agent-card-fleet-summary

Fleet-analyze a directory of A2A AgentCard documents. Counts by autonomy / memory; surfaces the governance gaps that hurt an audit — autonomous agents missing incident-response URIs, persistent-memory agents with no refusal taxonomy, destructive tools on non-autonomous agents.

Status: v0.1.0 — Node 20/22 supported, library + CLI.

What it flags

Code	Severity	Rule
autonomous-without-incident-response-uri	🔴	autonomy_level=autonomous but safety_posture.incident_response_uri is missing (the spec requires the conjunction).
no-evaluations-on-autonomous	🔴	Autonomous agent with no evaluations[] entries.
destructive-tool-on-non-autonomous	🟠	Agent declares destructive tools without being marked autonomous — confirm gating.
persistent-memory-without-refusal-taxonomy	🟠	memory_persistence=persistent but no refusal categories declared.
empty-refusal-taxonomy	🟡	No refusal categories at all.
no-evaluations	🟡	No evaluations on a non-autonomous agent.
missing-homepage	ℹ️	agent.homepage not set.

CLI

npx agent-card-fleet-summary <cards-dir>
    [--format json|markdown|summary]
    [--now <iso>]
    [--fail-on-high]
    [--out FILE]

Library

import { summarize, toMarkdown } from "agent-card-fleet-summary";

const report = summarize(cards);
console.log(report.byAutonomy);      // { assistive, supervised, autonomous }
console.log(report.findings);
console.log(toMarkdown(report));

Composes with

• agent-cards-spec — the schema this reads.
• agent-card-diff — diff two cards across versions.
• agent-card-tool-coverage — verify a card's declared tools against an MCP server's tools/list.
• mcp-tool-card-summary — sibling for the MCP Tool Card side.

Develop

npm install
npm run lint && npm run typecheck && npm run coverage && npm run build
npm run demo

License

AGPL-3.0-or-later