[johnnichev]

Closes #182 (items 1, 2, 3 and the loop test; item 5, a native Responses provider, is left for its own release).

What changed

StreamEnd (selectools.types, @beta): a provider may yield one as the last item of stream() / astream(). The agent consumes it in all three stream consumers (_streaming_call, _astreaming_call, the native path in Agent.astream):

• usage goes into agent.usage, to on_llm_end / on_usage, and onto the LLM_CALL trace step. Streamed turns used to hit on_llm_end with None and never counted.
• provider_state is copied onto the assistant Message the agent assembles from chunks.

Consumers that pattern-match on str / ToolCall ignore it. Fallback and Router re-yield it untouched. The OpenAI-compatible provider now emits it from the usage chunk it already requested on astream(), and requests it on stream() too. A usage object only counts if it has an integer prompt_tokens, so usage: null from compatible servers never produces a bogus item.

Message.provider_state: Optional[Dict[str, Any]]: opaque, JSON-safe. Carried through the tool loop (the tool-call turn is appended as-is; the final no-tool turn now copies it when the message is rebuilt), ConversationMemory._copy_message (deep copy), to_dict() / from_dict(), and the cache key (only keyed when set, so existing cache entries stay valid).

temperature: Optional[float] on AgentConfig, the Provider protocol, and every built-in provider. None means omit; the OpenAI-compatible and Anthropic request builders skip the field, Gemini already tolerated None. Default stays 0.0, so nobody's output changes. 0.0 is still sent (zero/falsy trap covered by a test).

Tests

• tests/core/test_provider_state.py: field default, dict round trip, memory deep copy, cache key.
• tests/agent/test_stream_end.py: usage + observers + trace on run, arun, astream; provider_state attachment; legacy providers unchanged.
• tests/providers/test_openai_stream_end.py: mocked Chat Completions stream, sync and async, cached tokens, no-usage case, ordering after tool calls.
• tests/providers/test_optional_temperature.py: agent boundary, cache key, OpenAI (4 methods), Anthropic, Gemini.
• tests/agent/test_provider_state_roundtrip.py: the loop test from the issue. A fake Responses-style provider returns reasoning items in provider_state with a function call; after the tool runs, the second provider call sees the assistant message with its state intact, on run, streamed run, arun, and astream.
• One existing test updated: test_astreaming_call_sync_fallback unpacked the helper's old two-tuple.

Full non-e2e suite: 7551 passed, 66 skipped. mypy, ruff, bandit clean.

Notes for review

• Streamed turns on run() / arun() now record an LLM_CALL trace step; previously only astream() did. Listed under Changed in the changelog.
• No version bump here; left for the release commit.
• Docs: PROVIDERS.md (protocol + new section), STREAMING.md, changelog, and examples/115_stream_end_provider.py (offline).