sdsrss/agentsmd/agentsmd-scorecard

agentsmd-scorecard

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" "$SKILL_MD" "$@"
}

Generate the operator report or versioned JSON capture:

agentsmd_skill_run --days=30
agentsmd_skill_run --days=30 --json
agentsmd_skill_run --days=30 --compare=scorecard-previous.json
agentsmd_skill_run --days=30 --outcomes=/absolute/agentsmd-outcomes.json
agentsmd_skill_run --days=30 --conformance-candidate=/absolute/candidate.json
agentsmd_skill_run --days=30 --conformance-candidate=/absolute/candidate.json --conformance-binding=/absolute/binding.json

• Run from the project whose AGENTS.md prompt budget and worktree inventory should be measured.
• Treat test and qa data classes as visible provenance, not field evidence.
• Treat missing dimensions, stale captures, no-opportunity, insufficient opportunity, and unmeasured/partial/invalid false-block outcomes as gaps.
• Read compatibility.dimension_join_attribution before diagnosing missing dimensions. pre-first-observed-only is retained-window ordering debt, not proof of historical schema or a current emitter defect; post-first-observed-present warrants current SessionStart investigation, while no-dimension-reference cannot attribute a version or surface. Invalid/absent identities stay unjoinable and outside the missing-session count.
• Read false_blocks.state before interpreting its rate. The denominator contains only reviewed external true-block plus false-block outcomes; legacy, duplicate, self, test, QA, unknown, unreviewed, unmeasurable, mismatched, and future-dated evidence stays outside it.
• Use agentsmd outcomes list/review for explicit bounded human review. The scorecard is read-only, never fabricates labels from conformance near-negatives, and never rewrites raw telemetry or the review sidecar.
• Read automation.fail_open_causes as a complete category split whose total must equal fail_open_events; retain audit for exact reasons and version attribution.
• Read conformance.provenance before recommending a model run: packaged release evidence is historical, a source/input mismatch is not current-tree proof, and missing evidence calls for a bounded evidence source before an unconditional rerun.
• When an operator supplies an external evidence directory and a source checkout is available, use npm run captures:inventory -- --root=/absolute/evidence-dir --json there. This explicit bounded read cannot write/import evidence; use the existing candidate/binding flags only after inspecting the located records. Never scan home/config directories to guess an evidence location. Inventory hashes do not prove current applicability, and formal SLO captures do not replace the packaged reference baseline.
• Read conformance.provenance.evidence_phase when present: local-candidate matches a candidate artifact but is not publication proof; published-binding also binds the exact candidate to verified release/registry bytes and npm provenance. The command never fetches either file implicitly.
• Treat a published-binding as offline consistency evidence, not an independent signature audit; release provenance authenticity and input acquisition remain separate release-closure checks.
• Treat sampling calibration as a structural proxy and memory cite-recall as engagement only; neither proves semantic adherence.
• Use --compare only with a prior JSON capture from this command. The reader rejects symlinks, oversized input, malformed JSON, unknown fields, and unsupported schema versions.
• Read recommended actions as operator prompts. The command is read-only and never promotes/demotes rules, cleans worktrees, changes the spec, or runs external canaries.
• Preserve the machine-readable JSON directly; do not reconstruct it from the human report.