JSON-RPC Security Error Codes

JSON-RPC Security Error Codes 🛑

When FaradAiWire’s Policy Enforcement Point (PEP) intercepts an adversarial payload, policy violation, or rate limit breach, it returns a standard JSON-RPC 2.0 error response with a deterministic negative integer code (pkg/mcp/protocol.go).

{
  "jsonrpc": "2.0",
  "id": "req-004",
  "error": {
    "code": -32004,
    "message": "Security policy violation: AWS secret key detected in arguments (aws_credentials)",
    "data": {
      "filter": "dlp",
      "verdict": "BLOCKED",
      "rule": "aws_credentials"
    }
  }
}

Gateway Security Error Catalog

Code	Constant	Meaning	Typical Trigger
`-32000`	`ErrGateway`	Gateway internal error or rate limit burst exceeded	Token-bucket rate limit exhausted.
`-32001`	`ErrPrefillBlocked`	Assistant prefill injection detected	Agent input contains conversational prefill hijack tokens.
`-32002`	`ErrLoopDetected`	Infinite reasoning loop detected	The exact same tool was called ≥ max_repeat times with identical arguments.
`-32003`	`mcp.ErrPolicyDenied`	Policy authorization denied	Tool is blacklisted in `agents/<agent>.yaml` (Standardized in ADR-0050).
`-32004`	`ErrDLPBlocked`	Sensitive data pattern detected	Argument contains private keys, `.env` references, or credentials.
`-32005`	`ErrHITLRejected`	Operator rejected or approval timed out	Human operator declined execution of a destructive tool or failed to confirm within SLA.

Standard JSON-RPC 2.0 Protocol Codes

FaradAiWire also enforces standard JSON-RPC 2.0 specification errors:

Code	Meaning	Reason
`-32700`	`Parse error`	Inbound payload was not valid JSON.
`-32600`	`Invalid Request`	JSON-RPC frame missing `"jsonrpc": "2.0"` or `"id"`.
`-32601`	`Method not found`	Requested MCP method (e.g. `unknown/method`) is not implemented.
`-32602`	`Invalid params`	Arguments fail JSON Schema validation or mandatory fields are missing.
`-32603`	`Internal error`	Upstream MCP tool crashed or returned malformed output.

Recommended Agent Client Error Handling

When developing autonomous LLM agents or prompt scaffolding, guide the model’s behavior based on error codes:

Upon -32004 (DLP Blocked): Prompt the LLM: “Your previous tool call was blocked because it contained credentials. Strip the secret and retry without exposing sensitive data.”
Upon -32002 (Loop Detected): Prompt the LLM: “You are repeating the same tool call without progress. Change your approach or ask the user for clarification.”
Upon -32005 (HITL Rejected): Prompt the LLM: “The human operator refused authorization for this destructive action. Do not attempt to retry this tool.”