The Watchman skill keeps inference work evidence-first. It distinguishes a read-only hardware snapshot from an active canary, a bounded saturation experiment, a local artifact inspection, and a capacity estimate.

The registry distributes Watchman as watchman. Install the checksum- verified binary through the hosted path, then add the agent skill through the Skills CLI.

12curl -fsSL https://tools.yshubham.com/watchman/install.sh | sh
npx skills add bas3line/rool-repo --skill watchman --agent '*' --global --yes

Keep the tool local to the evidence source. Run it where the GPUs, process table, endpoint reachability, and secret-file paths are actually available; do not turn a support report into a remote credential collection channel.

Begin with read-only evidence. Use snapshot, ps, top, or doctor to understand hardware and telemetry coverage. Invoke canary only when a real synthetic generation request is appropriate; invoke benchmark saturation only when intentionally measuring an explicit fixed-concurrency ladder.

Capacity is a model, not live allocation telemetry. Use it to state a topology and conservative bounds, then retain its caveats in the rollout record. Read health and collection and inference and capacity before choosing a command.

Save the report that justified the deployment choice. Offline benchmark and canary comparison use complete, compatible reports; missing or incompatible evidence remains not-evaluable instead of turning into a green gate.

For service deployment, history rotation, remote-listen safety, and rollback thinking, use the Watchman rollout guide.