Replace the hand-rolled push/pop contextvar stack with _run_context.py: one AgentRun dataclass on a token-reset contextvar, managed by a single agent_run_scope() context manager so pairing no longer relies on manual try/finally in three places. The streaming wrapper's manual scope/span/stack bookkeeping collapses into one ExitStack, which also stops leaking the isolation scope and invoke_agent span when entering the wrapped context manager fails, and the token reset tolerates streaming runs that exit in a different asyncio task.

The request hooks deliberately keep pairing chat spans through RunContext.metadata: contextvar pairing can mispair overlapping run_stream context managers in the same task, while the metadata dict is correct per run by construction. Review-driven hardening folded in: hook bodies are wrapped in capture_internal_exceptions so span-data failures cannot abort the user's model request, invoke_agent spans now receive model data from the agent argument instead of the not-yet-pushed contextvar, the model-settings mapping has a single source of truth in _extract.py, and the usage extractor passes record_token_usage kwargs directly.