Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
Workflow Output PVC Upgrade Note
This upgrade changes the default WorkflowRecipe output PVC ownership model for
spec.output.destination: pvc.
New Default
When spec.output.claimName is omitted, WRC now creates one workflow output PVC
per parent WorkflowRecipe:
<recipe-name>-workflow-output
Triggered child runs reuse that parent recipe PVC and write run bytes under:
workflow-output/<parentRecipe>/<runId>/
This avoids RWO Multi-Attach failures by co-locating the output anchor,
coordinator, mcp-host, snippet runner, and artifact reader on the same node.
It does not create PVCs per run.
For WRC-managed output PVCs, each triggered child run gets a one-shot
workflow-output-prepare pod before runtime pods start. That pod runs on the
same node as the output anchor, validates the run-scoped subPath, creates only
that directory chain, and makes it writable by the non-root workflow UID/GID
1000. If the prepare pod is pending or fails, WRC keeps the run in
initializing/failed state instead of starting coordinator or artifact-reader
pods against a path that cannot be written.
Existing Global PVC Artifacts
Older deployments used the shared global PVC:
clerum-workflow-output
Existing runs whose artifact bytes live only on that shared PVC will not be
served by a recipe after it moves to the new per-recipe PVC. The DB run history
remains, but download attempts for missing PVC bytes return artifact_gone
(410 Gone through the API).
Before rollout, operators must choose one path for recipes with historical artifacts that need to remain downloadable. WRC does not automatically adopt legacy output PVCs that lack the new ownership/scope labels, because doing so would make it impossible to distinguish old WRC data from a same-name external claim.
Keep a single recipe on the legacy global PVC by setting:
spec: output: destination: pvc claimName: clerum-workflow-outputWRC treats
claimNameas external storage: it mounts the PVC but does not create, resize, or delete it. The recipe will expose aWorkflowOutputLegacyGlobalClaimcondition as a compatibility warning. External claims are also responsible for their own filesystem permissions; WRC does not run the output prepare pod or mutate ownership on operator-provided claims. The external PVC must be explicitly labeled for the recipe scope before rollout:kubectl --context=<context> -n sandbox-recipes label pvc clerum-workflow-output \ clerum.io/workflow-output-external=true \ clerum.io/workflow-output-claim=clerum-workflow-output \ clerum.io/workflow-output-scope=<recipe-name>For long names, use the same RFC1123 hash/truncation value that WRC uses for workflow output labels. Do not point
spec.output.claimNameat a PVC labeledclerum.io/managed-by=wrc; removeclaimNameto use WRC-managed lifecycle and output preparation.The legacy global PVC can safely represent only one exact
clerum.io/workflow-output-scopeat a time under this compatibility model. Do not use this path to preserve downloads for multiple parent recipes on the same shared claim.Accept that old bytes are non-migrated historical data and let the recipe use the new WRC-managed per-recipe PVC. Retrigger runs that need fresh artifacts.
For multiple recipes that must keep historical downloads, create one operator-owned external PVC per parent recipe, copy that recipe's historical
workflow-output/<parentRecipe>/...bytes fromclerum-workflow-output, label each destination PVC with its exact claim and scope labels, and set that recipe'sspec.output.claimNameto its per-recipe external claim.Manually copy historical bytes from
clerum-workflow-outputinto the new WRC-managed per-recipe PVC using the same run-scoped path layout before relying on downloads.
Do not set claimName for workload/service storage PVCs. spec.output.claimName
is only for the workflow output PVC mounted at /output; PVCs declared under
workloads[].volumes keep their own lifecycle and are not workflow output
storage.
When a parent WorkflowRecipe is deleted, WRC deletes its managed output PVC and
the artifact bytes it contains. The database run and artifact metadata remain
historical records, and download attempts for missing bytes return
artifact_gone.