AutoCoder branch-validation run 8617 attempted cache restore at 0.0.0.0:8088 and received ECONNREFUSED. The live cache Deployment launched act_runner cache-server --host 0.0.0.0 --port 8088; its log said it was listening on http://0.0.0.0:8088. That wildcard address cannot be reached as a destination from Docker job containers. This caused repeated cold Rust restores and undermined the candidate delivery latency target.
The repository-managed manifest documentation/configs/autocoder-engine/gitea-migration/gitea-act-runner-shared-cache.yaml now advertises gitea-act-runner-cache.gitea.svc.cluster.local while retaining port 8088 and the existing worker-local hostPath cache storage. The administrator applied that manifest to the gitea namespace on the verified kubernetes-admin@kubernetes context. Only the cache Deployment changed; Service and NetworkPolicy were unchanged. The Deployment rolled out successfully.
The replacement server logged http://gitea-act-runner-cache.gitea.svc.cluster.local:8088. The Service EndpointSlice was ready, and an active Actions job container resolved and reached the endpoint (HTTP 404 on /, proving connectivity). Run 8621 then saved Cargo and Rust toolchain caches successfully. Run 8624 restored both caches successfully and its branch validation passed in 2 minutes 27 seconds (16:44:20–16:46:47 UTC). This proves a real restore, not the candidate-to-healthy-test p95 SLA. To roll back, restore the previous manifest revision and apply it, but that reinstates the known invalid advertised URL. Cache entries were not intentionally deleted.