Deploy the system-bus-worker to the joelclaw Kubernetes cluster from local machine. Use when syncing changes in packages/system-bus to k8s, especially because the GitHub Actions deploy job targets a non-existent self-hosted runner and cannot complete deploys automatically.
86%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
High
Do not use without reviewing
Deploy system-bus-worker to the local joelclaw k8s cluster (Talos v1.12.4 / k8s v1.35.0).
Important: .github/workflows/system-bus-worker-deploy.yml has a deploy job on self-hosted. That runner does not exist, so deploys must be completed locally.
The publish script handles everything — build, auth, push, k8s apply, rollout, verification:
cd ~/Code/joelhooks/joelclaw
k8s/publish-system-bus-worker.shOptional: pass a tag (defaults to timestamp):
k8s/publish-system-bus-worker.sh a6de1e0publish-system-bus-worker.sh now authenticates in this order:
GHCR_TOKEN env var (if provided)secrets lease ghcr_pat (agent-secrets)gh auth token fallbackIf your gh auth token lacks read:packages/write:packages, push will 403. Use ghcr_pat.
agent-secrets lease ghcr_pat; falls back to gh auth token) with temp Docker configghcr.io/joelhooks/system-bus-worker:${TAG} and :latestk8s/system-bus-worker.yamlkubectl apply the manifest--timeout=180s)joelclaw refresh # Re-register functions with Inngest
joelclaw functions | grep "<new-function>" # Verify new function appears
joelclaw status # Full health check
joelclaw runs --count 3 # Confirm runs are flowingThe worker is stateless between Inngest steps. Each step is a separate HTTP call; Inngest stores step output server-side. This means k8s rolling restarts are safe — Inngest retries the in-flight step against the new pod.
Critical rule: NEVER set retries: 0 on Inngest functions. With retries: 0, a worker restart during step execution kills the run permanently. With retries ≥ 1, Inngest retries and hits the new pod.
Current story-pipeline has retries: 2 specifically to survive the ~1s restart window during deploys.
Step executing on old pod → old pod terminates → step fails (SDK unreachable)
→ Inngest retries after backoff → new pod handles retry → step completesAll previously completed steps are memoized. Only the in-flight step reruns.
If a deploy kills a codex step mid-execution, the step reruns from scratch on the new pod (5-10 min wasted but not fatal). For time-critical deploys during active loops, check joelclaw loop status first and deploy between stories.
cd ~/Code/joelhooks/joelclaw
TAG=$(git rev-parse --short HEAD)
IMAGE="ghcr.io/joelhooks/system-bus-worker:${TAG}"
docker build --platform linux/arm64 -t "$IMAGE" -t ghcr.io/joelhooks/system-bus-worker:latest -f packages/system-bus/Dockerfile .gh auth token | docker login ghcr.io -u $(gh api user -q .login) --password-stdin
docker push "$IMAGE"
docker push ghcr.io/joelhooks/system-bus-worker:latestkubectl -n joelclaw set image deployment/system-bus-worker system-bus-worker="$IMAGE"
kubectl -n joelclaw rollout status deployment/system-bus-worker --timeout=180sjoelclaw refresh
joelclaw statusVerify the rollout and keep the image/digest plus command output as the deploy receipt. Runtime deploy telemetry belongs in canonical OTel; durable behavior/config changes belong in the relevant Brain .svx. The former slog journal is retired.
Talon is a Rust binary that supervises the worker process. It leases secrets from agent-secrets and injects them as env vars. When adding new webhook secrets or changing supervision behavior:
# 1. Add secret to agent-secrets
secrets add my_new_secret --value "the-secret-value"
# 2. Update Talon source — add mapping to SECRET_MAPPINGS array
# File: ~/Code/joelhooks/joelclaw/infra/talon/src/worker.rs
# ("my_new_secret", "MY_NEW_SECRET_ENV_VAR"),
# 3. Recompile (fast — ~3s incremental)
export PATH="$HOME/.cargo/bin:$PATH"
cd ~/Code/joelhooks/joelclaw/infra/talon
cargo build --release
# 4. Install + re-sign (macOS kills unsigned binaries)
cp target/release/talon ~/.local/bin/talon
codesign -fs - ~/.local/bin/talon
# 5. Restart via launchd
launchctl bootout gui/$(id -u)/com.joel.talon
sleep 1
launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.joel.talon.plist
sleep 12
# 6. Verify
curl -s http://localhost:3111/ | jq '.status'
curl -X PUT http://localhost:3111/api/inngest # Force function sync| Secret Name | Env Var |
|---|---|
claude_oauth_token | CLAUDE_CODE_OAUTH_TOKEN |
todoist_client_secret | TODOIST_CLIENT_SECRET |
todoist_api_token | TODOIST_API_TOKEN |
front_rules_webhook_secret | FRONT_WEBHOOK_SECRET |
front_api_token | FRONT_API_TOKEN |
vercel_webhook_secret | VERCEL_WEBHOOK_SECRET |
joelclaw_webhook_secret | JOELCLAW_WEBHOOK_SECRET |
revalidation_secret | REVALIDATION_SECRET |
| What | Path |
|---|---|
| Binary | ~/.local/bin/talon |
| Source | ~/Code/joelhooks/joelclaw/infra/talon/src/ |
| LaunchAgent plist | ~/Library/LaunchAgents/com.joel.talon.plist |
| Logs | ~/.local/log/talon.log / talon.err |
| ADR | ~/Vault/docs/decisions/0159-talon-worker-manager.md |
codesign -fs - is requiredAfter cargo build, the binary has adhoc linker-signed signature. macOS launchd may SIGKILL:9 it. Re-signing with codesign -fs - fixes this.
| Problem | Cause | Fix |
|---|---|---|
exec format error in pod | Built for amd64, not arm64 | Rebuild with --platform linux/arm64 |
GHCR push fails with 403 Forbidden on blob HEAD | gh auth token missing package scopes | Use ghcr_pat via agent-secrets or export GHCR_TOKEN with package scope |
docker-credential-desktop error | Docker config has credsStore | Script uses temp config dir — if manual, remove "credsStore": "desktop" |
| Function missing after deploy | Not in the index for the worker that should own it | Add it to index.cluster.ts for cluster-owned callbacks or index.host.ts for host-only work; do not blindly add to both or Inngest will create duplicate runs |
| Function still missing | Stale Inngest registration | joelclaw refresh then check again |
| "Unable to reach SDK URL" | Worker pod not ready | Wait for rollout, then joelclaw refresh |
| Runs stuck after deploy | retries: 0 on the function | Set retries: 2 minimum (ADR-0156) |
| Stale app registrations | Multiple apps registered | Delete old registrations in Inngest dashboard (:8289) |
| What | Path |
|---|---|
| Publish script | k8s/publish-system-bus-worker.sh |
| Dockerfile | packages/system-bus/Dockerfile |
| k8s manifest | k8s/system-bus-worker.yaml |
| Host function index | packages/system-bus/src/inngest/functions/index.host.ts |
| Cluster function index | packages/system-bus/src/inngest/functions/index.cluster.ts |
| Worker entry | packages/system-bus/src/serve.ts |
| GH Actions workflow | .github/workflows/system-bus-worker-deploy.yml |
| ADR-0156 | ~/Vault/docs/decisions/0156-graceful-worker-restart.md |
joelhooks/joelclaw
ee2c272
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.