Chat errors have a special cruelty: they interrupt a story at the moment you are inside it. The good news is that the error zoo in the Janitor AI application is small — empty replies, stuck generation, outright error messages, and the quiet catastrophe of a character who suddenly stops making sense — and each species has a known habitat and a known cure.
This guide works through them in the order you should check: the simple and local first, the account and configuration causes next, and the service-side possibilities last, because most errors are born within arm's reach of your own settings.
Keep notes when an error recurs. A fault with a pattern is a diagnosis waiting to be read, and your notes are the text it is written in.
Empty or Blank Replies
An empty reply usually means the model generated nothing the app could show: a content filter intervened, the response was cut by a length or token limit, or the request silently failed. Retry once; a single retry distinguishes a hiccup from a condition.
If empties persist, shorten your message, cool any extreme settings, and check whether a particular topic or phrasing triggers them. A pattern tied to content points one way; a pattern tied to nothing points to the connection and configuration checks below.
Stuck or Endless Generation
Generation that spins forever is usually a request that never completed its journey: a stalled connection, a busy service, or an app session that has lost its thread. Back out, wait a breath, and retry; then restart the app if the stall repeats.
Long contexts and maximal length settings lengthen every journey. If stalls cluster in your longest chat, the story has outgrown its current configuration, and the chapter-and-summary method in this library is the structural cure.
Error Messages, Decoded Calmly
Error text is terse but rarely meaningless: authentication errors indict the key or sign-in, rate and quota errors indict usage limits, server errors indict the service's day rather than yours. Read the message literally before doing anything dramatic.
Match the response to the indictment. Re-enter credentials for authentication, wait or reduce usage for quota, and wait for server errors while checking whether other users report the same — a crowd experiencing your error means the fix is not on your phone. The credential side is covered carefully in Janitor AI API and model settings explained.
The Character Suddenly Stops Making Sense
A character who derails mid-story — wrong facts, wrong voice, wrong relationship — is usually a context event, not a possession: the window filled with recent noise, an earlier contradiction compounded, or a setting change altered the model's behaviour mid-stream.
Recover by re-anchoring: restate the scene's facts briefly, correct the specific error in-world, and continue. If derailment repeats at similar depths in a chat, the story needs chaptering, which the memory guide in this library treats in full.
When Chats Will Not Load or Send
Interface-level failures — chats that will not open, messages that will not send — belong to the app layer: restart it, check your connection by loading anything else, confirm you are signed in, and clear the cache if the behaviour survives all three.
The complete local sequence lives in Janitor AI not working: common problems and fixes; this page handles the errors that occur inside conversations, while that one handles the app refusing to hold them at all.
Protect the Story Before You Troubleshoot
Before any aggressive fix — cache clears, reinstalls, configuration resets — copy out anything irreplaceable: your summaries, character notes, and the last good stretch of the story. Most fixes are harmless to data, but the one that is not will choose the day you skipped this step.
Thirty seconds of copying converts troubleshooting from a risk into a routine. Professionals back up before they repair in every field, and story-keepers are professionals of exactly this.
Knowing When It Is the Service
The final diagnostic is social: if the app fails identically on a fresh chat, a different character, and a different network, and other users are reporting the same hour's misery, the fault is the service's and the cure is patience.
Spend the wait usefully — update your summaries and character notes — and let the story resume when the platform does. A well-kept story loses nothing to an evening's outage.
Read the error, check the local causes in order, protect the story first, and let patterns name the deeper faults. That sequence resolves nearly every interrupted conversation you will ever meet here.
An error log of three lines per incident will make you, within a season, the person other users ask. That is how expertise in any tool actually accumulates.
There is a composure skill underneath the whole page: an interrupted scene feels like loss and is almost always only interruption. Users who internalise that distinction troubleshoot calmly, lose less story, and enjoy the platform more — because they have stopped treating weather as climate.
Keep the error log beside the story summaries, because the two documents explain each other: the night the replies emptied is the night the chapter break moved, and future troubleshooting reads both. Your archive is small; its organisation is what makes it wisdom rather than clutter.
Calm is a diagnostic instrument. Keep yours calibrated.
