fix(anthropic): recreate stream on retry

[he-yufeng]

Summary

• create a fresh Anthropic streaming request for each retry attempt
• avoid re-awaiting the same coroutine after a transient stream creation failure
• add a no-network regression test for the retry path

Fixes #5805.

Tests

• uv run pytest tests\test_plugin_anthropic.py -q
• uv run ruff check livekit-plugins\livekit-plugins-anthropic\livekit\plugins\anthropic\llm.py tests\test_plugin_anthropic.py
• python -m py_compile livekit-plugins\livekit-plugins-anthropic\livekit\plugins\anthropic\llm.py tests\test_plugin_anthropic.py
• git diff --check

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[devin-ai-integration]

Devin Review found 1 potential issue.

View 4 additional findings in Devin Review.

[devin-ai-integration]

🔴 Instance state not reset between retries in _run(), causing stale _ignoring_cot and _tool_call_id on retry

Now that retries actually work (the core purpose of this PR), _run() can be called multiple times by _main_task() (livekit-agents/livekit/agents/llm/llm.py:215-223). However, _run() does not reset instance state variables (_tool_call_id, _fnc_name, _fnc_raw_arguments, _ignoring_cot, _input_tokens, _output_tokens, etc.) at the start of each attempt. If a first attempt partially processes stream events before failing with a retryable error, the stale state carries into the retry.

Two concrete failure modes:

1. _ignoring_cot stuck True: If the first attempt receives a <thinking> text delta (line 367-368) before failing, _ignoring_cot stays True on retry, silently dropping all text content until a </thinking> tag appears — which may never happen in the retry's fresh response.
2. Stale _tool_call_id causes bogus tool calls: If the first attempt receives a content_block_start with type tool_use (lines 356-359) before failing, _tool_call_id remains set on retry. When the retry's text block hits content_block_stop (line 385), the if self._tool_call_id is not None check is True, emitting a spurious tool call with stale name/arguments.

Both of these are retryable scenarios (no ChatChunk was emitted, so retryable stays True). Other plugins like Mistral (livekit-plugins/livekit-plugins-mistralai/livekit/plugins/mistralai/llm.py:222-224) and OpenAI (livekit-plugins/livekit-plugins-openai/livekit/plugins/openai/responses/llm.py:402-403) reset their per-attempt state at the top of _run().

Suggested change

	async def _run(self) -> None:
	retryable = True
	try:
	if not self._anthropic_stream:
	self._anthropic_stream = await self._awaitable_anthropic_stream
	async with self._anthropic_stream as stream:
	async with await self._create_anthropic_stream() as stream:
	async def _run(self) -> None:
	# Reset per-attempt state so retries start clean
	self._tool_call_id = None
	self._fnc_name = None
	self._fnc_raw_arguments = None
	self._request_id = ""
	self._ignoring_cot = False
	self._input_tokens = 0
	self._cache_creation_tokens = 0
	self._cache_read_tokens = 0
	self._output_tokens = 0
	retryable = True
	try:
	async with await self._create_anthropic_stream() as stream:

---

Was this helpful? React with 👍 or 👎 to provide feedback.