sdsrss/agentsmd/agentsmd-sparkline

agentsmd-sparkline

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" "$SKILL_MD" "$@"
}

agentsmd audit reports one window's hit counts; it cannot see a rule that WAS firing and silently STOPPED — a regressed or unwired hook keeps its healthy lifetime total, so the point-count still looks fine. This buckets each spec_section's enforcement hits (block / deny / advisory / bypass — the same hit definition audit uses) across N equal windows and shows the shape: ↗ rising, ↘ falling, ≈ flat, plus a went-silent flag = fired earlier in the window, zero in the most recent bucket.

Run:

agentsmd_skill_run --windows=6 --bucket-days=7

• went silent (sorted to the top of the report, ⚠ in the markdown table) is the actionable signal — a section that stopped emitting. Check the emitting hook is still registered in hooks.json and building; a regressed hook keeps its lifetime total, so only the trend catches it.
• --markdown emits a CHANGELOG-ready table (section · trend · sparkline · recent/older · total).
• --include-test folds back rows tagged AGENTSMD_TELEMETRY_TAG=test (excluded by default so verify/smoke runs don't fabricate a trend).

Advisory only — sparse windows are noisy; the trend needs a few sessions of data to mean anything. From the repo instead of an install: node scripts/sparkline.js. For a single window's point counts plus fail-open / project-class splits, use agentsmd audit.