When a stage fails
When a stage can't finish, the run's status becomes Failed and the stage shows Failed. To find out why, open the stage's Attempts and read the log. The last lines say what happened, such as Stage failed: …. For a stage that repeats for each item, the log also has an Item failed: … line for each failed item, and the stage's own line starts with the number of the first item that failed.
Reading it is easier if you know the kinds of failure.
What the attempts say
Each attempt has an outcome:
| Outcome | Meaning |
|---|---|
| Success | It worked |
| Awaiting Approval | It worked and is waiting for your review |
| Check Failed | Its output failed a check |
| Qc Failed | Quality control scored it too low |
| Qc Error | The quality control judge couldn't produce a score |
| Qc Budget Exhausted | The judge would have gone over its cap |
| Provider Error | The AI service returned an error |
| Provider Timeout | The AI service didn't answer in time |
| Infra Error | Something went wrong inside Reelcraft |
| Budget Blocked | The next step would have gone over a budget limit |
| Rejected | You rejected it |
| Cancelled | The run was cancelled |
An attempt that failed isn't always the end: Reelcraft tries again, within the limits described in Stages, and the stage only fails once those run out.
Common failures and fixes
check_failed after N attempts
The output kept failing a check. The message names the checks. Open the attempts to see each one's message.
- If the checks are right, improve the stage's prompt, using the failure messages, or raise Max check attempts.
- If a check is wrong, such as a limit that's too tight, fix it in the blueprint.
- A check that is itself broken, such as a script with an error, is reported as an authoring problem.
qc_failed after N attempts
The judge kept scoring the output below the threshold, and the message ends with its last critique. Sharpen the prompt around the critique, lower the Threshold, or choose Hand off to human review so you decide.
Model reported …
The model said it couldn't do the task, with a code such as input_missing, and a reason. Fix the input or
the instructions it names. See Prompts.
QC could not run
Quality control itself failed to run, for example because the judge model didn't answer or can't hear audio and no Deepgram key is set. The stage doesn't fail for this: the run pauses as Paused Approval with the output waiting, and the review says Quality control could not run and why. See Review output.
Provider errors and timeouts
The AI service refused or didn't answer. Typical causes are a missing or wrong key, an account out of credit, a model that's no longer available, or the service being down.
- Check the key in Settings. See AI provider keys.
- Check you have credit with the provider.
- Try again later, then Continue the run.
A stage retries these automatically up to its Retries. After that, the stage fails.
ChatGPT, Codex and Claude jobs that stall (a ChatGPT tab that sits idle, or a job still running long past its own limit) are cancelled and started again by Reelcraft itself, without using up Retries. Only repeated stalls fail the stage.
unknown provider "undefined"
The stage has no model. Set one, or a default. See Models.
The run paused instead of failing
Running into a budget limit pauses the run, it doesn't fail it. See Budget and costs.
After a failure
- Fix the cause: the prompt, the check, the key, the budget.
- If you changed the blueprint on the canvas, Continue the test run there: it applies your edits. A production run always uses its saved version; save a new version and start a new run with it. See Test runs on the canvas.
- If you only fixed something outside the blueprint, such as a key or your credit, select Continue, or Run this stage only to run just that stage while keeping passed items. Choose Run from this stage to the end only when you want to regenerate completed items too. See Retries.
Still stuck? See Troubleshooting.