v1.0.21

Fixed

• SootheConfig now exposes a secret_resolver property (resolving ${ENV_VAR} via the shared _resolve_env helper), so MCPRegistry construction in the daemon (server/core.py) and the per-thread runner (runner/_thread_manager.py) no longer raises AttributeError: 'SootheConfig' object has no attribute 'secret_resolver' — the error that left the MCP registry empty and gave the agent zero tools.
• fetch_tools_and_descriptors is now async and awaits client.get_tools(...) when it returns a coroutine (langchain-mcp-adapters 0.3.x), instead of iterating the coroutine as a list. Previously every MCP server logged 'coroutine' object is not iterable, the coroutine was never awaited, and zero tools loaded despite Initialized: N connected, 0 failed.

Changed

• Upgraded langchain ecosystem pins to latest stable: langchain 1.4.4, langchain-core 1.6.9, langchain-mcp-adapters 0.3.2 (pin widened from <0.3.1 to <0.4.0 — 0.3.0 caused the async get_tools regression in #22), langchain-openai 1.7.0 (0.x→1.x), langchain-community 0.4.2, langgraph 1.2.14, mcp 1.30. langchain-typesafe[experimental] bumped to >=0.0.1a3.
• Adapted the three TypeSafe decision modules (risk_classifier, typesafe_intent, typesafe_eval_decision) to langchain-typesafe 0.0.1a3: TypeSafeClassifier no longer binds questions at construction — they travel per-request via ClassifierRequest — and the response field is .answers (was .choices). build_typesafe_classifier drops the questions param.
• Bump soothe-nano floor to 1.3.3 (pulls in the secret_resolver property + async get_tools fixes for #22).

Full changelog: v1.0.20...v1.0.21