More comprehensive handling in count_tokens(force_accurate==True): In case model is neither OpenAI nor Anthropic Claude, simply use an elbow room factor in order to force a more conservative estimate.

pr_agent/algo/token_handler.py

@@ -1,6 +1,7 @@
 from threading import Lock
 
 from jinja2 import Environment, StrictUndefined
+from math import ceil
 from tiktoken import encoding_for_model, get_encoding
 
 from pr_agent.config_loader import get_settings
@@ -114,6 +115,22 @@ class TokenHandler:
         Returns:
         The number of tokens in the patch string.
         """
-        if force_accurate and 'claude' in get_settings().config.model.lower() and get_settings(use_context=False).get('anthropic.key'):
+        encoder_estimate = len(self.encoder.encode(patch, disallowed_special=()))
+        if not force_accurate:
+            return encoder_estimate
+        #else, need to provide an accurate estimation:
+
+        model = get_settings().config.model.lower()
+        if force_accurate and 'claude' in model and get_settings(use_context=False).get('anthropic.key'):
             return self.calc_claude_tokens(patch) # API call to Anthropic for accurate token counting for Claude models
-        return len(self.encoder.encode(patch, disallowed_special=()))
+        #else: Non Anthropic provided model
+
+        import re
+        model_is_from_o_series = re.match(r"^o[1-9](-mini|-preview)?$", model)
+        if ('gpt' in get_settings().config.model.lower() or model_is_from_o_series) and get_settings(use_context=False).get('openai.key'):
+            return encoder_estimate
+        #else: Model is neither an OpenAI, nor an Anthropic model - therefore, cannot provide an accurate token count and instead, return a higher number as best effort.
+
+        elbow_factor = 1 + get_settings().get('config.model_token_count_estimate_factor', 0)
+        get_logger().warning(f"{model}'s expected token count cannot be accurately estimated. Using {elbow_factor} of encoder output as best effort estimate")
+        return ceil(elbow_factor * encoder_estimate)