Storage Sinks

Sinks are the output layer. You select and configure them entirely from YAML, no code required. Each sink receives events as they flow through the pipeline, plus optional telemetry spans and usage records, and a failing sink is isolated so it cannot block the others.

To react to events in your own code instead of writing a full sink, use CallbackSink: it delegates to async callables you supply.

Implementations

Sink	YAML `type`	Output
`CallbackSink`	n/a (SDK only)	Delegates to user-provided async callables.
`ConsoleSink`	`console`	Pretty-printed events / assessments to the terminal.
`JsonlSink`	`jsonl`	Append-only JSONL files, optional size-based rotation.
`SqliteOutputSink`	`sqlite`	Local SQLite with WAL mode, schema migration, batch inserts.
`S3Sink`	`s3`	Cloud object storage with buffered upload and key formatting (requires `boto3`).
`ParquetSink`	n/a (SDK only)	One columnar Parquet file per session for analytics (built on `pyarrow`, included in the base install).
`OtelExporterSink`	`otel`	Export events / spans / usage as OTLP/HTTP JSON to an OpenTelemetry collector.
`WebhookSink`	`webhook`	POST events / assessments to a webhook URL.

Configuration

Sinks are a list on each pipeline, mix and match freely:

sinks:
  - type: sqlite
    path: ./events.db
  - type: jsonl
    path: ./output/events.jsonl
    rotate_size_mb: 100
  - type: s3
    bucket: my-traces
    prefix: agents/
    region: us-east-1

:::note Title updates Live titling emits append-only TitleUpdate records out-of-band. Sink.on_title_update defaults to a one-time warning rather than a silent drop, because titles are primary output; every in-repo sink overrides it. See Live Structuring. :::

Implementations​

Configuration​

Implementations

Configuration