# Fabric Harness Tooling Connector Spec

A tooling connector observes Fabric events, traces, metrics, or eval results without changing agent behavior. Produce a project-local module that uses public Fabric event, telemetry, or eval APIs.

## Requirements

1. Initialize the vendor SDK before Fabric runtime construction and dispose/flush it during graceful shutdown.
2. Integrate from `onEvent`, `TelemetryExporter`, `createOpenTelemetryObserver()`, `runEvalSuite()`, or another documented public boundary. Do not monkey-patch model providers or session internals.
3. Keep API keys in environment variables or a secret manager. Never add them to event attributes, errors, model context, or artifacts.
4. Disable prompt, output, reasoning, tool argument/result, attachment, and customer-content capture by default. Require an explicit redaction/retention policy to enable any of it.
5. Use stable low-cardinality job, agent, model, environment, and outcome attributes. Do not use session, submission, tenant, or user IDs as metric dimensions.
6. Bound queues, retries, payload size, and shutdown time so telemetry failure cannot block agent execution indefinitely.
7. Keep evaluation datasets and grading outputs separate from production session storage and apply access/retention controls.
8. Add tests for redaction, exporter failure, flush behavior, disabled-content defaults, and correlation fields.