Roadmap & Future Design
UniTrack’s roadmap is tracked as GitHub epics and issues on the project repository. The P0–P2 feature set — ingestion, trends, flaky detection, quality gates, coverage flags, report merging, failure clustering, triage, GitHub checks, the dashboard, and accounts/API tokens — is delivered. The pages in this site document that shipped functionality.
The sections below summarise what is planned next.
1. Accounts, auth & settings (epic #38)
Build on the existing account system:
-
Per-project settings, persisted — move quality-gate and GitHub config from global properties into per-project records with a settings UI (#44, #58–#61).
-
Roles & project membership (RBAC) — global roles plus Owner/Write/Read membership, a members admin page, and write authorisation by membership (#45, #62–#65).
-
Notifications — email infrastructure, token-expiry reminders, gate-failure alerts, and per-user preferences (#46, #66–#69).
2. More formats & quality signals (epic #48)
Broaden what UniTrack can ingest and the signals it derives:
-
More coverage formats — a parser abstraction with auto-detection, plus Cobertura, LCOV, and Coverlet/OpenCover/coverage.py parsers (#49, #70–#73).
-
Performance-test ingestion — a performance data model and parsers for JMeter JTL, Gatling and k6, with trend views (#50, #74–#77).
-
Unit-test performance — slowest-tests views and per-test / suite duration trends (#51, #78–#80).
-
Performance regression detection — latency/throughput and slow-test regressions vs a baseline (#52, #81–#82).
-
Test regression report — a new-failures / new-passes diff against the baseline with its own UI and API (#53, #83–#84).
3. Dashboard & integrations
-
Configurable dashboard widgets + real-time ingest (#21).
-
MCP server via Spring AI — expose UniTrack’s data to AI agents through the Model Context Protocol (#54).
4. AI analysis (epic #346)
Layer LLM reasoning on top of the deterministic signals — only where it beats heuristics (flaky/coverage/clustering stay exact SQL):
-
Failure root-cause + fix direction and a run/PR narrative — the wedge: turn a failure cluster (stacktrace + test source + diff) into "likely a missing null check in
X— look here", and a one-paragraph quality summary posted on the commit/PR. -
Flaky diagnosis, regression explanation over history (consumes the broken-since signal), and coverage-gap prioritisation follow.
Cost stays negligible: each failure signature is analysed once and cached (UniTrack already clusters by signature), only failing runs are analysed, the model is tiered, and context is prompt-cached — so cost scales with unique problems, not build volume. Self-host is bring-your-own-key; analysis can run in-app or be delegated to the MCP server (#54).