Autopilot & trackers

Autopilot — the overnight researcher

Every time Mave answers with low confidence, that question lands in an auto-generated research backlog (the low-confidence queue on GET /api/v1/quality/chat). Autopilot burns that backlog down overnight inside a credit budget you set — you wake up to a digest in the chat: "While you were away I ran 3 studies on questions I couldn't answer with data."

curl -X PUT "$BASE/api/v1/autopilot" -H "x-api-key: $MAVERA_API_KEY" \
  -H "content-type: application/json" -d '{"budget": 500}'

curl "$BASE/api/v1/autopilot" -H "x-api-key: $MAVERA_API_KEY"

POST /api/v1/autopilot runs a pass immediately (same thing the nightly cron does). Each queue item gets a bounded pipeline — design (one cheap structured call) → audience build → run_study at n=50 — with the budget checked before each item and every step metered. Completed items link back to the original question, so the queue shrinks honestly.

Trackers — living, scheduled studies

A tracker re-fields the same study on a cadence and trends the code-computed metrics. When a metric moves more than your threshold vs the previous wave, Mave alerts you in the chat: "📈 NPS moved -6.2% since last wave."

curl -X POST "$BASE/api/v1/trackers" -H "x-api-key: $MAVERA_API_KEY" \
  -H "content-type: application/json" -d '{
  "name": "Collar NPS monthly",
  "toolName": "run_study",
  "input": { "audienceId": "aud_...", "topic": "collar satisfaction", "questions": [...], "n": 50 },
  "cadence": "monthly",
  "alertPct": 10
}'

• GET /api/v1/trackers — list with the latest wave
• GET /api/v1/trackers/{id}/runs — the full trend series (chart feed: metrics + deltas per wave)
• PUT (enable/disable, cadence, threshold) and DELETE ?id= round it out

Waves run on the nightly worker. Deltas are computed in code from the engine's aggregates — no model touches the numbers.