.gitea/workflows/local-provision-e2e.yml

@@ -74,6 +74,10 @@ jobs:
     env:
       PG_CONTAINER: pg-lpe2e-${{ github.run_id }}-${{ github.run_attempt }}
       REDIS_CONTAINER: redis-lpe2e-${{ github.run_id }}-${{ github.run_attempt }}
+      # Hard-code dev mode at the job level so the platform server ALWAYS sees it,
+      # even if the runner's $GITHUB_ENV propagation is flaky (#2468 RCA).
+      MOLECULE_ENV: development
+      SECRETS_ENCRYPTION_KEY: lpe2e-test-encryption-key-32bytes!!
     steps:
       - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
       - uses: actions/setup-go@40f1582b2485089dde7abd97c1529aa768e1baff # v5
@@ -254,6 +258,10 @@ jobs:
     env:
       PG_CONTAINER: pg-lpe2e-real-${{ github.run_id }}-${{ github.run_attempt }}
       REDIS_CONTAINER: redis-lpe2e-real-${{ github.run_id }}-${{ github.run_attempt }}
+      # Hard-code dev mode at the job level so the platform server ALWAYS sees it,
+      # even if the runner's $GITHUB_ENV propagation is flaky (#2468 RCA).
+      MOLECULE_ENV: development
+      SECRETS_ENCRYPTION_KEY: lpe2e-test-encryption-key-32bytes!!
     steps:
       - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
       - uses: actions/setup-go@40f1582b2485089dde7abd97c1529aa768e1baff # v5

.gitea/workflows/publish-workspace-server-image.yml

@@ -248,16 +248,36 @@ jobs:
             --tag "${STAGING_TENANT_IMAGE_NAME}:${TAG_LATEST}"
           )
 
-          docker buildx build \
-            --file ./workspace-server/Dockerfile.tenant \
-            --build-arg NEXT_PUBLIC_PLATFORM_URL= \
-            --build-arg GIT_SHA="${GIT_SHA}" \
-            --label "org.opencontainers.image.source=https://git.moleculesai.app/molecule-ai/${REPO}" \
-            --label "org.opencontainers.image.revision=${GIT_SHA}" \
-            --label "org.opencontainers.image.created=$(date -u +%Y-%m-%dT%H:%M:%SZ)" \
-            --label "molecule.workflow.run_id=${GITHUB_RUN_ID}" \
-            "${build_tags[@]}" \
-            --push .
+          # Retry loop: buildkit EOF (internal#2468) is often transient on the
+          # publish runner under memory pressure. Up to 3 attempts with a fresh
+          # builder each time so a crashed buildkit doesn't poison the next try.
+          for attempt in 1 2 3; do
+            echo "::notice::Tenant image build attempt ${attempt}/3 ..."
+            builder="tenant-builder-${GITHUB_RUN_ID}-${attempt}"
+            docker buildx create --name "${builder}" --use >/dev/null 2>&1 || true
+            if docker buildx build \
+                --builder "${builder}" \
+                --file ./workspace-server/Dockerfile.tenant \
+                --build-arg NEXT_PUBLIC_PLATFORM_URL= \
+                --build-arg GIT_SHA="${GIT_SHA}" \
+                --label "org.opencontainers.image.source=https://git.moleculesai.app/molecule-ai/${REPO}" \
+                --label "org.opencontainers.image.revision=${GIT_SHA}" \
+                --label "org.opencontainers.image.created=$(date -u +%Y-%m-%dT%H:%M:%SZ)" \
+                --label "molecule.workflow.run_id=${GITHUB_RUN_ID}" \
+                "${build_tags[@]}" \
+                --push .; then
+              docker buildx rm "${builder}" >/dev/null 2>&1 || true
+              echo "::notice::Tenant image build succeeded on attempt ${attempt}"
+              break
+            fi
+            echo "::warning::Tenant image build attempt ${attempt} failed — cleaning builder and retrying"
+            docker buildx rm "${builder}" >/dev/null 2>&1 || true
+            sleep 10
+            if [ "$attempt" -eq 3 ]; then
+              echo "::error::Tenant image build failed after 3 attempts"
+              exit 1
+            fi
+          done
 
   # bp-exempt: production deploy side-effect; merge is gated by CI / all-required and this job waits for push CI before acting.
   deploy-production:

tests/e2e/test_local_provision_lifecycle_e2e.sh

@@ -488,6 +488,12 @@ echo ""
 # Step 5 — proxy reach (ws-<id>:8000 Docker-DNS rewrite, end to end).
 # ----------------------------------------------------------------------------
 echo "--- Step 5: proxy reach (POST /workspaces/$WSID/a2a) ---"
+# Debug: print the workspace URL the platform stored so SSRF failures are
+# actionable (#2468 RCA).
+WS_DEBUG=$(admin_curl "$BASE/workspaces/$WSID")
+WS_URL_DEBUG=$(ws_field "$WS_DEBUG" "url")
+WS_STATUS_DEBUG=$(ws_field "$WS_DEBUG" "status")
+echo "  workspace url=$WS_URL_DEBUG status=$WS_STATUS_DEBUG"
 # In minimax mode we send a DETERMINISTIC known-answer prompt and assert the
 # model echoes the answer back — proving a real LLM round-trip, not just
 # reachability. Otherwise a plain "ping".