feat(openai): support automatic server-side compaction (#35212)

libs/langchain/tests/unit_tests/chat_models/test_base.py

@@ -147,6 +147,7 @@ def test_configurable() -> None:
             "reasoning_effort": None,
             "verbosity": None,
             "frequency_penalty": None,
+            "context_management": None,
             "include": None,
             "seed": None,
             "service_tier": None,

libs/langchain_v1/tests/unit_tests/chat_models/test_chat_models.py

@@ -192,6 +192,7 @@ def test_configurable() -> None:
             "reasoning_effort": None,
             "verbosity": None,
             "frequency_penalty": None,
+            "context_management": None,
             "include": None,
             "seed": None,
             "service_tier": None,

libs/partners/openai/langchain_openai/chat_models/base.py

@@ -799,6 +799,11 @@ class BaseChatOpenAI(BaseChatModel):
     passed in the parameter during invocation.
     """
 
+    context_management: list[dict[str, Any]] | None = None
+    """Configuration for
+    [context management](https://developers.openai.com/api/docs/guides/compaction).
+    """
+
     include: list[str] | None = None
     """Additional fields to include in generations from Responses API.
 
@@ -1096,6 +1101,7 @@ class BaseChatOpenAI(BaseChatModel):
             "reasoning_effort": self.reasoning_effort,
             "reasoning": self.reasoning,
             "verbosity": self.verbosity,
+            "context_management": self.context_management,
             "include": self.include,
             "service_tier": self.service_tier,
             "truncation": self.truncation,
@@ -1482,6 +1488,7 @@ class BaseChatOpenAI(BaseChatModel):
             return self.use_responses_api
         if (
             self.output_version == "responses/v1"
+            or self.context_management is not None
             or self.include is not None
             or self.reasoning is not None
             or self.truncation is not None
@@ -3894,6 +3901,7 @@ def _use_responses_api(payload: dict) -> bool:
         _is_builtin_tool(tool) for tool in payload["tools"]
     )
     responses_only_args = {
+        "context_management",
         "include",
         "previous_response_id",
         "reasoning",
@@ -4273,6 +4281,7 @@ def _construct_responses_api_input(messages: Sequence[BaseMessage]) -> list:
                                 )
                         elif block_type in (
                             "reasoning",
+                            "compaction",
                             "web_search_call",
                             "file_search_call",
                             "function_call",
@@ -4475,6 +4484,7 @@ def _construct_lc_result_from_responses_api(
             tool_calls.append(tool_call)
         elif output.type in (
             "reasoning",
+            "compaction",
             "web_search_call",
             "file_search_call",
             "computer_call",
@@ -4683,6 +4693,7 @@ def _convert_responses_chunk_to_generation_chunk(
             }
         )
     elif chunk.type == "response.output_item.done" and chunk.item.type in (
+        "compaction",
         "web_search_call",
         "file_search_call",
         "computer_call",

libs/partners/openai/pyproject.toml

@@ -24,7 +24,7 @@ version = "1.1.9"
 requires-python = ">=3.10.0,<4.0.0"
 dependencies = [
     "langchain-core>=1.2.11,<2.0.0",
-    "openai>=1.109.1,<3.0.0",
+    "openai>=2.20.0,<3.0.0",
     "tiktoken>=0.7.0,<1.0.0",
 ]
 

