mirror of
https://github.com/bytedance/deer-flow.git
synced 2026-05-21 15:36:48 +00:00
feat: static system prompt with DynamicContextMiddleware for prefix-cache optimization (#2801)
* feat(middleware): inject dynamic context via DynamicContextMiddleware
Move memory and current date out of the system prompt and into a
dedicated <system-reminder> HumanMessage injected once per session
(frozen-snapshot pattern) via a new DynamicContextMiddleware.
This keeps the system prompt byte-exact across all users and sessions,
enabling maximum Anthropic/Bedrock prefix-cache reuse.
Key design decisions:
- ID-swap technique: reminder takes the first HumanMessage's ID
(replacing it in-place via add_messages), original content gets a
derived `{id}__user` ID (appended after). Preserves correct ordering.
- hide_from_ui: True on reminder messages so frontend filters them out.
- Midnight crossing: date-update reminder injected before the current
turn's HumanMessage when the conversation spans midnight.
- INFO-level logging for production diagnostics.
Also adds prompt-caching breakpoint budget enforcement tests and
updates ClaudeChatModel docs to reference the new pattern.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(token-usage): log input/output token detail breakdown in middleware
Extend the LLM token usage log line to include input_token_details and
output_token_details (cache_creation, cache_read, reasoning, audio, etc.)
when present. Adds tests covering Anthropic cache detail logging from
both usage_metadata and response_metadata.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix: fix nginx
* fix(middleware): always inject date; gate memory on injection_enabled
Date injection is now unconditional — it is part of the static system
prompt replacement and should always be present. Memory injection
remains gated by `memory.injection_enabled` in the app config.
Previously the entire DynamicContextMiddleware was skipped when
injection_enabled was False, which also suppressed the date.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(lint): format files and correct test assertions for token usage middleware
- ruff format dynamic_context_middleware.py and test_claude_provider_prompt_caching.py
- Remove unused pytest import from test_dynamic_context_middleware.py
- Fix two tests that asserted response_metadata fallback logic that
doesn't exist: replace with tests that match actual middleware behavior
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(middleware): address Copilot review comments on DynamicContextMiddleware
- Use additional_kwargs flag for reminder detection instead of content
substring matching, so user messages containing '<system-reminder>'
are not mistakenly treated as injected reminders
- Generate stable UUID when original HumanMessage.id is None to prevent
ambiguous 'None__user' derived IDs and message collisions
- Downgrade per-turn no-op log to DEBUG; keep actual injection events at INFO
- Add two new tests: missing-id UUID fallback and user-text false-positive
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
@@ -0,0 +1,312 @@
|
||||
"""Tests for DynamicContextMiddleware.
|
||||
|
||||
Verifies that memory and current date are injected as a <system-reminder> into
|
||||
the first HumanMessage exactly once per session (frozen-snapshot pattern).
|
||||
"""
|
||||
|
||||
from types import SimpleNamespace
|
||||
from unittest import mock
|
||||
|
||||
from langchain_core.messages import AIMessage, HumanMessage
|
||||
|
||||
from deerflow.agents.middlewares.dynamic_context_middleware import (
|
||||
_DYNAMIC_CONTEXT_REMINDER_KEY,
|
||||
DynamicContextMiddleware,
|
||||
)
|
||||
|
||||
_SYSTEM_REMINDER_TAG = "<system-reminder>"
|
||||
|
||||
|
||||
def _make_middleware(**kwargs) -> DynamicContextMiddleware:
|
||||
return DynamicContextMiddleware(**kwargs)
|
||||
|
||||
|
||||
def _fake_runtime():
|
||||
return SimpleNamespace(context={})
|
||||
|
||||
|
||||
def _reminder_msg(content: str, msg_id: str) -> HumanMessage:
|
||||
"""Build a reminder HumanMessage the way the middleware would produce it."""
