Eval health & longitudinal trends

GET /evals/trends tracks your agent pipeline quality over time — not just whether a single run passed, but whether your team is getting better or worse across weeks or months.

Why this matters

A one-off eval score tells you nothing about drift. You change a prompt, re-run, see 0.83 — is that good? You need the baseline. Eval health gives you a rolling view:

• Average score over the last N days
• Pass rate — what fraction of evals met all expectations
• Trend — improving, declining, or stable, computed by comparing the most recent third of data points against the oldest third

The dashboard Eval health panel shows the sparkline automatically once you have at least one completed eval.

API

Get eval trends

GET /evals/trends?team=DevTeam&days=90&limit=200

Param	Default	Description
team	(all)	Filter by team name
days	90	Lookback window in days
limit	200	Maximum data points returned

{
  "team": "DevTeam",
  "days": 90,
  "data": [
    {
      "date": "2026-06-01",
      "eval_id": "abc123",
      "name": "Build JWT auth",
      "team": "DevTeam",
      "overall_score": 0.82,
      "passed": true,
      "cost_usd": 0.0041,
      "elapsed_ms": 14200
    }
  ],
  "summary": {
    "total": 22,
    "avg_score": 0.84,
    "pass_rate": 0.91,
    "passed": 20,
    "failed": 2,
    "trend": "improving"
  }
}

Data is returned in chronological order (oldest first) — suitable for direct rendering in time-series charts.

Trend calculation

trend compares the mean overall_score of the oldest 33% of data points against the most recent 33%:

Condition	trend
New mean > old mean by >3 pp	"improving"
New mean < old mean by >3 pp	"declining"
Otherwise	"stable"

Requires at least 6 data points to compute a meaningful trend; returns "stable" otherwise.

Scheduled evals for continuous tracking

Use EvalSchedule to run evals automatically on a recurring basis, so the trend chart stays current without manual intervention:

POST /evals/schedules/
{
  "name": "DevTeam nightly",
  "team": "DevTeam",
  "request": "Build a JWT authentication module with refresh tokens",
  "interval_hours": 24
}

Each scheduled run persists its overall_score and passed state, feeding the trends chart automatically.

Dashboard panel

The Eval health panel appears automatically in the dashboard once at least one eval has been completed. It shows:

• A sparkline of overall_score over time (green dots = passed, red = failed)
• Trend badge (↑ Mejorando / ↓ Bajando / — Estable)
• Average score and pass rate for the period
• The 5 most recent evals with individual scores

No configuration required — the panel fetches GET /evals/trends?days=90 on page load.