// Product
AI triage
What triage writes on every ticket, how credits work, and why running out never blocks your pipeline.
Every report arrives pre-worked: summarized, classified, and pointed at code — generated right after capture, before anyone opens the ticket.
What triage writes
- A plain-English summary with reproduction steps and a severity — low, medium or high — with a confidence score.
- A classification: bug, change request, question, or other. Downstream formatting adapts — change requests are never framed as defects.
- From → to extraction for change requests: the AI pulls out exactly what should change and to what, preserving the original strings verbatim so you can search for them.
- A suggested title — but only when the reporter didn't write one. A derived title gets replaced; a human-written title is never touched.
- A suspected source location — see Source attribution.
Summaries are written in English regardless of the report's language, but quoted UI strings and error messages stay in their original language so they remain searchable.
Summaries that link to evidence
Sentences in the summary are clickable: "POST /orders returned 500" opens that exact network entry, "TypeError follows" opens that console line, and file references point at the file and line. The summary is a map of the capture, not a paraphrase of it.
Duplicate detection
Tickets are embedded semantically and matched against the last 90 days. Near-duplicates (0.86 cosine similarity) are marked and linked to the canonical ticket, and the canonical tracker issue gets a cross-reference comment so related reports stay visible from the original issue.
Duplicates are detected and linked — your queue stays a list of problems, not a list of complaints.
Credits
Simple metering, no surprises:
| Fact | Value |
|---|---|
| 1 credit | 1 AI-triaged ticket |
| Included | 200 credits per project, per month |
| Free forever | Semantic search, duplicate matching, embeddings |
| Meter | Live usage meter in Settings |
| Notices | Owner emails at 80% and 100% |
| Failures | Credits are refunded if the model call fails |
A no-signal gate runs before the model does: reports with nothing to work with are detected early, so empty reports never burn a credit.
Re-running triage
Owners and team members can re-run the AI on any ticket — after a failure, or once credits refill. Re-triage regenerates the summary only; it never re-creates tracker issues or duplicate comments, and it costs one credit like a normal run.
The Debug view
Any ticket thread morphs into a Debug panel: the AI triage on top, tabs for steps, console, network and environment underneath, and the suspected source — in the widget, on the page. There is no separate dashboard to open.
What the AI sees
Triage reads a scrubbed, size-budgeted slice of the capture — never the raw payload, and never anything that was masked at the source. Your data is never used to train models.