| name | wandb-onboarding |
| description | Use when signing up or 가입 for W&B Cloud, configuring a personal API key, checking authentication, or creating a first experiment-tracking run or 실험 추적 project. |
W&B Onboarding
Overview
Guide signup with browser assistance, hand sensitive identity steps to the user, then verify the SDK offline before any upload. Read the current official workflow for URLs and commands. For official examples and autoresearch analysis, also read the W&B official ecosystem.
Preflight
Classify the target as signup, SDK authentication, offline tutorial, or online run. Confirm browser and uv availability without inspecting credentials, secret files, cookies, browser storage, or environment-variable values.
Safety Contract
Never ask the user to paste an API key, password, MFA code, or OAuth token into chat. Do not read cookies, browser storage, password managers, ~/.netrc, environment variable values, or raw credential files.
The user must complete identity-provider selection, credentials, OAuth consent, CAPTCHA, MFA, email verification, legal acceptance, final signup, and API-key copy. Use an interactive prompt for the key; never place it in a command argument, source file, generated asset, or repository.
Workflow
- Open the official W&B signup page with the available Browser skill. Inspect the current visible options; do not guess selectors when the page changes. If browser control is unavailable, give the exact manual link and checkpoint list.
- Recommend the Cloud-hosted Free plan and personal entity for an individual tutorial. Present any discovered organization/team separately and let the user choose.
- Hand off the sensitive signup steps. After the user returns, verify only the visible logged-in state.
- Guide the user to User Settings → API Keys. The user creates and stores the key privately.
- For existing credentials, run
wandb login --verify. For a new key, explain the local credential write, obtain explicit confirmation, then run wandb login --relogin --cloud --verify. The user enters the key only in the interactive prompt.
- Copy
assets/wandb-quickstart.py into a disposable uv workspace. Run it first with WANDB_MODE=offline and inspect the completed local run.
- Before any online run or offline-run sync, show the target entity, project, visibility, metrics/config fields,
group, job_type, code/Git/console/system capture settings, every offline run directory, and files or data that would upload. Set or verify project visibility in the official UI, bind the approved destination through WANDB_ENTITY and WANDB_PROJECT, then obtain explicit confirmation.
- Prefer syncing the already verified offline run with
wandb sync --entity ... --project ... --skip-console --no-sync-tensorboard <run-directory>. Verify the authoritative run URL and visible fields after upload. Report what was uploaded and how to remove or change access if needed.
Autoresearch integration
Keep VESSL as the sole compute scheduler and W&B as the local-first observability layer. Do not modify the official benchmark to import W&B and do not install a W&B key in VESSL. After each VESSL Job finishes, fetch vesslctl job logs locally and perform local post-processing with the auto-research recorder in WANDB_MODE=offline; create an offline run with group=<run_tag>, job_type=autoresearch-trial, and val_bpb summarized by its minimum.
The upload allowlist contains only the approved run/trial, Git/cookbook, VESSL Job, exact A100 resource/model/count, cache/evidence hashes, metric, time, VRAM, parameter, and status fields. Disable console, code, Git, requirements, machine metrics/info, automatic Job creation, datasets, checkpoints, and artifacts. W&B 0.28.0 still writes internal SDK metadata to an offline run: the recorder uses a sanitized host, but SDK/Python/platform versions, timestamps/runtime, and SDK telemetry records remain. Sync only after the user confirms entity, project, visibility, allowlist, exact offline directories, and that metadata disclosure.
W&B ARIA is an optional official analysis surface after approved sync. It may inspect results and propose a falsifiable single-variable probe; do not use ARIA Launch, W&B Launch, Sweeps, or Jobs to schedule this VESSL campaign.
Completion Levels
| Level | Evidence |
|---|
| Account ready | User confirms signup and visible logged-in state |
| SDK ready | wandb login --verify succeeds without exposing the key |
| Local tutorial ready | Offline run finishes and creates a local W&B run directory |
| Online tutorial ready | User-approved offline run is synced to the intended entity/project and visibility |
Stop Conditions
Stop before signup submission, team join/create, trial or payment selection, API-key creation/revocation, credential storage, or online upload unless the user has authorized that exact action. If authentication or upload partially succeeds, inspect status before retrying to avoid duplicate keys or runs.
Verification
- Verify authentication with
wandb login --verify without printing the key.
- For offline work, confirm the synthetic run completes and creates a local run directory.
- For online work, verify the authoritative W&B URL, entity, project, visibility, config, metrics, group/job type, capture settings, and every uploaded file or data field.
- Report what was uploaded and the remaining privacy, access-control, deletion, or cleanup work.
Output
Return the completion level, verified evidence, user handoff, exact planned or completed side effect, authoritative run URL when applicable, and remaining privacy or cleanup work.
Next Steps
- Offline complete: review the upload card and stop for confirmation.
- Online approved: sync the reviewed offline run and verify it.
- Online declined: preserve the local offline artifact and stop.
Common Mistakes
- Key in a shell argument → use
wandb login --verify and its interactive prompt.
- Immediate online test → prove the script with
WANDB_MODE=offline first.
- Project name without entity/visibility → show all three before confirmation.
- “No dataset” assumed safe → config, metrics, logs, code, and system metadata can still upload.
- Putting W&B inside the VESSL benchmark → keep benchmark execution unchanged and record fetched logs locally.
- Using Launch or Sweeps for this campaign → VESSL is the only scheduler; keep the experiment loop bounded and sequential.
