Errors

DimiLinks error responses follow the upstream protocol in use:

Endpoints under https://api-direct.dimilinks.com/v1/* (OpenAI-compatible) return the OpenAI-style error wrapper.
https://api-direct.dimilinks.com/v1/messages (native Anthropic) returns the native Anthropic format {"type":"error","error":{...}}.

OpenAI style (/v1/* endpoints):

{
  "error": {
    "message": "prompt must not be empty",
    "type": "invalid_request_error",
    "code": "invalid_request_error"
  }
}

Anthropic style (/v1/messages endpoint):

{
  "type": "error",
  "error": {
    "type": "invalid_request_error",
    "message": "messages must not be empty"
  }
}

Common errors

HTTP	code	Meaning	Recommended caller action
`400`	`invalid_request_error` / `invalid_reference_image`	Invalid parameters, an empty prompt, or an invalid reference image	Ask the user to correct the parameters
`401`	`missing_api_key` / `invalid_api_key`	Missing or invalid key	Check the key configuration
`402`	`insufficient_balance`	Insufficient balance	Ask the user to add funds or contact an administrator
`403`	`model_not_allowed`	The current key cannot access the model, or its available groups do not intersect the model’s `enable_groups`	Check the model allowlist; see Model groups & pricing to change `group` or use a different key
`404`	`not_found`	The task does not exist or does not belong to the current user	Stop polling
`429`	`rate_limit_rpm` / `rate_limit_tpm`	Rate limit exceeded	Retry with backoff and reduce concurrency
`500`	`internal_error` / `billing_error`	Internal service error	Record logs and retry later
`502` / `503`	`upstream_error` / `service_unavailable`	The upstream route is temporarily unavailable	Allow the user to retry

For 400, show error.message directly so the user can adjust the input.
For 401 / 403, ask the user to check the key or model permissions; do not retry silently.
For 429, use exponential backoff and reduce concurrency; do not retry immediately at high frequency.
For 5xx / 502 / 503, allow retries, but record logs for upstream troubleshooting.
Stop polling immediately when a task endpoint returns 404 to avoid wasting requests.
If the same application code calls both protocols, add a shared HTTP client abstraction that reads either error.message or error.error.message instead of duplicating compatibility logic in every caller.