LangChain
Published reports for this stack, with severity, affected version, and a direct path into each diagnostic.
In langchain-xai, ChatXAI.validate_environment creates two separate openai.OpenAI clients (and two openai.AsyncOpenAI) for each instance, rather than deriving client from root_client like sibling integrations. This causes duplicated resources, inconsistent close behavior, and potential connection pool waste. The fix is to build root clients first and set client/async_client as views.
The convert_to_openai_messages utility in langchain-core omits AIMessage.invalid_tool_calls from the converted OpenAI message, while the corresponding ToolMessage replies are retained, leading to orphaned tool_call_id references in exported conversations.
The UsageMetadataCallbackHandler in langchain_core only records token usage when the response's response_metadata contains a 'model_name' key. If model_name is absent, usage_metadata is silently discarded, defeating the purpose of the callback and causing silent data loss in usage tracking for custom or third-party chat models.
In langchain-anthropic's _format_messages, when an AIMessage contains both a tool_use content block and a matching tool_call in the structured tool_calls field, the tool_use block is rebuilt from tool_calls and any extra keys on the original block, notably cache_control, are discarded. This causes prompt-cache breakpoints placed on tool_use blocks to be silently lost.
The constructor of HTMLHeaderTextSplitter sorts headers_to_split_on using a bare int(tag[1:]) conversion, which fails for any tag outside h1-h6 (e.g., div -> 'iv'). The _generate_documents method already contains a fallback level of 9999 for such tags, causing an internal inconsistency that crashes at initialization.
In langchain-openai 1.6.2, Responses API usage metadata for 'priority' and 'flex' model providers double-counts cache_write_tokens: the non-cached input bucket only subtracts cached_tokens, not cache_write_tokens, causing inflated cost estimates.
Report of numerous test failures in langchain-core 1.6.5 on FreeBSD 15.1 (Python 3.12), likely due to platform-specific environmental issues such as event loop policy or locale settings.
ChatAnthropic.bind_tools() warns and removes forced tool_choice when thinking is set on the constructor, but the same thinking supplied via invoke() is not checked. This sends an invalid combination (thinking enabled + tool_choice={'type':'any'}) to Anthropic API, resulting in HTTP 400.
When converting a finalized AIMessage into the v3 protocol event stream using message_to_events(), the generated message-finish event does not include the AIMessage.additional_kwargs field. This causes provider-specific data to be lost during cache replay, making replayed execution inconsistent with the original live streaming path.
When using ChatOpenAI with use_responses_api=True and output_version='v1', only the first chunk (response.created) carries the provider ID resp_...; subsequent content chunks get id=None and are filled with lc_run--<run_id>. The final merged AIMessage prefers the provider ID, causing duplicate messages in LangGraph stream_mode='messages' and frontends that reconcile by ID.
The public ChatOpenAI.stream() API raises AttributeError when an OpenAI streaming chunk contains a tool-call delta whose `function` field is null. The parser assumes `rtc["function"]` is a dict and calls `.get` on None, causing a crash. Even though the repro uses a mocked response, the public API should handle malformed or unexpected provider payloads gracefully.
The langchain.mcp adapter drops the optional `_meta` field from MCP `CallToolResult` when converting to a LangChain `ToolMessage`, causing loss of out-of-band response metadata. The hotfix stores this metadata under a dedicated `mcp_meta` key in `response_metadata`, without conflating it with the static tool definition `_meta`.
XMLOutputParser.parse() raises OutputParserException for valid XML where a non-declaration element has an encoding attribute and the opening tag is followed by a newline. The regex used to remove the XML declaration is too broad and matches such element tags, causing the parser to remove the opening tag and produce malformed XML.
Streaming large tool call arguments through (prompt | model).stream() or with_structured_output(...).stream() triggers repeated parse_partial_json calls on every AIMessageChunk addition, causing quadratic CPU time as arguments grow.
ContextThreadPoolExecutor.map() calls len(iterables[0]) to preallocate contexts, which raises TypeError for generators and other unsized iterables, even though the standard ThreadPoolExecutor.map() accepts them. The proposed fix delegates directly to super().map(), relying on the existing submit() override for context propagation.
