Chat Errors Now Say What to Do Next
October 2, 2026 · InfraScout Team
When a chat turn fails, you should be able to tell from the message alone whether to press Retry, reply "continue", or ask an administrator to change a setting. Every chat and workflow error now says those three things: what happened, who can fix it, and where. Answers cut short by a connection's output-token cap are flagged, so they no longer pass as complete.
Answers that hit the output-token cap
Each AI provider connection has a Max Tokens setting that limits how much the model can write in one reply. Every model in the picker reasons before it answers, and that reasoning counts against the same budget. On a long assessment summary, the model could use up the whole cap before it finished.
Before this release, that went wrong in two ways. An answer cut off mid-sentence was saved as if it were complete, so nothing told you the last section was missing. A turn that spent its whole budget on reasoning showed an error that never mentioned the cap.
Both cases now say what happened. A cut-off answer is kept and gets a warning card underneath. A turn with no answer at all shows an error. Both tell an administrator to raise Max Tokens under Admin › AI Integrations › AI Providers and suggest a value: double the current cap, at least 32,768, and never more than the model can produce. This works the same way for Anthropic, OpenAI, and Azure AI Foundry connections.
A larger default, and caps the model can actually use
New connections now start at a Max Tokens of 32,768 instead of 8,192. Existing connections keep their current value.
Check your existing connections
If any of your connections were created with the old default of 8,192, consider raising them now. Reasoning models regularly need more than that for long answers. Open Admin › AI Integrations › AI Providers, look at the Max Tokens column, and edit any connection that is still at 8,192.
When you save a connection, InfraScout now checks Max Tokens against the model's own output limit. A value that is too high is refused, and the message names the limit. Before, the connection saved and the next chat failed. Current Claude 5.x models, Sonnet 4.6, Opus 4.6 and later, and the GPT-5 and GPT-6 families allow up to 128,000 output tokens. Haiku 4.5, Sonnet 4.5, and Opus 4.5 allow up to 64,000. A connection you saved earlier still saves when you rename it or rotate its key. The check runs only when you change the model or the cap.
Every error tells you who acts and where
Provider failures now start with a plain sentence of our own, not the provider's raw text. Each one names the cause: credentials that don't work, a model or deployment that can't be found, used-up quota or credit, a rate limit, or a content filter. If an administrator has to fix it, the message names Admin › AI Integrations › AI Providers. If you can retry, it says so.
Messages from the chat itself are clearer too. When a turn reaches the step limit, the message says it stopped after 40 steps and that you can reply "continue" to pick up where it left off. A message that fails to send shows plain text, not raw server output. If your browser can't reach the service at all, you see "Couldn't reach InfraScout".
Provider detail is for administrators only
A provider's own error text is useful when you're fixing a connection. It can also name deployments, endpoint hosts, organization IDs, and billing details that most users shouldn't see. That raw text now goes to administrators only, as a separate Provider detail line under the message. Everyone else sees only our sentence. This applies both while the error is happening and when someone reopens the conversation later.
Task Agents follow the same rule. If a Task Agent can't start because its AI provider can't be set up, the task card no longer shows infrastructure details. It says the provider couldn't be set up and points an administrator to AI Providers.
Memory rollup says why it was refused
Administrators can run a memory rollup by hand from a conversation. When that was refused, all you saw was "Memory rollup failed". The message now gives the reason and the fix. The most common reason is that no connection has Memory Rollups turned on. That setting needs an Anthropic connection, so if you only have OpenAI or Azure AI Foundry connections, add an Anthropic one first.
If you see that message, it is worth acting on even if you never run a rollup by hand: without a rollup connection, the background pass skips your tenant entirely, so none of your conversations have been folded into memory so far.
If a rollup is already running for the conversation, you now see a notice, not an error. When the request succeeds, the message says the rollup has started rather than finished, because rollups are queued and run in the background.
Find out more
The AI providers overview explains how connections, models, and providers fit together. The AI provider connections reference covers each setting on a connection, including Max Tokens and Memory Rollup.
Questions or feedback? Reach us at info@infrascout.cloud.