Watchman answers the questions that arrive first in a GPU incident. Is the node healthy? Who owns VRAM? Is the serving runtime moving work? What evidence would make the next rollout safe enough to continue?

It is a Rust library with a thin watchman CLI. Its read-only hardware collection, explicit active canaries, bounded saturation experiments, capacity planner, and machine-readable reports are kept separate on purpose.

The baseline collection is evidence, not an opaque score. Watchman inventories NVIDIA GPUs and joins ownership, process, container/Kubernetes attribution, health signals, topology, and optional inference-endpoint telemetry.

It emits a human terminal view plus stable JSON and NDJSON contracts. Optional sources are recorded as ok, partial, unavailable, or skipped so a missing capability does not become a silently empty result.

Health and capacity are different questions. A snapshot helps triage current hardware and runtime state. A canary tests one real OpenAI-compatible generation path. A saturation ladder characterizes explicit concurrency points. Capacity planning models a stated topology and assumptions without pretending to measure live allocation.

Read health and collection, inference + capacity, and rollouts + deployment before using a result as a deployment gate.

The hosted installer publishes the watchman binary for supported macOS and Linux architectures and verifies release checksums:

12curl -fsSL https://tools.yshubham.com/watchman/install.sh | sh
watchman version

The public source repository, forge mirror, and agent skill guide are the right starting points for a local installation or automation integration.