feat(backend): BACKEND_SLOT_PINNING — per-phase engine slot ownership

Each pipeline phase owns one llama-server slot (steward 0, orchestrator
1, synthesizer 2), carried as id_slot in extra_body through the same
mechanism tool_choice already uses, so the phase's stable prompt prefix
stays in that slot's KV cache and a turn re-prefills only its new
tokens. Off by default; a no-op on the Claude backend and ignored by
Ollama, so the flag is safe on any backend and the cutover itself stays
a pure env swap.

The merge helper preserves existing extra_body keys — mutation-checked
(dropping the merge fails exactly the test written for it).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

CHANGELOG.md

@@ -7,6 +7,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Added
+
+- `BACKEND_SLOT_PINNING` (default off): pins each pipeline phase to a
+  llama-server slot (steward=0, orchestrator=1, synthesizer=2) so a
+  phase's stable prompt prefix stays in that slot's KV cache. No-op on
+  Claude; ignored by Ollama.
+
 ## [2.5.0] - 2026-09-11
 
 ### Added

src/agents/steward/agent.py

@@ -162,15 +162,20 @@ class StewardAgent:
         OpenAI-compatible server (Ollama, llama-server) can sit behind
         OLLAMA_HOST without this method knowing which.
         """
+        payload: dict = {
+            "model": self.ollama_model,
+            "messages": [{"role": "user", "content": prompt}],
+            "temperature": 0.3,  # Lower = more consistent
+            "top_p": 0.9,
+        }
+        if config.BACKEND_SLOT_PINNING:
+            # The steward owns engine slot 0 (see BACKEND_SLOT_PINNING)
+            payload["id_slot"] = 0
+
         async with httpx.AsyncClient(timeout=self.timeout) as client:
             response = await client.post(
                 f"{self.ollama_host}/v1/chat/completions",
-                json={
-                    "model": self.ollama_model,
-                    "messages": [{"role": "user", "content": prompt}],
-                    "temperature": 0.3,  # Lower = more consistent
-                    "top_p": 0.9,
-                },
+                json=payload,
             )
 
             response.raise_for_status()

src/agents/tatlock.py

@@ -536,13 +536,13 @@ class TatlockAgent(AgentInterface):
 
         # Run with scoped tools and tracker
         # Force tool_choice to make LLM actually call tools
-        from src.anthropic.model_selector import get_tool_choice_settings
+        from src.anthropic.model_selector import get_tool_choice_settings, with_slot_pinning
 
         result = await scoped_agent.run(
             enriched_message,
             message_history=pydantic_history if pydantic_history else None,
             deps=tool_tracker,
-            model_settings=get_tool_choice_settings(),
+            model_settings=with_slot_pinning(get_tool_choice_settings(), slot=1),
         )
 
         logger.info(
@@ -725,13 +725,13 @@ class TatlockAgent(AgentInterface):
                 pydantic_history.append(ModelResponse(parts=[TextPart(content=content)]))
 
         # Run with scoped tools and tracker
-        from src.anthropic.model_selector import get_tool_choice_settings
+        from src.anthropic.model_selector import get_tool_choice_settings, with_slot_pinning
 
         result = await scoped_agent.run(
             enriched_message,
             message_history=pydantic_history if pydantic_history else None,
             deps=tool_tracker,
-            model_settings=get_tool_choice_settings(),
+            model_settings=with_slot_pinning(get_tool_choice_settings(), slot=1),
         )
 
         # Extract tool calls and results from the agent's messages
@@ -888,9 +888,12 @@ class TatlockAgent(AgentInterface):
                 pydantic_history.append(ModelResponse(parts=[TextPart(content=content)]))
 
         # Run synthesis
+        from src.anthropic.model_selector import with_slot_pinning
+
         result = await synthesis_agent.run(
             synthesis_prompt,
             message_history=pydantic_history if pydantic_history else None,
+            model_settings=with_slot_pinning(None, slot=2),
         )
 
         logger.info(

src/anthropic/model_selector.py

@@ -259,6 +259,28 @@ def get_tool_choice_settings() -> ModelSettings:
         return ModelSettings(extra_body={"tool_choice": "required"})
 
 
+def with_slot_pinning(settings: ModelSettings | None, slot: int) -> ModelSettings | None:
+    """
+    Merge llama-server slot pinning into model settings when enabled.
+
+    Each pipeline phase owns one engine slot (steward=0, orchestrator=1,
+    synthesizer=2), so the phase's stable prompt prefix stays in that
+    slot's KV cache and a turn re-prefills only its new tokens. Off by
+    default (BACKEND_SLOT_PINNING); a no-op on the Claude backend, and
+    Ollama ignores the field, so enabling it is safe on any backend.
+    """
+    if not config.BACKEND_SLOT_PINNING or resolve_backend() == "claude":
+        return settings
+
+    from pydantic_ai.settings import ModelSettings
+
+    merged = dict(settings or {})
+    extra_body = dict(merged.get("extra_body") or {})
+    extra_body["id_slot"] = slot
+    merged["extra_body"] = extra_body
+    return ModelSettings(**merged)
+
+
 def get_sampling_settings(temperature: float) -> ModelSettings:
     """
     Get model_settings with a sampling temperature where the backend allows it.

src/core/config.py

@@ -147,6 +147,10 @@ class Config(BaseSettings):
         default=None,
         description="OpenAI-compatible embeddings host; falls back to OLLAMA_HOST so gen and embed can live on different servers",
     )
+    BACKEND_SLOT_PINNING: bool = Field(
+        default=False,
+        description="Pin each pipeline phase to a llama-server slot (steward=0, orchestrator=1, synthesizer=2) so a phase's stable prompt prefix stays in that slot's KV cache; no-op on Claude, ignored by Ollama",
+    )
 
     # Redis Memory Database
     REDIS_MEMORY_DB: int = Field(default=1, description="Redis database number for memory cache")

tests/anthropic/test_model_selector.py

@@ -91,3 +91,32 @@ class TestGetModelInfo:
         info = model_selector.get_model_info()
         assert info["backend"] == "claude"
         assert info["model"] == config.ANTHROPIC_MODEL
+
+
+class TestWithSlotPinning:
+    def test_disabled_returns_settings_unchanged(self, local_first, monkeypatch):
+        monkeypatch.setattr(config, "BACKEND_SLOT_PINNING", False)
+        base = model_selector.get_tool_choice_settings()
+        assert model_selector.with_slot_pinning(base, slot=1) is base
+        assert model_selector.with_slot_pinning(None, slot=2) is None
+
+    def test_enabled_merges_slot_with_existing_extra_body(self, local_first, monkeypatch):
+        monkeypatch.setattr(config, "BACKEND_SLOT_PINNING", True)
+        settings = model_selector.with_slot_pinning(
+            model_selector.get_tool_choice_settings(), slot=1
+        )
+        extra_body = settings["extra_body"]
+        assert extra_body["id_slot"] == 1
+        # tool_choice from the base settings survives the merge
+        assert extra_body["tool_choice"] == "required"
+
+    def test_enabled_pins_bare_settings(self, local_first, monkeypatch):
+        monkeypatch.setattr(config, "BACKEND_SLOT_PINNING", True)
+        settings = model_selector.with_slot_pinning(None, slot=2)
+        assert settings["extra_body"] == {"id_slot": 2}
+
+    def test_claude_backend_never_pinned(self, local_first, monkeypatch):
+        monkeypatch.setattr(config, "BACKEND_SLOT_PINNING", True)
+        monkeypatch.setattr(config, "PREFER_CLOUD_BACKEND", True)
+        base = model_selector.get_tool_choice_settings()
+        assert model_selector.with_slot_pinning(base, slot=1) is base