fix: validate and size governed patch outputs

packages/casan-harness/scripts/bash/model-call.py

@@ -56,6 +56,17 @@ PROVIDER_LOG = os.path.join(REPO_ROOT, ".specify/logs/level5/provider-usage.json
 REQUEST_TIMEOUT = max(1, int(os.environ.get("CASAN_MODEL_TIMEOUT_SEC", "60")))
 
 
+def generation_max_tokens(role):
+    """Return a bounded generation budget shared by every model provider."""
+    if role in ("classify", "judge"):
+        return 16
+    try:
+        configured = int(os.environ.get("CASAN_MODEL_GENERATE_MAX_TOKENS", "1400"))
+    except ValueError:
+        configured = 1400
+    return min(max(configured, 64), 8192)
+
+
 def enforce_call_budget():
     """Fail closed when a run exceeds CASAN_MODEL_MAX_CALLS. The count is tracked in
     CASAN_MODEL_CALL_COUNTER_FILE so it spans the many per-step model-call processes
@@ -205,7 +216,7 @@ def call_openai(model_name, prompt, role):
         "model": model_name,
         "messages": [{"role": "user", "content": prompt}],
         "temperature": 0 if role in ("classify", "judge") else 0.2,
-        "max_tokens": 16 if role in ("classify", "judge") else 512,
+        "max_tokens": generation_max_tokens(role),
     }
     data = json.dumps(body).encode()
     req = urllib.request.Request(
@@ -272,7 +283,7 @@ def call_openai_compatible(model_name, prompt, role):
         "model": model_name,
         "messages": [{"role": "user", "content": prompt}],
         "temperature": 0 if role in ("classify", "judge") else 0.2,
-        "max_tokens": 16 if role in ("classify", "judge") else 512,
+        "max_tokens": generation_max_tokens(role),
     }
     data = json.dumps(body).encode()
     req = urllib.request.Request(
@@ -335,7 +346,7 @@ def call_anthropic(model_name, prompt, role):
     url = f"https://{host}/v1/messages"
     body = {
         "model": model_name,
-        "max_tokens": 16 if role in ("classify", "judge") else 512,
+        "max_tokens": generation_max_tokens(role),
         "messages": [{"role": "user", "content": prompt}],
     }
     data = json.dumps(body).encode()