Provider troubleshooting

Symptom	Check
401 / 403	Provider, complete key, validity, permissions; repeated requests do not fix an invalid key
429	Quotas/rate limits; allow backoff or reduce request frequency
404 / 405	Base URL/protocol; initial auto can fall back, selected-protocol 405 has bounded retry
5xx	Temporary upstream failure; retain error evidence
Premature stream end	Incomplete result; output streams are not automatically replayed
Broken tool JSON	Bounded repair results; broken arguments never execute
Empty/failed model list	Enter the exact ID manually; discovery and inference are separate endpoints

Web model-list failures show correlation IDs for JSONL web_model_list_started / web_model_list_finished, recording domain/timing/errors without upstream bodies or keys. The post-save connection test performs real model inference without device commands.

For special providers try explicit chat_completions / responses, first a short request then a read-only Agent task. Override known context limits when metadata is missing; unknown token usage is not zero.

See custom providers and proxies.