LIA

🔌 User MCP tools now work in ReAct mode, exactly like the pipeline

Tool-resolution parity (fix). ReAct built its tool palette from the per-request manifests but resolved tool instances only through the process-global registry — so user MCP tools (held in the per-request user_mcp_tools_ctx ContextVar) were silently dropped and the loop fell back to a generic web search. A new shared tool_resolution.py (global registry → hallucinated-suffix strip → user MCP ContextVar) is now the single source of truth used by ReAct, the pipeline executor (3 inline fallbacks unified) and display-metadata lookup. HITL is read from the in-hand manifest so user MCP mutation tools keep their approval gate.

Iterative expansion (feature). Because the ReAct loop is itself iterative, iterative user MCP servers are expanded to their individual tools (the model picks them by description) instead of the opaque per-server task tool — except MCP App servers (e.g. Excalidraw), which keep their dedicated agent. Gated by REACT_MCP_EXPAND_ITERATIVE_ENABLED (default true, instant rollback). The pipeline is unchanged.

Strict-server reliability (fix). Optional MCP parameters left unset were forwarded as null and rejected by strictly-typed (Go) servers (parameter X is not of type string, is <nil>). Both adapters now drop None-valued arguments before the call.

No schema change, no migration. New env var REACT_MCP_EXPAND_ITERATIVE_ENABLED (default true).

📚 Docs & public surfaces refreshed

README, /how & /why guides (6 languages), FAQ changelog (6 languages) and the landing "Technical" section refreshed: LLM model catalogue aligned to the pricing seed (DeepSeek V4, Anthropic Opus 4.6/4.5 · Sonnet 4.6 · Haiku 4.5, Gemini 3.1/3, GPT-5.4, Qwen3.5), version → 1.20.20, metrics (81 ADRs, 484 test files, ~10,000 tests, 280+ docs).

Tests: TDD (5 new test files); Ruff / Black / MyPy strict clean; 8438 unit tests green; frontend tsc + eslint clean; i18n key parity preserved. See ADR-070 (amendment 2026-06-02).