perf: preserve conversation prefixes across agent turns

artemis/services/llm.py

@@ -58,6 +58,7 @@ from artemis.llm.reliability import (
     retry_policy_for,
 )
 from artemis.llm.messages import adapt_responses_messages
+from artemis.llm.prefix_cache import PromptPrefixTracker
 from artemis.llm.router import ModelEndpoint, ModelFactory, ModelProvider
 from artemis.services.token_meter import record_llm_usage
 from artemis.llm.usage import usage_from_message
@@ -671,6 +672,27 @@ class RobustChatModelWrapper:
         record_llm_usage(engine, response, source=self._endpoint_key())
 
     def _diagnostics(self, args, kwargs):
+        prefix = None
+        if self.ctx is not None:
+            tracker = getattr(self.ctx, "prompt_prefix_tracker", None)
+            if not isinstance(tracker, PromptPrefixTracker):
+                tracker = PromptPrefixTracker()
+                self.ctx.prompt_prefix_tracker = tracker
+            bound_options = getattr(self.base_model, "kwargs", {})
+            options = {
+                **(bound_options if isinstance(bound_options, dict) else {}),
+                **{
+                    key: kwargs[key]
+                    for key in ("tools", "tool_choice", "response_format")
+                    if key in kwargs
+                },
+            }
+            identity = f"{kwargs.get('prompt_cache_key')}:{self.role}:{self._endpoint_key()}"
+            prefix = tracker.observe(identity, args[0] if args else kwargs.get("input"), options)
+            ledger = getattr(self.ctx, "transcript_ledger", None)
+            if ledger is not None:
+                prefix["context_phase"] = getattr(ledger, "context_phase", 0)
+                prefix["last_compaction"] = getattr(ledger, "last_compaction", None)
         return {
             "provider": self._provider_value(),
             "model": self.endpoint.model_name if self.endpoint else None,
@@ -680,6 +702,7 @@ class RobustChatModelWrapper:
             "is_fallback": self.is_fallback,
             "images": input_images(args, kwargs),
             "prompt_cache_key": kwargs.get("prompt_cache_key"),
+            "prefix_cache": prefix,
         }
 
     async def _invoke_attempt(self, *args, **kwargs):