atlas-iac/services/hermes/execution-worker-statefulset.yaml

296 lines
13 KiB
YAML

apiVersion: apps/v1
kind: StatefulSet
metadata:
name: hermes-execution-worker
namespace: hermes
labels:
app: hermes-execution-worker
spec:
serviceName: hermes-execution-worker
replicas: 3
podManagementPolicy: Parallel
revisionHistoryLimit: 2
selector:
matchLabels:
app: hermes-execution-worker
updateStrategy:
type: RollingUpdate
rollingUpdate:
partition: 0
template:
metadata:
labels:
app: hermes-execution-worker
app.kubernetes.io/name: hermes-execution-worker
app.kubernetes.io/part-of: hermes
annotations:
ai.bstein.dev/role: fenced-execution-only
ai.bstein.dev/scm-boundary: mediated-pr14-broker-with-completion-gates
ai.bstein.dev/model-policy: Switchyard effort selection on the Claude subscription lane
ai.bstein.dev/storage: durable workspace and ordinal-private provider session state
ai.bstein.dev/config-rev: "20260822-claude-setup-token-v2"
vault.hashicorp.com/agent-inject: "true"
vault.hashicorp.com/role: hermes-execution-worker
vault.hashicorp.com/agent-inject-containers: stage-worker-access
vault.hashicorp.com/agent-service-account-token-volume-name: vault-auth-token
vault.hashicorp.com/agent-inject-secret-claude-oauth-token: kv/data/atlas/hermes/agent-tokens
vault.hashicorp.com/agent-inject-perms-claude-oauth-token: "0600"
vault.hashicorp.com/agent-inject-template-claude-oauth-token: |
{{- with secret "kv/data/atlas/hermes/agent-tokens" -}}
{{ .Data.data.claude_oauth_token }}
{{- end }}
vault.hashicorp.com/agent-pre-populate-only: "true"
vault.hashicorp.com/agent-init-first: "true"
vault.hashicorp.com/agent-requests-cpu: 2m
vault.hashicorp.com/agent-requests-mem: 16Mi
vault.hashicorp.com/agent-limits-cpu: 100m
vault.hashicorp.com/agent-limits-mem: 128Mi
spec:
serviceAccountName: hermes-execution-worker
priorityClassName: scavenger
automountServiceAccountToken: false
enableServiceLinks: false
terminationGracePeriodSeconds: 30
securityContext:
fsGroup: 10000
fsGroupChangePolicy: OnRootMismatch
seccompProfile:
type: RuntimeDefault
affinity:
nodeAffinity:
requiredDuringSchedulingIgnoredDuringExecution:
nodeSelectorTerms:
- matchExpressions:
- {key: kubernetes.io/arch, operator: In, values: [arm64]}
- {key: node-role.kubernetes.io/worker, operator: In, values: ["true"]}
- {key: kubernetes.io/hostname, operator: NotIn, values: [titan-04, titan-13, titan-14, titan-17, titan-18, titan-19, titan-22, titan-24]}
preferredDuringSchedulingIgnoredDuringExecution:
- weight: 100
preference:
matchExpressions:
- {key: node-role.kubernetes.io/accelerator, operator: Exists}
- weight: 50
preference:
matchExpressions:
- {key: hardware, operator: In, values: [rpi5]}
podAntiAffinity:
requiredDuringSchedulingIgnoredDuringExecution:
- labelSelector:
matchLabels:
app: hermes-execution-worker
topologyKey: kubernetes.io/hostname
topologySpreadConstraints:
- maxSkew: 1
topologyKey: kubernetes.io/hostname
whenUnsatisfiable: DoNotSchedule
labelSelector:
matchLabels:
app: hermes-execution-worker
initContainers:
- name: stage-worker-access
image: registry.bstein.dev/bstein/hermes-agent@sha256:81970563e542f0720773e72297810b3a844b83e381e278f25c0916c78d930107
imagePullPolicy: IfNotPresent
