v0.19.0 — LangChain / deepagents adapter

Added

• LangChain / deepagents integration: new optional [langchain] extra exposes create_langchain_tools() returning a list[BaseTool] consumable by deepagents.create_deep_agent(tools=...) or any langchain.agents.create_agent(tools=...). Each tool wraps the existing ToolDef via StructuredTool.from_function(args_schema=<JSON Schema dict>, response_format="content_and_artifact") — the JSON Schema reaches the agent verbatim and the original MCP envelope is preserved on ToolMessage.artifact. BLOCKED — envelopes from the underlying tools (tools/factory.py, tools/middleware.py) are surfaced as ToolException, which the agent loop renders as ToolMessage(status="error").
• ContractMiddleware(AgentMiddleware) for graph-level enforcement. Pair with create_langchain_tools(..., apply_middleware=False) to avoid duplicate work when the middleware is installed at the deepagents/LangGraph boundary. Runs Validator.validate(sql) + ContractSession.check_limits() and short-circuits with ToolMessage(status="error").
• Top-level lazy re-exports of create_langchain_tools and ContractMiddleware from agentic_data_contracts. The base install (without the [langchain] extra) keeps both names defined as None, so from agentic_data_contracts import … continues to work unchanged for existing Claude Agent SDK users.

Notes

• Auto-applied in-tool enforcement intentionally does not validate SQL on every call (unlike contract_middleware); doing so would block inspect_query's purpose of reporting violations as JSON. SQL validation is left to run_query's self-checks at tools/factory.py:632-672, which still surface as ToolException via the canonical BLOCKED — prefix sniff.
• Backward-compatible: zero changes to tools/factory.py, tools/sdk.py, tools/middleware.py, or any existing top-level import. Pinned to current stable: langchain-core>=1.3.3, langchain>=1.2.17.