@miller-tech/uap 1.20.17 → 1.20.18

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@miller-tech/uap",
-  "version": "1.20.17",
+  "version": "1.20.18",
   "description": "Autonomous AI agent memory system with CLAUDE.md protocol enforcement",
   "type": "module",
   "main": "dist/index.js",

package/tools/agents/scripts/anthropic_proxy.py

@@ -42,6 +42,12 @@ Configuration (Environment Variables)
     PROXY_READ_TIMEOUT   Read timeout in seconds for upstream LLM streaming
                          Default: 600 (10 minutes)
 
+    PROXY_TOOL_TURN_MAX_TOKENS   Max tokens for tool-call turns (0 to disable)
+                                Default: 8192
+
+    PROXY_TOOL_TURN_MAX_TOKENS_GARBLED   Max tokens after garbled/malformed output
+                                         Default: 4096
+
     PROXY_MAX_CONNECTIONS   Max concurrent connections to upstream
                             Default: 20
 
@@ -194,6 +200,10 @@ PROXY_STREAM_REASONING_MAX_CHARS = int(
     os.environ.get("PROXY_STREAM_REASONING_MAX_CHARS", "240")
 )
 PROXY_MAX_TOKENS_FLOOR = int(os.environ.get("PROXY_MAX_TOKENS_FLOOR", "16384"))
+PROXY_TOOL_TURN_MAX_TOKENS = int(os.environ.get("PROXY_TOOL_TURN_MAX_TOKENS", "8192"))
+PROXY_TOOL_TURN_MAX_TOKENS_GARBLED = int(
+    os.environ.get("PROXY_TOOL_TURN_MAX_TOKENS_GARBLED", "4096")
+)
 PROXY_TOOL_NARROWING = os.environ.get("PROXY_TOOL_NARROWING", "off").lower() not in {
     "0",
     "false",
@@ -631,6 +641,7 @@ class SessionMonitor:
     tool_state_review_cycles: int = 0
     last_tool_fingerprint: str = ""
     cycling_tool_names: list = field(default_factory=list)
+    last_response_garbled: bool = False  # previous turn had garbled/malformed output
     finalize_turn_active: bool = False
     completion_required: bool = False
     completion_pending: bool = False
@@ -1457,6 +1468,11 @@ async def lifespan(app: FastAPI):
         int(PROXY_GENERATION_TIMEOUT),
         int(PROXY_SLOT_HANG_TIMEOUT),
     )
+    logger.info(
+        "Tool turn max_tokens: cap=%d garbled_cap=%d",
+        PROXY_TOOL_TURN_MAX_TOKENS,
+        PROXY_TOOL_TURN_MAX_TOKENS_GARBLED,
+    )
 
     yield
     await http_client.aclose()
@@ -2316,6 +2332,23 @@ def build_openai_request(
                 )
                 requested_max = PROXY_OPUS46_MAX_TOKENS_HIGH_CTX
 
+        # Option 1+3+4: Cap max_tokens for tool turns to prevent 32K waste.
+        # Tool call responses rarely need more than a few thousand tokens.
+        # After garbled/malformed output, use an even lower cap.
+        if has_tools and PROXY_TOOL_TURN_MAX_TOKENS > 0:
+            if monitor.last_response_garbled and PROXY_TOOL_TURN_MAX_TOKENS_GARBLED > 0:
+                tool_cap = PROXY_TOOL_TURN_MAX_TOKENS_GARBLED
+            else:
+                tool_cap = PROXY_TOOL_TURN_MAX_TOKENS
+            if requested_max > tool_cap:
+                logger.info(
+                    "TOOL TURN MAX_TOKENS cap: %d -> %d (garbled_prev=%s)",
+                    requested_max,
+                    tool_cap,
+                    monitor.last_response_garbled,
+                )
+                requested_max = tool_cap
+
         openai_body["max_tokens"] = requested_max
     if "temperature" in anthropic_body:
         openai_body["temperature"] = anthropic_body["temperature"]
@@ -4260,6 +4293,7 @@ async def _apply_malformed_tool_guardrail(
             monitor.malformed_tool_streak = 0
             monitor.invalid_tool_call_streak = 0
             monitor.required_tool_miss_streak = 0
+            monitor.last_response_garbled = False
         if repair_count > 0:
             monitor.arg_preflight_repairs += repair_count
             logger.info(
@@ -4269,6 +4303,9 @@ async def _apply_malformed_tool_guardrail(
             )
         return working_resp
 
