feat: classify provider prompt-cache strategy (explicit/implicit/none)

EvoScientist/llm/models.py

@@ -333,11 +333,28 @@ def _drop_unsupported_chat_model_kwargs(kwargs: dict[str, Any]) -> None:
             model_kwargs.pop(key, None)
 
 
+_IMPLICIT_CACHE_PROVIDERS = frozenset({"zhipu", "zhipu-code", "siliconflow", "nvidia"})
+
+
+def _cache_strategy(provider: str, model_id: str) -> str:
+    """Return the provider's prompt-cache mechanism.
+
+    ``explicit`` — needs Anthropic-style ``cache_control`` markers (OpenRouter
+    Claude routes only, matching the pre-existing condition).
+    ``implicit`` — provider prefixes-cache automatically; no markers, but the
+    prompt prefix must stay byte-stable for hits (see memory injection order).
+    ``none`` — no cache model to declare.
+    """
+    if provider == "openrouter" and model_id.startswith(("anthropic/", "~anthropic/")):
+        return "explicit"
+    if provider in _IMPLICIT_CACHE_PROVIDERS:
+        return "implicit"
+    return "none"
+
+
 def _supports_openrouter_anthropic_prompt_cache(provider: str, model_id: str) -> bool:
     """Return whether EvoScientist should declare OpenRouter Claude caching."""
-    return provider == "openrouter" and model_id.startswith(
-        ("anthropic/", "~anthropic/")
-    )
+    return _cache_strategy(provider, model_id) == "explicit"
 
 
 def _has_cache_control_override(kwargs: dict[str, Any]) -> bool:

tests/test_llm.py

@@ -965,6 +965,18 @@ class TestThirdPartyRouting:
         assert "thinking" not in call_kwargs
 
 
+def test_cache_strategy_classifies_providers():
+    from EvoScientist.llm.models import _cache_strategy
+
+    assert _cache_strategy("openrouter", "anthropic/claude-sonnet-4-6") == "explicit"
+    assert _cache_strategy("openrouter", "x-ai/grok-4.3") == "none"
+    assert _cache_strategy("anthropic", "claude-sonnet-4-6") == "none"
+    assert _cache_strategy("zhipu", "glm-5.2") == "implicit"
+    assert _cache_strategy("zhipu-code", "glm-5.2") == "implicit"
+    assert _cache_strategy("siliconflow", "glm-5.2") == "implicit"
+    assert _cache_strategy("ollama", "llama3.1:8b") == "none"
+
+
 # =============================================================================
 # Test MiniMax provider
 # =============================================================================