|
||||
return HumanMessage(
|
||||
content=content,
|
||||
id=msg_id,
|
||||
additional_kwargs={"hide_from_ui": True, _DYNAMIC_CONTEXT_REMINDER_KEY: True},
|
||||
)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Basic injection
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
def test_injects_system_reminder_into_first_human_message():
|
||||
mw = _make_middleware()
|
||||
state = {"messages": [HumanMessage(content="Hello", id="msg-1")]}
|
||||
|
||||
with mock.patch("deerflow.agents.lead_agent.prompt._get_memory_context", return_value=""), mock.patch("deerflow.agents.middlewares.dynamic_context_middleware.datetime") as mock_dt:
|
||||
mock_dt.now.return_value.strftime.return_value = "2026-05-08, Friday"
|
||||
result = mw.before_agent(state, _fake_runtime())
|
||||
|
||||
assert result is not None
|
||||
updated_msgs = result["messages"]
|
||||
assert len(updated_msgs) == 2
|
||||
|
||||
reminder_msg = updated_msgs[0]
|
||||
assert isinstance(reminder_msg, HumanMessage)
|
||||
assert reminder_msg.id == "msg-1" # takes the original ID (position swap)
|
||||
assert reminder_msg.additional_kwargs.get(_DYNAMIC_CONTEXT_REMINDER_KEY) is True
|
||||
assert _SYSTEM_REMINDER_TAG in reminder_msg.content
|
||||
assert "<current_date>2026-05-08, Friday</current_date>" in reminder_msg.content
|
||||
assert "Hello" not in reminder_msg.content # reminder only — no user text
|
||||
|
||||
user_msg = updated_msgs[1]
|
||||
assert isinstance(user_msg, HumanMessage)
|
||||
assert user_msg.id == "msg-1__user" # derived ID
|
||||
assert user_msg.content == "Hello"
|
||||
|
||||
|
||||
def test_memory_included_when_present():
|
||||
mw = _make_middleware()
|
||||
state = {"messages": [HumanMessage(content="Hi", id="msg-1")]}
|
||||
|
||||
with (
|
||||
mock.patch(
|
||||
"deerflow.agents.lead_agent.prompt._get_memory_context",
|
||||
return_value="<memory>\nUser prefers Python.\n</memory>",
|
||||
),
|
||||
mock.patch("deerflow.agents.middlewares.dynamic_context_middleware.datetime") as mock_dt,
|
||||
):
|
||||
mock_dt.now.return_value.strftime.return_value = "2026-05-08, Friday"
|
||||
result = mw.before_agent(state, _fake_runtime())
|
||||
|
||||
# Reminder is the first returned message; user query is the second
|
||||
reminder_content = result["messages"][0].content
|
||||
assert "User prefers Python." in reminder_content
|
||||
assert "<current_date>2026-05-08, Friday</current_date>" in reminder_content
|
||||
assert result["messages"][1].content == "Hi"
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Frozen-snapshot: no re-injection within a session
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
def test_skips_injection_if_already_present():
|
||||
"""Second turn: separate reminder message already present → no update."""
|
||||
mw = _make_middleware()
|
||||
reminder_content = "<system-reminder>\n<current_date>2026-05-08, Friday</current_date>\n</system-reminder>"
|
||||
state = {
|
||||
"messages": [
|
||||
_reminder_msg(reminder_content, "msg-1"),
|
||||
HumanMessage(content="Hello", id="msg-1__user"),
|
||||
AIMessage(content="Hi there"),
|
||||
HumanMessage(content="Follow-up", id="msg-2"),
|
||||
]
|
||||
}
|
||||
|
||||
with mock.patch("deerflow.agents.middlewares.dynamic_context_middleware.datetime") as mock_dt:
|
||||
mock_dt.now.return_value.strftime.return_value = "2026-05-08, Friday"
|
||||
result = mw.before_agent(state, _fake_runtime())
|
||||
|
||||
assert result is None # no update needed
|
||||
|
||||
|
||||
def test_injects_only_into_first_human_message_not_later_ones():
|
||||
"""Reminder targets the first HumanMessage; subsequent messages are not touched."""
|
||||
mw = _make_middleware()
|
||||
state = {
|
||||
"messages": [
|
||||
HumanMessage(content="First", id="msg-1"),
|
||||
AIMessage(content="Reply"),
|
||||
HumanMessage(content="Second", id="msg-2"),
|
||||
]
|
||||
}
|
||||
|
||||
with mock.patch("deerflow.agents.lead_agent.prompt._get_memory_context", return_value=""), mock.patch("deerflow.agents.middlewares.dynamic_context_middleware.datetime") as mock_dt:
|
||||
mock_dt.now.return_value.strftime.return_value = "2026-05-08, Friday"
|
||||
result = mw.before_agent(state, _fake_runtime())
|
||||
|
||||
assert result is not None
|
||||
msgs = result["messages"]
|
||||
# Only the two injected messages are returned (reminder + original first query)
|
||||
assert len(msgs) == 2
|
||||
assert msgs[0].id == "msg-1" # reminder takes first message's ID
|
||||
assert msgs[0].additional_kwargs.get(_DYNAMIC_CONTEXT_REMINDER_KEY) is True
|
||||
assert _SYSTEM_REMINDER_TAG in msgs[0].content
|
||||
assert msgs[1].id == "msg-1__user" # original content with derived ID
|
||||
assert msgs[1].content == "First"
|
||||
# "Second" (msg-2) is not in the returned update — it is left unchanged
|
||||
assert all(m.id != "msg-2" for m in msgs)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Edge cases
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
def test_no_messages_returns_none():
|
||||
mw = _make_middleware()
|
||||
result = mw.before_agent({"messages": []}, _fake_runtime())
|
||||
assert result is None
|
||||
|
||||
|
||||
def test_no_human_message_returns_none():
|
||||
mw = _make_middleware()
|
||||
state = {"messages": [AIMessage(content="assistant only")]}
|
||||
with mock.patch("deerflow.agents.lead_agent.prompt._get_memory_context", return_value=""):
|
||||
result = mw.before_agent(state, _fake_runtime())
|
||||
assert result is None
|
||||
|
||||
|
||||
def test_list_content_message_handled_as_separate_reminder():
|
||||
"""List-content (e.g. multi-modal) messages remain intact; reminder is a separate message."""
