fix: discard stale todo reminders during context compaction (#5614)

* fix: rebuild todo reminders after context compaction

* chore: remove implementation plan from PR

* refactor: share todo reminder message name

* docs: keep agent guidance within CI size budgets

---------

Co-authored-by: Willem Jiang <willem.jiang@gmail.com>

README.md

@@ -1495,6 +1495,12 @@ When your role lacks `runs:create`, the Web UI rejects a new task or `/goal <com
 
 ### Manual Context Compaction
 
+Automatic and manual compaction exclude old todo reminder messages from both the
+summary input and retained context. The current todo list stays in thread state.
+In planning mode, if the original `write_todos` call is no longer visible,
+DeerFlow adds a reminder using the latest task statuses before the next model
+call. Skipped or failed compaction leaves the existing messages unchanged.
+
 Optional `pii_redaction.enabled` redacts detected identifiers in user messages,
 remote tool results, compaction input, reinjected summaries, and configured
 LLM title input. It is off by default. Existing summary placeholders reserve

backend/app/gateway/AGENTS.md

@@ -5,8 +5,7 @@ credential fields and creates a normal MCP connection. The MCP API accepts the
 exact isolated interpreter/module/provider launcher generated by
 `deerflow.capabilities.business`, including the exact credential environment key
 set. Do not allow arbitrary Python commands, trust manifest metadata to bypass
-execution policy, or put credentials into tool schemas. Existing admin checks,
+execution policy, or put credentials into tool schemas. Keep admin checks, masked edits, atomic configuration writes and MCP cache reloads.
-masked edits, atomic configuration writes and MCP cache reloads remain owners.
 
 Memory shutdown resolves hot-reloaded config and the backend, flushes, then
 closes as one `await_drained` operation. Keep config resolution inside the

backend/docs/summarization.md

@@ -14,6 +14,20 @@ The summarization feature uses LangChain's `SummarizationMiddleware` to monitor
 4. Maintains AI/Tool message pairs together for context continuity
 5. Stores the summary in `ThreadState.summary_text` and projects it ephemerally through durable context data
 
+## Todo reminders
+
+Compaction filters `HumanMessage(name="todo_reminder")` snapshots after the
+trigger check and before selecting the retained tail. They enter neither summary
+generation/pre-compaction hooks nor retained messages, while `state["todos"]`
+remains unchanged. Only a successful compaction commits the removal; no-op and
+failure paths keep the original state. The following `TodoMiddleware.before_model`
+rebuilds one reminder from current todos when no `write_todos` call is still
+visible; empty todos need no reminder. Both automatic and manual compaction use
+this shared preparation path. Coverage: `tests/test_todo_compaction.py`.
+`todo_middleware.py::TODO_REMINDER_MESSAGE_NAME` owns the backend message name;
+the producer, presence check, and compaction filter share it. Its value remains
+`todo_reminder` for compatibility with the frontend's hidden-message filtering.
+
 ## Configuration
 
 Summarization is configured in `config.yaml` under the `summarization` key:

backend/packages/harness/deerflow/agents/middlewares/AGENTS.md

@@ -7,6 +7,8 @@ mixed windows keep normal anchoring and final-message fallback. Budget raw text
 before escaping/wrapping; pass `trim_tokens_to_summarize=None` to avoid the
 LangChain default.
 
+Todo compaction: [contract](../../../../../docs/summarization.md#todo-reminders).
+
 Delegation verdicts are untrusted: revalidate persisted values, ignore malformed
 ones, and treat completed work as reusable evidence rather than acceptance.
 

