Inspiration Every software team starts the week the same way: a backlog full of issues and no shared answer to "what do we fix first?" Triage is slow, inconsistent, and usually done from memory. We wanted the first answer to be instant, explainable, and available wherever the team already works — in Jira.

What it does TriagePilot triages any Jira issue and returns four things:

  1. Priority — a deterministic score from 0–100 with the reasons behind it (base 20, plus issue type, crash/outage/security keywords, hot labels, and staleness). Bands: P0 ≥ 70, P1 ≥ 50, P2 ≥ 30, else P3.
  2. Likely duplicates — weighted token similarity on summary (0.7) + description (0.3), plus a same-project bonus, threshold 0.30.
  3. Suggested assignee — candidates (assignee, reporter, recent commenters via Teamwork Graph, frequent project assignees) ranked by open workload and component/label match.
  4. Next steps — a draft triage summary and concrete actions (check the duplicate first, assign, confirm reproduction steps the same day).

It is available three ways, all powered by the same engine: a Rovo agent ("Triage MSD-19"), a Jira issue panel, and an MCP server exposing the same three actions (triage-issue, find-duplicates, suggest-assignee). Live demo: a crash bug scores 65/100 → P1 High, flags its near-duplicate at 47% similarity, and suggests an assignee with next steps — identically in the Rovo chat and in the issue panel.

How we built it Forge app (Node 24, UI Kit 2). One pure, dependency-free triage engine (src/triage.js, 12 unit tests) is shared by the Rovo agent, the issue panel, and the MCP tools, so the three surfaces can never disagree. Jira data is read through Forge route templates as the user (api.asUser()), so results are permission-filtered. Teamwork Graph is queried with curated GraphQL for the people around a work item (comments, team). The MCP endpoint is https://mcp.atlassian.com/forge/1458ed72-0d6c-4014-b720-3e7904a96ac9, OAuth 2.1 protected. Read-only by design: every action is a GET — no writes, no state, no external APIs.

Challenges we ran into

  • The issue panel first failed in production with "Disallowing path manipulation attempt": a pre-built path string had been interpolated into Forge's route tag. Fixed by building every route inline in the template literal with single-value interpolation, plus a regression test.
  • Rovo Chat is not available on the free Cloud Developer Bundle site at all (/rovo 404s, zero agents listed) — even after redeploying to work around Atlassian bug ROVO-1239. We verified the agent instead on a Forge demo site (forge site provision, enterprise editions including Rovo), where the agent is listed and answers triage requests end to end.

Accomplishments that we're proud of

  • One engine, three surfaces, zero disagreement between them.
  • Deterministic, explainable scoring — the agent never invents a priority, a duplicate, or a person; missing sources are reported as unread, never as "nothing there".
  • Live-verified on real Jira sites: panel on two sites, Rovo agent on the demo site, MCP endpoint OAuth-protected.

What we learned

  • Forge's route tag is strict about path construction for a reason — and the error message points exactly at the fix.
  • Rovo availability depends on the tenant, not on the app: the same production deployment is invisible on a free dev site and fully working on a demo site.
  • A demo video is only as honest as its numbers: we verified every figure in the video live before recording.

What's next for TriagePilot

  • Confluence support (triage against related docs), JSM queue triage, configurable scoring weights per project, and a Marketplace listing.

Built With

Share this project:

Updates

Submission history