diff --git a/config/forge-saw/README.md b/config/forge-saw/README.md
index 6f1ea23..b2b566c 100644
--- a/config/forge-saw/README.md
+++ b/config/forge-saw/README.md
@@ -148,6 +148,24 @@ oc get secret openshell-credentials -n guy-ziv-evalflow -o json \
 `aeh-openshell-eval` TCP-checks `:17670` and optionally runs `openshell sandbox list`.
 If SAW is down the **run fails**. That step does not install SAW.
 
+## NetworkPolicy (Forge shared namespace)
+
+When SAW VMs and Tekton share a namespace with default-deny policies, evaluate
+pods must be allowed to reach the agent gateway on `:17670`. Use the canonical
+Pipeline name `abevalflow-pipeline-openshell` (so `tekton.dev/pipeline` matches)
+and label PipelineRuns with `app.kubernetes.io/part-of: abevalflow`.
+
+Same-namespace template:
+
+```bash
+sed "s/NAMESPACE/${EVAL_NS}/g" config/forge-saw/networkpolicy-ci-openshell.yaml \
+  | oc apply -f -
+```
+
+Do not create ad-hoc copies of the Pipeline under a different name unless those
+PipelineRuns also carry the `part-of=abevalflow` label and the NetworkPolicies
+above are applied — otherwise gateway preflight times out.
+
 ## Image
 
 Stock `agent-eval-harness:v1.0.x` cannot import `agent_eval.openshell`. Point
diff --git a/config/forge-saw/networkpolicy-ci-openshell.yaml b/config/forge-saw/networkpolicy-ci-openshell.yaml
new file mode 100644
index 0000000..5bd6f18
--- /dev/null
+++ b/config/forge-saw/networkpolicy-ci-openshell.yaml
@@ -0,0 +1,213 @@
+# NetworkPolicies for OpenShell OpenClaw evals when SAW VMs and the Tekton
+# pipeline share one namespace (Forge workspace style).
+#
+# Select evaluate pods by either:
+#   - tekton.dev/pipeline=abevalflow-pipeline-openshell  (canonical Pipeline name)
+#   - app.kubernetes.io/part-of=abevalflow               (set on PipelineRun labels)
+#
+# Do NOT rename/copy the Pipeline to an ad-hoc name without also labeling the
+# PipelineRun with app.kubernetes.io/part-of=abevalflow — otherwise default-deny
+# blocks TCP to the gateway on :17670 and evaluate fails preflight.
+#
+# Apply after substituting NAMESPACE (and AGENT_VM / INTEG_VM if different):
+#   sed "s/NAMESPACE/forge-nommen/g" networkpolicy-ci-openshell.yaml | oc apply -f -
+---
+apiVersion: networking.k8s.io/v1
+kind: NetworkPolicy
+metadata:
+  name: abevalflow-allow-ci-agent-gateway
+  namespace: NAMESPACE
+  labels:
+    app.kubernetes.io/part-of: abevalflow
+    app.kubernetes.io/component: forge-saw
+spec:
+  podSelector:
+    matchLabels:
+      app.kubernetes.io/name: openshell-saw-agent
+      vm.kubevirt.io/name: openshell-saw-agent
+  policyTypes: [Ingress]
+  ingress:
+    - from:
+        - podSelector:
+            matchLabels:
+              tekton.dev/pipeline: abevalflow-pipeline-openshell
+        - podSelector:
+            matchLabels:
+              app.kubernetes.io/part-of: abevalflow
+      ports:
+        - protocol: TCP
+          port: 17670
+---
+apiVersion: networking.k8s.io/v1
+kind: NetworkPolicy
+metadata:
+  name: abevalflow-allow-ci-integ-gateway
+  namespace: NAMESPACE
+  labels:
+    app.kubernetes.io/part-of: abevalflow
+    app.kubernetes.io/component: forge-saw
+spec:
+  podSelector:
+    matchLabels:
+      app.kubernetes.io/name: openshell-saw-integ
+      vm.kubevirt.io/name: openshell-saw-integ
+  policyTypes: [Ingress]
+  ingress:
+    - from:
+        - podSelector:
+            matchLabels:
+              tekton.dev/pipeline: abevalflow-pipeline-openshell
+        - podSelector:
+            matchLabels:
+              app.kubernetes.io/part-of: abevalflow
+      ports:
+        - protocol: TCP
+          port: 17670
+        - protocol: TCP
+          port: 18082
+---
+apiVersion: networking.k8s.io/v1
+kind: NetworkPolicy
+metadata:
+  name: abevalflow-allow-ci-egress-by-pipeline
+  namespace: NAMESPACE
+  labels:
+    app.kubernetes.io/part-of: abevalflow
+    app.kubernetes.io/component: forge-saw
+spec:
+  podSelector:
+    matchLabels:
+      tekton.dev/pipeline: abevalflow-pipeline-openshell
+  policyTypes: [Egress]
+  egress:
+    - to:
+        - podSelector:
+            matchLabels:
+              app.kubernetes.io/name: openshell-saw-agent
+              vm.kubevirt.io/name: openshell-saw-agent
+      ports:
+        - protocol: TCP
+          port: 17670
+        - protocol: TCP
+          port: 8443
+    - to:
+        - podSelector:
+            matchLabels:
+              app.kubernetes.io/name: openshell-saw-integ
+              vm.kubevirt.io/name: openshell-saw-integ
+      ports:
+        - protocol: TCP
+          port: 17670
+        - protocol: TCP
+          port: 18082
+        - protocol: TCP
+          port: 8443
+    - to:
+        - podSelector:
+            matchLabels:
