@norskvideo/ctl-dev-kit 0.1.65 → 0.1.66

package/conventions/smoke.yml

@@ -102,37 +102,47 @@ jobs:
       # gracefully, and this workflow cancels its own superseded runs -- a killed
       # daemon reaps nothing and the port stays held for every later job here.
       #
-      # On 2026-08-30 something held it on one runner from 11:18 to ~16:00 and
+      # On 2026-08-30 something held it on ar-arm-vm from 11:18 to ~16:00 and
       # took out the next nine smoke and integration jobs across four repos.
       # What held it was never established: the job that last bound it PASSED
-      # and shut down cleanly, so it may not have been a control-plane container
-      # at all -- a stale docker-proxy or dangling endpoint reports the same
-      # "already allocated". So this reaps what it can name AND names whatever
-      # still holds the port, rather than exiting quiet and leaving the next
-      # occurrence as much of a mystery as this one.
+      # and shut down cleanly, and every add-path failure already removes its
+      # own container -- so the holder may not have been a control plane at all.
+      # "already allocated" is also what a dangling endpoint or a stranded
+      # docker-proxy reports after an unclean removal.
       #
-      # Gated on the daemon being this runner's own. The GPU box runs several
-      # runners as sibling CONTAINERS against one host daemon, where a blanket
-      # reap would kill a concurrent job's control plane.
+      # Hence: kill by PORT rather than by name, prune the endpoints a removal
+      # can strand, and then PROVE the port binds instead of assuming it. This
+      # branch only runs where the docker daemon is this runner's own and jobs
+      # are serial, so there is nothing else here to protect. The GPU box runs
+      # several runners as sibling CONTAINERS against one host daemon and is
+      # excluded -- there, a blanket reap would kill a concurrent job.
       - name: Reap a stranded product control plane
         if: ${{ env.NORSK_RUNNER_OWNS_DOCKER == 'true' }}
         run: |
           set -uo pipefail
-          stale="$(docker ps -aq --filter 'name=^norsk-product-')"
-          if [ -n "$stale" ]; then
-            echo "::warning::reaping a stranded product control plane -- a previous job did not tear its own down"
-            docker inspect --format '{{.Name}} {{.State.Status}} {{.Config.Image}}' $stale || true
-            docker rm -f $stale || true
+          port=14321
+          bindable() { docker run --rm -p "127.0.0.1:$port:1" alpine true >/dev/null 2>&1; }
+
+          held="$(docker ps -a --filter "publish=$port" --format '{{.Names}} {{.Status}} {{.Image}}')"
+          if [ -n "$held" ]; then
+            echo "::warning::$port was held on entry by: $held"
+            docker rm -f $(docker ps -aq --filter "publish=$port") || true
           fi
-          # Whatever is left on the port is the thing worth seeing: if this is
-          # ever non-empty after the reap above, the holder is not a control
-          # plane and `docker rm` was never going to clear it.
-          holder="$(docker ps -a --filter 'publish=14321' --format '{{.Names}} {{.Status}} {{.Image}}')"
-          if [ -n "$holder" ]; then
-            echo "::warning::14321 is still published by: $holder"
-          else
-            echo "14321 free; no container publishes it"
+          # A control plane that is stopped, or one whose publish docker no
+          # longer reports, still gets removed: nothing here outlives a job.
+          leftover="$(docker ps -aq --filter 'name=^norsk-product-')"
+          [ -z "$leftover" ] || docker rm -f $leftover || true
+
+          if ! bindable; then
+            echo "::warning::$port still not bindable after removing containers -- pruning dangling endpoints"
+            docker network prune -f || true
+          fi
+          if ! bindable; then
+            echo "::error::cannot bind 127.0.0.1:$port; the journey would fail later and more obscurely"
+            docker ps -a --filter "publish=$port" || true
+            exit 1
           fi
+          echo "$port is bindable"
 
       - name: Write the Norsk license (from the org secret)
         env:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@norskvideo/ctl-dev-kit",
-  "version": "0.1.65",
+  "version": "0.1.66",
   "type": "module",
   "exports": {
     "./create-product": "./create-product/create-product.ts",