@miller-tech/uap 1.96.4 → 1.96.6

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@miller-tech/uap",
-  "version": "1.96.4",
+  "version": "1.96.6",
   "description": "Autonomous AI agent memory system with CLAUDE.md protocol enforcement",
   "type": "module",
   "main": "dist/index.js",

package/tools/agents/scripts/anthropic_proxy.py

@@ -305,6 +305,13 @@ _WRITE_TOOL_CLASS = frozenset({
     "write", "edit", "multiedit", "notebookedit",
     "str_replace", "str_replace_editor", "str_replace_based_edit_tool",
     "create_file", "applypatch", "apply_patch",
+    # OpenAI-style tool-loop names (uap deliver's agentic executor and other
+    # OpenAI-compat agents). Without these, every agentic round that WRITES
+    # files still counted as a no-write turn, the recon-convergence streak
+    # grew unbounded, and the proxy injected "stop exploring, synthesize"
+    # directives into a loop that was mid-build — observed live at
+    # no_write_streak=62 with the model then emitting files as plain text.
+    "write_file", "edit_file", "save_file",
 })
 
 PROXY_GUARDRAIL_RETRY = os.environ.get("PROXY_GUARDRAIL_RETRY", "on").lower() not in {
@@ -4269,6 +4276,13 @@ def build_openai_request(
         # retries (observed Shannon: max_tokens=512 + tools=7 -> ~5 retries
         # per turn). Bump up to THINKING_MIN_FOR_TOOLS for these requests.
         THINKING_MIN_FOR_TOOLS = 2048
+        # No-tool turns need thinking headroom too: evaluator calls (acceptance
+        # judge, critic, ideation) request ~4096 max_tokens, Qwen spends all of
+        # it inside <think>, the EMPTY-OUTPUT GUARD promotes truncated
+        # reasoning as the body, and the caller gets an unparseable verdict —
+        # observed live as every acceptance judgment failing. Bump small
+        # no-tool budgets so the model can finish thinking AND answer.
+        THINKING_MIN_NO_TOOLS = int(os.environ.get("PROXY_THINKING_MIN_NO_TOOLS", "8192"))
         if skip_floor:
             requested_max = requested_raw
             # Even when skipping the big floor, bump small tool-turn
@@ -4285,6 +4299,18 @@ def build_openai_request(
                     requested_raw,
                     requested_max,
                 )
+            elif (
+                not has_tools
+                and THINKING_MIN_NO_TOOLS > 0
+                and requested_raw < THINKING_MIN_NO_TOOLS
+                and requested_raw > 16  # leave true preflight alone
+            ):
+                requested_max = THINKING_MIN_NO_TOOLS
+                logger.info(
+                    "MAX_TOKENS thinking-floor (no-tool): %d -> %d (Qwen mandatory thinking; evaluator verdicts need room after <think>)",
+                    requested_raw,
+                    requested_max,
+                )
             elif requested_raw < PROXY_MAX_TOKENS_FLOOR and PROXY_MAX_TOKENS_FLOOR > 0:
                 logger.info(
                     "MAX_TOKENS floor skipped: has_tools=%s thinking_active=%s requested=%d floor=%d",