Skip to content

Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure

TypeScript

fix(branch-profile): refuse a profile context that is not this local Minikube

Alfredo Lopez467ab776c528
diff --git a/scripts/minikube-profiles/branch-profile.sh b/scripts/minikube-profiles/branch-profile.sh
index fc6b51fcc..130f6656b 100755
--- a/scripts/minikube-profiles/branch-profile.sh
+++ b/scripts/minikube-profiles/branch-profile.sh
@@ -28,6 +28,7 @@ STATE_TEMP_PROFILE=""
STATE_TEMP_PORTS=""
STATE_NEW_CACHE_DIR=false
PORT_FORWARD_OWNER_LOADED=false
+CONTEXT_IDENTITY_LOADED=false
die() {
printf 'ERROR: %s\n' "$*" >&2
@@ -102,10 +103,16 @@ normalize_cache_root() {
fi
}
-# Refuses the kube contexts shared across sessions or clusters: the Clerum dev
-# and production contexts, the default minikube and Docker Desktop contexts,
-# every GKE context (gke_<project>_<zone>_<cluster>, covered by *gke*) and the
-# protected names profile-owner.sh also refuses. Matching is case-insensitive.
+# Refuses by name the kube contexts shared across sessions or clusters: the
+# Clerum dev and production contexts, the default minikube and Docker Desktop
+# contexts, GKE contexts under their generated gke_<project>_<zone>_<cluster>
+# name (covered by *gke*) and the protected names profile-owner.sh also
+# refuses. Matching is case-insensitive. A name list cannot recognise a renamed
+# remote context, so this is only the first barrier: the profile must also be
+# owned (profile-owner.sh) and in the clerum-* namespace, and before any
+# minikube -p or kubectl --context call its kube context must point at a local
+# API server (require_local_context_endpoint) and, once the cluster answers,
+# identify this Minikube profile (require_profile_cluster_identity).
refuse_shared_profile_name() {
local name="$1" normalized
normalized="$(printf '%s' "${name}" | LC_ALL=C tr '[:upper:]' '[:lower:]')"
@@ -520,6 +527,7 @@ init_profile() {
normalize_cache_root
PROFILE_OWNER_SCRIPT="${REPO_DIR}/scripts/minikube/profile-owner.sh"
PORT_FORWARD_OWNER_SCRIPT="${REPO_DIR}/scripts/minikube/port-forward-owner.sh"
+ CONTEXT_IDENTITY_SCRIPT="${REPO_DIR}/scripts/minikube/context-identity.sh"
DOCKER_CLI_ENV_SCRIPT="${REPO_DIR}/scripts/minikube/docker-cli-env.sh"
DEADLINE_RUNNER="${REPO_DIR}/scripts/minikube/run-with-deadline.mjs"
[[ -f "${PROFILE_OWNER_SCRIPT}" && -x "${PROFILE_OWNER_SCRIPT}" && ! -L "${PROFILE_OWNER_SCRIPT}" ]] ||
@@ -636,6 +644,69 @@ load_port_forward_owner() {
PORT_FORWARD_OWNER_LOADED=true
}
+load_context_identity() {
+ [[ "${CONTEXT_IDENTITY_LOADED}" == true ]] && return 0
+ require_command kubectl python3
+ if [[ ! -f "${CONTEXT_IDENTITY_SCRIPT}" || ! -r "${CONTEXT_IDENTITY_SCRIPT}" || -L "${CONTEXT_IDENTITY_SCRIPT}" ]]; then
+ die "CONTEXT_IDENTITY_UNAVAILABLE: refusing to address a cluster without ${CONTEXT_IDENTITY_SCRIPT}"
+ fi
+
+ # The predicates the T2 lane applies in t2_context_check and
+ # t2_profile_context_identity_check (scripts/minikube/t2-common.sh).
+ # shellcheck source=/dev/null
+ . "${CONTEXT_IDENTITY_SCRIPT}"
+ local function_name
+ for function_name in kube_endpoint_host kube_endpoint_is_local minikube_nodes_identify_profile; do
+ declare -F "${function_name}" >/dev/null ||
+ die "CONTEXT_IDENTITY_API_INVALID: missing ${function_name} in ${CONTEXT_IDENTITY_SCRIPT}"
+ done
+ CONTEXT_IDENTITY_LOADED=true
+}
+
+# kube_context_server <context>: prints the API server of the kubeconfig
+# context named <context>, or nothing when there is no such context. It reads
+# the kubeconfig only and never contacts a cluster.
+kube_context_server() {
+ local context="$1" contexts server
+ contexts="$(kubectl config get-contexts -o name)" ||
+ die 'BRANCH_PROFILE_REMOTE_CONTEXT: unable to read the kubeconfig contexts'
+ grep -Fqx -- "${context}" <<<"${contexts}" || return 0
+ server="$(kubectl config view --raw --minify "--context=${context}" -o 'jsonpath={.clusters[0].cluster.server}')" ||
+ die "BRANCH_PROFILE_REMOTE_CONTEXT: unable to read the API server of kube context ${context}"
+ [[ -n "${server}" ]] ||
+ die "BRANCH_PROFILE_REMOTE_CONTEXT: kube context ${context} has no API server"
+ printf '%s' "${server}"
+}
+
+# minikube -p and kubectl --context address a cluster by context name alone, and
+# `minikube -p <p> delete` on a profile minikube does not know removes the
+# kubeconfig context of that name. So a context named after the profile must
+# point at this machine before either runs. A missing context is allowed:
+# minikube start creates it, and kubectl cannot reach a cluster through it.
