vLLM: tool_choice="required" with hermes structural tag returns only first of several parallel tool calls

Finding: In vLLM main c23ca06b6d (0.31.1rc1.dev150, xgrammar 0.2.8), serving a Hermes-style model with --tool-call-parser hermes --enable-auto-tool-choice, a request with tool_choice="required" reportedly returns only one tool call when several are needed, whereas tool_choice="auto" returns all of them (https://github.com/vllm-project/vllm/issues/60769). This is an external report; KFA did not reproduce it. Reported conditions: Qwen/Qwen3-0.6B with a single get_weather(city) tool, the prompt asking for weather in Paris and Tokyo, temperature 0, and enable_thinking false. Auto gave 2 calls separated by a newline in the raw output; required gave 1 call (Paris) with finish_reason tool_calls. A three-city prompt gave 3 calls under auto and 1 under required, and a prompt needing two different tools returned only the first. Streaming and non-streaming behave the same. Single-call required and named tool choice are reported fine. The reporter also saw the same on Qwen2.5-1.5B-Instruct and Hermes-3-Llama-3.2-3B, and on both the Python and Rust frontends. The reporter further says that on the Python frontend, auto with a tool marked strict: true plus a response_format reuses the required list and drops calls the same way (same issue). Reporter's hypothesis (not maintainer-confirmed): the required-mode Hermes structural tag uses tags_with_separator with an empty separator (reported at structural_tag_registry.py lines 373-406, separator at line 402). Hermes-style templates place a newline between calls, so the grammar rejects that newline but still allows EOS, ending greedy decoding after the first call. The issue includes a CPU-only xgrammar reproducer that reportedly accepts one call but rejects two newline-separated calls on main; KFA has not run it. A commenter reports running a repro on vLLM v0.30.0 where the grammar rejected two newline-separated calls, suggesting the behavior is not limited to current main (https://github.com/vllm-project/vllm/issues/60769#issuecomment-6085340771). That commenter states it was not a confirmed workaround. Status: the issue is open with no maintainer response, no fix PR, and no release identified in the supplied sources. No verified workaround exists. Expected behavior per the reporter is that required should accept the template's newline framing between calls when parallel_tool_calls is enabled. Related note: the reporter says ERNIE 4.5 needs different framing handling, so a single newline separator would not cover it. Limitations: external report only, no KFA reproduction, cause unconfirmed by maintainers, and comment coverage may omit middle comments.