+    # Mark garbled state for progressive max_tokens reduction on next turn
+    monitor.last_response_garbled = True
+
     if issue.kind == "malformed_payload":
         monitor.malformed_tool_streak += 1
     elif issue.kind == "invalid_tool_args":
@@ -4354,6 +4391,7 @@ async def _apply_malformed_tool_guardrail(
             monitor.malformed_tool_streak = 0
             monitor.invalid_tool_call_streak = 0
             monitor.required_tool_miss_streak = 0
+            monitor.last_response_garbled = False
             logger.info(
                 "TOOL RESPONSE RETRY success: kind=%s attempt=%d/%d",
                 current_issue.kind,

package/tools/agents/tests/test_anthropic_proxy_streaming.py

@@ -3696,3 +3696,90 @@ class TestGarbledArgsRetry(unittest.TestCase):
     def test_env_sync_malformed_retry_max(self):
         """PROXY_MALFORMED_TOOL_RETRY_MAX should be 3."""
         self.assertEqual(proxy.PROXY_MALFORMED_TOOL_RETRY_MAX, 3)
+
+
+class TestToolTurnMaxTokensCap(unittest.TestCase):
+    """Tests for tool turn max_tokens capping to prevent 32K waste."""
+
+    def test_tool_turn_max_tokens_constant(self):
+        """PROXY_TOOL_TURN_MAX_TOKENS should default to 8192."""
+        self.assertEqual(proxy.PROXY_TOOL_TURN_MAX_TOKENS, 8192)
+
+    def test_tool_turn_max_tokens_garbled_constant(self):
+        """PROXY_TOOL_TURN_MAX_TOKENS_GARBLED should default to 4096."""
+        self.assertEqual(proxy.PROXY_TOOL_TURN_MAX_TOKENS_GARBLED, 4096)
+
+    def test_tool_turn_caps_high_max_tokens(self):
+        """Tool turn with max_tokens=32000 should be capped to 8192."""
+        body = {
+            "model": "test-model",
+            "max_tokens": 32000,
+            "messages": [{"role": "user", "content": "test"}],
+            "tools": [
+                {
+                    "name": "Bash",
+                    "description": "run command",
+                    "input_schema": {"type": "object"},
+                }
+            ],
+        }
+        monitor = proxy.SessionMonitor(context_window=262144)
+        openai_body = proxy.build_openai_request(body, monitor)
+        self.assertLessEqual(openai_body["max_tokens"], proxy.PROXY_TOOL_TURN_MAX_TOKENS)
+
+    def test_tool_turn_garbled_reduces_cap(self):
+        """After garbled output, max_tokens should use the lower garbled cap."""
+        body = {
+            "model": "test-model",
+            "max_tokens": 32000,
+            "messages": [{"role": "user", "content": "test"}],
+            "tools": [
+                {
+                    "name": "Bash",
+                    "description": "run command",
+                    "input_schema": {"type": "object"},
+                }
+            ],
+        }
+        monitor = proxy.SessionMonitor(context_window=262144)
+        monitor.last_response_garbled = True
+        openai_body = proxy.build_openai_request(body, monitor)
+        self.assertLessEqual(
+            openai_body["max_tokens"], proxy.PROXY_TOOL_TURN_MAX_TOKENS_GARBLED
+        )
+
+    def test_non_tool_request_not_capped(self):
+        """Non-tool requests should not be affected by tool turn cap."""
+        body = {
+            "model": "test-model",
+            "max_tokens": 32000,
+            "messages": [{"role": "user", "content": "test"}],
+        }
+        monitor = proxy.SessionMonitor(context_window=262144)
+        openai_body = proxy.build_openai_request(body, monitor)
+        # Should not be capped to 8192 (may be capped by context window logic)
+        self.assertGreater(openai_body["max_tokens"], proxy.PROXY_TOOL_TURN_MAX_TOKENS)
+
+    def test_last_response_garbled_cleared_on_clean(self):
+        """SessionMonitor.last_response_garbled should default to False."""
+        monitor = proxy.SessionMonitor(context_window=262144)
+        self.assertFalse(monitor.last_response_garbled)
+
+    def test_small_max_tokens_stays_within_cap(self):
+        """If client requests less than the cap, result should not exceed cap."""
+        body = {
+            "model": "test-model",
+            "max_tokens": 4096,
+            "messages": [{"role": "user", "content": "test"}],
+            "tools": [
+                {
+                    "name": "Bash",
+                    "description": "run command",
+                    "input_schema": {"type": "object"},
+                }
+            ],
+        }
+        monitor = proxy.SessionMonitor(context_window=262144)
+        openai_body = proxy.build_openai_request(body, monitor)
+        # The tool turn cap should ensure we don't exceed PROXY_TOOL_TURN_MAX_TOKENS
+        self.assertLessEqual(openai_body["max_tokens"], proxy.PROXY_TOOL_TURN_MAX_TOKENS)