@miller-tech/uap 1.20.25 → 1.20.26

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@miller-tech/uap",
-  "version": "1.20.25",
+  "version": "1.20.26",
   "description": "Autonomous AI agent memory system with CLAUDE.md protocol enforcement",
   "type": "module",
   "main": "dist/index.js",

package/tools/agents/scripts/anthropic_proxy.py

@@ -2281,11 +2281,13 @@ def _resolve_state_machine_tool_choice(
                 1, PROXY_TOOL_STATE_FORCED_BUDGET // 2
             )
             # If stagnation cleared during review, the model tried a
-            # different approach — reward by reducing cycle pressure.
+            # different approach — reward by reducing cycle pressure and
+            # lifting persistent tool exclusion.
             if monitor.tool_state_stagnation_streak == 0 and monitor.tool_state_review_cycles > 0:
                 monitor.tool_state_review_cycles = max(0, monitor.tool_state_review_cycles - 1)
+                monitor.cycling_tool_names = []
                 logger.info(
-                    "TOOL STATE MACHINE: review_cycles decremented to %d (stagnation cleared)",
+                    "TOOL STATE MACHINE: review_cycles decremented to %d, cycling exclusion lifted (stagnation cleared)",
                     monitor.tool_state_review_cycles,
                 )
             return "required", "review_complete"
@@ -2589,31 +2591,41 @@ def build_openai_request(
             monitor.no_progress_streak = (
                 0 if last_user_has_tool_result else monitor.no_progress_streak + 1
             )
-            # Option 1: Inject cycle-break instruction when entering review
+            # Inject cycle-break instruction when entering review
+            # Option 3 (Cycle 14): Escalate hint text based on review cycle count
             if (
                 monitor.tool_turn_phase == "review"
                 and state_reason in {"cycle_detected", "stagnation"}
                 and monitor.cycling_tool_names
             ):
                 cycling_names = ", ".join(monitor.cycling_tool_names)
-                cycle_hint = (
-                    f"You have been repeatedly calling the same tool(s): {cycling_names}. "
-                    "This is not making progress. Use a DIFFERENT tool to advance the task, "
-                    "or call a tool that produces your final answer."
-                )
+                cycles = monitor.tool_state_review_cycles
+                if cycles <= 1:
+                    cycle_hint = (
+                        f"You have been repeatedly calling the same tool(s): {cycling_names}. "
+                        "This is not making progress. Use a DIFFERENT tool to advance the task, "
+                        "or call a tool that produces your final answer."
+                    )
+                else:
+                    cycle_hint = (
+                        f"CRITICAL: You have cycled {cycling_names} for {cycles} review rounds without progress. "
+                        "State what you have accomplished so far and what the next DIFFERENT action should be. "
+                        "Do NOT call the same tool again. Choose a completely different approach or "
+                        "produce your final answer now."
+                    )
                 messages = openai_body.get("messages", [])
                 messages.append({"role": "user", "content": cycle_hint})
                 openai_body["messages"] = messages
                 logger.warning(
-                    "CYCLE BREAK: injected hint about cycling tools: %s",
+                    "CYCLE BREAK: injected hint about cycling tools: %s (escalation=%d)",
                     cycling_names,
+                    cycles,
                 )
-            # Option 2: Narrow tools during review to exclude cycling tools
-            # Option 1 enhancement: if any cycling tool is read-only, exclude
-            # the entire read-only class to prevent tool-hopping (read→glob→grep)
+            # Narrow tools to exclude cycling tools
+            # Option 1 (Cycle 13): if any cycling tool is read-only, exclude entire class
+            # Option 1 (Cycle 14): persist exclusion during act phase too, not just review
             if (
-                monitor.tool_turn_phase == "review"
-                and monitor.cycling_tool_names
+                monitor.cycling_tool_names
                 and "tools" in openai_body
             ):
                 exclude_set = set(monitor.cycling_tool_names)

package/tools/agents/tests/test_anthropic_proxy_streaming.py

@@ -4406,3 +4406,172 @@ class TestReadOnlyCycleClassExclusion(unittest.TestCase):
 
         dup, _ = monitor.has_duplicate_read_target(threshold=3)
         self.assertFalse(dup)
+
+
+class TestPersistentCycleExclusion(unittest.TestCase):
+    """Tests for Cycle 14: persistent exclusion, escalating hints, and
+    exclusion across review→act transitions."""
+
+    def _make_body_with_tools(self, tool_names, active_tool="bash", active_input=None):
+        tools = [
+            {"name": n, "description": f"{n} tool", "input_schema": {"type": "object"}}
+            for n in tool_names
+        ]
+        inp = active_input or {"command": "ls"}
+        return {
+            "model": "test",
+            "messages": [
+                {"role": "user", "content": "do something"},
+                {
+                    "role": "assistant",
+                    "content": [
+                        {"type": "tool_use", "id": "t1", "name": active_tool, "input": inp}
+                    ],
+                },
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "tool_result", "tool_use_id": "t1", "content": "ok"}
+                    ],
+                },
+            ],
+            "tools": tools,
+        }
+
+    def test_exclusion_persists_through_act_phase(self):
+        """Option 1: cycling_tool_names exclusion persists in act phase after review."""
+        old_vals = {}
+        for k in ["PROXY_TOOL_STATE_MACHINE", "PROXY_TOOL_STATE_MIN_MESSAGES",
+                   "PROXY_TOOL_STATE_FORCED_BUDGET", "PROXY_TOOL_STATE_CYCLE_WINDOW",
+                   "PROXY_TOOL_STATE_STAGNATION_THRESHOLD"]:
+            old_vals[k] = getattr(proxy, k)
+        try:
+            setattr(proxy, "PROXY_TOOL_STATE_MACHINE", True)
+            setattr(proxy, "PROXY_TOOL_STATE_MIN_MESSAGES", 3)
+            setattr(proxy, "PROXY_TOOL_STATE_FORCED_BUDGET", 6)
+            setattr(proxy, "PROXY_TOOL_STATE_CYCLE_WINDOW", 3)
+            setattr(proxy, "PROXY_TOOL_STATE_STAGNATION_THRESHOLD", 2)
+
+            all_tools = ["bash", "read", "write", "edit"]
+            body = self._make_body_with_tools(all_tools)
+            monitor = proxy.SessionMonitor(context_window=262144)
+
+            # Simulate bash cycling that triggers review
+            monitor.cycling_tool_names = ["bash"]
+            monitor.tool_turn_phase = "act"
+            monitor.tool_state_forced_budget_remaining = 5
+
+            openai = proxy.build_openai_request(body, monitor)
+
+            # In act phase with cycling_tool_names set, bash should be excluded
+            remaining = [t["function"]["name"] for t in openai.get("tools", [])]
+            self.assertNotIn("bash", remaining)
+            self.assertIn("read", remaining)
+            self.assertIn("write", remaining)
+        finally:
+            for k, v in old_vals.items():
+                setattr(proxy, k, v)
+
+    def test_exclusion_cleared_on_stagnation_clear(self):
+        """Option 1: cycling exclusion is lifted when stagnation clears in review."""
+        monitor = proxy.SessionMonitor(context_window=262144)
+        monitor.tool_turn_phase = "review"
+        monitor.tool_state_review_cycles = 1
+        monitor.tool_state_stagnation_streak = 0  # stagnation cleared
+        monitor.cycling_tool_names = ["bash"]
+        monitor.tool_state_auto_budget_remaining = 0
+        monitor.tool_state_forced_budget_remaining = 6
+
+        # This should transition review→act and clear cycling names
+        old_vals = {}
+        for k in ["PROXY_TOOL_STATE_MACHINE", "PROXY_TOOL_STATE_MIN_MESSAGES",
+                   "PROXY_TOOL_STATE_FORCED_BUDGET"]:
+            old_vals[k] = getattr(proxy, k)
+        try:
+            setattr(proxy, "PROXY_TOOL_STATE_MACHINE", True)
+            setattr(proxy, "PROXY_TOOL_STATE_MIN_MESSAGES", 3)
+            setattr(proxy, "PROXY_TOOL_STATE_FORCED_BUDGET", 6)
+
+            body = self._make_body_with_tools(["bash", "read", "write"])
+            proxy.build_openai_request(body, monitor)
+
+            self.assertEqual(monitor.tool_turn_phase, "act")
+            self.assertEqual(monitor.cycling_tool_names, [])
+        finally:
+            for k, v in old_vals.items():
+                setattr(proxy, k, v)
+
+    def test_escalated_hint_on_cycle_2(self):
+        """Option 3: cycle 2+ gets escalated CRITICAL hint text."""
+        old_vals = {}
+        for k in ["PROXY_TOOL_STATE_MACHINE", "PROXY_TOOL_STATE_MIN_MESSAGES",
+                   "PROXY_TOOL_STATE_FORCED_BUDGET", "PROXY_TOOL_STATE_CYCLE_WINDOW",
+                   "PROXY_TOOL_STATE_STAGNATION_THRESHOLD"]:
+            old_vals[k] = getattr(proxy, k)
+        try:
+            setattr(proxy, "PROXY_TOOL_STATE_MACHINE", True)
+            setattr(proxy, "PROXY_TOOL_STATE_MIN_MESSAGES", 3)
+            setattr(proxy, "PROXY_TOOL_STATE_FORCED_BUDGET", 20)
+            setattr(proxy, "PROXY_TOOL_STATE_CYCLE_WINDOW", 3)
+            setattr(proxy, "PROXY_TOOL_STATE_STAGNATION_THRESHOLD", 2)
+
+            all_tools = ["bash", "read", "write"]
+            body = self._make_body_with_tools(all_tools)
+            monitor = proxy.SessionMonitor(context_window=262144)
+            # Pre-set as if we've already been through 1 review cycle
+            monitor.tool_turn_phase = "act"
+            monitor.tool_state_review_cycles = 1
+            monitor.tool_state_forced_budget_remaining = 20
+            monitor.tool_state_stagnation_streak = 3
+            fp = "bash:781c24ad"
+            monitor.tool_call_history = [fp, fp, fp]
+            monitor.last_tool_fingerprint = fp
+
+            openai = proxy.build_openai_request(body, monitor)
+
+            # Should now be in review with cycles=2 and escalated hint
+            self.assertEqual(monitor.tool_turn_phase, "review")
+            self.assertEqual(monitor.tool_state_review_cycles, 2)
+            messages = openai.get("messages", [])
+            last_user = [m for m in messages if m.get("role") == "user"][-1]
+            self.assertIn("CRITICAL", last_user["content"])
+            self.assertIn("2 review rounds", last_user["content"])
+        finally:
+            for k, v in old_vals.items():
+                setattr(proxy, k, v)
+
+    def test_mild_hint_on_cycle_1(self):
+        """Option 3: cycle 1 gets mild hint, not escalated."""
+        old_vals = {}
+        for k in ["PROXY_TOOL_STATE_MACHINE", "PROXY_TOOL_STATE_MIN_MESSAGES",
+                   "PROXY_TOOL_STATE_FORCED_BUDGET", "PROXY_TOOL_STATE_CYCLE_WINDOW",
+                   "PROXY_TOOL_STATE_STAGNATION_THRESHOLD"]:
+            old_vals[k] = getattr(proxy, k)
+        try:
+            setattr(proxy, "PROXY_TOOL_STATE_MACHINE", True)
+            setattr(proxy, "PROXY_TOOL_STATE_MIN_MESSAGES", 3)
+            setattr(proxy, "PROXY_TOOL_STATE_FORCED_BUDGET", 20)
+            setattr(proxy, "PROXY_TOOL_STATE_CYCLE_WINDOW", 3)
+            setattr(proxy, "PROXY_TOOL_STATE_STAGNATION_THRESHOLD", 2)
+
+            body = self._make_body_with_tools(["bash", "read", "write"])
+            monitor = proxy.SessionMonitor(context_window=262144)
+            monitor.tool_turn_phase = "act"
+            monitor.tool_state_review_cycles = 0
+            monitor.tool_state_forced_budget_remaining = 20
+            monitor.tool_state_stagnation_streak = 3
+            fp = "bash:781c24ad"
+            monitor.tool_call_history = [fp, fp, fp]
+            monitor.last_tool_fingerprint = fp
+
+            openai = proxy.build_openai_request(body, monitor)
+
+            self.assertEqual(monitor.tool_turn_phase, "review")
+            self.assertEqual(monitor.tool_state_review_cycles, 1)
+            messages = openai.get("messages", [])
+            last_user = [m for m in messages if m.get("role") == "user"][-1]
+            self.assertNotIn("CRITICAL", last_user["content"])
+            self.assertIn("DIFFERENT tool", last_user["content"])
+        finally:
+            for k, v in old_vals.items():
+                setattr(proxy, k, v)