libs/partners/openai/tests/integration_tests/chat_models/test_responses_api.py

@@ -1107,3 +1107,96 @@ def test_custom_tool(output_version: Literal["responses/v1", "v1"]) -> None:
         full = chunk if full is None else full + chunk
     assert isinstance(full, AIMessageChunk)
     assert len(full.tool_calls) == 1
+
+
+@pytest.mark.default_cassette("test_compaction.yaml.gz")
+@pytest.mark.vcr
+@pytest.mark.parametrize("output_version", ["responses/v1", "v1"])
+def test_compaction(output_version: Literal["responses/v1", "v1"]) -> None:
+    """Test the compation beta feature."""
+    llm = ChatOpenAI(
+        model="gpt-5.2",
+        context_management=[{"type": "compaction", "compact_threshold": 10_000}],
+        output_version=output_version,
+    )
+
+    input_message = {
+        "role": "user",
+        "content": f"Generate a one-sentence summary of this:\n\n{'a' * 50000}",
+    }
+    messages: list = [input_message]
+
+    first_response = llm.invoke(messages)
+    messages.append(first_response)
+
+    second_message = {
+        "role": "user",
+        "content": f"Generate a one-sentence summary of this:\n\n{'b' * 50000}",
+    }
+    messages.append(second_message)
+
+    second_response = llm.invoke(messages)
+    messages.append(second_response)
+
+    content_blocks = second_response.content_blocks
+    compaction_block = next(
+        (block for block in content_blocks if block["type"] == "non_standard"),
+        None,
+    )
+    assert compaction_block
+    assert compaction_block["value"].get("type") == "compaction"
+
+    third_message = {
+        "role": "user",
+        "content": "What are we talking about?",
+    }
+    messages.append(third_message)
+    third_response = llm.invoke(messages)
+    assert third_response.text
+
+
+@pytest.mark.default_cassette("test_compaction_streaming.yaml.gz")
+@pytest.mark.vcr
+@pytest.mark.parametrize("output_version", ["responses/v1", "v1"])
+def test_compaction_streaming(output_version: Literal["responses/v1", "v1"]) -> None:
+    """Test the compation beta feature."""
+    llm = ChatOpenAI(
+        model="gpt-5.2",
+        context_management=[{"type": "compaction", "compact_threshold": 10_000}],
+        output_version=output_version,
+        streaming=True,
+    )
+
+    input_message = {
+        "role": "user",
+        "content": f"Generate a one-sentence summary of this:\n\n{'a' * 50000}",
+    }
+    messages: list = [input_message]
+
+    first_response = llm.invoke(messages)
+    messages.append(first_response)
+
+    second_message = {
+        "role": "user",
+        "content": f"Generate a one-sentence summary of this:\n\n{'b' * 50000}",
+    }
+    messages.append(second_message)
+
+    second_response = llm.invoke(messages)
+    messages.append(second_response)
+
+    content_blocks = second_response.content_blocks
+    compaction_block = next(
+        (block for block in content_blocks if block["type"] == "non_standard"),
+        None,
+    )
+    assert compaction_block
+    assert compaction_block["value"].get("type") == "compaction"
+
+    third_message = {
+        "role": "user",
+        "content": "What are we talking about?",
+    }
+    messages.append(third_message)
+    third_response = llm.invoke(messages)
+    assert third_response.text

libs/partners/openai/uv.lock

@@ -1,5 +1,5 @@
 version = 1
-revision = 3
+revision = 2
 requires-python = ">=3.10.0, <4.0.0"
 resolution-markers = [
     "python_full_version >= '3.13' and platform_python_implementation == 'PyPy'",
@@ -580,7 +580,7 @@ requires-dist = [
 provides-extras = ["community", "anthropic", "openai", "azure-ai", "google-vertexai", "google-genai", "fireworks", "ollama", "together", "mistralai", "huggingface", "groq", "aws", "deepseek", "xai", "perplexity"]
 
 [package.metadata.requires-dev]
-lint = [{ name = "ruff", specifier = ">=0.14.11,<0.16.0" }]
+lint = [{ name = "ruff", specifier = ">=0.15.0,<0.16.0" }]
 test = [
     { name = "blockbuster", specifier = ">=1.5.26,<1.6.0" },
     { name = "langchain-openai", editable = "." },
@@ -610,7 +610,7 @@ typing = [
 
 [[package]]
 name = "langchain-core"
-version = "1.2.11"
+version = "1.2.12"
 source = { editable = "../../core" }
 dependencies = [
     { name = "jsonpatch" },
@@ -641,7 +641,7 @@ dev = [
     { name = "jupyter", specifier = ">=1.0.0,<2.0.0" },
     { name = "setuptools", specifier = ">=67.6.1,<83.0.0" },
 ]
-lint = [{ name = "ruff", specifier = ">=0.14.11,<0.16.0" }]
+lint = [{ name = "ruff", specifier = ">=0.15.0,<0.16.0" }]
 test = [
     { name = "blockbuster", specifier = ">=1.5.18,<1.6.0" },
     { name = "freezegun", specifier = ">=1.2.2,<2.0.0" },
@@ -718,7 +718,7 @@ typing = [
 [package.metadata]
 requires-dist = [
     { name = "langchain-core", editable = "../../core" },
-    { name = "openai", specifier = ">=1.109.1,<3.0.0" },
+    { name = "openai", specifier = ">=2.20.0,<3.0.0" },
     { name = "tiktoken", specifier = ">=0.7.0,<1.0.0" },
 ]
 
@@ -791,7 +791,7 @@ requires-dist = [
 ]
 
 [package.metadata.requires-dev]
-lint = [{ name = "ruff", specifier = ">=0.14.11,<0.16.0" }]
+lint = [{ name = "ruff", specifier = ">=0.15.0,<0.16.0" }]
 test = [{ name = "langchain-core", editable = "../../core" }]
 test-integration = []
 typing = [