@miller-tech/uap 1.20.23 → 1.20.24

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@miller-tech/uap",
-  "version": "1.20.23",
+  "version": "1.20.24",
   "description": "Autonomous AI agent memory system with CLAUDE.md protocol enforcement",
   "type": "module",
   "main": "dist/index.js",

package/tools/agents/scripts/anthropic_proxy.py

@@ -4581,6 +4581,34 @@ async def _apply_malformed_tool_guardrail(
         )
         current_issue = retry_issue
 
+    # Option 2 (PR #154): When retries exhaust during review phase, reset to
+    # bootstrap instead of returning guardrail fallback. This re-enables all
+    # tools (including previously excluded cycling ones) and gives the model
+    # a clean shot. The cycle detector will catch re-cycling if it recurs.
+    if monitor.tool_turn_phase == "review":
+        logger.warning(
+            "TOOL RESPONSE review-phase reset: session=%s retries exhausted in review "
+            "(kind=%s malformed=%d), resetting to bootstrap for fresh attempt",
+            session_id,
+            current_issue.kind or issue.kind,
+            monitor.malformed_tool_streak,
+        )
+        monitor.reset_tool_turn_state(reason="review_retry_exhausted")
+        monitor.malformed_tool_streak = 0
+        monitor.invalid_tool_call_streak = 0
+        # Return the best response we have — even if degraded — to keep
+        # the conversation moving rather than returning a guardrail stub.
+        degraded_text = _sanitize_tool_call_apology_text(
+            _openai_message_text(working_resp)
+        ).strip()
+        if degraded_text and not _looks_malformed_tool_payload(degraded_text):
+            return _build_safe_text_openai_response(
+                working_resp, degraded_text, finish_reason="tool_calls",
+            )
+        return _build_clean_guardrail_openai_response(
+            working_resp, finish_reason="tool_calls",
+        )
+
     logger.error(
         "TOOL RESPONSE issue persisted after retries (session=%s kind=%s malformed=%d invalid=%d required_miss=%d); returning clean guardrail response",
         session_id,

package/tools/agents/tests/test_anthropic_proxy_streaming.py

@@ -4178,3 +4178,45 @@ class TestFinalizePingPongFix(unittest.TestCase):
             proxy.TOOL_CALL_GRAMMAR_TOOLS_COMPATIBLE = old_compat
             proxy.PROXY_TOOL_CALL_GRAMMAR = old_flag
             proxy.TOOL_CALL_GBNF = old_gbnf
+
+
+class TestReviewPhaseBootstrapReset(unittest.TestCase):
+    """Tests for bootstrap reset after exhausted retries in review phase (PR #154)."""
+
+    def _make_monitor_in_review(self):
+        m = proxy.SessionMonitor()
+        m.set_tool_turn_phase("review", reason="test")
+        m.malformed_tool_streak = 3
+        m.invalid_tool_call_streak = 0
+        return m
+
+    def _make_monitor_in_act(self):
+        m = proxy.SessionMonitor()
+        m.set_tool_turn_phase("act", reason="test")
+        m.malformed_tool_streak = 3
+        return m
+
+    def test_review_phase_resets_to_bootstrap(self):
+        """After retries exhaust in review, monitor resets to bootstrap."""
+        m = self._make_monitor_in_review()
+        self.assertEqual(m.tool_turn_phase, "review")
+        self.assertEqual(m.malformed_tool_streak, 3)
+
+        # Simulate what happens after retry exhaustion: the code checks
+        # monitor.tool_turn_phase == "review" and resets
+        if m.tool_turn_phase == "review":
+            m.reset_tool_turn_state(reason="review_retry_exhausted")
+            m.malformed_tool_streak = 0
+            m.invalid_tool_call_streak = 0
+
+        self.assertEqual(m.tool_turn_phase, "bootstrap")
+        self.assertEqual(m.malformed_tool_streak, 0)
+        self.assertEqual(m.tool_state_stagnation_streak, 0)
+        self.assertEqual(m.cycling_tool_names, [])
+
+    def test_act_phase_does_not_reset(self):
+        """In act phase, retries exhaustion should NOT trigger bootstrap reset."""
+        m = self._make_monitor_in_act()
+        # The bootstrap reset only triggers for review phase
+        self.assertNotEqual(m.tool_turn_phase, "review")
+        # In act phase, the normal guardrail fallback path runs instead