> ## Documentation Index
> Fetch the complete documentation index at: https://docs.custral.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Troubleshooting

> When auto-fill goes wrong, here is how to read the run trace.

Every property agent run is recorded with the trigger kind, the status, what the AI produced, the value that was written, the prior cell value, the duration, and any error message. Open a run from the agent panel to see the full trace, or right-click the cell in a table view and choose **View AI runs for this cell** to see the history for that exact record and property.

<Note>
  The run trace shows what the AI **produced** and the value that was **written**, not the full prompt that was sent. Use what the AI produced to see what it actually said when the written value looks unexpected.
</Note>

## Common failures

A run is either successful or failed; failures carry a descriptive message.

| Symptom                                                   | Likely cause                                                                      | Fix                                                                                                    |
| --------------------------------------------------------- | --------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------ |
| Error that the AI's output didn't match the property type | The AI returned text that doesn't fit the property type.                          | Tighten the prompt; for a select, list the valid options in the prompt.                                |
| Error that no value could be determined                   | The record (and the web, if enabled) didn't contain enough to answer.             | Add more context to the record, refine the instructions, or enable the web source.                     |
| Error that the AI declined to answer                      | The model refused.                                                                | Reword the prompt with more context; enable web research if appropriate.                               |
| Error about an image                                      | An image auto-fill couldn't find, validate, download, or attach a suitable image. | Refine the search hint in the instructions; the error message says which stage failed.                 |
| Cell updated but the value is wrong                       | The value passed validation but was plausible-but-wrong (rare).                   | Add validation on the property; consider a Classify kind with explicit options, or tighten the prompt. |

## A run that didn't write a value

Not every "no write" is a failure. With **only-if-empty** mode, a run succeeds with a "Skipped: property already has a value" note. With **required properties** configured, a run succeeds with a "Skipped: required property is empty" note (and no AI call) when a prerequisite property is empty. These are intentional skips, not errors.

## Bulk reruns

If you change the prompt and want to re-evaluate, run the agent across a bounded set of records. Don't run unbounded bulk reruns. The cost adds up quickly, and Custral processes large runs in batches.

## Cost

Property agent runs use the same AI models as chat. Each run records how long it took, and your total workspace spend is visible in the Usage block.


## Related topics

- [Troubleshooting](/comms/sms/troubleshooting.md)