backend/packages/harness/deerflow/agents/middlewares/summarization_middleware.py

@@ -20,6 +20,7 @@ from langgraph.runtime import Runtime
 from deerflow.agents.middlewares.dynamic_context_middleware import is_dynamic_context_reminder
 from deerflow.agents.middlewares.message_utils import is_genuine_user_message
 from deerflow.agents.middlewares.pii_redaction_middleware import redact_text
+from deerflow.agents.middlewares.todo_middleware import TODO_REMINDER_MESSAGE_NAME
 from deerflow.config.app_config import get_app_config
 from deerflow.config.summarization_config import DEFAULT_KEEP
 from deerflow.config.task_continuity_config import TaskContinuityConfig
@@ -576,6 +577,13 @@ class DeerFlowSummarizationMiddleware(SummarizationMiddleware):
         if not force and not self._should_summarize(trigger_messages, total_tokens):
             return None
 
+        # Todo reminders are snapshots of state["todos"], not conversation history.
+        # Exclude them before partitioning so neither the summary nor the retained
+        # tail contains stale task statuses. TodoMiddleware restores current context
+        # when needed before the next model call. Keep state untouched if compaction
+        # is skipped or summary generation fails.
+        messages = [message for message in messages if not (isinstance(message, HumanMessage) and message.name == TODO_REMINDER_MESSAGE_NAME)]
+
         cutoff_index = self._determine_cutoff_index(messages)
         if cutoff_index <= 0:
             return None

backend/packages/harness/deerflow/agents/middlewares/todo_middleware.py

@@ -31,6 +31,8 @@ from langgraph.runtime import Runtime
 
 from deerflow.agents.thread_state import ThreadState
 
+TODO_REMINDER_MESSAGE_NAME = "todo_reminder"
+
 
 def _todos_in_messages(messages: list[Any]) -> bool:
     """Return True if any AIMessage in *messages* contains a write_todos tool call."""
@@ -45,7 +47,7 @@ def _todos_in_messages(messages: list[Any]) -> bool:
 def _reminder_in_messages(messages: list[Any]) -> bool:
     """Return True if a todo_reminder HumanMessage is already present in *messages*."""
     for msg in messages:
-        if isinstance(msg, HumanMessage) and getattr(msg, "name", None) == "todo_reminder":
+        if isinstance(msg, HumanMessage) and getattr(msg, "name", None) == TODO_REMINDER_MESSAGE_NAME:
             return True
     return False
 
@@ -150,7 +152,7 @@ class TodoMiddleware(TodoListMiddleware):
         # Inject a reminder as a HumanMessage so the model stays aware.
         formatted = _format_todos(todos)
         reminder = HumanMessage(
-            name="todo_reminder",
+            name=TODO_REMINDER_MESSAGE_NAME,
             additional_kwargs={"hide_from_ui": True},
             content=(
                 "<system_reminder>\n"

backend/tests/test_todo_compaction.py

@@ -0,0 +1,149 @@
+"""Todo context survives compaction through state, not stale reminder snapshots."""
+
+from copy import deepcopy
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+from langchain.agents import create_agent
+from langchain_core.language_models import BaseChatModel
+from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, ToolMessage
+from langchain_core.outputs import ChatGeneration, ChatResult
+from pydantic import PrivateAttr
+
+from deerflow.agents.middlewares.dynamic_context_middleware import _DYNAMIC_CONTEXT_REMINDER_KEY
+from deerflow.agents.middlewares.summarization_middleware import DeerFlowSummarizationMiddleware
+from deerflow.agents.middlewares.todo_middleware import TODO_REMINDER_MESSAGE_NAME, TodoMiddleware
+from deerflow.agents.thread_state import ThreadState
+
+
+def _reminder(msg_id: str) -> HumanMessage:
+    return HumanMessage(
+        id=msg_id,
+        name=TODO_REMINDER_MESSAGE_NAME,
+        content="STALE TODO: [pending] Inspect the Gateway routes",
+        additional_kwargs={"hide_from_ui": True},
+    )
+
+
+def _todos() -> list[dict[str, str]]:
+    return [
+        {"content": "Inspect the Gateway routes", "status": "completed"},
+        {"content": "Document the authentication endpoints", "status": "completed"},
+    ]
+
+
+def _messages(*, visible_write_todos: bool = False) -> list:
+    name = "write_todos" if visible_write_todos else "read_file"
+    args = {"todos": _todos()} if visible_write_todos else {"path": "backend/app/gateway/routers/auth.py"}
+    return [
+        SystemMessage(content="Document the project accurately.", id="system"),
+        HumanMessage(content="Document the Gateway routes.", id="old-user"),
+        _reminder("old-todo"),
+        AIMessage(content="I inspected the route definitions.", id="old-ai"),
+        HumanMessage(content="Include the authentication endpoints.", id="current-user"),
+        AIMessage(content="", tool_calls=[{"name": name, "args": args, "id": "recent-call"}], id="recent-ai"),
+        ToolMessage(content="Current endpoint documentation is ready.", tool_call_id="recent-call", name=name, id="recent-tool"),
+        _reminder("recent-todo"),
+    ]
+
+
+def _middleware(*, trigger: int = 4, text: str = "Gateway route documentation is ready.", hooks=None):
+    model = MagicMock()
+    model.invoke.return_value = SimpleNamespace(text=text)
+    model.ainvoke = AsyncMock(return_value=SimpleNamespace(text=text))
+    model.with_config.return_value = model
+    return DeerFlowSummarizationMiddleware(model=model, trigger=("messages", trigger), keep=("messages", 3), token_counter=len, before_summarization=hooks)
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("asynchronous", [False, True])
+@pytest.mark.parametrize("force", [False, True], ids=["automatic", "manual"])
+async def test_compaction_excludes_todo_reminders_from_summary_and_retained_context(asynchronous, force):
+    events = []
+    middleware = _middleware(trigger=1000 if force else 4, hooks=[events.append])
+    messages = _messages()
+    memory = HumanMessage(content="User prefers detailed API examples.", id="memory", additional_kwargs={"hide_from_ui": True, _DYNAMIC_CONTEXT_REMINDER_KEY: True})
+    # An unrelated hidden message must not be removed with the todo reminders.
+    messages.insert(1, memory)
+    state = {"messages": messages, "todos": _todos()}
+    original = deepcopy(state)
+    runtime = SimpleNamespace(context={"thread_id": "docs", "run_id": "docs-run"})
+
+    result = await middleware.acompact_state(state, runtime, force=force) if asynchronous else middleware.compact_state(state, runtime, force=force)
+
+    assert result is not None
+    assert [m.id for m in result.messages_to_summarize] == ["old-user", "old-ai"]
+    assert [m.id for m in result.preserved_messages] == ["system", "memory", "current-user", "recent-ai", "recent-tool"]
+    call = middleware.model.ainvoke.call_args if asynchronous else middleware.model.invoke.call_args
+    assert "STALE TODO" not in str(call)
+    assert "Gateway routes" in str(call)
+    assert events[0].messages_to_summarize == result.messages_to_summarize
+    assert events[0].preserved_messages == result.preserved_messages
+    assert state == original
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("asynchronous", [False, True])
+@pytest.mark.parametrize("reason", ["below-trigger", "summary-failed", "no-history"])
+async def test_unsuccessful_compaction_leaves_todo_reminders_and_state_untouched(asynchronous, reason):
+    events = []
+    middleware = _middleware(trigger=1000 if reason == "below-trigger" else 4, text=" " if reason == "summary-failed" else "summary", hooks=[events.append])
+    messages = _messages()
+    if reason == "no-history":
+        messages = [SystemMessage(content="Document the project.", id="system"), HumanMessage(content="Document auth.", id="user"), _reminder("todo-1"), _reminder("todo-2")]
+    state = {"messages": messages, "todos": _todos(), "summary_text": "previous summary"}
+    original = deepcopy(state)
+    runtime = SimpleNamespace(context={"thread_id": "docs", "run_id": "docs-run"})
+
+    result = await middleware.abefore_model(state, runtime) if asynchronous else middleware.before_model(state, runtime)
+
+    assert result is None
+    assert state == original
+    assert events == []
+
+
+class _CapturingModel(BaseChatModel):
+    _seen: list[list] = PrivateAttr(default_factory=list)
+
+    @property
+    def _llm_type(self):
+        return "todo-compaction-test"
+
+    def bind_tools(self, tools, **kwargs):
+        return self
+
+    def _generate(self, messages, stop=None, run_manager=None, **kwargs):
+        self._seen.append(list(messages))
+        return ChatResult(generations=[ChatGeneration(message=AIMessage(content="The Gateway documentation is complete."))])
+
+    async def _agenerate(self, messages, stop=None, run_manager=None, **kwargs):
+        return self._generate(messages, stop=stop, run_manager=run_manager, **kwargs)
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("asynchronous", [False, True])
+@pytest.mark.parametrize("context", ["rebuild", "empty-todos", "visible-write-todos"])
+async def test_next_model_gets_current_todos_after_compaction(asynchronous, context):
+    model = _CapturingModel()
+    summary = _middleware()
+    graph = create_agent(model=model, middleware=[summary, TodoMiddleware()], state_schema=ThreadState)
+    todos = [] if context == "empty-todos" else _todos()
+    state = {"messages": _messages(visible_write_todos=context == "visible-write-todos"), "todos": todos}
+    runtime_context = {"thread_id": "docs", "run_id": "docs-run"}
+
+    result = await graph.ainvoke(state, context=runtime_context) if asynchronous else graph.invoke(state, context=runtime_context)
+
+    assert result["todos"] == todos
+    assert result["summary_text"] == "Gateway route documentation is ready."
+    assert len(model._seen) == 1
+    request = model._seen[0]
+    assert all("STALE TODO" not in str(m.content) for m in request)
+    reminders = [m for m in request if isinstance(m, HumanMessage) and m.name == TODO_REMINDER_MESSAGE_NAME]
+    persisted = [m for m in result["messages"] if isinstance(m, HumanMessage) and m.name == TODO_REMINDER_MESSAGE_NAME]
+    assert len(reminders) == len(persisted) == (1 if context == "rebuild" else 0)
+    if reminders:
+        assert "[completed] Inspect the Gateway routes" in reminders[0].content
+        assert "[completed] Document the authentication endpoints" in reminders[0].content
+        assert reminders[0].additional_kwargs["hide_from_ui"] is True
+    assert {"recent-ai", "recent-tool"}.issubset({m.id for m in request})

