scripts/32_gate2_toolcalls.py

@@ -35,6 +35,10 @@ HOLDOUT_PATH = REPO_ROOT / "data" / "holdout_prompts.jsonl"
 RESULTS_PATH = REPO_ROOT / "data" / os.environ.get("GATE2_RESULTS_FILENAME", "gate2_results.json")
 BASE_URL = os.environ.get("VLLM_EVAL_URL", "http://localhost:8001")
 MODEL_NAME = os.environ.get("VLLM_EVAL_MODEL", "qwen3.6-35b-a3b-mcp-bf16")
+# 1024 dejaba cortar la respuesta a mitad de razonamiento en modelos con
+# --reasoning-parser activo antes de emitir el tool_call -- ver hallazgo de
+# Fase 5 (misma causa que el fix de gate3, max_tokens 512->2048).
+GATE2_MAX_TOKENS = int(os.environ.get("GATE2_MAX_TOKENS", "2048"))
 
 
 def load_holdout():
@@ -92,13 +96,13 @@ def validate_tool_call(tool_call, tools):
     return True, None
 
 
-def call_vllm(prompt, tools, timeout=120):
+def call_vllm(prompt, tools, timeout=240):
     payload = {
         "model": MODEL_NAME,
         "messages": [{"role": "user", "content": prompt}],
         "tools": to_openai_tools(tools),
         "tool_choice": "auto",
-        "max_tokens": 1024,
+        "max_tokens": GATE2_MAX_TOKENS,
         "temperature": 0.0,
     }
     resp = requests.post(f"{BASE_URL}/v1/chat/completions", json=payload, timeout=timeout)
@@ -133,11 +137,24 @@ def main():
             continue
 
         message = response["choices"][0]["message"]
+        # Se guarda siempre el texto completo (content + reasoning) para poder auditar
+        # con criterio humano los casos que fallan o quedan sin tool_call -- antes no se
+        # guardaba nada de esto, lo que hacia imposible diagnosticar truncamiento.
+        content = message.get("content") or ""
+        reasoning = message.get("reasoning") or ""
         tool_calls = message.get("tool_calls") or []
         if not tool_calls:
             stats[mcp]["no_tool_call"] += 1
             stats["__global__"]["no_tool_call"] += 1
-            results.append({"mcp": mcp, "prompt": ex["prompt"], "tool_calls": None, "valid": None})
+            results.append({
+                "mcp": mcp,
+                "prompt": ex["prompt"],
+                "tool_calls": None,
+                "valid": None,
+                "content": content,
+                "reasoning": reasoning,
+                "finish_reason": response["choices"][0].get("finish_reason"),
+            })
             continue
 
         all_valid = True
@@ -161,6 +178,9 @@ def main():
             "tool_calls": [tc["function"]["name"] for tc in tool_calls],
             "valid": all_valid,
             "errors": errors,
+            "content": content,
+            "reasoning": reasoning,
+            "finish_reason": response["choices"][0].get("finish_reason"),
         })
 
         if (i + 1) % 20 == 0: