| name | dgx-station-diagnose |
| description | Run and interpret the complete read-only dgx-assist diagnostic suite for NVIDIA DGX Station GB300, correlate findings with pinned NVIDIA playbooks, export a redacted support bundle, and apply one separately approved allowlisted fix. Use when the user reports a Station, CUDA, GPU health, coherency, vsloshd, Docker, CDI, MIG, cache, port, or owned inference-service failure. |
DGX Station diagnostics
Diagnose first. Do not mutate as part of diagnosis.
Workflow
- Run
scripts/dgx-assist diagnose run --json.
- Report the detected compatibility profile, then findings in severity order with their stable IDs and evidence. Preserve
unknown states. On Software 1.0, do not reinterpret intentionally skipped Software 2.0 service checks as faults.
- Search the pinned playbooks for each high or critical finding; cite the relevant URL, heading, lines, and commit.
- If no passage overlaps, say so and avoid inventing a platform fix.
- Offer
diagnose bundle --report-id "<id>" when escalation is appropriate.
- Offer at most one automatic fix at a time, and only when
fix_id is present.
- Preview with
diagnose fix --report-id "<id>" --finding "<id>" --dry-run.
- Explain exact actions, impact, privilege, and reboot state. Obtain explicit approval.
- Repeat with
--yes only after approval and report the action receipt.
Safety requirements
- Keep
diagnose run read-only.
- Never install packages, rewrite Docker configuration, change power caps or driver parameters, kill workloads, or modify MIG through a diagnostic fix.
- Never stop a service without current
dgx-assist ownership evidence.
- Never use Fabric Manager as a routine Station check or remediation.
- Re-run diagnostics when finding evidence is stale.
- Never reveal secrets or unredacted home paths in a bundle.
- Do not execute an unregistered remediation.
- Treat
--yes only as approval already obtained.
Read references/findings.md before proposing a fix or support bundle. Read references/bringup.md when the problem concerns physical deployment, BMC or firmware verification, driver bring-up, power braking, or support escalation.