v0.4.0

Security

• Fixed a read-only guard bypass: the SQL extracted from the LLM's final response executed against the database with no guard at all; the destructive-statement regex also missed writable CTEs (WITH d AS (DELETE ...)) and DML nested in subqueries. Both are now closed, with regression tests. Full writeup in docs/security-analysis.md. Thanks to @Mann1ng for finding this and submitting the fix (#1).

Framework-agnostic core

• The default agent loop no longer requires LangChain/deepagents — it runs natively against the raw Anthropic or OpenAI SDK (pip install "text2sql-framework[anthropic]" or [openai]).
• OpenAI-compatible endpoints (OpenRouter, Ollama, vLLM, ...) now work via OPENAI_BASE_URL.
• LangChain/Deep Agents remains available as an opt-in backend (agent_backend="langchain", [langchain] extra), and Text2SqlMiddleware for LangChain's create_agent is unchanged.
• Public API (TextSQL, SQLResult, tracing) is unchanged.

Fixes

• [langchain] extra now installs its own provider integrations (langchain-anthropic, langchain-openai) instead of silently depending on them being pulled in transitively.
• Corrected Homepage/Repository package metadata (was pointing at a fork).

Docs

• Added install/quickstart section, badges, and a use-case issue template.