package.json

{
  "name": "token-proctor",
  "displayName": "Token Proctor",
  "description": "Prompt validation, model routing, and token/turn cost projection for GitHub Copilot and other token-priced LLM plans.",
  "version": "0.4.2",
  "icon": "media/icon.png",
  "publisher": "token-proctor",
  "license": "MIT",
  "repository": {
    "type": "git",
    "url": "https://github.com/navintkr/token-proctor.git"
  },
  "homepage": "https://github.com/navintkr/token-proctor#readme",
  "bugs": {
    "url": "https://github.com/navintkr/token-proctor/issues"
  },
  "engines": {
    "vscode": "^1.95.0",
    "node": ">=20"
  },
  "categories": [
    "AI",
    "Chat",
    "Other"
  ],
  "keywords": [
    "copilot",
    "model routing",
    "prompt validation",
    "cost",
    "token",
    "llm",
    "agent"
  ],
  "activationEvents": [],
  "main": "./out/extension.js",
  "bin": {
    "token-proctor-mcp": "./out/mcp-server.js"
  },
  "contributes": {
    "chatParticipants": [
      {
        "id": "token-proctor.proctor",
        "name": "proctor",
        "fullName": "Token Proctor",
        "description": "Validate prompt, route to best model, project turns and token cost.",
        "isSticky": true,
        "commands": [
          { "name": "route",    "description": "Pick the best model for this prompt" },
          { "name": "validate", "description": "Score prompt completeness; ask follow-ups if weak" },
          { "name": "cost",     "description": "Estimate tokens, turns and $ for this prompt" },
          { "name": "explain",  "description": "Explain the routing/validation decision" },
          { "name": "confirm",  "description": "Confirm the previous recommendation and forward to the chosen model" },
          { "name": "cancel",   "description": "Cancel a pending confirmation — no model will be called" }
        ]
      }
    ],
    "commands": [
      {
        "command": "token-proctor.diagnoseModelCommands",
        "title": "Token Proctor: Diagnose Model Commands",
        "category": "Token Proctor"
      }
    ],
    "configuration": {
      "title": "Token Proctor",
      "properties": {
        "tokenProctor.completenessThreshold": {
          "type": "number",
          "default": 60,
          "minimum": 0,
          "maximum": 100,
          "description": "Below this score, Token Proctor asks follow-up questions instead of invoking a model."
        },
        "tokenProctor.autoForward": {
          "type": "boolean",
          "default": true,
          "description": "After routing, forward the prompt to the chosen model and stream the answer."
        },
        "tokenProctor.requireConfirmation": {
          "type": "boolean",
          "default": true,
          "description": "Show the recommended model and wait for user confirmation (Accept & hand off button or /confirm) before calling the model."
        },
        "tokenProctor.handoffToCopilot": {
          "type": "boolean",
          "default": true,
          "description": "When confirmed, hand the prompt off to the GitHub Copilot agent (which has file-edit tools) instead of streaming a plain text answer from vscode.lm. User should pick the recommended model in the chat model picker first."
        },
        "tokenProctor.handoffParticipant": {
          "type": "string",
          "default": "github.copilot",
          "description": "Chat participant id to hand off to when handoffToCopilot is enabled. Defaults to the GitHub Copilot agent."
        },
        "tokenProctor.preferCheap": {
          "type": "boolean",
          "default": false,
          "description": "Bias routing toward the cheapest model that clears the quality bar."
        },
        "tokenProctor.optimizeFor": {
          "type": "string",
          "enum": ["tokens", "turns", "balanced"],
          "default": "tokens",
          "markdownDescription": "How to weigh cost axes in routing:\n- `tokens` — **default** — minimize $/M token price (classic); best for one-shot tasks.\n- `turns` — minimize premium-multiplier × turns; best when agent loops run many turns (e.g. `code_large`, `agentic`).\n- `balanced` — weighted compromise between the two."
        },
        "tokenProctor.exactTokenCounts": {
          "type": "boolean",
          "default": true,
          "description": "Use js-tiktoken for exact token counts. Falls back to a chars/4 heuristic if unavailable."
        },
        "tokenProctor.llmJudge.enabled": {
          "type": "boolean",
          "default": true,
          "description": "When rule-based classifier confidence is low, escalate to an LLM judge that auto-picks the cheapest available model via vscode.lm."
        },
        "tokenProctor.llmJudge.confidenceThreshold": {
          "type": "number",
          "default": 0.85,
          "minimum": 0,
          "maximum": 1,
          "description": "Rule-based confidence below this triggers the LLM-judge. Default 0.85 = judge is consulted on most prompts. 0 = always consult, 1 = never."
        }
      }
    }
  },
  "scripts": {
    "compile": "tsc -p .",
    "watch": "tsc -w -p .",
    "vscode:prepublish": "npm run compile",
    "mcp": "node ./out/mcp-server.js"
  },
  "devDependencies": {
    "@types/node": "^20.11.0",
    "@types/vscode": "^1.95.0",
    "typescript": "^5.4.0"
  },
  "dependencies": {
    "@modelcontextprotocol/sdk": "^1.0.4",
    "js-tiktoken": "^1.0.21"
  }
}