcher in the same shell. The launcher preserves selected bundle → manifest-owned standalone → versioned CLI identity checks and exports plugin context only for a verified selected bundle.

SKILL_MD="<selected SKILL.md absolute path from the live skills list>"
agentsmd_skill_run() {
  node "$(dirname "$SKILL_MD")/scripts/agentsmd-run.js" "$SKILL_MD" "$@"
}

memory-prompt-hint.sh records a suggest event with the memory files it surfaced. A hint is a relevance candidate, not a requirement to read every suggested file; §7 requires consultation when the task actually matches the memory. This audit asks whether a non-user transcript row named a suggested file after the hint. A mention proves neither a successful read nor semantic relevance.

• applied — a suggested file is named by a non-user transcript row after the hint.
• bypassed — transcript found, no such reference.
• unmeasurable — no transcript for that session (surfaced separately so the % isn't overclaimed).

Run:

agentsmd_skill_run --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.