@miller-tech/uap 1.20.9 → 1.20.11

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@miller-tech/uap",
-  "version": "1.20.9",
+  "version": "1.20.11",
   "description": "Autonomous AI agent memory system with CLAUDE.md protocol enforcement",
   "type": "module",
   "main": "dist/index.js",

package/tools/agents/scripts/anthropic_proxy.py

@@ -147,10 +147,10 @@ PROXY_TOOL_STATE_FORCED_BUDGET = int(
 )
 PROXY_TOOL_STATE_AUTO_BUDGET = int(os.environ.get("PROXY_TOOL_STATE_AUTO_BUDGET", "2"))
 PROXY_TOOL_STATE_STAGNATION_THRESHOLD = int(
-    os.environ.get("PROXY_TOOL_STATE_STAGNATION_THRESHOLD", "9")
+    os.environ.get("PROXY_TOOL_STATE_STAGNATION_THRESHOLD", "5")
 )
 PROXY_TOOL_STATE_CYCLE_WINDOW = int(
-    os.environ.get("PROXY_TOOL_STATE_CYCLE_WINDOW", "8")
+    os.environ.get("PROXY_TOOL_STATE_CYCLE_WINDOW", "4")
 )
 PROXY_TOOL_STATE_FINALIZE_THRESHOLD = int(
     os.environ.get("PROXY_TOOL_STATE_FINALIZE_THRESHOLD", "18")
@@ -280,9 +280,6 @@ PROXY_FORCED_TOOL_DAMPENER_REJECTIONS = int(
 PROXY_TOOL_STARVATION_THRESHOLD = int(
     os.environ.get("PROXY_TOOL_STARVATION_THRESHOLD", "5")
 )
-PROXY_CONTEXT_HIGH_RELAXATION_THRESHOLD = float(
-    os.environ.get("PROXY_CONTEXT_HIGH_RELAXATION_THRESHOLD", "0.70")
-)
 PROXY_SESSION_CONTAMINATION_BREAKER = os.environ.get(
     "PROXY_SESSION_CONTAMINATION_BREAKER", "on"
 ).lower() not in {
@@ -2056,7 +2053,7 @@ def _resolve_state_machine_tool_choice(
                 monitor.tool_state_stagnation_streak,
                 monitor.tool_state_review_cycles,
             )
-            return "auto", reason
+            return "required", reason
 
         if monitor.tool_state_forced_budget_remaining <= 0:
             monitor.set_tool_turn_phase("review", reason="forced_budget_exhausted")
@@ -2071,7 +2068,7 @@ def _resolve_state_machine_tool_choice(
                 "TOOL STATE MACHINE: forced budget exhausted, entering review (cycles=%d)",
                 monitor.tool_state_review_cycles,
             )
-            return "auto", "forced_budget_exhausted"
+            return "required", "forced_budget_exhausted"
 
         monitor.tool_state_forced_budget_remaining -= 1
         return "required", "act"
@@ -2091,7 +2088,7 @@ def _resolve_state_machine_tool_choice(
                 1, PROXY_TOOL_STATE_FORCED_BUDGET // 2
             )
             return "required", "review_complete"
-        return "auto", "review"
+        return "required", "review"
 
     if monitor.tool_turn_phase == "finalize":
         if monitor.tool_state_auto_budget_remaining <= 0:
@@ -2388,22 +2385,6 @@ def build_openai_request(
             if not has_tool_results:
                 monitor.reset_tool_turn_state(reason="no_tool_results")
 
-        # CONTEXT-AWARE RELAXATION: when context utilization is high and
-        # tool_choice was forced to required, relax to auto to let the model
-        # emit shorter text responses instead of consuming more tokens.
-        if openai_body.get("tool_choice") == "required":
-            ctx_utilization = (
-                monitor.last_input_tokens / monitor.context_window
-                if monitor.context_window > 0
-                else 0.0
-            )
-            if ctx_utilization >= PROXY_CONTEXT_HIGH_RELAXATION_THRESHOLD:
-                openai_body["tool_choice"] = "auto"
-                logger.warning(
-                    "CONTEXT-AWARE RELAXATION: tool_choice=auto (utilization=%.1f%% >= %.0f%% threshold)",
-                    ctx_utilization * 100,
-                    PROXY_CONTEXT_HIGH_RELAXATION_THRESHOLD * 100,
-                )
 
         if PROXY_DISABLE_THINKING_ON_TOOL_TURNS:
             openai_body["enable_thinking"] = False

package/tools/agents/tests/test_anthropic_proxy_streaming.py

@@ -1861,7 +1861,8 @@ class TestToolTurnControls(unittest.TestCase):
 
             self.assertEqual(openai_1.get("tool_choice"), "required")
             self.assertEqual(openai_2.get("tool_choice"), "required")
-            self.assertEqual(openai_3.get("tool_choice"), "auto")
+            # Review phase now keeps required to prevent end-turn escape
+            self.assertEqual(openai_3.get("tool_choice"), "required")
         finally:
             setattr(proxy, "PROXY_TOOL_STATE_MACHINE", old_state)
             setattr(proxy, "PROXY_TOOL_STATE_MIN_MESSAGES", old_min_msgs)
@@ -1938,7 +1939,8 @@ class TestToolTurnControls(unittest.TestCase):
             }
 
             openai = proxy.build_openai_request(body, monitor)
-            self.assertEqual(openai.get("tool_choice"), "auto")
+            # Review phase now keeps required to prevent end-turn escape
+            self.assertEqual(openai.get("tool_choice"), "required")
             self.assertEqual(monitor.tool_turn_phase, "review")
         finally:
             setattr(proxy, "PROXY_TOOL_STATE_MACHINE", old_state)
@@ -2067,7 +2069,8 @@ class TestToolTurnControls(unittest.TestCase):
             }
 
             openai = proxy.build_openai_request(body, monitor)
-            self.assertEqual(openai.get("tool_choice"), "auto")
+            # Review phase now keeps required to prevent end-turn escape
+            self.assertEqual(openai.get("tool_choice"), "required")
             self.assertEqual(monitor.tool_turn_phase, "review")
             self.assertEqual(monitor.tool_state_review_cycles, 1)
         finally:
@@ -3166,44 +3169,6 @@ class TestToolStarvationBreaker(unittest.TestCase):
         self.assertIn("tools", result)
 
 
-class TestContextAwareRelaxation(unittest.TestCase):
-    """Tests for context-aware tool_choice relaxation."""
-
-    def test_relaxes_at_high_utilization(self):
-        monitor = proxy.SessionMonitor()
-        monitor.context_window = 100000
-        monitor.last_input_tokens = 75000  # 75% > 70% threshold
-        body = {
-            "model": "qwen3.5",
-            "messages": [
-                {"role": "user", "content": "hello"},
-                {"role": "assistant", "content": "text only"},
-                {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "x", "content": "ok"}]},
-            ],
-            "tools": [{"name": "Bash", "input_schema": {"type": "object", "properties": {"command": {"type": "string"}}}}],
-        }
-        result = proxy.build_openai_request(body, monitor)
-        # Should be auto, not required
-        self.assertEqual(result.get("tool_choice"), "auto")
-
-    def test_no_relaxation_below_threshold(self):
-        monitor = proxy.SessionMonitor()
-        monitor.context_window = 100000
-        monitor.last_input_tokens = 50000  # 50% < 70%
-        body = {
-            "model": "qwen3.5",
-            "messages": [
-                {"role": "user", "content": "hello"},
-                {"role": "assistant", "content": "text only"},
-                {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "x", "content": "ok"}]},
-            ],
-            "tools": [{"name": "Bash", "input_schema": {"type": "object", "properties": {"command": {"type": "string"}}}}],
-        }
-        result = proxy.build_openai_request(body, monitor)
-        # Should still be required (state machine forces it)
-        self.assertEqual(result.get("tool_choice"), "required")
-
-
 if __name__ == "__main__":
     unittest.main()