{
  "schema": "agent-buildprint/publication.v1",
  "publish": true,
  "fileExcludes": [],
  "slug": "complete-agent-skills-evaluation-os",
  "title": "Complete Agent Skills Evaluation OS",
  "creator": "Agent Buildprint",
  "category": "Workflow OS",
  "tier": "agent-grade",
  "status": "validated",
  "runtime": [
    "Any coding agent",
    "JavaScript proof",
    "CI-ready adapters"
  ],
  "stack": [
    "Static lint",
    "Loadout inventory",
    "Skill tests",
    "Activation evals",
    "Transcript checks",
    "Multi-agent safety"
  ],
  "iconKeys": [
    "json",
    "md",
    "typescript"
  ],
  "difficulty": "Advanced",
  "featured": true,
  "summary": "Complete evaluation stack for agent skill setups: validate config, measure loadout cost, test skills, check activation, enforce transcript process, and score multi-agent safety.",
  "plainDescription": "A blueprint for evaluating an entire coding-agent setup, not just one skill: installed skills, agents, commands, hooks, MCP, routers, subagents, workflow discipline, token cost, safety, and CI evidence.",
  "promise": "A validated Buildprint for evaluating complete agent+skills installations from static validity through real behavior, with skill-eval-runner as the core module but not the whole system.",
  "whatYouGet": [
    "Setup snapshot and install parity model",
    "Static lint gates for agent config files",
    "Loadout and token-cost inventory",
    "Skill unit/regression test harness pattern",
    "Activation and routing eval pattern",
    "Transcript/process invariant checks",
    "E2E task benchmark model",
    "Multi-agent/subagent safety gates",
    "Weighted scorecard and CI report contract"
  ],
  "whatYouNeed": [
    "A target agent setup to evaluate",
    "Offline fixture cases for deterministic mode",
    "Optional live agent/provider credentials for live adapters",
    "A safety policy for external/destructive actions",
    "A list of critical workflow invariants"
  ],
  "architectureFlow": [
    "Snapshot",
    "Lint",
    "Inventory",
    "Skill tests",
    "Activation",
    "Transcript",
    "Scorecard"
  ],
  "includes": [
    "Deep tool comparison",
    "Static lint layer",
    "Loadout inventory layer",
    "Skill unit eval layer",
    "Activation eval layer",
    "Transcript process eval layer",
    "E2E task bench",
    "Multi-agent safety layer",
    "Safety policy",
    "Weighted scorecard",
    "Offline JavaScript proof"
  ],
  "risks": [
    "Mistaking per-skill tests for full setup proof",
    "Ignoring activation failures",
    "Skipping transcript/order evidence",
    "Measuring a drifted install",
    "Token bloat from unused loadout",
    "Unsafe external actions in live tests",
    "Subagent file ownership collisions"
  ],
  "checks": [
    "Static-invalid setups fail before expensive behavior tests",
    "Loadout inventory exposes loaded token tax and dormant artifacts",
    "Skill unit tests are separated from activation tests",
    "Activation evals include positive and negative prompts",
    "Transcript checks enforce skill-before-action and approval-before-risky-action invariants",
    "Multi-agent cases check parent context, output schema, and file ownership",
    "Safety hard-fails override good final outputs",
    "Offline proof passes 8/8 tests without live agents, providers, or network calls"
  ],
  "trustBadges": [
    {
      "label": "Deep stack design",
      "detail": "Combines skill-eval-runner, cc-plugin-eval, Superpowers-style transcripts, agnix, loadout inventory, and multi-agent safety layers.",
      "tone": "success"
    },
    {
      "label": "Offline proof passed",
      "detail": "Deterministic JavaScript proof validates snapshot, lint, loadout, activation, transcript, and multi-agent scoring.",
      "tone": "success"
    },
    {
      "label": "Not just skill tests",
      "detail": "Skill unit tests are one module; activation, process compliance, safety, and loadout cost are separate gates.",
      "tone": "info"
    },
    {
      "label": "Safety hard-fails",
      "detail": "Secrets, external writes, destructive actions, and fabricated evidence override aggregate scores.",
      "tone": "warning"
    }
  ],
  "originGithubUrl": "https://github.com/balyakin/skill-eval-runner",
  "originLabel": "balyakin/skill-eval-runner + eval stack",
  "copyPrompt": "Use the Complete Agent Skills Evaluation OS Buildprint. First bootstrap exact snapshots: agb start https://agent-buildprint.com/buildprints/complete-agent-skills-evaluation-os/package.json . If agb is not installed, clone https://github.com/DomEscobar/agent-buildprint and run node agent-buildprint/bin/agb.js start https://agent-buildprint.com/buildprints/complete-agent-skills-evaluation-os/package.json . Then read .buildprint/next-agent.md and continue. Do not write Buildprint snapshots manually. Build the evaluation stack from the packet; use offline fixtures by default and ask before live external actions."
}