LangChain
Published reports for this stack, with severity, affected version, and a direct path into each diagnostic.
On Windows, FilesystemFileSearchMiddleware returns virtual paths containing native backslashes (e.g., '/src\\nested\\example.py') instead of the contract's forward-slash format, causing inconsistent and non-portable tool outputs.
On Windows, path validation in FilesystemClaudeTextEditorMiddleware and FilesystemClaudeMemoryMiddleware rejects valid nested virtual paths due to backslash vs forward slash mismatch in allowed_prefixes comparison.
Three security regression tests in langchain-core's prompt loading tests unconditionally create symlinks, causing OSError [WinError 1314] on default Windows environments. This makes the core test suite fail for contributors without elevated privileges or Developer Mode, and the suite has no Windows CI to catch it.
The max_length parameter is not applied when using the lightweight 'sentencizer' pipeline, causing spaCy's default limit of 1,000,000 characters to be used instead of the user-specified limit.
ChatDeepSeek only overrides sync `_generate` and `_stream` to translate `JSONDecodeError` from the OpenAI SDK into a DeepSeek-specific message. The inherited async `_agenerate` and `_astream` are not wrapped, so async callers still receive the opaque `json.decoder.JSONDecodeError: Expecting value: line 1 column 1 (char 0)`.
The parse_json_markdown function in langchain_core.utils.json uses a case-sensitive regex to identify fenced code blocks, causing parsing failures when models output fences with uppercase 'JSON' or 'Json'. This affects users of JsonOutputParser and other utilities relying on this function.
MarkdownHeaderTextSplitter incorrectly includes CommonMark optional closing hash sequences in header metadata values, e.g. '# Title ##' yields metadata 'Title ##' instead of 'Title'. This breaks filtering and grouping by header.
When a piece cannot be split further because no separators remain, it is appended directly to the output without applying strip_whitespace, causing inconsistent leading/trailing whitespace compared to merged chunks.
26 unit tests across test_shell_tool.py and test_shell_execution_policies.py fail on Windows because ShellToolMiddleware hardcodes /bin/bash, uses POSIX-only signals and the resource module, and the tests lack platform skip markers or proper patching of the _HAS_RESOURCE flag.
The standard chat model integration tests in langchain-tests 1.1.9 verify that input_tokens, output_tokens, and total_tokens are integers but do not assert that total_tokens equals the sum of input_tokens and output_tokens. A model can report inconsistent usage metadata (e.g., input_tokens=10, output_tokens=105, total_tokens=15) and still pass both test_usage_metadata and test_usage_metadata_streaming.
Blob.as_string() uses the default 'utf-8' encoding, which does not strip a UTF-8 byte order mark (BOM). Files written by Windows tools (Notepad, PowerShell, etc.) often include a BOM, and this BOM is returned as part of the string, contaminating downstream processing such as Markdown heading parsing.
The `_convert_content_block` function in `langchain.mcp` silently discards metadata fields (name, title, description, size, and URI for embedded resources) when converting MCP `ResourceLink` and `EmbeddedResource` objects to LangChain content blocks, leading to loss of information.
ChatGroq in langchain-groq allows n=2 during construction for non-streaming requests, sending an invalid parameter to the Groq API which returns HTTP 400.
ModelRouterMiddleware._latest_human_message uses next() without a default, causing StopIteration to be converted into an opaque RuntimeError on both sync and async paths when no HumanMessage exists in the state. The error does not identify the middleware or the remedy.
langchain-typesafe's ModelRouterMiddleware adds model_route to the agent input JSON schema even though the middleware unconditionally overwrites any caller-provided value in before_agent. This advertises an input field that is silently ignored, causing schema-driven clients to pass a route choice that has no effect.
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.