perf: preserve conversation prefixes across agent turns

artemis/agents/flash/runner.py

@@ -18,9 +18,9 @@ Executes autonomous reactive mobile workflows through OpenAI Responses.
 
 Conversation shape (shared with the Pro operator, history redesign §3.2):
 the prompt is built every turn from a session :class:`TranscriptLedger` —
-a byte-stable system prefix, the committed earlier turns (append-only; old UI
-lists stripped at depth 1, screenshots resolved to visual summaries at depth
-K, long spans chunk-compressed), and a fresh tail carrying the current
+a byte-stable system prefix, the committed earlier turns (append-only within
+a context phase; old observations and long spans compacted together only at
+pressure thresholds), and a fresh tail carrying the current
 observation under a ``# CURRENT OBSERVATION [T+mm:ss]`` header. Actions run
 synchronously and their tool messages already carry the outcome; only a turn
 whose action failed ends with an ``--- Action Execution Result (T+mm:ss) ---``
@@ -30,7 +30,7 @@ session recording.
 
 Per-turn notices on the tail (the reasoning reminder after a silent turn, the
 no-tool-call nudge, the final-turn warning, the user-guidance wrapper) are
-marked ephemeral: the ledger's scrub edge deletes them once the message leaves
+marked ephemeral: the ledger's compaction pass deletes them when it rewrites
 the recent window, so they never reach the frozen history or a chunk capsule.
 The verbatim body of a user instruction is the exception: it stays in the
 active window until its turn is chunked, whose user lines then carry it on.
@@ -83,7 +83,12 @@ from artemis.llm.router import ModelEndpoint
 from artemis.llm.structured import ParseFailure, parse_structured
 from artemis.mcp.action_executor import McpActionExecutor
 from artemis.mcp.observation import DeviceObservationError, require_observation
-from artemis.memory.transcript import PRO_UI_LIST_MARKER, TranscriptLedger, mark_ephemeral
+from artemis.memory.transcript import (
+    OBSERVATION_SCOPE_RULE,
+    PRO_UI_LIST_MARKER,
+    TranscriptLedger,
+    mark_ephemeral,
+)
 from artemis.services.llm import (
     RobustChatModelWrapper,
     acomplete,
@@ -272,6 +277,8 @@ class FlashRunner:
             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,
         )
         if engine is not None:
@@ -1247,7 +1254,13 @@ class FlashRunner:
         """Installs the static prefix and captures the initial device state."""
         ledger = self._build_ledger()
         ledger.set_static_prefix(
-            [SystemMessage(content=self._render_system_prompt(tools_declaration))]
+            [
+                SystemMessage(
+                    content=self._render_system_prompt(tools_declaration)
+                    + "\n"
+                    + OBSERVATION_SCOPE_RULE
+                )
+            ]
         )
 
         # Capture Initial State (Screenshot + UI Tree)
@@ -1369,7 +1382,7 @@ class FlashRunner:
             )
             pending_notices = []
 
-            # S + F + A + tail: chunk compression and the scrub edge advance here.
+            # S + F + A + tail: history changes only at a compaction boundary.
             messages = ledger.render([tail])
             turn_base = len(messages) - 1
             current_tools = report_only_tools if is_final else self._get_tools()
@@ -1427,7 +1440,7 @@ class FlashRunner:
             # ratio — only from a provider measurement, never from the
             # estimate (same as the Pro operator).
             ledger.record_prompt_tokens(
-                step_token_usage.get("prompt_tokens"),
+                step_token_usage.get("prompt_tokens") if measured else None,
                 messages=messages[:-1] if measured else None,
             )