Read pipeline progress
Read each document target's stage state, unit counts, warnings, and failure reasons.
Read documents[].targets[].stages in a job response to understand what is happening for one file and language. Stages appear after the engine configuration is captured at start; the array can be empty before that.
Each entry contains stage, status, and units. A progress excerpt:
{
"stage": "humanEdit",
"status": { "_tag": "Running" },
"units": { "finished": 3, "total": 8 }
}Here, three of the target's eight units have completed or skipped human review. Inspect the review to see which units still require a decision.
Stage identifiers #
| Identifier | Work |
|---|---|
preEdit | Refine source text before translation. |
localize | Produce the target translation. |
rephrase | Refine target phrasing. |
qualityGate | Find errors and use their severity to decide whether human review is required. |
humanEdit | Wait for internal or external review. |
postEdit | Refine reviewed output. |
contentLength | Adjust wording toward configured length limits. |
backTranslation | Check meaning through reverse translation and correction. |
Configure stages on the engine. Only enabled stages retained at job start are listed.
Stage status #
status._tag | Meaning and additional fields |
|---|---|
Pending | Work has not begun. |
Planning | Work is being prepared. |
Waiting | Waiting for preceding work. |
Running | Work is active; human review can remain here while decisions are pending. |
Completed | Finished; includes finishedAt. |
CompletedWithWarnings | Finished with retained warnings; includes finishedAt and nonempty warnings. |
Skipped | No execution was needed; includes finishedAt and nonempty reasons. |
Cancelled | Work stopped; includes finishedAt and nonempty reasons. |
Failed | Stage failure; includes finishedAt and nonempty messages. |
finishedAt is an ISO timestamp. units.finished counts completed or skipped units, and units.total counts the target's units. Before document preparation, both can be zero. These are content-unit counts, not model-request counts, elapsed time, or a billing estimate.
Decide whether to accept an output #
A required translation failure prevents the affected output from completing. Optional stages can retain preceding text with warnings. Consequently, a completed job still needs its stage warnings checked if your publication policy depends on review, length, or another quality requirement.
A human-review timeout can preserve AI-generated text without a human approval. Later stages can also modify a person's wording. Read the review policy and verify the final downloaded file before publishing.
The API returns progress here. Use Studio's run view to investigate the model attempts and detailed execution behind a stage.