security: strip untrusted identity headers when no auth backend configured

config/config.yaml

@@ -1156,6 +1156,7 @@ global:
           max_models: 50
     authz:
       fail_open: false
+      trust_identity_headers: true
       identity:
         user_id_header: x-user-id
         user_groups_header: x-user-groups

e2e/config/config.session-affinity-demo.yaml

@@ -0,0 +1,189 @@
+# Session Affinity Demo Config
+# Demonstrates the route-bouncing problem in multi-turn conversations.
+#
+# Setup:
+#   - "expensive" model: Claude via claude-code-proxy on port 11480
+#   - "cheap" model: qwen2.5:0.5b (0.5B params) on Ollama port 11434
+#
+# Routing logic:
+#   - Complex queries (complexity signal "hard") → Claude (expensive)
+#   - Simple queries (complexity signal "easy") → qwen2.5:0.5b (cheap)
+#
+# The demo scenario:
+#   1. User sends complex coding question → routes to Claude (expensive)
+#   2. Claude responds with great code and asks "want me to add tests?"
+#   3. User says "yes" → VSR sees only "yes" → classifies as easy
+#   4. Routes to qwen2.5:0.5b (cheap) → 0.5B model has NO IDEA what "yes" means
+
+# Disable features we don't need for the demo
+semantic_cache:
+  enabled: false
+
+tools:
+  enabled: false
+
+prompt_guard:
+  enabled: true
+  model_id: "models/mom-jailbreak-classifier"
+  threshold: 0.7
+  use_cpu: true
+jailbreak_mapping_path: "models/mom-jailbreak-classifier/label_mapping.json"
+
+# Jailbreak rules for signal-based detection
+jailbreak_rules:
+  - name: "jailbreak_high"
+    threshold: 0.7
+    description: "High-confidence jailbreak detection"
+
+hallucination_mitigation:
+  enabled: false
+
+# Domain classifier (required by VSR startup)
+classifier:
+  category_model:
+    model_id: "models/mom-domain-classifier"
+    threshold: 0.6
+    use_cpu: true
+    category_mapping_path: "models/mom-domain-classifier/category_mapping.json"
+
+# Embedding models for complexity signal
+embedding_models:
+  qwen3_model_path: "models/mom-embedding-pro"
+  use_cpu: true
+  hnsw_config:
+    model_type: "qwen3"
+    preload_embeddings: true
+    target_dimension: 1024
+    enable_soft_matching: true
+    min_score_threshold: 0.5
+
+# Two backends
+vllm_endpoints:
+  - name: "ollama"
+    address: "127.0.0.1"
+    port: 11434
+    weight: 1
+  - name: "claude-proxy"
+    address: "127.0.0.1"
+    port: 11480
+    weight: 1
+
+# Model-to-endpoint mapping
+model_config:
+  "expensive-model":
+    preferred_endpoints: ["claude-proxy"]
+  "qwen2.5:0.5b":
+    preferred_endpoints: ["ollama"]
+
+# Complexity signal: distinguish hard vs easy prompts
+complexity_rules:
+  - name: "prompt_complexity"
+    threshold: 0.15
+    description: "Classify prompt complexity for routing decisions"
+    hard:
+      candidates:
+        - "Implement a concurrent lock-free data structure with memory ordering guarantees"
+        - "Design a distributed consensus algorithm for fault-tolerant systems"
+        - "Write a compiler optimization pass for loop vectorization"
+        - "Build a real-time stream processing pipeline with exactly-once semantics"
+        - "Implement a B+ tree with concurrent readers and writers"
+        - "Design a garbage collector with generational collection and compaction"
+        - "Write an optimized matrix multiplication kernel with cache tiling"
+        - "Implement a raft consensus protocol with log compaction"
+        - "Explain the mathematical proof of the P vs NP problem"
+        - "Analyze the time complexity of this recursive algorithm with memoization"
+    easy:
+      candidates:
+        - "yes"
+        - "no"
+        - "ok"
+        - "sure"
+        - "thanks"
+        - "got it"
+        - "sounds good"
+        - "please do"
+        - "go ahead"
+        - "that works"
+        - "hello"
+        - "hi"
+        - "what is a variable"
+        - "how do I print hello world"
+        - "what does this error mean"
+
+# Categories (minimal, required by classifier)
+categories:
+  - name: computer_science
+    description: "Computer science and programming"
+    mmlu_categories: ["computer_science"]
+  - name: other
+    description: "General topics"
+    mmlu_categories: ["other"]
+
+# Routing strategy
+strategy: "priority"
+
+# Two decisions: complex → expensive, simple → cheap
+decisions:
+  - name: "jailbreak_block"
+    description: "Block jailbreak attempts"
+    priority: 999
+    rules:
+      operator: "AND"
+      conditions:
+        - type: "jailbreak"
+          name: "jailbreak_high"
+    modelRefs:
+      - model: "qwen2.5:0.5b"
+        use_reasoning: false
+    plugins:
+      - type: "fast_response"
+        configuration:
+          enabled: true
+          message: "Request blocked: jailbreak attempt detected."
+          status_code: 403
+
+  - name: "complex_query"
+    description: "Complex queries that need a powerful model"
+    priority: 200
+    rules:
+      operator: "AND"
+      conditions:
+        - type: "complexity"
+          name: "prompt_complexity:hard"
+    modelRefs:
+      - model: "expensive-model"
+        use_reasoning: false
+
+  - name: "simple_query"
+    description: "Simple queries that a cheap model can handle"
+    priority: 100
+    rules:
+      operator: "AND"
+      conditions:
+        - type: "complexity"
+          name: "prompt_complexity:easy"
+    modelRefs:
+      - model: "qwen2.5:0.5b"
+        use_reasoning: false
+
+  - name: "fallback"
+    description: "Fallback for unmatched queries"
+    priority: 1
+    rules:
+      operator: "AND"
+      conditions:
+        - type: "domain"
+          name: "other"
+    modelRefs:
+      - model: "qwen2.5:0.5b"
+        use_reasoning: false
+
+# Default model when no decision matches
+default_model: "qwen2.5:0.5b"
+
+# Observability
+observability:
+  metrics:
+    enabled: true
+  tracing:
+    enabled: false

e2e/config/envoy.session-affinity-demo.yaml

@@ -0,0 +1,109 @@
+# Envoy config for session affinity demo
+# Stripped down: no ext_authz, just ext_proc + dynamic routing
+static_resources:
+  listeners:
+  - name: listener_0
+    address:
+      socket_address:
+        address: 0.0.0.0
+        port_value: 8801
+    filter_chains:
+    - filters:
+      - name: envoy.filters.network.http_connection_manager
+        typed_config:
+          "@type": type.googleapis.com/envoy.extensions.filters.network.http_connection_manager.v3.HttpConnectionManager
+          stat_prefix: ingress_http
+          access_log:
+          - name: envoy.access_loggers.stdout
+            typed_config:
+              "@type": type.googleapis.com/envoy.extensions.access_loggers.stream.v3.StdoutAccessLog
+              log_format:
+                json_format:
+                  time: "%START_TIME%"
+                  request_method: "%REQ(:METHOD)%"
+                  request_path: "%REQ(X-ENVOY-ORIGINAL-PATH?:PATH)%"
+                  response_code: "%RESPONSE_CODE%"
+                  upstream_host: "%UPSTREAM_HOST%"
+                  selected_model: "%REQ(X-SELECTED-MODEL)%"
+                  destination: "%REQ(X-VSR-DESTINATION-ENDPOINT)%"
+          route_config:
+            name: local_route
+            virtual_hosts:
+            - name: local_service
+              domains: ["*"]
+              routes:
+              - match:
+                  prefix: "/"
+                route:
+                  cluster: dynamic_backend
+                  timeout: 300s
+          http_filters:
+          - name: envoy.filters.http.ext_proc
+            typed_config:
+              "@type": type.googleapis.com/envoy.extensions.filters.http.ext_proc.v3.ExternalProcessor
+              grpc_service:
+                envoy_grpc:
+                  cluster_name: extproc_service
+              allow_mode_override: true
+              processing_mode:
+                request_header_mode: "SEND"
+                response_header_mode: "SEND"
+                request_body_mode: "BUFFERED"
+                response_body_mode: "BUFFERED"
+                request_trailer_mode: "SKIP"
+                response_trailer_mode: "SKIP"
+              failure_mode_allow: true
+              message_timeout: 300s
+          - name: envoy.filters.http.router
+            typed_config:
+              "@type": type.googleapis.com/envoy.extensions.filters.http.router.v3.Router
+              suppress_envoy_headers: true
+          http2_protocol_options:
+            max_concurrent_streams: 100
+          stream_idle_timeout: "300s"
+          request_timeout: "300s"
+          common_http_protocol_options:
+            idle_timeout: "300s"
+
+  clusters:
+  - name: extproc_service
+    connect_timeout: 300s
+    type: STATIC
+    lb_policy: ROUND_ROBIN
+    typed_extension_protocol_options:
+      envoy.extensions.upstreams.http.v3.HttpProtocolOptions:
+        "@type": type.googleapis.com/envoy.extensions.upstreams.http.v3.HttpProtocolOptions
+        explicit_http_config:
+          http2_protocol_options:
+            connection_keepalive:
+              interval: 300s
+              timeout: 300s
+    load_assignment:
+      cluster_name: extproc_service
+      endpoints:
+      - lb_endpoints:
+        - endpoint:
+            address:
+              socket_address:
+                address: 127.0.0.1
+                port_value: 50051
+
+  # Dynamic backend using original destination (VSR sets x-vsr-destination-endpoint)
+  - name: dynamic_backend
+    connect_timeout: 300s
+    type: ORIGINAL_DST
+    lb_policy: CLUSTER_PROVIDED
+    original_dst_lb_config:
+      use_http_header: true
+      http_header_name: "x-vsr-destination-endpoint"
+    typed_extension_protocol_options:
+      envoy.extensions.upstreams.http.v3.HttpProtocolOptions:
+        "@type": type.googleapis.com/envoy.extensions.upstreams.http.v3.HttpProtocolOptions
+        explicit_http_config:
+          http_protocol_options: {}
+
+admin:
+  address:
+    socket_address:
+      address: "127.0.0.1"
+      port_value: 19000

e2e/config/router-runtime.json

@@ -0,0 +1,6 @@
+{
+  "phase": "ready",
+  "ready": true,
+  "message": "Router models are ready. Starting router services...",
+  "updated_at": "2026-03-06T18:09:13Z"
+}