Run Utilization Report

autocontext provides a descriptive report for parallel autoresearch runs. It does not gate runs.

Window semantics

window.started_at and window.completed_at are the earliest and latest known telemetry timestamps from run events and role usage rows. duration_seconds is null when no telemetry has timestamps.

Metrics

• mean_runner_utilization: active_runner_seconds / runner_capacity_seconds.
• runner_capacity_seconds: window.duration_seconds * max_parallel_branches.
• mean_token_utilization: model_active_seconds / runner_capacity_seconds.
• token_throughput_per_second: total_tokens / model_active_seconds.
• tokens_to_success: tokens from role usage rows at or before the first success event.
• verifier_idle_seconds: runner_capacity_seconds - verifier_active_seconds.
• eval_throughput_per_second: completed evaluations per window second.

Unknown telemetry is represented as null; token counts degrade to 0.

Shared schema: run-utilization-report.json. Parity fixture: run-utilization-report-parity-fixture.json.