SDK follow-ups from 2026-10-05 dependency update #858

Follow-up SDK work found while reviewing the release notes for the 2026-10-05 weekly dependency update (#849). Every bump replays cleanly; none of these block that PR.

Integration changes (highest value first)

• OpenAI: Agents session stream get_final_result() / with_result_collection() skip tracing #859 OpenAI: Agents session stream helpers skip tracing (bug). openai 3.23 added AgentSessionEventStream.get_final_result() and with_result_collection(). Our _TracedStream / _AsyncTracedStream in py/src/braintrust/integrations/openai/tracing.py passes unknown attributes through to the real stream, so calling either method skips _AgentSessionTrace.observe() / finish(). The openai.agents.sessions.create span then gets no output or metrics and is never closed. Fix: drain the traced generator before calling the real stream's get_final_result(), and have with_result_collection() return the traced wrapper. Needs a new cassette-backed test.
• OpenRouter: GA client.responses is not traced by wrap_openrouter(), and client.beta.responses is not auto-instrumented #860 OpenRouter: GA client.responses is never traced. wrap_openrouter() in py/src/braintrust/integrations/openrouter/tracing.py only wraps client.beta.responses, which is now a separate, deprecated BetaResponses class. auto_instrument() covers the GA client.responses but not BetaResponses. Fix: patch BetaResponses.send/send_async only when that module exists, and make wrap_openrouter wrap client.responses. Add cassette-backed coverage, including a Responses check in test_auto_openrouter.py.
• Anthropic: capture prompt cache diagnostics (cache_miss_reason, cache_missed_input_tokens) #861 Anthropic: capture cache diagnostics (GA in 1.9.0). Add diagnostics to METADATA_PARAMS in py/src/braintrust/integrations/anthropic/tracing.py. Log message.diagnostics.cache_miss_reason.type as metadata and cache_missed_input_tokens as a metric (e.g. prompt_cache_missed_tokens). Read the fields with getattr so older versions need no gating; this also covers BetaMessage.diagnostics.
• agno: workflow StepProgress event text leaks into workflow span output #862 agno: keep StepProgress text out of workflow output (3.1.0). _aggregate_workflow_chunks in py/src/braintrust/integrations/agno/tracing.py appends .content from every non-WorkflowCompleted chunk, so text from the new StepProgress events will end up in the span output. Skip event == "StepProgress" there, optionally record progress in metadata, and add a focused test.
• google-adk: instrument Workflow nodes and workflow tool calls #863 google-adk: no spans for Workflow nodes and their tool calls (predates this bump). ADK 2.x Workflow is a BaseNode, not a BaseAgent, and _ToolNode calls tool.run_async directly, so workflow tool calls get no tool span (MCP tools excepted). ADK now steers users toward Workflow over SequentialAgent / ParallelAgent. Candidate patch targets: google.adk.workflow._base_node.BaseNode.run and _tool_node._ToolNode._run_impl.

Lower priority

• [bot] Google GenAI: streaming responses drop url_context_metadata that non-streaming responses preserve #774 google-genai: _aggregate_generate_content_chunks drops the new Candidate.continuation_token (2.26+). Add status and continuation_token to _extract_interaction_metadata. Stop reading the deprecated response_mime_type / response_modalities fields, which raise DeprecationWarnings. Folded into [bot] Google GenAI: streaming responses drop url_context_metadata that non-streaming responses preserve #774, which has the same root cause (url_context_metadata dropped by the same allowlist).
• pydantic-ai: capture web-search and grounding counts from RequestUsage.details #866 pydantic-ai: capture the new web-search and grounding counts in RequestUsage.details (2.52–2.54) as metrics.
• Temporal: add workflow.Info.original_execution_run_id to workflow span metadata #868 temporal: add workflow.Info.original_execution_run_id (1.34.0) to workflow span metadata when available.
• Mistral: record usage.service_tier as metadata #867 mistral: optionally record usage.service_tier as metadata.
• OpenAI: trace computer-use Agents session items and capture Responses access_programs #869 openai: optionally add computer-use session items (computer_use_call, 3.22) to _AGENT_TOOL_ITEM_INPUT_KEYS and capture access_programs (3.20) in Responses metadata.

Cassettes that couldn't be re-recorded

These kept their previous latest recordings in #849 because a fresh recording isn't possible right now:

• Anthropic: test_anthropic_beta_messages_create_captures_compaction_metadata can't be re-recorded #864 anthropic, test_anthropic_beta_messages_create_captures_compaction_metadata: the live API now stops with max_tokens instead of compaction for that prompt, so I kept the old cassette. The test prompt or assertion needs updating before it can be re-recorded.
• LiteLLM: test_litellm_text_completion_metrics[*] can't be re-recorded (model retired) #865 litellm, test_litellm_text_completion_metrics[*]: the model it calls (gpt-3.5-turbo-instruct) has been retired. Point the test at a supported model or retire it.
• mistral 3.0.0: the chat, agents and OCR latest cassettes are still the 2.10.1 recordings, because the API key kept returning 429 (rate limited). Re-record them once the quota clears.

Nothing worth doing

openai-agents 0.23, ai-sdk 0.8, huggingface-hub 2.1, transformers 5.18, harbor 0.24. litellm 1.104.0's Windows wheel still ships CRLF tiktoken files, so the ignore_hosts workaround in test_litellm.py stays.

🤖 Generated with Claude Code