When something goes wrong, StewAI shows you exactly what happened and where.
Spotting failures
- In Orders, failed runs show a red progress bar
- On the canvas, failed steps have a red status indicator
- Steps that depended on a failed step stay gray (they never ran)
Reading error messages
Click on a failed step. The side panel shows an error callout — a red box with:
- Error title — A short description like “Template rendering failed” or “Provider error”
- Message — More detail about what went wrong
- Details (expandable) — Technical information you can copy if you need help
Common errors and fixes
”Template rendering failed”
What it means: A token reference in your prompt points to something that doesn’t exist or didn’t produce output.
How to fix:
- Check that all token pills in the step’s prompt are valid
- Make sure the referenced step exists and hasn’t been deleted
- If you’re inside a loop, make sure you’re referencing the right scope
”Provider error”
What it means: The AI model returned an error — usually because the service is temporarily overloaded or the API key is invalid.
How to fix:
- Wait a minute and re-run (often it’s a temporary issue)
- Try a different model (switch from GPT-4o to Claude Sonnet, for example)
“Rate limited”
What it means: Too many AI requests in a short time.
How to fix: Wait a moment and try again. The error message includes how long to wait.
”Invalid step parameters”
What it means: A step’s configuration is missing something required, like a prompt or a schema.
How to fix: Open the step’s settings panel and check for empty required fields.
”Configuration error”
What it means: Something in the step’s setup doesn’t match what’s expected.
How to fix: Review the step’s settings. Common causes: wrong data type for a field, referencing a deleted schema, or an empty prompt.
”Compute error”
What it means: A Compute step’s calculation failed — usually because the input data was in an unexpected format.
How to fix: Check that the upstream step produced the data the Compute step expects.
Debugging strategies
Click on upstream steps first. If Step C failed, check Steps A and B. Often the problem is in the data that was passed in, not the failing step itself.
Use Input steps as breakpoints. Add a temporary Input step between two steps. When the run pauses there, you can see what data was produced so far before deciding to continue.
Simplify to isolate. If a complex recipe fails, try a minimal version with just the failing step and its direct inputs. This helps you isolate whether the problem is in the step or the data.
Check the model. Some models handle complex tasks better than others. If a step fails on GPT-4o Mini, try GPT-4o or Claude Sonnet.
Re-running after a fix
After fixing the issue:
- Open the recipe in the Cookbook and make your changes
- Create a new run from Orders → +
- The new run uses the updated recipe
See also
