Typesafe Batch

Bounded batch triage for @swamp/typesafe-ai. Adds one method, triage_batch, that evaluates many compact {id, state} items against one shared set of typed questions (noul, choice, score) while holding the model lock once. It writes a single triage-batch-<name> resource with each item's answers, a SHA-256 fingerprint of the input, aggregate token usage, and one failure record per item that could not be evaluated. Raw item state and question text are never stored.

Authentication

Uses the @swamp/typesafe-ai global arguments: apiKey (prefer a vault reference), falling back to the TYPESAFE_API_KEY environment variable.

Quick Start

swamp extension pull @webframp/typesafe-batch
swamp model create @swamp/typesafe-ai jev
swamp model method run jev triage_batch --input-file batch.yaml
swamp data get jev triage-batch-latest --json

2026.09.26.1

Added: triage_batch, an extension method on @swamp/typesafe-ai, moved here from @webframp/operator-briefing 2026.09.25.1 so packages other than the briefing can depend on it. The method name, arguments, and the triageBatch resource (triage-batch-<name>) are the same, so existing workflows and stored batches keep working. The only behavior change is the request User-Agent, now webframp-typesafe-batch/1.