fix(anthropic): support the Claude 4.7+ and 5 model generations

[sharkfabri]

Problem

The Anthropic plugin decides what to put in a request from hardcoded model ids, and Anthropic has changed the request contract twice since Claude 4.6. Every model released after that list was written now sends a request the API rejects.

1. Prefilling. _NO_PREFILL_PATTERNS lists claude-sonnet-4-6 and claude-opus-4-6 literally, so on Opus 4.7/4.8, Opus 5 and Sonnet 5 the plugin still sends a trailing assistant message and Anthropic 400 Error on Claude 4.6 - "Prefilling assistant messages is no longer supported" #4907 ("Prefilling assistant messages is no longer supported") comes back.

2. Sampling parameters. temperature, top_p and top_k are rejected with a 400 from Claude 4.7 onwards, so a documented constructor argument breaks every request:

   anthropic.LLM(model="claude-opus-5", temperature=0.7)

3. Thinking. On Sonnet 5 and Opus 5 thinking is on unless the request turns it off, and max_tokens caps thinking and the reply together — with the plugin's 1024 default a voice turn can spend most of its budget reasoning and come back truncated. On the way out, thinking_delta and signature_delta fall through _parse_event, so the reasoning is generated and billed but never surfaces anywhere.

Separately, LLMStream._run is re-entered on the same instance on every retry and nothing clears the state of the failed attempt.

Fix

Two commits.

reset per-attempt state before a stream retry — the framework only retries when the attempt emitted no chunk, which is exactly when a <thinking> block or a tool call was left open. _ignoring_cot left set makes the retry's answer get swallowed as chain of thought; a half-read tool call keeps collecting arguments from the next attempt. Both are cleared at the top of _run.

support the Claude 4.7+ and 5 generations — the hardcoded id list is replaced by the generation read out of the model id, and the three behaviours are gated on it:

	applies from	behaviour
prefilling	4.6	a trailing assistant message is closed with a user turn
sampling params	4.7	temperature/top_p/top_k are dropped, warned once per instance
thinking	4.6	thinking: {"type": "disabled"} is sent; extra_kwargs still wins

Version parsing anchors on the family name, so gateway and snapshot ids keep resolving to the underlying model: anthropic.claude-opus-5, claude-opus-4-5@20251101, claude-opus-5[1m], gw-1-claude-opus-5. When the family cannot be found in the id — a proxy alias such as my-claude-proxy — no version is returned and every guard keeps the behaviour it had before, rather than guessing at the model's capabilities.

thinking_delta and signature_delta are now recognised: the reasoning is kept out of the caller's text (a voice agent would speak it) and logged at debug level, with a one-time warning if thinking appears when it was not requested.

models.py gains Opus 4.5/4.7/4.8, Opus 5, Sonnet 4.5, Sonnet 5 and Haiku 4.5, and marks the ids that now return 404. claude-fable-5 and claude-mythos-5 are deliberately left out — thinking cannot be turned off on them and a single request can run for minutes, which does not fit a voice session — but the runtime guard still recognises them, since model accepts any string.

Behaviour change

Requests to 4.6+ models now carry thinking: {"type": "disabled"} where they previously carried nothing. That is a no-op on 4.6/4.7/4.8, where thinking was already off by default, and a real change on Sonnet 5 / Opus 5, where it was on. Callers who want the model to reason opt back in with chat(extra_kwargs={"thinking": {"type": "adaptive"}}); note that thinking blocks are not kept in the chat context, so a turn carrying a tool call is replayed without them, which Anthropic can reject — the plugin logs a warning when that combination is used.

Tests

tests/test_plugin_anthropic.py moves from the plugin category to unit. It is hermetic — fake client, no API key, no network — and no CI job runs --plugin, so nothing in it was running; the other hermetic plugin tests (cerebras, aws, openai, google) are already marked unit.

uv run pytest tests/test_plugin_anthropic.py --unit   # 94 passed
uv run pytest --unit                                  # 2055 passed, 5 skipped
make format-check && make lint
uv run mypy -p livekit.plugins.anthropic

Coverage: version parsing including gateway/snapshot/unparseable ids, the prefill guard as a regression test for #4907, sampling parameters from both the constructor and extra_kwargs, the thinking configuration per model family, max_tokens defaults, the stream-side thinking blocks, and the per-attempt state reset.

[CLAassistant]

All committers have signed the CLA.