release v2.6.0

The cutover's first real lesson ships: tool_choice "required" was a
workaround for a backend that ignored it, and became an unbreakable
tool loop on a backend that obeys it every request (~80 s arithmetic
turns, observed). The health check now learns which server answers
behind OLLAMA_HOST (/props is llama-server's own surface) and only
Ollama gets the advisory nudge. Probed unforced on llama-server: 3/3
tool calls via the --jinja template.

Also carries BACKEND_SLOT_PINNING (default off) for the P3 pilot.

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

src/anthropic/model_selector.py

@@ -28,6 +28,11 @@ logger = get_logger(__name__)
 # Cached health check results (set once at startup)
 _claude_available: bool | None = None
 _ollama_available: bool | None = None
+# Which server answers behind OLLAMA_HOST: "llama-server" or "ollama".
+# They disagree on tool_choice semantics (see get_tool_choice_settings),
+# so the health check probes /props — served by llama-server only — and
+# caches the answer here.
+_local_flavor: str | None = None
 
 
 async def check_ollama_health() -> bool:
@@ -40,7 +45,7 @@ async def check_ollama_health() -> bool:
     Returns:
         True if Ollama is reachable and OLLAMA_DEFAULT_MODEL is pulled.
     """
-    global _ollama_available
+    global _ollama_available, _local_flavor
 
     host = str(config.OLLAMA_HOST).rstrip("/")
     model = config.OLLAMA_DEFAULT_MODEL
@@ -53,12 +58,20 @@ async def check_ollama_health() -> bool:
             response.raise_for_status()
             names = [m.get("id", "") for m in response.json().get("data", [])]
 
+            # /props is llama-server's own surface; Ollama 404s it.
+            try:
+                props = await client.get(f"{host}/props")
+                _local_flavor = "llama-server" if props.status_code == 200 else "ollama"
+            except httpx.HTTPError:
+                _local_flavor = "ollama"
+
         if model in names or f"{model}:latest" in names:
             _ollama_available = True
             logger.info(
                 "ollama_health_check_passed",
                 host=host,
                 model=model,
+                flavor=_local_flavor,
             )
             return True
 
@@ -244,19 +257,30 @@ def get_model(prefer_cloud: bool | None = None) -> AnthropicModel | OpenAIChatMo
 
 def get_tool_choice_settings() -> ModelSettings:
     """
-    Get model_settings for forcing tool calls on the first request.
+    Get model_settings for tool calling on the orchestration phase.
 
-    For Claude: PydanticAI handles tool_choice natively, so no extra_body needed.
-    For Ollama: Pass tool_choice="required" via extra_body to force tool calling.
+    Claude: PydanticAI handles tool_choice natively — no extra_body.
+
+    Ollama: tool_choice="required" via extra_body. Advisory there (Ollama
+    ignores it), but it nudges gemma4 to actually call tools, which the
+    persona-suppression gotcha made necessary.
+
+    llama-server: NO tool_choice. It enforces "required" on every request
+    in the run, so after a tool returns, the next generation is again
+    forced to call a tool — an unbreakable tool loop (~80 s turns,
+    observed at cutover). Its --jinja template renders tool definitions
+    the way gemma4 was trained, and the model calls tools reliably
+    unforced (probed 3/3).
     """
     from pydantic_ai.settings import ModelSettings
 
     if resolve_backend() == "claude":
         # PydanticAI's Anthropic model handles tool_choice internally
         return ModelSettings()
-    else:
-        # Ollama needs explicit tool_choice via extra_body
-        return ModelSettings(extra_body={"tool_choice": "required"})
+    if _local_flavor == "llama-server":
+        return ModelSettings()
+    # Ollama needs explicit tool_choice via extra_body
+    return ModelSettings(extra_body={"tool_choice": "required"})
 
 
 def with_slot_pinning(settings: ModelSettings | None, slot: int) -> ModelSettings | None: