diff --git a/5_django_backend/apps/chat/stream.py b/5_django_backend/apps/chat/stream.py
index d4c8139..e971fd5 100644
--- a/5_django_backend/apps/chat/stream.py
+++ b/5_django_backend/apps/chat/stream.py
@@ -228,6 +228,7 @@ async def _finalize(session: ChatSession, state: TurnState, started: float, *, f
     usage: dict | None = None
     title = state.title
     tool_parts: list[dict] = []
+    model = ""
     try:
         msgs = await opencode_service.list_messages_a(session.id) or []
         assistant = [m for m in msgs if (m.get("info") or {}).get("role") == "assistant"]
@@ -236,6 +237,7 @@ async def _finalize(session: ChatSession, state: TurnState, started: float, *, f
             parts = assistant[-1].get("parts") or []
             # 이 턴의 assistant 메시지가 여러 개(도구 왕복마다 하나)라 도구 파트는 전부 모음
             tool_parts = [p for m in assistant for p in (m.get("parts") or []) if p.get("type") == "tool"]
+            model = info.get("modelID") or ""
             full = "".join(p.get("text", "") for p in parts if p.get("type") == "text" and not p.get("synthetic"))
             if full:
                 content = full
@@ -276,9 +278,9 @@ async def _finalize(session: ChatSession, state: TurnState, started: float, *, f
         t0 = datetime.fromtimestamp(started, timezone.utc).isoformat()
         root = langfuse.trace(
             tid, "chat", userId=user_email, sessionId=session.id, input=state.user_text, output=content,
-            tags=["codeassist"], start=t0, end=langfuse.now_iso(),
-            usage=langfuse.usage_of(usage["input"], usage["output"]) if usage else None,
-            metadata={"elapsed_ms": usage["elapsed_ms"]} if usage else {}, error=failed,
+            tags=["codeassist", *langfuse.usage_tags(tool_parts)], start=t0, end=langfuse.now_iso(),
+            usage=langfuse.usage_of(usage["input"], usage["output"]) if usage else None, model=model,
+            metadata={"model": model, **({"elapsed_ms": usage["elapsed_ms"]} if usage else {})}, error=failed,
         )
         langfuse.send_later([root, *_tool_spans(tid, root, tool_parts)])
     return usage, title
diff --git a/5_django_backend/apps/gateway/langfuse.py b/5_django_backend/apps/gateway/langfuse.py
index 439cefa..205a4a1 100644
--- a/5_django_backend/apps/gateway/langfuse.py
+++ b/5_django_backend/apps/gateway/langfuse.py
@@ -150,6 +150,35 @@ def tool_span(trace_id: str, name: str, *, parent: dict, input=None, output=None
     return _span(trace_id, name, attrs, start=start, end=end, parent=parent["spanId"], error=error)
 
 
+def usage_tags(tool_parts: list[dict]) -> list[str]:
+    """이 턴에 뭘 써서 답했는지 태그로 — Phoenix 에서 metadata.tags 로 필터. 순서 유지·중복 제거.
+    skill:<이름> = OpenCode skill 도구, 또는 read 로 skills/<이름>/SKILL.md 를 연 것(ABAP_OPENCODE 방식)
+    wiki        = wiki/ 아래 파일을 읽음(CodeAssist 위키)
+    mcp:<함수>  = MCP 도구 호출. OpenCode 는 "<서버>_<함수>" 로 이름 붙임 → 서버 접두어 떼고 함수만"""
+    import re
+
+    out: list[str] = []
+    for p in tool_parts:
+        tool = p.get("tool") or ""
+        inp = (p.get("state") or {}).get("input") or {}
+        path = str(inp.get("filePath") or inp.get("path") or "").replace("\\", "/")
+        name = inp.get("name") if tool == "skill" else None
+        if not name and tool == "read":
+            m = re.search(r"skills/([^/]+)/SKILL\.md$", path)
+            name = m.group(1) if m else None
+        if name:
+            out.append(f"skill:{name}")
+        if "wiki/" in path:
+            out.append("wiki")
+        for server in MCP_SERVERS:
+            if tool.startswith(server + "_"):
+                out.append("mcp:" + tool[len(server) + 1:])
+    return list(dict.fromkeys(out))
+
+
+MCP_SERVERS = ("sap-icf",)  # opencode.json 의 mcp 이름들. 새 MCP 붙이면 여기 추가
+
+
 def ms_iso(ms: int | float | None) -> str | None:
     """OpenCode 의 epoch ms → ISO. 없으면 None(=지금)."""
     return datetime.fromtimestamp(ms / 1000, timezone.utc).isoformat() if ms else None
diff --git a/5_django_backend/tests/test_langfuse.py b/5_django_backend/tests/test_langfuse.py
index 809bccd..6b42c09 100644
--- a/5_django_backend/tests/test_langfuse.py
+++ b/5_django_backend/tests/test_langfuse.py
@@ -88,8 +88,9 @@ async def test_finalize_sends_user_trace(lf_server, django_user_model, monkeypat
     async def _msgs(_sid):
         tool = {"type": "tool", "tool": "sap-icf_get_program_source", "state": {"status": "completed", "input": {"program_name": "ZTEST"}, "output": "REPORT ztest.", "title": "ZTEST", "time": {"start": 1100, "end": 1300}}}
         bad = {"type": "tool", "tool": "read", "state": {"status": "error", "input": {"filePath": "x"}, "error": "no such file", "time": {"start": 1300, "end": 1310}}}
-        return [{"info": {"role": "assistant"}, "parts": [tool, bad]},
-                {"info": {"role": "assistant", "tokens": {"input": 3, "output": 4, "reasoning": 1}, "time": {"created": 1000, "completed": 1500}}, "parts": [{"type": "text", "text": "답"}]}]
+        wiki = {"type": "tool", "tool": "read", "state": {"status": "completed", "input": {"filePath": "/workspace/wiki/programs/ZTEST.md"}, "output": "x", "time": {"start": 1310, "end": 1320}}}
+        return [{"info": {"role": "assistant"}, "parts": [tool, bad, wiki]},
+                {"info": {"role": "assistant", "modelID": "z-ai/glm-5.2", "tokens": {"input": 3, "output": 4, "reasoning": 1}, "time": {"created": 1000, "completed": 1500}}, "parts": [{"type": "text", "text": "답"}]}]
 
