support muse glimmer tool/reasoning parser

src/cli/groups/add.py

@@ -63,10 +63,10 @@
     type=float,
     help='Confidence threshold for accepting draft tokens.')
 @click.option('--tool-call-parser',
-    type=click.Choice(['qwen35', 'hermes', 'gemma4']),
+    type=click.Choice(['qwen35', 'hermes', 'gemma4', 'museglimmer']),
     required=False,
     default=None,
-    help='Tool-call output format for this model (qwen35 XML, hermes JSON, or gemma4 call syntax). llm/vlm only; required for tool calling.')
+    help='Tool-call output format for this model (qwen35 XML, hermes JSON, gemma4 call syntax, or museglimmer Harmony atem). llm/vlm only; required for tool calling.')
 @click.pass_context
 def add(ctx, model_path, model_name, engine, model_type, device, runtime_config, scheduler_config, cache_dir, draft_model_path, draft_device, num_assistant_tokens, assistant_confidence_threshold, tool_call_parser):
     """- Add a model configuration to the config file."""

src/engine/ov_genai/llm.py

@@ -102,9 +102,12 @@ async def generate_text(self, gen_config: OVGenAI_GenConfig) -> AsyncIterator[Un
 
         perf_metrics = result.perf_metrics
         decoder_tokenizer = self.model.get_tokenizer()
-        # gemma4 protocol tags are special=True: keep them in the decoded text
-        # so the route-level parse_generation can split reasoning/tool calls.
-        keep_special = getattr(gen_config, "tool_call_parser", None) == "gemma4"
+        # gemma4/museglimmer protocol tags are special=True: keep them in the
+        # decoded text so the route-level parse_generation can split
+        # reasoning/tool calls.
+        keep_special = getattr(gen_config, "tool_call_parser", None) in (
+            "gemma4", "museglimmer",
+        )
         text = (
             decoder_tokenizer.decode(result.tokens, skip_special_tokens=not keep_special)[0]
             if getattr(result, "tokens", None) else ""

src/engine/ov_genai/streamers.py

@@ -5,6 +5,7 @@
 from openvino_genai import StreamerBase
 from src.server.schemas.modeling.contract_ovgenai_llm_and_vlm import OVGenAI_GenConfig
 from src.engine.ov_genai.tool_parse import gemma4 as gemma4_tool_parse
+from src.engine.ov_genai.tool_parse import museglimmer as museglimmer_tool_parse
 from src.engine.ov_genai.tool_parse import qwen35 as qwen35_tool_parse
 
 
@@ -101,7 +102,10 @@ def select_streamer(tokenizer, gen_config: OVGenAI_GenConfig) -> StreamerBase:
     (token-ID block boundaries, parsed OpenAI deltas). gemma4 requests use
     Gemma4ToolCallStreamer whenever tools are requested OR thinking is enabled
     (its thought-channel tags are special=True, so the plain text path cannot
-    split reasoning); everything else uses ChunkStreamer. All of them enqueue
+    split reasoning). museglimmer requests ALWAYS use MuseGlimmerToolCallStreamer:
+    its Harmony 'to=' channel routing is token-ID-only, so even a plain content
+    response opens with ' to=user<|message|>' header text that only the engine
+    streamer can strip. Everything else uses ChunkStreamer. All of them enqueue
     on .text_queue, so consumers are unaffected.
     """
     parser_name = getattr(gen_config, "tool_call_parser", None)
@@ -113,4 +117,6 @@ def select_streamer(tokenizer, gen_config: OVGenAI_GenConfig) -> StreamerBase:
             thinking = bool(gen_config.chat_template_kwargs.get("enable_thinking", True))
         if gen_config.tools or thinking:
             return gemma4_tool_parse.Gemma4ToolCallStreamer(tokenizer, gen_config)
+    if parser_name == "museglimmer":
+        return museglimmer_tool_parse.MuseGlimmerToolCallStreamer(tokenizer, gen_config)
     return ChunkStreamer(tokenizer, gen_config)