v0.18.0

Added

• [harness] extra + a CodeMode batching recipe. Optional
  pip install django-ag-ui[harness] pulls pydantic-ai-harness (and its
  pydantic-monty sandbox) — lazy-imported, so the core install stays django +
  pydantic-ai-slim. Its CodeMode capability drops into the existing
  CAPABILITIES seam to collapse many tools (notably a large drf-mcp bridge) into
  one sandboxed run_code tool the model batches in a single round-trip. New
  CodeMode guide.

Changed

• The drf-mcp bridge now carries each tool's outputSchema onto
  ToolDefinition.return_schema. drf-mcp advertises an output schema by default
  (INCLUDE_OUTPUT_SCHEMA); the bridge previously dropped it. Propagating it means
  the tool's return type reaches the model — chiefly so CodeMode renders each
  bridged tool as a typed Python stub (-> <Model>) instead of -> Any. A
  service with no output serializer advertises no schema, so its stub stays
  untyped (unchanged). No effect on tool dispatch or results.

Documentation

• New "Tool approval (human-in-the-loop)" guide documenting the end-to-end
  approval flow: enabling TOOL_GUARD, what counts as destructive, what the user
  sees (approve / deny / resume), how a custom (non-web-component) client drives
  the interrupt/resume loop, and the ask_user typed-question tool. Cross-linked
  from the TOOL_GUARD configuration section.