Merge pull request #377 from alichherawalla/fix/remote-drop-max-tokens-cap

fix(remote): stop sending max_tokens/num_predict to remote servers

Author: dishit-wednesday

src/services/providers/openAICompatibleProvider.ts

@@ -99,7 +99,9 @@ export class OpenAICompatibleProvider implements LLMProvider {
       messages: openaiMessages,
       stream: true,
       ...(options.temperature !== undefined && { temperature: options.temperature }),
-      ...(options.maxTokens !== undefined && { max_tokens: options.maxTokens }),
+      // max_tokens intentionally omitted — the remote server owns output limits.
+      // A client-side cap (default 1024) silently truncates reasoning models that
+      // need a larger budget for <think> blocks (Qwen3, DeepSeek-R1, etc).
       ...(options.topP !== undefined && { top_p: options.topP }),
       ...(options.tools && options.tools.length > 0 && { tools: options.tools, tool_choice: 'auto' }),
       // LM Studio only: control Qwen3 thinking per-request via chat_template_kwargs.

src/services/providers/openAICompatibleStream.ts

@@ -289,7 +289,8 @@ export async function generateOllamaChatImpl(
     ...(options.tools && options.tools.length > 0 && { tools: options.tools }),
     options: {
       ...(options.temperature !== undefined && { temperature: options.temperature }),
-      ...(options.maxTokens !== undefined && { num_predict: options.maxTokens }),
+      // num_predict intentionally omitted — Ollama defaults to -1 (until natural stop).
+      // A client-side cap truncates reasoning models mid-<think>.
       ...(options.topP !== undefined && { top_p: options.topP }),
     },
   };