v1.4.0 — StreamEnd, Provider State & New SDK Majors

Streamed turns now report usage and can carry opaque provider state for
verbatim replay (#182); temperature becomes optional for reasoning models;
and the supported SDK ranges widen to openai 2.x/3.x and anthropic 1.x.

Added

• StreamEnd: usage and provider state at the end of a native stream
  (#182). Providers may yield one selectools.types.StreamEnd as the last
  item of stream() / astream(). Its usage is counted into agent.usage,
  passed to on_llm_end / on_usage, and recorded on the LLM_CALL trace
  step; before this, every streamed turn reached on_llm_end with None
  usage and was missing from the totals. Its provider_state is copied onto
  the assistant Message the agent assembles from chunks. The agent consumes
  the item; consumers that pattern-match on str / ToolCall are unaffected.
  The OpenAI provider now emits it from the usage chunk it already requested
  on astream(), and requests it on stream() too.
• Message.provider_state — an opaque, JSON-safe dict a provider attaches
  to an assistant message so it can replay that turn verbatim on the next
  request (stateless Responses-style APIs with encrypted reasoning items).
  Carried through the tool loop, ConversationMemory copies,
  to_dict() / from_dict(), and the cache key. Replaces smuggling such
  state through ToolCall.thought_signature.

Changed

• temperature is Optional[float] on AgentConfig and every provider
  method. None means "send no temperature", which reasoning models
  require; the OpenAI-compatible and Anthropic providers omit the field
  instead of relying on the 400-and-retry fallback. The default stays 0.0,
  so no existing output changes.
• Streamed turns on run() / arun() now record an LLM_CALL trace step
  (previously only astream() did).
• Supported SDK ranges are now openai>=1.55.3,<4 and
  anthropic>=0.40.0,<2 (previously >=1.30,<2 and >=0.28,<1). The new
  majors are drop-in for the surface selectools uses (chat.completions,
  embeddings, images, messages); their breaking changes are the httpx
  to httpx2 swap (only matters if you hand httpx objects to the SDK
  clients) and, for anthropic 1.0, the removal of temperature / top_p /
  top_k from the messages.create() signature. The floors were raised
  because the old ones never resolved to a working install: with httpx 0.28
  (which a fresh resolve picks) openai<1.55.3 and anthropic<0.40 fail at
  client construction with unexpected keyword argument 'proxies'.
  anthropic 2.x stays unsupported: it sends API version 2026-10-21, which
  returns 400 for any temperature (even the default), and the provider
  sends temperature=0.0 by default.
• The Anthropic provider now sends temperature through extra_body instead
  of as a keyword argument, so a non-None temperature no longer raises
  TypeError (surfaced as ProviderError) on anthropic >= 1.0. The request
  JSON is identical on every supported SDK version. New wire-level tests
  (tests/providers/test_sdk_compat.py, from #184) drive the real SDK
  clients through a mock HTTP transport so signature drift like this fails
  CI, and pin the Anthropic API version the installed SDK sends.

---

pip install -U selectools==1.4.0