Bug Triage Agent is a Claude plugin that triages a ticket the way a careful engineer would, from evidence rather than the reporter's wording. It decides what the ticket is before giving it a priority, gathers evidence from the tracker, releases, metrics, warehouse and code in parallel, correlates the bug with the release that most likely caused it, finds a workaround, and scores its own confidence by script. One shared rubric serves every team; each team owns a single config file.

Bug triage decides what engineers work on next, and it is usually done from the ticket's wording. Non-defects get scored as if they were bugs, priority follows whoever shouts loudest, and nobody checks which release changed the affected area or how many customers are hit. Every team builds its own rubric, so there is no single quality bar across an organisation.
Classify before scoring: a ticket is a defect, expected behaviour, a config or data issue, a duplicate or a customer request before any priority is assigned, because scoring a non-defect is worse than not triaging it. Keep one shared, versioned rubric and let each team configure rather than fork it. Put the parts that must be exact into scripts rather than prompts, and turn every mistake seen in a live run into a recorded eval case so the guard that should have caught it cannot quietly regress.
A preflight check and an information gate run first and spend no AI tokens. Disposition comes next; anything that is not a defect stops there and is routed, with no priority. For defects, a connectivity gate checks enough sources are connected, then parallel subagents read the tracker history, releases and deploys, metrics and logs, the warehouse and the code, each through a pinned, read-only connector. The agent correlates the bug with a release, looks for a workaround and assigns P1 to P4 with a stated reason, while a script computes the confidence score and audits which tools were used. Output is an HTML report with a timeline and evidence links, a fix brief for the engineer, an optional tracker comment, and an audit log that records every human override.
Bug Triage Agent turns triage from an opinion into a decision a team can check: every verdict carries its sources, its release, its affected accounts and a computed confidence. In a live run on synthetic demo data it traced a CSV export defect to a wrong bound in csvExport.ts introduced by a specific pull request, sized it at eight accounts, and scored the result from the evidence it actually read.