Use skills

Sync skill content into the workspace and attach it to a PromptPack so the model can activate skills on demand.

You create a SkillSource, publish a PromptPack version that lists it in spec.skills, then check that the agent pod loads the manifest.

Prerequisites

Workspace content storage. The chart enables it, and the operator receives --workspace-content-path (default /workspace-content).
A PromptPack from a ConfigMap, already running on an AgentRuntime.
A kubectl that can reach the workspace namespace (dev-agents below).

Step 1: create a SkillSource

Save as skills.yaml. The repository URL, tag and skill names are illustrative; use your own.

apiVersion: omnia.altairalabs.ai/v1alpha1
kind: SkillSource
metadata:
  name: support-skills
  namespace: dev-agents
spec:
  type: git
  git:
    url: https://github.com/anthropic/skills
    ref: { tag: v1.4.0 }
  interval: 1h
  targetPath: skills/anthropic
  filter:
    include: ["ai-safety", "pdf-processing"]

Apply it and wait for the first fetch:

kubectl apply -f skills.yaml
kubectl get skillsource support-skills -n dev-agents -w

The source is synced when PHASE is Ready and SKILLS matches the filtered set. Then check that every retained SKILL.md parsed:

kubectl get skillsource support-skills -n dev-agents \
  -o jsonpath='{.status.conditions[?(@.type=="ContentValid")]}'

status: "True" means every SKILL.md parsed and no two skills in the source share a name.

See SkillSource for the oci and configmap source types and every filter field.

Step 2: attach to a PromptPack

A published PromptPack version is immutable, so kubectl patch on spec.skills is rejected. Publish a new version that lists the source. Save as pack-v2.yaml:

apiVersion: omnia.altairalabs.ai/v1alpha1
kind: PromptPack
metadata:
  name: support-pack-1-1-0
  namespace: dev-agents
spec:
  packName: support-pack
  version: "1.1.0"
  source:
    type: configmap
    configMapRef:
      name: support-pack-prompts   # the ConfigMap your current version uses
  skills:
    - source: support-skills
      include: ["ai-safety"]
      mountAs: compliance
  skillsConfig:
    maxActive: 5

To offer the model only some skills, add selector: tag and a tags list under skillsConfig. A skill is offered when its SKILL.md metadata tags shares an entry with that list.

kubectl apply -f pack-v2.yaml

An AgentRuntime that pins promptPackRef.version keeps serving the old version until you change the pin.

The PromptPack controller resolves the references, writes a JSON manifest into the workspace content volume, and sets three conditions:

kubectl get promptpack support-pack-1-1-0 -n dev-agents \
  -o jsonpath='{range .status.conditions[*]}{.type} {.status} {.reason}{"\n"}{end}'

Condition	Reason when `True`	Meaning
`SkillsResolved`	`AllSkillsResolved`	Every `spec.skills[].source` exists and has synced
`SkillsValid`	`NoCollisions`	No two referenced sources publish a skill with the same name
`SkillToolsResolved`	`AllToolsResolved`	Every skill’s `allowed-tools` is a tool the pack declares

If SkillToolsResolved is False, add the tool to the pack’s tools or remove it from the skill’s allowed-tools.

See PromptPack for every skills and skillsConfig field.

Step 3: verify the runtime loads the manifest

The operator rebuilds the pod of each AgentRuntime that serves the pack. The runtime container gets:

Your workspace’s part of the content volume, mounted read-only at /workspace-content. The pod sees no other workspace’s content.
The env var OMNIA_PROMPTPACK_MANIFEST_PATH=/workspace-content/manifests/<promptpack-name>.json, where <promptpack-name> is the PromptPack’s metadata.name (support-pack-1-1-0 above).

Inspect the pod:

kubectl exec <agent-pod> -n dev-agents -c runtime -- env | grep OMNIA_PROMPTPACK
kubectl exec <agent-pod> -n dev-agents -c runtime -- cat /workspace-content/manifests/support-pack-1-1-0.json
kubectl logs <agent-pod> -n dev-agents -c runtime | grep "skill manifest loaded"

The log line lists skillNames, skillMounts, maxActive and selector.

Step 4: test from a session

Open a session and ask about something a skill covers (for ai-safety, handling sensitive content). The runtime gives the model an index of the available skills, and the model calls the skill__activate tool to load the one it needs before answering.

Confirm the activation in the session’s tool calls:

curl -s <session-api-url>/api/v1/sessions/<session-id>/tool-calls \
  | jq '.[] | select(.name=="skill__activate")'

skill__activate is a server-side tool, so it appears here and in the runtime logs, not as a tool_call message on the WebSocket.

Troubleshooting

Symptom	Likely cause	Fix
`SkillsResolved=False`, reason `LookupFailed`, message `SkillSource ... not found`	The `SkillSource` is not in the PromptPack’s namespace. Lookup does not cross namespaces	`kubectl get skillsource -n <pack-ns>`
`SkillsResolved=False`, reason `LookupFailed`, message `... has no synced artifact yet`	The `SkillSource` has not finished its first fetch	Wait for `status.phase` to reach `Ready`
`SkillsResolved=False`, reason `ContentStorageUnavailable`	The operator started with an empty `--workspace-content-path`	Set `--workspace-content-path` on the operator and restart it. The Helm chart always sets it.
`SkillsValid=False`, reason `NameCollision`	Two referenced sources publish skills with the same `name`	Add `include` to one of the `skills` entries
`SkillToolsResolved=False`, reason `UnknownTool`	A `SKILL.md` `allowed-tools` entry is not declared in the pack	Declare the tool in the pack or drop it from the skill
`kubectl patch` on the PromptPack is rejected	Published versions are immutable	Apply a new version, as in step 2
The runtime pod has no `workspace-content` volume	The pack declares no `spec.skills`, or the operator has no content storage	Check step 2 and the operator’s `--workspace-content-path`