When `HumanInTheLoopMiddleware` is configured with an `InterruptOnConfig` that includes an explicit `args_schema`, the generated `ReviewConfig` in the interrupt payload omits that schema. This is a silent configuration failure: the schema is accepted but not propagated, preventing schema-driven client-side handling.
The splitter terminates fenced code blocks whenever it sees any line of three or more backticks or tildes, ignoring CommonMark rules that require a closing fence to match the opening character and have at least the opening length. This creates extra chunks and wrong `Code` metadata when the content includes an inner shorter fence or a mismatched fence.
When `ensure_ascii=False` is passed to `split_text`, `_json_size` still measures using `json.dumps` with default `ensure_ascii=True`, causing non-ASCII chunks to be far below `max_chunk_size` and producing more chunks than expected.
When an input raises an exception not in exceptions_to_handle, batch/abatch re-raise immediately, skipping the cleanup loop that closes root callback runs, causing them to remain pending forever in tracers like LangSmith.
In langchain-core's test_convert_to_openai_data_block, the Responses API file/url case sets up test data but never calls the conversion function or asserts the result, leaving the branch untested. The fix adds the missing call and assertion.
When ModelFallbackMiddleware falls back to ChatAnthropicMantle (langchain-aws), cache_control markers are silently removed from both model_settings and message content blocks, disabling prompt caching and increasing costs. The root cause is a missing _llm_type in _ANTHROPIC_LLM_TYPES.
Current batched HITL interrupts force approval UIs to infer tool action context from positional lists and skip response validation, leading to fragile integrations and potential stuck threads. This change adds per-call interrupts with typed tool_approval values and per-tool response_schema validation.
PromptTemplate.from_template("Value: {}") is accepted silently, reports an empty input variable, and then raises IndexError on every .format() call. It should be rejected at creation time with a clear ValueError, consistent with the all-digit `{0}` case.
Two unit tests in the langchain package (tests/unit_tests/test_version.py and tests/unit_tests/test_dependencies.py) open pyproject.toml using Path.open() without specifying encoding. On Windows systems with a non-UTF-8 locale (e.g., cp936), Python defaults to the locale encoding, causing UnicodeDecodeError when encountering UTF-8 bytes. The fix is to pass encoding="utf-8" to the open calls.
Multiple unit test call sites read pyproject.toml without specifying encoding='utf-8'. On Windows with a non-UTF-8 locale (e.g., cp936), Python defaults to the ANSI code page, leading to UnicodeDecodeError because the file is valid UTF-8 and contains non-ASCII characters. Issue #40588 fixed two call sites, but two remain: one active in langchain_v1 summarization tests and one latent in langchain-classic dependencies tests.
The ImportError messages raised when importing moved REPL tools/agents in langchain-classic contain outdated and dead links to SECURITY.md and a GitHub discussion, and have missing spaces causing link concatenation. This leads users to 404 pages or broken links, reducing safety guidance.
Docstring examples in langchain-core reference modules 'langchain_core.output_parser', 'langchain_core.runnable', and 'langchain_core.documents' which do not exist, causing ModuleNotFoundError or ImportError when copied. The correct modules are 'langchain_core.output_parsers', 'langchain_core.runnables', and 'langchain_core.documents.base'.
When importing a moved agent (e.g., create_csv_agent) from langchain_classic.agents.agent_toolkits, the __getattr__ helper incorrectly calls as_import_path without specifying relative_to, causing a ValueError from langchain_core instead of the intended ImportError with migration guidance.
ChatOpenAI with extra_body={"enable_thinking": true} and bind_tools(..., tool_choice="get_weather") sends a named tool-choice object to Alibaba DashScope, which rejects it with InternalError.Algo.InvalidParameter. LangChain should validate this incompatible combination before sending the request.
The @wrap_tool_call decorator does not accept or propagate a context_schema parameter, so the returned AgentMiddleware always types ContextT as None. When used with create_agent(context_schema=AgentContext), the invariant ContextT generic causes a type error, even though the runtime behavior may be correct.