don't abort reasoning when emitting a tool call

Author: aviallon

common/chat-parser.cpp

@@ -1513,17 +1513,24 @@ common_chat_msg common_chat_parse(const std::string & input, bool is_partial, co
         return common_chat_peg_parse(syntax.parser, input, is_partial, syntax);
     }
     common_chat_msg_parser builder(input, is_partial, syntax);
+    bool partial_exception_caught = false;
     try {
         common_chat_parse(builder);
     } catch (const common_chat_msg_partial_exception & ex) {
         LOG_DBG("Partial parse: %s\n", ex.what());
+        partial_exception_caught = true;
         if (!is_partial) {
             builder.clear_tools();
             builder.move_to(0);
             common_chat_parse_content_only(builder);
         }
     }
     auto msg = builder.result();
+    // Mark tool_call_in_progress if we caught a partial exception during partial parsing
+    // and there are tool calls in progress (indicates incomplete tool call parsing)
+    if (is_partial && partial_exception_caught && !msg.tool_calls.empty()) {
+        msg.tool_call_in_progress = true;
+    }
     if (!is_partial) {
         LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat<json>({msg}).at(0).dump().c_str());
     }

common/chat.h

@@ -51,6 +51,7 @@ struct common_chat_msg {
     std::vector<common_chat_tool_call> tool_calls;
     std::string reasoning_content;
     common_chat_reasoning_status reasoning_status;
+    bool tool_call_in_progress = false;
     std::string tool_name;
     std::string tool_call_id;
 

tools/server/server-context.cpp

@@ -1210,7 +1210,11 @@ struct server_context_impl {
             if (slot.reasoning == REASONING_STATE_REASONING) {
                 slot.n_reasoning_tokens++;
 
-                if (slot.n_reasoning_tokens >= reasoning_budget) {
+                // Detect if we are in the middle of emitting a tool call this step.
+                // The parser sets tool_call_in_progress when it catches a partial exception
+                // while parsing tool calls, indicating incomplete tool call parsing.
+                // We also check for tool call diffs in this token as a fallback.
+                if (!parsed_msg.tool_call_in_progress && slot.n_reasoning_tokens >= reasoning_budget) {
                     SLT_INF(slot, "reasoning budget exceeded, forcing close with '%s', n_reasoning_tokens = %d, reasoning_budget = %d\n",
                         slot.reasoning_end_tag.c_str(), slot.n_reasoning_tokens, reasoning_budget);