+require_local_context_endpoint() {
+ load_context_identity
+ local server
+ server="$(kube_context_server "${PROFILE}")" || exit 1
+ [[ -n "${server}" ]] || return 0
+ kube_endpoint_is_local "${server}" ||
+ die "BRANCH_PROFILE_REMOTE_CONTEXT: kube context ${PROFILE} points at a non-local API server ($(kube_endpoint_host "${server}")); refusing to run minikube or kubectl against it. That context may belong to a cluster in use: stop and ask before renaming or removing it"
+}
+
+# A local endpoint is necessary but not sufficient: the context may name
+# another local Minikube profile, or a LAN cluster on a private address. The
+# cluster must have a node labelled minikube.k8s.io/name=<profile> at the IP
+# `minikube -p <profile> ip` reports.
+require_profile_cluster_identity() {
+ load_context_identity
+ local expected_ip nodes_json
+ expected_ip="$(run_bounded minikube-ip "${MINIKUBE_STATUS_TIMEOUT_SECONDS}" minikube -p "${PROFILE}" ip)" ||
+ die "BRANCH_PROFILE_CONTEXT_IDENTITY: minikube did not report an IP for profile ${PROFILE}"
+ nodes_json="$(kubectl "--context=${PROFILE}" "--request-timeout=${KUBECTL_REQUEST_TIMEOUT}" get nodes -o json)" ||
+ die "BRANCH_PROFILE_CONTEXT_IDENTITY: unable to read the nodes of kube context ${PROFILE}"
+ minikube_nodes_identify_profile "${nodes_json}" "${PROFILE}" "${expected_ip}" ||
+ die "BRANCH_PROFILE_CONTEXT_IDENTITY: kube context ${PROFILE} does not identify Minikube profile ${PROFILE} at ${expected_ip:-<no IP>}; refusing to use it"
+}
+
probe_service() {
local namespace="$1" service="$2" output
if ! output="$(kubectl "--context=${PROFILE}" "--request-timeout=${KUBECTL_REQUEST_TIMEOUT}" \
@@ -927,6 +998,7 @@ ensure_shims() {
cmd_start() {
require_command git minikube kubectl helm shasum
+ require_local_context_endpoint
check_docker_ready
check_all_ports_free
persist_state
@@ -964,12 +1036,14 @@ cmd_start() {
fi
run_bounded minikube-status "${MINIKUBE_STATUS_TIMEOUT_SECONDS}" \
minikube -p "${PROFILE}" status
+ require_profile_cluster_identity
kubectl "--context=${PROFILE}" "--request-timeout=${KUBECTL_REQUEST_TIMEOUT}" cluster-info
kubectl "--context=${PROFILE}" "--request-timeout=${KUBECTL_REQUEST_TIMEOUT}" get nodes
}
cmd_status() {
require_existing_profile
+ require_local_context_endpoint
printf 'profile: %s\n\n' "${PROFILE}"
# `minikube status` exits non-zero for a stopped cluster. Report that state
# and the reachability line, then return the status code so a stopped
@@ -979,6 +1053,7 @@ cmd_status() {
minikube -p "${PROFILE}" status || minikube_status=$?
printf '\n'
if cluster_reachable; then
+ require_profile_cluster_identity
kubectl "--context=${PROFILE}" "--request-timeout=${KUBECTL_REQUEST_TIMEOUT}" get nodes
printf '\n'
kubectl "--context=${PROFILE}" "--request-timeout=${KUBECTL_REQUEST_TIMEOUT}" get deploy -A
@@ -995,10 +1070,12 @@ cmd_pf() {
require_command git kubectl curl shasum
require_existing_profile
load_port_forward_owner
+ require_local_context_endpoint
if ! cluster_reachable; then
printf 'ERROR: cluster not reachable for context %s\n' "${PROFILE}" >&2
exit 1
fi
+ require_profile_cluster_identity
persist_state
# Start every forward before judging the set: stopping at the first failure
@@ -1034,10 +1111,12 @@ cmd_pf() {
cmd_health() {
require_existing_profile
load_port_forward_owner
+ require_local_context_endpoint
if ! cluster_reachable; then
printf 'ERROR: cluster not reachable for context %s\n' "${PROFILE}" >&2
exit 1
fi
+ require_profile_cluster_identity
# Probe every service so one dead forward does not hide the state of the
# rest; the command still fails when any probe failed.
@@ -1214,6 +1293,7 @@ report_kept_pf_records() {
cmd_stop() {
require_existing_profile
+ require_local_context_endpoint
# Clear the verified port-forward records first: a stopped cluster leaves
# records naming dead kubectl processes, and the next T2 preflight refuses
# them with PORT_FORWARD_CONFLICT. A record that cannot be verified is kept
@@ -1237,6 +1317,7 @@ cmd_setup() {
require_command git minikube kubectl helm shasum perl
require_existing_profile
require_profile_confirmation setup
+ require_local_context_endpoint
check_docker_ready
persist_state
ensure_shims
@@ -1258,6 +1339,7 @@ cmd_delete() {
printf 'ERROR: refusing delete. Re-run with CONFIRM_DELETE=%s\n' "${PROFILE}" >&2
exit 1
fi
+ require_local_context_endpoint
# The registry (pids/*.pid) outlives the cluster; clear the port-forward
# records so the deleted profile does not leave PORT_FORWARD_CONFLICT behind.
# As in cmd_stop, a kept record is reported and fails the command, but does