All CI workflows across all repos were broken. Three root causes:
1. Runner labels pointed to bare Alpine image (forgejo/runner:6) which has
no Node.js, no docker, no Go/Rust, no root, no apt-get. Every workflow
step failed immediately.
Fix: Change labels to official Debian language images:
golang → docker://golang:1.26-bookworm
node → docker://node:22-bookworm
rust → docker://rust:1-bookworm
2. Docker socket not shared between dind sidecar and runner container.
dind creates /run/docker.sock in its own filesystem. Runner container
has a separate filesystem and can't see it. Workflow containers that
run 'docker build' get 'Cannot connect to Docker daemon'.
Fix: Share /run between dind and runner via emptyDir volume.
Add docker_host: automount to runner config so the socket is
mounted into workflow containers automatically.
Note: /var/run is a symlink to /run in Alpine — must mount at /run.
3. Init container skipped re-registration if .runner file existed on PVC.
Changing labels in values.yaml had no effect until PVC was manually
deleted — not GitOps-friendly.
Fix: Always rm .runner and re-register on pod start. Labels stay
in sync with values.yaml automatically.
Also removes custom Dockerfiles and build-runner-images workflow — no longer
needed since official images provide the language tools directly.
Problem: In-cluster pods (portfolio, services) couldn't resolve
api.riotpiao.com because it was missing from CoreDNS rewrite rules.
This broke LLM API calls from portfolio → api gateway even with valid JWT.
Solution: Add rewrite rule to route api.riotpiao.com through nginx ingress
(TLS termination + Host header preservation), matching pattern for other
internal hostnames (authentik.riotpiao.com, minio.riotpiao.com, etc).
Impact:
- Portfolio pod now successfully resolves api.riotpiao.com
- LLM API calls proceed to auth/permission checking
- Applies to all in-cluster services needing LLM gateway
rock