command: [/opt/hermes/.venv/bin/python, /opt/coordinator/stage_runtime_access.py, execution-worker]
env:
- {name: HERMES_WORKER_ROOT, value: /workspace}
- {name: HERMES_PROVIDER_ACCESS_ROOT, value: /provider-access}
- name: HERMES_WORKER_ORDINAL
valueFrom:
fieldRef:
fieldPath: metadata.labels['apps.kubernetes.io/pod-index']
securityContext:
allowPrivilegeEscalation: false
runAsUser: 0
runAsGroup: 0
seccompProfile: {type: RuntimeDefault}
volumeMounts:
- {name: claude-oauth-access, mountPath: /claude-oauth-access}
- {name: workspace, mountPath: /workspace}
- {name: provider-access, mountPath: /provider-access}
- {name: coordinator, mountPath: /opt/coordinator, readOnly: true}
resources:
requests: {cpu: 2m, memory: 32Mi}
limits: {cpu: 100m, memory: 64Mi}
- name: install-worker-go
image: registry.bstein.dev/bstein/hermes-agent@sha256:81970563e542f0720773e72297810b3a844b83e381e278f25c0916c78d930107
imagePullPolicy: IfNotPresent
command: [/bin/sh, -ec, "timeout 300 /bin/sh /opt/coordinator/install_worker_go.sh"]
env:
- {name: HERMES_WORKER_TOOLS_DIR, value: /worker-data/tools}
securityContext:
allowPrivilegeEscalation: false
capabilities: {drop: [ALL]}
runAsNonRoot: true
runAsUser: 10000
runAsGroup: 10000
seccompProfile: {type: RuntimeDefault}
volumeMounts:
- {name: tools, mountPath: /worker-data/tools}
- {name: coordinator, mountPath: /opt/coordinator, readOnly: true}
resources:
requests: {cpu: 5m, memory: 64Mi}
limits: {cpu: "1", memory: 256Mi}
- name: install-provider-clis
image: registry.bstein.dev/bstein/hermes-agent@sha256:81970563e542f0720773e72297810b3a844b83e381e278f25c0916c78d930107
imagePullPolicy: IfNotPresent
command: [/bin/sh, -ec]
args:
# The tools volume is durable, so this install runs once per pinned
# version rather than on every Pod start. The version marker is only
# trusted when both binaries are actually present and executable, so
# a partial or pruned cache reinstalls instead of failing closed
# forever. The install itself is time-bounded so a stalled registry
# surfaces as a fast init failure with backoff instead of an
# open-ended wait that a Flux health timeout would have to absorb.
- |
tools=/worker-data/tools
marker="${tools}/.cli-versions-0.147.0-2.1.226"
mkdir -p "${tools}/bin"
if [ ! -f "${marker}" ] || [ ! -x "${tools}/bin/codex" ] \
|| [ ! -x "${tools}/bin/claude" ]; then
rm -f "${marker}"
timeout 900 npm install --global --omit=dev --no-audit --no-fund \
--fetch-timeout=120000 --fetch-retries=2 --prefix "${tools}" \
@openai/codex@0.147.0 @anthropic-ai/claude-code@2.1.226
touch "${marker}"
fi
test -x "${tools}/bin/codex"
test -x "${tools}/bin/claude"
securityContext:
allowPrivilegeEscalation: false
capabilities: {drop: [ALL]}
runAsNonRoot: true
runAsUser: 10000
runAsGroup: 10000
seccompProfile: {type: RuntimeDefault}
volumeMounts:
- {name: tools, mountPath: /worker-data/tools}
resources:
requests: {cpu: 5m, memory: 64Mi}
limits: {cpu: "1", memory: 1Gi}
containers:
- name: execution-worker
image: registry.bstein.dev/bstein/hermes-agent@sha256:81970563e542f0720773e72297810b3a844b83e381e278f25c0916c78d930107
imagePullPolicy: IfNotPresent
command: [/opt/hermes/.venv/bin/python, /opt/coordinator/execution_pool_worker.py]
env:
- {name: HERMES_HOME, value: /worker-data}
- {name: HERMES_WORKER_ROOT, value: /workspace}
- {name: HOME, value: /worker-data/home}
- {name: CODEX_HOME, value: /provider-access/codex}
- {name: CLAUDE_CONFIG_DIR, value: /provider-access/claude}
- {name: CLAUDE_CODE_OAUTH_TOKEN_FILE, value: /claude-oauth-access/token}
- {name: HERMES_CLAUDE_BIN, value: /opt/coordinator/claude_oauth_exec}
- {name: HERMES_CLAUDE_NATIVE_BIN, value: /worker-data/tools/bin/claude}
# Codex OAuth is deliberately not copied from Hermes' shared refresh lineage.
- {name: HERMES_EXECUTION_DISABLED_PROVIDER, value: codex}
- {name: HERMES_AUTO_ROUTER_PROFILE, value: agent}
- {name: PYTHONPATH, value: /opt/hermes}
- {name: PATH, value: /worker-data/tools/bin:/opt/coordinator:/opt/hermes/.venv/bin:/usr/local/bin:/usr/bin:/bin}
- name: HERMES_WORKER_ORDINAL
valueFrom:
fieldRef:
fieldPath: metadata.labels['apps.kubernetes.io/pod-index']
- name: HERMES_WORKER_NODE
valueFrom:
fieldRef:
fieldPath: spec.nodeName
startupProbe:
exec:
command: [/bin/sh, -ec, "test -w /workspace && test -r /claude-oauth-access/token"]
periodSeconds: 5
failureThreshold: 60
readinessProbe:
exec:
command: [/bin/sh, -ec, "test -w /workspace && test -r /claude-oauth-access/token"]
periodSeconds: 10
securityContext:
allowPrivilegeEscalation: false
capabilities: {drop: [ALL]}
readOnlyRootFilesystem: true
runAsNonRoot: true
runAsUser: 10000
runAsGroup: 10000
seccompProfile: {type: RuntimeDefault}
volumeMounts:
- {name: claude-oauth-access, mountPath: /claude-oauth-access, readOnly: true}
- {name: workspace, mountPath: /workspace}
- {name: worker-data, mountPath: /worker-data}
- {name: tools, mountPath: /worker-data/tools, readOnly: true}
- {name: provider-access, mountPath: /provider-access}
- {name: coordinator, mountPath: /opt/coordinator, readOnly: true}
- {name: routing-catalog, mountPath: /routing-catalog, readOnly: true}
- {name: tmp, mountPath: /tmp}
resources:
requests: {cpu: 5m, memory: 128Mi, ephemeral-storage: 1Gi}
limits: {cpu: "2", memory: 4Gi, ephemeral-storage: 8Gi}
volumes:
- name: claude-oauth-access
emptyDir: {medium: Memory, sizeLimit: 1Mi}
- name: worker-data
emptyDir: {sizeLimit: 128Mi}
- name: coordinator
configMap:
name: hermes-execution-pool
defaultMode: 0555
- name: routing-catalog
persistentVolumeClaim:
claimName: hermes-routing-catalog
readOnly: true
- name: tmp
emptyDir: {sizeLimit: 2Gi}
- name: vault-auth-token
projected:
defaultMode: 0600
sources:
- serviceAccountToken:
audience: vault
expirationSeconds: 3600
path: token
- configMap:
name: kube-root-ca.crt
items:
- {key: ca.crt, path: ca.crt}
- downwardAPI:
items:
- {path: namespace, fieldRef: {fieldPath: metadata.namespace}}
volumeClaimTemplates:
# The workspace is shared with this ordinal's mediator, which performs every
# SCM operation on it. ReadWriteMany removes the cross-node Multi-Attach
# deadlock that a drain or preemption used to create between the two Pods,
# so each can be evicted and rescheduled independently. Longhorn already
# serves the hermes-chat tenant workspaces this way on the same class.
- metadata:
name: workspace
labels:
app: hermes-execution-worker
spec:
accessModes: [ReadWriteMany]
storageClassName: astreae
resources:
requests:
storage: 30Gi
# Provider CLIs are installed once per pinned version onto a durable volume
# instead of being re-downloaded into an emptyDir on every Pod start.
- metadata:
name: tools
labels:
app: hermes-execution-worker
spec:
accessModes: [ReadWriteOnce]
storageClassName: astreae
resources:
requests:
storage: 2Gi
# Claude session state remains ordinal-private and survives Pod replacement.
- metadata:
name: provider-access
labels:
app: hermes-execution-worker
spec:
accessModes: [ReadWriteOnce]
storageClassName: astreae
resources:
requests:
storage: 1Gi