v0.2.0

Added

• @tool(confirm="…") / ToolSpec.confirm — an optional human-readable
  confirmation prompt for a destructive tool, stamped into the JSON Schema as
  the x-confirm extension (X_CONFIRM_KEY) for the frontend to display.
• DJANGO_AG_UI["API_KEY"] and ["PROVIDER"] — supply the model's API key (or
  a full Provider instance / dotted path) explicitly instead of inferring it
  from the environment, while keeping the built-in toolset wiring. MODEL may
  also be a pre-built Model instance.
• Provider extras django-ag-ui[anthropic], [openai], [google].
• Skills — SkillRegistry + SkillSpec (pre-defined prompts) and a
  read-only catalog endpoint mounted by get_urls(view, skills=registry) at
  <prefix>skills/, serving the JSON the web component consumes.
• DjangoAGUIView(require_authenticated=True) fails closed (401) for
  unauthenticated requests, and a get_user(request) hook establishes the user
  (assigned to request.user) before tools run — closing the "tools run as
  AnonymousUser" footgun. The contract is documented on the view.
• @tool(summary="…") / ToolSpec.summary → x-summary (X_SUMMARY_KEY): a
  short display label the frontend shows on the tool-call card.

Changed

• DEFAULT_SYSTEM_PROMPT now steers the model to call destructive tools
  directly and rely on the client's explicit confirmation step, instead of
  asking the user to confirm in prose.
• Dependency: pydantic-ai[ag-ui] → pydantic-ai-slim[ag-ui]. Drops the
  full meta-package's logfire / fastmcp / temporalio / otel footprint. Action
  required: install a provider — pip install django-ag-ui[anthropic] (or
  [openai] / [google], or the provider lib directly) — to use a
  "provider:model" MODEL string.

Notes

• The AG-UI endpoint now emits a one-time RuntimeWarning when served over WSGI
  (SSE can't stream there); deploy under ASGI (Daphne / Uvicorn).