Troubleshooting

Where the logs are

CrewForge writes to its own output channel. Open View > Output and choose CrewForge from the drop-down. It holds page errors, Lint failures, and the output of the helm and kubectl commands behind a deploy, redeploy, or undeploy. A failed deploy opens it for you. A failed kmctl create shows kmctl’s own message in a dialog.

For a bug report, run CrewForge: Show Connection Info and use its Copy button. It gives CrewForge’s version, the Kubemoot operator’s version, the Kubernetes server version, and the context and server address.

A dashboard says “Reading…” or “Still reading”

A dashboard shows Reading… while it loads, and Still reading from context X… above the page when a read takes more than a moment. The trees show the same line above their items. The cluster is answering slowly; the page updates when it does.

If the cluster does not answer, the page gives up after 45 seconds with No answer from context X in 45 s. Is the cluster running? Refresh reads again. A request to the API server that gets no answer for 20 seconds fails the same way in the trees. Check that the cluster is up and that you can reach it (network, VPN, kubectl get nodes from a terminal), then press Refresh.

If a page never leaves Reading… and then says This page did not start, its script did not load. Run Developer: Reload Window. If the page stays empty, reinstall the extension and report it with the output of Show Connection Info.

The cluster is unreachable

CrewForge names the context and server it tried and what failed: refused, timed out, host name not found, connection cut, an untrusted certificate, rejected credentials, a refused request, or a login plugin that failed. The full list, with what each means, is in Install and connect.

Every one of those messages offers Select Kubernetes Context. Common fixes:

• The cluster is not running, or a kubectl port-forward or tunnel in front of it is down. Start it and press Refresh.
• The kubeconfig points at another cluster than the one you expect. Check the status bar’s context, or run CrewForge: Select Kubeconfig File.
• The token expired. Log in again, or get a fresh kubeconfig.
• Your account is not allowed to read crews. See what your account needs, and set crewforge.namespaces if you may read only some namespaces.

Deployed Crews is empty

The context is reachable but shows no crews. Check in this order:

1. The status bar’s context is the one you mean.
2. The Kubemoot operator is installed on that cluster (kubectl get crds | grep kubemoot).
3. crewforge.namespaces is empty, or includes the namespace you are looking for.
4. Your account may list crews.kubemoot.ai there.

Crew Sources is empty

CrewForge looks in the folders open in the workspace for a Helm chart whose templates declare a Kubemoot Crew, and for a folder of plain manifests that includes one. Open the folder that holds the crew, or create one with New Kubemoot Crew Here. A source whose render fails is listed with the error as an item; click it to open the file and line.

A tool is missing

Create Crew and Lint check the tool before they ask you anything, and each reports a missing tool differently:

• Create Crew shows a dialog that stays until you dismiss it.
• Lint shows a warning notification in the corner of the window, not a dialog; it can go unseen if you are looking elsewhere. Lint you run yourself shows it every time. Lint that runs when you save shows it once per session. The same text is written to the CrewForge output channel, which keeps it.

A schema mismatch

CrewForge checks your Kubemoot objects against the schemas your cluster serves. When your source and your cluster’s operator are different versions, you may see:

• Lint: “spec has no field X; the API server would drop it.” The field is not in the CRD your operator serves. It is a typo, or the source is newer than the operator. Fix the name, or upgrade the operator.
• Lint: “The Kubemoot schema check was skipped.” CrewForge could not read the schemas from the cluster, so it ran helm lint and the render only. Fix the connection and lint again.
• Compare with Live shows a difference you did not make. The API server drops fields its CRD does not define, so the live object lacks them. Treat it as the same mismatch.
• Pause, Resume, and Stop are missing on the fitness dashboard. Your operator’s CrewFitnessSuite CRD does not have spec.suspend and spec.cancel. Upgrade the operator.

Check the operator’s version with Show Connection Info. For the fields a resource supports, see the reference pages.

Deploy and redeploy

• A deploy is refused because Flux or a bundle already owns the crew. Two channels must not fight over the same objects. Change the crew through the channel it came by, or undeploy it first. A Flux-managed crew changes only through git.
• CrewForge warns that the crew came from another source or developer. The Crew’s recorded source or owner differs from yours. Confirm only if you mean to replace it.
• Redeploy goes to the wrong namespace. Choose Change the Namespace Redeploy Uses… from the status bar menu, or deploy again with Deploy to Namespace….
• The deploy finished but the crew never becomes ready. Cancel the progress notification and open the crew dashboard’s Overview for its phase and conditions. A new crew also needs a model provider that the scheduler can bind its Models to; see Models and scheduling.

The chat will not send

When the send button is missing and a line above the input gives a reason, the crew cannot take a question yet:

• The crew is not ready: phase X. Wait for the crew, or open its dashboard to see which agent is not ready.
• Can’t reach the crew’s discussion gateway. The crew’s <crew>-discussion Service has no ready address. Check the crew’s pods. CrewForge reads the crew again every 15 seconds while the chat is visible and after each turn.
• A reason that names the cluster is a connection problem; see above.

If a turn ends with a red card or a notice under the answer, an agent failed, could not run, or had not finished. Under the answer, N agents took part shows each agent’s last word. When every GPU is busy, an agent waits for one with room for its model; see Models and scheduling.