Diagnostics

Read-only local and cluster diagnostics for stuck runner, placement, resource, and process inspection.

📄️ Get local node diagnostics

Return a read-only diagnostic bundle for this API node, including runtime identity, placement analysis, live runner-supervisor state, local resources, and relevant OS processes.

📄️ Get local telemetry-plane diagnostics

Return aggregate bounded-admission, coalescing, drop, queue, and publish metrics for this node's isolated telemetry transport.

📄️ Get this node's performance envelopes

Return the observe-only throughput-and-latency-versus-concurrency curve this API node has measured for each (hardware class x model x engine x quantization) it has served, including a simple knee estimate. Adaptive-concurrency Phase 0: data only, no behavior change.

📄️ Get performance envelopes across the cluster

Fan out to every reachable cluster member and return each one's performance-envelope report. Unreachable members appear as explicit failures rather than vanishing.

📄️ Request direct local runner cancellation

Request cooperative cancellation for one live task on one local runner supervisor. This is a best-effort control path intended for diagnostics and may not stop a runner wedged in native code.

📄️ Capture local node diagnostic bundle

Collect an on-demand diagnostic bundle for this node, optionally focused on one live runner or task. The bundle includes current node diagnostics, runner flight-recorder entries, the latest MLX memory snapshot, and best-effort macOS process samples.

📄️ Get cluster diagnostics

Fan out to reachable peer APIs and return read-only diagnostic bundles for the local node and peers. Unreachable peers are reported as partial failures instead of failing the whole request. Peer bundles tolerate additive diagnostic fields, and aggregate plus per-node versionStatus values expose mixed-build rollouts.

📄️ Get cross-rank runner timeline

Stitch every reachable node's runner-supervisor diagnostics into one cross-rank chronological view. Returns a per-runner synopsis (rank, current phase, seconds-in-phase, MLX memory) and every flight-recorder entry across all ranks merged and sorted by wall-clock timestamp. The intended use is debugging distributed deadlocks where the rank-disagreement signature is invisible from any single node's local diagnostics.