summarize - the numbers, deterministically

Aggregate records in one pass. Free - never calls a model. A compact grammar: count(), avg(), percentiles, and by FIELD grouping.

cat orders.jsonl \
| smartpipe summarize 'count(), avg(total), p95(total) by region'
# → {"region":"EU","count":812,"avg_total":74.2,"p95_total":189.0}
# → {"region":"US","count":310,"avg_total":61.8,"p95_total":140.5}

# nested records: by-fields and aggregation args take field paths
cat events.jsonl \
| smartpipe summarize 'count(), avg(metrics.score) by user.plan'
# → {"user.plan":"pro","count":812,"avg_metrics.score":8.7}

Aggregation	Output name
count()	count
sum(f) avg(f) min(f) max(f)	sum_f, avg_f, …
p50(f) p90(f) p95(f) p99(f)	p95_f, …
dcount(f) (exact distinct count)	dcount_f

Semantics worth knowing: groups sort largest first; a record missing the by field groups under null, visibly; non-numeric values in numeric aggregations are skipped and counted on stderr (the run continues); a group with no numeric values reports null, not zero. Percentile aggregations hold each group's values in memory - everything else streams.

Time buckets

by bin(ts, 1h) groups by UTC time bucket (buckets: 1m 5m 15m 1h 6h 1d). Limitation: timestamps parse as ISO-8601 or epoch seconds/milliseconds only - anything else groups under null and any other format is a preprocessing job for jq/date. Date-only values (a {due date} field) bin as their UTC midnight, so count() by bin(due, 1d) just works. chart --by-time ts:1h draws the same buckets chronologically, zero-filled (empty buckets are shown, not dropped).

The natural pairs: map "…{label}" | summarize 'count() by label' for the numbers, | chart label for the picture.