[kaxil]

pydantic-ai 2.43 and later print an ASCII-art banner to stderr on the first agent run in a process. pydantic-ai shows it when stderr is a terminal, or when an environment variable says a coding agent is running the process (CLAUDECODE, CODEX_*, CURSOR_AGENT, AI_AGENT and others). Task processes inherit the environment of whatever started Airflow. So anyone who runs airflow standalone, airflow dags test or breeze from Claude Code, Cursor or Codex gets the banner in every AgentOperator / LLMOperator task log. The supervisor logs task stderr at ERROR, so a run that succeeded shows about ten ERROR lines, and they come back after each deferral resume because each resume is a new process. This was seen while testing 1.0.0rc1 (#74741).

This sets pydantic_ai.BANNER_ENABLED = False when hooks/pydantic_ai.py is imported. Every agent the provider runs is built through PydanticAIHook.create_agent, so the operators, @task.agent / @task.llm and the model-backed retry policies all import it before an agent runs. BANNER_ENABLED is pydantic-ai's own switch for programs that want to own their output, and pydantic-ai reads it at display time. On pydantic-ai older than 2.43 the assignment adds an unused attribute, so the 2.36 floor is unaffected.

Why not set PYDANTIC_AI_NO_BANNER from the provider? It does the same thing, but writing to os.environ would leak into user code and every subprocess a task starts. Filtering one library's stderr in core would be the wrong layer.

A task that builds its own Agent(...) and never imports the hook isn't covered. The "run your own agent" guide now tells those users to set PYDANTIC_AI_NO_BANNER=1 in the worker environment.

The test sets CLAUDECODE=1, clears PYTEST_VERSION and CI (pydantic-ai suppresses the banner under both on its own), and resets pydantic-ai's once-per-process claim. Without the hook change it fails with the banner text on stderr.

Before: