add model affinity docs to llm_router guide, config reference, and routing API

docs/routing-api.md

@@ -120,6 +120,49 @@ routing_preferences:
 
 ---
 
+## Model Affinity
+
+In agentic loops where the same session makes multiple LLM calls, send an `X-Model-Affinity` header to pin the routing decision. The first request routes normally and caches the result. All subsequent requests with the same affinity ID return the cached model without re-running routing.
+
+```json
+POST /v1/chat/completions
+X-Model-Affinity: a1b2c3d4-5678-...
+
+{
+  "model": "openai/gpt-4o-mini",
+  "messages": [...]
+}
+```
+
+The routing decision endpoint also supports model affinity:
+
+```json
+POST /routing/v1/chat/completions
+X-Model-Affinity: a1b2c3d4-5678-...
+```
+
+Response when pinned:
+```json
+{
+  "models": ["anthropic/claude-sonnet-4-20250514"],
+  "route": "code generation",
+  "trace_id": "...",
+  "session_id": "a1b2c3d4-5678-...",
+  "pinned": true
+}
+```
+
+Without the header, routing runs fresh every time (no breaking change).
+
+Configure TTL and cache size:
+```yaml
+routing:
+  session_ttl_seconds: 600    # default: 10 min
+  session_max_entries: 10000  # upper limit
+```
+
+---
+
 ## Version Requirements
 
 | Version | Top-level `routing_preferences` |

docs/source/guides/llm_router.rst

@@ -376,6 +376,44 @@ For the canonical Plano Kubernetes deployment (ConfigMap, Secrets, Deployment YA
 `demo README <https://github.com/katanemo/plano/tree/main/demos/llm_routing/model_routing_service/README.md>`_.
 
 
+.. _model_affinity:
+
+Model Affinity
+--------------
+
+In agentic loops — where a single user request triggers multiple LLM calls through tool use — Plano's router classifies each turn independently. Because successive prompts differ in intent (tool selection looks like code generation, reasoning about results looks like analysis), the router may select different models mid-session. This causes behavioral inconsistency and invalidates provider-side KV caches, increasing both latency and cost.
+
+**Model affinity** pins the routing decision for the duration of a session. Send an ``X-Model-Affinity`` header with any string identifier (typically a UUID). The first request routes normally and caches the result. All subsequent requests with the same affinity ID skip routing and reuse the cached model.
+
+.. code-block:: python
+
+    import uuid
+    from openai import OpenAI
+
+    client = OpenAI(base_url="http://localhost:12000/v1", api_key="EMPTY")
+    affinity_id = str(uuid.uuid4())
+
+    # Every call in the loop uses the same header
+    response = client.chat.completions.create(
+        model="gpt-4o-mini",
+        messages=messages,
+        tools=tools,
+        extra_headers={"X-Model-Affinity": affinity_id},
+    )
+
+Without the header, routing runs fresh on every request — no behavior change for existing clients.
+
+**Configuration:**
+
+.. code-block:: yaml
+
+    routing:
+      session_ttl_seconds: 600    # How long affinity lasts (default: 10 min)
+      session_max_entries: 10000  # Max cached sessions (upper limit: 10000)
+
+To start a new routing decision (e.g., when the agent's task changes), generate a new affinity ID.
+
+
 Combining Routing Methods
 -------------------------
 

docs/source/resources/includes/plano_config_full_reference.yaml

@@ -174,6 +174,11 @@ overrides:
   # Model used for agent orchestration (must be listed in model_providers)
   agent_orchestration_model: Plano-Orchestrator
 
+# Model affinity — pin routing decisions for agentic loops
+routing:
+  session_ttl_seconds: 600    # How long a pinned session lasts (default: 600s / 10 min)
+  session_max_entries: 10000  # Max cached sessions before eviction (upper limit: 10000)
+
 # State storage for multi-turn conversation history
 state_storage:
   type: memory            # "memory" (in-process) or "postgres" (persistent)