LLMAgent.cleanup() previously iterated over hard-coded provider SDK client
attributes (root_async_client, root_client, _async_client, _client) and
called close() on each. Recent langchain-anthropic and langchain-openai
versions cache their default httpx clients via module-level lru_cache
functions. Closing those clients poisoned the cache: every subsequent
ChatAnthropic/ChatOpenAI instance in the same process received the same
closed httpx client and failed with a connection error.
Fix: cleanup() now only releases the agent's own reference to the chat
model (self._chat_model = None). The removed _KNOWN_CLIENT_ATTRS class
variable has been deleted and ClassVar removed from the typing import.
The concurrent-idempotency guarantee is preserved: the lock is still
acquired before nulling _chat_model, so two concurrent cleanup() calls
cannot both see a non-None model and attempt conflicting operations.
The four provider SDK client-closing scenarios in credential_injection.feature
and llm_missing_coverage.feature are removed as they tested the old (buggy)
behaviour. Their step definitions are removed from credential_cleanup_steps.py
(now only carries the resolve_class_ref patch step) and
llm_missing_coverage_steps.py is updated with the corrected assertions.
Six new regression BDD scenarios tagged @tdd_issue @tdd_issue_57 are added
in features/llm_cleanup_shared_client.feature, covering all four provider
SDK client attribute paths (Anthropic _async_client/_client, OpenAI
root_async_client/root_client) and two end-to-end two-agent scenarios that
prove a second agent can run successfully after the first is cleaned up.
ISSUES CLOSED: #57
hurui200320