Cost guard

{
  "id": "dev3d.cost-guard",
  "name": "Cost guard",
  "author": "dev3d",
  "license": "MIT",
  "version": "1.1.0",
  "settings": [
    {
      "key": "aggressiveness",
      "type": "select",
      "label": "How hard to push down-tier",
      "default": "balanced",
      "options": [
        "gentle",
        "balanced",
        "aggressive"
      ],
      "description": "Balanced honours the role's policy and only overrides the cheapest task classes. Aggressive pulls everything except coding and review down a tier."
    },
    {
      "key": "localOnly",
      "type": "boolean",
      "label": "Never prefer a hosted provider",
      "default": false,
      "description": "When on, the cheap task classes are pinned to the local provider even if a hosted model is cheaper."
    },
    {
      "key": "maxPromptTokens",
      "max": 200000,
      "min": 1000,
      "type": "number",
      "label": "Warn above this prompt size",
      "default": 12000,
      "description": "Only used for the plugin's own logging; it does not truncate anything."
    }
  ],
  "apiVersion": "1",
  "contributes": {
    "models": [
      {
        "id": "local/qwen2.5-coder-7b-instruct",
        "tier": "small",
        "label": "Qwen 2.5 Coder 7B (local)",
        "strengths": [
          "summarize",
          "intake",
          "routing"
        ],
        "providerId": "local",
        "capabilities": {
          "tools": true,
          "vision": false,
          "reasoning": false,
          "streaming": true
        },
        "contextWindow": 32768,
        "costPerMTokIn": 0,
        "costPerMTokOut": 0,
        "maxOutputTokens": 4096
      }
    ],
    "skills": [
      {
        "id": "cost-aware-delegation",
        "body": "# Cost-aware delegation\n\nBefore you reach for a strong model, answer three questions in writing:\n\n1. **What does this task actually require?** Parsing, summarising and re-formatting need none of the reasoning a frontier model is priced for.\n2. **What is the cost of being wrong?** A wrong intake summary costs a re-run. A wrong migration costs a day.\n3. **Is the expensive model buying anything?** If you cannot name the specific step that needs the extra capability, you do not need it.\n\n## Say this\n\n> This is a summarise-and-reformat task. A small model is sufficient; I would spend the budget on the review stage instead.\n\n## Not this\n\n> I will use the most capable model available to be safe.\n\nSafety is not a reason to escalate. Specific, named difficulty is.",
        "name": "Cost-aware delegation",
        "tags": [
          "cost",
          "process",
          "routing"
        ],
        "description": "Decide whether a task needs a frontier model at all, and say which tier is actually justified.",
        "taskClasses": [
          "routing",
          "summarize",
          "planning"
        ]
      }
    ],
    "pipelines": [
      {
        "id": "spend-review",
        "name": "Spend review",
        "stages": [
          {
            "kind": "review",
            "mode": "single",
            "name": "Audit",
            "roleIds": [
              "cost-auditor"
            ]
          },
          {
            "kind": "report",
            "mode": "single",
            "name": "Report",
            "roleIds": [
              "cost-auditor"
            ]
          }
        ],
        "description": "Read the runs the office has already paid for and say which spend bought nothing. Two stages, no code written."
      }
    ],
    "routingRules": [
      {
        "id": "cheap-first-for-mechanical-work",
        "tier": "small",
        "taskClass": "intake",
        "description": "Intake and summarising are mechanical: prefer the cheapest capable model, and prefer a local runtime when one is configured.",
        "preferProviderIds": [
          "local",
          "deepseek"
        ]
      },
      {
        "id": "no-frontier-for-summaries",
        "tier": "nano",
        "taskClass": "summarize",
        "description": "A summary never needs the strongest tier.",
        "avoidModelIds": [
          "gpt-4o",
          "claude-3-5-sonnet-latest",
          "o3-mini"
        ]
      }
    ],
    "roleTemplates": [
      {
        "id": "cost-auditor",
        "rank": 2,
        "title": "Cost auditor",
        "roomId": null,
        "seatId": null,
        "mission": "keep the office honest about what it spends, and say when a cheaper route would do",
        "persona": {
          "voice": "dry, specific, allergic to waste",
          "values": [
            "evidence",
            "thrift",
            "candour"
          ]
        },
        "skillIds": [
          "cost-aware-delegation",
          "technical-writing",
          "code-review"
        ],
        "reportsTo": null,
        "seniority": "senior",
        "appearance": {
          "height": 1,
          "bodyColor": "#facc15",
          "accentColor": "#78350f"
        },
        "canDelegate": false,
        "displayName": "Mira",
        "modelPolicy": {
          "maxTier": "strong",
          "minTier": "small",
          "defaultTier": "standard"
        },
        "allowedTools": [
          "think",
          "read_file",
          "list_dir",
          "search_files"
        ],
        "departmentId": "frontend",
        "maxDirectReports": 0,
        "maxTurnsPerStage": 4,
        "responsibilities": [
          "read a run's turns and name the ones that bought nothing",
          "propose the cheapest model that would still have finished the task",
          "refuse to call a summary a research problem"
        ]
      }
    ]
  },
  "description": "Keeps the cheap work cheap: routes intake, summarising and research to the least expensive capable model, adds a small local model to the catalog for it to aim at, and offers a cost-auditor role and a spend-review pipeline.",
  "permissions": [
    "models",
    "routing",
    "skills",
    "agents",
    "pipelines",
    "settings"
  ]
}