Get Predictions

Evaluate all predictive forecasts and return active alerts.

Checks three forecast dimensions:

• Budget exhaustion: At current spend velocity, when will the budget cap be reached?
• Completion rate decline: Is the task completion rate trending downward, indicating the run will take longer than expected?
• Run duration overrun: Based on current throughput, will the run exceed the configured time window?

Use budget_cap to enable the budget forecast. The run duration forecast requires at least one completed task.

Both numeric parameters are echoed back in the response body, so both refuse non-finite values with a 422 instead of admitting them: a range bound alone does not exclude them (inf >= 0.0 is true, and every comparison against NaN is false), and the JSON renderer cannot serialise either one.

The budget forecast is scoped to tenant_id: the spend series it is built from is narrowed to cost points recorded for the caller's tenant, the same way the rest of the cost surface is (see load_cost_history). Cost points written before per-tenant attribution existed are treated as the default tenant's spend, so a legacy single-tenant install keeps its existing numbers.

Returns a list of alerts ordered by severity (critical first). Each alert has: kind, severity, message, minutes_until_impact, confidence.