swarmr-kube

Kubernetes incident response team for swarmr. Diagnoses a live cluster and proves the root cause.

Read-only twice over: the credential grants only get/list/watch, and filesystem permissions deny writes outside evidence/. Nothing about the cluster is hardcoded — the team profiles it at startup and injects what it found into every prompt.

Install

Core and this team must land in the same environment.

uv tool install "swarmr[kube]" --with-executables-from swarmr-kube

That exposes teams, teams-mcp and incident-credentials on your PATH. --with-executables-from is required for the third. From a checkout: uv tool install ../swarmr --with ../swarmr-kube --with-executables-from swarmr-kube.

teams --target k8s_incident
teams k8s_incident "payments.demo.local returns 502, namespace demo"

Over MCP the tool is start_k8s_incident.

Credentials

This team needs a read-only credential. It refuses your ambient kubeconfig, which is usually cluster-admin.

incident-credentials --list
incident-credentials --context archdev
incident-credentials --print-manifest

That creates the incident-reader ServiceAccount and ClusterRole, mints an 8h token, writes .incident-reader.<context>.kubeconfig at mode 600, then asks the API server to confirm the credential can list pods and cannot delete them.

Each cluster gets its own credential file. Selection is explicit: INCIDENT_KUBECONFIG (exact path), else INCIDENT_CONTEXT (context name), else the single minted credential. Several credentials with no choice expressed is an error rather than a guess.

The 8h token refreshes itself: every run reads the expiry before opening a connection and re-mints when under five minutes remain. Only files this team minted are ever rewritten. INCIDENT_NO_REFRESH=1 turns it off.

Design notes and internals: CLAUDE.md.