|
||||
mw = _make_middleware()
|
||||
original_content = [{"type": "text", "text": "Hello"}]
|
||||
state = {"messages": [HumanMessage(content=original_content, id="msg-1")]}
|
||||
|
||||
with mock.patch("deerflow.agents.lead_agent.prompt._get_memory_context", return_value=""), mock.patch("deerflow.agents.middlewares.dynamic_context_middleware.datetime") as mock_dt:
|
||||
mock_dt.now.return_value.strftime.return_value = "2026-05-08, Friday"
|
||||
result = mw.before_agent(state, _fake_runtime())
|
||||
|
||||
assert result is not None
|
||||
msgs = result["messages"]
|
||||
assert len(msgs) == 2
|
||||
# Reminder is a plain string message with the flag set
|
||||
assert isinstance(msgs[0].content, str)
|
||||
assert msgs[0].additional_kwargs.get(_DYNAMIC_CONTEXT_REMINDER_KEY) is True
|
||||
assert _SYSTEM_REMINDER_TAG in msgs[0].content
|
||||
# Original list-content message is untouched
|
||||
assert msgs[1].content == original_content
|
||||
|
||||
|
||||
def test_reminder_uses_original_id_user_message_uses_derived_id():
|
||||
"""Reminder takes original ID (position swap); user message gets {id}__user."""
|
||||
mw = _make_middleware()
|
||||
original_id = "original-id-abc"
|
||||
state = {"messages": [HumanMessage(content="Hello", id=original_id)]}
|
||||
|
||||
with mock.patch("deerflow.agents.lead_agent.prompt._get_memory_context", return_value=""), mock.patch("deerflow.agents.middlewares.dynamic_context_middleware.datetime") as mock_dt:
|
||||
mock_dt.now.return_value.strftime.return_value = "2026-05-08, Friday"
|
||||
result = mw.before_agent(state, _fake_runtime())
|
||||
|
||||
assert result["messages"][0].id == original_id
|
||||
assert result["messages"][1].id == f"{original_id}__user"
|
||||
|
||||
|
||||
def test_message_without_id_gets_stable_uuid():
|
||||
"""If the original HumanMessage has no ID, a UUID is generated and used consistently."""
|
||||
mw = _make_middleware()
|
||||
state = {"messages": [HumanMessage(content="Hello", id=None)]}
|
||||
|
||||
with mock.patch("deerflow.agents.lead_agent.prompt._get_memory_context", return_value=""), mock.patch("deerflow.agents.middlewares.dynamic_context_middleware.datetime") as mock_dt:
|
||||
mock_dt.now.return_value.strftime.return_value = "2026-05-08, Friday"
|
||||
result = mw.before_agent(state, _fake_runtime())
|
||||
|
||||
assert result is not None
|
||||
reminder_id = result["messages"][0].id
|
||||
user_id = result["messages"][1].id
|
||||
assert reminder_id is not None
|
||||
assert reminder_id != "None"
|
||||
assert user_id == f"{reminder_id}__user"
|
||||
|
||||
|
||||
def test_user_message_containing_system_reminder_tag_does_not_prevent_injection():
|
||||
"""A user message containing '<system-reminder>' must not be mistaken for a reminder."""
