ellmer 0.4.1

• ellmer is now instrumented with OpenTelemetry, so that traces are emitted whenever the (suggested) otel package is installed and a tracer is active.

  Each call to $chat(), $chat_async(), $stream(), or $stream_async() produces a top-level invoke_agent span that wraps one or more child chat <model> spans (one per request to the provider) and execute_tool <tool> spans (one per tool invocation). Chat spans record the provider name, request model, response model, and response id, plus input and output token usage; tool spans record the tool name, description, call id, and any error raised during execution. HTTP spans from httr2 are automatically nested under the chat spans (#526).

  Chat spans can additionally record conversation content as gen_ai.input.messages, gen_ai.output.messages, and gen_ai.system_instructions. This is opt-in via the OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT environment variable (set to "true"), since these payloads may contain user data.

• ellmer now distinguishes text content from thinking content while streaming, allowing downstream packages like shinychat to provide specific UI for thinking content (@simonpcouch, #909).

• claude_tool_web_search(), openai_tool_web_search(), and other built-in tools now include description and annotations properties, making their metadata consistent with user-defined tools created by tool() (#942).

• chat_anthropic() no longer fails when streaming web search results: citations_delta events are now handled correctly and server_tool_use input is parsed from JSON during streaming (#941).

• chat_anthropic() no longer applies a hidden 1.25x pricing weight to cache-creation tokens; the input token counts reported by token_usage() and parallel_chat() are now raw counts. Cost calculations are unchanged.

• chat_aws_bedrock() now supports reasoning/thinking content. To enable thinking in Anthropic Claude models, see the api_args argument in ?chat_aws_bedrock for an example (#964).

• chat_aws_bedrock() gains a cache parameter for prompt caching. The default, "auto", enables caching for models known to support it (Anthropic Claude and Amazon Nova) and disables it otherwise (#954).

• chat_openai_compatible() now extracts reasoning_content from model responses (both streaming and non-streaming) as ContentThinking objects. A new preserve_thinking parameter controls whether reasoning content is sent back to the API in multi-turn conversations; it defaults to FALSE (matching DeepSeek's requirement) but is set to TRUE for chat_openrouter() (#972).

• chat_databricks() (and other chat_openai_compatible() providers) no longer fail with HTTP 400 when the conversation history contains empty ContentText("") objects, which can occur during tool calling (@JamesHWade, #932).

• chat_github() now uses chat_openai_compatible() for improved compatibility and models_github() now supports custom base_url configuration (@D-M4rk, #877).

• chat_groq() now supports structured chat (@CoryMcCartan, #930).

• New chat_lmstudio() and models_lmstudio() provide support for LM Studio, a local model server with an OpenAI-compatible API (#963).

• chat_ollama() now supports params(top_k = ) (@frankiethull, #896).

• chat_openai() no longer fails when streaming web search results for web_search_call action types other than search (e.g. open_page, find_in_page) (#941).

• chat_openai() now uses the default prices if the service tier is missing (@trangdata, #903).

• chat_snowflake() now correctly handles tool calling. Previously, when Snowflake's streaming API sent a tool-use chunk as the very first response (with no preceding text), the chunk merging logic produced malformed content, causing "argument is of length zero" errors (#938).

• default_google_credentials() no longer skips application default credentials (e.g. GOOGLE_APPLICATION_CREDENTIALS) in interactive sessions, instead falling through to the OAuth browser flow only when no gargle token is available (@stefanlinner, #922).

• models_anthropic() (and models_claude()) gains a credentials argument for consistency with chat_anthropic() and other models_*() functions (@jcrodriguez1989, #917).

• New stream_controller() enables programmatic cancellation of streaming chat responses, e.g. from a Shiny "Cancel" button with chat$stream() or chat$stream_async(). Streaming turns are now saved incrementally so that partial responses survive cancellation, interrupts (Ctrl-C), and errors. Incomplete turns are recorded as AssistantPartialTurn objects, display as interrupted in the chat history, and are included in subsequent model context like complete turns (#643).