core/agent: todo review nag as developer role, not fake user

Internal control uses DeveloperChatMessage so retry/history treat it as
non-human. Responses bridge maps developer mid-turn; compact/history label it.

src/plyngent/agent/compact.py

@@ -8,6 +8,7 @@ from plyngent.lmproto.openai_compatible.model import (
     AssistantChatMessage,
     AssistantFunctionToolCall,
     ChatCompletionsParam,
+    DeveloperChatMessage,
     SystemChatMessage,
     ToolChatMessage,
     UserChatMessage,
@@ -43,12 +44,14 @@ _SEED_MESSAGE_TEMPLATE = (
 )
 
 
-def format_transcript(messages: Sequence[AnyChatMessage]) -> str:
+def format_transcript(messages: Sequence[AnyChatMessage]) -> str:  # noqa: C901
     """Render messages as plain text for a summarization prompt."""
     lines: list[str] = []
     for msg in messages:
         if isinstance(msg, SystemChatMessage):
             lines.append(f"[system] {msg.content}")
+        elif isinstance(msg, DeveloperChatMessage):
+            lines.append(f"[developer] {msg.content}")
         elif isinstance(msg, UserChatMessage):
             lines.append(f"[user] {msg.content}")
         elif isinstance(msg, AssistantChatMessage):

src/plyngent/agent/loop.py

@@ -13,10 +13,10 @@ from plyngent.lmproto.openai_compatible.model import (
     AssistantChatMessage,
     AssistantFunctionToolCall,
     ChatCompletionsParam,
+    DeveloperChatMessage,
     StreamOptions,
     StreamToolCallDelta,
     ToolChatMessage,
-    UserChatMessage,
 )
 from plyngent.typedef import Unset  # noqa: TC001
 
@@ -377,7 +377,8 @@ async def run_chat_loop(
                     and not todo_review_injected
                 ):
                     todo_review_injected = True
-                    messages.append(UserChatMessage(content=todo_stack.review_prompt()))
+                    # Non-user control identity so retry/history don't treat this as human input.
+                    messages.append(DeveloperChatMessage(content=todo_stack.review_prompt()))
                     continue
                 return
             async for event in _execute_tool_calls(

src/plyngent/agent/responses_bridge.py

@@ -29,6 +29,7 @@ from plyngent.lmproto.openai_compatible.model import (
     ChatCompletionsParam,
     ChunkChoice,
     DeltaMessage,
+    DeveloperChatMessage,
     StreamFunctionDelta,
     StreamToolCallDelta,
     SystemChatMessage,
@@ -97,6 +98,10 @@ def chat_messages_to_responses_input(
         if isinstance(message, SystemChatMessage):
             if message.content.strip():
                 instructions_parts.append(message.content)
+        elif isinstance(message, DeveloperChatMessage):
+            # Keep mid-turn control as input developer messages (not folded into instructions).
+            if message.content.strip():
+                items.append(ResponseEasyInputMessage(role="developer", content=message.content))
         elif isinstance(message, UserChatMessage):
             items.append(ResponseEasyInputMessage(role="user", content=message.content))
         elif isinstance(message, AssistantChatMessage):

src/plyngent/agent/todo_stack.py

@@ -99,9 +99,9 @@ class TodoStack:
         return "\n".join(lines)
 
     def review_prompt(self) -> str:
-        """User-message body injected when the model finishes without todo ops."""
+        """Control-message body (developer role) when the model finishes without todo ops."""
         return (
-            "[system: todo stack review]\n"
+            "Todo stack review (internal control — not a human message).\n"
             "Open sub-tasks remain and you did not call any todo_* tools this turn.\n"
             "Review the stack: mark finished items done, update in_progress, "
             "push new sub-tasks, or pop/cancel obsolete ones. Then continue the user work.\n\n"

src/plyngent/cli/slash.py

@@ -15,6 +15,8 @@ from plyngent.cli.selection import select_model, select_provider
 from plyngent.lmproto.openai_compatible.model import (
     AssistantChatMessage,
     AssistantFunctionToolCall,
+    DeveloperChatMessage,
+    SystemChatMessage,
     ToolChatMessage,
     UserChatMessage,
 )
@@ -973,6 +975,10 @@ def _preview_content(text: str | None) -> str:
 def _format_history_message(index: int, message: AnyChatMessage) -> str:
     if isinstance(message, UserChatMessage):
         return f"{index}. user: {_preview_content(message.content)}"
+    if isinstance(message, DeveloperChatMessage):
+        return f"{index}. developer: {_preview_content(message.content)}"
+    if isinstance(message, SystemChatMessage):
+        return f"{index}. system: {_preview_content(message.content)}"
     if isinstance(message, AssistantChatMessage):
         parts: list[str] = []
         if isinstance(message.content, str) and message.content:

tests/test_agent/test_todo_stack.py

@@ -13,6 +13,7 @@ from plyngent.lmproto.openai_compatible.model import (
     ChatCompletionChunk,
     ChatCompletionResponse,
     ChatCompletionsParam,
+    DeveloperChatMessage,
     UserChatMessage,
 )
 from plyngent.memory import MemoryStore
@@ -102,7 +103,7 @@ class ScriptedClient:
         text = "ok" if self.calls > 1 else "done without todos"
         # Detect review message in history
         for msg in param.messages:
-            if isinstance(msg, UserChatMessage) and "todo stack review" in msg.content:
+            if isinstance(msg, DeveloperChatMessage) and "Todo stack review" in msg.content:
                 text = "reviewed stack"
                 break
         return ChatCompletionResponse(
@@ -142,7 +143,11 @@ async def test_loop_injects_todo_review_when_untouched() -> None:
             pass
         assert client.calls >= 2
         assert any(
-            isinstance(m, UserChatMessage) and "todo stack review" in m.content for m in agent.messages
+            isinstance(m, DeveloperChatMessage) and "Todo stack review" in m.content
+            for m in agent.messages
+        )
+        assert not any(
+            isinstance(m, UserChatMessage) and "Todo stack review" in m.content for m in agent.messages
         )
     finally:
         set_todo_stack(None)