|
||||
mw = _make_middleware()
|
||||
state = {
|
||||
"messages": [
|
||||
HumanMessage(content="What is <system-reminder>?", id="msg-1"),
|
||||
]
|
||||
}
|
||||
|
||||
with mock.patch("deerflow.agents.lead_agent.prompt._get_memory_context", return_value=""), mock.patch("deerflow.agents.middlewares.dynamic_context_middleware.datetime") as mock_dt:
|
||||
mock_dt.now.return_value.strftime.return_value = "2026-05-08, Friday"
|
||||
result = mw.before_agent(state, _fake_runtime())
|
||||
|
||||
# Injection must happen — the user message does NOT carry the reminder flag
|
||||
assert result is not None
|
||||
assert result["messages"][0].additional_kwargs.get(_DYNAMIC_CONTEXT_REMINDER_KEY) is True
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Midnight crossing
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
def test_midnight_crossing_injects_date_update_as_separate_message():
|
||||
"""When the date has changed, a separate date-update reminder is injected before
|
||||
the current turn's HumanMessage using the ID-swap technique."""
|
||||
mw = _make_middleware()
|
||||
reminder_content = "<system-reminder>\n<current_date>2026-05-08, Friday</current_date>\n</system-reminder>"
|
||||
state = {
|
||||
"messages": [
|
||||
_reminder_msg(reminder_content, "msg-1"),
|
||||
HumanMessage(content="Hello", id="msg-1__user"),
|
||||
AIMessage(content="Response"),
|
||||
HumanMessage(content="Good morning", id="msg-2"),
|
||||
]
|
||||
}
|
||||
|
||||
with mock.patch("deerflow.agents.middlewares.dynamic_context_middleware.datetime") as mock_dt:
|
||||
mock_dt.now.return_value.strftime.return_value = "2026-05-09, Saturday"
|
||||
result = mw.before_agent(state, _fake_runtime())
|
||||
|
||||
assert result is not None
|
||||
msgs = result["messages"]
|
||||
assert len(msgs) == 2
|
||||
|
||||
# Date-update reminder takes the current message's ID
|
||||
assert msgs[0].id == "msg-2"
|
||||
assert msgs[0].additional_kwargs.get(_DYNAMIC_CONTEXT_REMINDER_KEY) is True
|
||||
assert _SYSTEM_REMINDER_TAG in msgs[0].content
|
||||
assert "<current_date>2026-05-09, Saturday</current_date>" in msgs[0].content
|
||||
assert "Good morning" not in msgs[0].content # reminder only
|
||||
|
||||
# Original user text appended with derived ID
|
||||
assert msgs[1].id == "msg-2__user"
|
||||
assert msgs[1].content == "Good morning"
|
||||
|
||||
|
||||
def test_midnight_crossing_id_swap():
|
||||
"""Date-update reminder uses original ID; user message uses {id}__user."""
|
||||
mw = _make_middleware()
|
||||
reminder_content = "<system-reminder>\n<current_date>2026-05-08, Friday</current_date>\n</system-reminder>"
|
||||
state = {
|
||||
"messages": [
|
||||
_reminder_msg(reminder_content, "msg-1"),
|
||||
HumanMessage(content="Next day message", id="msg-2"),
|
||||
]
|
||||
}
|
||||
|
||||
with mock.patch("deerflow.agents.middlewares.dynamic_context_middleware.datetime") as mock_dt:
|
||||
mock_dt.now.return_value.strftime.return_value = "2026-05-09, Saturday"
|
||||
result = mw.before_agent(state, _fake_runtime())
|
||||
|
||||
assert result["messages"][0].id == "msg-2"
|
||||
assert result["messages"][1].id == "msg-2__user"
|
||||
|
||||
|
||||
def test_no_second_midnight_injection_once_date_updated():
|
||||
"""After a midnight update is persisted, the same-day path skips re-injection."""
|
||||
mw = _make_middleware()
|
||||
date_update_content = "<system-reminder>\n<current_date>2026-05-09, Saturday</current_date>\n</system-reminder>"
|
||||
state = {
|
||||
"messages": [
|
||||
_reminder_msg(
|
||||
"<system-reminder>\n<current_date>2026-05-08, Friday</current_date>\n</system-reminder>",
|
||||
"msg-1",
|
||||
),
|
||||
HumanMessage(content="Hello", id="msg-1__user"),
|
||||
AIMessage(content="Response"),
|
||||
_reminder_msg(date_update_content, "msg-2"),
|
||||
HumanMessage(content="Good morning", id="msg-2__user"),
|
||||
AIMessage(content="Good morning!"),
|
||||
HumanMessage(content="Third turn", id="msg-3"),
|
||||
]
|
||||
}
|
||||
|
||||
with mock.patch("deerflow.agents.middlewares.dynamic_context_middleware.datetime") as mock_dt:
|
||||
mock_dt.now.return_value.strftime.return_value = "2026-05-09, Saturday"
|
||||
result = mw.before_agent(state, _fake_runtime())
|
||||
|
||||
assert result is None # same day as last injected date → no update
|
||||
Reference in New Issue
Block a user