From 4ca9c589200ea9020b4502ccc1d4f829dc26595b Mon Sep 17 00:00:00 2001
From: Joey Yakimowich-Payne <jyapayne@pm.me>
Date: Sun, 15 Mar 2026 19:36:05 -0600
Subject: [PATCH] feat: detect context window size dynamically from model ID

Hardcoded 200K window caused 101% pressure at 201K tokens on 1M
models. Now detects model from request payload and sets window_size
accordingly (1M for opus-4-6/sonnet-4-6/sonnet-4-5, 200K for others).
Falls back to 200K for unknown models.
---
 src/mnemosyne/gateway.py | 43 ++++++++++++++++++++++++++++++++++++++++
 1 file changed, 43 insertions(+)

diff --git a/src/mnemosyne/gateway.py b/src/mnemosyne/gateway.py
index 0b903de..1afeb1d 100644
--- a/src/mnemosyne/gateway.py
+++ b/src/mnemosyne/gateway.py
@@ -1137,6 +1137,8 @@ class Session:
             self.page_store,
             log_path=log_dir / f"violations_{sid}.jsonl" if log_dir else None,
         )
+        # Default 200K; updated dynamically from actual model context window
+        # after the first API response via _update_window_size().
         self.fidelity_manager = FidelityManager(window_size=200_000)
         # Maps content keys → fidelity object IDs for lookup during apply
         self._fidelity_content_map: dict[str, str] = {}
@@ -1210,6 +1212,7 @@ class Session:
         # Summary cache: (object_id, FidelityLevel) → summary text
         # Avoids re-summarizing the same content on repeated _apply_fidelity calls
         self._summary_cache: dict[tuple[str, int], str] = {}
+        self._window_size_detected: bool = False
 
     def track_usage(self, usage: dict) -> None:
         """Update token state from API response usage."""
@@ -2337,6 +2340,31 @@ def create_app(
                 file=sys.stderr,
             )
 
+    # Model → context window size mapping.  Used to set the FM's window_size
+    # dynamically after the first API response reveals which model is in use.
+    _MODEL_WINDOW_SIZES: dict[str, int] = {
+        # 1M context models
+        "claude-opus-4-6": 1_000_000,
+        "claude-sonnet-4-6": 1_000_000,
+        "claude-sonnet-4-5": 1_000_000,  # 1M in beta
+        # 200K context models
+        "claude-opus-4-5": 200_000,
+        "claude-sonnet-4": 200_000,
+        "claude-haiku-4-5": 200_000,
+        "claude-haiku-3-5": 200_000,
+    }
+
+    def _detect_window_size(model: str) -> int | None:
+        """Return context window size for a model, or None if unknown."""
+        # Try exact match first, then prefix match (strips date suffix)
+        if model in _MODEL_WINDOW_SIZES:
+            return _MODEL_WINDOW_SIZES[model]
+        # Strip date suffix: "claude-sonnet-4-6-20260217" → "claude-sonnet-4-6"
+        for prefix, size in _MODEL_WINDOW_SIZES.items():
+            if model.startswith(prefix):
+                return size
+        return None
+
     def _update_fidelity_pressure(usage: dict, session: Session) -> None:
         """Update FidelityManager window understanding from actual API usage.
 
@@ -2351,6 +2379,19 @@ def create_app(
         fm = session.fidelity_manager
         turn = session.token_state.get("turn", 0)
 
+        # Detect window size from model on first response
+        if not getattr(session, "_window_size_detected", False):
+            model = session.token_state.get("model", "")
+            detected = _detect_window_size(model)
+            if detected and detected != fm.window_size:
+                fm.window_size = detected
+                print(
+                    f"  {_DIM}[{session.id}] window_size set to {detected:,} "
+                    f"for model {model}{_RESET}",
+                    file=sys.stderr,
+                )
+            session._window_size_detected = True
+
         pressure_ratio = input_tokens / fm.window_size if fm.window_size > 0 else 1.0
 
         if pressure_ratio >= fm.threshold_caution:
@@ -2527,6 +2568,8 @@ def create_app(
 
         req = adapter.normalize_request(payload)
         req = pipeline.run(req)
+        # Store model for dynamic window_size detection
+        session.token_state["model"] = req.model
         duplication_score = _duplication_score(req)
         outgoing_body = adapter.denormalize_request(req)
         if endpoint == "count_tokens":