{
  "name": "cost-guard",
  "version": "0.2.0",
  "description": "Real-time USD cost enforcement for Claude Code. Halts runaway agentic loops mid-step before they burn through your budget.",
  "author": {
    "name": "OGGEH, Inc"
  },
  "license": "MIT",
  "keywords": [
    "cost",
    "budget",
    "rate-limit",
    "observability",
    "safety"
  ],
  "userConfig": {
    "enabled": {
      "type": "boolean",
      "title": "Enable cost-guard halting",
      "description": "Master switch. When off, no halts fire (ledger still records).",
      "default": true
    },
    "max_usd_per_session": {
      "type": "number",
      "title": "Per-session cost cap (USD)",
      "description": "Halt new work when this session's total spend reaches this value. Primary defense against a session burning too much over time.",
      "default": 10.0,
      "min": 0.1
    },
    "max_usd_per_step": {
      "type": "number",
      "title": "Per-step cost cap (USD)",
      "description": "Halt the agentic loop after one Claude step (one user prompt \u2192 Claude's tool calls + final answer) has spent this many dollars. Primary defense against a single greedy step.",
      "default": 2.0,
      "min": 0.05
    },
    "rate_per_min_usd": {
      "type": "number",
      "title": "Burn-rate limit (USD/min)",
      "description": "Halt mid-step when the trailing 5-minute session burn rate exceeds this value.",
      "default": 0.5,
      "min": 0.01
    },
    "max_subagent_spawns_per_step": {
      "type": "number",
      "title": "Max subagent spawns per step",
      "description": "Deny further Agent/Task tool calls within one step once this count is reached. Bounds fan-out cost.",
      "default": 8,
      "min": 1,
      "max": 100
    },
    "warn_at_usd": {
      "type": "number",
      "title": "Per-session warning threshold (USD)",
      "description": "Emit an amber warning on UserPromptSubmit when session spend crosses this dollar amount (always below max_usd_per_session).",
      "default": 7.0,
      "min": 0.05
    },
    "seed_from_history": {
      "type": "boolean",
      "title": "Seed ledger from existing transcripts on first run",
      "description": "On first install, scan recent ~/.claude/projects transcripts to populate the rolling cost ledger. Gives accurate trailing rates on the first prompt.",
      "default": true
    },
    "log_decisions": {
      "type": "boolean",
      "title": "Log halt decisions",
      "description": "Write an audit line to halt-log.jsonl for every halt/warn decision. Useful for diagnosing false positives.",
      "default": true
    }
  }
}