From 65e4e38a98ffcab0b8c6c75f2294e11f81d59365 Mon Sep 17 00:00:00 2001
From: Joey Yakimowich-Payne <jyapayne@pm.me>
Date: Fri, 13 Mar 2026 21:13:38 -0600
Subject: [PATCH] fix: scale FM window_size to match real API pressure for
 fidelity degradation
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The FidelityManager's internal pressure calculation uses its own tracked
object tokens divided by window_size, which is always tiny compared to
the real context. Temporarily scale window_size so the FM's pressure
matches the actual API input_tokens/window ratio, triggering L0→L1→L2
degradations when context exceeds 50%.
---
 src/mnemosyne/gateway.py | 22 +++++++++++++++-------
 1 file changed, 15 insertions(+), 7 deletions(-)

diff --git a/src/mnemosyne/gateway.py b/src/mnemosyne/gateway.py
index 89fa156..e8057de 100644
--- a/src/mnemosyne/gateway.py
+++ b/src/mnemosyne/gateway.py
@@ -2122,8 +2122,8 @@ def create_app(
         """Update FidelityManager window understanding from actual API usage.
 
         After receiving the response, we know the real input token count.
-        Update the fidelity manager's window_size understanding and schedule
-        degradation if pressure is above NORMAL for the next turn.
+        Scale the FM's window_size so its internal pressure calculation
+        reflects the real API token usage, then trigger degradation.
         """
         input_tokens = usage.get("input_tokens", 0)
         if input_tokens <= 0:
@@ -2132,14 +2132,22 @@ def create_app(
         fm = session.fidelity_manager
         turn = session.token_state.get("turn", 0)
 
-        # The FidelityManager tracks its own token budget via registered objects.
-        # Here we use the real API token count to check if we need proactive degradation.
-        # If real usage exceeds the fidelity window threshold, trigger degradation now
-        # so the NEXT turn benefits from reduced content.
         pressure_ratio = input_tokens / fm.window_size if fm.window_size > 0 else 1.0
 
         if pressure_ratio >= fm.threshold_caution:
-            transitions = fm.degrade(turn)
+            # The FM's internal pressure uses total_tokens()/window_size, but
+            # total_tokens() only counts registered objects (a fraction of the
+            # real context).  Temporarily scale window_size down so the FM's
+            # pressure matches the real API pressure, then restore it.
+            obj_tokens = fm.total_tokens()
+            if obj_tokens > 0:
+                # Set window_size so obj_tokens/window_size == pressure_ratio
+                saved_ws = fm.window_size
+                fm.window_size = max(1, int(obj_tokens / pressure_ratio))
+                transitions = fm.degrade(turn)
+                fm.window_size = saved_ws
+            else:
+                transitions = fm.degrade(turn)
             if transitions:
                 zone = fm.current_pressure()
                 for _obj_id, old_level, new_level in transitions: