AI Scrum Master guide
Draft plans, health score, velocity forecast, rebalancer and the approval gate.
AI Scrum Master guide
The AI Scrum Master lives at /dashboard/ai-scrum-master. It produces drafts and explanations. It does not change your board on its own.
The approval gate
Every recommendation is a draft until a human accepts it. Accepting writes the change through the normal task and sprint endpoints, which means it produces the same notifications and activity entries as a manual edit. Dismissing a recommendation moves it to the dismissed archive on the same page, where it stays readable.
What is on the page
- Sprint health score: a single gauge built from completion, blocker count and scope movement.
- Meeting & PTO load (next 7 days): calendar pressure per member, so a plan can account for absence.
- Burndown projection: actual, ideal and AI-projected lines on one chart.
- Velocity forecast: what the team is likely to complete, based on finished story points from recent sprints.
- Workload rebalancer: suggests moving work away from overloaded members and names who could take it.
- Automation log: every run, what it produced, and what you did with it.
- Dismissed archive: recommendations you rejected, with the reason recorded.
Generating a draft
- Open /dashboard/ai-scrum-master.
- Run a recommendation. The call goes to /api/dashboard/ai-recommendation and reads capacity, history and skills.
- Read the rationale. Each suggested item comes with a plain-language reason, not just a score.
- Accept, edit or dismiss.
Insights
/api/ai-insights stores generated insights as rows you can list, edit and delete: GET /api/ai-insights, then PATCH or DELETE /api/ai-insights/<id>. The assistant surfaces are the workspace assistant (/api/assistant/workspace, from /dashboard/assistant and the top-bar panel) and the public site assistant (/api/assistant/site); /dashboard/chat is message storage only and makes no AI call.
AI provider
The AI provider is Google Gemini (lib/ai/gemini.ts). Set GOOGLE_GEMINI_API_KEY in your environment; unset = AI controls render disabled, log/reports/exports work fully. GEMINI_MODEL is optional; unset = auto-discover first text-capable generation model on cold start. The generation endpoint (/api/ai/generate) returns usage metadata; error kinds (auth, quota, blocked, timeout, transport, badjson) map to UI-friendly states.
Retrospectives
/dashboard/retros runs the retro side of the cycle: notes, votes on notes, and action items with owners. Each write goes through /api/retros/<id>/notes, /api/retros/<id>/notes/<noteId>/vote and /api/retros/<id>/actions.