Chat tells you when it cannot answer, instead of guessing

Two things could go wrong in an AI chat round without the thread ever saying so.

Losing your earlier messages, silently. When the conversation's prior turns could not be loaded in time, the round carried on regardless — the agent saw only your newest message and answered as though the thread had just been created. Nothing on screen said anything had gone missing, so a long triage or research thread could quietly repeat work or contradict what it had already established, and the round still finished marked as successful. A round that cannot read the conversation now stops and says so, and invites you to submit again; it no longer answers on a blank slate.

Unreadable provider errors. When the language-model provider refused a round — the account out of credit, or a model that no longer exists at the configured endpoint — the thread showed the raw error the provider's client library produced (HTTP 402 (: ) followed by an English paragraph about token budgets). That text was untranslated and mostly meaningless unless you administer the deployment. Those two refusals now read as a plain, translated sentence naming the model and what to do about it — top the provider account up, or check the model id in Settings → Language Models. The full technical detail still goes to the logs for whoever operates the portal.