perf: preserve conversation prefixes across agent turns

artemis/agents/operator/operator.py

@@ -330,6 +330,8 @@ class OperatorNode:
             image_scrub_depth_relaxed=getattr(cfg, "image_scrub_depth_relaxed", None),
             context_budget_tokens=getattr(cfg, "context_budget_tokens", None),
             start_ratio=getattr(cfg, "start_ratio", None),
+            soft_ratio=getattr(cfg, "soft_ratio", 0.7),
+            hard_ratio=getattr(cfg, "hard_ratio", 0.9),
             session_start=session_start if isinstance(session_start, (int, float)) else None,
         )
         # L2/L3 chunk compression (M3) rides the same flag; without a
@@ -392,7 +394,7 @@ class OperatorNode:
         """
         from langchain_core.messages import HumanMessage
 
-        from artemis.memory.transcript import mark_ephemeral
+        from artemis.memory.transcript import OBSERVATION_SCOPE_RULE, mark_ephemeral
 
         ledger = self._ensure_transcript_ledger(state)
 
@@ -407,6 +409,7 @@ class OperatorNode:
             if getattr(self.ctx, "plugins", None):
                 prefix.append(self.ctx.plugins.prompt())
             prefix.append(static_text)
+            prefix.append(OBSERVATION_SCOPE_RULE)
             ledger.set_static_prefix([SystemMessage(content="\n".join(prefix))])
 
         # 2. F region cold start: an empty ledger over an existing step record
@@ -496,7 +499,7 @@ class OperatorNode:
                 tail_content.append(part)
         tail = HumanMessage(content=tail_content)
         # Per-turn notices (reminders, hints, guidance, incident blocks) are
-        # deleted by the scrub edge once the turn leaves the recent window.
+        # deleted only when phase compaction rewrites the recent window.
         if builder.ephemeral_indices:
             mark_ephemeral(tail, builder.ephemeral_indices)
 
@@ -635,12 +638,8 @@ class OperatorNode:
                     transcript_ledger.record_prompt_tokens(
                         usage.get("prompt_tokens"), messages=current_messages
                     )
-            elif hasattr(transcript_ledger, "record_prompt_tokens"):
-                # No usage metadata from the provider: record an estimate so
-                # the start gate and thresholds still see a context base.
-                from artemis.memory.transcript import estimate_prompt_tokens
-
-                transcript_ledger.record_prompt_tokens(estimate_prompt_tokens(current_messages))
+            # Without provider usage the ledger estimates the complete next
+            # prompt at render time; do not label that estimate as a measurement.
 
             if response.content:
                 if isinstance(response.content, str):