---

name: cli-and-local-workflows description: "Use W&B CLI and local workflows safely: login, init, status, offline/online/disabled modes, syncing offline runs, cache cleanup, Docker wrapping, self-hosted verification, and local state inspection." disable-model-invocation: true metadata: disco-role: operating license: MIT

W&B CLI and Local Workflows

Use this sub-skill when an agent needs to configure a local W&B checkout or runtime environment without writing Python tracking code: authentication, workspace settings, local/offline run recovery, cache hygiene, Docker command wrapping, or self-hosted connectivity checks.

Start Here

1. Confirm the CLI is available with wandb --help or python -m wandb --help; wb is an equivalent console-script entry point when installed.
2. Inspect current local state before changing anything:
   • wandb status
   • python sub-skills/cli-and-local-workflows/scripts/inspect_wandb_local_state.py --root .
3. Choose the least-destructive mode/configuration command for the user's goal:
   • wandb init --project <project> --entity <entity> for workspace defaults.
   • wandb offline or wandb init --mode offline to keep future run data local.
   • wandb online to clear offline mode for future runs.
   • wandb disabled only when the user wants no W&B logging/syncing.
4. For offline recovery, run wandb sync first to summarize candidates; sync explicit run directories or use wandb sync --sync-all only after confirming destination entity, project, and server URL.

Reference Map

• CLI reference: command routing, command behavior, safe dry-run/help checks, sync and Docker workflows.
• Configuration: settings files, environment variables, mode precedence, offline run layout, and non-interactive CI setup.
• Troubleshooting: API keys, self-hosted host mismatch, sync path mistakes, destructive cleanup, local server caveats, and cache cleanup.
• Local-state inspector: read-only helper that reports CLI availability, selected WANDB_* variables with secrets redacted, workspace settings, and candidate run directories.

Boundaries

• Artifact commands such as wandb artifact ... and artifact cache details belong in the artifacts/registries sub-skill.
• Sweep and launch commands belong in the sweeps/launch sub-skill.
• Public API automation and Python-side run querying belong in the public API sub-skill.
• Do not include API keys in commands, logs, examples, or generated files; prefer environment injection by the caller or secret manager.