From 2554636ded46f87430cac6c20a35c828801585c8 Mon Sep 17 00:00:00 2001
From: Wasim Yousef Said <wasimysdev@gmail.com>
Date: Tue, 9 Jun 2026 18:11:38 +0200
Subject: [PATCH] Studio: follow-up fix for GGUF developer prompts (#6115)

* Studio: merge developer prompts for GGUF chat

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/routes/inference.py            | 13 ++--
 .../tests/test_openai_tool_passthrough.py     | 60 +++++++++++++++++++
 2 files changed, 68 insertions(+), 5 deletions(-)

diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index a2e0e5a811..81bc9ff628 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -2050,8 +2050,8 @@ def _extract_content_parts(messages: list) -> tuple[str, list[dict], "Optional[s
     first_image_b64: Optional[str] = None
 
     for msg in messages:
-        # ── System messages → extract as system_prompt ────────
-        if msg.role == "system":
+        # ── System / developer messages → extract as system_prompt ────────
+        if msg.role in ("system", "developer"):
             if isinstance(msg.content, str):
                 system_parts.append(msg.content)
             elif isinstance(msg.content, list):
@@ -3053,6 +3053,7 @@ async def openai_chat_completions(
             payload,
             llama_backend.is_vision,
         )
+        gguf_messages = _set_or_prepend_system_message(gguf_messages, system_prompt)
         image_b64 = None
 
         cancel_event = threading.Event()
@@ -5190,9 +5191,9 @@ def _set_or_prepend_system_message(
     if not system_prompt:
         return safe_messages
 
-    # Drop existing system turns so the backend never sees duplicate or
-    # conflicting system instructions, then prepend the resolved prompt.
-    others = [dict(msg) for msg in safe_messages if msg.get("role") != "system"]
+    # Drop existing system/developer turns so the backend never sees duplicate
+    # or conflicting system instructions, then prepend the resolved prompt.
+    others = [dict(msg) for msg in safe_messages if msg.get("role") not in ("system", "developer")]
     return [{"role": "system", "content": system_prompt}, *others]
 
 
@@ -6415,6 +6416,8 @@ def _build_openai_passthrough_body(payload, backend_ctx = None) -> dict:
     leak to the backend.
     """
     messages = _openai_messages_for_passthrough(payload)
+    system_prompt, _, _ = _extract_content_parts(payload.messages)
+    messages = _set_or_prepend_system_message(messages, system_prompt)
     tool_choice = payload.tool_choice if payload.tool_choice is not None else "auto"
     # When the caller asked for a specific reasoning mode, forward it via
     # chat_template_kwargs so the Jinja template renders with (or without) the
diff --git a/studio/backend/tests/test_openai_tool_passthrough.py b/studio/backend/tests/test_openai_tool_passthrough.py
index 495ebc434d..db4c07cdc8 100644
--- a/studio/backend/tests/test_openai_tool_passthrough.py
+++ b/studio/backend/tests/test_openai_tool_passthrough.py
@@ -33,6 +33,7 @@ from core.inference.anthropic_compat import (
     anthropic_tool_choice_to_openai,
 )
 from routes.inference import (
+    _build_openai_passthrough_body,
     _build_passthrough_payload,
     _clamp_finish_reason,
     _effective_max_tokens,
@@ -634,6 +635,24 @@ class TestBuildPassthroughPayloadToolChoice:
         assert body.get("repeat_penalty") == 1.1
         assert "repetition_penalty" not in body
 
+    def test_passthrough_body_merges_system_and_developer_messages(self):
+        payload = ChatCompletionRequest(
+            model = "default",
+            messages = [
+                {"role": "system", "content": "original system"},
+                {"role": "developer", "content": "developer rules"},
+                {"role": "user", "content": "hi"},
+            ],
+            tools = self._args()["openai_tools"],
+        )
+
+        body = _build_openai_passthrough_body(payload, backend_ctx = 4096)
+
+        assert body["messages"] == [
+            {"role": "system", "content": "original system\n\ndeveloper rules"},
+            {"role": "user", "content": "hi"},
+        ]
+
 
 # =====================================================================
 # OpenAI API compatibility helpers — verified spec edge cases
@@ -1092,6 +1111,47 @@ class TestGgufVisionToolRouting:
 
         assert captured["kwargs"]["disable_parallel_tool_use"] is True
 
+    def test_standard_gguf_merges_system_and_developer_messages(self, monkeypatch):
+        import routes.inference as inf_mod
+
+        captured = {}
+
+        def _generate(**kwargs):
+            captured["messages"] = kwargs["messages"]
+            yield "done"
+            yield {
+                "type": "metadata",
+                "usage": {"prompt_tokens": 3, "completion_tokens": 1, "total_tokens": 4},
+                "finish_reason": "stop",
+            }
+
+        backend = SimpleNamespace(
+            is_loaded = True,
+            is_vision = False,
+            supports_tools = False,
+            model_identifier = "test-gguf",
+            generate_chat_completion = _generate,
+        )
+        monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+        payload = ChatCompletionRequest(
+            model = "default",
+            messages = [
+                {"role": "system", "content": "original system"},
+                {"role": "developer", "content": "developer rules"},
+                {"role": "user", "content": "hi"},
+            ],
+        )
+
+        self._drive(
+            openai_chat_completions(payload, request = self._Request(), current_subject = "test")
+        )
+
+        assert captured["messages"] == [
+            {"role": "system", "content": "original system\n\ndeveloper rules"},
+            {"role": "user", "content": "hi"},
+        ]
+
     @pytest.mark.parametrize(
         ("seed", "expected"),
         [