+              app.kubernetes.io/name: litellm
+      ports:
+        - protocol: TCP
+          port: 4000
+    - to:
+        - podSelector:
+            matchLabels:
+              app.kubernetes.io/name: mlflow
+              app.kubernetes.io/part-of: abevalflow
+      ports:
+        - protocol: TCP
+          port: 5000
+    - to:
+        - podSelector:
+            matchLabels:
+              app.kubernetes.io/name: minio
+              app.kubernetes.io/part-of: abevalflow
+      ports:
+        - protocol: TCP
+          port: 9000
+    - to:
+        - namespaceSelector:
+            matchLabels:
+              kubernetes.io/metadata.name: openshift-dns
+      ports:
+        - protocol: UDP
+          port: 53
+        - protocol: TCP
+          port: 53
+        - protocol: UDP
+          port: 5353
+        - protocol: TCP
+          port: 5353
+---
+apiVersion: networking.k8s.io/v1
+kind: NetworkPolicy
+metadata:
+  name: abevalflow-allow-ci-egress-by-part-of
+  namespace: NAMESPACE
+  labels:
+    app.kubernetes.io/part-of: abevalflow
+    app.kubernetes.io/component: forge-saw
+spec:
+  podSelector:
+    matchLabels:
+      app.kubernetes.io/part-of: abevalflow
+  policyTypes: [Egress]
+  egress:
+    - to:
+        - podSelector:
+            matchLabels:
+              app.kubernetes.io/name: openshell-saw-agent
+              vm.kubevirt.io/name: openshell-saw-agent
+      ports:
+        - protocol: TCP
+          port: 17670
+        - protocol: TCP
+          port: 8443
+    - to:
+        - podSelector:
+            matchLabels:
+              app.kubernetes.io/name: openshell-saw-integ
+              vm.kubevirt.io/name: openshell-saw-integ
+      ports:
+        - protocol: TCP
+          port: 17670
+        - protocol: TCP
+          port: 18082
+        - protocol: TCP
+          port: 8443
+    - to:
+        - podSelector:
+            matchLabels:
+              app.kubernetes.io/name: litellm
+      ports:
+        - protocol: TCP
+          port: 4000
+    - to:
+        - podSelector:
+            matchLabels:
+              app.kubernetes.io/name: mlflow
+              app.kubernetes.io/part-of: abevalflow
+      ports:
+        - protocol: TCP
+          port: 5000
+    - to:
+        - podSelector:
+            matchLabels:
+              app.kubernetes.io/name: minio
+              app.kubernetes.io/part-of: abevalflow
+      ports:
+        - protocol: TCP
+          port: 9000
+    - to:
+        - namespaceSelector:
+            matchLabels:
+              kubernetes.io/metadata.name: openshift-dns
+      ports:
+        - protocol: UDP
+          port: 53
+        - protocol: TCP
+          port: 53
+        - protocol: UDP
+          port: 5353
+        - protocol: TCP
+          port: 5353
diff --git a/config/forge-saw/networkpolicy-gateway-from-abeval.yaml b/config/forge-saw/networkpolicy-gateway-from-abeval.yaml
index af1aaef..9071391 100644
--- a/config/forge-saw/networkpolicy-gateway-from-abeval.yaml
+++ b/config/forge-saw/networkpolicy-gateway-from-abeval.yaml
@@ -1,6 +1,11 @@
 # Allow evaluate Task pods to reach the OpenShell gateway on :17670.
 # bootstrap.sh rewrites namespace / from-namespace (SAW_NS, EVAL_NS).
 # Apply in the SAW namespace, not from the evaluate PipelineRun.
+#
+# Prefer matching Tekton pods by the canonical Pipeline name or the
+# app.kubernetes.io/part-of=abevalflow label on PipelineRuns. A bare
+# podSelector: {} from the whole EVAL_NS is broader than needed when SAW
+# and eval share a Forge workspace namespace — see networkpolicy-ci-openshell.yaml.
 apiVersion: networking.k8s.io/v1
 kind: NetworkPolicy
 metadata:
@@ -20,7 +25,15 @@ spec:
         - namespaceSelector:
             matchLabels:
               kubernetes.io/metadata.name: guy-ziv-evalflow
-        - podSelector: {}
+          podSelector:
+            matchLabels:
+              tekton.dev/pipeline: abevalflow-pipeline-openshell
+        - namespaceSelector:
+            matchLabels:
+              kubernetes.io/metadata.name: guy-ziv-evalflow
+          podSelector:
+            matchLabels:
+              app.kubernetes.io/part-of: abevalflow
       ports:
         - protocol: TCP
           port: 17670
diff --git a/config/forge-saw/openshell-eval-alias.yaml b/config/forge-saw/openshell-eval-alias.yaml
new file mode 100644
index 0000000..d8ac338
--- /dev/null
+++ b/config/forge-saw/openshell-eval-alias.yaml
@@ -0,0 +1,20 @@
+# Stable, namespace-local endpoint for evaluation jobs.  Selecting the agent
+# VMI directly avoids baking a deployment namespace into Pipeline defaults.
+# The gateway server leaf must include "openshell" as a DNS SAN before the
+# evaluation task can disable OPENSHELL_GATEWAY_INSECURE.
+apiVersion: v1
+kind: Service
+metadata:
+  name: openshell
+  labels:
+    app.kubernetes.io/name: openshell-eval-alias
+    app.kubernetes.io/part-of: abevalflow
+spec:
+  selector:
+    app.kubernetes.io/name: openshell-saw-agent
+    vm.kubevirt.io/name: openshell-saw-agent
+  ports:
+    - name: gateway
+      port: 17670
+      protocol: TCP
+      targetPort: 17670
diff --git a/config/litellm/configmap.yaml b/config/litellm/configmap.yaml
index b352a41..c01eca0 100644
--- a/config/litellm/configmap.yaml
+++ b/config/litellm/configmap.yaml
@@ -2,7 +2,6 @@ apiVersion: v1
 kind: ConfigMap
 metadata:
   name: litellm-config
-  namespace: gz-forge-eval
   labels:
     app.kubernetes.io/part-of: abevalflow
     app.kubernetes.io/name: litellm
@@ -15,6 +14,25 @@ data:
           api_base: os.environ/INFERENCE_ENDPOINT_URL
           api_key: os.environ/INFERENCE_API_KEY
 
+      - model_name: rits/zai-org/GLM-5-3-Flash
+        litellm_params:
+          model: openai/rits/zai-org/GLM-5-3-Flash
+          api_base: os.environ/INFERENCE_ENDPOINT_URL
+          api_key: os.environ/INFERENCE_API_KEY
+
+      # Judge route. The LLM judges call /v1/messages (Anthropic dialect) —
+      # score.py has no OpenAI dialect. For an "openai/" model LiteLLM converts
+      # that to the upstream's /responses API, which this upstream rejects with
+      # 403 "Authentication parameters missing", so every scoring judge errors
+      # and mean_reward is unavailable. Routing the same upstream as hosted_vllm
+      # keeps the call on /chat/completions. Select it with
+      # aeh-judge-model-override=judge-glm-5-3.
+      - model_name: judge-glm-5-3
+        litellm_params:
+          model: hosted_vllm/rits/zai-org/GLM-5-3-Flash
+          api_base: os.environ/INFERENCE_ENDPOINT_URL
+          api_key: os.environ/INFERENCE_API_KEY
+
     litellm_settings:
       drop_params: true
       set_verbose: false
diff --git a/config/litellm/deployment.yaml b/config/litellm/deployment.yaml
index 8146bab..46c015a 100644
--- a/config/litellm/deployment.yaml
+++ b/config/litellm/deployment.yaml
@@ -2,7 +2,6 @@ apiVersion: apps/v1
 kind: Deployment
 metadata:
   name: litellm
-  namespace: gz-forge-eval
   labels:
     app.kubernetes.io/part-of: abevalflow
     app.kubernetes.io/name: litellm
diff --git a/config/litellm/service.yaml b/config/litellm/service.yaml
index bc12954..885c92b 100644
--- a/config/litellm/service.yaml
+++ b/config/litellm/service.yaml
@@ -2,7 +2,6 @@ apiVersion: v1
 kind: Service
 metadata:
   name: litellm
-  namespace: gz-forge-eval
   labels:
     app.kubernetes.io/part-of: abevalflow
     app.kubernetes.io/name: litellm
diff --git a/config/litellm/serviceaccount.yaml b/config/litellm/serviceaccount.yaml
index 5ea3455..78ef60d 100644
--- a/config/litellm/serviceaccount.yaml
+++ b/config/litellm/serviceaccount.yaml
@@ -2,7 +2,6 @@ apiVersion: v1
 kind: ServiceAccount
 metadata:
   name: litellm
-  namespace: gz-forge-eval
   labels:
     app.kubernetes.io/part-of: abevalflow
     app.kubernetes.io/name: litellm
diff --git a/pipeline/pipelines/ci-pipeline-openshell.yaml b/pipeline/pipelines/ci-pipeline-openshell.yaml
index 07a41c5..65d7d85 100644
--- a/pipeline/pipelines/ci-pipeline-openshell.yaml
+++ b/pipeline/pipelines/ci-pipeline-openshell.yaml
@@ -2,9 +2,9 @@ apiVersion: tekton.dev/v1
 kind: Pipeline
 metadata:
   name: abevalflow-pipeline-openshell
-  namespace: ab-eval-flow
   labels:
     app.kubernetes.io/name: abevalflow
+    app.kubernetes.io/part-of: abevalflow
     app.kubernetes.io/version: "2.0"
     app.kubernetes.io/profile: aeh-openshell-openclaw
 spec:
@@ -38,7 +38,7 @@ spec:
       description: Must stay aeh_openshell_openclaw for this profile
     - name: llm-model
       type: string
-      default: "rits/zai-org/glm-5-3"
+      default: "rits/zai-org/GLM-5-3-Flash"
     - name: llm-api-base
       type: string
       default: "http://litellm.gz-forge-eval.svc:4000"
@@ -50,7 +50,7 @@ spec:
       default: "http://litellm.ab-eval-flow.svc:4000/v1"
     - name: llm-generation-model
       type: string
-      default: "rits/zai-org/glm-5-3"
+      default: "rits/zai-org/GLM-5-3-Flash"
     - name: enable-ai-generation
       type: string
       default: "false"
@@ -73,10 +73,13 @@ spec:
       default: "-1.0"
     - name: aeh-model-override
       type: string
-      default: ""
+      default: "inference/rits/zai-org/GLM-5-3-Flash"
     - name: aeh-judge-model-override
       type: string
-      default: ""
+      # Judges call /v1/messages (Anthropic dialect); see the judge-glm-5-3
+      # entry in config/litellm/configmap.yaml for why a plain openai/ model
+      # 403s there and hosted_vllm does not.
+      default: "judge-glm-5-3"
     - name: aeh-mode
       type: string
       default: "single"
@@ -102,17 +105,20 @@ spec:
       default: "https://github.com/GuyZivRH/agent-eval-harness.git"
     - name: agent-eval-harness-repo-revision
       type: string
-      default: "feat/aeh-openshell-openclaw"
-      description: Harness revision with agent_eval.openshell (feature-branch pin)
+      default: "main"
+      description: >-
+        GuyZivRH/agent-eval-harness main includes the OpenClaw brief-reader and
+        gateway fixes merged in #9. Pin a commit SHA for reproducible runs.
     - name: openshell-cli-version
       type: string
       default: "0.0.116"
     - name: openshell-gateway-endpoint
       type: string
-      default: "https://openshell.openshell.svc.cluster.local:17670"
+      # Installed by config/forge-saw/openshell-eval-alias.yaml in the run namespace.
+      default: "https://openshell:17670"
     - name: openshell-sandbox-image
       type: string
-      default: "ghcr.io/rh-forge/openclaw-saw-agent@sha256:bcc55e9b7a36d5f65e8ffc75962496f8b3617762a4cdb37fd1cf54611b72d41a"
+      default: "ghcr.io/rh-forge/openclaw-saw-agent@sha256:b47b92a6b3fd03327c1f2093a5c28aba0fdf3cb620e9154335688900191fe2b9"
     - name: openshell-provider
       type: string
       default: "forge-ai-gateway,m365-read-intervm,slack-read-proxy,drafts-service-agent"
diff --git a/pipeline/runs/openshell-openclaw-pipelinerun.yaml b/pipeline/runs/openshell-openclaw-pipelinerun.yaml
index f46241e..fd5f8c2 100644
--- a/pipeline/runs/openshell-openclaw-pipelinerun.yaml
+++ b/pipeline/runs/openshell-openclaw-pipelinerun.yaml
@@ -1,10 +1,20 @@
 # OpenShell OpenClaw profile — prepare → evaluate → analyze → store (no test cube).
-# Apply the Pipeline first: oc apply -f pipeline/pipelines/ci-pipeline-openshell.yaml
+#
+#   oc apply  -n <namespace> -f pipeline/pipelines/ci-pipeline-openshell.yaml
+#   oc create -n <namespace> -f pipeline/runs/openshell-openclaw-pipelinerun.yaml
+#
+# Service endpoints below use short names, which resolve inside the namespace
+# the run is created in, so this file is namespace-agnostic.
 # Do not put M365 tokens in this file; use Secret openshell-credentials in the namespace.
 apiVersion: tekton.dev/v1
 kind: PipelineRun
 metadata:
   generateName: aeh-openshell-openclaw-
+  labels:
+    # Required when NetworkPolicies select on part-of (and recommended always).
+    # Keep pipelineRef.name=abevalflow-pipeline-openshell so tekton.dev/pipeline
+    # also matches the canonical CI NetworkPolicies.
+    app.kubernetes.io/part-of: abevalflow
 spec:
   pipelineRef:
     name: abevalflow-pipeline-openshell
@@ -16,54 +26,52 @@ spec:
       value: "openclaw-forge"
     - name: eval-engine
       value: "aeh_openshell_openclaw"
-    # Until this branch is merged, pass the feature revision so store/parse scripts match.
-    # After merge to main, omit these (Pipeline defaults are main).
+    # This example targets the merged upstream main for both submission and
+    # Pipeline helper source. Replace either value with a commit SHA to pin a run.
     - name: revision
-      value: "feat/aeh-openshell-openclaw"
+      value: "main"
     - name: pipeline-repo-revision
-      value: "feat/aeh-openshell-openclaw"
+      value: "main"
     # [Optional] Add a harness repository or revision for this run; these replace Pipeline defaults.
     # - name: agent-eval-harness-repo-url
     #   value: "<harness-repo-url>"
     # - name: agent-eval-harness-repo-revision
     #   value: "<harness-commit-or-branch>"
     - name: openshell-gateway-endpoint
-      # Certificate-valid hostname, resolved to this namespace's Service below.
-      value: "https://host.containers.internal:17670"
+      # Stable, certificate-valid alias installed by
+      # config/forge-saw/openshell-eval-alias.yaml in the run namespace.
+      value: "https://openshell:17670"
     - name: openshell-mtls-secret
-      value: "openshell-gateway-mtls"
+      value: "openshell-mtls"
     - name: openshell-ai-gateway-ca-secret
-      value: "forge-agent-upstream-tls"
+      value: "forge-ai-gateway-ca"
     # To use another sandbox image, replace the value below with <image@sha256:digest>.
     - name: openshell-sandbox-image
-      value: "ghcr.io/rh-forge/openclaw-saw-agent@sha256:bcc55e9b7a36d5f65e8ffc75962496f8b3617762a4cdb37fd1cf54611b72d41a"
+      value: "ghcr.io/rh-forge/openclaw-saw-agent@sha256:b47b92a6b3fd03327c1f2093a5c28aba0fdf3cb620e9154335688900191fe2b9"
     - name: openshell-provider
       value: "forge-ai-gateway,m365-read-intervm,slack-read-proxy,drafts-service-agent"
     - name: aeh-openshell-image
       value: "registry.access.redhat.com/ubi9/python-311:9.6"
     - name: llm-model
-      value: "rits/zai-org/glm-5-3"
+      value: "rits/zai-org/GLM-5-3-Flash"
     - name: llm-api-base
-      value: "http://litellm.gz-forge-eval.svc.cluster.local:4000"
+      value: "http://litellm:4000"
     - name: llm-api-key
       value: "mock"
     - name: aeh-model-override
-      value: "inference/rits/zai-org/glm-5-3"
+      value: "inference/rits/zai-org/GLM-5-3-Flash"
     - name: aeh-judge-model-override
-      value: "rits/zai-org/glm-5-3"
+      value: "judge-glm-5-3"
     - name: enable-mlflow
       value: "true"
     - name: mlflow-tracking-uri
-      value: "http://abevalflow-mlflow.gz-forge-eval.svc.cluster.local:5000"
+      value: "http://abevalflow-mlflow:5000"
   taskRunTemplate:
     serviceAccountName: pipeline
-    podTemplate:
-      hostAliases:
-        - ip: "172.30.183.80"
-          hostnames: ["host.containers.internal"]
   timeouts:
-    pipeline: 2h0m0s
-    tasks: 1h30m0s
+    pipeline: 3h0m0s
+    tasks: 2h30m0s
+    finally: 15m0s
   workspaces:
     - name: shared-workspace
       volumeClaimTemplate:
diff --git a/pipeline/tasks/eval-stack-probe.yaml b/pipeline/tasks/eval-stack-probe.yaml
new file mode 100644
index 0000000..cee3d4c
--- /dev/null
+++ b/pipeline/tasks/eval-stack-probe.yaml
@@ -0,0 +1,48 @@
+apiVersion: tekton.dev/v1
+kind: Task
+metadata:
+  name: eval-stack-probe
+  # Deliberately not labelled app.kubernetes.io/part-of: abevalflow — the probe
+  # must be seen by NetworkPolicy as an ordinary Tekton pod, otherwise it passes
+  # through the stack's own allow-rules and reports connectivity the pipeline
+  # does not actually have.
+spec:
+  description: >-
+    Probes every dependency an ABEvalFlow pipeline pod needs. Run it before and
+    after NetworkPolicy changes to see exactly which paths are blocked.
+  steps:
+    - name: probe
+      image: registry.access.redhat.com/ubi9/python-311:9.6
+      script: |
+        #!/usr/bin/env bash
+        NS="$(cat /var/run/secrets/kubernetes.io/serviceaccount/namespace 2>/dev/null || echo current)"
+        fail=0
+        probe() { # name url
+          if curl -sS --max-time 8 -o /dev/null -w "%{http_code}" $2 >/tmp/code 2>/tmp/err; then
+            echo "  PASS  $1  (HTTP $(cat /tmp/code))"
+          else
+            echo "  FAIL  $1  -> $(tr -d '\n' </tmp/err | cut -c1-110)"; fail=1
+          fi
+        }
+        tcp_probe() { # name host port
+          if timeout 8 bash -c ": < /dev/tcp/$2/$3" 2>/tmp/tcp-err; then
+            echo "  PASS  $1  (TCP $2:$3)"
+          else
+            echo "  FAIL  $1  -> $(tr -d '\n' </tmp/tcp-err | cut -c1-110)"; fail=1
+          fi
+        }
+        echo "=== eval stack probe from a Tekton pod in $NS ==="
+        probe "mlflow   api"   "http://abevalflow-mlflow.$NS.svc.cluster.local:5000/api/2.0/mlflow/experiments/search?max_results=1"
+        probe "litellm  health" "http://litellm.$NS.svc.cluster.local:4000/health/liveliness"
+        probe "minio    health" "http://minio.$NS.svc.cluster.local:9000/minio/health/live"
+        probe "k8s api"         "--cacert /var/run/secrets/kubernetes.io/serviceaccount/ca.crt https://kubernetes.default.svc/version"
+        probe "github (clone)"  "https://github.com"
+        probe "pypi index"      "https://pypi.org/simple/pyyaml/"
+        probe "python wheels"   "https://files.pythonhosted.org"
+        probe "github releases" "https://release-assets.githubusercontent.com"
+        probe "github raw"      "https://raw.githubusercontent.com"
+        tcp_probe "openshell agent" "openshell-saw-agent-gateway.$NS.svc.cluster.local" 17670
+        tcp_probe "openshell integ" "openshell-saw-integ-gateway.$NS.svc.cluster.local" 17670
+        tcp_probe "postgres" "postgres.$NS.svc.cluster.local" 5432
+        echo "=== $( [ $fail -eq 0 ] && echo 'ALL PASSED' || echo 'FAILURES PRESENT' ) ==="
+        exit $fail
diff --git a/pipeline/tasks/phases/evaluate.yaml b/pipeline/tasks/phases/evaluate.yaml
index 827346c..f562960 100644
--- a/pipeline/tasks/phases/evaluate.yaml
+++ b/pipeline/tasks/phases/evaluate.yaml
@@ -2,7 +2,6 @@ apiVersion: tekton.dev/v1
 kind: Task
 metadata:
   name: evaluate
-  namespace: ab-eval-flow
   labels:
     app.kubernetes.io/name: abevalflow
     app.kubernetes.io/component: phase
@@ -184,7 +183,7 @@ spec:
       default: "https://github.com/GuyZivRH/agent-eval-harness.git"
     - name: agent-eval-harness-repo-revision
       type: string
-      default: "feat/aeh-openshell-openclaw"
+      default: "main"
       description: >-
         Branch or tag of GuyZivRH/agent-eval-harness to clone. Must include the
         OpenShell create keep-alive fix (`--detach` + `sleep infinity`).
@@ -196,7 +195,7 @@ spec:
         Override only if using a different gateway (e.g. forge-saw :17670).
     - name: openshell-sandbox-image
       type: string
-      default: "ghcr.io/rh-forge/openclaw-saw-agent@sha256:bcc55e9b7a36d5f65e8ffc75962496f8b3617762a4cdb37fd1cf54611b72d41a"
+      default: "ghcr.io/rh-forge/openclaw-saw-agent@sha256:b47b92a6b3fd03327c1f2093a5c28aba0fdf3cb620e9154335688900191fe2b9"
       description: >-
         Immutable SAW Chief-of-Staff image with the daily-briefing skill and
         governed M365/Slack tools.
@@ -1817,6 +1816,46 @@ spec:
         PY
           chmod 600 "$HOME/.config/openshell/gateways/$GW_NAME/oidc_token.json"
           echo "OpenShell OIDC token cache populated via client credentials"
+          # A case can run for 15 minutes, while the client-credentials token
+          # expires after roughly five. Refresh the CLI cache throughout this
+          # step so later artifact downloads and sandbox creation stay authorized.
+          OIDC_TOKEN_CACHE="$HOME/.config/openshell/gateways/$GW_NAME/oidc_token.json"
+          (
+            umask 077
+            trap 'rm -f "${OIDC_TOKEN_JSON}.refresh"' EXIT
+            while sleep 120; do
+              if curl -ksSf --max-time 20 \
+                -u "${OPENSHELL_OIDC_CLIENT_ID}:${OPENSHELL_OIDC_CLIENT_SECRET}" \
+                -d grant_type=client_credentials \
+                "${OPENSHELL_OIDC_ISSUER%/}/protocol/openid-connect/token" \
+                >"${OIDC_TOKEN_JSON}.refresh" 2>/dev/null; then
+                python3 - "${OIDC_TOKEN_JSON}.refresh" "$OIDC_TOKEN_CACHE" \
+                  "$OPENSHELL_OIDC_ISSUER" "$OPENSHELL_OIDC_CLIENT_ID" <<'REFRESH' || true
+        import json, os, sys, tempfile, time
+        src, dst, issuer, client_id = sys.argv[1:]
+        with open(src, encoding="utf-8") as stream:
+            payload = json.load(stream)
+        cache = {"access_token": payload["access_token"],
+                 "expires_at": int(time.time()) + int(payload.get("expires_in", 300)),
+                 "issuer": issuer, "client_id": client_id}
+        fd, tmp = tempfile.mkstemp(dir=os.path.dirname(dst), prefix=".oidc-token-")
+        try:
+            with os.fdopen(fd, "w", encoding="utf-8") as stream:
+                json.dump(cache, stream)
+            os.chmod(tmp, 0o600)
+            os.replace(tmp, dst)
+        finally:
+            if os.path.exists(tmp):
+                os.unlink(tmp)
+        REFRESH
+              else
+                echo "WARN: OpenShell OIDC token refresh failed; retrying" >&2
+              fi
+            done
+          ) &
+          OIDC_REFRESH_PID=$!
+          trap 'kill "$OIDC_REFRESH_PID" 2>/dev/null || true' EXIT
+          echo "OpenShell OIDC token refresher started"
           export OPENSHELL_GATEWAY="$GW_NAME"
           echo "OpenShell OIDC gateway configured: ${GW_NAME}"
         else
@@ -2061,6 +2100,13 @@ spec:
         echo "Model: $MODEL"
 
         export AGENT_EVAL_HARNESS_ROOT="$HARNESS_DIR"
+        # A published Forge briefing needs installation identity. Keep this
+        # synthetic fixture scoped to submissions that supply one; an explicit
+        # harness setting takes precedence for other deployments.
+        if [ -z "${AGENT_EVAL_FORGE_USER_FILE:-}" ] && [ -f "$SUBMISSION_DIR/fixtures/USER.md" ]; then
+          export AGENT_EVAL_FORGE_USER_FILE="$SUBMISSION_DIR/fixtures/USER.md"
+          echo "Using submission installation profile fixture"
+        fi
         if [ "$(params.enable-mlflow)" = "true" ] && [ -n "$(params.mlflow-tracking-uri)" ]; then
           export MLFLOW_TRACKING_URI="$(params.mlflow-tracking-uri)"
           echo "MLFLOW_TRACKING_URI set for AEH harness + CI logger"
diff --git a/pipeline/tasks/phases/prepare.yaml b/pipeline/tasks/phases/prepare.yaml
index 33f0f21..7e85c46 100644
--- a/pipeline/tasks/phases/prepare.yaml
+++ b/pipeline/tasks/phases/prepare.yaml
@@ -2,7 +2,6 @@ apiVersion: tekton.dev/v1
 kind: Task
 metadata:
   name: prepare
-  namespace: ab-eval-flow
   labels:
     app.kubernetes.io/name: abevalflow
     app.kubernetes.io/component: phase
@@ -122,16 +121,21 @@ spec:
         echo "=== PREPARE PHASE: Clone Pipeline Repo ==="
         
         PIPELINE_DIR="$(workspaces.source.path)/_pipeline"
+        PIPELINE_REV="$(params.pipeline-repo-revision)"
         if [ -d "$PIPELINE_DIR/.git" ]; then
           echo "Pipeline repo already cloned, updating..."
           cd "$PIPELINE_DIR"
-          git fetch origin "$(params.pipeline-repo-revision)"
-          git checkout "$(params.pipeline-repo-revision)" 2>/dev/null || git checkout FETCH_HEAD
+          git fetch origin "$PIPELINE_REV" --depth 1 2>/dev/null \
+            || git fetch origin "$PIPELINE_REV"
+          git checkout "$PIPELINE_REV" 2>/dev/null || git checkout FETCH_HEAD
         else
-          git clone --depth 1 --branch "$(params.pipeline-repo-revision)" \
-            "$(params.pipeline-repo-url)" "$PIPELINE_DIR"
+          if ! git clone --depth 1 --branch "$PIPELINE_REV" \
+            "$(params.pipeline-repo-url)" "$PIPELINE_DIR" 2>/dev/null; then
+            git clone "$(params.pipeline-repo-url)" "$PIPELINE_DIR"
+            git -C "$PIPELINE_DIR" checkout "$PIPELINE_REV"
+          fi
         fi
-        echo "Pipeline repo ready at $(params.pipeline-repo-revision)"
+        echo "Pipeline repo ready at $(git -C "$PIPELINE_DIR" rev-parse --short HEAD)"
 
     # Step 3: Generate tests (conditional)
     - name: generate-tests
diff --git a/pipeline/tasks/post/analyze-and-check-degradation.yaml b/pipeline/tasks/post/analyze-and-check-degradation.yaml
index 3140f6e..10ab0ae 100644
--- a/pipeline/tasks/post/analyze-and-check-degradation.yaml
+++ b/pipeline/tasks/post/analyze-and-check-degradation.yaml
@@ -2,7 +2,6 @@ apiVersion: tekton.dev/v1
 kind: Task
 metadata:
   name: analyze-and-check-degradation
-  namespace: ab-eval-flow
   labels:
     app.kubernetes.io/name: abevalflow
 spec:
diff --git a/pipeline/tasks/post/cleanup_pvc.yaml b/pipeline/tasks/post/cleanup_pvc.yaml
index 9ba6c65..c050958 100644
--- a/pipeline/tasks/post/cleanup_pvc.yaml
+++ b/pipeline/tasks/post/cleanup_pvc.yaml
@@ -2,7 +2,6 @@ apiVersion: tekton.dev/v1
 kind: Task
 metadata:
   name: cleanup-pvc
-  namespace: ab-eval-flow
 spec:
   description: >-
     Deletes the workspace PVC after a pipeline run completes.
@@ -11,36 +10,38 @@ spec:
     - name: pipelinerun-name
       type: string
       description: Name of the PipelineRun whose PVCs should be cleaned up
-  steps:
-    - name: delete-pvc
-      image: registry.redhat.io/openshift4/ose-cli:latest
-      script: |
-        #!/usr/bin/env bash
-        set -u
+  steps:
+    - name: delete-pvc
+      image: image-registry.openshift-image-registry.svc:5000/openshift/cli:latest
+      onError: continue
+      script: |
+        #!/usr/bin/env bash
+        set -uo pipefail
 
         # Get current namespace from service account
         NAMESPACE=$(cat /var/run/secrets/kubernetes.io/serviceaccount/namespace)
 
         echo "Cleaning up PVCs for PipelineRun $(params.pipelinerun-name) in namespace ${NAMESPACE}..."
-        PVC_LIST=""
-        for attempt in 1 2 3 4 5; do
-          if PVC_LIST=$(oc get pvc -n "${NAMESPACE}" \
-            -o jsonpath="{range .items[?(@.metadata.ownerReferences[0].name=='$(params.pipelinerun-name)')]}{.metadata.name}{'\\n'}{end}" 2>/tmp/cleanup-pvc.err); then
+        pvcs=""
+        for attempt in 1 2 3; do
+          if pvcs=$(oc get pvc -n "${NAMESPACE}" --request-timeout=20s \
+            -o jsonpath='{range .items[?(@.metadata.ownerReferences[0].name=="$(params.pipelinerun-name)")]}{.metadata.name}{"\n"}{end}'); then
             break
           fi
-          echo "WARNING: Kubernetes API unavailable (attempt ${attempt}/5); retrying..." >&2
-          sleep 10
+          echo "WARNING: PVC lookup attempt ${attempt}/3 failed"
+          sleep $((attempt * 5))
         done
-        if [ -z "${PVC_LIST}" ] && [ -s /tmp/cleanup-pvc.err ]; then
-          cat /tmp/cleanup-pvc.err >&2
-          echo "WARNING: PVC cleanup deferred because the Kubernetes API was unreachable" >&2
+
+        if [ -z "$pvcs" ]; then
+          echo "No owned PVCs found, or the API remained unavailable; cleanup is best-effort"
           exit 0
         fi
+
         while read -r pvc; do
-          if [ -n "${pvc}" ]; then
-            echo "Deleting PVC: ${pvc}"
-            oc delete pvc "${pvc}" -n "${NAMESPACE}" --wait=false || \
-              echo "WARNING: failed to delete ${pvc}; it may be retried later" >&2
+          if [ -n "$pvc" ]; then
+            echo "Deleting PVC: $pvc"
+            oc delete pvc "$pvc" -n "${NAMESPACE}" --wait=false --request-timeout=20s \
+              || echo "WARNING: failed to delete PVC $pvc"
           fi
-        done <<< "${PVC_LIST}"
+        done <<< "$pvcs"
         echo "PVC cleanup complete"
diff --git a/pipeline/tasks/post/store.yaml b/pipeline/tasks/post/store.yaml
index 32d27aa..bf0dcc3 100644
--- a/pipeline/tasks/post/store.yaml
+++ b/pipeline/tasks/post/store.yaml
@@ -2,7 +2,6 @@ apiVersion: tekton.dev/v1
 kind: Task
 metadata:
   name: store
-  namespace: ab-eval-flow
 spec:
   description: >-
     Persists A/B evaluation results and publishes artifacts. Step 1 writes
diff --git a/submissions/openclaw-forge/fixtures/USER.md b/submissions/openclaw-forge/fixtures/USER.md
new file mode 100644
index 0000000..e88fe80
--- /dev/null
+++ b/submissions/openclaw-forge/fixtures/USER.md
@@ -0,0 +1,7 @@
+# Synthetic evaluation identity
+
+- Display name: Alex Rivera
+- Role: Engineering executive
+- Initials: AR
+- Primary email: tbx-demo2@dev.mscloud.ibm.com
+- Time zone: America/New_York
diff --git a/tests/test_openshell_pipeline_profile.py b/tests/test_openshell_pipeline_profile.py
index b44f222..e3640ff 100644
--- a/tests/test_openshell_pipeline_profile.py
+++ b/tests/test_openshell_pipeline_profile.py
@@ -42,14 +42,60 @@ def test_openshell_defaults(self):
         assert defaults["aeh-runner"] == "openshell"
         assert defaults["aeh-openshell-image"] == "registry.access.redhat.com/ubi9/python-311:9.6"
         assert defaults["openshell-sandbox-image"] == (
-            "ghcr.io/rh-forge/openclaw-saw-agent@sha256:bcc55e9b7a36d5f65e8ffc75962496f8b3617762a4cdb37fd1cf54611b72d41a"
+            "ghcr.io/rh-forge/openclaw-saw-agent@sha256:b47b92a6b3fd03327c1f2093a5c28aba0fdf3cb620e9154335688900191fe2b9"
         )
+        assert defaults["llm-model"] == "rits/zai-org/GLM-5-3-Flash"
+        assert defaults["llm-generation-model"] == "rits/zai-org/GLM-5-3-Flash"
+        assert defaults["aeh-model-override"] == "inference/rits/zai-org/GLM-5-3-Flash"
+        # A plain openai/ judge model 403s on /v1/messages; the judge-glm-5-3
+        # hosted_vllm route in config/litellm/configmap.yaml stays on
+        # /chat/completions. See that file for the full rationale.
