Skip to content

The agent engineering platform.

fix(langchain): preserve content-block shape in PIIMiddleware redaction (#39894)

Sydney Runkle4a66e355da07
diff --git a/libs/langchain_v1/langchain/agents/middleware/pii.py b/libs/langchain_v1/langchain/agents/middleware/pii.py
index c3b99eee8a..e4c7612731 100644
--- a/libs/langchain_v1/langchain/agents/middleware/pii.py
+++ b/libs/langchain_v1/langchain/agents/middleware/pii.py
@@ -5,7 +5,7 @@ from __future__ import annotations
from functools import partial
from typing import TYPE_CHECKING, Any, ClassVar, Literal
-from langchain_core.messages import AIMessage, AnyMessage, BaseMessage, HumanMessage, ToolMessage
+from langchain_core.messages import AIMessage, BaseMessage, HumanMessage, ToolMessage
from langgraph.stream import StreamTransformer
from typing_extensions import override
@@ -660,13 +660,39 @@ class PIIMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT, ResponseT])
"""Name of the middleware."""
return f"{self.__class__.__name__}[{self.pii_type}]"
- def _process_content(self, content: str) -> tuple[str, list[PIIMatch]]:
- """Apply the configured redaction rule to the provided content."""
- matches = self.detector(content)
- if not matches:
- return content, []
- sanitized = apply_strategy(content, matches, self.strategy)
- return sanitized, matches
+ def _process_content(
+ self, content: str | list[str | dict[str, Any]]
+ ) -> tuple[str | list[str | dict[str, Any]], bool]:
+ """Apply the configured redaction rule, preserving the shape of `content`.
+
+ Only string leaves are redacted: a plain-string `content` directly, and
+ the `text` of each block in a block-list `content`. Non-text blocks pass
+ through untouched, so block-list content keeps its structure instead of
+ collapsing into the `repr` of the list.
+
+ Args:
+ content: The message content to redact.
+
+ Returns:
+ The redacted content and whether anything was rewritten.
+ """
+ if isinstance(content, str):
+ matches = self.detector(content)
+ if not matches:
+ return content, False
+ return apply_strategy(content, matches, self.strategy), True
+
+ new_blocks: list[str | dict[str, Any]] = []
+ changed = False
+ for block in content:
+ new_block: str | dict[str, Any] = block
+ text = block if isinstance(block, str) else block.get("text")
+ if isinstance(text, str) and (matches := self.detector(text)):
+ redacted = apply_strategy(text, matches, self.strategy)
+ new_block = redacted if isinstance(block, str) else {**block, "text": redacted}
+ changed = True
+ new_blocks.append(new_block)
+ return (new_blocks, True) if changed else (content, False)
@hook_config(can_jump_to=["end"])
@override
@@ -710,18 +736,12 @@ class PIIMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT, ResponseT])
break
if last_user_idx is not None and last_user_msg and last_user_msg.content:
- # Detect PII in message content
- content = str(last_user_msg.content)
- new_content, matches = self._process_content(content)
+ new_content, changed = self._process_content(last_user_msg.content)
- if matches:
- updated_message: AnyMessage = HumanMessage(
- content=new_content,
- id=last_user_msg.id,
- name=last_user_msg.name,
+ if changed:
+ new_messages[last_user_idx] = last_user_msg.model_copy(
+ update={"content": new_content}
)
-
- new_messages[last_user_idx] = updated_message
any_modified = True
# Check tool results if enabled
@@ -742,21 +762,12 @@ class PIIMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT, ResponseT])
if not tool_msg.content:
continue
- content = str(tool_msg.content)
- new_content, matches = self._process_content(content)
+ new_content, changed = self._process_content(tool_msg.content)
- if not matches:
+ if not changed:
continue
- # Create updated tool message
- updated_message = ToolMessage(
- content=new_content,
- id=tool_msg.id,
- name=tool_msg.name,
- tool_call_id=tool_msg.tool_call_id,
- )
-
- new_messages[i] = updated_message
+ new_messages[i] = tool_msg.model_copy(update={"content": new_content})
any_modified = True
if any_modified:
@@ -824,24 +835,13 @@ class PIIMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT, ResponseT])
