Ingestion Pipeline

The authoritative ingestion path is continueConversation(...) in packages/db/src/index.ts.

What it persists

• conversations
• messages
• turns
• inference_requests
• inference_events

All data lands in one SQLite database.

Continue Conversation sequence

For POST /conversations/:id/messages (and /stream), the Runtime API calls continueConversation(...).

1. Confirm Conversation exists.
2. Insert user Message.
3. If missing, derive and persist Conversation title from first user Message.
4. Insert pending Turn.
5. Load Conversation message history as inference input.
6. Insert pending Inference Request with inputPreview.
7. Stream model output and persist canonical lifecycle events once each with ordered sequenceNumber.
8. Backfill missing response_start, usage, request_end.
9. On success:
   • finalize Inference Request (completed, outputPreview, optional raw JSON)
   • insert assistant Message
   • mark Turn completed with committedAssistantMessageId
10. On failure:

• mark Inference Request and Turn as failed
• rethrow error to Runtime API

Telemetry persistence policy

Defaults are intentionally safe:

• inputPreview / outputPreview are truncated to 200 chars by default
• optional preview redaction hook can transform or drop preview content
• raw request/response JSON capture is off by default
• raw capture is enabled only with telemetry.captureRawPayloads: true

When raw capture is enabled but provider hooks are absent, fallback request/response payloads are persisted.

Inspection outputs

packages/db exposes:

• listInferenceRequests()
• getInferenceRequestInspection(id)
• getInferenceRequestMetrics(id)

Inspection includes parsed canonical events plus summary fields:

• eventCount
• firstTokenLatencyMs
• totalDurationMs
• usage

Note: current persisted canonical events are response_start, first_token, usage, request_end. text_delta is used to assemble assistant content but is not persisted as an inference event record.