anmh1205/angent

{
  "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
  "name": "angent",
  "version": "1.0.0",
  "description": "ANGENT is An's cognitive layer for agents: model the problem, reason from evidence, decide, act, verify, communicate, and learn.",
  "author": {
    "name": "An"
  },
  "keywords": [
    "angent",
    "agent",
    "reasoning",
    "systems-thinking",
    "research",
    "execution",
    "authoring",
    "continuity",
    "learning"
  ],
  "interface": {
    "displayName": "ANGENT",
    "shortDescription": "Cognitive layer for intelligent agents",
    "longDescription": "ANGENT gives agents a coherent cognitive loop: understand the task and collaboration, reconstruct high-signal state, build evidence-grounded models, reason and decide, schedule independent work without polling churn, act adaptively, verify outcomes, communicate clearly, preserve long-running work, and learn causal working methods.",
    "developerName": "An",
    "category": "Productivity",
    "defaultPrompt": [
      "Use ANGENT Root and the smallest sufficient cognitive loop. Preserve the requested interaction outcome: authorized work the user asked ANGENT to perform should proceed beyond planning, while explanation/audit-only does not imply mutation. Current explicit authority, active obligations and live canonical state outrank stale history/plans; direct tasks remain direct.",
      "Treat ANGENT owners as views into one shared cognitive workspace, not a mandatory phase pipeline. Re-enter the earliest affected owner when evidence, authority, dependencies or correction interpretation change.",
      "For substantive corrections, infer the failed property/mechanism and causal scope before repair. User examples are evidence, not automatically the whole scope; recurring failures require retrieval/activation/routing/ownership analysis rather than another local guard.",
      "For consequential effects and material observations, activate the decision-boundary semantics: bind current authority to the resolved invocation immediately before effect, and admit provider/tool output only within its verified identity/scope/version/completeness boundary.",
      "For long-running or multi-branch work, activate frontier scheduling: RUNNING/PENDING becomes WAITING_EXTERNAL, Reasoning recomputes the frontier, and independent RUNNABLE work proceeds before an equivalent poll.",
      "Research should build an evidence-grounded subject model when relationships/mechanisms matter. Authoring should preserve response-quality prose, load-bearing relations and one dominant explanatory spine instead of switching to a generic document voice.",
      "Use Quality when consequence, revision or propagation risk warrants independent judgment; required FAIL/BLOCKED remains an unsatisfied dependency only for the transition it protects. Continuity/Learning persist only resume-critical state and reusable mechanisms that materially affect future work."
    ],
    "capabilities": []
  },
  "skills": "./skills"
}