+        assert defaults["aeh-judge-model-override"] == "judge-glm-5-3"
+        # Alias Service from config/forge-saw/openshell-eval-alias.yaml; keeps
+        # the default namespace-agnostic.
+        assert defaults["openshell-gateway-endpoint"] == "https://openshell:17670"
         assert defaults["enable-mlflow"] == "true"
-        assert defaults["mlflow-tracking-uri"] == ("http://abevalflow-mlflow.gz-forge-eval.svc.cluster.local:5000")
         analyze = next(t for t in _load(PIPELINE)["spec"]["tasks"] if t["name"] == "analyze")
         scan = next(p for p in analyze["params"] if p["name"] == "security-scan-mode")
         assert scan["value"] == "disabled"
 
+    def test_litellm_configmap_has_flash_and_judge_routes(self):
+        config = _load(REPO / "config" / "litellm" / "configmap.yaml")
+        models = yaml.safe_load(config["data"]["config.yaml"])["model_list"]
+        names = {entry["model_name"] for entry in models}
+        assert "rits/zai-org/GLM-5-3-Flash" in names
+        assert "judge-glm-5-3" in names
+        judge = next(e for e in models if e["model_name"] == "judge-glm-5-3")
+        assert judge["litellm_params"]["model"].startswith("hosted_vllm/")
+
+    def test_gateway_alias_service_selects_agent_vmi(self):
+        alias = _load(REPO / "config" / "forge-saw" / "openshell-eval-alias.yaml")
+        assert alias["spec"]["selector"] == {
+            "app.kubernetes.io/name": "openshell-saw-agent",
+            "vm.kubevirt.io/name": "openshell-saw-agent",
+        }
+        ports = alias["spec"]["ports"]
+        assert ports == [{"name": "gateway", "port": 17670, "protocol": "TCP", "targetPort": 17670}]
+
+    def test_example_run_is_namespace_agnostic(self):
+        run = _load(REPO / "pipeline" / "runs" / "openshell-openclaw-pipelinerun.yaml")
+        params = {p["name"]: p["value"] for p in run["spec"]["params"]}
+        for name in ("openshell-gateway-endpoint", "llm-api-base", "mlflow-tracking-uri"):
+            value = params[name]
+            assert ".svc.cluster.local" not in value, f"{name} pins a namespace: {value}"
+        pod_template = run["spec"]["taskRunTemplate"].get("podTemplate", {})
+        assert "hostAliases" not in pod_template
+        timeouts = run["spec"]["timeouts"]
+        assert timeouts["pipeline"] == "3h0m0s"
+        assert timeouts["tasks"] == "2h30m0s"
+
+    def test_cleanup_task_is_best_effort(self):
+        task = _load(REPO / "pipeline" / "tasks" / "post" / "cleanup_pvc.yaml")
+        step = task["spec"]["steps"][0]
+        assert step.get("onError") == "continue"
+        assert "--request-timeout=20s" in step["script"]
+        assert step["image"].startswith("image-registry.openshift-image-registry.svc:5000/")
+
     def test_evaluate_openshell_step_logs_mlflow(self):
         evaluate = (REPO / "pipeline" / "tasks" / "phases" / "evaluate.yaml").read_text()
         openshell = evaluate.split("name: aeh-openshell-eval", 1)[1].split("name: aeh-pairwise", 1)[0]
@@ -80,6 +126,23 @@ def test_openshell_eval_uses_llm_param_instead_of_inference_secret(self):
             assert env["value"] == "$(params.llm-api-key)"
             assert "valueFrom" not in env
 
+    def test_forge_briefing_has_installation_user_fixture(self):
+        fixture = REPO / "submissions" / "openclaw-forge" / "fixtures" / "USER.md"
+        fields = {}
+        for line in fixture.read_text().splitlines():
+            if line.startswith("- ") and ": " in line:
+                key, value = line[2:].split(": ", 1)
+                fields[key.lower()] = value.strip()
+        assert all(fields.get(key) and not fields[key].startswith("<") for key in ("display name", "role", "initials"))
+        scene = _load(REPO / "submissions" / "openclaw-forge" / "scenes" / "monday-acquisition.yaml")
+        assert fields["primary email"] == scene["m365"]["user"]
+
+        task = _load(REPO / "pipeline" / "tasks" / "phases" / "evaluate.yaml")
+        step = next(s for s in task["spec"]["steps"] if s["name"] == "aeh-openshell-eval")
+        script = step["script"]
+        assert 'AGENT_EVAL_FORGE_USER_FILE="$SUBMISSION_DIR/fixtures/USER.md"' in script
+        assert script.index('[ -f "$SUBMISSION_DIR/fixtures/USER.md" ]') < script.index("scripts/run_aeh.py")
+
     def test_harbor_profiles_still_include_test(self):
         for path in (CI, CI_DEV):
             names = _task_names(_load(path)["spec"])