Build Week Judge Troubleshooting¶
Guardian Review says setup required¶
- Confirm
GUARDIANNODE_GUARDIAN_REVIEW_ENABLED=truein the backend service environment and restart the backend. - For the no-key judge path, select provider
mockand enable demo mode. - For live mode, confirm the key is in the server environment—not the browser— and that the explicit ZDR confirmation matches verified account controls.
codexdeliberately reports unavailable in this candidate; use mock or the advanced direct Responses API path.
Missing or invalid API key¶
The incident remains local and usable. Guardian Review stores a sanitized failure and does not delete the alert. Correct the backend secret and request a fresh assessment. Never paste keys into screenshots, issues, logs, or demo recordings.
Timeout, rate limit, or network outage¶
The request is durable. Leave the page or refresh; history continues to show queued/running/failed state. Only eligible transient failures are retried, with bounded attempts. Local detection and the dashboard remain available.
Duplicate click or browser refresh¶
Consent is bound to the preview digest and idempotency identity. A duplicate submission returns the existing review. A deliberate fresh assessment must be requested explicitly.
Backend restart or database lock¶
Restart the service after releasing the lock. A queued/running review is
recoverable from local state; the original incident is independent of the
external result. Check free disk space and use /api/health/ready before retrying.
Agent disconnected¶
Existing incidents and demo scenarios remain reviewable. Check the Devices page, Windows service status, configured backend URL, trusted firewall/VPN path, and pairing state. Do not expose the backend directly to the public internet.
Corrupt incident or invalid model output¶
Guardian Review fails safely with a controlled error and preserves the incident. Malformed output is never rendered as an assessment and is not blindly retried. Use the local incident evidence and, for demo recovery, reset and trigger a new synthetic scenario.
Port conflict or stale process¶
The backend defaults to loopback port 8787. Stop the old GuardianNode service or change the configured port consistently. Do not terminate unrelated processes. Verify the dashboard and agent use the same trusted backend address.
Uninstall and reinstall¶
Export a portable encryption-key backup and database backup first. Record whether the installer preserves or removes the data directory. A reinstall must not silently replace keys for existing encrypted evidence. See the clean-install checklist before release promotion.