optz: Goal orchestrator

2026-07-11 12:18:59 +09:00
parent b56f7d357e
commit 4fc72332f5
34 changed files with 850 additions and 452 deletions
@@ -171,6 +171,11 @@ def call_ollama(model_name, prompt, role):
        # small budget is consumed by reasoning and `response` comes back empty.
        body["think"] = False
        body["options"]["num_predict"] = 16  # terse final answer + fast
+    else:
+        # Prevent an unconstrained local generation from consuming the whole
+        # request window. Goal Orchestrator can tune this without weakening the
+        # shorter classifier/judge budgets.
+        body["options"]["num_predict"] = max(64, int(os.environ.get("CASAN_MODEL_GENERATE_MAX_TOKENS", "1400")))
    data = json.dumps(body).encode()
    req = urllib.request.Request(url, data=data, headers={"Content-Type": "application/json"})
    t0 = time.time()