if last_ai_idx is None or not last_ai_msg or not last_ai_msg.content:
return None
- # Detect PII in message content
- content = str(last_ai_msg.content)
- new_content, matches = self._process_content(content)
+ new_content, changed = self._process_content(last_ai_msg.content)
- if not matches:
+ if not changed:
return None
- # Create updated message
- updated_message = AIMessage(
- content=new_content,
- id=last_ai_msg.id,
- name=last_ai_msg.name,
- tool_calls=last_ai_msg.tool_calls,
- )
-
- # Return updated messages
new_messages = list(messages)
- new_messages[last_ai_idx] = updated_message
+ new_messages[last_ai_idx] = last_ai_msg.model_copy(update={"content": new_content})
return {"messages": new_messages}
diff --git a/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_pii.py b/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_pii.py
index 4945e49a5f..a86ed93164 100644
--- a/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_pii.py
+++ b/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_pii.py
@@ -551,6 +551,89 @@ class TestPIIMiddlewareIntegration:
messages = result["messages"]
assert any("[REDACTED_EMAIL]" in str(msg.content) for msg in messages)
+ def test_input_list_content_preserved(self) -> None:
+ """List-of-content-blocks input is redacted in place, not stringified.
+
+ Regression: `str(msg.content)` flattened block-list content into its
+ `repr`, so the stored content became the literal string
+ `"[{'type': 'text', 'text': '...'}]"`.
+ """
+ middleware = PIIMiddleware("email", strategy="redact")
+ state = AgentState[Any](
+ messages=[
+ HumanMessage(content=[{"type": "text", "text": "my email is [email protected]"}])
+ ]
+ )
+
+ result = middleware.before_model(state, Runtime())
+
+ assert result is not None
+ content = result["messages"][0].content
+ assert content == [{"type": "text", "text": "my email is [REDACTED_EMAIL]"}]
+ assert "[email protected]" not in str(content)
+
+ def test_output_list_content_preserved(self) -> None:
+ """List-of-content-blocks AI output is redacted in place, not stringified."""
+ middleware = PIIMiddleware(
+ "email", strategy="redact", apply_to_input=False, apply_to_output=True
+ )
+ state = AgentState[Any](
+ messages=[AIMessage(content=[{"type": "text", "text": "reach me at [email protected]"}])]
+ )
+
+ result = middleware.after_model(state, Runtime())
+
+ assert result is not None
+ content = result["messages"][0].content
+ assert content == [{"type": "text", "text": "reach me at [REDACTED_EMAIL]"}]
+ assert "[email protected]" not in str(content)
+
+ def test_redaction_preserves_message_fields(self) -> None:
+ """Redacting keeps fields the hooks never enumerated when rebuilding messages."""
+ middleware = PIIMiddleware("email", strategy="redact")
+ original = HumanMessage(
+ content="my email is [email protected]",
+ id="msg_1",
+ name="alice",
+ additional_kwargs={"source": "web"},
+ )
+ state = AgentState[Any](messages=[original])
+
+ result = middleware.before_model(state, Runtime())
+
+ assert result is not None
+ redacted = result["messages"][0]
+ assert redacted.content == "my email is [REDACTED_EMAIL]"
+ assert redacted.id == "msg_1"
+ assert redacted.name == "alice"
+ assert redacted.additional_kwargs == {"source": "web"}
+
+ def test_tool_result_list_content_preserved(self) -> None:
+ """List-of-content-blocks tool results are redacted in place."""
+ middleware = PIIMiddleware(
+ "email", strategy="redact", apply_to_input=False, apply_to_tool_results=True
+ )
+ state = AgentState[Any](
+ messages=[
+ HumanMessage("Search for user"),
+ AIMessage(
+ content="",
+ tool_calls=[ToolCall(name="search", args={}, id="call_1", type="tool_call")],
+ ),
+ ToolMessage(
+ content=[{"type": "text", "text": "found: [email protected]"}],
+ tool_call_id="call_1",
+ ),
+ ]
+ )
+
+ result = middleware.before_model(state, Runtime())
+
+ assert result is not None
+ content = result["messages"][2].content
+ assert content == [{"type": "text", "text": "found: [REDACTED_EMAIL]"}]
+ assert "[email protected]" not in str(content)
+
class TestCustomDetector:
"""Test custom detector functionality."""