fix(deploy): reclaim superseded images so the host stops filling up

Every build pushes a new api + web image and every deploy pulls both onto
galactus, but nothing ever removed the pair they replaced. That reached 63
images / 83.85GB, of which 79.26GB was unused, and filled the 98GB root
filesystem to 100%.

The symptom was not a disk alert. It was "re-import is broken": the
Operaciones REIMPORT job leads with a mysqldump safety backup, that write
had nowhere to go, and PIPEFAIL took the job down before it touched the
database. Nothing in the ops_jobs log pointed at the disk.

Prune runs last, after the verify step, because Docker refuses to prune an
image that a container references — the running stack is what protects the
release just shipped. `until` adds a grace window on top so a rollback
dispatch stays a stack swap instead of a re-pull, but note it filters on
image creation time rather than pull time, so it does NOT cover rolling
back to an old tag; the running-container rule is what does.

continue-on-error: housekeeping that fails leaves a fat host, not a broken
release.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

.gitea/workflows/deploy-galactus.yml

@@ -25,6 +25,10 @@
 #                           through this workflow at all.
 #   4. app (api + web)      the new images.
 #   5. verify               ask the running API what it actually is.
+#   6. prune images         reclaim the superseded api/web images. LAST, and
+#                           after verify: Docker will not prune an image a
+#                           container references, so the running stack is what
+#                           protects the release we just shipped.
 #
 # Rollback = re-dispatch with an older `tag`. That rolls back CODE only; the
 # schema stays forward. This is exactly why every schema change must be
@@ -386,3 +390,24 @@ jobs:
               echo "dispatched '$WANT'; tiers report '$API_VER' (not directly comparable)"
               ;;
           esac
+
+      # --- housekeeping ------------------------------------------------------
+      # Runs LAST, and only after the verify step proved the new containers are
+      # up. See deploy/scripts/prune-images.mjs: Docker refuses to prune an
+      # image a container references, so "the stack is running" is what makes
+      # the current images safe. Pruning earlier would have nothing holding
+      # them.
+      #
+      # continue-on-error: reclaiming disk is not what the deploy is for. A
+      # prune that fails leaves a fat host, not a broken release.
+      - name: Prune unused images
+        continue-on-error: true
+        env:
+          PORTAINER_URL: ${{ secrets.PORTAINER_URL_GALACTUS }}
+          PORTAINER_API_KEY: ${{ secrets.PORTAINER_API_KEY_GALACTUS }}
+          PORTAINER_ENDPOINT_ID: ${{ secrets.PORTAINER_ENDPOINT_ID_GALACTUS }}
+          # Grace window. Keeps the previous few releases on disk so a rollback
+          # dispatch is a stack swap instead of a re-pull.
+          KEEP_HOURS: "168"
+          NODE_TLS_REJECT_UNAUTHORIZED: "0"
+        run: node deploy/scripts/prune-images.mjs

deploy/scripts/prune-images.mjs

@@ -0,0 +1,106 @@
+#!/usr/bin/env node
+/**
+ * Delete unused images from the target host after a successful deploy.
+ *
+ * This exists because nothing else reclaims them. Every build.yml run pushes a
+ * new api + web image, every deploy pulls both onto the host, and the previous
+ * pair is left behind untagged-but-present forever. On galactus that reached
+ * 63 images / 83.85GB (79.26GB of it unused) and filled the 98GB root
+ * filesystem to 100% on 2026-08-20 — which surfaced as "re-import is broken",
+ * because the Operaciones REIMPORT job leads with a mysqldump that could no
+ * longer write its safety backup.
+ *
+ * Two things keep this from eating a live deployment:
+ *
+ *  - Docker never prunes an image that a container references, running or
+ *    stopped. The five images the prod stacks use are therefore untouchable
+ *    for as long as their containers exist.
+ *  - `until` gives a grace window on top of that, so a rollback target stays
+ *    on disk instead of forcing a re-pull from the registry.
+ *
+ * TRAP: `until` filters on the image's CREATION time, not when the host pulled
+ * it. Rolling back to an old tag pulls an image that is already older than the
+ * window, so the grace period does NOT protect it — the running-container rule
+ * is what does. That is why this step must run AFTER the app stack is deployed
+ * and verified, never before.
+ *
+ * Required env:
+ *   PORTAINER_URL, PORTAINER_API_KEY, PORTAINER_ENDPOINT_ID
+ * Optional:
+ *   KEEP_HOURS   grace window in hours (default 168 = 7 days)
+ *
+ * TLS: Portainer here is self-signed; the caller sets
+ * NODE_TLS_REJECT_UNAUTHORIZED=0 for this step.
+ */
+
+function required(name) {
+  const v = process.env[name];
+  if (!v) {
+    console.error(`missing required env: ${name}`);
+    process.exit(1);
+  }
+  return v;
+}
+
+const PORTAINER_URL = required("PORTAINER_URL").replace(/\/+$/, "");
+const API_KEY = required("PORTAINER_API_KEY");
+const ENDPOINT_ID = required("PORTAINER_ENDPOINT_ID");
+const KEEP_HOURS = process.env.KEEP_HOURS || "168";
+
+const DOCKER = `${PORTAINER_URL}/api/endpoints/${ENDPOINT_ID}/docker`;
+
+// `dangling: ["false"]` is what makes this `docker image prune -a` rather than
+// the default, which only collects untagged layers. The tagged-but-superseded
+// api/web images are the whole problem, and the default filter walks straight
+// past them.
+const FILTERS = JSON.stringify({
+  dangling: ["false"],
+  until: [`${KEEP_HOURS}h`],
+});
+
+function human(bytes) {
+  if (!bytes) return "0B";
+  const units = ["B", "KB", "MB", "GB", "TB"];
+  let i = 0;
+  let n = bytes;
+  while (n >= 1024 && i < units.length - 1) {
+    n /= 1024;
+    i += 1;
+  }
+  return `${n.toFixed(i === 0 ? 0 : 2)}${units[i]}`;
+}
+
+async function main() {
+  const url = `${DOCKER}/images/prune?filters=${encodeURIComponent(FILTERS)}`;
+  const res = await fetch(url, {
+    method: "POST",
+    headers: { "X-API-Key": API_KEY },
+  });
+  const body = await res.text();
+  if (!res.ok) {
+    throw new Error(`prune -> HTTP ${res.status} ${body.slice(0, 300)}`);
+  }
+
+  let report;
+  try {
+    report = JSON.parse(body);
+  } catch {
+    throw new Error(`prune returned non-JSON: ${body.slice(0, 300)}`);
+  }
+
+  const deleted = report.ImagesDeleted ?? [];
+  const reclaimed = report.SpaceReclaimed ?? 0;
+  console.log(
+    `pruned images older than ${KEEP_HOURS}h and unused by any container`,
+  );
+  console.log(`  entries removed : ${deleted.length}`);
+  console.log(`  space reclaimed : ${human(reclaimed)}`);
+}
+
+main().catch((err) => {
+  // Non-fatal by contract: the step that calls this sets continue-on-error, so
+  // housekeeping never turns a good deploy red. Exit non-zero anyway so the
+  // failure is visible in the run rather than swallowed.
+  console.error(`::warning::image prune FAILED: ${err.message}`);
+  process.exit(1);
+});