[bot] OpenAI and Anthropic generation parameters not captured in span metadata

从 braintrust-sdk/src/main/java/dev/braintrust/instrumentation/InstrumentationSemConv.java 中的 tagOpenAIRequest 和 tagAnthropicRequest 开始，然后比较 Google GenAI 的 BraintrustApiClient.java handler 中的参数提取方式。检查现有的 OpenAI 和 Anthropic 测试文件，并为列出的生成参数和提供商特定配置添加覆盖。完成的标准是：一致地捕获受支持的请求元数据，并且测试能够对此进行断言。

Summary

The OpenAI and Anthropic request taggers in InstrumentationSemConv only extract model and request routing info (request_path, request_base_uri, request_method) into span metadata. Generation parameters like temperature, max_tokens, tools, response_format, and provider-specific config like Anthropic's thinking are silently dropped.

In contrast, the Google GenAI handler in the same repo (BraintrustApiClient.tagSpan()) extracts temperature, topP, topK, maxOutputTokens, tools, toolConfig, safetySettings, responseMimeType, responseSchema, and more into metadata. This is an inconsistency within the repo — the Google GenAI handler provides materially more instrumentation detail for the same class of information.

What is missing

OpenAI — tagOpenAIRequest() (lines 78–108)

Currently captures only model in metadata. The following request parameters are silently dropped:

Field	Purpose
temperature	Sampling temperature
max_tokens / max_completion_tokens	Output length limit
top_p	Nucleus sampling
frequency_penalty, presence_penalty	Repetition control
tools	Tool/function definitions
response_format	Structured output config (JSON mode, JSON Schema)
reasoning_effort	Reasoning effort for o-series models
logprobs, top_logprobs	Log probability settings
stop	Stop sequences

For the Responses API, additional fields are missing: instructions, tools (with web_search, file_search, code_interpreter configs), reasoning (with effort and summary).

Anthropic — tagAnthropicRequest() (lines 163–205)

Currently captures only model in metadata. The following request parameters are silently dropped:

Field	Purpose
max_tokens	Output length limit (required parameter)
temperature	Sampling temperature
top_p, top_k	Sampling parameters
tools	Tool definitions
thinking	Extended thinking config (type, budget_tokens)
stop_sequences	Stop sequences
metadata	User-provided metadata (e.g., user_id)

The thinking config is particularly important: when a user enables extended thinking with a specific budget, this information is lost from the span. The Python SDK equivalent (issue braintrustdata/braintrust-sdk-python#107, now closed) captures thinking metadata.

Google GenAI — already captures these (for comparison)

BraintrustApiClient.tagSpan() (lines 64–95) extracts into metadata:

• systemInstruction, tools, toolConfig, safetySettings, cachedContent
• From generationConfig: temperature, topP, topK, candidateCount, maxOutputTokens, stopSequences, responseMimeType, responseSchema

Impact

Users viewing traces in Braintrust can see generation parameters for Google GenAI calls but not for OpenAI or Anthropic calls. This makes it impossible to understand model configuration from the trace alone for the two most commonly used providers.

Braintrust docs status

• Braintrust tracing docs at https://www.braintrust.dev/docs/guides/tracing state that LLM spans show "the model, messages, parameters, token usage, and cost" — supported (parameters are documented as captured, but not implemented for OpenAI/Anthropic in this Java SDK)
• The Braintrust OpenAI docs mention temperature handling for GPT-5 models, implying it is a tracked parameter: unclear

Upstream sources

• OpenAI Chat Completions API: https://platform.openai.com/docs/api-reference/chat/create — documents all request parameters including temperature, max_tokens, tools, response_format, reasoning_effort
• OpenAI Responses API: https://platform.openai.com/docs/api-reference/responses/create — documents instructions, tools, reasoning
• Anthropic Messages API: https://docs.anthropic.com/en/api/messages — documents max_tokens, temperature, tools, thinking, top_p, top_k, stop_sequences
• Anthropic extended thinking: https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking — documents thinking parameter with type and budget_tokens

Local files inspected

• braintrust-sdk/src/main/java/dev/braintrust/instrumentation/InstrumentationSemConv.java — lines 78–108 (tagOpenAIRequest: only model extracted), lines 163–205 (tagAnthropicRequest: only model extracted)
• braintrust-sdk/instrumentation/genai_1_18_0/src/main/java/com/google/genai/BraintrustApiClient.java — lines 64–95 (tagSpan: comprehensive parameter extraction into metadata)
• braintrust-sdk/instrumentation/openai_2_8_0/src/test/java/dev/braintrust/instrumentation/openai/v2_8_0/BraintrustOpenAITest.java — no test asserts generation parameters in metadata
• braintrust-sdk/instrumentation/anthropic_2_2_0/src/test/java/dev/braintrust/instrumentation/anthropic/v2_2_0/BraintrustAnthropicTest.java — no test asserts generation parameters in metadata