-
Notifications
You must be signed in to change notification settings - Fork 1
Troubleshooting
- Is the repository and checkout correct?
- Is the Controller healthy?
- Is the stable ingress healthy?
- Is the active Gateway healthy?
- Does MCP initialize and list tools?
- Is the external tunnel reachable?
A single 502 does not identify which layer failed.
Check the durable Work, Job, Run, or managed process before retrying. Replaying a write can create duplicate effects even when the original HTTP request failed.
Compare local ingress health with external MCP health. A tunnel process can remain alive after all data-plane connections are lost. Confirm the active transport and registered HA connections, then verify the authenticated MCP flow.
Inspect immutable release contents, entrypoints, process liveness, and revision coherence. Do not keep restarting a release that cannot cold-start; preserve evidence and switch only through the Supervisor's bounded rollout or rollback path.
Compare the canonical source commit, manifest revision, active slot, installed service definition, Supervisor, Controller Daemon, and Gateway. Short hashes are not sufficient release identity.
A previous release must have exact path, revision, and manifest evidence recorded as known-good. Never weaken matching merely to force a rollback.
Detailed guides: