diff --git a/.forgejo/workflows/auto-retrain-knowledge.yml b/.forgejo/workflows/auto-retrain-knowledge.yml
index 44f0414..680a512 100644
--- a/.forgejo/workflows/auto-retrain-knowledge.yml
+++ b/.forgejo/workflows/auto-retrain-knowledge.yml
@@ -55,13 +55,24 @@ jobs:
         env:
           FORCE_TRAINING: ${{ inputs.force }}
 
-      - name: Trigger training on VPS
+      - name: Trigger training on GPU workstation
         if: steps.check.outputs.should_train == 'true'
         run: |
-          echo "Triggering training on VPS via systemd..."
+          echo "Triggering training on GPU workstation via webhook..."
 
-          # In production, this would SSH to the VPS and trigger systemd service
-          # For now, create a marker file that monitoring can pick up
+          # Trigger via webhook (GPU workstation must run training-webhook-server.py)
+          curl -X POST http://${{ secrets.GPU_WORKSTATION_HOST }}:8888/trigger-training \
+            -H "Authorization: Bearer ${{ secrets.TRAINING_WEBHOOK_TOKEN }}" \
+            -f -s -o /tmp/trigger-response.json || {
+              echo "::error::Failed to trigger training on GPU workstation"
+              exit 1
+            }
+
+          # Show response
+          cat /tmp/trigger-response.json
+          echo "::notice::Training triggered on GPU workstation"
+
+          # Create marker file for record-keeping
           mkdir -p .artifacts/training-triggers
           cat > .artifacts/training-triggers/trigger-$(date +%Y%m%d-%H%M%S).json << EOF
           {