v0.44.0

Changed

• Vendors @artooi/ag-ui-web-component 0.39.0, and renders the signed-in
  user's name as user-name. Under PLACEMENT="page", an empty conversation
  now greets the user as Hello, {name} with the composer centred beneath it,
  and the composer slides to its usual place when the first message is sent.
  The name is the one the admin's own header welcomes them by --
  get_short_name(), else get_username() -- so the page never greets
  somebody differently from the header above it. It is presentation only: the
  component never sends it and scopes nothing by it, unlike user-key. The
  default floating panel, "sidebar" and "embedded" show no greeting, so for
  them the attribute changes nothing.

  What else 0.39.0 changes in the admin, with nothing to configure:

  • Every tool call gets an answer before the next request. A call cut off by
    Stop, or naming a tool the page does not have, used to go out with no
    result, which some model providers refuse; it is now answered as not
    finished, and its card says so rather than claiming it was done. (A call
    cut off by a run error was already answered by django-ag-ui.)
  • A reload while a tool call waits for approval restores that card as not
    finished, where it showed a spinner that never stopped.
  • The conversation list slides over the transcript at every width, where a
    wide "page" panel used to dock it beside the transcript and push the
    conversation sideways.
  • Under "sidebar" and "page", the corners inside the panel stay round
    while its frame stays square.
  • Focus moves into the composer on New chat, and into the conversation list
    or the checkpoints panel when either opens.

  A project overriding noResult in STRINGS should move that text to
  callNotFinished. 0.39.0 no longer draws noResult, so the override is
  silently unused; toolInterrupted and interruptedLabel label the new card
  status.

• django-ag-ui is floored at >=0.60 (was >=0.59) and
  django-pydantic-agent at >=0.23 (was >=0.21), so a refused tool from
  AdminAgentServer(drf_mcp_server=...) renders as failed. That server's tools
  reach the agent through django-pydantic-agent's in-process bridge, and below
  0.23 the bridge returned a refused call as the tool's value, which pydantic-ai
  records as a success: TOOL_CALL_RESULT carried no outcome, and the vendored
  card settled the call to done. 0.23 raises ToolFailed with the server's
  sentence and the refusal's code, so the card shows it failed. django-ag-ui
  0.60 floors django-pydantic-agent at 0.23 and holds this with a test. The
  configuration page now says that the failed call's text is the server's own
  and that TOOL_FAILURE.INCLUDE_DETAIL does not redact it, where its note on
  refusal reasons had said every non-permission refusal withholds its reason by
  default. The floor stated in README.md, CLAUDE.md, docs/installation.md
  and the upstream-drift workflow moves with it.

• The [mcp] extra is floored at djangorestframework-mcp-server>=0.45 (was
  >=0.37). Below 0.44, once a bridged server's specs declared drf-services
  affordances, a chain tool ran a step whose service's own affordances refuse it
  and reported success, and the outputSchema handed to the model as its return
  schema left out the affordances object each rendered item carries. 0.45
  serves a refusal's code, which the bridge appends to the failed call's text,
  and resolves a chain tool's RETRIEVE step to its row before the object-level
  permission judges it: below it a selector returning a queryset skipped that
  check, and a chain step rendered without an output serializer answered with
  the text of the row the rule refuses. It also refuses a many=True service
  spec at registration. 0.45 floors djangorestframework-services at 0.52.1 in
  turn. The floor stated in CLAUDE.md, docs/installation.md and the
  upstream-drift workflow moves with it.