     async def _sess(_sid):
         return {"title": "MARA 조회"}
@@ -103,11 +104,14 @@ async def test_finalize_sends_user_trace(lf_server, django_user_model, monkeypat
         await asyncio.gather(*list(langfuse._pending))
     assert len(lf_server) == 1
     spans = lf_server[0]["body"]["resourceSpans"][0]["scopeSpans"][0]["spans"]
-    assert [sp["name"] for sp in spans] == ["chat", "sap-icf_get_program_source", "read"]
+    assert [sp["name"] for sp in spans] == ["chat", "sap-icf_get_program_source", "read", "read"]
     t = _attrs(spans[0])
     assert t["langfuse.user.id"] == "u@x.com" and t["langfuse.session.id"] == "ses_1"
     assert t["langfuse.trace.input"] == "질문" and t["langfuse.trace.output"] == "답"
     assert t["langfuse.observation.type"] == "generation"  # 루트가 곧 답변 — 토큰은 여기
+    assert t["llm.model_name"] == "z-ai/glm-5.2"
+    meta = json.loads(t["metadata"])
+    assert meta["model"] == "z-ai/glm-5.2" and meta["tags"] == ["codeassist", "mcp:get_program_source", "wiki"]
     assert json.loads(t["langfuse.observation.usage_details"]) == {"input": 3, "output": 5, "total": 8}
     a1, a2 = _attrs(spans[1]), _attrs(spans[2])
     assert spans[1]["parentSpanId"] == spans[0]["spanId"] and a1["langfuse.observation.type"] == "tool"
@@ -131,3 +135,19 @@ async def test_phoenix_target_no_auth_and_openinference_attrs(lf_server):
     assert r["llm.model_name"] == "581" and r["llm.token_count.total"] == "10" and r["session.id"] == "s9" and r["user.id"] == "u@x.com"
     assert tl["openinference.span.kind"] == "TOOL" and tl["tool.name"] == "read" and json.loads(tl["input.value"]) == {"filePath": "a.md"}
 
+
+def test_usage_tags_skill_wiki_mcp():
+    parts = [
+        {"tool": "skill", "state": {"input": {"name": "classic-abap-fs"}}},
+        {"tool": "read", "state": {"input": {"filePath": "/workspace/skills/abap-program-analyze/SKILL.md"}}},
+        {"tool": "read", "state": {"input": {"filePath": "C:\\ws\\wiki\\programs\\Z.md"}}},
+        {"tool": "sap-icf_get_table_fields", "state": {"input": {"table": "MARA"}}},
+        {"tool": "sap-icf_get_program_source", "state": {"input": {"program_name": "Z"}}},
+        {"tool": "read", "state": {"input": {"filePath": "/workspace/skills/classic-abap-fs/glossary.md"}}},  # SKILL.md 아님
+        {"tool": "grep", "state": {"input": {"pattern": "x"}}},  # 태그 없음
+        {"tool": "skill", "state": {"input": {"name": "classic-abap-fs"}}},  # 중복
+    ]
+    assert langfuse.usage_tags(parts) == [
+        "skill:classic-abap-fs", "skill:abap-program-analyze", "wiki", "mcp:get_table_fields", "mcp:get_program_source",
+    ]
+