backend/tests/test_todo_middleware.py

@@ -14,6 +14,7 @@ from deerflow.agents.middlewares.model_length_finish_reason_middleware import (
     ModelLengthFinishReasonMiddleware,
 )
 from deerflow.agents.middlewares.todo_middleware import (
+    TODO_REMINDER_MESSAGE_NAME,
     TodoMiddleware,
     _format_todos,
     _has_tool_call_intent_or_error,
@@ -28,7 +29,7 @@ def _ai_with_write_todos():
 
 
 def _reminder_msg():
-    return HumanMessage(name="todo_reminder", content="reminder")
+    return HumanMessage(name=TODO_REMINDER_MESSAGE_NAME, content="reminder")
 
 
 class _CapturingFakeMessagesListChatModel(FakeMessagesListChatModel):
@@ -173,7 +174,7 @@ class TestBeforeModel:
         msgs = result["messages"]
         assert len(msgs) == 1
         assert isinstance(msgs[0], HumanMessage)
-        assert msgs[0].name == "todo_reminder"
+        assert msgs[0].name == TODO_REMINDER_MESSAGE_NAME
 
     def test_reminder_contains_formatted_todos(self):
         mw = TodoMiddleware()
@@ -198,7 +199,7 @@ class TestAbeforeModel:
         }
         result = asyncio.run(mw.abefore_model(state, _make_runtime()))
         assert result is not None
-        assert result["messages"][0].name == "todo_reminder"
+        assert result["messages"][0].name == TODO_REMINDER_MESSAGE_NAME
 
 
 def _todo_completion_reminders(messages):