Read the response contract
A fictional merchant asks FASHN Agent for four outfits on the same model and receives status: completed but sees only two usable final looks. This is not automatically an API transport failure. The experimental October 2026 Agent contract defines completed as having made at least one image. Its predictions array lists operations, including intermediate and failed work. A client should not equate HTTP success or top-level completion with fulfillment of a merchant’s four-item deliverable list.
Reconcile operations and outcomes
For each prediction, inspect operation ID, tool, status, results and credits_used. Completed intermediate images are billed even if they are not selected for publication; failed operations consume zero operation credits. The response’s top-level credits_used is the request total, not a campaign-wide bill. The Agent has an eight-operation cap and a fifteen-minute operation-start window, so a run can return what it made without fulfilling every requested output. Do not assume every successful result is a final look.
| Condition or role | Decision or evidence | Next check |
|---|---|---|
| Top-level completed | At least one image produced | Not full lookbook approval |
| Prediction completed | Operation produced results | Classify final versus intermediate |
| Prediction failed | No usable result from that operation | Investigate error |
| credits_used | Provider operation/request cost | Do not bill only selected finals |
Illustrative four-look ledger
Suppose three operations succeeded and one failed. The reviewer maps two completed images to approved looks, classifies the third as an intermediate candidate and records the fourth look as missing. This fictional state would be two deliverables accepted, one intermediate retained, one unresolved, regardless of top-level completed. The reviewer must inspect actual garments, not infer look numbers from result-array position. No Agent job was submitted, timed or billed for this example.
Make a bounded recovery decision
Record the missing SKU, approved garment source and intended model/location reference. If another request is authorized, target the unresolved look rather than blindly rerunning all four. The multi-SKU release matrix helps track which outputs are approved; campaign budget planning reminds the team that generated intermediates are not final deliverables. Preserve run ID, per-operation images and actual credits. Hold lookbook publication when promised looks are missing, even if the API says completed.